Machine learning-based apparatus for engineering meso-scale peptides and methods and system for the same

Abstract

To provide improved devices and methods for computational design of proteins that mimic a target protein structure.SOLUTION: Provided herein are methods for design of engineered polypeptides that recapitulate molecular structure features of a predetermined portion of a reference protein structure, e.g., an antibody epitope or a protein binding site. A Machine Learning (ML) model is trained by labeling blueprint records generated from a reference target structure with scores calculated on the basis of computational protein modeling of polypeptide structures generated by the blueprint records. The method may include training the ML model on the basis of a first set of blueprint records, or representations thereof, and a first set of scores. Then, a set of engineered polypeptides are generated on the basis of the second set of blueprint records.SELECTED DRAWING: Figure 1

Description

CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application was filed on May 31, 2019 and is entitled "Meso-Scale Engine ered Peptides and Methods of Selection” This application claims priority to and the benefit of U.S. patent application Ser. No. 62 / 855,767, entitled "Compounds for a cellular endothelial cell that is a cellular organophosphate endothelium." , which is incorporated herein by reference in its entirety. [Technical field]

[0002] The present disclosure relates generally to the field of artificial intelligence / machine learning, and in particular to methods for engineering peptides. The present invention relates to a method and apparatus for training and using machine learning models. [Background technology]

[0003] Computational design is a method to mimic natural proteins. or to design new therapeutic proteins that target desired epitopes(s) from pathogenic antigens. ) can be used to design vaccines that exhibit For example, a library of antibodies (e.g., , a phage display library) is panned against the designed protein bait ( The clones that bind to the bait can be selected by panning, or the experimental animals can be , can be immunized with the engineered immunogen to generate novel antibodies.

[0004] Although there are others, the primary modeling platform for computational design is Rosetta (D (Assam and Baker, 2008). This platform is Correia et al. Structure ure 18:1116-26 (2010) for conformational stabilization and immune presentation A method for designing epitope-scaffolds by grafting continuous structural epitopes onto scaffold proteins. A general calculation method is disclosed in Olek et al. PNAS USA 107:1 7880-87 (2010) identified epitopes from the HIV-1 gp41 protein as It is disclosed that the selected recipient is grafted onto a scaffold.

[0005] Conventional computational design techniques typically involve the identification of a target protein structure (e.g., an epitope). It relies on transplanting parts onto existing scaffolds. The form is a large-scale model that can reproduce the vast topological space of proteins for a given protein structure. It is computationally too intensive to adequately explore a large topological space. New and improved devices for the computational design of proteins that mimic protein structures There is a need for a method. Summary of the Invention

[0006] In general, in some variations, an apparatus includes a processor that stores code representing instructions to be executed by the processor. The code may include a non-transitory processor-readable medium for storing the code. - Based on a first set of print records, or a representation thereof, and a first set of scores The first record of the blueprint may include a code for training a machine learning model based on the Each Blueprint record from this set is associated with each score from the first set of scores. After training, the medium executes the machine learning model to generate at least one desired The method may further include code for generating a second set of blueprint records having scores. The second set of Blueprint records is received as input in computational protein modeling. generating an engineered polypeptide based on the second set of blueprint records. The device may be configured to:

[0007] The medium may include code for causing a processor to receive a reference target structure. a processor for generating a first set of blueprint records from a predetermined portion of the reference target structure; Each blueprint from the first set of blueprint records may include code for generating The record includes a target residue position and a scaffold residue position, and each target residue from the set of target residues is The group position corresponds to one target residue from a set of target residues. In any one blueprint record, the target residue positions are non-contiguous. In at least one blueprint record, the target residue position is are in a different order than the order of the target residue positions of

[0008] The media performs computational protein modeling on each Blueprint recording to identify polypeptides. Generating structures, calculating scores for the polypeptide structures, and grouping the scores and associating the first section of the blueprint record with the first section of the blueprint record. In some variations, the computational protein model may include code for labeling the set of samples. The rings are also based on de novo design without matching a reference target structure to a template. In some variations, each score is derived from an energy term and a representation of the reference target structure. and a structural constraint match term, which may be determined using one or more structural constraints.

[0009] The medium may also include a second set of blueprint records for the processor. Code that determines whether to retrain a machine learning model by computing the In response to the determining, the medium may include: (1) a second (1) a retraining blueprint record including a second set of scores; and (2) a retraining blueprint record including a second set of scores. Further code may be included for retraining the machine learning model based on the training score. .

[0010] The medium is a processor, after retraining the machine learning model, blueprint recording Concatenate the first set of Blueprint records and the second set of Blueprint records to create a Blueprint record. and generating a retrain score. Each Blueprint record from the retraining of the blueprint record is associated with a score from the retraining score. In some variations, the at least one desired score may be a preset value. In some variations, the at least one desired score may be determined dynamically.

[0011] In some variations, the machine learning model may be a supervised machine learning model. The machine learning models include ensembles of decision trees, boosted decision tree algorithms, and eX Treme Gradient Boosting (XGBoost) model, or Random Forest In some variations, the supervised machine learning model may include a support vector machine ( SVM), feedforward machine learning models, recurrent neural networks (RNN) ), Convolutional Neural Networks (CNN), Graph Neural Networks (GN N), or a Transformer Neural Network.

[0012] In some variations, the machine learning model may include an inductive machine learning model. In some embodiments, the machine learning model may include a generative machine learning model.

[0013] The medium causes the processor to generate a computational protein model on the second set of blueprint records. The polypeptide may include code for performing the gene expression vector to produce an engineered polypeptide.

[0014] The medium is manipulated by the processor through a comparison of the static structure to a representation of a reference target structure. The method may include code for filtering the polypeptides that are detected.

[0015] The medium may also include a processor for processing a representation of each of the reference target structure and the engineered polypeptide. Ratio of dynamic structure to a representation of a reference target structure using molecular dynamics (MD) simulations The method may include code for filtering engineered polypeptides by comparison. For some deformations, MD simulations were performed using Symmetric Multiprocessing (SMP). These are carried out in parallel. [Brief description of the drawings]

[0016] [Figure 1] FIG. 1 is a schematic diagram of an exemplary engineered polypeptide design device. [Diagram 2] FIG. 1 is a schematic diagram of an exemplary machine learning model for engineered polypeptide design. [Diagram 3] FIG. 1 is a schematic diagram of an exemplary method for engineered polypeptide design. [Figure 4] FIG. 1 is a schematic diagram of an exemplary method for engineered polypeptide design. [Diagram 5] FIG. 1 is a schematic diagram of an exemplary method for preparing data for an engineered polypeptide design device. [Figure 6] FIG. 1 is a schematic diagram of an exemplary method for engineered polypeptide design. [Figure 7] FIG. 1 is a schematic diagram of an exemplary performance of a machine learning model for engineered polypeptide design. [Figure 8] FIG. 1 is a schematic diagram of an exemplary method of using a machine learning model for engineered polypeptide design. [Figure 9] FIG. 1 is a schematic diagram of an exemplary performance of a machine learning model for engineered polypeptide design. [Figure 10A] 1 shows an exemplary method for performing molecular dynamics simulations to validate engineered polypeptides. [Figure 10B] 1 shows an exemplary method for performing molecular dynamics simulations to validate engineered polypeptides. [Figure 10C] 1 shows an exemplary method for performing molecular dynamics simulations to validate engineered polypeptides. [Figure 10D] 1 shows an exemplary method for performing molecular dynamics simulations to validate engineered polypeptides. [Figure 11] 1 shows an exemplary method for performing molecular dynamics simulations to validate engineered polypeptides. [Figure 12] FIG. 1 is a schematic diagram of an exemplary method for parallelizing a molecular dynamics simulation. [Figure 13] FIG. 1 is a schematic diagram of an exemplary method for validating a machine learning model for engineered polypeptide design. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0017] Non-limiting examples of various aspects and variations of the present invention are described herein and illustrated in the accompanying drawings.

[0018] Described herein are methods for designing engineered polypeptides and methods for producing engineered polypeptides. Compositions containing the engineered peptides and methods of using the engineered peptides are provided. Provided herein are methods for using engineered peptides in the in vitro selection of antibodies. In some embodiments, a user (or a program) may create a Select a target protein that is suitable for the synthesis and use the target protein as input for the design of the engineered polypeptide. The target protein may be an antigen (or a promoter) from a pathogenic organism. constant antigens), proteins, enzymes, and signaling molecules involved in cellular functions associated with disease Any protein for which an engineered polypeptide that reproduces a portion of a protein is desired. The engineered polypeptides may be useful in antibody discovery, vaccination, diagnostics, The target protein may be intended for use in methods of therapy, biomanufacturing, or other applications. The "protein" may be two or more proteins, such as a multimeric protein complex, in a transformation. For simplicity, the present disclosure refers to target proteins, but the methods also apply to multimeric structures. In a variant, the target protein is two or more distinct proteins or proteins. For example, the methods disclosed herein can be used to prepare a storage complex for, e.g., antibody selection. Mimicking common attributes of proteins from diverse species to target defined epitopes It can be used to design engineered peptides.

[0019] A computational record of the protein's topology is derived, referred to herein as the "reference target structure." A reference target structure may be, for example, a triad of all (or most) atoms in a protein. D coordinates, or the 3D coordinates of selected atoms (e.g., the coordinates of the Cβ atoms of each protein residue) ) or a structural model of a conventional protein structure represented by The reference target structure may be determined computationally (e.g., from molecular dynamics simulations) or experimentally. It may also include dynamical terms derived (eg, from spectroscopy, crystallography, or electron microscopy).

[0020] A predetermined portion of the target protein is a blueprint having target and scaffold residue positions. Each position is converted to a fixed amino acid residue identity or a variable identity (e.g., of amino acids, or have desired physicochemical properties - polar / non-polar, hydrophobic, size, etc. In a variant, a given portion of the target protein may be assigned to either Each amino acid derived from the target protein is sequenced so that it has the same amino acid identity as it does in the target protein. The target residue position is mapped to a single target-residue position that is assigned to the corresponding It may be sequential and / or sequenced. However, some variations are possible. In this form, the advantage is that the target-residue positions are non-contiguous (interrupted by scaffold-residue positions). The problem is that the target protein cannot be ordered (in a different order than the target protein). Unlike the approach, in some variations, the order of residues is not constrained. , discontinuous parts of the target protein (e.g., different parts of the same protein or different proteins) Even one protein chain can contribute to one epitope (discontinuous epitopes) .

[0021] A scaffold-residue position in the blueprint can be assigned to have any amino acid at that position. In a variant, the scaffold-residue position may be: A subset of possible natural or unnatural amino acids (e.g., small polar amino acids The amino acid residues are assigned by selecting from the group consisting of large hydrophobic amino acid residues, large hydrophobic amino acid residues, etc. Lint may also be applied to any target residue position and / or scaffold residue position. Additionally, the blueprint may allow for the insertion or deletion of a residue position. For example, a target residue position Alternatively, scaffold residue positions may be assigned as present or absent, or their positions may be The position can be 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 or more residues. may be assigned.

[0022] A subset of the blueprints can then be used to perform computational modeling, e.g. , using the scores calculated for each polypeptide structure derived from the reference target structure. Generate the corresponding polypeptide structure using energy terms and topological constraints Machine learning (ML) models can be used to generate scores and blueprints, or blueprints. It may be trained using a representation of the print (e.g., a vector representing the blueprint). The ML model may then be run to generate further blueprints. The points are much more complex than can be explored by iterative modeling of many blueprints. The topological space covered by many blueprints is explored by an ML model. It is possible to search for it.

[0023] The present disclosure relates to an output blueprint for an engineered polypeptide sequence and / or structure. These engineered polypeptides are then compared using static comparison, dynamic comparison, or both. and use these comparisons to filter polypeptides. Further provided are methods and associated devices for monitoring.

[0024] The method and apparatus includes a set of blueprint records, a set of scores, a set of energy terms, set of molecular dynamics energies, a set of energy terms, or an energy function Although described herein as processing data from a set, in some instances , a polypeptide design device 10 operated as shown and described with respect to FIG. 1 to generate a set of blueprint records, a set of scores, a set of energy terms, A set of molecular dynamics energies, or a set of energy functions, can be generated. Thus, the engineered polypeptide design device 101 can store data, events, and may be used to generate or process any collection or stream of data and / or objects. For example, the engineered polypeptide design device 101 can be any character string, number, name, Images, videos, executables, datasets, spreadsheets, data files, May process and / or generate blueprint files, and / or the like. For further example, the engineered polypeptide design device 101 may be any software Software code, web pages, data files, model files, source files, scripts, As another example, the system may process and / or generate scripts, and / or the like. The engineered polypeptide design device 101 receives a data stream, an image data stream, data streams, text data streams, numeric data streams, computer-aided design (CA D) may process and / or generate file streams, and / or the like.

[0025] FIG. 1 is a schematic diagram of an exemplary engineered polypeptide design device 101. The polypeptide design device is adapted to generate a set of engineered polypeptide designs. The engineered polypeptide design device 101 includes a memory 102, a communication interface 103, and a The engineered polypeptide design data includes an interface 103, and a processor 104. The device 101 communicates with the back-end service platform 102 via the network 150. 160, optionally connected (without intervening components) or (without intervening components) The polypeptides may be linked together (with or without intervening components). The measurement device 101 may be, for example, a desktop computer, a server computer, a Frame computers, quantum computing devices, parallel computing devices devices, desktop computers, laptop computers, smartphone devices hardware-based computing, such as an ensemble of The present invention may also be applied to a communication device.

[0026] The memory 102 of the engineered polypeptide design device 101 may, for example, include a memory buffer. Random Access Memory (RAM), Read Only Memory (ROM), Erasable Program Programmable Read-Only Memory (EPROM), Embedded Multi-Time Programmable ( MTP) memory, embedded multimedia card (eMMC), universal flash The memory 10 may include a USB storage device, a USB 2.0 storage device, a USB 3.0 ... 2, for example, to the processor 104 of the engineered polypeptide design device 101 The above processes or functions (e.g., data preparation module 105, computational protein modeling Ring module 106, machine learning model 107, and / or molecular dynamics simulation one or more software modules including instructions for implementing the and / or code. The memory 102 may store engineered polypeptide designs. The machine learning model 107 may include data generated by the machine learning model 107 during operation of the device 101. A file associated with (e.g., generated by running) the learning model 107 In some instances, a set of rules associated with the machine learning model 107 may be stored. The file set is generated during operation of the engineered polypeptide design device 101. Temporary variables, return memory addresses, variables, graphs of machine learning models 107 (e.g., machine learning A set of arithmetic operations, or a representation of a set of arithmetic operations, used by the learning model 107 , graph metadata, assets (e.g. external files), digital signatures (e.g. exports) (type of machine learning model 107 to be supported, and input / output tensors) and and / or the like.

[0027] The communication interface 103 of the engineered polypeptide design device 101 is operatively coupled to and used by the processor 104 and / or memory 102. The hardware components of the engineered polypeptide design device 101 may be The communication interface 103 may be, for example, a network interface card ( NIC), Wi-Fi TM Module, Bluetooth (registered trademark) module, optical communication communication module, and / or any other suitable wired and / or wireless communication interface. The communication interface 103 may include a The engineered polypeptide design device 101 is connected to a network 150 as described in In some instances, the communication interface 103 may be configured to connect to The network 150 may facilitate receiving or transmitting data. In some implementations, the communication interface 103 may be, for example, set, set of scores, set of energy terms, set of molecular dynamics energies, A set of energy terms, or a set of energy functions, is added to the network. Received or transmitted from or to a back-end service platform 160 through In some instances, the communication interface 103 may facilitate the transmission of The transmitted data is processed by the processor 104 as described in more detail herein. It may be processed or stored in memory 102 .

[0028] The processor 104 may be, for example, a hardware-based integrated circuit (IC) or or any other suitable software configured to implement and / or execute a set of code. For example, the processor 104 may be a general purpose processor, a central processor, Processing Unit (CPU), Graphical Processing Unit (GPU), Tensor Processing Unit (TPU), Accelerated Processing Unit (APU), Application Specific Application-specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), Programmable Logic Arrays (PLA), Complex Programmable Logic Devices (CPLD), Programmable The processor may be a programmable logic controller (PLC), and / or the like. The processor 104 may include a system bus (e.g., an address bus, a data bus, and / or a control bus). The controller 102 is operatively coupled to the memory 102 via a bus.

[0029] The processor 104 includes a data preparation module 105, a computational protein modeling module 106, and The processor 104 may optionally include a machine learning model 107. The data preparation module 108 may include a molecular dynamics simulation module 108. 05, Computational Protein Modeling Module 106, Machine Learning Models 107, or Molecular Each of the dynamics simulation modules 108 is stored in the memory 102 and For example, the machine learning model 10 7, code for generating a set of blueprint records is stored in memory 102, Similarly, the data preparation module 105, the calculation Protein modeling module 106, machine learning model 107, or molecular dynamics simulation Each of the simulation modules 108 may be a hardware-based device. For example, a process for having a machine learning model 107 generate a set of blueprint records. may be implemented on individual integrated circuit (IC) chips.

[0030] The data preparation module 105 includes receiving a reference target structure for the reference target. , the set of data is stored (e.g., in memory 102 or in a back-end service platform) The data preparation module 105 may be configured to receive the reference A set of blueprint records (e.g., a table of alphanumeric data) is generated from a given portion of the target structure. The method may be further configured to generate a blueprint file (a blueprint file encoded in a In some instances, each blueprint record from the set of blueprint records may include a target residue. Each target residue position may include a group position and a scaffold residue position, and each target residue position may include a group position and a scaffold residue position from a set of target residues. Corresponding to one target residue.

[0031] In some instances, the data preparation module 105 creates a blueprint of the reference target structure. The data preparation module 1 may be further configured to encode the data into a loop print record. 05 provides a representation of blueprint records that are generally suitable for use in machine learning models, called blueprints. The print record may be further transformed. In some instances, the representation is a one-dimensional vector of numerical values. , a 2D matrix of alphanumeric data, or a 3D tensor of normalized numbers. Specifically, in some instances, the representations are ordered lists of the number of intervening scaffold residue positions. Such a representation is useful because the target-residue order can be inferred from the target structure. The present invention may be used, so that the representation specifies the amino acid identity of the target-residue position. An example of such a display is further described with respect to FIG.

[0032] In some instances, the data preparation module 105 may generate a set of blueprint records, set of energy terms, set of molecular dynamics energies, set of energy terms A set of data, and / or an energy function may be generated and / or processed. The preparation module 105 includes a set of blueprint records, a set of scores, and an energy term. set of molecular dynamics energies, set of energy terms, or energy functions It may be configured to extract information from a set of numbers.

[0033] In some instances, the data preparation module 105 may include a code for a set of blueprint records. For example, ASCII, UTF-8, UTF-16, Guobiao, Big5, U to have a common character encoding such as Unicode, or any other suitable character encoding In yet other instances, the data preparation module 105 may convert, for example, For example, a portion of a Blueprint record or a Blueprint record that is important for the operation of a polypeptide. By identifying the indications of the Blueprint Record Characteristics and / or Blueprint In some instances, the data preparation model may be further configured to extract a representation of the data record. Joule 105 is a set of blueprint records, a set of scores, a set of energy terms. a set of molecular dynamics energies, a set of energy terms, or, for example, A set of energy functions from English units such as feet, inches, and / or similar. units of measurement, e.g. kilometers, meters, centimeters and / or similar. It may also be converted to the International System of Units (SI), such as:

[0034] The computational protein modeling module 106 generates a protein sequence of interest from a predetermined portion of the reference target structure. A blueprint that can serve as a starting template for the computational optimization process described in the specification. In one embodiment, the computational task may be configured to generate an initial set of candidate task records. The protein modeling module 106 can be a Rosetta modeler. Variations of the method include molecular dynamics simulations, ab initio fragment assembly, and te Carlo fragment assembly, AlphaFold or trRosetta etc. Machine learning structure prediction, protein folding supported by structural knowledge base, Neural network protein folding, sequence-based recursive or trans transformer network protein folding, antagonistic network protein Protein structure generation, Markov Chain Monte Carlo protein fall Other modeling apps, including but not limited to dummy models, dummy models, and / or the like. The initial candidate structures generated using the Rosetta Modeler are It may be used as a training set for a machine learning model. Joule 106 is the energy calculation for each blueprint from the initial candidates of the blueprint record. The Gee term can then be further determined computationally. Next, the data preparation module 105 , may be configured to generate a score from the energy terms. In one embodiment, the score is , can be the normalized value of the energy term. The normalized value is a number between 0 and 1. , a number between -1 and -1, a normalized value between 0 and 100, or any other numeric range. In some variations, the computational protein modeling module 106 can Without matching the structure to a template or with weak distance constraints nce restraints), or based on de novo design, where, for example, For example, the distance between target residues is within 1 angstrom of the target-residue distance in the target structure. Weak distance constraints are constrained to be: Restrictions that allow for Gaussian noise with a particular variance around a particular mean and distance limit In some variations, the computational protein modeling module 106 may include any of the following: and / or by smoothing or adding fluctuating noise to the distance limits of The computational protein model is designed so that it is not penalized severely when the condition is not satisfied. In some instances, the objective function of the quality model may be used. The computational protein modeling module 106 generates smoothed labels for the energy terms. The advantage of this method is that it smooths the energy term labels. In this way, the machine learning model 107 can predict the topology covered by the blueprint to be explored. The advantage of this approach is that it makes it easier to optimize the cull space.

[0035] The machine learning model 107 has been improved compared to the initial set of candidates for blueprint records. The machine learning model 107 may be used to generate a calculated blueprint record. The initial blueprint record calculated by the protein modeling module 106 A supervised machine learning model configured to receive a set of candidates and a set of scores. Each score from the set of scores can be an initial candidate for a blueprint record. The processor 104 then processes each corresponding score into a corresponding blueprint record from the corresponding set. associate the dataset and blueprint records to generate a set of labeled training data The method may be configured to perform the above steps.

[0036] In some instances, the machine learning model 107 may be an inductive machine learning model and / or a generative The machine learning model may include a boosted decision tree algorithm. , ensembles of decision trees, and eXtreme Gradient Boosting (XGBoost) models , Random Forest, Support Vector Machine (SVM), Feedforward Mechanics Learning Models, Recurrent Neural Networks (RNNs), and Convolutional Neural Networks (CNN), Graph Neural Network (GNN), Adversarial Network Model, instance-based training models, transformer neural networks, and / or or the like. Once trained, the machine learning model 107 may be It can generate scores from Blueprint recordings, or run in Generate mode A set of weights that can be run to generate a Blueprint record from the scores, Model parameters, including a set of biases and / or a set of activation functions It may be configured to include a set of meters.

[0037] In one embodiment, the machine learning model 107 includes an input layer, an output layer, and multiple hidden layers (e.g., Deep learning models, including those with 5 layers, 10 layers, 20 layers, 50 layers, 100 layers, 200 layers, etc. The multiple hidden layers are a normalization layer, a fully connected layer, an activation layer, and a convolution layer. Between a set of convolutional layers, recurrent layers, and / or blueprint records and a set of scores , and each score may include an energy term represent.

[0038] In one embodiment, the machine learning model 107 may be, for example, a boosted model in the XGBoost model. The number of rounds or trees, the maximum allowable number of rounds from the tree root to the leaves of the XGBoost model. A set of hyperparameters, such as the maximum depth defining the number of threads, and / or the like An XGBoost model can be an XGBoost model that includes a set of trees. We take a set of nodes, a set of weights, a set of biases, and an XGBoost model. Other parameters that are useful for explanation may also be included.

[0039] In some implementations, machine learning models107 (e.g., deep learning models, XGBoost models) Dell, Inc., and / or the like) will create a Blueprint for each Blueprint from the set of Blueprint Records. The blueprint record may be configured to repeatedly receive the blueprint record and generate an output. Each Blueprint record from the set is associated with one score from the set of scores. The outputs and scores are then multiplied by the objective function (the cost function) to generate the first training loss value. The objective function may be, for example, the mean square error, the mean Mean absolute error, mean absolute error rate, logcosh, categorical cross-entry The set of model parameters may include multiple The first objective function may be modified in each iteration such that the first training loss value is greater than or equal to a first predetermined training threshold. is performed in each iteration until it converges to a value (e.g., 80%, 85%, 90%, 97%, etc.). This is also fine.

[0040] In some implementations, the machine learning model107 iteratively receives each score from a set of scores. Each blueprint from a set of blueprint records may be configured to receive and generate an output. A loop print record is associated with one score from a set of scores. The blueprint records are compared using the objective function to generate a second training loss value. The set of model parameters may be varied in multiple iterations to achieve the first objective. The function performs each iteration of the multiple iterations until the second training loss value converges to a second predetermined training threshold. may be performed.

[0041] Once trained, the machine learning model107 is run to improve Blueprint record security. A set of improved blueprint records can be generated for the blueprint In some instances, the candidate set of records may be expected to have a higher score than the initial candidate set of records. The machine learning model 107 is the first to rank the design space and score for the first set of blueprint records. The first set of scores is used to represent the correlation with the set of set (e.g., Rosetta Energy of a Blueprint record from a set of Blueprint records) The first of the blueprint records corresponding to each score (each score having an energy term corresponding to the energy A generative machine trained on a set (e.g., generated using the Rosetta Remodeler) Once trained, the machine learning model 107 can Generate a second set of blueprint records having a second set of associated scores. In some implementations, a computational protein modeling module 106 is used to By calculating a set of energy terms for the second set of print records, A second set of loop print records and a second set of scores can be verified. The set of energy terms is the ground truth for the second set of blueprint records. A subset of blueprint records may be used to generate a set of score scores. Each Blueprint record from the subset of Blueprint records is a grand total that exceeds the threshold. A second set of blueprint records may be selected from the second set of blueprint records to have a truth score. In some instances, the threshold value may be, for example, the user of the engineered polypeptide design device 101. In some other instances, the threshold may be a number predetermined by the user. It may be a number that is dynamically determined based on a set of truth scores.

[0042] Optionally, the machine learning model 107 is executed to generate a second set of blueprint records. After the molecular dynamics simulation module 108 is used to simulate the machine learning model 1 The output of the polypeptide design device 101 can be verified. generating an engineered polypeptide based on a second set of loopprint records; Molecular dynamics (MD) simulations of the structure and each representation of the engineered polypeptide structure By performing a comparison of the dynamic structure against a representation of the reference target structure using A subset of the second blueprint records may be filtered. For example, The biological simulation module 108 may generate a sequence of several (e.g., 10 A second set of Blueprint records may be used to select a number of Blueprint records that are less than a hit. In some instances, MD simulations are performed using boundary conditions, constraints, and / or In some instances, MD simulations may be performed at the model preparation stage. Steps for equilibration (e.g., temperature 100K to 300K), force field parameters, and and / or solvent model parameters for each of the reference target structure and the engineered polypeptide structure. and may be carried out under solution conditions. D simulations can be performed using constrained minimization (e.g., to mitigate structural clashes), constrained controlled heating (e.g., controlled heating for 100 picoseconds and a stepwise increase to ambient temperature), Relaxed constraints (e.g., relaxing the constraints for 100 picoseconds and gradually removing the skeletal constraints) ), and / or the like.

[0043] In some implementations, the machine learning model 107 is an inductive machine learning model. Such machine learning models107 are used, for example, to calculate the blueprint score. Numerical methods (e.g., computational protein modeling modules, density functional theory-based molecular dynamics) Dynamic Energy Simulator, and / or similar) to create blueprints Record-based scores can be predicted in a fraction of the time it would normally take. Use machine learning models107 to quickly generate scores for a set of blueprint records Estimate the optimization speed of the optimization algorithm (e.g., 50% faster, 2x faster, 10x faster). , 100x faster, 1000x faster, 1,000,000x faster, 1,000,000,0 00 times faster, and / or the like. The machine learning model 107 may generate a first set of scores for a first set of blueprint records. The processor 104 of the engineered polypeptide design device 101 can generate a set of executes the code that represents the set of instructions and returns the top 10 scores (for example, the first set of scores) % of the first set of scores, e.g., in the top 2% of the first set of scores, and / or the like. The top performers of the first set of Blueprint records may be selected. The processor 104 may select the top performers from the first set of blueprint records. In some variations, the method may further include code for verifying the score of the corresponding verified score. If the score obtained has a value greater than any of the first set of scores, the blueprint is The top performers in the first set of event records can be generated as output. In some variations, the machine learning model 107 may include a second set of blueprint records, and A second set of scores including the Blueprint records and the scores of the top performers The model can be retrained based on new datasets, including

[0044] The network 150 includes digital servers and / or computing devices. The network may be a single-server communications network. A computing device is, for example, a device that provides data storage or computing power. One or more wired or wireless communication networks (not shown) may be used to share resources such as The network may be connected via a server and / or computer A wired or wireless communication network between operating devices may include one or more communication channels (e.g. , radio frequency (RF) communication channels, fiber optic communication channels, and / or the like. The network may include, for example, the Internet, an intranet, Local Area Network (LAN), Wide Area Network (WAN), Metro Return Area Network (MAN), a global network for microwave access networks Interoperability (WiMAX), virtual networks, any other suitable communication The network may be a signaling system, a signaling system, and / or a combination of such networks.

[0045] The back-end service platform 160 is, for example, a service such as the Internet. to a digital communications network of bars and / or computing devices; and and / or a computing device operably coupled in a digital communications network. In some variations, the back-end service platform may be Platform 160 is, for example, software as a service (SaaS), service and Platform as a Service (PaaS), Infrastructure as a Service (Ia may include cloud-based services such as cloud-based services (e.g., ... In one embodiment, the backend service platform 1 60 protein structure, blueprint record, rosette energy, molecular dynamics energy Data storage for storing large amounts of data, including data, graphics, and / or the like In another embodiment, the backend service platform 16 0 is used for computational protein modeling, molecular dynamics simulations, training machine learning models, and / or provide high speed computing for executing a set of the same It is possible.

[0046] In some variations, the procedures of the computational protein module 106 described herein are performed using a cloud Back-end service platform that provides computing services160 In such a variation, the engineered polypeptide design device 101 may be , using the communication interface 103 to transmit the signal to the back-end service platform. The blueprint record set may be configured to be sent to system 160 to generate a set of blueprint records. The backend service platform 160 generates a set of blueprint records. Then, the back-end server can execute a computational protein modeling process. The service platform 160 communicates with the Blueprint record via the network 150. The set can be sent to an engineered polypeptide design device 101 .

[0047] In some variations, the engineered polypeptide design device 101 comprises a machine learning model 10 7 to a remote user computer from the operated polypeptide design device 101. The user's computing device (not shown) may then transmit the information to the user's computing device. The design devices are then evaluated to determine whether they meet the design criteria (e.g., have a desired score) and whether they meet the blueprint criteria. In some variations, the user computing device may be configured to generate a set of records. The design device receives a reference target structure from the engineered polypeptide design device 101. The user computing device then determines whether each blueprint record corresponds to a target residue position and A first subdivision of the blueprint record from a predetermined portion of the reference target structure to include the scaffold residue positions. Each target residue position may be matched to one target residue from the set of target residues. The user computing device receives a first set of blueprint records, Further train the machine learning model based on the representation and the first set of scores. The user computing device can then run the machine learning model after training. The blocks are then selected to have at least one desired score (e.g., meet certain design criteria). A second set of blueprint records may be generated. is received as input in computational protein modeling and is added to the second section of the Blueprint record. Based on this set, engineered polypeptides may be generated.

[0048] FIG. 2 illustrates an exemplary machine learning model 202 for engineered polypeptide design (FIG. 1). 1 is a schematic diagram of a machine learning model 107 as described and illustrated in The machine learning model 202 analyzes the design space of blueprint records by The scores are then correlated with the energy terms of the polypeptides constructed based on the recorded points. The machine learning model may be a supervised machine learning model that operates in a generate mode and and / or may have a recursive mode of operation.

[0049] In a generate mode of operation, the machine learning model 202 generates a first set of blueprint records 2 01 and the first set of scores 203. Once trained, machine learning model 2 02 is a set of scores that are statistically higher (e.g., have a higher average) than the first set of scores. A second set of blueprint records is generated having a second set of inductive behavior models. In the model, the machine learning model 202 also generates a first set of blueprint records 201 and The machine learning model 202 is trained on the first set of classifications and scores 203. Once trained, the machine learning model 202 Generate a second set of scores for the second set of blueprint records. The second set is the historical training data (e.g., the first set of blueprint records and the The first set of cores) are a set of prediction scores based on computational protein modeling ( Computational protein modeling module 106 as shown and described with respect to FIG. ) or molecular dynamics simulations (as shown and described with respect to FIG. 1 ). Numerically calculated scores and and / or significantly faster than the energy term (e.g., 50% faster, 2x faster, 10x faster). Faster, 100x faster, 1000x faster, 1,000,000x faster, 1,000,000, 000 times faster, and / or the like) are generated.

[0050] FIG. 3 is a schematic diagram of an exemplary method of engineered polypeptide design 300. The method of polypeptide design 300 can be implemented, for example, by using an engineered polypeptide design device (FIG. 1. An engineered polypeptide design device 101 as shown and described with respect to The engineered polypeptide design 300 method may include Optionally, step 301 includes receiving a reference target structure for the reference target. The method of engineered polypeptide design 300 optionally includes, in step 302, generating a first set of blueprint records from a predetermined portion of the target structure; Each blueprint record from the first set of blueprint records is Each target residue position corresponds to one target residue from a set of target residues. In some instances, the target residues are non-contiguous. The method of engineered polypeptide design 300 includes, in step 303, Based on a first set of print records, or a representation thereof, and a first set of scores , a machine learning model (such as the machine learning model 107 shown and described with respect to FIG. 1 ); and training a first set of blueprint records, each of which is similar to a first set of blueprint records. Each blueprint record is associated with a respective score from the first set of scores. The representation is processed by a data preparation module (data preparation as shown and described with respect to FIG. 1 ). module) based on the first set of blueprint records. The method of engineered polypeptide design 300 may include, in step 304, training Subsequently, the machine learning model is run to generate at least one desired score (e.g., one score and generating a second set of blueprint records having a score or scores. In some configurations, the machine learning model includes a generative machine learning model, including at least One desired score is determined by the user of the engineered polypeptide design device. In some configurations, the machine learning model uses the second The group includes an inductive machine learning model that predicts a set of prediction scores for a set of A subset of the second set of subsets of lint records is a subset of the blueprint records. Each Blueprint record from the team will have at least one score greater than the desired score. In some configurations, the at least one desired score may be dynamically selected. For example, at least one desired score may be determined based on 9 of the set of predicted scores. It may be determined to be the 0th percentile.

[0051] The method of engineered polypeptide design 300 can optionally include, at 305, e.g., a method for designing a polypeptide comprising: Modeller, ab initio molecular dynamics simulation, AlphaFold or The structure prediction is based on machine learning such as trRosetta, and the structure knowledge base is used to Protein folding, neural network protein folding, sequence-based Recurrent or transformer network protein folding, adversarial Generation of network protein structures, Markov Chain Monte Carlo lo protein folding, and / or the like, By computing a second set of cores (e.g., ground truth scores), This involves deciding whether to retrain the machine learning model. The chip design device compares the second set of scores to the set of predicted scores and Whether to retrain the machine learning model based on the deviation of the predicted scores from the second set The method of engineered polypeptide design 300 optionally includes, at 305, determining In response to this, (1) a retraining blueprint including a second set of blueprint records; (2) a machine learning model based on the retrained scores, the retrained scores including a set of predicted scores. In some configurations, the engineered polypeptide design device includes retraining the Link the first set of blueprint records with the second set of blueprint records to A trained blueprint record can be generated. The engineered polypeptide design device can Concatenate the first set of scores with the second set of scores to generate a further retraining score. In some configurations, retraining a Blueprint record requires a second session of the Blueprint record. The first set of scores contains only the first set of scores, and the retraining scores contains only the second set of scores.

[0052] FIG. 4 is a schematic diagram of an exemplary method of engineered polypeptide design 400. The method of polypeptide design 400 can be implemented, for example, by using an engineered polypeptide design device (FIG. 1. An engineered polypeptide design device 101 as shown and described with respect to The engineered polypeptide design 400 method may include In step 401, a first set of blueprint records, or a representation thereof, and a score are generated. Based on the first set of data, a machine learning model (as shown and described with respect to FIG. 1) is generated. This includes training a machine learning model (such as Each Blueprint record from the first set will have a score for each score from the first set. The representation is generated by the data preparation module (see FIG. 1 and described above). (similar to the Data Preparation module mentioned above) to the first step of the Blueprint record. The engineered polypeptide design 400 method may be generated based on a set of steps. At step 402, after training, the machine learning model is run to generate at least one desired score. The method further includes generating a second set of blueprint records having the engineered polynomials. The method of peptide design 400 optionally includes, in step 403, adding a second Performing computational protein modeling on the set to generate engineered polypeptides. In some configurations, the method of engineered polypeptide design 400 optionally includes steps In step 404, the engineered polypeptide is generated by comparing the static structure against a representation of the reference target structure. In some configurations, engineered polypeptide designs include filtering the polypeptides. The method of 00 optionally comprises, in step 405, determining whether the reference target structure and the engineered polypeptide are identical. A table of reference target structures using molecular dynamics (MD) simulations of each representation of the structure of Filtering engineered polypeptides by comparing dynamic structure to the current Includes.

[0053] FIG. 5 is a schematic diagram of an exemplary method for preparing data for an engineered polypeptide design device. On the left is a ribbon diagram of the target protein structure. The specified parts are shown in stick form. The amino acid residues are shown in darker colors with the side chains of the amino acid residues of the selected regions shown as diagrams. In this example, the predetermined portion is the desired target epitope of the target protein. To generate a polypeptide engineered to reproduce this epitope, This allows the production of antibodies that specifically bind to this portion of the target protein. It is expected.

[0054] The right panel of Figure 5 shows a diagram of a set of blueprints. Each circle represents a residue position. Scaffold-residue positions are in light grey and side chains are not shown. Target-residue positions are in darker grey. The side chains of each are shown in grey. The side chains are those of known naturally occurring amino acids. In some instances, the target-residue and / or the scaffold-residue are unnatural amino acids. In this example, each target-residue position is a positive integer of a given portion of the reference target structure of the target protein. The set of blueprints shown corresponds to exactly one residue in all figures. The target residues are "ordered" in that they are in the same order. The residues do not have to be in the same order as in the target protein sequence. Some blueprints have continuous target-residue positions, whereas other blueprints are discontinuous. At least one scaffold-residue position is located between the first and last target-residue positions. The letters N and C denote the amino (N) and amino-terminal ends of a polypeptide that matches a given blueprint. and the carboxyl (C) terminus.

[0055] The five blueprints shown in Figure 5 are a set of possible It is a member of a large set of blueprints that have 35 positions. For the 35-mer polypeptide, we assume that the target residues are ordered. Then, the potential blue The total number of prints is given. Even with the use of the Rosetta Modeler, calculations of all possible 35-mers have taken years. Therefore, it is necessary to spend time on direct computational modeling of each blueprint. are not individually computable using current computing devices and methods. .

[0056] FIG. 6 is a schematic diagram of an exemplary method for engineered polypeptide design. The part is how to record a Blueprint (e.g., a suitable Blueprint for use as input). (not shown) may be used in conjunction with other programs (including but not limited to Rosetta Remodeler) 1. The computational protein modeling program 106 shown and described with respect to FIG. (similar to The score generally indicates how well the model can generate a score for the Reflects the energy terms used by the program. In the case of the Rosetta Modeler, this The score reflects the folding of the designed polypeptide generated from the blueprint. The energy terms that correspond to the predicted structure of the designed polypeptide and the structural similarity to the target protein are compared. and a structure constraint match term that reflects the known structure of a given portion of a reference target structure of the protein. Other modeling programs and other scoring functions may be used.

[0057] The left part of the schematic shows how the Blueprint is converted into a Blueprint representation. The representation is generated based on a machine learning model (such as the machine learning model shown and described with respect to FIG. 1). The representation may be any representation suitable for use in a learning model (e.g., a learning model 107), where the representation is a vector. More specifically, the vector is a function of the number of intervening scaffold residues between the target and residue positions. is an ordered list. The order of target-residue positions is fixed in this representation, so This representation may be used, and therefore the representation specifies the amino acid identity of the target residue position. It is not necessary to specify the sequence of the target residues; that information is implied. The order of the target-residue positions is determined by the The first element of the vector, 8, is the first target-residue position. The second element of the vector, 1, indicates that there are eight scaffold-residue positions before the first position. One target-residue position is followed by one scaffold-residue position before the second target-residue position. A subsequent element of 0, 1, 2, or 3 indicates that there are no intervening scaffold-residue positions; there is one intervening scaffold-residue position; there are two intervening scaffold-residue positions; or 3 intervening scaffold-residue positions. The last element of the vector, 4 indicates that the last four positions in the blueprint are scaffold-residue positions.

[0058] The advantage of this variation of the representation of a Blueprint record is that, apart from the first and last elements, the vector Frameshift invariant. That is, the machine learning model can be has available information regarding the relative positions of target residues, regardless of the position of the target residue in This allows the identification of similarly structured proteins with variable structured / unstructured regions at the N- and C-termini. Design becomes possible.

[0059] FIG. 7 is a schematic diagram of an exemplary performance of a machine learning model for engineered polypeptide design. The scatter plot shows how the machine learning model (as shown and explained in Figure 1) A machine learning model (such as 107) predicts a score for a set of blueprint records. Each dot in the scatter plot represents a Blueprint record's The horizontal axis represents the blueprint records from the set, e.g., Rosetta Remodeler, Ab Numerical methods such as initio molecular dynamics simulations and / or the like represents the ground truth score of a set of blueprint records, which can be calculated by The axis can be substantially faster than the numerical method (e.g., 50% faster, 2 times faster, 10 times faster, 100 times faster, etc.). x faster, 1000x faster, 1,000,000x faster, 1,000,000,000x higher The business model is generated / predicted by machine learning models that operate (fast, fast, and / or the like). represents the predicted score for a set of loop print records. Ideally, the predicted score is , corresponds to (e.g., equals, is close to) the ground truth score. If the land truth score does not correspond, the machine learning model will not update the new The newly generated prediction scores for the newly generated set are added to the newly generated The set of blueprint records is updated until it corresponds to the ground truth score of the set created. The scores may be retrained with the previous and ground truth scores. For example, the energy terms in the Rosetta Energy Function 2015 (REF15) and Fig. 6 The score may include both structural constraint match terms as described with respect to FIG. As shown in Figure 1, the low scores of the Blueprint records are due to the low molecular dynamics energy and the blueprints. It may be defined to reflect a greater stability of the loop print record. The score is based on the Blueprint record's high score. It may be defined to generally reflect a greater stability of a given polypeptide.

[0060] FIG. 8 illustrates an exemplary method of using a machine learning model for engineered polypeptide design. As shown in FIG. 8, a first set of blueprint records and scores The initial data set including the first set (e.g., Rosetta Energy or Molecular Dynamics Energy) The data preparation module (shown in relation to Figure 1) generates the energy terms, such as the and further preparing the data by a data preparation module 105 as described above. A machine learning model (such as the machine learning model shown and described with respect to FIG. 1) can be used. A group of neural networks (similar to the neural network 107) may be trained based on the initial data set. The second set of lint records is used as input to generate a second set of scores. A score above a predetermined value (e.g., a desired score) may be fed to the learning model. A second set of blueprint records or a part of the second set of blueprint records The portion may be verified against a ground truth score. A second set of scores may be , with sufficient accuracy (e.g., with accuracy above 95%), to the ground truth score. If applicable, a second set of Blueprint records or a second A portion of the set may be presented to the user, or the blueprint record may be Machine learning using the second set, or a portion of the second set of Blueprint records The model may be retrained. In some instances, it may be necessary to retrain the model to achieve the blueprint with the desired score. To achieve this, a third set of Blueprint records, a fourth set of Blueprint records, A larger number of blueprint records may be generated, each of which may be repeated. As many sets of Blueprints as needed to achieve the desired score are added to the Blueprint By repeatedly retraining the machine learning model on new sets of lint and scores Train a machine learning model to generate engineered polypeptide designs. And an example code snippet showing the procedure for use is as follows: training_energies=Rosetta(training_scaff olds)##The Rosetta energy is calculated for an initial training set of scaffolds. While training_energies has not converged:##Rosetta Energy has been improved Repeated until goodness stops Predict training_energies from training_scaffolds Train xgboost to measure Rosetta energy from the scaffold training set Train XGBoost to predict predicted_scaffolds = Top predicted scaffolds from xgboost Predicting optimal footing with XGBoost new_energies=Rosetta(predicted_scaffold s)##Rosetta energies are calculated for predicted scaffolds predicted_scaffolds to training_scaffolds Add##Add the predicted scaffolds to the training set Add new_energies to training_energies##Prediction Add the scaffold energies obtained from the training set

[0061] FIG. 9 is a schematic diagram of an exemplary performance of a machine learning model for engineered polypeptide design. As described with respect to FIG. 5, an exemplary blueprint ( For records that correspond to a 35-mer polypeptide, it is assumed that the target residues are ordered. Then, the potential blue The total number of prints is given. Therefore, each one using brute force detection / optimization Direct computational modeling of blueprints is a novel approach to computational modeling that is based on current computing devices and methods. Using the Using a data-driven approach, such as the machine learning models described herein, Such discovery / optimization times (e.g., weeks, days, hours, minutes, and / or It is possible to reduce the

[0062] 10A-D show molecular dynamics simulations performed to verify the engineered polypeptide. 1 illustrates an exemplary method for implementing a machine learning model (as shown and described with respect to FIG. 1 ). Machine learning models, such as those described above107, are trained and improved / optimized (e.g., The generated bulls (those that meet the design criteria, have the desired scores, and / or the like) After being run to generate a set of print records, the engineered polypeptide design data The device (as shown and described with respect to FIG. 1) is a set of generated design records. can be verified.

[0063] The engineered polypeptide design device is then run on a set of generated blueprint records. Computational protein modeling (computational design model as shown and described with respect to FIG. 1) (using a coding module 106) to produce an engineered polypeptide. In some implementations, the engineered polypeptide design device then performs a step of generating a polypeptide sequence of the reference target structure. A comparison of the static structure to the representation is performed to identify the subset of engineered polypeptides. The data may be filtered.

[0064] In some implementations, the engineered polypeptide design device then comprises a reference target structure and We performed molecular dynamics (MD) simulations of each representation of the engineered polypeptide structure. By comparing the dynamic structure to a representation of the reference target structure, the engineered polypeptide For example, engineered polypeptide design devices may be used to filter out a subset of the The selection may include selecting a few (e.g., less than 10 hits) of the engineered polypeptides. In some instances, MD simulations are performed after model preparation, equilibration (e.g., 100K ~300 K temperature), and the solution including the unconstrained MD simulation steps. Dynamic expression of each of the reference target structure and the engineered polypeptide structure under solution conditions In some instances, MD simulations can be used to determine the force field parameters. The data and / or solvent model parameters are then calculated for the reference target structure and the engineered polypeptide. In some instances, MD simulations may be performed to represent each of the structures. is a constrained minimization (e.g., to mitigate structural clashes) for 1000 cycles, controlled heating (e.g., controlled heating for 100 picoseconds and then gradually increasing to ambient temperature) , relax the constraints (e.g., for 100 picoseconds), and gradually remove the skeletal constraints. (to be able to receive the same benefits)

[0065] FIG. 11 shows a molecular dynamics simulation performed to verify the engineered polypeptide. 10 illustrates an exemplary method for performing the method according to the present invention. In some implementations, in addition or alternatively, For the methods described in connection with this, the MD simulations can be time-limited. For example, MD simulations were performed for 30 ns of unconstrained dynamics. In some implementations, the MD simulation may additionally or alternatively include For example, MD simulations can be performed to obtain such structural information. To obtain 80% of the observed structural information in any given time frame required to achieve Some implementations allow for a trade-off between throughput and accuracy of MD simulations. The metric for determining the simulation time is the reference target structure and the manipulated pole. The cosine similarity score of the simulated representation of each peptide structure was calculated. This may be done.

[0066] FIG. 12 is a schematic diagram of an exemplary method for parallelizing a molecular dynamics simulation. In some examples, the engineered polypeptide designs are many (e.g., 100s, 1000s , 10,000 s, and / or the like) molecular dynamics simulations. In such an embodiment, an engineered polypeptide design device (as shown in relation to FIG. 1) may be used. and the processor 104 of the engineered polypeptide design device 101 as described. The processors in the NVIDIA NVIDIA GPU (Graphical Processing Unit), NVIDIA GPU (Accelerated Processing Unit), NVIDIA GPU (Agile NVIDIA GPU ... and / or any other processing unit capable of performing calculations in parallel. A GPU may include a set of Symmetric Multiprocessing Units (SMPs). Therefore, the GPU can be used to run molecular dynamics simulations using a set of SMPs. to process a number of parallel processes (e.g., 10s, 100s, and / or the like). In some variations, the cloud computing platform (see FIG. 1. A back-end service platform 160 as shown and described with respect to Use multi-core processing units on the may be processed.

[0067] FIG. 13 illustrates an exemplary method for validating a machine learning model for engineered polypeptide design. 1 is a schematic diagram of a method for scoring each engineered polypeptide. In some implementations, the scoring method comprises: To evaluate the structure, we used molecular dynamics (MD) simulation results of the representation of the reference target structure, and The score can be used to calculate the MD simulation results for each of the engineered polypeptides. The method may involve using root mean square deviation (RMSD),

number

number

[0068] The preceding description, for purposes of explanation, uses specific nomenclature to provide a thorough understanding of the present invention. However, it should be understood that specific details are not required to practice the invention. It will be apparent to those skilled in the art that the foregoing description of specific embodiments of the present invention is intended to be illustrative and not restrictive. These are presented for illustrative purposes only. They are not intended to be exhaustive or to be construed as limiting the scope of the present invention. nor is it intended to limit the invention to the precise forms disclosed. Many modifications and variations are possible in light of the teachings. It has been chosen and described in order to explain its practical application and to enable others skilled in the art to easily understand the invention. and utilize various embodiments with various modifications as appropriate for the particular use contemplated. The following claims and their equivalents define the scope of the invention. It is intended to define

[0069] Enumerated embodiments: Embodiment I-1. A method comprising: Based on the first plurality of blueprint records, or representations thereof, and the first plurality of scores. training a machine learning model based on the first plurality of blueprint records; each blueprint record is associated with a respective score from the first plurality of scores; To train and After training, the machine learning model is run to generate a second set of replicas having at least one desired score. and generating a number of blueprint records, A second plurality of blueprint records is received as input in a computational protein modeling. to generate an engineered polypeptide based on the second plurality of blueprint records. A method that is configured to:

[0070] Embodiment I-2. receiving a representation of a reference target structure for a reference target; generating a first plurality of blueprint records from a predetermined portion of the reference target structure; each blueprint record from the first plurality of blueprint records comprising a target residue position and Each target residue position corresponds to a target residue from a plurality of target residues. The method of embodiment I-1, comprising:

[0071] Embodiment I-3. In at least one blueprint record, the target residue position is a non The method according to embodiment I-1 or I-2, which is sequential.

[0072] In at least one blueprint record, the target residue position is Any of embodiments I-1 to I-3, wherein the target residue positions are in an order different from the order of the target residue positions in the reference target sequence. The method according to any one of the above.

[0073] Embodiment I-5. For each blueprint record from the first plurality of blueprint records, Computational protein modeling is performed on the blueprint records to generate polypeptide structures. Generating and calculating a score for the polypeptide structure; and associating the score with the blueprint record, thereby Any one of embodiments I-1 to I-4, including labeling the lint record. How to.

[0074] Embodiment I-6. Computational protein modeling matches templates to reference target structures The method according to any one of embodiments I-1 to I-5, which is based on de novo design without Law.

[0075] Embodiment I-7. Each score from the first plurality of scores is a function of an energy term and a reference target structure. a structural constraint match term determined using one or more structural constraints extracted from the representation of the structure; The method according to any one of embodiments I-1 to I-6, comprising:

[0076] Embodiment I-8. calculating a second plurality of scores for the second plurality of blueprint records; determining whether to retrain the machine learning model; and In response to determining, (1) generating a retraining group including a second plurality of blueprint records; (2) a machine learning algorithm based on the retraining scores, the retraining scores including a second plurality of scores. and retraining the model. Law.

[0077] Embodiment I-9. After retraining the machine learning model, the first plurality of blueprint records and the second plurality of blueprint records are Concatenate multiple Blueprint records to generate a retraining Blueprint record and generate a retraining score. generating a retraining blueprint record for each of the retraining blueprint records; The method of embodiment I-8, wherein the score is associated with a score from a training score.

[0078] Embodiment I-10. An embodiment in which at least one desired score is a preset value The method according to any one of I-1 to I-9.

[0079] Embodiment I-11. The at least one desired score is dynamically determined. The method according to any one of I-1 to I-9.

[0080] Embodiment I-12. The embodiment I-, in which the machine learning model is a supervised machine learning model The method according to any one of 1 to I-10.

[0081] Embodiment I-13. Supervised machine learning models include ensembles of decision trees, boosted decision tree algorithm, eXtreme Gradient Boosting (XGBoost) model, Or the method of embodiment I-12, comprising a random forest.

[0082] Embodiment I-14. The supervised machine learning model is a support vector machine (SVM), Feedforward machine learning models, recurrent neural networks (RNNs), convolution Neural Networks (CNNs), Graph Neural Networks (GNNs), and The method of embodiment I-12, wherein the comprises a Transformer neural network.

[0083] Embodiment I-15. The embodiment I-1, in which the machine learning model is an inductive machine learning model The method according to any one of claims 1 to 14.

[0084] Embodiment I-16. The machine learning model is a generative machine learning model, The method according to any one of claims I-14.

[0085] Embodiment I-17. Computational Protein Modeling on a Second Plurality of Blueprint Records 16. The method of claim 11, further comprising carrying out the method to produce an engineered polypeptide. The method according to any one of the preceding claims.

[0086] Embodiment I-18. Manipulated by comparison of a static structure to a representation of a reference target structure Any one of the embodiments I-1 to I-17, comprising filtering the polypeptide. The method according to claim 1.

[0087] Embodiment I-19. Representation of the reference target structure and the structure of the engineered polypeptide of the dynamic structure against a representation of the reference target structure using molecular dynamics (MD) simulations The embodiment I-, comprising filtering the engineered polypeptides by comparison The method according to any one of 1 to I-18.

[0088] Embodiment I-20. The MD simulation uses symmetric multiprocessing (SMP) The method of embodiment I-19, wherein the method is performed in parallel using

[0089] Embodiment I-21. The number of blueprint records in the second plurality of blueprint records is less than the number of blueprint records in the first plurality of blueprint records; The method according to any one of I-1 to I-20.

[0090] Embodiment I-22. A non-transitory device for storing code representing instructions to be executed by a processor A processor-readable medium, comprising code for causing a processor to: Based on the first plurality of blueprint records, or representations thereof, and the first plurality of scores. training a machine learning model based on the first plurality of blueprint records; each blueprint record is associated with a respective score from the first plurality of scores; To train and After training, the machine learning model is run to generate a second replica having at least one desired score. generating a number of blueprint records; A second plurality of blueprint records is received as input in a computational protein modeling. to generate an engineered polypeptide based on the second plurality of blueprint records. 23. A non-transitory processor-readable medium configured to:

[0091] Embodiment I-23. A processor comprising: receiving a representation of a reference target structure; generating a first plurality of blueprint records from a predetermined portion of the reference target structure; each blueprint record from the first plurality of blueprint records comprising a target residue position and each target residue position from the plurality of target residue positions comprising a scaffold residue position; and generating a target residue corresponding to one of the target residues of -The medium described in 22.

[0092] Embodiment I-24. In at least one blueprint record, the target residue position is: The medium of embodiment I-23, which is non-continuous.

[0093] Embodiment I-25. In at least one blueprint record, the target residue position is: In one embodiment, the target residue positions are in an order different from the order of the target residue positions in the reference target sequence. -The medium described in 24.

[0094] Embodiment I-26. A processor comprising: Computational protein modeling is performed on each blueprint record to generate a polypeptide structure. The score of the polypeptide structure is calculated, and the score is recorded in the blueprint. and labeling the first plurality of blueprint records by associating the first plurality of blueprint records with the first plurality of blueprint records. The medium of any one of embodiments I-23 to I-25, comprising a code for:

[0095] Embodiment I-27. Computational protein modeling matches templates with reference target structures The medium according to embodiment I-26, which is based on de novo design, without

[0096] Embodiment I-28. Each score is derived from an energy term and a representation of a reference target structure. and a structure constraint match term determined using one or more structure constraints. 6 or the medium described in I-27.

[0097] Embodiment I-29. A processor comprising: calculating a second plurality of scores for the second plurality of blueprint records; determining whether to retrain the machine learning model; and In response to determining, (1) generating a retraining group including a second plurality of blueprint records; (2) a machine learning algorithm based on the retraining scores, the retraining scores including a second plurality of scores. Any of embodiments I-22 to I-28 including a code for retraining the model. The medium described in any one of the above.

[0098] In an embodiment I-30, a processor is provided, After retraining the machine learning model, the first plurality of blueprint records and the second plurality of blueprint records are Concatenate multiple Blueprint records to generate a retraining Blueprint record and Each Blueprint record contains the code to generate a core and retrain the Blueprint record. The medium of embodiment I-29, wherein is associated with a score from the retraining score.

[0099] Embodiment I-31. The embodiment in which at least one desired score is a preset value A medium described in any one of I-22 to I-30.

[0100] Embodiment I-32. The at least one desired score is dynamically determined. A medium described in any one of I-22 to I-31.

[0101] 3. The embodiment I-3, wherein the machine learning model is a supervised machine learning model. The medium according to any one of 22 to I-32.

[0102] Embodiment I-34. Supervised machine learning models include ensembles of decision trees, boosted decision tree algorithm, eXtreme Gradient Boosting (XGBoost) model, or a random forest, Medium.

[0103] Embodiment I-35. The supervised machine learning model is a support vector machine (SVM); Feedforward machine learning models, recurrent neural networks (RNNs), convolution Neural Networks (CNNs), Graph Neural Networks (GNNs), and The medium of embodiment I-33, wherein the comprises a Transformer neural network.

[0104] Embodiment I-36. The embodiment I-2, in which the machine learning model is an inductive machine learning model. The medium described in any one of 2 to I-35.

[0105] Embodiment I-37. The embodiment I-22, in which the machine learning model is a generative machine learning model. A medium according to any one of claims 1 to 36.

[0106] Embodiment I-38. A processor comprising: Computational protein modeling was performed on the second plurality of blueprint records to engineer Any one of the embodiments I-22 to I-37, comprising a code for producing a polypeptide. The medium described.

[0107] Embodiment I-39. A processor comprising: Filter engineered polypeptides by comparison of their static structures against representations of reference target structures. The medium of embodiment I-38, comprising code for tagging.

[0108] Embodiment I-40. A processor comprising: Molecular dynamics (MD) simulations of the reference target structure and each representation of the engineered polypeptide The manipulated position is identified by comparing the dynamic structure to a representation of the reference target structure using a 3D model. The method according to any one of the preceding claims, including a code for filtering the polypeptides. Media.

[0109] EMBODIMENT I-41. MD simulation using symmetric multiprocessing (SMP) The medium of embodiment I-40, wherein the medium is carried out in parallel using

[0110] Embodiment I-42. The number of blueprint records in the second plurality of blueprint records is less than the number of blueprint records in the first plurality of blueprint records; A medium described in any one of I-22 to I-41.

[0111] Embodiment I-43. An apparatus for selecting engineered polypeptides, comprising: A processor; a first computing device having a memory; Memory is from a second computing device remote from the first computing device; Receiving a reference target structure; generating a first plurality of blueprint records from a predetermined portion of the reference target structure; each blueprint record from the first plurality of blueprint records comprising a target residue position and Each target residue position corresponds to a target residue from a plurality of target residues. To generate, Based on the first plurality of blueprint records, or representations thereof, and the first plurality of scores. training a machine learning model based on the first plurality of blueprint records; each blueprint record is associated with a respective score from the first plurality of scores; To train and After training, the machine learning model is run to generate a second replica having at least one desired score. and generating a number of blueprint records. It remembers the commands, A second plurality of blueprint records is received as input in a computational protein modeling. to generate an engineered polypeptide based on the second plurality of blueprint records. The device is configured to.

[0112] Embodiment I-44. A processor comprising: calculating a second plurality of scores for the second plurality of blueprint records; determining whether to retrain the machine learning model; and In response to determining, (1) generating a retraining group including a second plurality of blueprint records; (2) a machine learning algorithm based on the retraining scores, the retraining scores including a second plurality of scores. The apparatus of embodiment I-43, further comprising code for retraining the model.

[0113] Embodiment I-45. The embodiment I-43 or I-46, wherein the desired score is a preset value. - The device according to claim 44.

[0114] Embodiment I-46. The desired score is dynamically determined, embodiments I-43 to I-45 13. The apparatus according to claim 12,

[0115] Embodiment I-47. The embodiment I-, in which the machine learning model is a supervised machine learning model. The apparatus according to any one of claims 43 to I-46.

[0116] Embodiment I-48. Supervised machine learning models include ensembles of decision trees, boosted decision tree algorithm, eXtreme Gradient Boosting (XGBoost) model, Or the apparatus of embodiment I-47, comprising a random forest.

[0117] Embodiment I-49. The supervised machine learning model is a support vector machine (SVM); Feedforward machine learning models, recurrent neural networks (RNNs), convolution Neural Networks (CNNs), Graph Neural Networks (GNNs), and In the embodiment I-47 or I-48, the transformer neural network is included. The apparatus described in

[0118] Embodiment I-50. The embodiment I-4, in which the machine learning model is an inductive machine learning model. The apparatus according to any one of claims 3 to I-49.

[0119] Embodiment I-51. The embodiment I-43, in which the machine learning model is a generative machine learning model. 10. An apparatus according to any one of claims 1 to 50.

[0120] Embodiment I-52. A processor comprising: Computational protein modeling was performed on the second plurality of blueprint records to engineer Any one of the embodiments I-43 to I-51, comprising a code for producing a polypeptide. The apparatus described.

[0121] Embodiment I-53. A processor comprising: Filter engineered polypeptides by comparison of their static structures against representations of reference target structures. The apparatus of embodiment I-52, further comprising a code for tagging.

[0122] Embodiment I-54. A processor comprising: Molecular dynamics (MD) simulations of the reference target structure and each representation of the engineered polypeptide The manipulated position is identified by comparing the dynamic structure to a representation of the reference target structure using a 3D model. The embodiment I-52 or the embodiment I-53, comprising a code for filtering the peptides. 53. The apparatus according to claim 53.

[0123] Embodiment I-55. The MD simulation uses symmetric multiprocessing (SMP). The apparatus of embodiment I-54, wherein the apparatus is performed in parallel using

[0124] Embodiment I-56. The method according to any one of embodiments I-1 to I-21, The medium according to any one of I-22 to I-42 or the medium according to any one of I-43 to I-55 An engineered polypeptide design produced by the apparatus of any one of claims 1 to 4.

[0125] Embodiment I-57. An engineered peptide, wherein the engineered peptide is between 1 kDa and The engineered peptide has a molecular weight of 10 kDa and contains up to 50 amino acids. A combination of spatially related topological constraints, where one or more of the constraints Target-based constraints, Between 10% and 98% of the amino acids in the engineered peptide satisfy constraints derived from one or more reference targets. And, Amino acids satisfying one or more of the constraints from the reference target have a backbone average of less than 8.0 Å with respect to the reference target. Engineered peptides with root-squared deviation (RSMD) structural homology.

[0126] Embodiment I-58. Amino acids satisfying the constraints from one or more reference targets are identical to the reference targets The engineered peptide of embodiment I-57, having between 0% and 90% sequence homology.

[0127] Embodiment I-59. The combination includes constraints from at least two reference targets. The engineered peptide of form I-57 or I-58.

[0128] Embodiment I-60. Combinations are extracted from energy terms and representations of reference target structures. and a structure constraint match term determined using one or more structure constraints. 57 to 1-59.

[0129] Embodiment I-61. One or more non-reference target-derived constraints are used to determine desired structural, dynamic, or other properties. or any combination thereof. 1. An engineered peptide according to claim 1.

[0130] Embodiment I-62. The reference target is one or more associated with a biological response or function containing atoms of The origin of one or more atoms in an engineered peptide that is associated with a biological response or function. The molecular variation is the origin of one or more atoms in the reference target that are associated with a biological response or function. The engineered peptide according to any one of embodiments I-57 to I-61, which overlaps with the child variation. Chid.

[0131] Embodiment I-63. The overlap is greater than 0.25 root mean square dot product (RMSIP). The engineered peptide according to embodiment I-62.

[0132] Embodiment I-64. The overlap has a root mean square dot product (RMSIP) greater than 0.75. The engineered peptide according to embodiment I-62 or I-63, comprising:

[0133] Embodiment I-65. A method for selecting an engineered peptide, comprising: Identifying one or more topological properties of the reference target; Each target is then analyzed to generate a set of spatially related topological constraints derived from the reference target. Designing spatially related constraints on the physical properties; The spatially related topological properties of the candidate peptides are compared with the spatially related topological properties of the reference target. Comparing with combinations of topological constraints, A set of spatially related topological constraints that overlap with the set of spatially related topological constraints from the reference target. Select candidate peptides with associated topological properties to generate engineered peptides. and

[0134] Embodiment I-66. One or more constraints are per-residue energy and per-residue elemental The method of embodiment I-65, wherein the distance is derived from the child distance.

[0135] Embodiment I-67. The properties of one or more candidate peptides are determined by computer simulation. The method of any one of embodiments I-65 or I-66, wherein the method is determined by

[0136] Embodiment I-68. The computer simulation comprises a molecular dynamics simulation; Monte Carlo simulation, coarse-grained simulation, Gaussian network model, The method of embodiment I-67, comprising machine learning, or any combination thereof.

[0137] Embodiment I-69. Amino acids satisfying the constraints from one or more reference targets are identical to the reference targets Any one of embodiments I-65 to I-68 having 0% to 90% sequence homology. How to.

[0138] Embodiment I-70. One or more non-reference target-derived constraints are used to determine desired structural characteristics and / or The method of any one of embodiments I-65 to I-69, describing the dynamic characteristics.

Claims

1. An engineered peptide, said engineered peptide having a molecular weight between 1 kDa and 10 kDa and containing up to 50 amino acids, said engineered peptide comprising: a combination of spatially related topological constraints, one or more of which are derived from a reference target; between 10% and 98% of the amino acids of the engineered peptide satisfy constraints derived from the one or more reference targets; The engineered peptide, wherein the amino acids that satisfy the constraints from the one or more reference targets have a backbone root mean square deviation (RSMD) structural identity with the reference target of less than 8.0 Å.

2. the amino acids satisfying the constraints from the one or more reference targets have between 10% and 90% sequence identity with the reference targets; the combination includes constraints from at least two reference targets; the combination includes an energy term and a structural constraint match term determined using one or more structural constraints extracted from a representation of a reference target structure; the one or more non-reference target-derived constraints define desired structural properties, dynamic properties, or any combination thereof; the reference target comprises one or more atoms associated with a biological response or biological function, and an atomic variation of the one or more atoms in the engineered peptide associated with a biological response or biological function overlaps with an atomic variation of the one or more atoms in the reference target associated with a biological response or biological function; The overlap is a root mean square dot product (RMSIP) greater than 0.25, or the overlap has a root mean square dot product (RMSIP) greater than 0.

75.

2. The engineered peptide of claim 1, wherein said peptide is at least one of:

3. 1. A method for selecting an engineered peptide, comprising: Identifying one or more topological properties of a reference target; designing spatially related constraints for each topological property to generate a combination of spatially related topological constraints derived from the reference target; comparing the spatially related topological properties of the candidate peptide with a combination of the spatially related topological constraints from the reference target; selecting candidate peptides having spatially related topological properties that overlap with the combination of spatially related topological constraints from the reference target to generate the engineered peptide; The method comprising:

4. One or more constraints are derived from per-residue energies and per-residue atom distances, The properties of one or more candidate peptides are determined by computer simulation; or The computer simulation includes a molecular dynamics simulation, a Monte Carlo simulation, a coarse-grained simulation, a Gaussian network model, machine learning, or any combination thereof; the amino acids satisfying the constraints from the one or more reference targets have between 10% and 90% sequence identity with the reference targets; or The one or more non-reference target-derived constraints describe desired structural and / or dynamic properties. The method of claim 3, wherein the at least one of

Images