Systems and methods for generative design of custom biologics

WO2024216084A8PCT designated stage expired Publication Date: 2024-11-14PYTHIA LABS INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2024/024344
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-11-03
Filing Date
2024-04-12
Publication Date
2024-11-14

AI Technical Summary

Technical Problem

The complexity of biologics poses challenges for computational tools in accurately predicting their properties and molecular behavior, hindering the development of new drug candidates, which requires significant time and capital.

Method used

The development of generative machine learning models that create custom peptide backbones tailored to exhibit desired properties and bind to specific targets by leveraging existing protein and peptide structures, using a processor to generate de-novo peptide backbones through iterative updates of feature vectors based on velocity fields.

Benefits of technology

This approach enables the design of custom biologics with enhanced performance for target binding, potentially reducing the time and cost of drug development by generating accurate peptide backbones and sequences for improved drug candidates.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2024024344_14112024_PF_FP_ABST
    Figure US2024024344_14112024_PF_FP_ABST
Patent Text Reader

Abstract

Presented herein are systems and methods for generative design of custom biologics. In particular, in certain embodiments, generative biologic design technologies of the present disclosure utilize a machine learning models to create custom (e.g., de-novo) peptide backbones that, among other things, can be tailored to exhibit desired properties and / or bind to specified target molecules, such as other proteins (e.g., receptors). Generative machine learning models described herein may be trained on, and accordingly leverage, a vast landscape of existing protein and peptide structures. Once trained, however, these generative models may create wholly new (de-novo) custom peptide backbones that are expressly tailored to particular targets. These generated custom peptide backbones can, e.g., subsequently, be populated with amino acid sequences to generate final custom biologics providing enhanced performance for binding to desired targets.
Need to check novelty before this filing date? Find Prior Art

Description

Attorney Docket No.2013969-0048 SYSTEMS AND METHODS FOR GENERATIVE DESIGN OF CUSTOM BIOLOGICS CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to U.S. Provisional Application No.63 / 596,216, filed November 3, 2023, U.S. Provisional Application No.63 / 460,985, filed April 21, 2023, and U.S. Provisional Application No.63 / 459,124, filed April 13, 2023, the contents of each of which are incorporated herein by reference in their entirety. BACKGROUND

[0002] An increasing number of important drugs and vaccines are complex biomolecules referred to as biologics. For example, seven of the top ten best selling drugs as of early 2020 were biologics, including the monoclonal antibody adalimumab (Humira®). Biologics have much more complex structure than traditional small molecule drugs. The process of drug discovery, drug development, and clinical trials requires an enormous amount of capital and time. Typically, new drug candidates undergo in vitro testing, in vivo testing, then clinical trials prior to approval.

[0003] Software tools for in-silico design and testing of new drug candidates can cut the cost and time of the preclinical pipeline. However, biologics often have hard-to-predict properties and molecular behavior. To date, despite recent interest in software and computational tools (including artificial intelligence (AI) and machine learning) relating to biological molecules, the extraordinary complexity of biologics continues to challenge computational tools aiming to produce accurate predictions for biologics and advances are needed. SUMMARY

[0004] Presented herein are systems and methods for generative design of custom biologics. In particular, in certain embodiments, generative biologic design technologies of the present disclosure utilize a machine learning models to create custom (e.g., de-novo) peptide - 1 - 11901435v1Attorney Docket No.2013969-0048 backbones that, among other things, can be tailored to exhibit desired properties and / or bind to specified target molecules, such as other proteins (e.g., receptors). Generative machine learning models described herein may be trained on, and accordingly leverage, a vast landscape of existing protein and peptide structures. Once trained, however, these generative models may create wholly new (de-novo) custom peptide backbones that are expressly tailored to particular targets. These generated custom peptide backbones can, e.g., subsequently, be populated with amino acid sequences to generate final custom biologics providing enhanced performance for binding to desired targets.

[0005] In one aspect, the invention is directed to a method for computer generation of a de-novo peptide backbone of a custom biologic, the method comprising: (a) receiving and / or generating, by a processor of a computing device a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular backbone site and comprising position and / or orientation components representing a position and / or orientation of the particular backbone site [e.g., wherein values of the position and / or orientation components of each feature vector are set to initial starting values (e.g., selected at random from an initial probability (e.g., uniform) distribution)]; (b) determining, by the processor, using a machine learning model, one or more (e.g., a plurality of) velocity fields and, beginning with the seed set, updating (e.g., iteratively), by the processor, values of the position and / or orientation components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving the values of the position and / or orientation components of the plurality of feature vectors from a set of initial starting values into a set of final values representing positions and / or orientations of each backbone site of a generated peptide backbone; and (c) creating, by the processor, using, the set of final values, a generated scaffold model representing the de-novo peptide backbone [e.g., using the set of final values as the generated scaffold model representing the de-novo peptide backbone; e.g., determining, based on the set of final values, positions of backbone atoms (e.g., at each backbone site, a position of one or more backbone atoms (e.g., an N, alpha-Carbon, C, and O)) to be represented via the generated scaffold model]; and (d) storing and / or providing (e.g., for display and / or further processing), the generated scaffold model.

[0006] In certain embodiments, for each particular feature vector of the plurality of feature vectors, the position and / or orientation components of the particular feature vector represent a relative translation and / or rotation, respectively, of the corresponding backbone site - 2 - 11901435v1Attorney Docket No.2013969-0048 relative to another, neighbor backbone site corresponding to another feature vector in the seed set.

[0007] In certain embodiments, for each particular feature vector of the plurality of feature vectors, the position and / or orientation components of the particular feature vector represent a global translation and / or rotation, respectively, with respect to a single common reference frame (e.g., origin) [e.g., and wherein the machine learning model is equivariant with respect to three-dimensional (3D) translations and / or rotations].

[0008] In certain embodiments, values of the position and / or orientation components of each feature vector are set to initial starting values selected at random from an initial probability (e.g., uniform) distribution.

[0009] In certain embodiments, the method comprises generating, by the processor, using a preliminary machine learning model to determine a plurality of sets of initial starting values, each set of initial starting values comprising an initial position and / or orientation and, at step (a), for each feature vector of the plurality of feature vectors selecting and using a set of initial starting values as values for the position and / or orientation components of the feature vector.

[0010] In certain embodiments, the seed set comprises 50 or more feature vectors {e.g., wherein the seed set comprises 100 or more feature vectors [e.g., wherein the seed set comprises 200 or more feature vectors (e.g., wherein the seed set comprises 500 or more feature vectors)]}.

[0011] In certain embodiments, step (b) comprises determining the one or more velocity fields and updating the values of the position and / or orientation components of the plurality of feature vectors in an iterative fashion [e.g., wherein each of the one or more velocity fields is an initial velocity field associated with and determined at a first iteration and / or a current velocity field associated with and determined at a particular subsequent iteration].

[0012] In certain embodiments, determining the one or more velocity fields and updating the values of the position and / or orientation components of the plurality of feature vectors in an iterative fashion comprises: determining initial starting values for the position and / or orientation components of each of the plurality of feature vectors; beginning with the initial starting values of the position and / or orientation components of the plurality of feature vectors, at a first iteration: determining, by the processor, using the machine learning model, an initial velocity - 3 - 11901435v1Attorney Docket No.2013969-0048 field based on the initial starting values of the position and / or orientation components of the plurality of feature vectors; and updating values of the position and / or orientation components of the plurality of feature vectors according to the initial velocity field; and at subsequent iterations: using the updated values from a prior iteration as current values of the position and / or orientation components of the feature vectors; determining, using the machine learning model, a current velocity field based on the current values of the position and / or orientation components of the feature vectors; and updating values of the position and / or orientation components of the feature vectors according to the current velocity field; and upon reaching a final iteration, using the updated values of the position and / or orientation components of the feature vectors as the final set of values for creating the generated scaffold model representing the de-novo peptide backbone.

[0013] In certain embodiments, initial starting values of the position and / or orientation components of each feature vector are selected at random from an initial probability (e.g., uniform) distribution.

[0014] In certain embodiments, the method comprises determining, by the processor, the set of initial starting values using a preliminary machine learning model (e.g., a variational autoencoder).

[0015] In certain embodiments, each iteration corresponds to one of a plurality of time- points [e.g., separated according to a (e.g., pre-defined) time-step], and (i) the initial velocity field is determined based (e.g., further) at least in part on an initial time-point corresponding to the first iteration; and / or (ii) each current velocity field associated with and determined at a particular subsequent iteration is determined based (e.g., further) at least in part on a particular subsequent time-point corresponding to the subsequent iteration.

[0016] In certain embodiments, at each particular iteration, the machine learning model receives, as input, (i) the current values of the feature vectors and (ii) the time point corresponding to the particular iteration, and generates, as output, the current velocity field.

[0017] In certain embodiments, at each particular iteration, the machine learning model receives, as input, (i) the current values of the feature vectors and (ii) the time point corresponding to the particular iteration, and generates, as output, a set of prospective final values of the position and / or orientation components of the feature vectors (e.g., representing a - 4 - 11901435v1Attorney Docket No.2013969-0048 current estimate of positions and / or orientations of each backbone site of a generated peptide backbone) and the current velocity field is determined based on the set of prospective final values and the time point corresponding to the current iteration (e.g., generating a current estimate of x1as output of the machine learning model and computing the current velocity field based on x1and the current time point).

[0018] In certain embodiments, the current time point is used, within the machine learning model, as a conditioning value.

[0019] In certain embodiments, at step (b), the machine learning model generates at least a portion (e.g., each) of the one or more velocity fields as output.

[0020] In certain embodiments, at step (b), the machine learning model generates a set of prospective final values of the position and / or orientation components of the feature vectors (e.g., representing a current estimate of positions and / or orientations of each backbone site of a generated peptide backbone) as output and the current velocity field is determined based on the set of prospective final values and a current time point [e.g., corresponding to a current iteration (e.g., generating a current estimate of x1as output of the machine learning model and computing the current velocity field based on x1and the current time point)].

[0021] In certain embodiments, the set of prospective final values comprises, for each particular feature vector: (i) a prospective final value of the position component represents a translation relative to a position component of the (e.g., same) particular feature vector at the current time (e.g., thereby expressing a final position of the backbone site represented by the feature vector as a translation relative to its position at an earlier, e.g., the current, time point); and / or (ii) a prospective final value of the orientation component represents a rotation relative to an orientation component of the (e.g., same) particular feature vector at the current time (e.g., thereby expressing a final orientation of the backbone site represented by the feature vector as a rotation relative to its orientation at an earlier, e.g., the current, time point).

[0022] In certain embodiments, the machine learning model is or comprises a transformer-based model.

[0023] In certain embodiments, the machine learning model receives and / or operates on an input graph representation of a peptide backbone, the input graph representation comprising a - 5 - 11901435v1Attorney Docket No.2013969-0048 plurality of nodes and edges, each node corresponding to and representing a particular backbone (e.g., amino acid) site, and each edge associated with and relating two nodes [e.g., thereby representing (i) an interaction between and / or (ii) a relative position and / or orientation between two particular backbone (e.g., amino acid) sites], and wherein the machine learning model comprises at least one transformer-based edge retrieval layer that: (A) comprises one or more self-attention head(s), each of which determines a corresponding set of attention weights based at least in part on values of the feature vectors and / or node feature values determined therefrom, such that the edge retrieval layer determines one or more sets of attention weights; and (B) uses the one or more sets of attention weights to determine values of retrieved edge feature vectors.

[0024] In certain embodiments, the machine learning model is or comprises a graph neural network (GNN).

[0025] In certain embodiments, the machine learning model is or comprises an auto- regressive neural network.

[0026] In certain embodiments, the machine learning model receives, as input and / or determines, values of one or more edge features, each determined based on values of the position and / or orientation components of two or more feature vectors representing two or more different backbone sites.

[0027] In certain embodiments, the method comprises conditioning generation of the one or more velocity fields according to a set of one or more desired peptide backbone features.

[0028] In certain embodiments, the machine learning model receives, as input, and conditions generation of the one or more velocity fields on values one or more global property variables, each global property variable representing a desired property of a protein or peptide (e.g., having the generated peptide backbone).

[0029] In certain embodiments, the one or more global property variables comprise one or more of the following: (A) a protein family variable whose value identifies a particular one of a set of protein family types (e.g., a categorical variable that can take on one of a finite set of values, each identifying a particular protein family; e.g., finite length vector that represents a particular protein family via a one-hot encoding); (B) a thermostability variable whose value categorizes and / or measures (e.g., quantitatively) protein thermostability [e.g., a binary variable - 6 - 11901435v1Attorney Docket No.2013969-0048 classifying a protein as thermophilic or not (e.g., based on PDB classification); e.g., a continuous real-number measuring melting temperature]; (C) an immunology variable whose value classifies and / or measures a propensity and / or likelihood of provoking an immune response [e.g., in a particular host organism (e.g., a human)]; (D) a function variable whose value classifies protein function; (E) a solubility variable whose value classifies and / or measures a protein solubility; and (F) a pH sensitivity variable whose value classifies and / or measures protein pH sensitivity.

[0030] In certain embodiments, the machine learning model receives, as input and conditions generation of the one or more velocity fields on, values of one or more node property variables, each node property variable associated with an representing a particular property of a particular amino acid site.

[0031] In certain embodiments, the one or more node property tokens comprise one or more of the following:(A) a side chain type variable that identifies a particular type of amino acid side chain; (B) an amino acid polarity variable that identifies a polarity and / or charge of an amino acid site; (C) a buriedness variable that classifies and / or measures an extent to which a particular amino acid site is buried and / or surface-accessible (e.g., a binary core / surface value); (D) a contact hotspot variable classifying a particular amino acid site according to a desired and / or threshold distance from one or more portions of another (e.g., target) molecule; and (E) a secondary structure variable whose value classifies and / or measures a secondary structure motif at the particular amino acid site.

[0032] In certain embodiments, the machine learning model receives, as input, values of one or more edge property variables (e.g., a contact hotspot token classifying the edge as associated with one or more particular amino acid sites having a desired and / or within a threshold distance from one or more portions of another (e.g., target) molecule).

[0033] In certain embodiments, the method comprises conditioning generation of the one or more velocity fields according to a representation of at least a portion of a target molecule (e.g., protein and / or peptide) and / or one or more particular sub-regions (e.g., prospective binding sites) thereof, thereby creating a de-novo peptide backbone suitable for binding to the target molecule (e.g., at a location within or comprising at least one of the one or more prospective binding sites). - 7 - 11901435v1Attorney Docket No.2013969-0048

[0034] In certain embodiments, the method comprises conditioning generation of the one or more velocity fields according to a representation of at least a portion of a desired final protein and / or peptide (e.g., a scaffold model representing at least a portion of a desired peptide backbone; e.g., an amino acid sequence of at least a portion of a desired final protein and / or peptide).

[0035] In certain embodiments, the method comprises: using the generated scaffold model as input to an interface designer module (e.g., comprising a machine learning model) for generating an amino acid interface for binding to a target molecule; and populating, by the processor, at least a portion of the generated scaffold model with a plurality of amino acids selected, by the processor (e.g., using a machine learning model) based at least in part on (i) the generated scaffold model and (ii) a target model representing at least a portion of the target molecule.

[0036] In certain embodiments, the method comprises creating, based at least in part on the generated scaffold model, a custom biologic [e.g., a therapeutic biologic; e.g., a diagnostic biologic; e.g., a theranostic biologic; e.g., an enzyme (e.g., for use in manufacturing or other industrial processes, e.g., disposal, clean up, etc.); e.g., a biologic that acts as a structural element].

[0037] In certain embodiments, the method comprises determining an amino acid sequence based at least in part on (e.g., that will and / or is predicted to fold into a three- dimensional structure corresponding to) the generated scaffold model.

[0038] In certain embodiments, the method comprises providing a polynucleotide encoding a sequence (e.g., an amino acid sequence) determined based on the generated scaffold model (e.g., the polynucleotide having a nucleic acid sequence that encodes a corresponding amino acid sequence determined using the generated scaffold model) and delivering the polynucleotide, for transcription and translation thereof, to a plurality of host cells, thereby producing the custom biologic.

[0039] In certain embodiments, the method comprises determining (e.g., at step (b); e.g., subsequent to step (b), e.g., via repeated use of the machine learning model), using the machine learning model, one or more sequence velocity fields and using the one or more sequence - 8 - 11901435v1Attorney Docket No.2013969-0048 velocity fields to generate predicted sequence data representing an amino acid sequence of a protein and / or peptide having the de-novo peptide backbone.

[0040] In certain embodiments, the method comprises determining (e.g., at step (b); e.g., subsequent to step (b), via a repeated use of the machine learning model), using the machine learning model, one or more side chain geometry velocity fields and using the one or more side chain velocity fields to generate a prediction of a three-dimensional side chain geometry for an amino acid side chain (e.g., attached) at each of at least a portion of amino acid sites of the de- novo peptide backbone [e.g., wherein the three-dimensional side chain geometry at the portion of amino acid sites is represented via a set of side chain feature vectors, each corresponding to a particular side chain / amino acid site and having values representing a particular three- dimensional geometry of a side chain at that amino acid site (e.g., having four elements, each corresponding to one of four possible torsion angles), and the method comprises updating values of the feature vectors (e.g., repeatedly) according to the one or more determined side chain velocity fields].

[0041] In certain embodiments, the method comprises conditioning generation of the one or more velocity fields according to one or more protein fold representation(s) [e.g., wherein the machine learning model receives, as input, and conditions output on, the one or more protein fold representation(s)].

[0042] In certain embodiments, the one or more protein fold representation(s) are or comprise a set of secondary structure element (SSE) values (e.g., a SSE feature vector), each SSE value associated with a particular position (e.g., amino acid site) within a polypeptide chain of the custom biologic and having a value encoding a particular type of secondary structure [e.g., each value representing one of a plurality of possible secondary structure motifs / categories (e.g., helix, strand, loop, etc.)] at the particular position.

[0043] In certain embodiments, the one or more protein fold representation(s) are or comprise a block adjacency matrix [e.g., representing a spatial and orientational (e.g., angle) relationship between SSE’s at pairs of amino acid sides], said block adjacency matrix comprising a plurality of elements, each element of the block adjacency matrix associated with a particular pair of positions (e.g., a pair of amino acid sites) within a polypeptide chain of the custom biologic and having a value representing a level (e.g., one of two or more discrete levels) of - 9 - 11901435v1Attorney Docket No.2013969-0048 interaction between the particular pair of positions (e.g., that would occur in a desired 3D folded shape of the protein) (e.g., based on a proximity between the amino acid sites and / or a particular SSE value at each of the amino acid sites) [e.g., wherein each value of the block adjacency matrix is one of two possible binary values, specifying whether a pair of amino acid sites would interact (e.g., be within a particular cutoff distance) in a 3D structure according to the desired fold family].

[0044] In certain embodiments, the one or more protein fold representation(s) are or comprise a block adjacency matrix [e.g., representing a spatial and orientational (e.g., angle) relationship between SSE’s at pairs of amino acid sides], said block adjacency matrix comprising a plurality of elements, each element of the block adjacency matrix associated with a particular pair of positions (e.g., a pair of amino acid sites) within a polypeptide chain of the custom biologic and having one or more values representing a relative position and / or orientation secondary structural elements (SSEs) at the particular pair of positions [e.g., the topological classification value selected from one of a set of categories (e.g., parallel, antiparallel, vertical; e.g., each element having a plurality of values (e.g., different channels), each associated with a particular topological classification category and having a binary value representing whether or not (the 3D folds at) the pair of positions has / is oriented in accordance with the particular topological classification category; e.g., a (e.g., continuous numerical value) representing a relative distance and / or angle between SSEs].

[0045] In another aspect the invention is directed to a method for computer-aided generative design of de-novo polypeptides and / or complexes thereof, the method comprising: (a) receiving and / or generating, by a processor of a computing device, a seed representation of one or more peptide backbones and / or amino acid sequences, the seed representation comprising a plurality of feature values representing (i) (e.g., locations and / or orientations of) amino acid sites and / or (ii) individual backbone atoms and / or side chain types thereof; (b) determining, by the processor, using one or more machine learning models, one or more (e.g., a plurality of) velocity fields and, beginning with the seed set, updating (iteratively), by the processor, the plurality of feature values according to the one or more velocity fields, thereby evolving the representation of one or more peptide backbones and / or amino acid sequences from initial feature values of the seed representation to a set of final feature values representing amino acid sites, and / or individual backbone atoms and / or side chain types, of one or more generated peptide backbones - 10 - 11901435v1Attorney Docket No.2013969-0048 and / or amino acid sequences of the one or more de-novo polypeptide and / or a complex thereof; (c) creating, by the processor, using the set of final values, one or more generated scaffold model(s) and / or sequence string(s) representing the de-novo polypeptide(s) and / or complexes thereof; and (d) storing and / or providing (e.g., for display and / or further processing) the generated scaffold models and / or sequence strings.

[0046] In certain embodiments, the plurality of feature values comprise values representing locations (e.g., in Cartesian space) of one or more individual backbone atoms (e.g., N, C, alpha-Carbon, and O) for each of a plurality of amino acid sites of the de-novo polypeptide to-be and / or being generated.

[0047] In certain embodiments, the plurality of feature values comprise values representing positions and / or orientations of amino acid sites of the de-novo polypeptide to-be and / or being generated.

[0048] In certain embodiments, the plurality of feature values comprise sequence data representing an amino acid sequence of the de-novo polypeptide to-be and / or being generated.

[0049] In certain embodiments, the method comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ).

[0050] In certain embodiments, the method comprises obtaining (e.g., receiving and / or accessing and / or generating), by the processor, a template model representing a partial sequence and / or three-dimensional structure of at least a portion of a reference protein and / or peptide; at step (a), generating the seed representation comprising the plurality of feature values, wherein a first subset of the plurality of feature values correspond to pre-defined regions of the polypeptide(s) being designed and are populated with feature values representing (e.g., and having been determined based on) the partial sequence and / or three dimensional structure of the reference protein and / or peptide (e.g., using the template model) and wherein a second subset of the plurality of feature values correspond to custom, to-be-designed, regions of the polypeptide(s) and are populated with initial starting values; and at step (b), determining the one or more velocity fields based at least in part on the plurality of feature values, including the first subset (corresponding to pre-defined regions) and the second subset (corresponding to unknown region that are, e.g., iteratively, updated) [e.g., thereby conditioning generation of the one or more velocity fields at least in part on the first subset of feature values] and updating the second - 11 - 11901435v1Attorney Docket No.2013969-0048 subset of feature values based on the one or more velocity fields, while retaining / holding fixed the first subset of feature values.

[0051] In certain embodiments, the method comprises: obtaining (e.g., receiving and / or accessing and / or generating), by the processor, a template model representing a partial sequence and / or three-dimensional structure of one or more base portion(s) of a reference protein and / or peptide; determining, by the processor, one or more protein fold representation(s) based on the template model, wherein each of the one or more protein fold representation(s) represent (i) a secondary structure type of, and / or (ii) a relative proximity and / or orientation between amino acid sites of, the one or more base portion(s) of the reference protein and / or peptide [e.g., but do not include or represent definite three-dimensional coordinates of backbone atoms of the reference protein and / or peptide]; at step (a), generating the seed representation comprising the plurality of feature values, said plurality of feature values representing (i) locations and / or orientations of, and / or (ii) individual backbone atoms and / or side chain types of, a plurality of amino acid sites within a de-novo polypeptide (e.g., being designed), and wherein: a first subset of the plurality of feature values represent [e.g., (i) locations and / or orientations, and / or (ii) individual backbone atoms and / or side chain types of] amino acid sites within a first portion of the de-novo polypeptide, said first portion corresponding to the one or more base portion(s) of the reference polypeptide; and a second subset of the plurality of feature values represent [e.g., (i) locations and / or orientations, and / or (ii) individual backbone atoms and / or side chain types of] amino acid sites within a second portion of the de-novo polypeptide, said second portion adjacent to (e.g., having one or more amino acid sites adjacent to sites within) the first portion; and at step (b), determining the one or more velocity fields based at least in part on the one or more protein fold representation(s), such that the one or more generated peptide backbones and / or amino acid sequences comprise a first portion having a three-dimensional fold based on the one or more protein fold representation(s) and a second, custom, portion.

[0052] In certain embodiments, the reference protein and / or peptide is an antibody [e.g., and wherein the one or more base portions are or comprise heave and / or light chain region(s) of the antibody, but exclude complementary determining regions (CDRs) of the antibody].

[0053] In another aspect, the invention is directed to a method for (e.g., memory efficient) generating attention-based deep-learning predictions from input graph representations - 12 - 11901435v1Attorney Docket No.2013969-0048 comprising nodes and edges, the method comprising: (a) receiving and / or accessing, by a processor of a computing device, an initial graph representation comprising a plurality of node feature vectors, each node feature vector associated with a particular node of the initial graph representation and comprising one or more initial node feature values; (b) determining, by the processor, using a machine learning model, a predicted graph representation, wherein the machine learning model: (i) receives the initial graph representation as input; (ii) comprises at least one edge retrieval layer, wherein the at least one edge retrieval layer: (A) comprises one or more self-attention head(s), each of which determines a corresponding set of attention weights based at least in part on values of the feature vectors and / or node feature values determined therefrom, such that the edge retrieval layer determines one or more sets of attention weights; and (B) uses the one or more sets of attention weights to determine values of retrieved edge feature vectors; and (iii) generates, as output, based at least in part on the values of the retrieved edge feature vectors, a plurality of predicted node feature vectors and / or one or more velocity fields, and wherein the predicted graph representation is determined using the plurality of predicted node feature vectors and / or one or more velocity fields; and (c) storing and / or providing (e.g., for display and / or further processing), the predicted graph representation.

[0054] In another aspect, the invention is directed to a method for designing a custom biologic having a desired three-dimensional (e.g., fold) fold, the method comprising: (a) receiving and / or generating, by a processor of a computing device, one or more protein fold representations, each encoding desired three-dimensional structural features of the desired fold at / in at least a portion of locations within the custom biologic; (b) receiving and / or generating, by the processor, a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular (e.g., amino acid) site within the one or more variable region(s) of template model and comprising one or both of: (A) position and / or orientation components representing a position and / or orientation of (e.g., individual backbone atoms and / or a local frame representation of) the particular site; and (B) a side chain type component, representing likelihoods of one or more particular amino acid side chain types occupying the particular site [e.g., wherein values of (A) the position and / or orientation components and / or (B) side chain type components of each feature vector are set to initial starting values (e.g., selected at random from an initial probability (e.g., uniform) distribution)];(c) determining, by the processor, using a machine learning model, one or more (e.g., a plurality of) velocity fields based at least in part on - 13 - 11901435v1Attorney Docket No.2013969-0048 the one or more protein fold representations, and, beginning with the seed set, updating (e.g., iteratively), by the processor, values of one or both of (e.g., (A) and (B)) (A) the position and / or orientation components and (B) the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving them a set of initial starting values into a set of final values representing (i) positions and / or orientations of each site of generated peptide backbone structures of the one or more variable regions, and / or (ii) likelihoods of (e.g., amino acid) side chain types at locations within the one or more variable regions; (d) generating, by the processor, using the set of final values, one or both of: (A) a scaffold model representing a three-dimensional structure of a de-novo peptide backbone for the custom biologic, and (B) sequence data representing an amino acid sequence of the custom biologic; and (e) storing and / or providing, by the processor, the generated (A) scaffold model and / or (B) sequence data for display and / or further processing and / or use in designing the custom biologic.

[0055] In certain embodiments, the one or more protein fold representation(s) are or comprise a set of secondary structure element (SSE) values (e.g., a SSE feature vector), each SSE value associated with a particular position (e.g., amino acid site) within a polypeptide chain of the custom biologic and having a value encoding a particular type of secondary structure [e.g., each value representing one of a plurality of possible secondary structure motifs / categories (e.g., helix, strand, loop, etc.)] at the particular position.

[0056] In certain embodiments, the one or more protein fold representation(s) are or comprise a block adjacency matrix [e.g., representing a spatial and orientational (e.g., angle) relationship between SSE’s at pairs of amino acid sides], said block adjacency matrix comprising a plurality of elements, each element of the block adjacency matrix associated with a particular pair of positions (e.g., a pair of amino acid sites) within a polypeptide chain of the custom biologic and having a value representing a level (e.g., one of two or more discrete levels) of interaction between the particular pair of positions (e.g., that would occur in a desired 3D folded shape of the protein) (e.g., based on a proximity between the amino acid sites and / or a particular SSE value at each of the amino acid sites) [e.g., wherein each value of the block adjacency matrix is one of two possible binary values, specifying whether a pair of amino acid sites would interact (e.g., be within a particular cutoff distance) in a 3D structure according to the desired fold family]. - 14 - 11901435v1Attorney Docket No.2013969-0048

[0057] In certain embodiments, the one or more protein fold representation(s) are or comprise a block adjacency matrix [e.g., representing a spatial and orientational (e.g., angle) relationship between SSE’s at pairs of amino acid sides], said block adjacency matrix comprising a plurality of elements, each element of the block adjacency matrix associated with a particular pair of positions (e.g., a pair of amino acid sites) within a polypeptide chain of the custom biologic and having one or more values representing a relative position and / or orientation secondary structural elements (SSEs) at the particular pair of positions [e.g., the topological classification value selected from one of a set of categories (e.g., parallel, antiparallel, vertical; e.g., each element having a plurality of values (e.g., different channels), each associated with a particular topological classification category and having a binary value representing whether or not (the 3D folds at) the pair of positions has / is oriented in accordance with the particular topological classification category; e.g., a (e.g., continuous numerical value) representing a relative distance and / or angle between SSEs].

[0058] In certain embodiments, e.g., step (c) comprises performing the method of any one of various aspects and / or embodiments described herein, e.g., in paragraphs above (e.g., at paragraphs

[0005] -

[0053] ) [e.g., using flow-matching; e.g., using the machine learning model to update and evolve feature values representing (i) (e.g., locations and / or orientations of) amino acid sites and / or (ii) individual backbone atoms and / or side chain types thereof] to generate the sequence and / or the three-dimensional peptide backbone structure, and wherein} the machine learning model receives the one or more protein fold representations as input, thereby conditioning generation of the one or more velocity fields the desired fold.

[0059] In certain embodiments, the method comprises, at step (a) generating the one or more protein fold representations based on a portion of an antibody template representing at least a portion (e.g., a Fab region, a variable heavy chain, a variable light chain, substantially all, all, etc.) of a reference antibody structure.

[0060] In certain embodiments, the portion of the antibody template is a base portion, representing a region of the reference antibody about one or more complementary-determining regions (CDRs) but excluding the CDRs themselves. - 15 - 11901435v1Attorney Docket No.2013969-0048

[0061] In certain embodiments, steps (b) through (d) comprises generating one or both of (i) a custom sequence and (ii) a custom three-dimensional peptide backbone structure of the one or more CDRs.

[0062] In certain embodiments, the method comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ).

[0063] In another aspect, the invention is directed to a method for designing a custom antibody for binding to a target [e.g., a target antigen or portion (e.g., epitope) thereof], the method comprising: (a) receiving and / or generating, by a processor of a computing device, an antibody template representing a sequence and / or three-dimensional structure of at least a portion of a reference antibody, the antibody template comprising (i) a base portion located about one or more CDRs of the reference antibody, but (the base portion) excluding at least a portion of the one or more CDRs themselves; (b) determining, by the processor, one or more protein fold representations, each encoding and representing three-dimensional structural features of at least the base portion of the antibody template; (c) receiving and / or generating, by the processor, a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular (e.g., amino acid) site within one or more variable region(s) corresponding to, and representing to-be-designed custom versions of, CDR regions of the reference antibody and comprising one or both of: (A) position and / or orientation components representing a position and / or orientation of (e.g., individual backbone atoms and / or a local frame representation of) the particular site and (B) a side chain type component, representing likelihoods of one or more particular amino acid side chain types occupying the particular site [e.g., wherein values of (A) the position and / or orientation components and / or (B) side chain type components of each feature vector are set to initial starting values (e.g., selected at random from an initial probability (e.g., uniform) distribution)];(c) determining, by the processor, using a machine learning model, one or more (e.g., a plurality of) velocity fields based at least in part on the one or more protein fold representations, and, beginning with the seed set, updating (e.g., iteratively), by the processor, values of one or both of (e.g., (A) and (B)) (A) the position and / or orientation components and (B) the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving them a set of initial starting values into a set of final values representing (i) positions and / or orientations of each site of generated peptide backbone structures of the one or more variable regions, and / or (ii) likelihoods of (e.g., amino - 16 - 11901435v1Attorney Docket No.2013969-0048 acid) side chain types at locations within the one or more variable regions; (d) generating, by the processor, using the set of final values, one or both of: (A) a scaffold model representing a three- dimensional structure of a de-novo peptide backbone for one or more CDR regions of the custom antibody; and (B) sequence data representing an amino acid sequence of one or more CDR regions of the custom antibody; (e) storing and / or providing, by the processor, the generated (A) scaffold model and / or (B) sequence data, for display and / or further processing and / or use in designing the custom biologic.

[0064] In certain embodiments, the method comprises receiving and / or generating a target model representing a sequence and / or three-dimensional structure of at least a portion of the target (e.g., an epitope) and wherein step (c) comprises generating one or both of (i) the sequence and (ii) the three-dimensional peptide backbone structure based further on the target model [e.g., wherein the machine learning model receives, as input, and conditions output on, at least a portion of the target model and / or one or more features values derived therefrom].

[0065] In certain embodiments, the method comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ).

[0066] In another aspect, the invention is directed to a method for in-silico generation and / or prediction of three-dimensional (3D) side chain geometries of a polypeptide chain (e.g., a protein and / or peptide), the method comprising: (a) receiving and / or generating, by a processor of a computing device, (i) sequence data representing an amino acid sequence of at least a portion of the polypeptide chain and (ii) a representation of a 3D peptide backbone geometry, and / or folds thereof, for the portion of the polypeptide chain; (b) receiving and / or generating, by a processor of a computing device a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular amino acid site of the portion of the polypeptide chain and comprising side chain geometry components representing a (e.g., 3D) geometry [e.g., position and / or orientation (e.g., as determined via one or more torsion angles)] of a side chain at the particular amino acid site [e.g., wherein values of the one or more side chain geometry components of each feature vector are set to initial starting values (e.g., selected at random from an initial probability (e.g., uniform) distribution)]; (c) determining, by the processor, using a machine learning model, one or more (e.g., a plurality of) (side chain) velocity fields based at least in part on (i) the sequence data and (ii) the representation of the 3D peptide backbone - 17 - 11901435v1Attorney Docket No.2013969-0048 geometry and / or folds thereof and, beginning with the seed set, updating (e.g., iteratively), by the processor, values of the side chain geometry components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving the values of the side chain geometry components of the plurality of feature vectors from a set of initial starting values into a set of final values representing (e.g., final, determined) geometries of each amino acid site of the portion of the polypeptide chain; and (c) creating, by the processor, using, the set of final values, a 3D polypeptide model of the portion of the polypeptide chain representing the 3D geometries of amino acid side chains within the portion of the polypeptide chain [e.g., using the set of final values as the 3D model representing the polypeptide chain; e.g., determining, based on the set of final values, positions of side chain atoms at each amino acid site]; and (d) storing and / or providing (e.g., for display and / or further processing), the 3D polypeptide model.

[0067] In certain embodiments, the method comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ).

[0068] In another aspect, the invention is method for in-silico generation and / or prediction of an amino acid sequence of a polypeptide chain (e.g., a protein and / or peptide), the method comprising: (a) receiving and / or generating, by a processor of a computing device a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular (e.g., backbone) site of the portion of the polypeptide chain and a side chain type component representing a likelihood of one or more possible types of amino acid side chains [e.g., a likelihood of each of one of N (e.g., twenty) possible types of amino acid(s) (e.g., each feature vector having a plurality of elements, each corresponding to a particular amino acid type (e.g., of the N possible types, such as the twenty canonical amino acids) and having a value (e.g., between 0 and 1) representing a likelihood of that particular amino acid type occupying the particular backbone site) (e.g., each feature vector representing a distribution over a (e.g., N -1 dimension) simplex)] at the particular site [e.g., wherein values of the side chain type component of each feature vector are set to initial starting values (e.g., selected at random from an initial probability (e.g., uniform) distribution)]; (b) determining, by the processor, using a machine learning model, one or more (e.g., a plurality of) (sequence) velocity fields and, beginning with the seed set, updating (e.g., iteratively), by the processor, values of the side chain type components of the plurality of feature vectors according to the one or more (sequence) velocity fields, thereby evolving the values of the side chain type components of the plurality of feature vectors from a - 18 - 11901435v1Attorney Docket No.2013969-0048 set of initial starting values into a set of final values representing (e.g., final, determined) likelihoods of amino acid side chain types at each site of the portion of the polypeptide chain; and (c) determining, by the processor, using, the set of final values, sequence data representing an amino acid sequence of the portion of the polypeptide chain representing (e.g., optionally, projecting each final feature vector of the set of final values from a N-1 dimensional simplex to a N dimensional likelihood vector, and taking an argmax to determine a particular amin acidy side chain type at each location within the portion of the polypeptide chain); and (d) storing and / or providing (e.g., for display and / or further processing), the sequence data.

[0069] In certain embodiments, the method comprises: receiving and / or generating, by the processor, a representation of a 3D peptide backbone geometry, and / or folds thereof, for at least a portion of the polypeptide chain; and at step (b), determining the one or more velocity fields based at least in part on the representation of the 3D peptide backbone geometry and / or folds thereof.

[0070] In certain embodiments, the method comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ).

[0071] In another aspect, the invention is directed to a method for generating in-silico docking predictions for a plurality (e.g., two or more) of polypeptide chains (e.g., proteins and / or peptides), the method comprising: (a) receiving and / or generating, by a processor of a computing device, a first scaffold model representation of a first 3D peptide backbone of at least a portion of a first polypeptide chain; (b) receiving and / or generating, by a processor of a computing device, a second scaffold model representation of a second 3D peptide backbone of at least a portion of a second polypeptide chain; (c) receiving and / or generating, by a processor of a computing device a seed set comprising a plurality of feature vectors, the plurality of feature vectors comprising (i) a first set of feature vectors corresponding to and representing positions and / or orientations of (e.g., individual backbone sites of) the first 3D peptide backbone and / or (ii) a second set of feature vectors corresponding to and representing positions and / or orientations of (e.g., individual backbone sites of) the second 3D peptide backbone [e.g., wherein values of the position and / or orientation components of each feature vector are set to initial starting values (e.g., selected at random from an initial probability (e.g., uniform) distribution)]; (d) determining, by the processor, using a machine learning model, based at least in part on the first - 19 - 11901435v1Attorney Docket No.2013969-0048 scaffold model and the second scaffold model, one or more (e.g., a plurality of) velocity fields, by the processor, and, beginning with the seed set, updating (e.g., iteratively), by the processor, values of (i) the first set of feature vectors and / or (ii) the second set of feature vectors according to the one or more velocity fields, thereby evolving values of the first and / or second feature vector sets from a set of initial starting values into a set of final values representing final position(s) and / or orientation(s) of the first 3D peptide backbone and / or the second 3D peptide backbone docket (e.g., forming a polypeptide complex) with each other [e.g., wherein the method comprises using the machine learning model to (e.g., iteratively) determine, at each of a plurality of time points, xta velocity field v(xt,t) using: (i) a current x(t) as input, for which the first and second polypeptide chains are treated together – e.g., edge features relating sites within a single chain and relating sites in between chains are computed, and (ii) a conditioning input that is an “x1” for each chain – representing locations and orientations of each backbone site in the known structure of each chain as represented by the first and second scaffold models e.g., in which each chain is treated independently – e.g.., only edge features for relationships within a same chain are computed and used as conditioning input]; (e) creating, by the processor, using, the set of final values, a generated polypeptide complex model representing first and second 3D peptide backbones docked to form a polypeptide complex; and (f) storing and / or providing (e.g., for display and / or further processing), the generated polypeptide complex model.

[0072] In certain embodiments, the method comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ).

[0073] In another aspect, the invention is directed to a method for designing a custom biologic for binding to a target [e.g., a target antigen or portion (e.g., epitope) thereof], the method comprising: (a) receiving and / or generating, by the processor, a target model representing at least a portion of the target (e.g., comprising a desired epitope region, to target for binding); (b) receiving and / or generating, by a processor of a computing device, a template model representing a sequence and / or three-dimensional structure of at least a portion of a reference biologic (e.g., one or more polypeptide chains; e.g., a protein and / or peptide), the template model comprising a base portion representing a portion of the reference biologic located about (e.g., in proximity to, in regards to linear sequence and / or 3D folded structure of the reference biologic) one or more (e.g., predetermined) variable regions of the reference biologic designated (e.g., a-priori known, have been determined, or otherwise identified as - 20 - 11901435v1Attorney Docket No.2013969-0048 regions that do, will, or are intended / desired to interact with and mediate binding to the target) as modifiable for binding to the target; (c) receiving and / or generating, by the processor, a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular (e.g., amino acid) site within the one or more variable region(s) of template model and comprising one or both of (i) position and / or orientation components representing a position and / or orientation of (e.g., individual backbone atoms and / or a local frame representation of) the particular site and (ii) a side chain type component, representing likelihoods of one or more particular amino acid side chain types occupying the particular site [e.g., wherein values of (i) the position and / or orientation components and / or (ii) side chain type components of each feature vector are set to initial starting values (e.g., selected at random from an initial probability (e.g., uniform) distribution)]; (d) determining, by the processor, using a machine learning model, one or more (e.g., a plurality of) velocity fields based at least in part on (i) the target model [e.g., including an identification of a desired target epitope thereof to bind with, e.g., via a categorical variable used as node feature input to the machine learning model] and (ii) the base portion of the template model and, beginning with the seed set, updating (e.g., iteratively), by the processor, values of one or both of (i) the position and / or orientation and (ii) the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving them a set of initial starting values into a set of final values representing (i) positions and / or orientations of each site of generated peptide backbone structures of the one or more variable regions, and / or (ii) likelihoods of (e.g., amino acid) side chain types at locations within the one or more variable regions; (e) creating, by the processor, using, the set of final values, one or both of: (i) a generated scaffold model representing a de-novo peptide backbone for the one or more variable regions of the custom biologic [e.g., using the set of final values as the generated scaffold model representing the de-novo peptide backbone; e.g., determining, based on the set of final values, positions of backbone atoms (e.g., at each backbone site, a position of one or more backbone atoms (e.g., an N, alpha-Carbon, C, and O)) to be represented via the generated scaffold model] and (ii) sequence data representing a generated amino acid sequence of the one or more variable region(s) (e.g., optionally, projecting each final feature vector of the set of final values from a N- 1 dimensional simplex to a N dimensional likelihood vector, and taking an argmax to determine a particular amin acidy side chain type at each location within the portion of the polypeptide chain); and (f) storing and / or providing, by the processor, the peptide backbone structure and / or - 21 - 11901435v1Attorney Docket No.2013969-0048 the generated sequence for display and / or further processing and / or use in designing the custom biologic.

[0074] In certain embodiments, the method comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ).

[0075] In another aspect, the invention is directed to a system for computer generation of a de-novo peptide backbone of a custom biologic, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular backbone site and comprising position and / or orientation components representing a position and / or orientation of the particular backbone site; (b) determine, using a machine learning model, one or more velocity fields and, beginning with the seed set, updating, by the processor, values of the position and / or orientation components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving the values of the position and / or orientation components of the plurality of feature vectors from a set of initial starting values into a set of final values representing positions and / or orientations of each backbone site of a generated peptide backbone; and (c) create, using, the set of final values, a generated scaffold model representing the de-novo peptide backbone; and (d) store and / or provide the generated scaffold model.

[0076] In certain embodiments, the system comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ).

[0077] In another aspect, the invention is directed to a system for computer-aided generative design of de-novo polypeptides and / or complexes thereof, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate a seed representation of one or more peptide backbones and / or amino acid sequences, the seed representation comprising a plurality of feature values representing (i) amino acid sites and / or (ii) individual backbone atoms and / or side chain types thereof; (b) determine, using one or more machine learning models, one or more velocity fields and, beginning with the seed set, updating, by the processor, the plurality of feature values according to the one or more velocity fields, - 22 - 11901435v1Attorney Docket No.2013969-0048 thereby evolving the representation of one or more peptide backbones and / or amino acid sequences from initial feature values of the seed representation to a set of final feature values representing amino acid sites, and / or individual backbone atoms and / or side chain types, of one or more generated peptide backbones and / or amino acid sequences of the one or more de-novo polypeptide and / or a complex thereof; (c) create, using the set of final values, one or more generated scaffold model(s) and / or sequence string(s) representing the de-novo polypeptide(s) and / or complexes thereof; and (d) store and / or provide the generated scaffold models and / or sequence strings.

[0078] In certain embodiments, the system comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ).

[0079] In another aspect, the invention is directed to a system for generating attention- based deep-learning predictions from input graph representations comprising nodes and edges, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to (a) receive and / or access an initial graph representation comprising a plurality of node feature vectors, each node feature vector associated with a particular node of the initial graph representation and comprising one or more initial node feature values; (b) determine, using a machine learning model, a predicted graph representation, wherein the machine learning model: (i) receives the initial graph representation as input; (ii) comprises at least one edge retrieval layer, wherein the at least one edge retrieval layer: (A) comprises one or more self-attention head(s), each of which determines a corresponding set of attention weights based at least in part on values of the feature vectors and / or node feature values determined therefrom, such that the edge retrieval layer determines one or more sets of attention weights; and (B) uses the one or more sets of attention weights to determine values of retrieved edge feature vectors; and (iii) generates, as output, based at least in part on the values of the retrieved edge feature vectors, a plurality of predicted node feature vectors and / or one or more velocity fields, and wherein the predicted graph representation is determined using the plurality of predicted node feature vectors and / or one or more velocity fields; and (c) store and / or provide, the predicted graph representation. - 23 - 11901435v1Attorney Docket No.2013969-0048

[0080] In another aspect, the invention is directed to a system for designing a custom biologic having a three-dimensional structure corresponding to a desired fold, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate one or more protein fold representations, each encoding desired three- dimensional structural features of the desired fold at / in at least a portion of locations within the custom biologic; (b) receive and / or generate a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular site within the one or more variable region(s) of template model and comprising one or both of: (A) position and / or orientation components representing a position and / or orientation of the particular site and (B) a side chain type component, representing likelihoods of one or more particular amino acid side chain types occupying the particular site; (c) determine, using a machine learning model, one or more velocity fields based at least in part on the one or more protein fold representations, and, beginning with the seed set, update values of one or both of (A) the position and / or orientation components and (B) the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving them a set of initial starting values into a set of final values representing (i) positions and / or orientations of each site of generated peptide backbone structures of the one or more variable regions, and / or (ii) likelihoods of side chain types at locations within the one or more variable regions; (d) generate, using the set of final values, one or both of: (A) a scaffold model representing a three-dimensional structure of a de- novo peptide backbone for the custom biologic, and (B) sequence data representing an amino acid sequence of the custom biologic; and (e) store and / or provide the generated (A) scaffold model and / or (B) sequence data for display and / or further processing and / or use in designing the custom biologic.

[0081] In certain embodiments, the system comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ).

[0082] In another aspect, the invention is directed to a system for designing a custom antibody for binding to a target, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate an antibody template representing a sequence and / or three-dimensional structure of at least a portion of a reference antibody, the - 24 - 11901435v1Attorney Docket No.2013969-0048 antibody template comprising (i) a base portion located about one or more CDRs of the reference antibody, but excluding at least a portion of the one or more CDRs themselves; (b) determine one or more protein fold representations, each encoding and representing three-dimensional structural features of at least the base portion of the antibody template; (c) receive and / or generate a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular site within one or more variable region(s) corresponding to, and representing to-be- designed custom versions of, CDR regions of the reference antibody and comprising one or both of: (A) position and / or orientation components representing a position and / or orientation of (e.g., individual backbone atoms and / or a local frame representation of) the particular site and (B) a side chain type component, representing likelihoods of one or more particular amino acid side chain types occupying the particular site;(c) determine, using a machine learning model, one or more velocity fields based at least in part on the one or more protein fold representations, and, beginning with the seed set, update values of one or both of (A) the position and / or orientation components and (B) the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving them a set of initial starting values into a set of final values representing (i) positions and / or orientations of each site of generated peptide backbone structures of the one or more variable regions, and / or (ii) likelihoods of side chain types at locations within the one or more variable regions; (d) generate, using the set of final values, one or both of: (A) a scaffold model representing a three-dimensional structure of a de- novo peptide backbone for one or more CDR regions of the custom antibody; and (B) sequence data representing an amino acid sequence of one or more CDR regions of the custom antibody; and (e) store and / or provide the generated (A) scaffold model and / or (B) sequence data, for display and / or further processing and / or use in designing the custom biologic.

[0083] In certain embodiments, the system comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ).

[0084] In another aspect, the invention is directed to a system for in-silico generation and / or prediction of three-dimensional (3D) side chain geometries of a polypeptide chain, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate (i) sequence data representing an amino acid sequence of at least a portion of the polypeptide chain and (ii) a representation of a 3D peptide backbone geometry, - 25 - 11901435v1Attorney Docket No.2013969-0048 and / or folds thereof, for the portion of the polypeptide chain; (b) receive and / or generate a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular amino acid site of the portion of the polypeptide chain and comprising side chain geometry components representing a geometry of a side chain at the particular amino acid site; (c) determine, using a machine learning model, one or more velocity fields based at least in part on (i) the sequence data and (ii) the representation of the 3D peptide backbone geometry and / or folds thereof and, beginning with the seed set, update values of the side chain geometry components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving the values of the side chain geometry components of the plurality of feature vectors from a set of initial starting values into a set of final values representing geometries of each amino acid site of the portion of the polypeptide chain; and (c) create, using, the set of final values, a 3D polypeptide model of the portion of the polypeptide chain representing the 3D geometries of amino acid side chains within the portion of the polypeptide chain; and (d) store and / or provide the 3D polypeptide model.

[0085] In certain embodiments, the system comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ).

[0086] In another aspect, the invention is directed to a system for in-silico generation and / or prediction of an amino acid sequence of a polypeptide chain, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular site of the portion of the polypeptide chain and a side chain type component representing a likelihood of one or more possible types of amino acid side chains at the particular site; (b) determine, using a machine learning model, one or more velocity fields and, beginning with the seed set, update values of the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving the values of the side chain type components of the plurality of feature vectors from a set of initial starting values into a set of final values representing likelihoods of amino acid side chain types at each site of the portion of the polypeptide chain; and (c) determine, using, the set of final values, sequence data representing an amino acid sequence of the portion of the polypeptide chain; and (d) store and / or provide the sequence data. - 26 - 11901435v1Attorney Docket No.2013969-0048

[0087] In certain embodiments, the instructions cause the processor to: receive and / or generate a representation of a 3D peptide backbone geometry, and / or folds thereof, for at least a portion of the polypeptide chain and at step (b), determine the one or more velocity fields based at least in part on the representation of the 3D peptide backbone geometry and / or folds thereof.

[0088] In certain embodiments, the system comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ).

[0089] In another aspect, the invention is directed to a system for generating in-silico docking predictions for a plurality (e.g., two or more) of polypeptide chains, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate a first scaffold model representation of a first 3D peptide backbone of at least a portion of a first polypeptide chain; (b) receive and / or generate a second scaffold model representation of a second 3D peptide backbone of at least a portion of a second polypeptide chain; (c) receive and / or generate a seed set comprising a plurality of feature vectors, the plurality of feature vectors comprising (i) a first set of feature vectors corresponding to and representing positions and / or orientations of the first 3D peptide backbone and / or (ii) a second set of feature vectors corresponding to and representing positions and / or orientations of the second 3D peptide backbone; (d) determine, using a machine learning model, based at least in part on the first scaffold model and the second scaffold model, one or more velocity fields and, beginning with the seed set, update values of (i) the first set of feature vectors and / or (ii) the second set of feature vectors according to the one or more velocity fields, thereby evolving values of the first and / or second feature vector sets from a set of initial starting values into a set of final values representing final position(s) and / or orientation(s) of the first 3D peptide backbone and / or the second 3D peptide backbone docket with each other; (e) create, using, the set of final values, a generated polypeptide complex model representing first and second 3D peptide backbones docked to form a polypeptide complex; and (f) store and / or provide the generated polypeptide complex model

[0090] In certain embodiments, the system comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ). - 27 - 11901435v1Attorney Docket No.2013969-0048

[0091] In another aspect, the invention is directed to a system for designing a custom biologic for binding to a target, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate a target model representing at least a portion of the target; (b) receive and / or generate a template model representing a sequence and / or three-dimensional structure of at least a portion of a reference biologic, the template model comprising a base portion representing a portion of the reference biologic located about one or more variable regions of the reference biologic designated as modifiable for binding to the target;(c) receive and / or generate a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular site within the one or more variable region(s) of template model and comprising one or both of (i) position and / or orientation components representing a position and / or orientation of the particular site and (ii) a side chain type component, representing likelihoods of one or more particular amino acid side chain types occupying the particular site; (d) determine, using a machine learning model, one or more velocity fields based at least in part on (i) the target model and (ii) the base portion of the template model and, beginning with the seed set, update values of one or both of (i) the position and / or orientation and (ii) the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving them a set of initial starting values into a set of final values representing (i) positions and / or orientations of each site of generated peptide backbone structures of the one or more variable regions, and / or (ii) likelihoods of side chain types at locations within the one or more variable regions; (e) create, using, the set of final values, one or both of: (i) a generated scaffold model representing a de-novo peptide backbone for the one or more variable regions of the custom biologic; and (ii) sequence data representing a generated amino acid sequence of the one or more variable region(s); and (f) store and / or provide, by the processor, the peptide backbone structure and / or the generated sequence for display and / or further processing and / or use in designing the custom biologic.

[0092] In certain embodiments, the system comprises one or more features described herein, e.g., in paragraphs above (e.g., at paragraphs

[0006] -

[0044] ).

[0093] Features of embodiments described with respect to one aspect of the invention may be applied with respect to another aspect of the invention. - 28 - 11901435v1Attorney Docket No.2013969-0048 BRIEF DESCRIPTION OF THE DRAWING

[0094] The foregoing and other objects, aspects, features, and advantages of the present disclosure will become more apparent and better understood by referring to the following description taken in conjunction with the accompanying drawings, in which:

[0095] FIG.1A shows three sets of graphs (3D plots) illustrating use of a flow matching approach to push an initial distribution to a final target distribution, at t = 0, according to an illustrative embodiment.

[0096] FIG.1B shows three sets of graphs (3D plots) illustrating use of a flow matching approach to push an initial distribution to a final target distribution, corresponding to 33% of completion, according to an illustrative embodiment.

[0097] FIG.1C shows three sets of graphs (3D plots) illustrating use of a flow matching approach to push an initial distribution to a final target distribution, corresponding to 66% of completion, according to an illustrative embodiment.

[0098] FIG.1D shows three sets of graphs (3D plots) illustrating use of a flow matching approach to push an initial distribution to a final target distribution, corresponding to 100% of completion, according to an illustrative embodiment.

[0099] FIG.2A is a block flow diagram of an example machine learning model training process, according to an illustrative embodiment.

[0100] FIG.2B is a schematic illustrating a path from an initial starting point to a final point, according to an illustrative embodiment.

[0101] FIG.2C is a block flow diagram of an example process for using a machine learning model to generate de-novo peptide structures, according to an illustrative embodiment.

[0102] FIG.3A is a schematic illustrating an approach for parameterizing protein backbones, according to an illustrative embodiment.

[0103] FIG.3B is a schematic illustrating an approach for parameterizing protein backbones, according to an illustrative embodiment. - 29 - 11901435v1Attorney Docket No.2013969-0048

[0104] FIG.4 is an illustrative schematic illustrating how positions and orientations of reference frames representing protein backbone sites can be iteratively updated via flow matching methods described herein to generate a de-novo peptide backbone structure, according to an illustrative embodiment.

[0105] FIG.5 is a set of five graphs illustrating a use of flow matching approach on a 2- dimensional simplex path to push an initial uniform distribution to a final target distribution, according to an illustrative embodiment.

[0106] FIG.6 is a box plot showing accuracy of flow matching approach over time, according to an illustrative embodiment.

[0107] FIG.7A is a block flow diagram of an example process for generating attention- based deep-learning predictions from input graph representations, according to an illustrative embodiment.

[0108] FIG.7B is a schematic illustrating an example transformer-based machine learning model architecture, used in certain embodiments.

[0109] FIG.7C is a schematic illustrating an example transformer-based machine learning model architecture, used in certain embodiments.

[0110] FIG.7D is a block flow diagram of an example process for designing a custom biologic having a three-dimensional structure belonging to a desired fold family, according to an illustrative embodiment.

[0111] FIG.8 is a ribbon diagram illustrating TIM barrel (PDB: 8TIM) secondary structure.

[0112] FIG.9A is a heatmap showing Block Adjacency Matrix of TIM barrel, according to an illustrative embodiment.

[0113] FIG.9B is a heatmap showing multi-channel Block Adjacency Matrix of TIM barrel, according to an illustrative embodiment.

[0114] FIG.9C is a block flow diagram of an example process for designing a custom antibody for binding to a target, according to an illustrative embodiment. - 30 - 11901435v1Attorney Docket No.2013969-0048

[0115] FIG.10 is a block diagram of an exemplary cloud computing environment, used in certain embodiments.

[0116] FIG.11 is a block diagram of an example computing device and an example mobile computing device, used in certain embodiments.

[0117] FIG.12A is graph showing an example loss curve during training of a machine learning model used to estimate velocity fields in a flow-matching framework for generating de- novo peptide backbones, according to an illustrative embodiment.

[0118] FIG.12B is graph showing an example loss curve showing validation of a machine learning model used to estimate velocity fields in a flow-matching framework for generating de-novo peptide backbones, according to an illustrative embodiment.

[0119] FIG.12C is a series of snapshots showing computed atom positions at various time points as they are iteratively adjusted to create a generated scaffold model representing a generated de-novo peptide backbone, according to an illustrative embodiment.

[0120] FIG.13 is a schematic showing an example neural network architecture, used in certain embodiments.

[0121] FIG.14A is a schematic showing an example neural network architecture and approach for determining input features, according to an illustrative embodiment.

[0122] FIG.14B is a schematic showing an example neural network architecture and approach for determining edge features, according to an illustrative embodiment.

[0123] FIG.14C is an illustration of computed edge features, used in certain embodiments.

[0124] FIG.14D is a schematic showing an example neural network architecture and approach for determining edge features, according to an illustrative embodiment.

[0125] FIG.14E is a schematic showing an example architecture and approach for determining velocity predictions, according to an illustrative embodiment.

[0126] FIG.15A is a schematic illustrating various metrics for evaluating model performance and / or training, according to an illustrative embodiment. - 31 - 11901435v1Attorney Docket No.2013969-0048

[0127] FIG.15B is a schematic illustrating a metric for evaluating model performance and / or training, according to an illustrative embodiment.

[0128] FIG.15C is a graph showing a metric for evaluating model performance and / or training, according to an illustrative embodiment.

[0129] FIG.15D is a graph showing a metric for evaluating model performance and / or training, according to an illustrative embodiment.

[0130] FIG.15E is a schematic illustrating a metric for evaluating model performance and / or training, according to an illustrative embodiment.

[0131] FIG.16 shows an example batch creation procedure, used in certain embodiments.

[0132] FIG.17 is a schematic showing an example neural network architecture, used in certain embodiments.

[0133] FIG.18A is a schematic showing an example neural network architecture, used in certain embodiments.

[0134] FIG.18B is a schematic showing an example neural network architecture, used in certain embodiments.

[0135] FIG.19A is a graph showing a position loss function for a training set with increasing training epoch.

[0136] FIG.19B is a graph showing an orientation loss function for a training set with increasing training epoch.

[0137] FIG.19C is a graph showing a total loss function for a training set with increasing training epoch.

[0138] FIG.19D is a graph showing a position loss function for a validation set with increasing training epoch.

[0139] FIG.19E is a graph showing an orientation loss function for a validation set with increasing training epoch. - 32 - 11901435v1Attorney Docket No.2013969-0048

[0140] FIG.19F is a graph showing a total loss function for a validation set with increasing training epoch.

[0141] FIG.20 is a bar chart showing values of inter-residue clash scores (“between_clash_score”) and values of intra-residue clash scores (“within_clash_score”).

[0142] FIG.21 is a bar chart showing values for bond angle scores determined for N-Cα- C, Cα-C-N, and C-N- Cα bonds.

[0143] FIG.22 is a bar chart showing values for distance scores, in particular a consecutive residue Cα distance score (Cα- Cα distance score) and a C-N bond length score.

[0144] FIG.23 is a bar chart showing values for number of unsatisfied bonds, in particular a number of peptide bond lengths and Cα- Cα distances.

[0145] FIG.24 is a bar chart showing values for number of atoms clashing, in particular where inter-residue and intra-residue clash restrictions were.

[0146] FIG.25A is a box plot showing computed RMSD score distributions for generated structures.

[0147] FIG.25B is a box plot showing computed TM score distributions for generated structures.

[0148] FIG.26A is a histogram showing computed RMSD score distributions for generated structures.

[0149] FIG.26B is a histogram showing computed TM score distributions for generated structures.

[0150] FIG.27 is a box plot showing values of a secondary structure run (“SS run”) metric.

[0151] FIG.28 is a box plot showing values of a secondary structure composition (“SS composition”) metric.

[0152] FIG.29 is a box plot showing values of a secondary structure contact (“SS contact”) metric. - 33 - 11901435v1Attorney Docket No.2013969-0048

[0153] FIG.30A is a bar chart showing backbone dihedral (phi and psi angles) distribution metric.

[0154] FIG.30B is a heatmap showing aggregate residue phi-psi distributions for generated designable structures.

[0155] FIG.30C is a heatmap showing aggregate residue phi-psi distributions for a reference dataset.

[0156] FIG.31A is a ribbon diagram showing an example of a generated protein backbone structures with 100 amino acids.

[0157] FIG.31B is a ribbon diagram showing an example of a generated protein backbone structures with 150 amino acids.

[0158] FIG.31C is a ribbon diagram showing an example of a generated protein backbone structures with 200 amino acids.

[0159] FIG.31D is a ribbon diagram showing an example of a generated protein backbone structures with 250 amino acids.

[0160] FIG.32A is a ribbon diagram showing an example of a generated protein backbone structures with 300 amino acids.

[0161] FIG.32B is a ribbon diagram showing an example of a generated protein backbone structures with 350 amino acids.

[0162] FIG.32C is a ribbon diagram showing an example of a generated protein backbone structures with 400 amino acids.

[0163] FIG.32D is a ribbon diagram showing an example of a generated protein backbone structures with 450 amino acids.

[0164] FIG.33 is a ribbon diagram showing an example of a generated protein backbone structures with 500 amino acids.

[0165] FIG.34 shows ribbon diagrams illustrating predictions performed on PDB structure PDBID 1U2H with 23 residues masked and inpainted by the machine learning model. - 34 - 11901435v1Attorney Docket No.2013969-0048

[0166] FIG.35 shows ribbon diagrams illustrating predictions performed on PDB structure PDBID 6MFW with 48 residues masked and inpainted by the machine learning model.

[0167] FIG.36 shows ribbon diagrams illustrating predictions performed on PDB structure PDBID 4RUW with 50 residues masked and inpainted by the machine learning model.

[0168] FIG.37 shows ribbon diagrams illustrating predictions performed on PDB structure PDBID 4ES6 with 50 residues masked and inpainted by the machine learning model.

[0169] FIG.38 displays ribbon diagrams showing results for several PDB with the prediction shown alone along the top row and overlaid on ground truth.

[0170] FIG.39 shows ribbon diagrams of binders generated using inpainting approaches.

[0171] FIG.40A shows a ribbon diagram of a generated binder, where hotspot information was unknown, using example protein PDB ID 3nau.

[0172] FIG.40B shows a ribbon diagram of a generated binder, where hotspot information was unknown, using example protein PDB ID 2vkl.

[0173] FIG.40C shows a ribbon diagram of a generated binder, where hotspot information was unknown, using example protein PDB ID 4pwy.

[0174] FIG.40D shows a ribbon diagram of a generated binder, where hotspot information was unknown, using example protein PDB ID 5rdv.

[0175] FIG.41A shows a ribbon diagram of a generated binders, where hotspot information was known and provided, using example protein with PDB ID 4wja.

[0176] FIG.41B shows a ribbon diagram of a generated binders, where hotspot information was known and provided, using example protein with PDB ID 5rdv.

[0177] FIG.41C shows a ribbon diagram of a generated binders, where hotspot information was known and provided, using example protein with PDB ID 6ahp.

[0178] FIG.41D shows a ribbon diagram of a generated binders, where hotspot information was known and provided, using example protein with PDB ID 5xta.

[0179] FIG.42A is a ribbon diagram showing an example of a generated protein structures with 76 amino acids. - 35 - 11901435v1Attorney Docket No.2013969-0048

[0180] FIG.42B is a ribbon diagram showing an example of a generated protein structures with 107 amino acids.

[0181] FIG.42C is a ribbon diagram showing an example of a generated protein structures with 250 amino acids.

[0182] FIG.42D is a ribbon diagram showing an example of a generated protein structures with 365 amino acids.

[0183] FIG.43 is a ribbon diagram showing an example of a generated protein structures with 425 amino acids.

[0184] FIG.44 is a bar chart showing values of inter-residue clash scores (“between_clash_score”) and values of intra-residue clash scores (“within_clash_score”).

[0185] FIG.45 is a bar chart showing values for bond angle scores determined for N-Cα- C, Cα-C-N, and C-N- Cα bonds.

[0186] FIG.46 is a bar chart showing values for distance scores, in particular a consecutive residue Cα distance score (Cα- Cα distance score) and a C-N bond length score.

[0187] FIG.47 is a bar chart showing values for number of unsatisfied bonds, in particular a number of peptide bond lengths and Cα- Cα distances.

[0188] FIG.48 is a bar chart showing values for number of atoms clashing, in particular where inter-residue and intra-residue clash restrictions were.

[0189] FIG.49 is box plot showing statistics of per protein accuracy results for different tasks.

[0190] FIG.50 is box plot showing statistics of per protein similarity results for different tasks.

[0191] FIG.51 is box plot showing statistics of per protein accuracy results for various temperatures in binder design.

[0192] FIG.52 is box plot showing statistics of per protein similarity results for various temperatures in binder design

[0193] FIG.53 is box plot showing statistics of per protein accuracy on core and surface. - 36 - 11901435v1Attorney Docket No.2013969-0048

[0194] FIG.54 is a heatmap showing how amino acid probabilities vary in a 19- dimensional simplex at a particular site over the course of multiple time steps.

[0195] FIG.55 is a bar chart showing torsion angle accuracy for monomer side chain generation by percentage of surface residues.

[0196] FIG.56 is a bar chart showing torsion angle accuracy for monomer side chain generation by number of residues.

[0197] FIG.57 is a ribbon diagram showing side chain generation of PDB 4kdw using flow matching generative model overlaid with ground truth.

[0198] FIG.58 is a ribbon diagram showing side chain generation of PDB 3icw (chain B) using flow matching generative model overlaid with ground truth backbone and side chains.

[0199] FIG.59 is a box plot showing model distribution of violation values.

[0200] FIG.60 is a box plot showing PDB distribution of violation values.

[0201] FIG.61 is a ribbon diagram showing side chain generation of PDB 5nr7 (chain A) in orange as well as structure of backbone and side chains (chain B) in blue.

[0202] FIG.62 is a plot showing example trajectories of vector fields approximated by flow matching generative model on the complex plane to generate protein side chain torsion angles.

[0203] FIG.63 is a plot showing time step evolution of violation values of generated side chain conformation for various amino acid types.

[0204] FIG.64 is a sunburst plot showing selected hierarchy of CATH database used for fold conditioning.

[0205] FIG.65 is a box plot showing TM-scores for two models for fold conditioning.

[0206] FIG.66 is a box plot showing the highest TM-scores selected from ten generation for each fold family for two models used for fold conditioning.

[0207] FIG.67 is a box plot showing SSE accuracy for two models used for fold conditioning. - 37 - 11901435v1Attorney Docket No.2013969-0048

[0208] FIG.68 shows a collection of ribbon diagrams for various CATH proteins on the left and three different samples for each protein on the right.

[0209] Features and advantages of the present disclosure will become more apparent from the detailed description of certain embodiments that is set forth below, particularly when taken in conjunction with the figures, in which like reference characters identify corresponding elements throughout. In the figures, like reference numbers generally indicate identical, functionally similar, and / or structurally similar elements. CERTAIN DEFINITIONS

[0210] In order for the present disclosure to be more readily understood, certain terms are first defined below. Additional definitions for the following terms and other terms are set forth throughout the specification.

[0211] Comprising: A device, composition, system, or method described herein as “comprising” one or more named elements or steps is open-ended, meaning that the named elements or steps are essential, but other elements or steps may be added within the scope of the composition or method. To avoid prolixity, it is also understood that any device, composition, or method described as “comprising” (or which “comprises”) one or more named elements or steps also describes the corresponding, more limited composition or method “consisting essentially of” (or which “consists essentially of”) the same named elements or steps, meaning that the composition or method includes the named essential elements or steps and may also include additional elements or steps that do not materially affect the basic and novel characteristic(s) of the composition or method. It is also understood that any device, composition, or method described herein as “comprising” or “consisting essentially of” one or more named elements or steps also describes the corresponding, more limited, and closed-ended composition or method “consisting of” (or “consists of”) the named elements or steps to the exclusion of any other unnamed element or step. In any composition or method disclosed herein, known or disclosed equivalents of any named essential element or step may be substituted for that element or step.

[0212] A, an: As used herein, “a” or “an” with reference to a claim feature means “one or more,” or “at least one.” - 38 - 11901435v1Attorney Docket No.2013969-0048

[0213] Administration: As used herein, the term “administration” typically refers to the administration of a composition to a subject or system. Those of ordinary skill in the art will be aware of a variety of routes that may, in appropriate circumstances, be utilized for administration to a subject, for example a human. For example, in some embodiments, administration may be ocular, oral, parenteral, topical, etc. In some particular embodiments, administration may be bronchial (e.g., by bronchial instillation), buccal, dermal (which may be or comprise, for example, one or more of topical to the dermis, intradermal, interdermal, transdermal, etc.), enteral, intra-arterial, intradermal, intragastric, intramedullary, intramuscular, intranasal, intraperitoneal, intrathecal, intravenous, intraventricular, within a specific organ (e.g., intrahepatic), mucosal, nasal, oral, rectal, subcutaneous, sublingual, topical, tracheal (e.g., by intratracheal instillation), vaginal, vitreal, etc. In some embodiments, administration may involve dosing that is intermittent (e.g., a plurality of doses separated in time) and / or periodic (e.g., individual doses separated by a common period of time) dosing. In some embodiments, administration may involve continuous dosing (e.g., perfusion) for at least a selected period of time.

[0214] Affinity: As is known in the art, “affinity” is a measure of the tightness with which two or more binding partners associate with one another. Those skilled in the art are aware of a variety of assays that can be used to assess affinity, and will furthermore be aware of appropriate controls for such assays. In some embodiments, affinity is assessed in a quantitative assay. In some embodiments, affinity is assessed over a plurality of concentrations (e.g., of one binding partner at a time). In some embodiments, affinity is assessed in the presence of one or more potential competitor entities (e.g., that might be present in a relevant – e.g., physiological – setting). In some embodiments, affinity is assessed relative to a reference (e.g., that has a known affinity above a particular threshold [a “positive control” reference] or that has a known affinity below a particular threshold [a “negative control” reference”]). In some embodiments, affinity may be assessed relative to a contemporaneous reference; in some embodiments, affinity may be assessed relative to a historical reference. Typically, when affinity is assessed relative to a reference, it is assessed under comparable conditions.

[0215] Amino acid: The term “amino acid,” in its broadest sense, as used herein, refers to any compound and / or substance that can be incorporated into a polypeptide chain, e.g., through formation of one or more peptide bonds. In some embodiments, an amino acid has the - 39 - 11901435v1Attorney Docket No.2013969-0048 general structure H2N–C(H)(R)–COOH. In some embodiments, an amino acid is a naturally- occurring amino acid. In some embodiments, an amino acid is a non-natural amino acid; in some embodiments, an amino acid is a D-amino acid; in some embodiments, an amino acid is an L- amino acid. “Standard amino acid” refers to any of the twenty standard L-amino acids commonly found in naturally occurring peptides. “Nonstandard amino acid” refers to any amino acid, other than the standard amino acids, regardless of whether it is prepared synthetically or obtained from a natural source. In some embodiments, an amino acid, including a carboxy- and / or amino-terminal amino acid in a polypeptide, can contain a structural modification as compared with the general structure above. For example, in some embodiments, an amino acid may be modified by methylation, amidation, acetylation, pegylation, glycosylation, phosphorylation, and / or substitution (e.g., of the amino group, the carboxylic acid group, one or more protons, and / or the hydroxyl group) as compared with the general structure. In some embodiments, such modification may, for example, alter the circulating half-life of a polypeptide containing the modified amino acid as compared with one containing an otherwise identical unmodified amino acid. In some embodiments, such modification does not significantly alter a relevant activity of a polypeptide containing the modified amino acid, as compared with one containing an otherwise identical unmodified amino acid. As will be clear from context, in some embodiments, the term “amino acid” may be used to refer to a free amino acid; in some embodiments it may be used to refer to an amino acid residue of a polypeptide.

[0216] Antibody, Antibody polypeptide: As used herein, the terms “antibody polypeptide” or “antibody”, or “antigen-binding fragment thereof”, which may be used interchangeably, refer to polypeptide(s) capable of binding to an epitope. In some embodiments, an antibody polypeptide is a full-length antibody, and in some embodiments, is less than full length but includes at least one binding site (comprising at least one, and preferably at least two sequences with structure of antibody “variable regions”). In some embodiments, the term “antibody polypeptide” encompasses any protein having a binding domain which is homologous or largely homologous to an immunoglobulin-binding domain. In particular embodiments, “antibody polypeptides” encompasses polypeptides having a binding domain that shows at least 99% identity with an immunoglobulin binding domain. In some embodiments, “antibody polypeptide” is any protein having a binding domain that shows at least 70%, 80%, 85%, 90%, or 95% identity with an immuglobulin binding domain, for example a reference immunoglobulin - 40 - 11901435v1Attorney Docket No.2013969-0048 binding domain. An included “antibody polypeptide” may have an amino acid sequence identical to that of an antibody that is found in a natural source. Antibody polypeptides in accordance with the present invention may be prepared by any available means including, for example, isolation from a natural source or antibody library, recombinant production in or with a host system, chemical synthesis, etc., or combinations thereof. An antibody polypeptide may be monoclonal or polyclonal. An antibody polypeptide may be a member of any immunoglobulin class, including any of the human classes: IgG, IgM, IgA, IgD, and IgE. In certain embodiments, an antibody may be a member of the IgG immunoglobulin class. As used herein, the terms “antibody polypeptide” or “characteristic portion of an antibody” are used interchangeably and refer to any derivative of an antibody that possesses the ability to bind to an epitope of interest. In certain embodiments, the “antibody polypeptide” is an antibody fragment that retains at least a significant portion of the full-length antibody’s specific binding ability. Examples of antibody fragments include, but are not limited to, Fab, Fab’, F(ab’)2, scFv, Fv, dsFv diabody, and Fd fragments. Alternatively or additionally, an antibody fragment may comprise multiple chains that are linked together, for example, by disulfide linkages. In some embodiments, an antibody polypeptide may be a human antibody. In some embodiments, the antibody polypeptides may be a humanized. Humanized antibody polypeptides include may be chimeric immunoglobulins, immunoglobulin chains or antibody polypeptides (such as Fv, Fab, Fab', F(ab')2 or other antigen- binding subsequences of antibodies) that contain minimal sequence derived from non-human immunoglobulin. In general, humanized antibodies are human immunoglobulins (recipient antibody) in which residues from a complementary-determining region (CDR) of the recipient are replaced by residues from a CDR of a non-human species (donor antibody) such as mouse, rat or rabbit having the desired specificity, affinity, and capacity.

[0217] Approximately: As used herein, the term “approximately” or “about,” as applied to one or more values of interest, refers to a value that is similar to a stated reference value. In certain embodiments, the term “approximately” or “about” refers to a range of values that fall within 25%, 20%, 19%, 18%, 17%, 16%, 15%, 14%, 13%, 12%, 11%, 10%, 9%, 8%, 7%, 6%, 5%, 4%, 3%, 2%, 1%, or less in either direction (greater than or less than) of the stated reference value unless otherwise stated or otherwise evident from the context (except where such number would exceed 100% of a possible value). - 41 - 11901435v1Attorney Docket No.2013969-0048

[0218] Backbone, peptide backbone: As used herein, the term “backbone,” for example, as in a backbone or a peptide or polypeptide, refers to the portion of the peptide or polypeptide chain that comprises the links between amino acid of the chain but excludes side chains. In other words, a backbone refers to the part of a peptide or polypeptide that would remain if side chains were removed. In certain embodiments, the backbone is a chain comprising a carboxyl group of one amino acid bound via a peptide bond to an amino group of a next amino acid, and so on. Backbone may also be referred to as “peptide backbone”. It should be understood that, where the term “peptide backbone” is used, it is used for clarity, and is not intended to limit a length of a particular backbone. That is, the term “peptide backbone” may be used to describe a peptide backbone of a peptide and / or a protein.

[0219] Biologic: As used herein, the term “biologic” refers to a composition that is or may be produced by recombinant DNA technologies, peptide synthesis, or purified from natural sources and that has a desired biological activity. A biologic can be, for example, a protein, peptide, glycoprotein, polysaccharide, a mixture of proteins or peptides, a mixture of glycoproteins, a mixture of polysaccharides, a mixture of one or more of a protein, peptide, glycoprotein or polysaccharide, or a derivatized form of any of the foregoing entities. Molecular weight of biologics can vary widely, from about 1000 Da for small peptides such as peptide hormones to one thousand kDa or more for complex polysaccharides, mucins, and other heavily glycosylated proteins. In certain embodiments, a biologic is a drug used for treatment of diseases and / or medical conditions. Examples of biologic drugs include, without limitation, native or engineered antibodies or antigen binding fragments thereof, and antibody-drug conjugates, which comprise an antibody or antigen binding fragments thereof conjugated directly or indirectly (e.g., via a linker) to a drug of interest, such as a cytotoxic drug or toxin. In certain embodiments, a biologic is a diagnostic, used to diagnose diseases and / or medical conditions. For example, allergen patch tests utilize biologics (e.g., biologics manufactured from natural substances) that are known to cause contact dermatitis. Diagnostic biologics may also include medical imaging agents, such as proteins that are labelled with agents that provide a detectable signal that facilitates imaging such as fluorescent markers, dyes, radionuclides, and the like. In certain embodiments, biologics may, but need not necessarily be, used for medical applications. For example, biologics may include enzymes useful in e.g., industrial processes such as manufacturing, waste disposal, etc. In certain embodiments, a biologic may be a structural - 42 - 11901435v1Attorney Docket No.2013969-0048 protein or peptide (e.g., based on and / or analogues to collagen, keratin, etc.), which may be designed for medical, cosmetic, industrial, research, or other purposes.

[0220] In vitro: The term “in vitro” as used herein refers to events that occur in an artificial environment, e.g., in a test tube or reaction vessel, in cell culture, etc., rather than within a multi-cellular organism.

[0221] In vivo: As used herein, the term “in vivo” refers to events that occur within a multi-cellular organism, such as a human and a non-human animal. In the context of cell-based systems, the term may be used to refer to events that occur within a living cell (as opposed to, for example, in vitro systems).

[0222] Native, wild-type (WT): As used herein, the terms “native” and “wild-type” are used interchangeably to refer to biological structures and / or computer representations thereof that have been identified and demonstrated to exist in the physical, real world (e.g., as opposed to in computer abstractions). The terms, native and wild-type may refer to structures including naturally occurring biological structures, but do not necessarily require that a particular structure be naturally occurring. For example, the terms native and wild-type may also refer to structures including engineered structures that are man-made, and do not occur in nature, but have nonetheless been created and (e.g., experimentally) demonstrated to exist. In certain embodiments, the terms native and wild-type refer to structures that have been characterized experimentally, and for which an experimental determination of molecular structure (e.g., via x- ray crystallography) has been made.

[0223] Patient: As used herein, the term “patient” refers to any organism to which a provided composition is or may be administered, e.g., for experimental, diagnostic, prophylactic, cosmetic, and / or therapeutic purposes. Typical patients include animals (e.g., mammals such as mice, rats, rabbits, non-human primates, and / or humans). In some embodiments, a patient is a human. In some embodiments, a patient is suffering from or susceptible to one or more disorders or conditions. In some embodiments, a patient displays one or more symptoms of a disorder or condition. In some embodiments, a patient has been diagnosed with one or more disorders or conditions. In some embodiments, the disorder or condition is or includes cancer, or presence of one or more tumors. In some embodiments, the patient is receiving or has received certain therapy to diagnose and / or to treat a disease, disorder, or condition. - 43 - 11901435v1Attorney Docket No.2013969-0048

[0224] Peptide: The term “peptide” as used herein refers to a polypeptide that is typically relatively short, for example having a length of less than about 100 amino acids, less than about 50 amino acids, less than about 40 amino acids less than about 30 amino acids, less than about 25 amino acids, less than about 20 amino acids, less than about 15 amino acids, or less than 10 amino acids.

[0225] Polypeptide: As used herein refers to a polymeric chain of amino acids. In some embodiments, a polypeptide has an amino acid sequence that occurs in nature. In some embodiments, a polypeptide has an amino acid sequence that does not occur in nature. In some embodiments, a polypeptide has an amino acid sequence that is engineered in that it is designed and / or produced through action of the hand of man. In some embodiments, a polypeptide may comprise or consist of natural amino acids, non-natural amino acids, or both. In some embodiments, a polypeptide may comprise or consist of only natural amino acids or only non- natural amino acids. In some embodiments, a polypeptide may comprise D-amino acids, L- amino acids, or both. In some embodiments, a polypeptide may comprise only D-amino acids. In some embodiments, a polypeptide may comprise only L-amino acids. In some embodiments, a polypeptide may include one or more pendant groups or other modifications, e.g., modifying or attached to one or more amino acid side chains, at the polypeptide’s N-terminus, at the polypeptide’s C-terminus, or any combination thereof. In some embodiments, such pendant groups or modifications may be selected from the group consisting of acetylation, amidation, lipidation, methylation, pegylation, etc., including combinations thereof. In some embodiments, a polypeptide may be cyclic, and / or may comprise a cyclic portion. In some embodiments, a polypeptide is not cyclic and / or does not comprise any cyclic portion. In some embodiments, a polypeptide is linear. In some embodiments, a polypeptide may be or comprise a stapled polypeptide. In some embodiments, the term “polypeptide” may be appended to a name of a reference polypeptide, activity, or structure; in such instances it is used herein to refer to polypeptides that share the relevant activity or structure and thus can be considered to be members of the same class or family of polypeptides. For each such class, the present specification provides and / or those skilled in the art will be aware of exemplary polypeptides within the class whose amino acid sequences and / or functions are known; in some embodiments, such exemplary polypeptides are reference polypeptides for the polypeptide class or family. In some embodiments, a member of a polypeptide class or family shows significant sequence - 44 - 11901435v1Attorney Docket No.2013969-0048 homology or identity with, shares a common sequence motif (e.g., a characteristic sequence element) with, and / or shares a common activity (in some embodiments at a comparable level or within a designated range) with a reference polypeptide of the class; in some embodiments with all polypeptides within the class). For example, in some embodiments, a member polypeptide shows an overall degree of sequence homology or identity with a reference polypeptide that is at least about 30-40%, and is often greater than about 50%, 60%, 70%, 80%, 90%, 91%, 92%, 93%, 94%, 95%, 96%, 97%, 98%, 99% or more and / or includes at least one region (e.g., a conserved region that may in some embodiments be or comprise a characteristic sequence element) that shows very high sequence identity, often greater than 90% or even 95%, 96%, 97%, 98%, or 99%. Such a conserved region usually encompasses at least 3-4 and often up to 20 or more amino acids; in some embodiments, a conserved region encompasses at least one stretch of at least 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 or more contiguous amino acids. In some embodiments, a relevant polypeptide may comprise or consist of a fragment of a parent polypeptide. In some embodiments, a useful polypeptide as may comprise or consist of a plurality of fragments, each of which is found in the same parent polypeptide in a different spatial arrangement relative to one another than is found in the polypeptide of interest (e.g., fragments that are directly linked in the parent may be spatially separated in the polypeptide of interest or vice versa, and / or fragments may be present in a different order in the polypeptide of interest than in the parent), so that the polypeptide of interest is a derivative of its parent polypeptide.

[0226] Pose: As used herein, the term “pose” refers to a relative three-dimensional rotation and / or translation of one object, such as a polypeptide chain (e.g., protein) or peptide backbone thereof (e.g., represented by a scaffold model), with respect to another, such as a target molecule, e.g., a target protein. Accordingly, as used herein, for example in the context of molecular binding, where a particular protein and / or peptide backbone (e.g., a candidate peptide backbone) is referred to and / or described as being oriented at or according to a particular pose with respect to another protein (e.g., a target protein), it should be understood that the particular protein and / or peptide backbone may be rotated and / or translated in three dimensions relative to the other protein (e.g., target protein). A variety of manners and approaches may be used for representing poses, which may, for example, include representing rotations and / or transformations relative to a reference, such as the target and / or an initial pose of a scaffold - 45 - 11901435v1Attorney Docket No.2013969-0048 representation, using one or more fixed coordinate systems, etc. For example, in certain embodiments, a particular pose may be represented as a combination one or more rotations and / or translations in three-dimensional space, relative to a particular, reference object (e.g., such as a target protein) or coordinate (e.g., a location at or within a target protein representation), such as values of three rotational angles and / or three distances / angles defining a 3D translation (e.g., defined in a particular coordinate system, such as rectangular, cylindrical, spherical). In certain embodiments, a target and a scaffold model may be represented in a particular coordinate system, with the scaffold model oriented with respect to the target at a particular initial pose. Additional poses may then be represented relative to the initial pose, for example as relative translations and / or rotations (in three dimensions) with respect to the initial pose.

[0227] Protein: As used herein, the term “protein” refers to a polypeptide (i.e., a string of at least two amino acids linked to one another by peptide bonds). Proteins may include moieties other than amino acids (e.g., may be glycoproteins, proteoglycans, etc.) and / or may be otherwise processed or modified. Those of ordinary skill in the art will appreciate that a “protein” can be a complete polypeptide chain as produced by a cell (with or without a signal sequence), or can be a characteristic portion thereof. Those of ordinary skill will appreciate that a protein can sometimes include more than one polypeptide chain, for example linked by one or more disulfide bonds or associated by other means. Polypeptides may contain L-amino acids, D- amino acids, or both and may contain any of a variety of amino acid modifications or analogs known in the art. Useful modifications include, e.g., terminal acetylation, amidation, methylation, etc. In some embodiments, proteins may comprise natural amino acids, non-natural amino acids, synthetic amino acids, and combinations thereof. The term “peptide” is generally used to refer to a polypeptide having a length of less than about 100 amino acids, less than about 50 amino acids, less than 20 amino acids, or less than 10 amino acids. In some embodiments, proteins are antibodies, antibody fragments, biologically active portions thereof, and / or characteristic portions thereof.

[0228] Target: As used herein, the terms “target,” and “receptor” are used interchangeably and refer to one or more molecules or portions thereof to which a binding agent – e.g., a custom biologic, such as a protein or peptide, to be designed – binds. In certain embodiments, the target is or comprises a protein and / or peptide. In certain embodiments, the target is a molecule, such as an individual protein or peptide (e.g., a protein or peptide - 46 - 11901435v1Attorney Docket No.2013969-0048 monomer), or portion thereof. In certain embodiments, the target is a complex, such as a complex of two or more proteins or peptides, for example, a macromolecular complex formed by two or more protein or peptide monomers. For example, a target may be a protein or peptide dimer, trimer, tetramer, etc. or other oligomeric complex. In certain embodiments, the target is a drug target, e.g., a molecule in the body, usually a protein, that is intrinsically associated with a particular disease process and that could be addressed by a drug to produce a desired therapeutic effect. In certain embodiments, a custom biologic is engineered to bind to a particular target. While the structure of the target remains fixed, structural features of the custom biologic may be varied to allow it to bind (e.g., at high specificity) to the target.

[0229] Treat: As used herein, the term “treat” (also “treatment” or “treating”) refers to any administration of a therapeutic agent (also “therapy”) that partially or completely alleviates, ameliorates, eliminates, reverses, relieves, inhibits, delays onset of, reduces severity of, and / or reduces incidence of one or more symptoms, features, and / or causes of a particular disease, disorder, and / or condition. In some embodiments, such treatment may be of a patient who does not exhibit signs of the relevant disease, disorder and / or condition and / or of a patient who exhibits only early signs of the disease, disorder, and / or condition. Alternatively, or additionally, such treatment may be of a patient who exhibits one or more established signs of the relevant disease, disorder and / or condition. In some embodiments, treatment may be of a patient who has been diagnosed as suffering from the relevant disease, disorder, and / or condition. In some embodiments, treatment may be of a patient known to have one or more susceptibility factors that are statistically correlated with increased risk of development of a given disease, disorder, and / or condition. In some embodiments the patient may be a human.

[0230] Machine learning module, machine learning model: As used herein, the terms “machine learning module” and “machine learning model” are used interchangeably and refer to a computer implemented process (e.g., a software function) that implements one or more particular machine learning algorithms, such as an artificial neural networks (ANN), convolutional neural networks (CNNs), random forest, decision trees, support vector machines, and the like, in order to determine, for a given input, one or more output values. In some embodiments, machine learning modules implementing machine learning techniques are trained, for example using curated and / or manually annotated datasets. Such training may be used to determine various parameters of machine learning algorithms implemented by a machine - 47 - 11901435v1Attorney Docket No.2013969-0048 learning module, such as weights associated with layers in neural networks. In some embodiments, once a machine learning module is trained, e.g., to accomplish a specific task such as determining scoring metrics as described herein, values of determined parameters are fixed and the (e.g., unchanging, static) machine learning module is used to process new data (e.g., different from the training data) and accomplish its trained task without further updates to its parameters (e.g., the machine learning module does not receive feedback and / or updates). In some embodiments, machine learning modules may receive feedback, e.g., based on user review of accuracy, and such feedback may be used as additional training data, for example to dynamically update the machine learning module. In some embodiments, a trained machine learning module is a classification algorithm with adjustable and / or fixed (e.g., locked) parameters, e.g., a random forest classifier. In some embodiments, two or more machine learning modules may be combined and implemented as a single module and / or a single software application. In some embodiments, two or more machine learning modules may also be implemented separately, e.g., as separate software applications. A machine learning module may be software and / or hardware. For example, a machine learning module may be implemented entirely as software, or certain functions of a ANN module may be carried out via specialized hardware (e.g., via an application specific integrated circuit (ASIC), field programmable gate arrays (FPGAs), and the like).

[0231] Substantially: As used herein, the term “substantially” refers to the qualitative condition of exhibiting total or near-total extent or degree of a characteristic or property of interest.

[0232] Scaffold Model: As used herein, the term “scaffold model” refers to a computer representation of at least a portion of a peptide backbone of a particular protein and / or peptide. In certain embodiments, a scaffold model represents a peptide backbone of a protein and / or peptide and omits detailed information about amino acid side chains. Such scaffold models, may, nevertheless, include various mechanisms for representing sites (e.g., locations along a peptide backbone) that may be occupied by prospective amino acid side chains. In certain embodiments, a particular scaffold models may represent such sites in a manner that allows determining regions in space that may be occupied by prospective amino acid side chains and / or approximate proximity to representations of other amino acids, sites, portions of the peptide backbone, and other molecules that may interact with (e.g., bind, so as to form a complex with) a - 48 - 11901435v1Attorney Docket No.2013969-0048 biologic having the peptide backbone represented by the particular scaffold model. For example, in certain embodiments, a scaffold model may include a representation of a first side chain atom, such as a representation of a beta-carbon, which can be used to identify sites and / approximate locations of amino acid side chains. For example, a scaffold model can be populated with amino acid side chains (e.g., to create a ligand model that represents at least a portion of protein and / or peptide) by creating full representations of various amino acids about beta-carbon atoms of the scaffold model (e.g., the beta-carbon atoms acting as ‘anchors’ or ‘placeholders’ for amino acid side chains). In certain embodiments, locations of sites and / or approximate regions (e.g., volumes) that may be occupied by amino acid side chains may be identified and / or determined via other manners of representation for example based on locations of an alpha-carbons, hydrogen atoms, etc. In certain embodiments, scaffold models may be created from structural representations of existing proteins and / or peptides, for example by stripping amino acid side chains. In certain embodiments, scaffold models created in this manner may retain a first atom of stripped side chains, such as a beta-carbon atom, which is common to all side chains apart from Glycine. As described herein, retained beta-carbon atoms may be used, e.g., as a placeholder for identification of sites that can be occupied by amino acid side chains. In certain embodiments, where an initially existing side chain was Glycine, the first atom of glycine, which is hydrogen, can be used in place of a beta-carbon and / or, in certain embodiments, a beta carbon (e.g., though not naturally occurring in the full protein used to create a scaffold model) may be added to the representation (e.g., artificially). In certain embodiments, for example where hydrogen atoms are not included in a scaffold model, a site initially occupied by a Glycine may be identified based on an alpha-carbon. In certain embodiments, scaffold models may be computer generated (e.g., and not based on an existing protein and / or peptide). In certain embodiments, computer generate scaffold models may also include first side chain atoms, e.g., beta carbons, e.g., as placeholders of potential side chains to be added. DETAILED DESCRIPTION

[0233] It is contemplated that systems, architectures, devices, methods, and processes of the claimed invention encompass variations and adaptations developed using information from the embodiments described herein. Adaptation and / or modification of the systems, architectures, - 49 - 11901435v1Attorney Docket No.2013969-0048 devices, methods, and processes described herein may be performed, as contemplated by this description.

[0234] Throughout the description, where articles, devices, systems, and architectures are described as having, including, or comprising specific components, or where processes and methods are described as having, including, or comprising specific steps, it is contemplated that, additionally, there are articles, devices, systems, and architectures of the present invention that consist essentially of, or consist of, the recited components, and that there are processes and methods according to the present invention that consist essentially of, or consist of, the recited processing steps.

[0235] It should be understood that the order of steps or order for performing certain action is immaterial so long as the invention remains operable. Moreover, two or more steps or actions may be conducted simultaneously.

[0236] The mention herein of any publication, for example, in the Background section, is not an admission that the publication serves as prior art with respect to any of the claims presented herein. The Background section is presented for purposes of clarity and is not meant as a description of prior art with respect to any claim.

[0237] Documents are incorporated herein by reference as noted. Where there is any discrepancy in the meaning of a particular term, the meaning provided in the Definition section above is controlling.

[0238] Headers are provided for the convenience of the reader – the presence and / or placement of a header is not intended to limit the scope of the subject matter described herein.

[0239] Described herein are methods, systems, and architectures for Artificial- Intelligence (AI)-based design of custom biologics using one or more generative machine learning models (also referred to as “generative networks”). In particular, as described in further detail herein, in certain embodiments, technologies described herein apply a flow-matching technique to peptide backbone design, whereby machine learning models are trained and used to determine adjustments to positions and / or orientations of a collection of seed backbone sites and applied repeatedly to an initial starting distribution (e.g., an initial ‘guess’) to produce a final arrangement of backbone sites representing a newly generated de-novo peptide backbone. - 50 - 11901435v1Attorney Docket No.2013969-0048

[0240] Among other things, the present disclosure encompasses the insight that flow- matching, previously described for image generation in Y. Lipman et al., “Flow matching for generative learning,” the content of which is hereby incorporated by reference in its entirety, can be applied to protein structure design. Beginning with existing flow-matching framework, that have previously been limited to image generation, to create approaches suitable for protein design is non-trivial, and involved, among other things, creation of new featurization approaches to encode protein backbone structures as well as processes for training machine learning models to approximate velocity fields that are used to compute positional and / or orientation adjustments. These approaches are described in further detail herein. A. Flow Matching

[0241] Without wishing to be bound to any particular theory, turning to FIGs.1A-1D, given a dataset of samples, flow-matching techniques are based on the assumption that the samples of the dataset are or were generated according to (e.g., selected from) an underlying target probability distribution, p1. Accordingly, in certain embodiments, flow matching approaches seek to generate new samples, like those in the dataset, according to the underlying target probability distribution, p1, such that they have desired properties and / or are realistic, as embodied by the original examples of the dataset.

[0242] Flow matching methods address this challenge by beginning with an initial starting distribution, p0, and computing a target velocity field v(x, t) that, when applied to a sample x0selected from p0, over a period of time (e.g., from t = 0 to t = 1), will transform (e.g., “push”) it to a sample x1selected from target distribution p1. In particular, in certain embodiments, flow matching approaches train a machine learning model (e.g., a deep neural network) vθ(x, t) to approximate the target velocity field, v(x, t) (the subscript θ indicating / representing learnable parameters of the machine learning model).

[0243] FIGs.1A-1D are snapshots of graphs showing, at each of four time-steps, an example target distribution p1(left plot, red coloration), a current distribution (middle plot, blue coloration), and individual example points xt(right plot, blue dots). FIG.1A is a snapshot at t = 0 and, accordingly, shows an initial starting distribution, p0, which was selected to be a uniform distribution. FIGs.1B, 1C, and 1D show snapshots at subsequent time-steps, corresponding to - 51 - 11901435v1Attorney Docket No.2013969-0048 33%, 66%, and 100% of completion, illustrating how a current distribution and sample points evolve to match target distribution, p1.

[0244] Turning to FIG.2A, in certain embodiments, a machine learning model vθ(x, t) is trained to approximate a target velocity field using a training dataset, comprising a plurality of training examples, x1(i), and having a distribution pdata. At a particular training step, an initial seed sample, x0, is selected from initial starting distribution p0, a training example, x1(i)is selected from pdata, and a particular time point, t, selected at random, for example from a uniform distribution of times ranging from zero to one (e.g., U([0, 1])). In certain embodiments, seed sample x0, training example x1(i), and selected time t can be used to compute a ground truth intermediate position, xt(i), and velocity, v(xt(i), t), at the selected time point, along a path from x0to x1(i). For example, in certain embodiments, a target position and velocity are related via an ordinary differential equation (ODE), as shown below: ^^^^ ^^^^ = ^^^^( ^^^^, ^^^^) ^^^^ ^^^^ with x0 and x1(i)being known values of xt = 1, respectively.

[0245] In certain embodiments, a velocity field v(x, t) can be used to update values of feature vectors using a numerical approximation technique, e.g., for ODE solution, such as Euler’s method. In certain embodiments, one or more constraints may be placed on a path from x0to x1(i)and / or a target velocity field, for example to facilitate and / or obtain a particular form of v(x, t). For example, as shown in FIG.2B, in certain embodiments, a straight-line path – illustrated in 2D space, but, in certain embodiments, corresponding to a geodesic path along a 3D surface - may be used as a constraint (e.g., a hyperparameter) to facilitate evaluation of intermediate position xt(i)and velocity, v(x, t). In certain embodiments, velocity field may be assumed (e.g., constrained) to be constant with respect to time.

[0246] In certain embodiments, a machine learning model is used to compute, based on an intermediate position, xt(i)and time, t, an estimated velocity, vθ(xt(i), t). This estimated velocity (as computed by the machine learning model) can then be compared to the directly computed (as described above), ground truth, intermediate velocity and used to evaluate an error of the machine learning model’s estimated velocity, e.g., a loss function. For example, in certain - 52 - 11901435v1Attorney Docket No.2013969-0048 embodiments, a flow matching loss function may be determined according to the below equation: ^^^^ ^^^^ ^^^^ ^^^^ =‖^^^^ ^^^^(^^^^ ^^^^, ^^^^)− ^^^^ ^^^^‖2

[0247] In certain e.g., as described herein, may beused.

[0248] In certain embodiments, training parameters may be updated, and the above described process may be repeated for a plurality of training iterations, updating parameters of a machine learning model at each iteration (e.g., based on the loss function) to train the machine learning model to generate increasingly accurate estimates of target velocity field.

[0249] In certain embodiments, once trained, a machine learning model vθ(x, t) can be used for inference, to generate new data points that are not included within the original dataset, but fit within the target distribution, p1.

[0250] For example, turning to FIG.2C, in certain embodiments, at inference, a seed sample, having initial values selected, e.g., from a starting distribution (e.g., a uniform distribution) can be used as a starting point, x0~ p0. In inference process, rather than select an x1from a set of training examples, pdata, machine learning model is used to generate a new, not previously seen, x1by repeatedly computing velocity field estimates, which are then used to update feature values of seed sample, thereby evolving initial seed sample x0into a new, generated point x1. B. Peptide Backbone Encoding and Generation

[0251] Among other things, the present disclosure includes systems and methods for a flow-matching based framework that can be used to generate de-novo custom peptide backbones. In particular, approaches described herein include methods for encoding peptide backbone structures in feature vectors and for evolving these feature vectors via machine learning model deterred adjustments to create new custom peptide backbones that nonetheless follow rules and structural motifs of existing native peptides and, accordingly, are likely to be viable. - 53 - 11901435v1Attorney Docket No.2013969-0048 B.i Peptide Backbone Encoding

[0252] Turning to FIGs.3A and 3B, in certain embodiments, systems and methods of the present disclosure include approaches for representing peptide backbone sites. For example, as shown in FIG.3A, a peptide backbone site includes atoms that form a backbone portion of a peptide chain at each amino acid site (e.g., atoms that would remain if side chains were removed from each amino acid site). As shown in FIG.3A, a backbone site may include a nitrogen (N), an alpha-carbon (Cα), another carbon (C), and an oxygen (O), e.g., bound to each other as illustrated in FIG.3A. (a) Local Frame Representations

[0253] In certain embodiments, approaches described herein encode backbone sites using a plurality of feature vectors, each corresponding a particular backbone site and comprising a position and / or orientation component representing a position and / or orientation, respectively, of the corresponding backbone site. In this manner, in certain embodiments, a peptide backbone can be representing using plurality of feature vectors.

[0254] In certain embodiments, each feature vector comprises a position component, representing three-dimensional position of a backbone site. In certain embodiments, each feature vector comprises an orientation component, representing a 3D orientation of a corresponding backbone site. In certain embodiments, each feature vector comprises a position and anorientation component. For example, in certain embodiments, each feature vector has a form[ ^^^^, ^^^^], where T is a position component, representing a position of a corresponding backbonesite and R is an orientation component, representing a 3D orientation of the corresponding backbone site.

[0255] In certain embodiments, a local frame, Fi, can be determined for, and used to represent, each backbone site. In certain embodiments, local frames can be determined based on a three-dimensional [e.g., Cartesian, (x, y, z), coordinate] location of a Cαatom and orientations based on Cα-N and Cα-C axes. For example, in certain embodiments, a local frame can be determined as a frame having an origin coincident with Cαof a particular backbone site. For example, in certain embodiments, an orthonormal basis – a set of three vectors – can be used to - 54 - 11901435v1Attorney Docket No.2013969-0048 define a local frame for each backbone site. In certain embodiments, an orthonormal basis can be determined using positions of atoms in (e.g., represented via a structural model of a) backbone site as illustrated in FIG.3A. A first basis vector, v1, can be determined using positions of a Cαand N, as a normalized vector pointing from Cαto N. A second basis vector, v2, can be determined as a vector is orthogonal to v1and in a plane comprising v1and a Cα-C bond (and normalizing the result). A third basis vector, v3, can be determined by computing a cross product between v1and v2(and normalizing the result). In this manner, an orthonormal basis (v1, v2, v3) representing a local frame can be determined for each backbone site of a peptide chain. Additionally or alternatively, each local frame can be used to compute (e.g., approximate) positions of atoms of a corresponding backbone site, for example using ideal bond lengths and angles.

[0256] A peptide chain can, accordingly, be represented as a set of local frames, each representing a particular backbone site of the peptide chain and having a position and orientation. In certain embodiments, positions and orientations of each local frame Fiare represented in a relative fashion, from a point of view of each of one or more other local frames. For example, in certain embodiments, each local frame may be assigned an index (e.g., according to a sequence order of a corresponding backbone site that it represents), and its position and orientation represented as a relative translation and rotation from a point of view of a preceding local frame, Fi-1. That is, each feature vector encodes a position and orientation of a local frame according to ^^^^^^^^= [ ^^^^^^^^, ^^^^^^^^], where Tiand Riare a translation and rotation of Firelative to Fi-1. A translation may be determined using relative positions of the two origins of the two local frames, and a rotation may be determined using the two sets of orthonormal basis vectors.

[0257] For example, as illustrated in FIG.3B, local frame Fi+1can be represented via a translation and rotation relative to local frame Fi, local frame Fican be represented via a translation and rotation relative to local frame Fi-1, and so on. In this manner, backbone sites of a peptide chain can be represented in an entirely relative fashion that is invariant to global translation and / or rotation of the entire peptide chain. In certain embodiments, this invariance / equivariance with respect to global translations and / or rotations is advantageous, for example allowing use of machine learning (e.g., neural network) architectures that are not inherently equivariant. - 55 - 11901435v1Attorney Docket No.2013969-0048

[0258] In certain embodiments, each local frame is represented relative to multiple other frames. In this manner, multiple relative representations of each local frame can be fed into machine learning models described herein and used to generate multiple predictions of positions and / or orientations of backbone sites and / or adjustments thereto, allowing for a consensus or average of the predictions to be used / selected.

[0259] Additionally or alternatively, in certain embodiments, positions and / or orientations of local frames may be represented relative to their own positions and / or orientations at other time points / iterations. For example, a position and / or orientation of a particular local frame at a particular time point or iteration may be represented as a translation and / or rotation, respectively, relative to that particular local frame’s position and / or orientation and one or more previous time points or iterations.

[0260] In certain embodiments, each local frame’s position and orientation can be represented in a global fashion, relative to a single common global reference point / axis, such as a common origin and set of three unit vectors (e.g., Cartesian unit vectors (nx, ny, nz)).

[0261] Accordingly, in certain embodiments, a peptide chain can be parameterized using a set or list of feature vectors, each representing a translation and rotation of a particular local frame that represents a particular backbone site. Each feature vector may, e.g., be or be viewed as a row in a matrix, and take the form ^^^^( ^^^^)= ^^^^^^^^=[^^^^^^^^, ^^^^^^^^], for an ithrow / feature vector representing an ithbackbone site,and R are the translation and rotation of the local frame. A three-dimensional translation, T, may be represented via three values, each representing a shift along a particular coordinate, such as T = (Tx, Ty, Tz). A rotation R may be represented via a quaternion (e.g., a set of four values representing a rotation about an axis u by an angle theta), such as R = (q1, q2, q3, q4). Accordingly, in certain embodiments, each feature vector may be or comprise a seven-element vector, and a peptide chain having length k represented by k feature vectors, e.g., a 7 by k matrix. (b) Individual Backbone Atom Representations

[0262] In certain embodiments, (e.g., additionally or alternatively to using a local frame- based representation as described herein) atoms of backbone sites – i.e., N, Cα, C, and O atoms – - 56 - 11901435v1Attorney Docket No.2013969-0048 can be represented directly, via their location in three-dimensional space, for example via a Cartesian (x, y, z) coordinate for each atom. In this manner, each backbone site may be represented via a twelve-element coordinate vector, such as x(i)= [Ni, Cαi, Ci, Oi], where, for example, Ni, Cαi, Ci, and Oiare (x, y, z) coordinates of the N, Cα, C, and O atoms of the ithamino acid site. Accordingly, in certain embodiments, each feature vector may be or comprise a twelve-element vector representing a position of each of the four backbone atoms, and a peptide chain having length k represented by k feature vectors, e.g., a twelve by k matrix. (c) Characteristics and Benefits of Local Frame and Individual Backbone Atom Representations

[0263] In certain embodiments, an individual (four) backbone atom representation can be advantageous, since it represents backbone sites via a list of coordinates in 3D space. Each of these coordinate values thus lie in ℝ3(real coordinate space), which is relatively easy to manipulate and is linear, thereby avoiding more complex mathematical formulations which, for example rely on logarithmic and / or exponential mapping tools. Moreover, in certain embodiments, by representing each atom of each backbone site directly, an individual backbone atom representation approach does not use assumptions, such as perfect bond length and angles. Individual backbone atom representation approaches, however, do not inherently enforce a fixed relationship between relative positions of individual backbone sites, and, accordingly, increase a number of dimensions to be learned by machine learning models described herein. For example, in certain embodiments, learned velocities via flow matching techniques described herein use twelve (12) coordinates per amino acid site when an individual backbone representation is used, as opposed to e.g., six, when a local frame representation is used.

[0264] Accordingly, in certain embodiments, local frame representations can be advantageous in that they simplify the problem in terms of the number dimensions to use. Additionally or alternatively, in certain embodiments, local frame representations can facilitate creation of equivariant machine learning architectures from invariant features, for example, as described herein with respect to self-attention based SO(3) transformer versus applying frame rotation. - 57 - 11901435v1Attorney Docket No.2013969-0048

[0265] In certain embodiments, approaches described herein address challenges associated with use of a local frame representation. Among other things, in certain embodiments, where local frame representations are used, machine learning models described herein utilize a logarithmic mapping technique, which can become undefined when an angle of rotation approaches π. This can, in turn, cause velocities to be undefined, leading to numerical instabilities. Additionally, or alternatively, machine learning approaches described herein account for a two-to-one (2-to-1) mapping where quaternion representations are used (e.g., since q and -q represent a same rotation). As described herein, local frame representation approaches also may include techniques to recover individual atom positions, for example based on assumptions of (e.g., ideal) bond lengths and angles, from local frame Ca coordinates and rotations. B.ii Training and Inference

[0266] In certain embodiments, flow matching technologies for generating de-novo peptide backbones of the present disclosure include systems and methods for training machine learning models, such as neural networks (e.g., deep neural networks), to model velocity fields that can be used to adjust positions and / orientations of feature vectors representing peptide backbone sites to create new backbone structures. (a) Training

[0267] In certain embodiments, a machine learning model is trained to model a velocity field using 3D protein backbone structures, for example, selected as examples from data sources such as the protein data bank (PDB). For example, a training dataset may be created from 3D structures of proteins by converting each structure into a list of feature vectors matching a desired parameterization, such as the local frame representation described above.

[0268] In certain embodiments, at a particular training step, an example representation of a backbone structure, x1(i), is selected from a training dataset. As described herein, in certain embodiments, x1(i)is or comprises a plurality of feature vectors, each representing a corresponding backbone site via a local frame. In certain embodiments, e.g., as described herein - 58 - 11901435v1Attorney Docket No.2013969-0048 with respect to FIG.2A, a time point may be selected, for example at random from a uniform distribution.

[0269] In certain embodiments, a seed set comprising a plurality of feature vectors is used as an initial starting point, x0. Values of position and / or orientation components of each feature vector of seed set may be set to initial starting values. Initial starting values for each feature vector may be selected at random, for example from an initial probability distribution, such as a uniform distribution of position and / or orientations. In certain embodiments, initial starting values may be determined, for example, using a preliminary machine learning model (different from a machine learning model used to model velocity field vθ(x, t)). A preliminary machine learning model may be, for example, be trained to generate a prior distribution that more closely matches a target distribution p1than, e.g., random selection from a uniform distribution, thereby providing an improved ‘initial guess’ as a set of initial starting values.

[0270] In certain embodiments, based on the initial starting point x0(i), the selected time point t, and training example x1(i), an intermediate feature set xt(i)and velocity, v(xt(i), t), can be calculated (e.g., using x(i)(t = 0) = x0(i)to x(i)(t = 1) = x1(i)), for example using ODEsolution approaches such as Euler’s method and, additionally or alternatively, various constraints, such as particular forms of paths from x0to x1(i), functional forms of velocity field ^^^^ ^^^^( ^^^^) v(x, t), etc. In certain embodiments, a (e.g., analytical) derivative of position v(xt(i), t) =^^^^^^^^ ^^^^ can be used if a closed form exists. Additionally or alternatively, derivatives can be^^^^( ^^^^)−( ^^^^)for example via v(xt(i), t) ≈^^^^+∆ ^^^^^^^^^^^^∆ ^^^^ . Intermediate feature set xt(i)comprises updated values for position and / or orientation components of the plurality of feature vectors in the initial seed set, and, accordingly, represents intermediate positions and / or orientations of each backbone site of a peptide backbone being generated, at a position (in feature space) along a path from the initial starting values to the final positions and orientations of the backbone sites as they are in the selected training example, x1(i)– e.g., a real peptide chain. Determined velocity, v(xt(i), t), represents an adjustment to each feature vector – i.e., to a position and orientation of each backbone site, in a direction of their final positions and orientations.

[0271] In certain embodiments, intermediate feature set xt(i)and selected time t can be used as input to a machine learning model that determines, as output, an estimated velocity - 59 - 11901435v1Attorney Docket No.2013969-0048 vθ(xt(i), t) associated with intermediate feature set xt(i)and time t. This estimated velocity, as determined by the machine learning model (at a particular training iteration), can then be compared with directly computed velocity v(xt(i), t), to evaluate a loss or error function at the particular training step. This process can be repeated, for a number of training steps, as parameters of the machine learning model are adjusted to obtain increasingly accurate estimations of velocity, until a desired number of evaluations and / or a desired accuracy is reached. In this manner, a machine learning model can be trained to determine, for example, for a particular feature set (e.g., a position and orientation of a plurality of local frames representing peptide backbone sites) and time, a velocity that can be used to update the particular feature set and “push” its values towards those that mirror those of the training examples – i.e., real physical peptide chains.

[0272] In certain embodiments, training does not necessarily proceed one example at a time. For example, in certain embodiments, a batch of a plurality of training examples and a plurality of sets of initial starting values are selected. Each set of initial starting values may then be matched with a particular training example that it is nearest to, e.g., using a Sinkhorn algorithm, to improve efficiency of training. (b) Inference

[0273] Accordingly, in certain embodiments, a trained machine learning model can be used to compute velocities that can be used to update position and / or orientation component values of feature vectors so as to progressively ‘push’ them from initial values representing, e.g., a random distribution of backbone sites or an initial guess, to a set of final values that represent positions and orientations that are indicative / representative of real, physical peptide chains. In this manner, peptide backbones created via generative techniques described herein may be new, not previously seen, yet nonetheless have a high likelihood of being viable, functioning (e.g., properly folded / stable) peptide backbones.

[0274] For example, as shown in FIG.4, to generate a new, custom peptide backbone, approaches described herein may begin with a seed set comprising a plurality of feature vectors, each having position and / or orientation components representing positions and / or orientations, respectively, of a corresponding peptide backbone site. Initial values of position and / or - 60 - 11901435v1Attorney Docket No.2013969-0048 orientation components of each feature vector may be determined in a variety of manners, as described herein, such as at random, from an initial distribution. Accordingly, for example, seed set may begin, e.g., at time t = 0, by representing a plurality of peptide backbone sites having random positions and / or orientations – e.g., random noise. In certain embodiments, beginning with the initial starting values of the position and / or orientation components of the plurality of feature vectors, at a first iteration, a machine learning model trained as described herein to estimate a velocity field may be used to compute an initial velocity filed, based on the initial values of the position and orientation (e.g., as well as the selected initial time point, t = 0). The initial velocity field may then be used to update values of the position and / or orientation components of the plurality of feature vectors. These updated values may be used as current values of position and / or orientation components of the feature vectors at a subsequent iteration, which, in turn, can be used (e.g., together with an incremented, current time point) to determine a current velocity field. The current values of the position and / or orientation components can then be updated according to the current velocity field, and the process repeated in an iterative fashion, until a final iteration (e.g., time step) is reached.

[0275] Final values of the feature vectors can then be used to generate a scaffold model representing a final, generated, custom peptide backbone, for example by computing backbone atom positions from each local frame representation. (c) Approaches for Velocity Field Estimation

[0276] In certain embodiments, one or more (e.g., different) approaches may be used for determining velocity fields using a machine learning model. In certain embodiments, a machine learning model is used to directly estimate a velocity field at each iteration. For example, a machine learning model may receive, as input, a current set of feature vector values and a current time point and generate, as output, a current velocity field that can be used to update feature vector values, e.g., for use in subsequent iterations.

[0277] Additionally or alternatively, in certain embodiments, a machine learning model may not necessarily directly output an estimated velocity field, but, rather, may output a set of prospective final feature vector values – e.g., a current estimate of x1, representing a current estimate of positions and / or orientations of each backbone site of a generated peptide backbone, - 61 - 11901435v1Attorney Docket No.2013969-0048 based on current feature vector values at a current time point. A current velocity field may then be determined based on the set of prospective final feature vector values and a current time point [e.g., corresponding to a current iteration (e.g., generating a current estimate of x1as output of the machine learning model and computing the current velocity Field based on x1and the current time point)]. This approach may proceed iteratively, with the machine learning model repeatedly generating refined calculations of prospective final feature vector values, which are then used to compute a current velocity field and adjust position and / or orientation components of the feature vectors.

[0278] In certain embodiments, approaches wherein a machine learning model generates a prospective / current prediction of x1as output may use a single type of local frame representation and / or may combine multiple representations. For example, in certain embodiments, initial starting values of feature vector position and / or orientation components may represent position and / or orientations of local frames in a (e.g., purely) spatial fashion – e.g., relative to one or more neighboring frames and / or to a common global reference. In certain embodiments, a machine learning model may generate output predictions of x1using a same representation as the initial starting values, e.g., also in a spatial fashion. In certain embodiments, a machine learning model may generate output predictions of x1using a different representation as the initial starting values, for example by computing, for each frame i, a translation and / or rotation of that frame at time t = 1 (a final time point) relative to its position and / or orientation at a current time point. C. Amino Acid Sequence Encoding and Generation

[0279] Additionally or alternatively, generative, flow-matching-based technologies of the present disclosure may be used to generate (e.g., de-novo) amino acid sequences for custom proteins and / or peptides. In certain embodiments, flow-matching methods and systems for generation of amino acid sequences may be implemented and / or used as dedicated (e.g., separate) systems and methods, for example via dedicated machine learning models, distinct from those used for design of other polypeptide features, such as backbone design and / or side- chain packing, described herein. In certain embodiments, as described in further detail herein, flow-matching technologies for generative design of amino acid sequences may be combined - 62 - 11901435v1Attorney Docket No.2013969-0048 with other flow-matching prediction technologies, such as peptide backbone design and / or side chain packing techniques, via a single (e.g., unified) multi-input and / or multi-task model. In this manner, sequence design technologies of the present disclosure can be used alone and / or in connection with other design modules and / or techniques to provide a range of capabilities in the context of in-silico design of custom biologics. C.i Sequence Encoding

[0280] In certain embodiments, sequence design technologies of the present disclosure leverage particular representations of polypeptide sequences to address a key challenge in flow- matching-based generative design of amino acid sequences – namely, the categorical, discrete nature of amino acid sequences. That is, for example, proteins are large biomolecules composed of one or more lengthy polypeptide chains, each comprising a plurality of amino acid residues linked together via peptide bonds. Accordingly, an amino acid sequence of a polypeptide chain and / or an overall protein represents order and type of the multiple amino acid residues that the one or more chains of a protein and / or peptide molecule comprises. An amino acid sequence of a polypeptide chain can, accordingly, be represented as an ordered list of characters, numbers, or other variables, each having a particular value selected from a discrete set representing possible amino acid types. A set of possible amino acid types may be or comprise the twenty standard amino acids and non-standard amino acids.

[0281] For example, an amino acid sequence may be represented as a list of three- character strings, each representing and identifying a particular one of the twenty standard amino acid, e.g., “Glu Ala Gln Ile Thr Gly …Thr Leu…” Additionally or alternatively, amino acids may be represented via single characters, such that this same sequence may be represented via a string, “EAQITG…TL…”. In certain embodiments, an amino acid sequence of a polypeptide chain can be represented as an ordered list of integers, each integer identifying a particular one of the twenty standard amino acids. For example, an amino acid sequence may be represented asshown below{^^^^ ^^^^} ^^^^^^^^=1 , 0 s.t., ri ∈ {1, ⋯ , 20},- 63 - 11901435v1Attorney Docket No.2013969-0048 where n is a number of amino acid residues in the polypeptide chain and riis a categorical type of amino acid present at that position, constrained to the set {1,…, 20}, reflecting the 20 standard amino acids. This representation effectively maps the linear sequence of a protein to a numerical sequence, with each number corresponding to a distinct type of amino acid.

[0282] Applications of flow-matching techniques typically involve generating continuous variables, such as real numbers. Applying flow matching to discrete domains presents a complex challenge. This difficulty primarily arises because flow-matching is based on modeling continuous flow of probability paths and is facilitated by solution of differential equations in this, continuous, context. Setting up differential equations on discrete domains, such as polypeptide sequence representations, is non-trivial, since, among other things, derivatives are not straightforwardly defined on discrete domains. Accordingly, the continuous nature of generative modeling techniques based on flow-matching presents a fundamental challenge to their use where quantities to be modelled take on and / or are restricted to a finite number of discrete values.

[0283] Accordingly, in certain embodiments, technologies of the present disclosure for amino acid sequence generation include approaches for representing amino acid sequences on continuous domains. For example, in certain embodiments systems and methods of the present disclosure include steps of converting representations of amino acid residues from discrete to continuous values.

[0284] In particular, in certain embodiments, amino acid sequence generation technologies described herein utilize a relaxed continuous domain of simplices for modeling amino acid types, which facilitates use of flow matching-based machine learning models for generation of amino acid sequences of polypeptide chains. In certain embodiments, a domain of residue types is transformed from discrete values to a representation within a 19-dimensional simplex, thereby effectively treating each amino acid type as a point in a continuous probability distribution. This approach models rias an element in ℝ20, subject to conditions that ||ri||1= 1 and each component of the rivector lies (i.e., takes on a value) between 0 and 1. Among other things, this method establishes a one-to-one correspondence between 20-dimensional categorical probability distributions and points within the simplex, underscoring the suitability of the simplex domain for the sequence generation technologies of the present disclosure. In this - 64 - 11901435v1Attorney Docket No.2013969-0048 manner, a polypeptide (e.g., protein and / or peptide) sequence may be represented via a set of twenty-element feature vectors, each corresponding to and representing a particular amino acid type at a particular location (e.g., site) within a protein and / or peptide. Each of the twenty amino acid types may be associated with a particular index or location within a feature vector, and each of the twenty elements of a particular feature vector may have a value between 0 and 1, representing a probability of a particular type of amino acid. Accordingly, a sequence of a protein having k amino acids may be represented via a set of k 20-element feature vectors, or a k × 20 size matrix. Machine learning models may then be used to generate velocity fields for iteratively updating values of these, sequence, feature vectors, updating probabilities of various types of amino acids at sites across a polypeptide chain, until they, e.g., coalesce into final values peaked at a one or a few likely amino acid types. Feature vectors may then be converted into a final identification of a particular, single, amino acid type via, for example, an argmax function. C.ii Training and Inference (a) Training

[0285] Turning to FIG.5, sequence design approaches of the present disclosure, may, accordingly, utilize continuous representations of protein sequences, such as the simplex approach described herein, to leverage flow matching methods and systems described herein with regard to generation of peptide backbone structures. In particular, an interpolant or path constraint, ϕ, representing a path interpolating between an initial seed point, x0,selected from an initial starting distribution p0, and a final target, x1, selected from a dataset of training examples, pdata. may be used to map points from an initial starting distribution to a desired, target distribution represented and / or approximated by a dataset. See, e.g., [Chen, 2016, “Neural Ordinary Differential Equations”].

[0286] Adopting the linear continuous interpolation methodology (e.g., used for backbone Cαatoms, i.e. corresponding to Euclidean manifold) for sequence generation is facilitated via the simplex domain representation approach described above, which constitutes a convex set within the Euclidean space. In particular, in certain embodiments, a straight line, linear interpolation, - 65 - 11901435v1Attorney Docket No.2013969-0048 ∀ ^^^^ ∈ [0,1], ϕt(x0|x1) = (1-t)×x0+t×x1, may be used. Without wishing to be bound to any particular theory, it is believed that the entire interpolation path remains within the simplex for all times t, since the interpolated point is a convex combination of points, which, by definition of convex sets, is always contained within the simplex, underscoring the suitability of simplex as the domain for this application. In certain embodiments, any path deviation from a simplex (e.g., due to a numerical error) may be corrected (e.g., adjusted) to be back on the simplex. In certain embodiments, an initial starting distribution from which a seed points x0are selected is a uniform distribution on a simplex, for example, uniform distribution on the simplex: i.e., x0~ p0= ^^^^ir (α = 120= (1, 1, …, 1) ∈ ℝ20, where ^^^^ir denotes the Dirichlet distribution. In certain embodiments, other initial starting distributions (e.g., not necessarily uniform; distributions that a priori introduce specific bias (e.g., used for conditioning); different Dirichlet distributions; different distributions across amino acid types; completely different distribution on the domain of simplex distribution) may be used.

[0287] FIG.5 shows an example of a uniform distribution on a simplex evolving to a final value at x1= [0,0,1].

[0288] Accordingly, by virtue the continuous representation of amino acid sequences described herein, flow-matching techniques for sequence generation can be trained in a similar manner to those described above with respect to peptide backbone generation approaches. That is, a particular training example x1(i)may be selected from a training dataset, having a distribution pdata. An initial seed point, x0(i), may be selected from an initial starting distribution, here a uniform distribution on a 19-dimensional simplex, and a (e.g., arbitrary / random) time point t selected. In the context of sequence generation, points x1(i)and x0(i)comprise sequence feature sets, representing an example amino acid sequence selected from the training dataset and a random initial starting point, where at each amino acid site probabilities across different amino acid types are random, based on a uniform distribution. From these selected points, and the interpolant form, ϕ, an intermediate feature set xt(i)and intermediate velocity, v(xt,t) (e.g., a sequence velocity) may be determined. As with the peptide backbone generation approaches described herein, the machine learning model may then be tasked with determining velocity estimates, vθ(xt(i), t), (e.g., sequence velocity estimates) and / or outputs from which they - 66 - 11901435v1Attorney Docket No.2013969-0048 (sequence velocity estimates) can be determined, which can, in turn, be compared with the determined intermediate velocity, via a loss function.

[0289] In certain embodiments, for example when a simplex approach as described herein is used, an intermediate velocity v(xt, t) = ut= x1– x0may be determined directly from x1and x0.

[0290] A variety of approaches may be used to determine estimated velocity fields vθ(xt(i), t), from a machine learning model for sequence generation. In particular, a machine learning model, such as various invariant models described herein may generate an invariant representation, ht. In certain embodiments, representation htis invariant by virtue of a particular, invariant, machine learning architecture. This invariant representation may itself be used as output, for example such that the machine learning model directly generates a predicted velocity, or one or more additional operators (e.g., softmax, sp, clipping, etc.) may be applied to generate various desired forms of output.

[0291] Four example approaches for generating estimated velocity fields via a machine learning model are listed below, without limitation.

[0292] 1. Delta Prediction. In certain embodiments, an approach, referred to herein as delta prediction, is used whereby a machine learning model generates (e.g., as output) a prospective final sequence feature vector – i.e., an estimate of x1(i), denoted ^̅^^^1. Predicted target values may be obtained from an invariant representation, ht, generated within a machine learning model via one or more differentiable projection operations that map ht to a 19-dimensional simplex, such as a softmax operation. Sequence velocity fields may then be generated from the prospective final sequence representation according to a desired path, for example, using the linear path in Euclidean space described herein, as ^^^^(^^^^ ^^^^, ^^^^)=^̅^^^1− ^^^^ ^^^^1− ^^^^ .

[0293] 2. Double Prediction. In certain embodiments, an invariant representation, ht, may be divided into two separate vectors and projected onto a (e.g., continuous) representation of an amino acid sequence, such as a simplex. Without wishing to be bound to any particular theory, for example, it is believed that this double prediction approach mirrors the form of a velocity field in a linear approximation, where it is computed as a difference between two - 67 - 11901435v1Attorney Docket No.2013969-0048 sequence feature vectors. For example, in certain embodiments htis a 40-dimensional invariant representation, which may be projected onto a (e.g., 19 dimensional) simplex via a sp operator (e.g., any operator that transforms 20-dimensional vector into a simplex, e.g., such that, values of resultant vectors are between 0 and 1 and all sum to 1), as shown below: ^^^^(^^^^ ^^^^, ^^^^)= ^^^^ ^^^^(ℎ ^^^^[0: 20])− ^^^^ ^^^^(ℎ ^^^^[20: 40]).

[0294] 3. Raw Prediction. In certain embodiments, an invariant representation, ht, is used directly as output, such that vθ(xt(i), t) = ht. In this manner, the machine learning model learns (e.g., via training) to generate, directly velocity field vectors.

[0295] 4. Raw Prediction with max norm clipping. In certain embodiments, a raw prediction, e.g., vθ(xt(i), t) = ht, is clipped to predict vectors having a particular maximum norm, for example a maximum L2 norm of √2, to mirror and / or enforce a desired norm known to be true for a ground truth velocity field.

[0296] In certain embodiments, machine learning model-based predictions may be evaluated, e.g., during training, via a variety of loss functions, which, in turn, may be used to update model weights and refine predictions during a training procedure. Loss functions may be structured to determine based on velocity field estimates and / or directly based on prospective final sequence representations.

[0297] For example, in certain embodiments, where amino acid sequences are represented via simplexes, an expected velocity given a particular position and time takes on the below relationship, ^^^^(^^^^, ^^^^)= ^^^^[^^^^1 − ^^^^0|^^^^ ^^^^ = ^^^^]= ^^^^�^^^^1 − ^^^^−^^^^^^^^� ^^^^ ^^^^ = ^^^^�

[0298] sequence feature vector may be considered equivalent, and, accordingly, loss functions can be structured based on either approach.

[0299] For example, where a machine learning model is used to estimate a velocity field, the following two loss functions may be used: - 68 - 11901435v1Attorney Docket No.2013969-0048 1. MSE loss function: ^^^^ = ^^^^ ^^^^~[0,1], ^^^^0~ ^^^^0||(^^^^1 − ^^^^0)− ^^�^^(^^^^ ^^^^, ^^^^)||22.2. Cross entropy^^^^ = ^^^^^^^^~[0,1], ^^^^0~ ^^^^0^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^� ^^^^1, ^^^^^^^^+ ^^�^^( ^^^^^^^^, ^^^^) × (1 − ^^^^)�.

[0300] In certain embodiments, where a machine learning model is used to estimate a prospective final sequence representation, ^^�^^1, one or more of the following loss functions may be used: 1. MSE loss function: ^^^^^^^^ ^^^^ ^^^^= ^^^^^^^^~[0,1], ^^^^0~ ^^^^0|�( ^^^^1− ^^^^0) −^^�^^1( ^^^^ ^^^^, ^^^^)− ^^^^ ^^^^ 21− ^^^^� |2. 2. Cross^^^^ ^^^^ ^^^^ = ^^^^ ^^^^~[0,1], ^^^^0~ ^^^^0 ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^� ^^^^1, ^^�^^1(^^^^ ^^^^, ^^^^)�.3. Sum of^^^^^^^^ ^^^^ ^^^^= ^^^^ ^^^^^^^^ ^^^^ ^^^^+ ^^^^ ^^^^^^^^ ^^^^. 4. Cross entropy with raw prediction: ^^^^ = ^^^^^^^^~[0,1], ^^^^0~ ^^^^0^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^( ^^^^1,ℎ^^^^), where htrepresents the outputted invariant

[0301] In certain embodiments, a non-uniform time sampling is used during training, for example to account for the categorical distribution of amino acid sequences, particularly where times lie close to 1. In particular, here, at times close to 1 the argmax function effectively minimizes the loss function – accordingly, where these (close to 1) times are used, the machine learning model may be essentially trained replicate an argmax function.

[0302] For example FIG.6 shows a plot produced by sampling 100 points from a Dirichlet distribution (e.g., ^^^^ir (α = 120= (1, 1, …, 1) ∈ ℝ20) and an arbitrary, p1distribution, taken to be uniform ei∈ {1,…,20} to simulate a flow matching path pt. The graph plots accuracies determined as argmax(xt) = argmax(x1) for any discretization over 50 time steps. As shown in FIG.6, accuracy of the argmax function rapidly approaches 1 for times over about 0.12. Accordingly, sampling time uniformly between 0 and 1 can significantly slow down - 69 - 11901435v1Attorney Docket No.2013969-0048 training, since the functional form that minimizes the loss at later times is captured by the relatively simple argmax function, while desired, more complex behavior at earlier times may be under-represented in the training examples. In certain embodiments, sequence prediction training approaches of the present disclosure, accordingly, sample time non-uniformly during training, to place a greater emphasis on values closer to t → 0 and less on values approaching t→ 1. For example, in certain embodiments, a time value, t, may be sample from a particular distribution, such as a Beta distribution [e.g., B(α, β) where α < β], an exponential decay distribution [e.g., t = f(x) = exp(-λx), where λ > 1 and x ~ U(0, α) for any α > 100], and the like. (b) Inference

[0303] In certain embodiments, a machine learning model having been trained, for example as described herein, to determine predictions of velocity field and / or final sequence representations, may be used to generate polypeptide sequences via a flow-matching approach. In certain embodiments, generating (e.g., de-novo) polypeptide sequences comprises using a machine learning model to determine estimates of velocities and / or final sequence representations for use in an ordinary differential equation (ODE) solver, such as a Euler or Runge-Kutta method.

[0304] For example, in certain embodiments, beginning with an initial, starting, sequence representation, x0, (e.g., a continuous representation of an amino acid sequence, such as a simplex representation as described herein), an ODE solution approach may be used to integrate velocity predictions obtained via a machine learning model to evolve x0to a final, generated sequence representation, x1. In certain embodiments, determining a final sequence representation x1 comprises repeatedly generating velocity field predictions using a machine learning model and updating a sequence representation, e.g., from xtto xt+Δt.

[0305] In certain embodiments, a sequence representation xtis updated using a linear push forward function. In certain embodiments, where a machine learning model is used to determine and estimate of a velocity field, e.g., directly, sequence representations may be updated via a linear push forward function according to the following equation: ^^^^^^^^+∆ ^^^^= ^^^^^^^^+ ^^�^^( ^^^^, ^^^^)∆ ^^^^. - 70 - 11901435v1Attorney Docket No.2013969-0048

[0306] In certain embodiments, where a machine learning model estimates a prospective final sequence representation, the following equations may be used, e.g., to first determine a velocity field based on the prospective final sequence representation, and then used the determined velocity field to update the sequence representation: ^^�^^( ^^^^^^, ^^^^) =^�^^1 − ^^^^ ^^^^^^^ 1 − ^^^^^^^^^^^^+∆ ^^^^= ^^^^^^^^+ ^^�^^( ^^^^^^^^, ^^^^)∆ ^^^^.

[0307] In certain embodiments, linear push forward functions may be used in connection with machine learning models having been trained using loss functions such as, without limitation, a MSE for velocity fields and / or prospective final sequence representations, and a cross-entropy loss function for estimates of prospective final sequence representations.

[0308] In certain embodiments, a sequence representation xtis updated using a conditional path push forward function. As described, for example in Stark et al. (2024) “Dirichlet Flow Matching with Applications to DNA Sequence Design”, (which deals with simpler, nucleotide sequences), a conditional push forward function provides a parameterization of a velocity field as a weighted sum of conditional simplex velocities, where weights p(x1= ei| xt) are learned by a machine learning model (e.g., neural network).

[0309] In certain embodiments, where a machine learning model is used to predict p(x1| xt), the following equations may be used for a conditional push forward approach: ^^�^^(^^^^ , ^^^^)=∑20^^�^^(^^^^ = ^^^^|^^^^)^^^^(^^^^ = ^^^^|^^^^), where ^^^^(^^^^ = ^^^^| ) ∑20 ^^^^− ^^^^ ^^^^( ^^^^)^^^^ ^^^^=1 1 ^^^^ ^^^^ 1 ^^^^ ^^^^ 1 ^^^^ ^^^^ ^^^^ = ^^^^=1 1− ^^^^ ,where the conditional probability of xtleading to the category variable i and eiis a one hot vector the for ithcategory. p(x1= ei| xt) is the probability of the xtbelonging to the class i as given by the neural network prediction, which can be obtained e.g., using raw velocity prediction and apply operation to get an estimate of p(xt| x1= ei).

[0310] In certain embodiments, machine learning model generated predictions of v(x, t): ^^�^^(^^^^1|^^^^ ^^^^)= ^^^^ ^^^^ +(1 − t)^^�^^(^^^^, ^^^^),^�^^^( ^^^^^^^^, ^^^^) = ∑2^^^^=01 ^^�^^( ^^^^1= ^^^^^^^^^^^^^^^^) ^^^^( ^^^^1= ^^^^^^^^^^^^^^^^), where ^^^^( ^^^^1= ^^^^^^^^^^^^^^^^) = ∑2^^^^=0 ^^^^ ^^^^− ^^^^ ^^^^( ^^^^)1 ,- - 11901435v1Attorney Docket No.2013969-0048 ^^^^^^^^+∆ ^^^^= ^^^^^^^^+ ^�^^^( ^^^^^^^^, ^^^^)∆ ^^^^.

[0311] Accordingly, in the above, the probabilities may be determined based on outputs of the machine learning model, for example as values associated with each amino acid time based on a simplex representation. These probabilities can be viewed as magnitudes of vectors associated with each amino acid type, while the quantity c(…) provides a vector direction.

[0312] In certain embodiments, applying a push forward function to update sequence representations may result in points outside the simplex. Accordingly, in certain embodiments, for example, to guard against such possibilities, updating a sequence representation comprises projecting an initially updated point back to, for example, a nearest point on the simplex. In this manner, an ODE solver always obtains the input on the simplex.

[0313] In certain embodiments, probability distributions may be adjusted to, for example, to be increasingly peaky and / or smoother, for example using temperature annealing. For example, where a machine learning model generates an output htcorresponding to an invariant representation within a simplex, a probability p(x1= ei| xt) may be determined using the below equation: ^^�^^exp (ℎ ^^^^^^^^(^^^^1 = ^^^^ ^^^^|^^^^ ^^^^)=^^^^ )^^^^).

[0314] In certainrepresentation is obtained, for example, using a linear push forward and / or conditional push forward function, as described herein, and adjusted based on its proximity to a nearest corner of a simplex. For example, sequence feature values may be updated according to a sequence velocity field generated by a machine learning model and each amino acid feature connected with a nearest corner on a simplex (e.g., a corner corresponding to its maximum value) and adjusted towards or away from the nearest corner.

[0315] In certain embodiments, a norm of a velocity determined by a machine learning model may be adjusted by a scaling factor. Since we always project the final point after the push forward function, the resultant point will still represent a probability distribution over all categorical variables. - 72 - 11901435v1Attorney Docket No.2013969-0048

[0316] Accordingly, as described herein, the present disclosure provides methods and systems that allow for flow matching-based generation of protein and / or peptide sequences. As described in further detail herein sequence generation may be performed unconditionally, to generate de-novo amino acid sequences for arbitrary new protein or peptide structure, or may be conditioned on various desired protein properties, such as particular backbone geometries, partial sequence information, and various categorical, global, protein properties.

[0317] Additionally or alternatively, sequence generation may be performed via a dedicated machine learning model (e.g., that does not and / or has not been trained to perform other tasks), or via a multi-task machine learning model, that has been trained not only to generate outputs that can be used to determine sequence velocity fields as described above, but also to generate outputs that are associated with and can be used to generate other velocity fields, such as those used for peptide backbone generation, side chain packing, and / or other tasks.

[0318] A detailed experimental example demonstrating flow matching-based sequence generation in various contexts described herein is provided herein in Example 10. D. Side Chain Geometries

[0319] In certain embodiments, flow-matching technologies of the present disclosure include systems and methods for determining side chain packing in a protein structure – that is, generating viable three-dimensional arrangements of side chains in the context of particular protein peptide backbone and / or amino acid sequences. In certain embodiments, flow-matching technologies for side chain generation mirror those described herein for peptide backbone generation and / or sequence design is associated. As described herein, overall flow matching techniques aim effectively learn a joint distribution of protein backbone, sequence, and side chain geometries. In certain embodiments, flow-matching techniques for side chain generation leverage machine learning models to learn a conditional distributions of protein backbone, sequence, and side chains (e.g., instead of the joint distribution).

[0320] As described herein, flow-matching-based machine learning techniques for determining side chain geometries utilize, and evolve feature values that represent 3D side chain geometries along, probability paths interpolating between a prior and a final, target distribution. - 73 - 11901435v1Attorney Docket No.2013969-0048 A machine learning model may be trained and used (e.g., at inference) to determine velocity fields that are used to repeatedly update feature values, and thereby push them along these probability paths, for example by utilizing various solutions and / or numerical approaches for evaluating ODEs. As described herein, this machine learning models such as equivariant graph neural networks may be used in this context. D.i Side Chain Geometry Encoding

[0321] A protein structure comprises of a sequence of amino acids (residues) folded into a 3D geometry. This 3D geometry includes a peptide backbone geometry as well as 3D orientations of the individual amino acid side chains. That is, each residue has four backbone atoms N – Cα – C – O whose 3D geometry may be represented via a local frame representation, such as a two sets of values, (Cα, q), where Cα is the (centered) 3D coordinate of Cα and q ∈ SU(2) is a unit quaternion representing the frame (orthogonal matrix) of the residue under the map SU(2) → SO(3).

[0322] Protein side chains are attached to the Cα atoms and their three-dimensional geometries described by torsion angles, which represent rotations around the bonds within the side chain. These angles, often denoted as χ angles, define the orientation of each part of the side chain relative to the protein backbone and to each other. Each side chain, depending on its size and complexity, can have up to four torsion angles, typically denoted as χ angles (χ1, χ2, χ3, χ4). The χ1 angle describes the rotation around the bond between the α-carbon (Cα) and a first carbon of the side chain (usually β-carbon, Cβ). Subsequent χ angles (χ2, χ3, etc.) describe rotations around bonds further along the side chain. The specific number and values of these torsion angles determine the three-dimensional conformation of the side chain. By altering these angles, the side chain can adopt various spatial arrangements, affecting the protein’s overall structure and interaction capabilities. Precisely describing the side chains’ conformations using these torsion angles facilitates understanding protein function and for computational modeling of protein structures.

[0323] In certain embodiments, torsion angles may be represented in radians between 0 and 2π. However, in certain cases, this (0 to 2π) representation may be undesirable because values close to 0 and values close to 2π may be treated very differently by a machine learning - 74 - 11901435v1Attorney Docket No.2013969-0048 model, but in reality, should be considered close to each other. Among other things, this representation may produce unwanted discontinuities and resulting in unstable training and poor performance. Accordingly, additionally or alternatively, torsion angles may be represented as a pair of (cos(χ) , sin(χ)), which represents an angle χ as a point on the unit circle: S1 = {z ∈ C | ∥z∥ = 1}.

[0324] This is the coordinate representation of the complex number exp(iχ) using Euler’s formula, where i is the imaginary unit. In this manner, a 3D geometry of a side chain may be represented as a tuple of four complex numbers, one for each torsion angle. Since not all side chains require all four torsion angles (e.g., smaller side chains may be described using fewer angles), in certain embodiments, where a side chain less than the maximum, four, torsion angles remaining values may be set to (complex) 0. For example, the 3D geometry of Proline can be described via two torsion angles (χ1, χ2), which, in the four-value representation described above, may be represented as (exp(iχ1) , exp(iχ2), 0, 0). Among other things, this approach may be advantageous as it places non-existing and / or missing angles in a center of the complex plane, equidistant from all data points.

[0325] In certain embodiments, a flow matching technique for side chain generation may be setup comprising Riemannian flow on S1. Initial, seed values for side chain geometry features may be uniformly at random on a unit circle, and updated according to velocity fields to travel along interpolating probability paths following geodesics, i.e. connecting an initial prior distribution and a data distribution along arcs. This approach, however, may face challenges and / or suffer from volatility, for example, in situations where a prior and data are approximately opposite to each other on a unit circle, since the corresponding angles differ by roughly π. Accordingly, in certain embodiments, flow matching technologies for side chain generation of the present disclosure leverage this insight, utilize Euclidean geometries on the complex plane. This approach is facilitated via the 4-element complex tuple representation, and may lead to better results than Riemannian flow, allowing for stabilized training and reliable inference. - 75 - 11901435v1Attorney Docket No.2013969-0048 D.ii Training and Inference (a) Training

[0326] Accordingly, in certain embodiments, side chain geometry feature values are complex numbers within a unit disk centered at the origin. D = {z ∈ ℂ | ∥z∥ ≤ 1}.

[0327] Initial starting values, here complex numbers, z0, may be selected from an initial starting distribution, p0, which may, for example be a uniform distribution, p0= U (D), for each residue and each torsion angle on the disk. Training examples may be selected from a training dataset to obtain examples, z1(i), and used, together with a particular form (e.g., interpolation) of probability path to determine an intermediate feature set at a selected time, t, and corresponding velocity field. For example, in certain embodiments, straight paths between z0∼ p0and data z1∼ pdatamay be used, such that: zt= (1 − t)z0+ tz1, t ∈ [0, 1].

[0328] In certain embodiments, a constant velocity, z1-z0, such that velocities determined according to: vt(z) = ^^^^ [z1– z0| zt= z].

[0329] In certain embodiments, each torsion angle may be represented as a tuple, z. In certain embodiments, an initial seed, z0∼ p0,is sampled independently for each torsion angle and each residue of a protein.

[0330] In certain embodiments, as described herein, a neural network is trained to approximate the velocity field vt(z), that is, to determine, for a particular current time, t, and set of side chain geometry feature values, zt, an estimated velocity field, ^^�^^^^^^(z). During training, this estimated velocity field is compared with the velocity determined via knowledge of a desired final target – i.e., a particular training example, together with a selected / desired path and velocity function. For example, where linear paths and constant velocities are used, as described above, the below loss function is used: ℒ^^^^^^^^ ^^^^( ^^^^) = ^^^^ [�| ^^�^^^^^^− ( ^^^^1− ^^^^0)|�2]. 11901435v1Attorney Docket No.2013969-0048

[0331] In certain embodiments, training may also utilize dynamic weights, to account for the fact that the empirical distribution of the 20 amino acids (AAs) in proteins is not uniform, i.e. different amino acids occur with varying frequencies. In certain embodiments, non-uniform distribution of AAs in a protein affects training a model on protein data by directing the model to learn the properties of the frequently occurring AAs more accurately, while downplaying the intricacies of less frequently occurring AAs. Accordingly, to address this (e.g., bias), in certain embodiments, training approaches incorporate dynamic weights that scale loss based on amino acid type, in relation to their (e.g., empirical) frequency. For example, in certain embodiments, dynamical weighting comprises computing, for each batch b of proteins an empirical distribution β(b) and multiply the flow matching loss by the value of the inverse 1 / β(b) for each residue in the condition mask, according to its amino acid type.

[0332] In certain embodiments, additionally or alternatively, auxiliary loss terms may be used, for example to reflect, and penalize deviations from, known properties and / or requirements of physical protein structure. For example, an auxiliary loss may be used to enforce avoidance of steric clashes, since physical protein structures exhibit little to no steric clashes. In certain embodiments, an auxiliary loss function associated with steric clashes comprises an auxiliary loss γ(t)ℒ^^^^ ^^^^ ^^^^, which is a time dependent penalty for steric clashes. In certain embodiments, the is associated only with sufficiently large time values. In certain embodiments, the auxiliary loss is associated with t > 0.75. Such late cutoff is associated with the understanding / belief that a model should not be penalized at early stages of the probability paths when the model is still exploring various directions. (b) Inference

[0333] The model uses learned vector field ^^�^^^^^^(^^^^, ^^^^)to generate torsion angles following the ordinary differential equation (ODE) dynamics ^^^^ ^^̂^^ = ^^�^^ ^^^^(^^^^, ^^^^)^^^^ ^^^^.

[0334] The vector field determines the trajectory of the complex number ^^̂^^ on its way from an initial seed point ^^̂^^0to a final set of feature values ^^̂^^1as time t evolves from 0 to 1. At inference time, an initial starting distribution may be chosen freely to accommodate for - 77 - 11901435v1Attorney Docket No.2013969-0048 preferences regarding accuracy, novelty, or diversity of the generations. However, perturbing the starting distribution at inference time comes with the caveat of introducing a bias in the conditional distributions generated by following the learned vector field.

[0335] At each time step, a prediction of the torsion angle ^^̂^^ as the phase of the complex number is obtained, which amounts to projecting ^^̂^^ onto the unit circle S1and computing its angle, i.e. ^^̂^^|| ^^̂^^||= exp ( ^^^^ ^^^^).

[0336] For each ^^̂^^ along thetorsion angles ^^̂^^ was then used to generate the coordinates of all atoms in the protein structure. The amino acid type of each residue determines the number and type of the atoms in the side chain. Finally, the full 3D protein structure may be obtained by rotating the groups of atoms around their bonds, one group at a time, using the predicted torsion angles.

[0337] In certain embodiments, side chain geometries depend on a particular type of amino acid, and accordingly, approaches for predicting side chain geometries may first receive, as input, an amino acid sequence, e.g., thereby identifying a particular amino acid type at each location, whose detailed three-dimensional geometry may then be determined via the flow- matching technique described above. Amino acid sequences used in this manner may be determined in a variety of fashions, including, without limitation, via the sequence generation approaches described herein. For example, in certain embodiments, a single model may be trained to perform both sequence generation and side chain geometry determination, and used twice, e.g., in succession, once to generate an amino acid sequence and then to determine geometries for each side chain thereof. Additionally or alternatively, in certain embodiments, side chain geometries and sequence may be determined jointly, at the same time, by a single flow-matching model trained to perform both sequence generation and side chain geometry prediction, as described herein.

[0338] In certain embodiments, additionally or alternatively, the procedure described herein with regard to generation of χ angles, can be applied to the generation of bond angles. By adapting this procedure, capture the nuanced dynamics of bond angles, instead of building atoms positions based on the ideal angel assumptions. This extension not only demonstrates the - 78 - 11901435v1Attorney Docket No.2013969-0048 versatility of the original method but could enable more accurate modeling of molecular geometries. E. Machine Learning Model Architectures

[0339] A variety of machine learning architectures may be utilized in connection with peptide backbone generation technologies described herein. Particular architectures may include, without limitation, transformers, graph neural networks (GNNs), auto-regressive networks, encoders, decoders, and combinations thereof.

[0340] FIG.7A shows an example process 700 for (e.g., memory efficient) generating attention-based deep-learning predictions from input graph representations. An initial graph representation may be received and / or accessed 701, for example retrieved from memory, either locally or on a PACS server, cloud, etc. A predicted graph representation may be determined 702 using a machine learning model. This determination may involve the following steps. The machine learning model may receive 702i the initial graph representation as input. The machine learning model may comprise 702ii edge retrieval layers. The edge retrieval layers may, in turn, comprise self-attention heads to determine 702a attention weights. The attention weights may be used to determine 702b values of retrieved edge feature vectors. The machine learning model may generate 702iii as output predicted node feature vectors and / or velocity fields based at least in part on the retrieved edge feature vectors. The predicted node feature vectors and / or velocity fields determine 703 the predicted graph representation. Once determined, the predicted graph representation may be stored, provided for display, and / or provided for further processing 704.

[0341] FIGs.7B and 7C illustrate an example transformer-encoder architecture, used in certain embodiments. As shown in FIG.7B, block 710 receives, as input, a feature set comprising relative positions and orientations of each backbone site and a (e.g., current) time value. Each backbone site is represented via a translation and rotation of its local frame relative to that of a preceding backbone site. As described herein, translations and rotations may be represented via sets of three and four values, respectively, such that input xtto block 710 can be viewed as a list of seven element vectors, each corresponding to and representing a position and orientation of a particular backbone site. - 79 - 11901435v1Attorney Docket No.2013969-0048

[0342] In certain embodiments, input xtis encoded via a Multi-Layer Perceptron (MLP), e.g., before being fed to transformer block 710, as illustrated in the right-hand side of FIG.7B. In certain embodiments, velocity is generated as output from a transformer block 710, following a final MLP. In certain embodiments, a positional embedding is used to encode a position (e.g., location in a sequence) of each backbone site, for example using an approach as described in A. Vaswani et al., “Attention is all you need,” NIPS 2017 the content of which is hereby incorporated by reference in its entirety. In certain embodiments, a bi-directional positional encoding is used, such that for a particular backbone site i, of a peptide of length n, position i and position n-i are both encoded, so as to provide transformer block 710 information about how far a particular backbone site is from a start and end of a sequence. In certain embodiments, bidirectional positional encoding is used for edges. In certain embodiments, relative positional encoding is used for edges.

[0343] In certain embodiments, a current time value is received as input, e.g., as a feature. Additionally or alternatively, in certain embodiments, a current time value is not used as a feature and (e.g., instead) is inputted, separately, and used to condition transformer intermediate representations. For example, as shown in FIG.7B, a time value is received as input, encoded by a multi-layer perceptron (MLP) and the resultant encoding used to modulate intermediate features that are learned by a transformer.

[0344] In certain embodiments, for example as shown in FIG.7C, multiple (e.g., a plurality of) blocks 710 are stacked in sequence, with output from one block feeding into a next.

[0345] While FIGs.7B and 7C show a particular machine learning structure, as described herein, other machine learning structures may be used in various embodiments. F. Multi-Input and Multi-Modal Peptide / Protein Design

[0346] In certain embodiments, the present disclosure includes a multi-model polypeptide generation technology whereby scaffold model generation techniques described herein are adapted and extended to allow for generation of (i) scaffold models representing peptide backbones and / or (ii) sequence data, representing amino acid sequences. Moreover, in - 80 - 11901435v1Attorney Docket No.2013969-0048 certain embodiments, scaffold and / or polypeptide generation approaches described herein can receive a variety of inputs on which to condition scaffold model and / or sequence data generation.

[0347] Generative scaffold design technologies described herein can, for example be used to create scaffold models based on a variety of initial conditions, input information, and for a variety of purposes. F.i Unconditional generation

[0348] Scaffold models can, for example, be generated without an initial conditioning, for example to generate new scaffold models representing new peptide backbones of various sizes (e.g., a number of amino acid sites may be fixed). In certain embodiments, sequence data can also be generated, for example together with a scaffold model. A scaffold model and its sequence can, for example, be generated without conditioning on initial features or properties, so as to allow for creation of a de-novo protein and / or peptide including, not only a viable 3D structure, but also a corresponding sequence (e.g., predicted to fold into the 3D structure). F.ii Conditioning on Backbone Structure and / or Sequence

[0349] In certain embodiments, scaffold generation can be conditioned on various initial scaffold structures and / or sequences. For example, in certain embodiments, scaffold models can be generated based on – that is, conditioned on – various initial partial scaffolds.

[0350] For example, scaffold generation can be conditioned on an input target scaffold model that represents a known and desired target, such that generated scaffold models represent custom candidate peptide backbones that are favorable for binding to the target.

[0351] Additionally, or alternatively, scaffold generation can be conditioned on a pre- existing partial scaffold that represents a partially known and / or desired backbone structure, such that generated scaffolds fill in (e.g., inpaint) unknown and / or variable regions.

[0352] In certain embodiments, scaffold generation techniques can be used to dock two known protein scaffolds.

[0353] In certain embodiments, scaffolds can be conditioned on oligomeric state. - 81 - 11901435v1Attorney Docket No.2013969-0048

[0354] In certain embodiments, sequence data may be received as input and / or generated as output. For example, in certain embodiments, generative technologies described herein can be used to generate a sequence, conditioned on a particular scaffold model. In certain embodiments, scaffold models can be generated conditioned on an amino acid sequence (folding problem). In certain embodiments, scaffold models can be repeatedly generated, conditioned on a particular sequence. For each new scaffold model, a different initial condition is sampled from a prior / initial distribution and, accordingly, this statistical approach can be used to generate a plurality of scaffold models for a particular sequence, thereby sampling the conformational landscape of potential folded backbones associated with a particular sequence. These possible conformations can be used to assess energy, stability, and other properties of a particular molecule, for a variety of scientific and practical applications.

[0355] In certain embodiments, generative technologies of the present disclosure can be used to dock two known proteins having a particular known monomeric structure and / or sequence. F.iii Conditioning on Global Properties

[0356] In certain embodiments, categorical variables can also be received as input, and used for conditioning. These may include, for example, protein family, thermophilly, immunology, function, solubility, pH sensitivity, etc. F.iv Conditioning on Node Properties

[0357] Additional node properties can also be used for conditioning. These may include, for example, amino acid type (e.g., via a one-hot encoding), a polarity categorical variable (e.g., having 5 possible polarities – polar, apolar, negatively charged, positively charged), burriedness (e.g., a binary variable indicating whether a node is at a surface or core), as well as secondary structure (e.g., a variable encoding whether a node is part of a helix or beta-sheet).

[0358] In certain embodiment, a binary variable can be used to identify and / or pre-define those nodes that represent hotspots – i.e., amino-acid sites that lie within a particular threshold distance of another molecule, such as a target and / or other chain. Hotspot identification - 82 - 11901435v1Attorney Docket No.2013969-0048 approaches are described in further detail, e.g., in U.S. Patent No.11,450,407, issued September 20, 2022, the content of which is hereby incorporated by reference in its entirety.

[0359] For example, in certain embodiments, sites referred to as “hotspots” may be identified on a ligand and / or receptor. For a ligand, hotspots refer to sites which, when occupied by an amino acid side chain, place at least a portion of the amino acid side chain in proximity to one or more side chains and / or atoms of the receptor. Likewise, for a receptor, hotspots are sites which, when occupied via an amino acid side chain, place at least a portion of the amino acid side chain in proximity to one or more side chains and / or atoms of the ligand.

[0360] In certain embodiments, for example since size, geometry, and orientation of various acid side chains may vary, hotspots may be identified based on distances between beta carbon (Cβ) atoms of a ligand and receptor of a complex. For example, a ligand hotspot may be identified as a particular site on the ligand that, when occupied by an amino acid side chain, will place a Cβ atom of the side chain located at the site within a threshold distance of a Cβ atom of the receptor. Receptor hotspots may be identified analogously. Since Cβ atoms are common to every amino acid side chain apart from Glycine, this approach provides a uniform criterion for identifying hotspots, independent of a particular amino acid that occupies a particular site. In certain embodiments, in the singular case where a Glycine residue occupies a particular site, Glycine’s hydrogen atom may be used in place of a Cβ, but hotspots identified in an otherwise identical fashion. Additionally or alternatively, in certain embodiments, distances between alpha- carbons (Cα) associated with amino-acid sites may be determined, e.g., in a similar manner to which distances between Cβ atoms are determined. In this manner, Cα distances may be compared with various threshold values to identify hotspots.

[0361] Various threshold distances may be used for identification of hotspots. For example, in certain embodiments, a hotspot threshold distance of 8Å (i.e., 8 Ångstroms) is used. In some embodiments, other thresholds may be used for defining a hotspot (such as less than 3Å, less than 4Å, less than 5Å, less than 6Å, less than 7Å, less than 9Å, less than 10Å, less than 12Å, less than 15Å, less than 20Å, as well as other suitable thresholds).

[0362] In certain embodiments, hotspots may be identified based on comparison of values computed by various functions – e.g., of one or both of a Cα and Cβ distance – with one - 83 - 11901435v1Attorney Docket No.2013969-0048 or more threshold values. Such functions may take into account features such as bond angles, surface area, etc. F.v Conditioning on Protein Folds

[0363] Protein topological properties can also be used for conditioning. These may include properties of protein secondary structure, folding properties. In certain embodiments, protein topological properties may be represented by a block adjacency matrix.

[0364] A “fold family” refers to a group of proteins that share a similar three- dimensional structure or fold, despite potentially having different amino acid sequences. Proteins with similar folds often exhibit similar overall shapes and structural features, even if their primary sequences (the order of amino acids) differ. While the specific sequence of amino acids determines protein's unique structure and function, certain structural motifs, or folds, repeat themselves throughout different proteins. These folds can be highly conserved across evolution, and proteins with the same or similar folds often perform related functions.

[0365] Understanding fold families is crucial in protein design and engineering because it allows scientists to predict the structure and function of a newly designed protein based on the known characteristics of proteins with similar folds. This knowledge aids in designing proteins with desired functions, such as enzymes for specific biochemical reactions or proteins for therapeutic purposes. It also facilitates the exploration of structure-function relationships in proteins.

[0366] FIG.7D shows an example process 750 for designing a custom biologic having a three-dimensional (e.g., fold) structure belonging to a desired fold family according to various embodiments described herein. Protein fold representation may be received and / or accessed 751, for example retrieved from memory, either locally or on a PACS server, cloud, etc. The protein fold representation may encode and represent desired three-dimensional structural features of the desired fold family and present within the custom biologic. A sequence and / or a three- dimensional peptide backbone structure may be generated 752, using a machine learning model, based on the protein fold representation. The generated sequence and / or 3D peptide backbone structure may be stored, provided for display, and / or provided for further processing 753. - 84 - 11901435v1Attorney Docket No.2013969-0048

[0367] The block adjacency matrix (BAM) represents the adjacency relationships between different secondary structure segments of a protein based on a specified distance cutoff. In certain embodiments, BAM is calculated based on protein's Secondary Structure Elements (SSE), backbone coordinates, and a cutoff distance. SSE may be represented by a (e.g., one- dimensional) tensor, where values of the tensor represent a numeric encoding of SSEs at each position in the protein sequence. The numeric encoding may be as follows: 0 for Helix, 1 for Strand, 2 for Loop, and 3 for Masked regions. The backbone coordinates may be represented by a tensor containing (e.g., Cartesian) coordinates of the backbone N, Ca, and C atoms for each residue in the protein. The cutoff value (e.g., 6.0 Ångströms) is used to determine whether two residues are considered adjacent. In certain embodiments, BAM calculations comprise inclusion criteria. In certain embodiments, the inclusion criteria comprise a flag to denote whether certain regions (e.g., loop regions) should be included as contacts in the block adjacency matrix.

[0368] A block adjacency matrix may be represented by a (e.g., 2D binary) matrix (e.g., of size (L, L), where L is the length of the protein sequence). For example, each element in the matrix is set to 1 if the corresponding SSEs are considered adjacent based on the distance cutoff and the inclusion criteria.. For example, FIG.8 shows a TIM barrel secondary structure and FIG.9A shows a block adjacency matrix calculated for a TIM barrel according to methods described herein. In certain embodiments, BAM calculations, additionally or alternatively, produce a segment identification and / or a block adjacency construction. The segment identification comprises values associated with identification of continuous segments of the same SSE type and values associated with start and end indices of these continuous segments. The block adjacency construction comprises values associated with a check if any pair of residues (e.g., one from each segment) falls within the specified distance cutoff (e.g., if a pair distance is below the cutoff distance, the corresponding block in the adjacency matrix is filled with ones, indicating adjacency).

[0369] In certain embodiments, additional information about the topology of secondary structure elements (SSEs) in protein structures may be incorporated into (e.g., concatenated with) BAM, providing a comprehensive representation of the protein’s structural and topological features. For example, the topological relationships between SSEs (e.g., as parallel, antiparallel, or vertical) may be classified and included into BAM. BAM may be represented as a (e.g, binary) matrix (e.g., of size (L, L, 4), where L is the length of the protein sequence). FIG.9B - 85 - 11901435v1Attorney Docket No.2013969-0048 shows a multi-channel block adjacency matrix calculated for a TIM barrel according to methods described herein. As a result, BAM comprises (e.g., binary; binned; one hot vector) values associated with topology classification (e.g., classified as parallel, antiparallel, and vertical).

[0370] The axis of each secondary structure segment is calculated using the coordinates of the first five (e.g., three, ten) residues in the SSE and the SSE type (e.g., alpha or beta). The orientation between two secondary structure segments is determined by the angle between their axes. The classification criteria for topology classification may be as follows: (1) Parallel - if the angle between the axes is less than 60 (e.g., 30, 45) degrees; (2) Antiparallel - if the angle is greater than 120 (e.g., 150, 135, respectively) degrees; (3) Vertical - if the angle is between 60 (e.g., 30, 45) and 120 (e.g., 150, 135, respectively) degrees. G. Custom Biologic Design Via Flow-Matching

[0371] In certain embodiments, generative technologies described herein may also include conditioning approaches, so as to produce generated scaffold models that represent peptide backbones having desired features such as particular properties, structural motifs, etc. In certain embodiments, conditioning may be used in connection with a representation of one or more prospective binding sites of a target molecule (e.g., protein and / or peptide), thereby allowing for creation of a de-novo peptide backbone suitably for binding to the target molecule (e.g., at a location within or comprising at least one of the one or more prospective binding sites).

[0372] In certain embodiments, generated scaffold models representing de-novo peptide backbones may be used as input to an interface designer module (e.g., comprising a machine learning model) for generating an amino acid interface for binding to a target molecule. Interface designer approaches, that, for example, populate scaffold models with amino acid sequences in order to create custom interfaces for binding to particular targets are described in further detail, for example, in International publication WO 2023 / 004116 A1 of PCT / US22 / 38014, entitled “SYSTEMS AND METHODS FOR ARTIFICIAL INTELLIGENCE-GUIDED BIOMOLECULE DESIGN AND ASSESSMENT,” and filed July 22, 2022, included herein as an Appendix and the content of which is hereby incorporated by reference in its entirety. - 86 - 11901435v1Attorney Docket No.2013969-0048 G.i Application examples for sequence generation

[0373] This section describes frameworks of using flow matching on protein sequences for various applications. (a) Unconditional sequence generation

[0374] In certain embodiments, generative technologies described herein may be used to produce a protein sequence from scratch (e.g., without specifying any explicit conditions or context to guide the generation process). For example, a model may be trained on sequence data only (e.g., no structural information), and use as input: • Nodes simplicial • Sequence position • Time • Any conditioning on node and / or global properties is masked. (b) Conditional sequence generation

[0375] In certain embodiments, generative technologies described herein may be used to generate a protein sequence from scratch, by specifying explicitly conditions or context to guide the generation process. For example, a model may use as input: • Nodes simplicial representation (xt), i.e. a point in the 19-dimensional simplex as explained in the present disclosure • Sequence position • Time • Any conditioning on node and / or global properties.

[0376] In certain embodiments, generative technologies described herein may be used to generate a protein sequence that is predicted to fold into a specific three-dimensional structure (e.g., backbone). For example, a model may use as input: • Nodes simplicial representation (xt), i.e. a point in the 19-dimensional simplex as explained in the present disclosure - 87 - 11901435v1Attorney Docket No.2013969-0048 • Sequence position • Time • Position of amino acids (e.g., CA position) • Backbone orientation (e.g., Canocical frames) • Relatative information (e.g., edges, 6DOF) • Any conditioning on node and / or global properties. (c) Co-design

[0377] In certain embodiments, generative technologies described herein may be used to generate both a protein backbone and a protein sequence. For example, a model may be trained on the sum of two loss objectives associated with backbone generation and sequence generation. The model may perform all tasks as related to protein generation and sequence generation as described in the present disclosure. The model may receive all conditioning associated with protein generation tasks and sequence generation tasks as described in the present disclosure. Additionally or alternatively, the model may replace one-hot encoding representation of amino acids as used, for example, for backbone generation with the corresponding values on the simplex as used, for example, for sequence generation.

[0378] In certain embodiments, generative technologies described herein may be used to generate both a protein side chain and a protein sequence conditioned on a protein backbone. For example, a model may be trained on the sum of two loss objectives associated with side chain generation and sequence generation. The model may perform all tasks as related to side chain generation and sequence generation as described in the present disclosure. The model may receive all conditioning associated with side chain generation tasks and sequence generation tasks as described in the present disclosure. Additionally or alternatively, the model may replace one-hot encoding representation of amino acids as used, for example, for side chain generation with the corresponding values on the simplex as used, for example, for sequence generation.

[0379] In certain embodiments, generative technologies described herein may be used to generate a protein backbone, a protein side chain, and a protein sequence unconditionally. For example, a model may be trained on the sum of three loss objectives associated with backbone - 88 - 11901435v1Attorney Docket No.2013969-0048 generation, side chain geometry prediction, and sequence generation. In certain embodiments, for side chain geometry generation, for example, multiple predictions of side chain geometries may be generated assuming all 20 amino acids for each location and then combined using the probability of each amino acid type, at each location throughout the protein. In certain embodiments, loss objectives for three tasks may remain separated and / or be trained using different weights. G.ii Application examples for fold conditioning

[0380] This section describes frameworks of using flow matching with fold conditioning for applications in designing de novo antibody / nanobody that binds antigens. (a) Inpainting from a known pose

[0381] The determinant of the antigen, referred to as epitope, is the part of the antigen that is typically recognized by an antibody. Epitopes are typically linear (i.e. a continuous sequence segment) or conformational (i.e. discontinuous fragments in sequence space but adjacent in 3D space).

[0382] Given an antibody / epitope complex, the inpainting task refers to designing a complementary region of the antibody that can specifically and tightly bind to the epitope. The task involves manipulating the complementarity-determining regions (CDRs) of the antibody, which are the parts of the antibody that directly interact with the epitope. Any part that is masked is generated by the model.

[0383] CDRs design. In certain embodiments, generative technologies described herein may be used to design de novo CDR regions of an antibody that is interacting with a given target antigen with high specificity and affinity. For example, an input representation at t=0 may take form: 1. Antibody: a. Available: everything except CDRs b. Masked: CDR coordinates and sequence (optional) 2. Antigen: - 89 - 11901435v1Attorney Docket No.2013969-0048 a. Available: everything b. Masked: nothing The coordinates and (optionally) the sequence of the CDRs are generated as an output.

[0384] Redesign of CDRs can potentially improve many physico-chemical and PK properties of the antibody. For example: • Increasing Affinity: Modifying CDRs to enhance the strength of the interaction between the antibody and the antigen • Improving Specificity: Ensuring the antibody binds only to the intended antigen and / or not to similar structures on other molecules, thereby reducing off-target effects. • Improving thermodynamic and thermal stability: Designing CDRs that maintain their structure and function under physiological conditions, improving the durability and shelf- life of antibodies. • Reducing Immunogenicity: Modifying or “humanizing” CDRs in therapeutic antibodies derived from non-human sources to prevent them from being recognized as foreign by the human immune system. • Identification of critical hotspots through mutations: o Mutations: simulate how CDR’s mutations affect binding regions o Insertions / deletion: extend / shorten the length of CDR to affect the binding regions

[0385] Epitope relaxation. In certain embodiments, generative technologies described herein may be used to design de novo epitope interfaces of an antigen interacting with a given antibody with high specificity and affinity. This application opens new avenues in terms of targeting strategies of specific therapeutic targets by expanding the universe of possible epitopes. For example, input representation at t=0 may take form: 1. Antibody: a. Available: everything b. Masked: nothing - 90 - 11901435v1Attorney Docket No.2013969-0048 2. Antigen: a. Available: everything excepts epitope coordinates b. Masked: epitope coordinates The coordinates of the epitope may be generated as an output.

[0386] Examples of epitope relaxation: • Increased Specificity / Immunogenicity: Modify the epitope to ensure that the immune response is highly specific to the target antigen, reducing the risk of cross-reactivity with non-target antigens. • Improved Stability: Enhance the physical and chemical stability of the epitope, ensuring that it maintains its structure and immunogenicity under a wide range of conditions. • Mutations: Simulate how epitopes' mutation affect the binding regions.

[0387] Epitope relaxation + CDRs design. In certain embodiments, generative technologies described herein may be applied in various ways to achieve both objectives of epitope relaxation and CDRs design as described herein. For example, iteratively repeating for any arbitrary number of times: (1) Epitope relaxation objective then CDRs design objective or (2) CDRs design objective then Epitope relaxation objective. In certain embodiments, epitope relaxation as well as the CDRs design may be performed at the same time. For example, both objectives are achieved simultaneously. For example, input representation at t=0 may take form: 1. Antibody: a. Available: everything except CDRs b. Masked: coordinates and sequence 2. Antigen: a. Available: everything except epitope coordinates b. Masked: epitope coordinates (b) Antibody design from an unknown pose - 91 - 11901435v1Attorney Docket No.2013969-0048

[0388] Any information used from existing antibody (e.g., BAM, SSE) may be modified / fine-tuned to get full control on the conditioning. For example, the general fold of the antibody to design may be given (e.g., through BAM) or one may start with internal coordinates of the framework (i.e. non-CDRs regions) and simultaneously dock and design an antibody.

[0389] Moreover, the framework may be modified through insertion and / or deletion on different parts of the framework. It would then affect only the length of the framework. If insertion are used, it would just be considered as unknown for the model. If deletion are used, only the length will be affected but the information elsewhere will remain the same.

[0390] Any masked information sequence, position, and side chains for the antibody may be generated and (optionally) simultaneously position and side chains on the epitope can also be generated. Hence, multiple levels of given input to condition on (on top of the target itself) may be used.

[0391] FIG.9C shows an example process 900 for designing a custom antibody for binding to a target, according to various embodiments described herein. Antibody template representing a sequence and / or three-dimensional structure of a reference antibody may be received and / or accessed 901, for example retrieved from memory, either locally or on a PACS server, cloud, etc. The antibody template may comprise: (i) a base portion located about CDRs of the reference antibody, but (the base portion) excluding CDRs themselves, and (ii) CDRs portion, each CDR portion associated with and comprising a CDR of the reference antibody. Protein fold representation that encodes and represents three-dimensional structural features of the base portion of the antibody template may be determined 902. A sequence and / or a three- dimensional peptide backbone structure may be generated 903, using a machine learning model, based on the protein fold representation. The generated sequence and / or three-dimensional peptide backbone structure may include custom CDRs. The generated sequence and / or 3D peptide backbone structure may be stored, provided for display, and / or provided for further processing 904.

[0392] New antibody generation conditioned on fold. In certain embodiments, generative technologies described herein may be used to design de novo antibodies conditioned on specific fold properties to interact with a given antigen with high specificity and affinity. For example, by utilizing one or more existing antibodies, a BAM and / or SSE may be calculated to - 92 - 11901435v1Attorney Docket No.2013969-0048 then inform the conditioning of the generation process. In doing this, de novo antibodies may be created through loose conditioning, with the expectation of discovering new antibody folds. The CDRs are not conditioned to give them full freedom at generation time. For example, input representation at t=0 may take form: 1. Antibody: a. Available: partial BAM and SSE (e.g., masked CDRs) for heavy and / or light chains (e.g., structural elements of the antibody) b. Masked: coordinates and sequence 2. Antigen: a. Available: everything b. Masked: nothing The coordinates as well as the sequence of the antibody are generated as an output.

[0393] New antibody generation conditioned on fold with partial sequence. In certain embodiments, generative technologies described herein may be used to design de novo antibodies conditioned on specific fold properties with partial sequence to interact with a given antigen with high specificity and affinity. The inclusion of partial sequence may enforce even more the conditioning on parts of the antibody. For example, the sequence of an existing antibody from the light and / or heavy chains is partially kept. For example, input representation at t=0 may take form: 1. Antibody: a. Available: partial sequence, partial BAM and SSE (e.g. masked CDRs) for heavy and / or light chains b. Masked: coordinates and partial sequence 2. Antigen: a. Available: everything b. Masked: nothing The coordinates as well as the masked sequence of the antibody are generated as an output. - 93 - 11901435v1Attorney Docket No.2013969-0048

[0394] New antibody generation conditioned on scaffold. In certain embodiments, generative technologies described herein may be used to design de novo antibodies conditioned on specific scaffold properties to interact with a given antigen with high specificity and affinity. For example, by utilizing one or more existing antibodies, six degrees of freedom (6DOF) may be calculated, which may further inform the conditioning of the generation process. In doing this, de novo antibodies may be created through stronger conditioning, with the expectation of discovering new antibody folds. The CDRs are not conditioned to give them full freedom at generation time. For example, input representation at t=0 may take form: 1. Antibody: a. Available: partial 6 DOFs (e.g. masked CDRs) for heavy and / or light chains b. Masked: coordinates and sequence 2. Antigen: a. Available: everything b. Masked: nothing The coordinates as well as the sequence of the antibody are generated as an output.

[0395] New antibody generation conditioned on scaffold with partial sequence. In certain embodiments, generative technologies described herein may be used to design de novo antibodies conditioned on specific scaffold properties with partial sequence to interact with a given antigen with high specificity and affinity. For example, partially keeping the sequence of an existing antibody from the light and / or heavy chains may enforce the conditioning on parts of the antibody. For example, input representation at t=0 may take form: 1. Antibody: a. Available: partial sequence, partial 6 DOFs (e.g. masked CDRs) for heavy and / or light chains b. Masked: coordinates and partial sequence 2. Antigen: a. Available: everything - 94 - 11901435v1Attorney Docket No.2013969-0048 b. Masked: nothing The coordinates as well as the masked sequence of the antibody are generated as an output. (c) Docking

[0396] In certain embodiments, generative technologies described herein may be used to design a docking process (e.g., values associated with two polypeptides interacting with each other; e.g., optimal configuration of molecules that results in the highest binding affinity) for a given antibody and an antigen. For example, input representation at t=0 may take form: 1. Antibody: a. Available: BAM (optional), SSE (optional), Sequence, 6 DOF for heavy and / or light chains b. Masked: coordinates (optional) 2. Antigen: a. Available: BAM (optional), SSE (optional), Sequence, 6 DOF for heavy and / or light chains b. Masked: coordinates (optional) Generated output may comprise either: (i) the antibody coordinates (e.g., initially masked) and make the antigen coordinates available in input; (ii) the antigen coordinates (e.g., initially masked) and make the antibody coordinates available in input; or (iii) the antibody and antigen coordinates (e.g., initially masked). (d) Relaxation

[0397] In certain embodiments, generative technologies described herein for applications described herein may comprise (e.g., simultaneous) relaxation of an epitope of an antigen. For example, relaxation of the epitope may be performed by adapting the antigen representation as follows. For example, input representation at t=0 may take form: Antigen: 1. Available: everything except on epitope 2. Masked on antigen: coordinates of the epitope - 95 - 11901435v1Attorney Docket No.2013969-0048

[0398] Additionally or alternatively, relaxation may be performed first on an antibody as described in, for example, (Antibody design from an unknown pose) or (Docking) and then apply relaxation performed on an antigen as described in, for example (Epitope relaxation). G.iii Synthesis of Custom Biologics

[0399] In certain embodiments, custom proteins and / or peptides are synthesized according to in-silico designed – e.g., sequences – determined herein. For example, in certain embodiments, once an amino acid sequence of a particular custom biologic is designed, e.g., as described herein, polynucleotides may be designed that encode the desired amino acid sequence. Such custom sequences may be or comprise sequences of antibodies or antigen-binding fragments thereof, e.g., as provided herein. Accordingly, the present disclosure includes nucleic acids encoding one or more heavy chains, VH domains, heavy chain FRs, heavy chain CDRs, heavy chain constant domains, light chains, VL domains, light chain FRs, light chain CDRs, light chain constant domains, or other immunoglobulin-like sequences, antibodies, or antigen- binding fragments thereof disclosed herein.

[0400] In some embodiments, the present disclosure provides polynucleotides which encode at least one CDR region, such as at least two, such as at least three CDR regions from the heavy or light chain of an antibody provided herein. In some embodiments, polynucleotides encode all or substantially all of the variable region sequence of the heavy chain and / or the light chain of an antibody.

[0401] Nucleic acid sequences can be produced by de novo solid-phase DNA synthesis or by PCR mutagenesis of an existing sequence (e.g., sequences as provided in the Examples below) encoding an antibody or its binding fragment. Direct chemical synthesis of nucleic acids can be accomplished by methods known in the art, such as the phosphotriester method of Narang et al., 1979, Meth. Enzymol.68:90; the phosphodiester method of Brown et al., Meth. Enzymol. 68:109, 1979; the diethylphosphoramidite method of Beaucage et al., Tetra. Lett., 22:1859, 1981; and the solid support method of U.S. Patent No.4,458,066. Introducing mutations to a nucleic acid sequence by PCR can be performed as described in, e.g., PCR Technology: Principles and Applications for DNA Amplification, H.A. Erlich (Ed.), Freeman Press, NY, NY, 1992; PCR Protocols: A Guide to Methods and Applications, Innis et al. (Ed.), Academic Press, San Diego, - 96 - 11901435v1Attorney Docket No.2013969-0048 CA, 1990; Mattila et al., Nucleic Acids Res.19:967, 1991; and Eckert et al., PCR Methods and Applications 1:17, 1991.

[0402] Various expression vectors can be employed to express polynucleotides encoding custom biologics, such as antibodies, or antigen-binding fragments thereof. Both viral-based and nonviral expression vectors can be used to produce antibodies or antigen-binding fragments thereof in a mammalian host cell. Nonviral vectors and systems include plasmids, episomal vectors, typically with an expression cassette for expressing a protein or RNA, and human artificial chromosomes (see, e.g., Harrington et al., Nat Genet 15:345, 1997). Useful viral vectors include vectors based on retroviruses, adenoviruses, adenoassociated viruses, herpes viruses, vectors based on SV40, papilloma virus, HBP Epstein Barr virus, vaccinia virus vectors and Semliki Forest virus (SFV). See, Brent et al., supra; Smith, Annu. Rev. Microbiol.49:807, 1995; and Rosenfeld et al., Cell 68:143, 1992.

[0403] The choice of expression vector depends on the intended host cells in which the vector is to be expressed. Typically, the expression vectors contain a promoter and other regulatory sequences (e.g., enhancers) that are operably linked to the polynucleotides encoding an antibody or antigen-binding fragment thereof. In some embodiments, an inducible promoter is employed to prevent expression of inserted sequences except under inducing conditions. Inducible promoters include, e.g., arabinose, lacZ, metallothionein promoter or a heat shock promoter. Cultures of transformed organisms can be expanded under noninducing conditions without biasing the population for coding sequences whose expression products are better tolerated by the host cells. In addition to promoters, other regulatory elements may also be required or desired for efficient expression of custom biologics, such as an antibody or antigen- binding fragment thereof.

[0404] Host cells for harboring and expressing various custom biologics generated in accordance with systems and methods of the present disclosure may be either prokaryotic or eukaryotic. E. coli is one prokaryotic host useful for cloning and expressing polynucleotides. Other microbial hosts suitable for use include bacilli, such as Bacillus subtilis, and other enterobacteriaceae, such as Salmonella, Serratia, and various Pseudomonas species. In these prokaryotic hosts, one can also make expression vectors, which typically contain expression control sequences compatible with the host cell (e.g., an origin of replication). In addition, any - 97 - 11901435v1Attorney Docket No.2013969-0048 number of a variety of well-known promoters will be present, such as the lactose promoter system, a tryptophan (trp) promoter system, a beta-lactamase promoter system, or a promoter system from phage lambda. The promoters typically control expression, optionally with an operator sequence, and have ribosome binding site sequences and the like, for initiating and completing transcription and translation.

[0405] In some embodiments, mammalian host cells may be used. These include any normal mortal or normal or abnormal immortal animal or human cell. For example, a number of suitable host cell lines capable of secreting intact immunoglobulins have been developed including the CHO cell lines, various Cos cell lines, HeLa cells, myeloma cell lines, transformed B-cells and hybridomas. The use of mammalian tissue cell culture to express polypeptides is discussed generally in, e.g., Winnacker, FROM GENES TO CLONES, VCH Publishers, N.Y., N.Y., 1987. Expression vectors for mammalian host cells can include expression control sequences, such as an origin of replication, a promoter, and an enhancer (see, e.g., Queen, et al., Immunol. Rev.89:49-68, 1986), and necessary processing information sites, such as ribosome binding sites, RNA splice sites, polyadenylation sites, and transcriptional terminator sequences. These expression vectors usually contain promoters derived from mammalian genes or from mammalian viruses. Suitable promoters may be constitutive, cell type-specific, stage-specific, and / or modulatable or regulatable.

[0406] Methods for introducing expression vectors containing the nucleic acid sequences of interest vary depending on the type of cellular host. For example, calcium chloride transfection is commonly utilized for prokaryotic cells, whereas calcium phosphate treatment or electroporation may be used for other cellular hosts. Other methods include, e.g., electroporation, calcium phosphate treatment, liposome-mediated transformation, injection and microinjection, ballistic methods, virosomes, immunoliposomes, polycation:nucleic acid conjugates, naked DNA, artificial virions, fusion to the herpes virus structural protein VP22 (Elliot and O'Hare, Cell 88:223, 1997), agent-enhanced uptake of DNA, and ex vivo transduction. For long-term, high-yield production of recombinant proteins, stable expression will often be desired. Following the introduction of the vector, cells may be allowed to grow for 1-2 days in an enriched media before they are switched to selective media. The purpose of the selectable marker is to confer resistance to selection, and its presence allows growth of cells - 98 - 11901435v1Attorney Docket No.2013969-0048 which successfully express the introduced sequences in selective media. Resistant, stably transfected cells can be proliferated using tissue culture techniques appropriate to the cell type.

[0407] In certain embodiments, custom biologics of the present disclosure are or comprise monoclonal antibodies (mAbs), which can be produced by a variety of techniques, including conventional monoclonal antibody methodology e.g., the standard somatic cell hybridization technique of Kohler and Milstein, 1975 Nature 256: 495. Many techniques for producing mAbs can be employed e.g., viral or oncogenic transformation of B lymphocytes. H. System, and Network Environment

[0408] Turning to FIG.10, an implementation of a network environment 1000 for use in providing systems, methods, and architectures as described herein is shown and described. In brief overview, referring now to FIG.10, a block diagram of an exemplary cloud computing environment 1000 is shown and described. The cloud computing environment 1000 may include one or more resource providers 1002a, 1002b, 1002c (collectively, 1002). Each resource provider 1002 may include computing resources. In some implementations, computing resources may include any hardware and / or software used to process data. For example, computing resources may include hardware and / or software capable of executing algorithms, computer programs, and / or computer applications. In some implementations, exemplary computing resources may include application servers and / or databases with storage and retrieval capabilities. Each resource provider 1002 may be connected to any other resource provider 1002 in the cloud computing environment 1000. In some implementations, the resource providers 1002 may be connected over a computer network 1008. Each resource provider 1002 may be connected to one or more computing device 1004a, 1004b, 1004c (collectively, 1004), over the computer network 1008.

[0409] The cloud computing environment 1000 may include a resource manager 1006. The resource manager 1006 may be connected to the resource providers 1002 and the computing devices 1004 over the computer network 1008. In some implementations, the resource manager 1006 may facilitate the provision of computing resources by one or more resource providers 1002 to one or more computing devices 1004. The resource manager 1006 may receive a request for a computing resource from a particular computing device 1004. The resource manager 1006 - 99 - 11901435v1Attorney Docket No.2013969-0048 may identify one or more resource providers 1002 capable of providing the computing resource requested by the computing device 1004. The resource manager 1006 may select a resource provider 1002 to provide the computing resource. The resource manager 1006 may facilitate a connection between the resource provider 1002 and a particular computing device 1004. In some implementations, the resource manager 1006 may establish a connection between a particular resource provider 1002 and a particular computing device 1004. In some implementations, the resource manager 1006 may redirect a particular computing device 1004 to a particular resource provider 1002 with the requested computing resource.

[0410] FIG.11 shows an example of a computing device 1100 and a mobile computing device 1150 that can be used to implement the techniques described in this disclosure. The computing device 1100 is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The mobile computing device 1150 is intended to represent various forms of mobile devices, such as personal digital assistants, cellular telephones, smart- phones, and other similar computing devices. The components shown here, their connections and relationships, and their functions, are meant to be examples only, and are not meant to be limiting.

[0411] The computing device 1100 includes a processor 1102, a memory 1104, a storage device 1106, a high-speed interface 1108 connecting to the memory 1104 and multiple high- speed expansion ports 1110, and a low-speed interface 1112 connecting to a low-speed expansion port 1114 and the storage device 1106. Each of the processor 1102, the memory 1104, the storage device 1106, the high-speed interface 1108, the high-speed expansion ports 1110, and the low-speed interface 1112, are interconnected using various busses, and may be mounted on a common motherboard or in other manners as appropriate. The processor 1102 can process instructions for execution within the computing device 1100, including instructions stored in the memory 1104 or on the storage device 1106 to display graphical information for a GUI on an external input / output device, such as a display 1116 coupled to the high-speed interface 1108. In other implementations, multiple processors and / or multiple buses may be used, as appropriate, along with multiple memories and types of memory. Also, multiple computing devices may be connected, with each device providing portions of the necessary operations (e.g., as a server bank, a group of blade servers, or a multi-processor system). Thus, - 100 - 11901435v1Attorney Docket No.2013969-0048 as the term is used herein, where a plurality of functions are described as being performed by “a processor”, this encompasses embodiments wherein the plurality of functions are performed by any number of processors (one or more) of any number of computing devices (one or more). Furthermore, where a function is described as being performed by “a processor”, this encompasses embodiments wherein the function is performed by any number of processors (one or more) of any number of computing devices (one or more) (e.g., in a distributed computing system).

[0412] The memory 1104 stores information within the computing device 1100. In some implementations, the memory 1104 is a volatile memory unit or units. In some implementations, the memory 1104 is a non-volatile memory unit or units. The memory 1104 may also be another form of computer-readable medium, such as a magnetic or optical disk.

[0413] The storage device 1106 is capable of providing mass storage for the computing device 1100. In some implementations, the storage device 1106 may be or contain a computer- readable medium, such as a floppy disk device, a hard disk device, an optical disk device, or a tape device, a flash memory or other similar solid state memory device, or an array of devices, including devices in a storage area network or other configurations. Instructions can be stored in an information carrier. The instructions, when executed by one or more processing devices (for example, processor 1102), perform one or more methods, such as those described above. The instructions can also be stored by one or more storage devices such as computer- or machine- readable mediums (for example, the memory 1104, the storage device 1106, or memory on the processor 1102).

[0414] The high-speed interface 1108 manages bandwidth-intensive operations for the computing device 1100, while the low-speed interface 1112 manages lower bandwidth-intensive operations. Such allocation of functions is an example only. In some implementations, the high- speed interface 1108 is coupled to the memory 1104, the display 1116 (e.g., through a graphics processor or accelerator), and to the high-speed expansion ports 1110, which may accept various expansion cards (not shown). In the implementation, the low-speed interface 1112 is coupled to the storage device 1106 and the low-speed expansion port 1114. The low-speed expansion port 1114, which may include various communication ports (e.g., USB, Bluetooth®, Ethernet, wireless Ethernet) may be coupled to one or more input / output devices, such as a keyboard, a - 101 - 11901435v1Attorney Docket No.2013969-0048 pointing device, a scanner, or a networking device such as a switch or router, e.g., through a network adapter.

[0415] The computing device 1100 may be implemented in a number of different forms, as shown in the figure. For example, it may be implemented as a standard server 1120, or multiple times in a group of such servers. In addition, it may be implemented in a personal computer such as a laptop computer 1122. It may also be implemented as part of a rack server system 1124. Alternatively, components from the computing device 1100 may be combined with other components in a mobile device (not shown), such as a mobile computing device 1150. Each of such devices may contain one or more of the computing device 1100 and the mobile computing device 1150, and an entire system may be made up of multiple computing devices communicating with each other.

[0416] The mobile computing device 1150 includes a processor 1152, a memory 1164, an input / output device such as a display 1154, a communication interface 1166, and a transceiver 1168, among other components. The mobile computing device 1150 may also be provided with a storage device, such as a micro-drive or other device, to provide additional storage. Each of the processor 1152, the memory 1164, the display 1154, the communication interface 1166, and the transceiver 1168, are interconnected using various buses, and several of the components may be mounted on a common motherboard or in other manners as appropriate.

[0417] The processor 1152 can execute instructions within the mobile computing device 1150, including instructions stored in the memory 1164. The processor 1152 may be implemented as a chipset of chips that include separate and multiple analog and digital processors. The processor 1152 may provide, for example, for coordination of the other components of the mobile computing device 1150, such as control of user interfaces, applications run by the mobile computing device 1150, and wireless communication by the mobile computing device 1150.

[0418] The processor 1152 may communicate with a user through a control interface 1158 and a display interface 1156 coupled to the display 1154. The display 1154 may be, for example, a TFT (Thin-Film-Transistor Liquid Crystal Display) display or an OLED (Organic Light Emitting Diode) display, or other appropriate display technology. The display interface 1156 may comprise appropriate circuitry for driving the display 1154 to present graphical and - 102 - 11901435v1Attorney Docket No.2013969-0048 other information to a user. The control interface 1158 may receive commands from a user and convert them for submission to the processor 1152. In addition, an external interface 1162 may provide communication with the processor 1152, so as to enable near area communication of the mobile computing device 1150 with other devices. The external interface 1162 may provide, for example, for wired communication in some implementations, or for wireless communication in other implementations, and multiple interfaces may also be used.

[0419] The memory 1164 stores information within the mobile computing device 1150. The memory 1164 can be implemented as one or more of a computer-readable medium or media, a volatile memory unit or units, or a non-volatile memory unit or units. An expansion memory 1174 may also be provided and connected to the mobile computing device 1150 through an expansion interface 1172, which may include, for example, a SIMM (Single In Line Memory Module) card interface. The expansion memory 1174 may provide extra storage space for the mobile computing device 1150, or may also store applications or other information for the mobile computing device 1150. Specifically, the expansion memory 1174 may include instructions to carry out or supplement the processes described above, and may include secure information also. Thus, for example, the expansion memory 1174 may be provide as a security module for the mobile computing device 1150, and may be programmed with instructions that permit secure use of the mobile computing device 1150. In addition, secure applications may be provided via the SIMM cards, along with additional information, such as placing identifying information on the SIMM card in a non-hackable manner.

[0420] The memory may include, for example, flash memory and / or NVRAM memory (non-volatile random access memory), as discussed below. In some implementations, instructions are stored in an information carrier. The instructions, when executed by one or more processing devices (for example, processor 1152), perform one or more methods, such as those described above. The instructions can also be stored by one or more storage devices, such as one or more computer- or machine-readable mediums (for example, the memory 1164, the expansion memory 1174, or memory on the processor 1152). In some implementations, the instructions can be received in a propagated signal, for example, over the transceiver 1168 or the external interface 1162. - 103 - 11901435v1Attorney Docket No.2013969-0048

[0421] The mobile computing device 1150 may communicate wirelessly through the communication interface 1166, which may include digital signal processing circuitry where necessary. The communication interface 766 may provide for communications under various modes or protocols, such as GSM voice calls (Global System for Mobile communications), SMS (Short Message Service), EMS (Enhanced Messaging Service), or MMS messaging (Multimedia Messaging Service), CDMA (code division multiple access), TDMA (time division multiple access), PDC (Personal Digital Cellular), WCDMA (Wideband Code Division Multiple Access), CDMA2000, or GPRS (General Packet Radio Service), among others. Such communication may occur, for example, through the transceiver 1168 using a radio-frequency. In addition, short-range communication may occur, such as using a Bluetooth®, Wi-Fi™, or other such transceiver (not shown). In addition, a GPS (Global Positioning System) receiver module 1170 may provide additional navigation- and location-related wireless data to the mobile computing device 1150, which may be used as appropriate by applications running on the mobile computing device 1150.

[0422] The mobile computing device 1150 may also communicate audibly using an audio codec 1160, which may receive spoken information from a user and convert it to usable digital information. The audio codec 1160 may likewise generate audible sound for a user, such as through a speaker, e.g., in a handset of the mobile computing device 1150. Such sound may include sound from voice telephone calls, may include recorded sound (e.g., voice messages, music files, etc.) and may also include sound generated by applications operating on the mobile computing device 1150.

[0423] The mobile computing device 1150 may be implemented in a number of different forms, as shown in the figure. For example, it may be implemented as a cellular telephone 1180. It may also be implemented as part of a smart-phone 1182, personal digital assistant, or other similar mobile device.

[0424] Various implementations of the systems and techniques described here can be realized in digital electronic circuitry, integrated circuitry, specially designed ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least - 104 - 11901435v1Attorney Docket No.2013969-0048 one programmable processor, which may be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0425] Actions associated with implementing the systems may be performed by one or more programmable processors executing one or more computer programs. All or part of the systems may be implemented as special purpose logic circuitry, for example, a field programmable gate array (FPGA) or an application-specific integrated circuit (ASIC), or both. All or part of the systems may also be implemented as special purpose logic circuitry, for example, a specially designed (or configured) central processing unit (CPU), conventional central processing units (CPU) a graphics processing unit (GPU), and / or a tensor processing unit (TPU).

[0426] These computer programs (also known as programs, software, software applications or code) include machine instructions for a programmable processor, and can be implemented in a high-level procedural and / or object-oriented programming language, and / or in assembly / machine language. As used herein, the terms machine-readable medium and computer-readable medium refer to any computer program product, apparatus and / or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and / or data to a programmable processor, including a machine- readable medium that receives machine instructions as a machine-readable signal. The term machine-readable signal refers to any signal used to provide machine instructions and / or data to a programmable processor.

[0427] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input. - 105 - 11901435v1Attorney Docket No.2013969-0048

[0428] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a client computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0429] The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

[0430] In some implementations, modules described herein can be separated, combined or incorporated into single or combined modules. The modules depicted in the figures are not intended to limit the systems described herein to the software architectures shown therein. I. Examples I.i Example 1: Example Training Curves and Generation of De-Novo Structures

[0431] This example demonstrates training of a generative machine learning model according to various embodiments described herein and its use for generating a de-novo peptide backbone structure. FIG.12A is graph showing an example loss curve during training of a machine learning model used to estimate velocity fields in a flow-matching framework for generating de-novo peptide backbones.

[0432] FIG.12B is graph showing an example loss curve showing validation of a machine learning model used to estimate velocity fields in a flow-matching framework for generating de-novo peptide backbones, according to an illustrative embodiment.

[0433] FIG.12C is a series of snapshots showing computed atom positions at various time points (0%, 25%, 50%, 75%, and 100%) as they are iteratively adjusted to create a - 106 - 11901435v1Attorney Docket No.2013969-0048 generated scaffold model representing a generated de-novo peptide backbone. Overlays at a final, 100%, structure show generated helical motifs. I.ii Example 2: Example Metrics for Assessing Final Structure Quality

[0434] While FIGs.12A and 12B evaluate loss, during training and validation, based on determined velocity field values, in certain embodiments, additional and / or alternative validation metrics may be determined based on / using a final generated scaffold model. For example, in certain embodiments, a self-consistency template modeling score (SC TM) may be determined using a scaffold model generated via various embodiments described herein. In certain embodiments, a SC TM score may be determined by populating a generated scaffold model with amino acids to create a generated amino acid sequence. The generated amino acid sequence can then be used, for example via one or more protein folding models (e.g., machine learning models) to create a predicted 3D folded protein structure. The backbone of the predicted 3D folded protein structure can then be compared to determine a level of consistency / similarity, via a metric such as a TM score (template modeling score). I.iii Example 3: Example Equivariant Model

[0435] This example describes an example model that is equivariant with respect to three-dimensional rotations and / or translations of its input. The model of the present example leverages multiple representations of local frames, shifting between them, and makes use of an approach whereby positions and / or orientations of each local frame can be represented relative to its position and / or orientation at another, e.g., earlier, time point.

[0436] In certain embodiments, approaches described herein may utilize machine learning models alone or in combination with various feature vector representations (of peptide backbone sites) to compute quantities such as velocity fields and / or feature vector values representing positions and / or rotations of backbone sites in a manner that is invariant and / or equivariant with respect to three-dimensional translation and / or rotation of a received input. An invariant function or model, as used herein, refers to a function or model whose output is invariant with respect to certain operations, such as rotation or translation, on its input. That is, a - 107 - 11901435v1Attorney Docket No.2013969-0048 model that is invariant with respect to three-dimensional rotations and translations will, (i), for a particular input received, produce a particular corresponding output, and, (ii), if a rotated and / or translated version of that particular input is received, still produce that same particular corresponding output. An equivariant function or model, as used herein, refers to a function or model whose output is equivariant with respect to certain operations, such as rotation and / or translation, of its input. That is, a model that is equivariant with respect to three-dimensional rotations and translations will, (i), for a particular input received, produce a particular corresponding output, and, (ii), if a rotated and / or translated version of that particular input is received, produce a likewise rotated and / or translated version (i.e., having a same rotation and / or translation as the input) of the particular corresponding output. That is, in brief, an invariant model’s output does not vary upon rotation and / or translation of its input, while an equivariant model’s output rotates and / or translates with rotation and / or translation of its input.

[0437] Without wishing to be bound to any particular theory, as described in this example, in certain embodiments, an equivariant model facilitates certain formats of a flow- matching framework, and can be desirable / allow for improved accuracy, particularly with respect to modelling large proteins and / or their tertiary structure. (a) Equivariant Framework

[0438] In this example, several approaches for representing local frames that correspond to peptide backbone sites are used to allow for machine learning models to generate predictions of sets of prospective final feature vector values and, from them, compute velocity fields in an equivariant fashion with respect to rotations and / or translations of input.

[0439] In one representation, an orientation and position of a local frame, i, at a time t may be represented using a set of rotations and translations with reference to a common global frame as shown in Equation (1), below. Eq. (1)^^^^ ^^^^(^^^^)∈ ℝ4,

[0440] In Equation (1), qi(t)used to represent a rotation of frame i at time t and xi(t) is a three-dimensional translation of frame i at time t. The representation shown in - 108 - 11901435v1Attorney Docket No.2013969-0048 Equation (1) uses a global reference frame, such that rotations and / or translations of each frame i are defined with respect to a single common global coordinate system, and not relative to any other frame and / or time point.

[0441] In another representation, a local frame j at a time t may be represented from a point of view of another backbone site – i.e., another local frame, i, also at a same time t, as shown in Equation (2), below. ^^^^^^^^ ^^^^( ^^^^) = ^^^^ ( ^^^^)−1^^^^ ( ^^^^), Eq. (2) �^^^^ ^^^^^^^^ ^^^^ ^^^^ = ^^^^ ^^^^ ^^^^ ^^^^ − ^^^^ ^^^^ �

[0442] may be represented from a point of view of that same local frame at another time, t, as shown in Equation (3), below. Eq. (3) �^^^^ ^^^^( ^^^^, ^^^^′) = ^^^^^^^^( ^^^^)−1^^^^^^^^( ^^^^′), ^^^^ ^^^^ = ^^^^ ^^^^−1^^^^ ^^^^ − ^^^^ ^^^^ ]

[0443] The relations inbetween a global representation of local frames and various relative representations, e.g., representing a particular local frame relative to neighboring local frames and / or a position / orientation of that same local frame at another (e.g., earlier) time point.

[0444] As shown in the below, shifting between global and various relative local frame representations can be used to employ a machine learning model (e.g., a deep neural network) that generates predictions of a set of prospective final feature vectors representing final positions and / or orientations of each local frame.

[0445] In particular, a function f that is invariant with respect to translations and / or rotations of its input may be determined as shown in Equation (4), below. Eq. (4) ^^^^:[^^^^ ^^^^(^^^^), ^^^^ ^^^^(^^^^)]→� ^^^^ ^^^^ ^^^^(^^^^), ^^^^ ^^^^ ^^^^(^^^^)^^^^ ^^^^ [^^�^^ ^^^^(^^^^, 1), ^^�^^ ^^^^(^^^^, 1)]

[0446] , positions and orientations, a relative representation according to Equation (2) is created, and then used as input to a machine learning model (the subscript, θ, denoting learned parameters of the model), such as a deep neural network. Machine learning model fθgenerates predictions of a set of prospective final feature vector values, [ ^^�^^^^^^( ^^^^, 1), ^^�^^^^^^( ^^^^, 1)], as output, where t = 1 is a final time - 109 - 11901435v1Attorney Docket No.2013969-0048 point (t ranging from [0,1]). Accordingly, function f takes, as input, a representation of eachlocal frame at time t according to a common, global reference,[^^^^ ^^^^(^^^^), ^^^^ ^^^^(^^^^)]and generates, asoutput, a set of prospective (predicted) final feature vector values,[^^�^^ ^^^^(^^^^, 1), ^^�^^ ^^^^(^^^^, 1)], representedas a translation and rotation relative to that local frame’s representation at time t (i.e., the input values).

[0447] In order to produce an equivariant output, that shifts and / or rotates in line with its input, an additional transformation can be used to shift back to a global reference, as shown in Equation (5), below, where f is as defined in Equation (4):Eq. (5) ^^^^:[^^^^ ^^^^(^^^^), ^^^^ ^^^^(^^^^)]→^^^^[^^�^^ ^^^^(^^^^, 1), ^^�^^ ^^^^(^^^^, 1)]→[^^^^ ^^^^(^^^^)^^�^^ ^^^^(^^^^, 1)+ ^^^^ ^^^^(^^^^), ^^^^ ^^^^(^^^^)^^�^^ ^^^^( ^^^^, 1)]accordingly, is equivariant. Eq. (6) �^^^^ ^^^^( ^^^^) ^^�^^^^^^( ^^^^, 1) = ^^�^^^^^^(1), ^^^^ ^^^^ ^^^^ ^^^^, + ^^^^ ^^^^ = ^^^^ ^^^^)

[0449] g, in accordance with Equation (5), predictions of a set of final prospective feature vector values that represent positions and / or orientations of local frames can be generated using a machine learning model. (b) Velocity Computations

[0450] An equivariant model for generating predictions of prospective final feature vector values can be used, in turn, to generate predictions of a velocity field at a particular time point, t. As described herein, predicted velocity fields can be used (i) during training, to compare with a ground truth predicted velocity and evaluate a loss function in order to train a machine learning model and (ii) during inference, to adjust position and / or orientation components of feature vectors at each time point.

[0451] As described herein, in certain embodiments, a machine learning model can be used to generate velocity field estimates directly, as output.

[0452] In this example, however, as described above, a machine learning model is used to generate predictions of a set of prospective final feature vector values, representing predicted - 110 - 11901435v1Attorney Docket No.2013969-0048 positions and / or orientations of each backbone site’s position and / or orientation in a final generated peptide. Accordingly, at each training step and / or time point, a set of prospective final feature vector values is generated and then used to compute a corresponding velocity field estimate that, in turn, can be used to compute loss (for training) or adjust position and / or orientation components of a current feature vector, e.g., to advance to a next iteration in a process of generating a final peptide structure.

[0453] In terms of training, at each step, velocity predictions and loss can be calculated as follows: 1. Machine learning model f receives current feature vector values and a current time point as input and generates, as output, estimated final feature vector values (at time t = 1): ^^^^(^^^^ ^^^^, ^^^^)= ^�^^^12. A velocity field estimate, ^^�^^^^^^, at time t is computed based on the estimated final feature vector values, ^�^^^1, the current feature vector values, ^^^^^^^^, and the current time point, t; and current loss is computed based on the estimated ( ^^�^^t) and ground truth ( ^^^^^^^^) velocity fields: ^^^^ ^^^^ ^^^^ ^^^^( ^^�^^^^^^, ^^^^^^^^), with x in this and the following sections of this example denoting a feature vector (e.g., inclusive of both position and / or orientation components and not necessarily solely the position components).

[0454] Where paths between x0and x1are geodesics, a ground truth velocity field and an estimated velocity field can be calculated using the relation in Equation (7a), below: Eq. (7a) ^^^^ ^^^^ ^^^^^^^^= ^^^^( ^^^^, ^^^^0, ^^^^1), ^^^^ℎ ^^^^ ^^^^ ^^^^^^^^= ^^^^(0, ^^^^^^^^, ^^^^1) / (1 − ^^^^).

[0455] Equation (7a) can be used to compute estimated velocity fields from predicted final feature vector values generated by a machine learning model as well, for example as shown in Equation (7b), below: Eq. (7b) ^^�^^^^^^= ^^^^(0, ^^^^^^^^, ^^�^^1) / (1 − ^^^^). - 111 - 11901435v1Attorney Docket No.2013969-0048

[0456] In a Euclidean space, Equations (7a) and (7b) follow from a constraint that paths between x0and x1are geodesics, as follows: ^^^^ ^^^^ = ^^^^0 + ^^^^( ^^^^1 − ^^^^0)=(1 − ^^^^) ^^^^0 + ^^^^ ^^^^1;^^^^ =^^^^ ^^^^^^^^ ^^^^ = ^^^^1− ^^^^0; and ^^^^(0, ^^^^ ^^^^ , ^^^^1) = ^^^^1 − ^^^^ ^^^^= ^^^^1− ^^^^0− ^^^^( ^^^^1− ^^^^0)^^^^ ^^^^

[0457] On aarc between x0and x1of a circle resulting from an intersection of S3and a plane made by x0, x1, and O, where O is an origin. This arc can be parameterized using a complex unit circle and its classical 1 to 1 mapping. In this representation, x0becomes exp(iθ0), x1becomes exp(iθ1), and xtbecomes exp(iθt) = exp{i[tθ1+ (1-t)θ0]}. Without a loss of generality, these three points can be represented according to their complex 1 to 1 mapping.

[0458] Accordingly, ^^^^ ^^^^ = exp(^^^^ ^^^^ ^^^^)= exp{^^^^[^^^^ ^^^^1 +(1 − ^^^^)^^^^0]}^^^^^^^^ ^^^^^^^^ =^^^^ ^^^^ = ^^^^(^^^^1 − ^^^^0)exp{^^^^[^^^^ ^^^^1 +(1 − ^^^^)^^^^0]}^^^^ ^^^^

[0459] and their velocities, which are in a 3D Euclidean space (R3). Spherical equations shown in this example may be used for quaternions, which are on a 4D sphere (S3), and have speeds on tangent® planes. - 112 - 11901435v1Attorney Docket No.2013969-0048 (c) Example Training and Inference in Euclidean Space

[0460] Accordingly, in Euclidean space, the above described framework provides the below equations: Eq. (8a) vt= v(t, x0, x1) Eq. (8b) vt= v(0, xt, x1) / (1 – t) Eq. (8c) xt= tx1+ (1 – t)x0Eq. (8d) vt= x1– x0= v(t, x0, x1) Eq. (8e) vt= v(0, xt, x1) / (1 – t) = (x1– xt) / (1 – t)

[0461] These equations can be used, at training and inference steps, for example, as follows: 1. Example Training Steps: 1. Given: x0~ p0; t ~ U[0,1]; and x1~ pdata2. Compute ‘ground truth’ / analytical: xt= tx1+ (1 – t)x0via Eq. (8c) & vt= (x1– xt) / (1 – t) via Eq. (8e) [or, additionally or alternatively, Eq. (8d) – i.e., vt= x1– x0] 3. Use a machine learning model to generate predictions: (xt, t) → fθ→ x1(est)(superscript (est) indicating estimate) 4. Compute loss: vt(est)= (x1(est)– xt) / (1 – t) via Eq. (8e) [or, additionally or alternatively, Eq. (8d) – i.e., vt(est)= x1(est)– x0] Loss = L(vt(est), vt) (e.g., ||vt(est)– vt||2) 2. Example Inference Steps: 1. Given x0~ p0, N desired iterations, so some time interval dt = 1 / N - 113 - 11901435v1Attorney Docket No.2013969-0048 2. at iteration 0: (x0, 0) → fθ→ x1(0)(est) / (1 – 0) via Eq. (8e)[at this step, identical to eq. (8d), i.e., vt= x1– x0] xt(1)(est)= x0+ vt(0)(est)dt 3. for iterations i = 1 through N – 1: (xt(i), t(i)) → fθ→ x1(i)(est)(where t(i)= N dt) vt(i)(est)= (x1(i)(est)– xt(i)(est)) / (1 – t(i)) via (5) [or, additionally or alternatively Eq. (8d) – i.e., vt(i)(est)= x1(i)(est)– x0] xt(i+1)(est)= xt(i)(est)+ vt(i)(est)dt

[0462] for Euclidean space for illustrative purposes – analogous formulations can be used for other (e.g., spherical, cylindrical) formulations described herein.

[0463] Accordingly, this example describes a particular form of a flow matching approach in accordance with certain embodiments described herein that utilizes a machine learning model to generate predictions of final feature vector position and / or orientation components from which velocity estimates are computed. Without wishing to be bound to any particular theory, in certain embodiments, it is believed that this approach may provide advantages, particularly where large proteins are being generated and / or with respect to tertiary structure predictions. For example, during inference, velocities may be repeatedly calculated at each of a plurality of iterations and used to adjust positions and / or orientations of backbone sites to create a final generated backbone structure. Where velocities are directly computed by a machine learning model, even where errors in individual velocity predictions are small, these errors may compound, resulting in reduced accuracies in long range / high level structure of larger protein molecules, since several such predictions are used together to repeatedly adjust backbone site positions and orientations. By utilizing a machine learning model to output predictions of final backbone site positions and orientations, this compounding issue can be - 114 - 11901435v1Attorney Docket No.2013969-0048 reduced, since the machine learning model predictions at each iteration are made with respect to a final structure, rather than a localized adjustment. I.iv Example 4: Example Framework for Flow Matching in Euclidean Space and SO(3) Manifolds for Protein Generation

[0464] This example describes a generalized framework for using flow matching techniques in Euclidean space and / or on manifolds to generate protein structures. (a) Paths on manifolds

[0465] As described herein, flow matching-based protein generation techniques may utilize a local frame representation that includes rotational components that represent 3D orientations of local frames used to represent peptide backbone sites. Accordingly, among other things, training and inference procedures for models that use a local frame representation may be facilitated by approaches for generating straight paths in 3D rotational space, SO(3), which can be used to interpolate rotational velocity values during training and update rotational components via push-forward equations during inference and / or generation steps. See, e.g., [Lipman 2022] and [Chen and Lipman 2023]. In particular, exponential and logarithmic maps of Riemannian geometries can be used to generate convenient forms of equations representing linear interpolations between rotational frames, which, in turn, can be used to compute losses during training and push rotational components along a desired path during inference.

[0466] Equation (9a), below, describes a linear interpretation between two frame rotation components, R0and R1, as a function of time, where time varies from 0 to 1, Eq. (9a) ^^^^(^^^^)= ^^^^01− ^^^^^^^^^^^^, ^^^^ ∈[0, 1].

[0467] (R)] and denoting an initial velocity, ^^^^0 =log( ^^^^−10 ^^^^1), equation (9b) describes another interpolation between two frame rotationcomponents, which may, in certain embodiments, be more convenient to use in practice than Eq. (9a), Eq. (9b) ^^^^(^^^^)= ^^^^0exp ( ^^^^ ^^^^0).- 115 - 11901435v1Attorney Docket No.2013969-0048

[0468] Accordingly, by training a neural network to predict initial velocity, V0, equation (9b) can be used to update (e.g., push) rotational components of feature vectors in a local frame representation along a path, R(t). See, e.g., [Bose 2023] and [Yim 2023]. An inverse of initial frame, R0-1, which appears in the relation for V0can be computed via equations (9c) and (9d) for rotations represented via 3D rotation matrices (i.e., as a matrix transpose) and unit quaternions (i.e., as a conjugate), respectively. Eq. (9c) ^^^^0−1= ^^^^0^^^^for SO(3).(b) Invariance Properties

[0469] Physical proteins are rotationally and translationally invariant – that is, a rotated and translated (moved in 3D space) version of a particular protein is still that same particular protein (and not, e.g., a new, different, protein). Accordingly, flow matching approaches formulated for proteins can, in certain embodiments, enforce rotational and translational invariances.

[0470] For example, translational invariance can be accomplished by centering a position of each protein representation before processing.

[0471] Methods for ensuring rotational invariance include (i) randomly rotating protein representations at t = 0 and t = 1, before interpolating them, and / or (ii) aligning x1and x0before interpolating them. Alignment can be accomplished using approaches such as a Kabsch algorithm. By utilizing rotation and / or alignment approaches, neural networks can be trained to predict an equivariant velocity field, e.g., as shown below, and encode a desired invariance property. ^^^^(^^^^ ^^^^, ^^^^)= ^^^^ ^^^^(^^^^, ^^^^).

[0472] Example procedures for training a generative flow matching model using either method are described in the following, for both individual backbone atom representation and - 116 - 11901435v1Attorney Docket No.2013969-0048 local frame representation approaches for encoding protein backbone geometries. In the below, xCArefers to a position of a Cα atom, used in certain embodiments as a position of a local frame. 1. Example Training Steps – Individual Backbone Atom Representation 1. Given / sample: x0~ p0; t ~ U(0,1); and x1~ p1; 2. Center x0and x1; 3. Align and / or rotate x0and x1(for rotational equivariance): Option 1: Align x0and x1(e.g., via Kabsch algorithm), or Option 2: Randomly rotate x0and x1(e.g., sample two 3D rotations from a uniform distribution on SO(3)); 4. Determine a value of x(t) by interpolating along a geodesic from x0to x1, e.g., according to: x(t) = x0+ t (x1- x0); and 5. Determine an example velocity at origin as v0= (x1– x0). 2. Example Training Framework – Local Frame Representation 1. Sample x0CA~ p0,R0~ Q0, (x1CA, R1) ~ p1and t ~ U(0, 1); 2. Center3. Align and / or rotate x0CAand x1CA(rotational-equivariance): Option 1: Align x0CAand x1CA(e.g., via Kabsch algorithm), or Option 2: Randomly rotate x0CAand x1CA(e.g., sample two 3D rotations from a uniform distribution on SO(3));4. Determine a value of x(t) = (xtCA, Rt) by interpolating along geodesics from (i) x0CAto x1CAand (ii) R0to R1, e.g., according to: xtCA= (1 - t)×x0CA+ t×x1CA, and - 11901435v1Attorney Docket No.2013969-0048 Rt= R0exp(t log(R0-1R1)); 5. Determine velocities at an origin via: v0CA= x1– x0, and V0= log(R0-1R1).

[0473] Accordingly, steps 1 through 5, above, can be used, in connection with approaches that utilize an individual backbone atom representation and / or local frame representation of peptide backbones, to determine (1) velocity fields at an origin and (2) values of x(t) along a geodesic path from x0to x1at a sampled time point, e.g., t ~ U(0, 1), based on values of x(t) at t = 0 and t = 1 sampled from an initial seed (e.g., prior) distribution and a dataset of example protein structures, respectively. At each training step, a machine learning model may then be provided inputs of x(t), and t – e.g., (xt, t) or (xtCA, Rt, t), depending on whether an individual backbone atom or local frame representation is used – and generate, as output, predicted velocity fields at an origin. These predicted velocity fields can then be compared with directly computed velocity fields (e.g., according to step 5, above) and used to determine loss values (e.g., a mean-squared loss (MSE)) determined velocities determined and asked to predict velocities at the origin and compare with computed velocities at the origin based on the above. A (e.g., thus) determined loss may then be used to update machine learning model parameters. (c) Example Seed Distributions

[0474] As described herein, flow matching techniques sample feature vector components, such as individual backbone atom positions and / or alpha-Carbon and frame rotations (e.g., quaternions) from initial seed distributions during training as well as during inference, e.g., to obtain initial starting values. Flow matching techniques as described herein may work with a variety of seed distributions. Among other things, for example, machine learning models that use an individual backbone atom representation may sample locations of each of four backbone atoms (for each amino acid site) - N, C, O, Cα- from an isotropic Gaussian distribution. In certain embodiments, machine learning model that use local frame representations may sample alpha-Carbon positions from an isotropic Gaussian distribution and may sample rotational component values from a uniform distribution on SO(3). Other approaches for generating seeds - 118 - 11901435v1Attorney Docket No.2013969-0048 include use of random walks (e.g., a random walk Cα seed), using an isotropic Gaussian for Cα positions in a local frame representation and then using a traveling salesman optimization approach to find a shortest path between nodes, etc. (d) Generation / Inference

[0475] In certain embodiments, in order to generate a backbone, an ordinary differential equation (ODE) solution approach, such as a Euler method, a Runge-Kutta method, etc., can be used to update feature vector values, e.g., in an iterative manner, based on predicted velocity fields as time increments are stepped through. In certain examples described herein, such as GNN and transformer implementations demonstrated in Examples 7 and 8, an Euler method was used and found to provide appropriate performance. In certain implementations and embodiments, other ODE solution methods may be utilized. As described herein, 3D positions of individual backbone atoms and / or alpha-Carbon atoms evolve to traverse paths in Euclidian space and, additionally or alternatively, 3D rotation components traverse geodesic paths along a SO(3) manifold. Example Euler ODE formulations in Euclidean space and on a SO(3) manifold are shown in equations (9e) and (9f), respectively, below. Equation (9e) may be used as a “push- forward” function for evolving positions of individual backbone atoms and / or local frames (e.g., alpha-Carbon atom positions) and equation (9f) may be used as a push-forward function for evolving rotations. Eq. (9e) ^^^^( ^^^^ + ∆ ^^^^) = ^^^^( ^^^^) + ∆ ^^^^ × ^^^^( ^^^^). Eq. (9f) ^^^^( ^^^^ + ∆ ^^^^) = ^^^^( ^^^^) exp(∆ ^^^^ ^^^^0).

[0476] In equations (9e) and (9f), above, V(t) (in certain embodiments, V(t) = v0) and V0are predicted via a machine learning model, e.g., as described herein (e.g., above) and Δt is a time increment. In certain embodiments, while velocities shown in Equations (9e) and (9f) are constant values, they are predicted anew at each iteration by a machine learning model receiving updated input – accordingly, velocity predicted at each iteration may change significantly e.g., in norm and direction. - 119 - 11901435v1Attorney Docket No.2013969-0048

[0477] In certain embodiments, protein backbone generation techniques described herein may utilize a timestep scheduler, whereby a scheduling function changes a speed traveled along a geodesic. In particular, a path on a manifold can be written according to ^^^^( ^^^^) = ^^^^0exp ( ^^^^ ^^^^0).

[0478] In certain implementations, a scheduling function, α(t), can be used to adjust a speed along a geodesic. An adjusted path, incorporating a scheduling function, can, accordingly, be written as shown below. ^^^^( ^^^^) = ^^^^0exp ( ^^^^( ^^^^) ^^^^0).

[0479] A corresponding, adjusted, push-forward equation used to determine a rotation at a subsequent time step may then be written as shown below, where α’ is a time derivative of α(t). ^^^^(^^^^ + ∆ ^^^^)= ^^^^( ^^^^)exp( ^^^^′(^^^^)∆ ^^^^ × ^^^^0).Specifically, when Δt is small, ^^^^(^^^^ + ∆ ^^^^)≈ ^^^^(^^^^)+ ^^^^′(^^^^)∆ ^^^^,^^^^(^^^^)= ^^^^0exp ( ^^^^(^^^^)^^^^0)^^^^(^^^^ + ∆ ^^^^)= ^^^^0exp ( ^^^^(^^^^ + ∆ ^^^^)^^^^0)^^^^(^^^^ + ∆ ^^^^)≈ ^^^^(^^^^)+ ∆ ^^^^ ∗ ^^^^′( ^^^^),Hence we can deduce: ^^^^(^^^^ + ∆ ^^^^)≈ ^^^^0exp ( ^^^^(^^^^)^^^^0 + ^^^^′( ^^^^)∆ ^^^^ ^^^^0),^^^^(^^^^ + ∆ ^^^^)≈ ^^^^0 exp(^^^^(^^^^)^^^^0)× exp( ^^^^′( ^^^^)∆ ^^^^ ^^^^0),^^^^( ^^^^ + ∆ ^^^^) = ^^^^( ^^^^)exp ( ^^^^′( ^^^^)∆ ^^^^ × ^^^^0).

[0480] In certain implementations, it was found that accelerating a path of frame rotations at earlier time steps during inference improved a quality of generated backbones, albeit, in certain cases, at a cost of reducing a diversity of generated backbones. A variety of functional forms can be used. An example scheduling function used in certain examples described herein can be defined in terms of its derivative, - 120 - 11901435v1Attorney Docket No.2013969-0048 ^^^^′( ^^^^) = ^^^^(1 − ^^^^) and was found to generate high quality scaffolds for values of k such as 5, 7, and 10 (other values may be used). Other scheduler functions can also be used. For example, a scheduler with the derivative ^^^^′(^^^^)= ^^^^ was also evaluated and found to generate satisfying scaffolds for values of k from 2 to 5. (e) Example Use Cases

[0481] Generative approaches of the present disclosure can be used to create custom peptide backbone structures and / or amino acid sequences, in certain embodiments, from scratch (e.g., referred to as unconditioned generation), based on various partial input structures and / or amino acid sequences, and / or desired properties. Where technologies of the present disclosure receive particular input, such as desired (e.g., partial) structural and / or amino acid sequence features and / or other properties, and generate output scaffold models and / or sequences based therein, generation is referred to as “conditioned on” such particular input.

[0482] Several example use cases are listed below. Detailed examples describing approaches for training and then using machine learning models to generate output in accordance with various use cases (e.g., including, but not limited to, those listed below) are provided in Examples 5-8 below.

[0483] Unconditional generation. In certain implementations and / or uses, generative technologies of the present disclosure may be used to generate scaffold models representing de- novo protein backbones of a variety of sizes, without use of conditioning input. In certain implementations, additionally or alternatively, techniques of described herein may be used to generate amino acid sequences (e.g., together with a de-novo backbone, thereby providing a new backbone structure along with an amino acid sequence predicted to, e.g., fold into that backbone structure).

[0484] Generation conditioned on structural and / or sequence data. In certain implementations and / or uses, conditioning generation on various structural and / or sequence properties can be used to train model(s) that can perform (without limitation) one or more of the following tasks: - 121 - 11901435v1Attorney Docket No.2013969-0048 • Design a binder to a desired (e.g., known, pre-defined, etc.) target (e.g., protein and / or peptide, such as a particular receptor). Binder designs may be conditioned based on known hotspots (e.g., locations on a protein determined to be within a threshold distance of a binding partner and, accordingly, participate in binding) (e.g., on the target), or without knowledge of binding hotspots; • Sample a conformations of particular protein(s) and / or peptide(s) (e.g., sampling conformations can be used to evaluate energy, stability, and other properties of molecules); • Generate a sequence conditioned on a particular scaffold model (e.g., representing a peptide backbone); • Inpaint one or more subsets / portions of a provided partial (e.g., desired, a-priori known, etc.) peptide backbone and / or protein complex; • Dock two or more particular protein and / or peptides based on their monomeric structures (e.g., conditioned on scaffold models and / or sequence data representing peptide backbones and / or amino acid sequences of the particular proteins and / or peptide monomers); • Generate scaffold models and / or sequences thereof representing peptide backbones and / or amino acid sequences conditioned on a particular oligomeric state (e.g., monomer, dimer, multimer, etc.); • Generate a scaffold model representing a peptide backbone conditioned on a particular amino acid sequence (e.g., protein folding); • Generation of side chains (e.g., rotamer structure) from a known backbone and sequence.

[0485] Generation conditioned on global and / or node variables. In certain embodiments, input variables may be used to represent particular (e.g., desired) global properties – for example, these may include, without limitation, one or more of the following: • a protein family variable whose value identifies a particular one of a set of protein family types (e.g., a categorical variable that can take on one of a finite set of values, each - 122 - 11901435v1Attorney Docket No.2013969-0048 identifying a particular protein family; e.g., finite length vector that represents a particular protein family via a one-hot encoding); • a thermostability variable whose value categorizes and / or measures (e.g., quantitatively) protein thermostability [e.g., a binary variable classifying a protein as thermophilic or not (e.g., based on PDB classification); e.g., a continuous real-number measuring melting temperature]; • an immunology / immunogenicity variable whose value classifies and / or measures a propensity and / or likelihood of provoking an immune response [e.g., in a particular host organism (e.g., a human)]; • a function variable whose value classifies protein function; • solubility variable whose value classifies and / or measures a protein solubility; and • a pH sensitivity variable whose value classifies and / or measures protein pH sensitivity.

[0486] In certain implementations, additionally or alternatively, generation may be conditioned based on values of one or more node property variables, each node property variable associated with and representing a particular property of a particular amino acid site. Node property variables may include, without limitation, one or more of the following: • a side chain type variable that identifies a particular type of amino acid side chain (e.g., via a one-hot encoding scheme); • an amino acid polarity variable that identifies a polarity and / or charge of an amino acid site (e.g., a categorical variable having one of four values identifying a particular side chain as polar, apolar, negatively charged, positively charged or unknown (e.g., node to predict or modified / synthetic amino acids)); • a buriedness variable that classifies and / or measures an extent to which a particular amino acid site is buried and / or surface-accessible (e.g., a binary variable classifying a particular node as representing an amino acid site at a surface or core of a protein and / or peptide); • a contact hotspot variable classifying a particular amino acid site according to a desired and / or threshold distance from one or more portions of another (e.g., target) molecule [in - 123 - 11901435v1Attorney Docket No.2013969-0048 certain embodiments, similar and / or related edge property variables may be used, e.g., a contact hotspot variable classifying the edge as associated with one or more particular amino acid sites having a desired and / or within a threshold distance from one or more portions of another (e.g., target) molecule]; and • a secondary structure variable whose value classifies and / or measures a secondary structure motif (e.g., helical, beta-sheet, etc.) at the particular amino acid site. (f) Side Chain Design via Flow Matching

[0487] In certain embodiments, once a structure and sequence have been designed (e.g., to create a binder for instance) side chains may be generated, for example to assess designability of the generated protein structure. For example, whether sufficient space is present between neighboring amino acids to accommodate particular side chains may not be apparent / feasible to determine from backbone geometry alone.

[0488] Accordingly, in certain embodiments, side chain structure can be designed using flow matching generative approaches. In certain embodiments, for example, a side chain design approach may be implemented as an additional step, e.g., following generation of a backbone and / or sequence. Side chain generation approaches described herein may also be implemented and / or used separately from structure design approaches, for example to predict side chain geometries based on a priori known and / or otherwise determined (e.g., via a variety of methods) sequence and backbone geometries.

[0489] Approaches for side chain generation of the present disclosure may use various techniques for representing side chains, for example by representing individual amino acid side chain atoms in 3D space and / or using side chain dihedral (e.g., χ) angles.

[0490] In certain embodiments, where side chains are represented via 3D coordinates of individual constituent atoms, flow matching techniques for Euclidean spaces described herein can be used to predict and evolve locations of individual side chain atoms, in a manner similar to approaches described herein for generation of peptide backbone structures using representations of individual backbone atoms. Fixed and / or known backbone and amino acid sequence information can be provided as conditioning input, e.g., via an edge retrieval layer and / or node - 124 - 11901435v1Attorney Docket No.2013969-0048 embeddings, respectively, analogously to approaches described in examples below for performing inpainting tasks, using an equivariant model architecture. Side chain atom position seeds could be selected from, for example, a Gaussian distribution centered about the amino acid’s Cα backbone atom (e.g., keeping side chain atom points near a Cα atom).

[0491] Additionally or alternatively, models using a dihedral angle representation of amino acid side chain structure may perform flow matching on a manifold geometry, representing each angle as an element of a 2D sphere. Appropriate machine learning model architectures include invariant models (e.g., such as an invariant encoder portion used herein with respect to described transformer models), since dihedral angles are invariant with respect to rotations of an entire protein. Fixed and / or known backbone and amino acid sequence information can be provided as conditioning input, e.g., via an edge retrieval layer and / or node embeddings, respectively, analogously to approaches described in examples below for performing inpainting tasks, using an equivariant model architecture. Seed angle values may be sampled uniformly over a 2D sphere. Training and generation approaches analogous to those described with respect to local frame representations that make use of quaternion representations can be adapted to apply to side chain geometry prediction based on dihedral angles, which also describe angles on a sphere (a 2D sphere, as opposed to a 4D sphere). I.v Example 5: Example Model Architectures

[0492] This example describes several example machine learning model architectures that can be used in connection with flow-matching technologies for protein backbone generation described herein. (a) Building Block Components 1. SO(3) Equivariant Model Components

[0493] As described herein, machine learning models in accordance with certain embodiments aim to learn to generate predictions of a vector field (i.e., velocities) that is equivariant to rotations. Accordingly, model architectures used must also be equivariant to rotations. - 125 - 11901435v1Attorney Docket No.2013969-0048

[0494] An SO(3)-equivariant model can be built by stacking various blocks, such as (i) an SO(3)-invariant node encoder, and (ii) a mix of SO(3)-invariant layers and SO(3)-equivariant layers, interleaved such that a final output is equivariant. Node encoders can be used to encode node-level information, such as an amino acid side chain type embedding, geometric features (e.g., individual backbone atom relationships) at and / or within a particular amino acid site, as well as various node property variables described herein, which, may, for example, be used as conditioning input.

[0495] A variety of types of invariant layers can be used, including, without limitation, graph neural network layers, transformer layers, and layers used to retrieve edge information. Layers for retrieving edge information, referred to hereinafter as “edge retriever layers” are described in further detail below.

[0496] Equivariant layers may include equivariant graph neural network (EGNN) layers, for example as described in [Satorras 2021], as well as particular transformer variants of the present disclosure, for example described in further detail in Section (b), below.

[0497] In certain implementations, where a local frame representation is used, invariant layer outputs can be transformed into equivariant output via equation (10), below, Eq. (10) ^^^^^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^( ^^^^) = ^^^^ ( ^^^^^^ ^^^^^^^^, ^^^^ ^^^^ ^^^^ ^^^^ ^^^ ^^^^) ^^^^^^^^, ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^ ^^^^2. Global Conditioning

[0498] Machine learning models described herein may also receive global input variables, such as time, protein size, as well as various categorical variables that can be used to condition scaffold generation on. Global input can be incorporated via various techniques. For example, certain embodiments of machine learning models in accordance with the present disclosure may use a conditional gating approach, whereby at various (e.g., each) layer(s), global information is input to a feed-forward network (FFN), whose output is, in turn, used to modulate layer outputs within the model architecture. Conditional gating is described in further detail, for example, in [Peebles 2022]. Another approach to including global input is to use global information nodes, whereby artificial nodes are added to an input, and used as registers to store and process global variables. See, e.g., [Darcet 2023]. - 126 - 11901435v1Attorney Docket No.2013969-0048 (b) Example Transformer Architecture

[0499] Turning to FIG.13, among other things, the present disclosure provides an example transformer...

Claims

Attorney Docket No.2013969-0048 What is claimed is:

1. A method for computer generation of a de-novo peptide backbone of a custom biologic, the method comprising: (a) receiving and / or generating, by a processor of a computing device a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular backbone site and comprising position and / or orientation components representing a position and / or orientation of the particular backbone site; (b) determining, by the processor, using a machine learning model, one or more velocity fields and, beginning with the seed set, updating by the processor, values of the position and / or orientation components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving the values of the position and / or orientation components of the plurality of feature vectors from a set of initial starting values into a set of final values representing positions and / or orientations of each backbone site of a generated peptide backbone; and (c) creating, by the processor, using, the set of final values, a generated scaffold model representing the de-novo peptide backbone; and (d) storing and / or providing, the generated scaffold model.

2. The method of claim 1, wherein, for each particular feature vector of the plurality of feature vectors, the position and / or orientation components of the particular feature vector represent a relative translation and / or rotation, respectively, of the corresponding backbone site relative to another, neighbor backbone site corresponding to another feature vector in the seed set.

3. The method of claim 1, wherein, for each particular feature vector of the plurality of feature vectors, the position and / or orientation components of the particular feature vector - 188 - 11901435v1Attorney Docket No.2013969-0048 represent a global translation and / or rotation, respectively, with respect to a single common reference frame.

4. The method of any one of the preceding claims, wherein values of the position and / or orientation components of each feature vector are set to initial starting values selected at random from an initial probability distribution.

5. The method of any one of the preceding claims, comprising generating, by the processor, using a preliminary machine learning model to determine a plurality of sets of initial starting values, each set of initial starting values comprising an initial position and / or orientation and, at step (a), for each feature vector of the plurality of feature vectors selecting and using a set of initial starting values as values for the position and / or orientation components of the feature vector.

6. The method of any one of the preceding claims, wherein the seed set comprises 50 or more feature vectors.

7. The method of any one of the preceding claims, wherein step (b) comprises determining the one or more velocity fields and updating the values of the position and / or orientation components of the plurality of feature vectors in an iterative fashion.

8. The method of claim 7, wherein determining the one or more velocity fields and updating the values of the position and / or orientation components of the plurality of feature vectors in an iterative fashion comprises: determining initial starting values for the position and / or orientation components of each of the plurality of feature vectors; beginning with the initial starting values of the position and / or orientation components of the plurality of feature vectors, at a first iteration: - 189 - 11901435v1Attorney Docket No.2013969-0048 determining, by the processor, using the machine learning model, an initial velocity field based on the initial starting values of the position and / or orientation components of the plurality of feature vectors; and updating values of the position and / or orientation components of the plurality of feature vectors according to the initial velocity field; and at subsequent iterations: using the updated values from a prior iteration as current values of the position and / or orientation components of the feature vectors; determining, using the machine learning model, a current velocity field based on the current values of the position and / or orientation components of the feature vectors; and updating values of the position and / or orientation components of the feature vectors according to the current velocity field; and upon reaching a final iteration, using the updated values of the position and / or orientation components of the feature vectors as the final set of values for creating the generated scaffold model representing the de-novo peptide backbone.

9. The method of claim 8, wherein initial starting values of the position and / or orientation components of each feature vector are selected at random from an initial probability distribution.

10. The method of claim 8, comprising determining, by the processor, the set of initial starting values using a preliminary machine learning model.

11. The method of any one of claims 8 to 10, wherein each iteration corresponds to one of a plurality of time-points, and (i) the initial velocity field is determined based at least in part on an initial time-point corresponding to the first iteration; and / or - 190 - 11901435v1Attorney Docket No.2013969-0048 (ii) each current velocity field associated with and determined at a particular subsequent iteration is determined based at least in part on a particular subsequent time- point corresponding to the subsequent iteration.

12. The method of claim 11, wherein, at each particular iteration, the machine learning model receives, as input, (i) the current values of the feature vectors and (ii) the time point corresponding to the particular iteration, and generates, as output, the current velocity field.

13. The method of claim 11, wherein, at each particular iteration, the machine learning model receives, as input, (i) the current values of the feature vectors and (ii) the time point corresponding to the particular iteration, and generates, as output, a set of prospective final values of the position and / or orientation components of the feature vectors and the current velocity field is determined based on the set of prospective final values and the time point corresponding to the current iteration.

14. The method of claim 12 or 13, wherein the current time point is used, within the machine learning model, as a conditioning value.

15. The method of any one of the preceding claims, wherein at step (b), the machine learning model generates at least a portion of the one or more velocity fields as output.

16. The method of any one of the preceding claims, wherein at step (b), the machine learning model generates a set of prospective final values of the position and / or orientation components of the feature vectors as output and the current velocity field is determined based on the set of prospective final values and a current time point.

17. The method of claim 16, wherein the set of prospective final values comprises, for each particular feature vector: - 191 - 11901435v1Attorney Docket No.2013969-0048 (i) a prospective final value of the position component represents a translation relative to a position component of the particular feature vector at the current time; and / or (ii) a prospective final value of the orientation component represents a rotation relative to an orientation component of the particular feature vector at the current time.

18. The method of any one of the preceding claims, wherein the machine learning model is or comprises a transformer-based model.

19. The method of claim 18, wherein the machine learning model receives and / or operates on an input graph representation of a peptide backbone, the input graph representation comprising a plurality of nodes and edges, each node corresponding to and representing a particular backbone site, and each edge associated with and relating two nodes, and wherein the machine learning model comprises at least one transformer-based edge retrieval layer that: (A) comprises one or more self-attention head(s), each of which determines a corresponding set of attention weights based at least in part on values of the feature vectors and / or node feature values determined therefrom, such that the edge retrieval layer determines one or more sets of attention weights; and (B) uses the one or more sets of attention weights to determine values of retrieved edge feature vectors.

20. The method of any one of the preceding claims, wherein the machine learning model is or comprises a graph neural network (GNN).

21. The method of any one of the preceding claims, wherein the machine learning model is or comprises an auto-regressive neural network. - 192 - 11901435v1Attorney Docket No.2013969-0048 22. The method of any one of the preceding claims, wherein the machine learning model receives, as input and / or determines, values of one or more edge features, each determined based on values of the position and / or orientation components of two or more feature vectors representing two or more different backbone sites.

23. The method of any one of the preceding claims, comprising conditioning generation of the one or more velocity fields according to a set of one or more desired peptide backbone features.

24. The method of claim 23, wherein the machine learning model receives, as input, and conditions generation of the one or more velocity fields on values one or more global property variables, each global property variable representing a desired property of a protein or peptide.

25. The method of claim 24, wherein the one or more global property variables comprise one or more of the following: (A) a protein family variable whose value identifies a particular one of a set of protein family types; (B) a thermostability variable whose value categorizes and / or measures protein thermostability; (C) an immunology variable whose value classifies and / or measures a propensity and / or likelihood of provoking an immune response; (D) a function variable whose value classifies protein function; (E) a solubility variable whose value classifies and / or measures a protein solubility; and (F) a pH sensitivity variable whose value classifies and / or measures protein pH sensitivity. - 193 - 11901435v1Attorney Docket No.2013969-0048 26. The method of any one of claims 24-25, wherein the machine learning model receives, as input and conditions generation of the one or more velocity fields on, values of one or more node property variables, each node property variable associated with an representing a particular property of a particular amino acid site.

27. The method of claim 26, wherein the one or more node property tokens comprise one or more of the following: (A) a side chain type variable that identifies a particular type of amino acid side chain; (B) an amino acid polarity variable that identifies a polarity and / or charge of an amino acid site; (C) a buriedness variable that classifies and / or measures an extent to which a particular amino acid site is buried and / or surface-accessible; (D) a contact hotspot variable classifying a particular amino acid site according to a desired and / or threshold distance from one or more portions of another molecule; and (E) a secondary structure variable whose value classifies and / or measures a secondary structure motif at the particular amino acid site.

28. The method of any one of claims 23-27, wherein the machine learning model receives, as input, values of one or more edge property variables.

29. The method of any one of the preceding claims, comprising conditioning generation of the one or more velocity fields according to a representation of at least a portion of a target molecule and / or one or more particular sub-regions thereof, thereby creating a de-novo peptide backbone suitable for binding to the target molecule. - 194 - 11901435v1Attorney Docket No.2013969-0048 30. The method of any one of the preceding claims, comprising conditioning generation of the one or more velocity fields according to a representation of at least a portion of a desired final protein and / or peptide.

31. The method of claim 30, wherein the portion of the desired final protein and / or peptide is a portion of an antibody.

32. The method of claim 31, wherein the portion of the antibody is one or more members selected from the group consisting of a Fab region, a variable heavy chain, and a variable light chain.

33. The method of any one of the preceding claims, comprising: using the generated scaffold model as input to an interface designer module for generating an amino acid interface for binding to a target molecule; and populating, by the processor, at least a portion of the generated scaffold model with a plurality of amino acids selected, by the processor based at least in part on (i) the generated scaffold model and (ii) a target model representing at least a portion of the target molecule.

34. The method of any one of the preceding claims, comprising determining, using the machine learning model, one or more sequence velocity fields and using the one or more sequence velocity fields to generate predicted sequence data representing an amino acid sequence of a protein and / or peptide having the de-novo peptide backbone.

35. The method of any one of the preceding claims, comprising determining, using the machine learning model, one or more side chain geometry velocity fields and using the one or more side chain velocity fields to generate a prediction of a three-dimensional side chain geometry for an amino acid side chain at each of at least a portion of amino acid sites of the de- novo peptide backbone. - 195 - 11901435v1Attorney Docket No.2013969-0048 36. The method of any one of the preceding claims, comprising conditioning generation of the one or more velocity fields according to one or more protein fold representation(s).

37. The method of claim 36, wherein the one or more protein fold representation(s) are or comprise a set of secondary structure element (SSE) values, each SSE value associated with a particular position (e.g., amino acid site) within a polypeptide chain of the custom biologic and having a value encoding a particular type of secondary structure at the particular position.

38. The method of claim 36 or 37, wherein the one or more protein fold representation(s) are or comprise a block adjacency matrix, said block adjacency matrix comprising a plurality of elements, each element of the block adjacency matrix associated with a particular pair of positions within a polypeptide chain of the custom biologic and having a value representing a level of interaction between the particular pair of positions.

39. The method of any one of claims 36 to 38, wherein the one or more protein fold representation(s) are or comprise a block adjacency matrix, said block adjacency matrix comprising a plurality of elements, each element of the block adjacency matrix associated with a particular pair of positions within a polypeptide chain of the custom biologic and having one or more values representing a relative position and / or orientation secondary structural elements (SSEs) at the particular pair of positions.

40. A method for computer-aided generative design of de-novo polypeptides and / or complexes thereof, the method comprising: (a) receiving and / or generating, by a processor of a computing device, a seed representation of one or more peptide backbones and / or amino acid sequences, the seed representation comprising a plurality of feature values representing (i) amino acid sites and / or (ii) individual backbone atoms and / or side chain types thereof; - 196 - 11901435v1Attorney Docket No.2013969-0048 (b) determining, by the processor, using one or more machine learning models, one or more velocity fields and, beginning with the seed set, updating, by the processor, the plurality of feature values according to the one or more velocity fields, thereby evolving the representation of one or more peptide backbones and / or amino acid sequences from initial feature values of the seed representation to a set of final feature values representing amino acid sites, and / or individual backbone atoms and / or side chain types, of one or more generated peptide backbones and / or amino acid sequences of the one or more de-novo polypeptide and / or a complex thereof; (c) creating, by the processor, using the set of final values, one or more generated scaffold model(s) and / or sequence string(s) representing the de-novo polypeptide(s) and / or complexes thereof; and (d) storing and / or providing the generated scaffold models and / or sequence strings.

41. The method of claim 40, wherein the plurality of feature values comprise values representing locations (e.g., in Cartesian space) of one or more individual backbone atoms (e.g., N, C, alpha-Carbon, and O) for each of a plurality of amino acid sites of the de-novo polypeptide to-be and / or being generated.

42. The method of claim 40 or 41, wherein the plurality of feature values comprise values representing positions and / or orientations of amino acid sites of the de-novo polypeptide to-be and / or being generated.

43. The method of any one of claims 40 to 42, wherein the plurality of feature values comprise sequence data representing an amino acid sequence of the de-novo polypeptide to-be and / or being generated.

44. The method of any one of claims 40 to 43, comprising: - 197 - 11901435v1Attorney Docket No.2013969-0048 obtaining (e.g., receiving and / or accessing and / or generating), by the processor, a template model representing a partial sequence and / or three-dimensional structure of at least a portion of a reference protein and / or peptide; at step (a), generating the seed representation comprising the plurality of feature values, wherein a first subset of the plurality of feature values correspond to pre-defined regions of the polypeptide(s) being designed and are populated with feature values representing (e.g., and having been determined based on) the partial sequence and / or three dimensional structure of the reference protein and / or peptide (e.g., using the template model) and wherein a second subset of the plurality of feature values correspond to custom, to-be-designed, regions of the polypeptide(s) and are populated with initial starting values; and at step (b), determining the one or more velocity fields based at least in part on the plurality of feature values, including the first subset (corresponding to pre-defined regions) and the second subset (corresponding to unknown region that are, e.g., iteratively, updated) [e.g., thereby conditioning generation of the one or more velocity fields at least in part on the first subset of feature values] and updating the second subset of feature values based on the one or more velocity fields, while retaining / holding fixed the first subset of feature values.

45. The method of any one of claims 40 to 44, comprising: obtaining (e.g., receiving and / or accessing and / or generating), by the processor, a template model representing a partial sequence and / or three-dimensional structure of one or more base portion(s) of a reference protein and / or peptide; determining, by the processor, one or more protein fold representation(s) based on the template model, wherein each of the one or more protein fold representation(s) represent (i) a secondary structure type of, and / or (ii) a relative proximity and / or orientation between amino acid sites of, the one or more base portion(s) of the reference protein and / or peptide [e.g., but do not include or represent definite three-dimensional coordinates of backbone atoms of the reference protein and / or peptide]; at step (a), generating the seed representation comprising the plurality of feature values, said plurality of feature values representing (i) locations and / or orientations of, and / or (ii) - 198 - 11901435v1Attorney Docket No.2013969-0048 individual backbone atoms and / or side chain types of, a plurality of amino acid sites within a de- novo polypeptide (e.g., being designed), and wherein: a first subset of the plurality of feature values represent [e.g., (i) locations and / or orientations, and / or (ii) individual backbone atoms and / or side chain types of] amino acid sites within a first portion of the de-novo polypeptide, said first portion corresponding to the one or more base portion(s) of the reference polypeptide; and a second subset of the plurality of feature values represent [e.g., (i) locations and / or orientations, and / or (ii) individual backbone atoms and / or side chain types of] amino acid sites within a second portion of the de-novo polypeptide, said second portion adjacent to (e.g., having one or more amino acid sites adjacent to sites within) the first portion; and at step (b), determining the one or more velocity fields based at least in part on the one or more protein fold representation(s), such that the one or more generated peptide backbones and / or amino acid sequences comprise a first portion having a three-dimensional fold based on the one or more protein fold representation(s) and a second, custom, portion.

46. The method of claim 44 or 45, wherein the reference protein and / or peptide is an antibody [e.g., and wherein the one or more base portions are or comprise heave and / or light chain region(s) of the antibody, but exclude complementary determining regions (CDRs) of the antibody].

47. A method for generating attention-based deep-learning predictions from input graph representations comprising nodes and edges, the method comprising: (a) receiving and / or accessing, by a processor of a computing device, an initial graph representation comprising a plurality of node feature vectors, each node feature vector associated with a particular node of the initial graph representation and comprising one or more initial node feature values; (b) determining, by the processor, using a machine learning model, a predicted graph representation, wherein the machine learning model: - 199 - 11901435v1Attorney Docket No.2013969-0048 (i) receives the initial graph representation as input; (ii) comprises at least one edge retrieval layer, wherein the at least one edge retrieval layer: (A) comprises one or more self-attention head(s), each of which determines a corresponding set of attention weights based at least in part on values of the feature vectors and / or node feature values determined therefrom, such that the edge retrieval layer determines one or more sets of attention weights; and (B) uses the one or more sets of attention weights to determine values of retrieved edge feature vectors; and (iii) generates, as output, based at least in part on the values of the retrieved edge feature vectors, a plurality of predicted node feature vectors and / or one or more velocity fields, and wherein the predicted graph representation is determined using the plurality of predicted node feature vectors and / or one or more velocity fields; and (c) storing and / or providing the predicted graph representation.

48. A method for designing a custom biologic having a desired three-dimensional fold, the method comprising: (a) receiving and / or generating, by a processor of a computing device, one or more protein fold representations, each encoding desired three-dimensional structural features of the desired fold at / in at least a portion of locations within the custom biologic; (b) receiving and / or generating, by the processor, a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular site within the one or more variable region(s) of template model and comprising one or both of: (A) position and / or orientation components representing a position and / or orientation of the particular site; and - 200 - 11901435v1Attorney Docket No.2013969-0048 (B) a side chain type component, representing likelihoods of one or more particular amino acid side chain types occupying the particular site; (c) determining, by the processor, using a machine learning model, one or more velocity fields based at least in part on the one or more protein fold representations, and, beginning with the seed set, updating, by the processor, values of one or both of: (A) the position and / or orientation components and (B) the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving them a set of initial starting values into a set of final values representing (i) positions and / or orientations of each site of generated peptide backbone structures of the one or more variable regions, and / or (ii) likelihoods of side chain types at locations within the one or more variable regions; (d) generating, by the processor, using the set of final values, one or both of: (A) a scaffold model representing a three-dimensional structure of a de-novo peptide backbone for the custom biologic, and (B) sequence data representing an amino acid sequence of the custom biologic; and (e) storing and / or providing, by the processor, the generated (A) scaffold model and / or (B) sequence data for display and / or further processing and / or use in designing the custom biologic.

49. The method of claim 48, wherein the one or more protein fold representation(s) are or comprise a set of secondary structure element (SSE) values, each SSE value associated with a particular position within a polypeptide chain of the custom biologic and having a value encoding a particular type of secondary structure at the particular position.

50. The method of claims 48 or 49, wherein the one or more protein fold representation(s) are or comprise a block adjacency matrix, said block adjacency matrix comprising a plurality of elements, each element of the block adjacency matrix associated with a particular pair of positions within a polypeptide chain of the custom biologic and having a value representing a level of interaction between the particular pair of positions. - 201 - 11901435v1Attorney Docket No.2013969-0048 51. The method of any one of claims 48 to 50, wherein the one or more protein fold representation(s) are or comprise a block adjacency matrix, said block adjacency matrix comprising a plurality of elements, each element of the block adjacency matrix associated with a particular pair of positions within a polypeptide chain of the custom biologic and having one or more values representing a relative position and / or orientation secondary structural elements (SSEs) at the particular pair of positions.

52. The method of any one of claims 48 to 51, wherein the machine learning model receives the one or more protein fold representations as input, thereby conditioning generation of the one or more velocity fields the desired fold.

53. The method of any one of claims 48 to 52, comprising, at step (a) generating the one or more protein fold representations based on a portion of an antibody template representing at least a portion of a reference antibody structure.

54. The method of claim 53, wherein the portion of the antibody template is a base portion, representing a region of the reference antibody about one or more complementary-determining regions (CDRs) but excluding the CDRs themselves.

55. The method of claim 54, wherein steps (b) through (d) comprises generating one or both of (i) a custom sequence and (ii) a custom three-dimensional peptide backbone structure of the one or more CDRs.

56. A method for designing a custom antibody for binding to a target, the method comprising: (a) receiving and / or generating, by a processor of a computing device, an antibody template representing a sequence and / or three-dimensional structure of at least a portion of a - 202 - 11901435v1Attorney Docket No.2013969-0048 reference antibody, the antibody template comprising a base portion located about one or more CDRs of the reference antibody, but excluding at least a portion of the one or more CDRs themselves; (b) determining, by the processor, one or more protein fold representations, each encoding and representing three-dimensional structural features of at least the base portion of the antibody template; (c) receiving and / or generating, by the processor, a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular site within one or more variable region(s) corresponding to, and representing to-be-designed custom versions of, CDR regions of the reference antibody and comprising one or both of: (A) position and / or orientation components representing a position and / or orientation of (e.g., individual backbone atoms and / or a local frame representation of) the particular site; and (B) a side chain type component, representing likelihoods of one or more particular amino acid side chain types occupying the particular site; (c) determining, by the processor, using a machine learning model, one or more velocity fields based at least in part on the one or more protein fold representations, and, beginning with the seed set, updating, by the processor, values of one or both of (A) the position and / or orientation components and (B) the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving them a set of initial starting values into a set of final values representing (i) positions and / or orientations of each site of generated peptide backbone structures of the one or more variable regions, and / or (ii) likelihoods of side chain types at locations within the one or more variable regions; (d) generating, by the processor, using the set of final values, one or both of: (A) a scaffold model representing a three-dimensional structure of a de-novo peptide backbone for one or more CDR regions of the custom antibody; and (B) sequence data representing an amino acid sequence of one or more CDR regions of the custom antibody; and - 203 - 11901435v1Attorney Docket No.2013969-0048 (e) storing and / or providing, by the processor, the generated (A) scaffold model and / or (B) sequence data, for display and / or further processing and / or use in designing the custom biologic.

57. The method of claim 56, comprising receiving and / or generating a target model representing a sequence and / or three-dimensional structure of at least a portion of the target and wherein step (c) comprises generating one or both of (i) the sequence and (ii) the three- dimensional peptide backbone structure based further on the target model.

58. A method for in-silico generation and / or prediction of three-dimensional (3D) side chain geometries of a polypeptide chain, the method comprising: (a) receiving and / or generating, by a processor of a computing device, (i) sequence data representing an amino acid sequence of at least a portion of the polypeptide chain and (ii) a representation of a 3D peptide backbone geometry, and / or folds thereof, for the portion of the polypeptide chain; (b) receiving and / or generating, by a processor of a computing device a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular amino acid site of the portion of the polypeptide chain and comprising side chain geometry components representing a geometry of a side chain at the particular amino acid site; (c) determining, by the processor, using a machine learning model, one or more velocity fields based at least in part on (i) the sequence data and (ii) the representation of the 3D peptide backbone geometry and / or folds thereof and, beginning with the seed set, updating, by the processor, values of the side chain geometry components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving the values of the side chain geometry components of the plurality of feature vectors from a set of initial starting values into a set of final values representing geometries of each amino acid site of the portion of the polypeptide chain; and - 204 - 11901435v1Attorney Docket No.2013969-0048 (c) creating, by the processor, using, the set of final values, a 3D polypeptide model of the portion of the polypeptide chain representing the 3D geometries of amino acid side chains within the portion of the polypeptide chain; and (d) storing and / or providing, for display and / or further processing, the 3D polypeptide model.

59. A method for in-silico generation and / or prediction of an amino acid sequence of a polypeptide chain, the method comprising: (a) receiving and / or generating, by a processor of a computing device a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular site of the portion of the polypeptide chain and a side chain type component representing a likelihood of one or more possible types of amino acid side chains at the particular site; (b) determining, by the processor, using a machine learning model, one or more velocity fields and, beginning with the seed set, updating, by the processor, values of the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving the values of the side chain type components of the plurality of feature vectors from a set of initial starting values into a set of final values representing likelihoods of amino acid side chain types at each site of the portion of the polypeptide chain; (c) determining, by the processor, using, the set of final values, sequence data representing an amino acid sequence of the portion of the polypeptide chain; and (d) storing and / or providing the sequence data.

60. The method of claim 59, comprising: receiving and / or generating, by the processor, a representation of a 3D peptide backbone geometry, and / or folds thereof, for at least a portion of the polypeptide chain; and at step (b), determining the one or more velocity fields based at least in part on the representation of the 3D peptide backbone geometry and / or folds thereof. - 205 - 11901435v1Attorney Docket No.2013969-0048 61. A method for generating in-silico docking predictions for a plurality (e.g., two or more) of polypeptide chains, the method comprising: (a) receiving and / or generating, by a processor of a computing device, a first scaffold model representation of a first 3D peptide backbone of at least a portion of a first polypeptide chain; (b) receiving and / or generating, by a processor of a computing device, a second scaffold model representation of a second 3D peptide backbone of at least a portion of a second polypeptide chain; (c) receiving and / or generating, by a processor of a computing device a seed set comprising a plurality of feature vectors, the plurality of feature vectors comprising (i) a first set of feature vectors corresponding to and representing positions and / or orientations of the first 3D peptide backbone and / or (ii) a second set of feature vectors corresponding to and representing positions and / or orientations of the second 3D peptide backbone; (d) determining, by the processor, using a machine learning model, based at least in part on the first scaffold model and the second scaffold model, one or more velocity fields, by the processor, and, beginning with the seed set, updating, by the processor, values of (i) the first set of feature vectors and / or (ii) the second set of feature vectors according to the one or more velocity fields, thereby evolving values of the first and / or second feature vector sets from a set of initial starting values into a set of final values representing final position(s) and / or orientation(s) of the first 3D peptide backbone and / or the second 3D peptide backbone docked with each other; (e) creating, by the processor, using, the set of final values, a generated polypeptide complex model representing first and second 3D peptide backbones docked to form a polypeptide complex; and (f) storing and / or providing the generated polypeptide complex model 62. A method for designing a custom biologic for binding to a target, the method comprising: (a) receiving and / or generating, by the processor, a target model representing at least a portion of the target; - 206 - 11901435v1Attorney Docket No.2013969-0048 (b) receiving and / or generating, by a processor of a computing device, a template model representing a sequence and / or three-dimensional structure of at least a portion of a reference biologic, the template model comprising a base portion representing a portion of the reference biologic located about one or more variable regions of the reference biologic designated as modifiable for binding to the target; (c) receiving and / or generating, by the processor, a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular site within the one or more variable region(s) of template model and comprising one or both of (i) position and / or orientation components representing a position and / or orientation of the particular site and (ii) a side chain type component, representing likelihoods of one or more particular amino acid side chain types occupying the particular site; (d) determining, by the processor, using a machine learning model, one or more velocity fields based at least in part on (i) the target model and (ii) the base portion of the template model and, beginning with the seed set, updating, by the processor, values of one or both of (i) the position and / or orientation and (ii) the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving them a set of initial starting values into a set of final values representing (i) positions and / or orientations of each site of generated peptide backbone structures of the one or more variable regions, and / or (ii) likelihoods of side chain types at locations within the one or more variable regions; (e) creating, by the processor, using, the set of final values, one or both of: (i) a generated scaffold model representing a de-novo peptide backbone for the one or more variable regions of the custom biologic; and (ii) sequence data representing a generated amino acid sequence of the one or more variable region(s); and (f) storing and / or providing, by the processor, the peptide backbone structure and / or the generated sequence for display and / or further processing and / or use in designing the custom biologic. - 207 - 11901435v1Attorney Docket No.2013969-0048 63. A system for computer generation of a de-novo peptide backbone of a custom biologic, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular backbone site and comprising position and / or orientation components representing a position and / or orientation of the particular backbone site; (b) determine, using a machine learning model, one or more velocity fields and, beginning with the seed set, updating, by the processor, values of the position and / or orientation components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving the values of the position and / or orientation components of the plurality of feature vectors from a set of initial starting values into a set of final values representing positions and / or orientations of each backbone site of a generated peptide backbone; and (c) create, using, the set of final values, a generated scaffold model representing the de-novo peptide backbone; and (d) store and / or provide the generated scaffold model.

64. A system for computer-aided generative design of de-novo polypeptides and / or complexes thereof, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate a seed representation of one or more peptide backbones and / or amino acid sequences, the seed representation comprising a plurality of - 208 - 11901435v1Attorney Docket No.2013969-0048 feature values representing (i) amino acid sites and / or (ii) individual backbone atoms and / or side chain types thereof; (b) determine, using one or more machine learning models, one or more velocity fields and, beginning with the seed set, updating, by the processor, the plurality of feature values according to the one or more velocity fields, thereby evolving the representation of one or more peptide backbones and / or amino acid sequences from initial feature values of the seed representation to a set of final feature values representing amino acid sites, and / or individual backbone atoms and / or side chain types, of one or more generated peptide backbones and / or amino acid sequences of the one or more de- novo polypeptide and / or a complex thereof; (c) create, using the set of final values, one or more generated scaffold model(s) and / or sequence string(s) representing the de-novo polypeptide(s) and / or complexes thereof; and (d) store and / or provide the generated scaffold models and / or sequence strings.

65. A system for generating attention-based deep-learning predictions from input graph representations comprising nodes and edges, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to (a) receive and / or access an initial graph representation comprising a plurality of node feature vectors, each node feature vector associated with a particular node of the initial graph representation and comprising one or more initial node feature values; (b) determine, using a machine learning model, a predicted graph representation, wherein the machine learning model: (i) receives the initial graph representation as input; - 209 - 11901435v1Attorney Docket No.2013969-0048 (ii) comprises at least one edge retrieval layer, wherein the at least one edge retrieval layer: (A) comprises one or more self-attention head(s), each of which determines a corresponding set of attention weights based at least in part on values of the feature vectors and / or node feature values determined therefrom, such that the edge retrieval layer determines one or more sets of attention weights; and (B) uses the one or more sets of attention weights to determine values of retrieved edge feature vectors; and (iii) generates, as output, based at least in part on the values of the retrieved edge feature vectors, a plurality of predicted node feature vectors and / or one or more velocity fields, and wherein the predicted graph representation is determined using the plurality of predicted node feature vectors and / or one or more velocity fields; and (c) store and / or provide, the predicted graph representation.

66. A system for designing a custom biologic having a three-dimensional structure corresponding to a desired fold, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate one or more protein fold representations, each encoding desired three-dimensional structural features of the desired fold at / in at least a portion of locations within the custom biologic; (b) receive and / or generate a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular site within the one or more variable region(s) of template model and comprising one or both of: - 210 - 11901435v1Attorney Docket No.2013969-0048 (A) position and / or orientation components representing a position and / or orientation of the particular site and (B) a side chain type component, representing likelihoods of one or more particular amino acid side chain types occupying the particular site; (c) determine, using a machine learning model, one or more velocity fields based at least in part on the one or more protein fold representations, and, beginning with the seed set, update values of one or both of (A) the position and / or orientation components and (B) the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving them a set of initial starting values into a set of final values representing (i) positions and / or orientations of each site of generated peptide backbone structures of the one or more variable regions, and / or (ii) likelihoods of side chain types at locations within the one or more variable regions; (d) generate, using the set of final values, one or both of: (A) a scaffold model representing a three-dimensional structure of a de-novo peptide backbone for the custom biologic, and (B) sequence data representing an amino acid sequence of the custom biologic; and (e) store and / or provide the generated (A) scaffold model and / or (B) sequence data for display and / or further processing and / or use in designing the custom biologic.

67. A system for designing a custom antibody for binding to a target, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate an antibody template representing a sequence and / or three-dimensional structure of at least a portion of a reference antibody, the antibody template comprising (i) a base portion located about one or more CDRs of the reference antibody, but excluding at least a portion of the one or more CDRs themselves; - 211 - 11901435v1Attorney Docket No.2013969-0048 (b) determine one or more protein fold representations, each encoding and representing three-dimensional structural features of at least the base portion of the antibody template; (c) receive and / or generate a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular site within one or more variable region(s) corresponding to, and representing to-be-designed custom versions of, CDR regions of the reference antibody and comprising one or both of: (A) position and / or orientation components representing a position and / or orientation of (e.g., individual backbone atoms and / or a local frame representation of) the particular site and (B) a side chain type component, representing likelihoods of one or more particular amino acid side chain types occupying the particular site; (c) determine, using a machine learning model, one or more velocity fields based at least in part on the one or more protein fold representations, and, beginning with the seed set, update values of one or both of (A) the position and / or orientation components and (B) the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving them a set of initial starting values into a set of final values representing (i) positions and / or orientations of each site of generated peptide backbone structures of the one or more variable regions, and / or (ii) likelihoods of side chain types at locations within the one or more variable regions; (d) generate, using the set of final values, one or both of: (A) a scaffold model representing a three-dimensional structure of a de-novo peptide backbone for one or more CDR regions of the custom antibody; and (B) sequence data representing an amino acid sequence of one or more CDR regions of the custom antibody; and (e) store and / or provide the generated (A) scaffold model and / or (B) sequence data, for display and / or further processing and / or use in designing the custom biologic. - 212 - 11901435v1Attorney Docket No.2013969-0048 68. A system for in-silico generation and / or prediction of three-dimensional (3D) side chain geometries of a polypeptide chain, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate (i) sequence data representing an amino acid sequence of at least a portion of the polypeptide chain and (ii) a representation of a 3D peptide backbone geometry, and / or folds thereof, for the portion of the polypeptide chain; (b) receive and / or generate a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular amino acid site of the portion of the polypeptide chain and comprising side chain geometry components representing a geometry of a side chain at the particular amino acid site; (c) determine, using a machine learning model, one or more velocity fields based at least in part on (i) the sequence data and (ii) the representation of the 3D peptide backbone geometry and / or folds thereof and, beginning with the seed set, update values of the side chain geometry components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving the values of the side chain geometry components of the plurality of feature vectors from a set of initial starting values into a set of final values representing geometries of each amino acid site of the portion of the polypeptide chain; and (c) create, using, the set of final values, a 3D polypeptide model of the portion of the polypeptide chain representing the 3D geometries of amino acid side chains within the portion of the polypeptide chain; and (d) store and / or provide the 3D polypeptide model.

69. A system for in-silico generation and / or prediction of an amino acid sequence of a polypeptide chain, the system comprising: a processor of a computing device; and - 213 - 11901435v1Attorney Docket No.2013969-0048 memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular site of the portion of the polypeptide chain and a side chain type component representing a likelihood of one or more possible types of amino acid side chains at the particular site; (b) determine, using a machine learning model, one or more velocity fields and, beginning with the seed set, update values of the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving the values of the side chain type components of the plurality of feature vectors from a set of initial starting values into a set of final values representing likelihoods of amino acid side chain types at each site of the portion of the polypeptide chain; and (c) determine, using, the set of final values, sequence data representing an amino acid sequence of the portion of the polypeptide chain; and (d) store and / or provide the sequence data.

70. The system of claim 69, wherein the instructions cause the processor to: receive and / or generate a representation of a 3D peptide backbone geometry, and / or folds thereof, for at least a portion of the polypeptide chain and at step (b), determine the one or more velocity fields based at least in part on the representation of the 3D peptide backbone geometry and / or folds thereof.

71. A system for generating in-silico docking predictions for a plurality (e.g., two or more) of polypeptide chains, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate a first scaffold model representation of a first 3D peptide backbone of at least a portion of a first polypeptide chain; - 214 - 11901435v1Attorney Docket No.2013969-0048 (b) receive and / or generate a second scaffold model representation of a second 3D peptide backbone of at least a portion of a second polypeptide chain; (c) receive and / or generate a seed set comprising a plurality of feature vectors, the plurality of feature vectors comprising (i) a first set of feature vectors corresponding to and representing positions and / or orientations of the first 3D peptide backbone and / or (ii) a second set of feature vectors corresponding to and representing positions and / or orientations of the second 3D peptide backbone; (d) determine, using a machine learning model, based at least in part on the first scaffold model and the second scaffold model, one or more velocity fields and, beginning with the seed set, update values of (i) the first set of feature vectors and / or (ii) the second set of feature vectors according to the one or more velocity fields, thereby evolving values of the first and / or second feature vector sets from a set of initial starting values into a set of final values representing final position(s) and / or orientation(s) of the first 3D peptide backbone and / or the second 3D peptide backbone docket with each other; (e) create, using, the set of final values, a generated polypeptide complex model representing first and second 3D peptide backbones docked to form a polypeptide complex; and (f) store and / or provide the generated polypeptide complex model 72. A system for designing a custom biologic for binding to a target, the system comprising: a processor of a computing device; and memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to: (a) receive and / or generate a target model representing at least a portion of the target; (b) receive and / or generate a template model representing a sequence and / or three-dimensional structure of at least a portion of a reference biologic, the template - 215 - 11901435v1Attorney Docket No.2013969-0048 model comprising a base portion representing a portion of the reference biologic located about one or more variable regions of the reference biologic designated as modifiable for binding to the target; (c) receive and / or generate a seed set comprising a plurality of feature vectors, each feature vector corresponding to a particular site within the one or more variable region(s) of template model and comprising one or both of (i) position and / or orientation components representing a position and / or orientation of the particular site and (ii) a side chain type component, representing likelihoods of one or more particular amino acid side chain types occupying the particular site; (d) determine, using a machine learning model, one or more velocity fields based at least in part on (i) the target model and (ii) the base portion of the template model and, beginning with the seed set, update values of one or both of (i) the position and / or orientation and (ii) the side chain type components of the plurality of feature vectors according to the one or more velocity fields, thereby evolving them a set of initial starting values into a set of final values representing (i) positions and / or orientations of each site of generated peptide backbone structures of the one or more variable regions, and / or (ii) likelihoods of side chain types at locations within the one or more variable regions; (e) create, using, the set of final values, one or both of: (i) a generated scaffold model representing a de-novo peptide backbone for the one or more variable regions of the custom biologic; and (ii) sequence data representing a generated amino acid sequence of the one or more variable region(s); and (f) store and / or provide, by the processor, the peptide backbone structure and / or the generated sequence for display and / or further processing and / or use in designing the custom biologic. - 216 - 11901435v1