Method for in silico sequence-based design of a peptide which binds to a target

Abstract

The present invention provides an in silico method of designing peptides that bind a target. The method comprising providing a starting pool of parent peptides; subjecting the pool of parent peptides to in silico mutation and crossover to produce an in silico pool of offspring peptides. To identify candidate peptides, interaction scores and a fitness value is determined for each offspring peptide. Offspring peptides with the highest calculated fitness value are selected. Also provided are methods of producing and using the designed peptides.

Description

METHOD FOR IN SILICO SEQUENCE BASED DESIGN OF A PEPTIDE WHICH BINDS TO A TARGETFIELD

[0001] The present invention pertains to the field of peptide drug development, in particular, in silico sequence-based design of peptides which bind to a target.BACKGROUND

[0002] Small molecules represent the majority of the currently approved therapeutics designed to disrupt deleterious PPIs, inhibit overactive enzymes, or activate receptors on the cell’s surface. While small molecules often possess advantageous pharmacoavailability, they are often described as lacking specificity, and thus, are susceptible to off-target interactions, often leading to undesired side effects. Peptides, or short sequences of amino acids, in contrast, can be designed to exhibit exquisite binding specificity for their target. Dozens of peptide-based drugs have been approved by the FDA since the turn the century to treat a variety of conditions including diabetes, heart failure, obesity, and several cancers (Wang et al.2022).

[0003] Delivering peptides to the right target is one challenge, but another, perhaps even more significant challenge lies upstream: designing target-specific, bioactive peptide binders. Several experiment-based methods to develop peptide binders against specific targets already exist. These approaches are resource intensive, both in terms of costs and time, and may never reach regions of the peptide space where binders are located. For this reason, computational design of peptide binders has emerged as a convenient alternative to probe the peptide space for candidates to be validated in vitro.

[0004] Most computational drug research has focused on small molecules, for which there is greater biochemical and clinical data owing to the quantity of pharmacological data available for this class of drug. In contrast, there are only a few dozen peptides in use in clinical practice (Wang et al. 2022).

[0005] A number of groups have leveraged recent advances in deep learning to build algorithms specifically for peptide design but only a couple methods addressing the challenging problem of designing peptide binders using only the target protein sequence have been published. One such method is InSiPS (Schoenrock et al. 2015; Burnside et al. 2018), which frames the peptide design as an optimization problem and solves it using a genetic algorithmwhere the peptide candidate’s fitness is a function of its predicted interaction scores with the target protein and off-target proteins. The InSiPS calculates fitness as follows:f= sT[1 -max(Sε\T)]where f is the fitness, sTis the interaction score between the peptide candidate in the target, and Sε\Tis the set of interaction scores between the peptide candidate and all non-target proteins.

[0006] By merely considering the highest scoring off-target protein for a given peptide as a proxy for specificity, the InSiPS fitness score fails to account for all other off-target interactions and loses useful context. While maximizing this fitness function does account for both objectives of 1) maximizing the predicted probability that the peptide candidate binds to the target and 2) minimizing the probability of an off-target interaction, it fails to capture some important context. More specifically, this fitness function will generate candidates where sTis as high as possible and max(Sε\T) is as low as possible, but it considers no off-target interactions except for the highest-scoring off-target interaction. This is problematic because, in practice, interest is placed in how a peptide will interact with all the proteins to which it is exposed, not only the highest-scoring one.

[0007] In addition, InSiPS evolves peptide binders for a specified target using a genetic algorithm strategy that relies on the PIPE algorithm (more specifically, a “massively parallel” version called MP-PIPE) to calculate the fitness of the peptide candidates. PIPE embodies a significant advance in the field of PPI prediction, having been the first algorithm to generate the comprehensive human interactome (Schoenrock et al. 2011). However, this feat first reported in 2011 required significant computational resources: 3 months of computation, using 6,400 hardware-supported threads (Schoenrock et al. 2011). The authors of a recent article describing the design of peptide binders against the SARS-CoV-2 Spike protein report having used InSiPS (Hajikarimlou et al. 2022). They report that each run required about a week of computation using 10,000 processor cores. Of course, nearly all of that computation is spent scoring protein-peptide interactions with MP-PIPE.

[0008] Given the vastness of the peptide space, there is significant incentive to improve the efficiency of the optimization procedure. Accelerating PPI prediction is one way to achieve this, as this would allow for deeper exploration of the peptide space by enabling the both the size and number of GA peptide populations. This increases the probability of the algorithm converging towards suitable peptides binders with as few off-target interactions as possible.

[0009] The SPRINT algorithm (Li and Llie, 2017) borrows substantially from PIPE (Pitre et al. 2006; Pitre et al. 2008; Dick et al. 2020). SPRINT uses similarity windows (high-scoring segment pair (HSP)). An HSP is defined by two locations (which may or may not occur within the same protein) as well as the number of residues comprised in the window. In contrast with PIPE, SPRINT does not exhaustively score all possible window pairs in the proteins considered. Instead, it uses an approach where short, spaced seeds are used to identify candidate locations where similarity windows may occur. The SPRINT algorithm follows the following steps:

[0010] Step 1 - Extraction of all s-mers: A set of 4 space seeds of length 12 are used to extract all s-mers occurring in all protein sequences provided. The space seeds used by SPRINT are 11****11***1, 1**1***1**1, 1**1*1***1*1, and 1*1******111 positions with a 1 are positions that are considered (“care” positions) while positions denoted with a * are not considered (“don’t care” positions). These seeds are slid along the length of all proteins and used to extract all s-mers, each of which is encoded in an unsigned 64-bit integer, where each position takes 5 bits (highly memory-efficient representation). These seeds are used to populate a table that keeps track of where all s-mers occur in the sequences.

[0011] Step 2 - Identification of HSPs: The algorithm compares all extracted s-mers and finds ones that are similar at “care” positions, l.e. the ones that meet a tuneable score threshold, when scored with a substitution matrix. Protein locations that are associated with similar s-mers are considered potential HSPs and are then scored using all positions. If the similarity score threshold, THSP, is met, the similarity region is extended to the left and to the right until the similarity score is no longer met. If the similarity score threshold between two windows is not met, that pair is not a HSP and no attempt to extend the window is made.

[0012] Step 3 - Correction for frequent windows: SPRINT checks within every protein whether there are residues that occur frequently within HSPs. Residues that are involved in too many within HSPs (above a count Tcount) are removed from HSPs in which they occur, thus splitting HSPs into smaller ones. The rationale for this step is that residues involved in too many HSPs are likely to not be relevant to interactions. In this step, each amino acid within HSPs is considered independently, but frequently, contiguous residues meet the count threshold, causing HSPs the be split by a window of several residues rather than a single residue.

[0013] Step 4 - Interaction scoring: The last step in the SPRINT algorithm is the scoring of protein pairs (usually the complete interactome, or all pairs of proteins in a proteome). To achieve this, the algorithm iterates through a set of protein pairs assumed to engage in physical interactions (which we shall refer to as “training pairs”), referred to as a “training set”, and increments the interaction score for the pair (P1, P2) if P1shares a HSP with I1, and P2sharesan HSP with I2, I1and I2being known interactors in the training set. Unlike PIPE, however, SPRINT does not keep track of what residues within protein pairs contribute to their interaction score.

[0014] The use of HSPs allows SPRINT to achieve very significant speedups over PIPE. This allows SPRINT to predict the entire human interactome in a matter of one hour or so with fewer than 100 CPUs, while PIPE requires a computer cluster with thousands of cores and days or weeks of computation. In addition, SPRINT differs from PIPE in one other important way: instead of simply counting the number of similarities of (equal-length) windows in a query pair that are similar to known interactors, SPRINT weighs the score updates by the length-adjusted similarity of those windows.

[0015] This background information is provided for the purpose of making known information believed by the applicant to be of possible relevance to the present invention. No admission is necessarily intended, nor should be construed, that any of the preceding information constitutes prior art against the present invention.SUMMARY

[0016] The present application provides for in silico methods of designing peptides that bind a target protein. In some embodiments, the method comprises:a) providing a pool of parent peptides each having an amino acid sequence;b) subjecting the parent peptides to in silico mutation and crossover to produce an in silico pool of offspring peptides;c) determining an interaction score for each offspring peptide based on their sequence similarity to known interactors of the target protein or homologs thereof;d) determining a calculated fitness value based on the interaction score that quantifies the likelihood of interaction between each offspring peptide and the target protein relative to interaction between each offspring peptide and off-target proteins, and selecting offspring peptides with the highest calculated fitness value to generate a new in silico pool of parent peptides;e) repeating steps (b) to (d) until the interaction scores cease to increase; andf) selecting one or more offspring peptides with the highest fitness values as candidate peptides that bind the target protein.

[0017] In one embodiment, determining the calculated fitness value comprises calculating a numerical value with a rank-based fitness function configured to consider the rank of the target protein relative to a given set of possible protein interactions each offspring peptide may engage in. In one embodiment, rank is determined based on the interaction score. In one embodiment, the calculated fitness value of a given offspring peptide incorporates the interaction score for said offspring peptide. In one embodiment, the rank-based fitness function has a rank term for quantifying the likelihood of interaction between one offspring peptide and the target protein relative to off-target proteins. In one embodiment, determining the calculated fitness value comprises applying a fitness penalty to offspring peptides having a rank term below a given threshold.

[0018] In one embodiment, the rank-based fitness function (fRFF) has the formula:if max(Sp,ε\T) ≤ τfRFF(p, T, Sp, ε) = (1)otherwisewherein:sp,T∈ Spis a predicted peptide protein interaction score determined between the peptide (p) and the target protein (T),£ is the exposome consideredS£\Tis a set of all predicted interaction scores between p and all non-target proteins, hyperparameter T controls a penalty term used to constrain the search space, and rp Tis the rank of sPiTrelative to all other interactions on the exposome.

[0019] In one embodiment, selecting the offspring peptides with the highest interaction scores comprise calculating the interaction score with a selection fitness function (fRFF (selection)) having the formula:1 P’ fRFF (selection) (P> $p, TT(2)|ε|. wherein:sPiTe Spis the interaction score determined between the peptide (p) and the target protein (T),£ is the exposome consideredrp Tis the rank of sPiTrelative to all other interactions on the exposome.

[0020] In one embodiment, determining the interaction score comprises: i) extracting all s-mers occurring in the sequence of each offspring peptide, based one or more space seeds of a given length; ii) identifying high-scoring segment pairs (HSPs) based on a predetermined threshold; iii) correcting for frequent windows by removing residues that are involved in more than a given number of HSPs; and iv) scoring the interaction of each offspring peptide and all possible targets with a training set of known protein- protein interactions; and wherein the method further comprises identifying a potential binding region of the target protein based on the contributions of the HSPs to the interaction score, by calculating an individual residue contribution (c) for each amino acid residue of the candidate peptides using a set of four proteins comprising two proteins known to interact and two query proteins, where one of the query proteins shares a HSP with one of the known interacting proteins and the other query protein shares a HSP with the other known interacting protein, and wherein amino acid residues having higher individual residue contribution values are identified as having higher likelihood of comprising the binding region.

[0021] In one embodiment, calculating the individual residue contribution comprises first calculating contribution of the set of four proteins (c(he, hp)) based on formula:sh_e(|hp| - k + 1)sh_p(|he| - k + 1)c(he, hp) =|hp||he| wherein:the set of four proteins has one exposon-interactor HSP (he) and one peptide-interactor HSP (hp),shis a substitution matrix-derived similarity score of the HSP, and |h| is the length of the HSP.

[0022] In one embodiment, calculating the individual residue contribution comprises deriving a mean per-residue contribution (c̄(he,hp)) from the c(he,hp) based on formula:(4)and adding said mean per-residue contribution to the individual residue contribution.

[0023] In some embodiment, systems are provided comprising one or more processors operating in conjunction with a memory storage device, the one or more processors configured to perform the methods described herein.

[0024] In one embodiment, the memory storage device stores data representing s-mers as unsigned 64-bit integers in a hash table optimized for single instruction multiple data (SIMD) operations. In one embodiment, the data representing s-mers further comprise corresponding location data in the exposome and / or known interactors. In one embodiment, only peptide interaction scores between offspring peptide and protein target pairs are stored in the memory storage device. In one embodiment, the memory storage device comprises a shared memory module configured to store the interaction scores as a score matrix, the score matrix comprising a 1-D array of floats and a second dimension of a plurality of threads tensor. In one embodiment, the plurality of threads are assigned an index and increment the interaction scores along a single slice along the threads axis, and wherein the shared memory module is configured to collapse the tensor back into a 2-D matrix by summing the interaction scores along the threads axis.

[0025] In some embodiments, a non-transitory computer readable medium is provided storing machine interpretable instructions, which when executed by a processor, cause the processor to perform the methods described herein.

[0026] In one embodiment, the systems described herein further comprises one or more of: a) a display device for visually representing interactions between candidate proteins and the target protein, b) a peptide synthesizer, c) a protein assay screening apparatus, and d) a peptide sequencer. In one embodiment, the peptide synthesizer is a solid-phase peptide synthesis (SPPS) apparatus and / or solution-phase peptide synthesis (SPS) apparatus. In one embodiment, the display device displays the interactions with off / on-target plots and / or or one-to-all curves.

[0027] Another object of the present application is to provide a method for in silica sequence-based design of a peptide which binds to a target. In accordance with an aspect of the present invention, there is provided an in silico method of designing peptides that bind a target, the method comprises:(a) providing an in silica pool of parent peptides;(b) subjecting the in silica pool of parent peptides to in silico mutation and crossover to produce an in silico pool of offspring peptides;(c) determining the interaction (fitness) score of each of the offspring peptides with a target; (d) selecting the offspring peptides with the highest interaction (fitness) scores to generate a new in silico pool of parent peptides;(e) repeating steps (b) to (d) until the interaction (fitness) scores cease to increase; and (f) selecting peptides with the highest interaction (fitness) scores as a potential peptide binder of the target.

[0028] In certain embodiments, step (c) utilizes a similarity-based sequence-based method of determining interaction with a target.

[0029] In certain embodiments, the similarity-based sequence-based method comprises the steps of extraction of all s-mers; identification of HSPs; correction for frequent windows; and interaction scoring.

[0030] In certain embodiments, the fitness of peptides is calculated with:ifmax(Sp T) < Tfp. FF (P’ T, Sp, 8) —otherwisewhere T is a score cut-off hyperparameter used to constrain the search space, and rPiTis the rank of sPiTon the one-to-all curve i.e. when ranking all scores in Spin decreasing order, and £ is the set of all proteins the peptide may interact with due to co-localization exposome; orfllFF (selectiori) (. P’ $p, T |£ | ]

[0031] In certain embodiments, selecting the offspring peptides utilizes tournament selection.

[0032] In accordance with another aspect of the present disclosure, there is provided a system for computationally efficient sequence-based peptide binder design for a target protein, the system comprising one or more processors operating in conjunction with computer memory, the one or more processors configured to execute the above method.

[0033] In accordance with another aspect of the present disclosure, there is provided a non-transitory computer readable medium storing machine interpretable instructions, which when executed by a processor, cause the processor to perform the above computationally efficient method of sequence-based peptide binder design for a target protein.BRIEF DESCRIPTION OF THE FIGURES

[0034] In the figures, embodiments are illustrated by way of example. It is to be expressly understood that the description and figures are only for the purpose of illustration and as an aid to understanding.

[0035] Embodiments will now be described, by way of example only, with reference to the attached figures, wherein in the figures:

[0036] Fig. 1 provides a genetic algorithm overview. The genetic algorithm optimization procedure consists of three steps repeated iteratively. First, the fitness of individuals in an initial (usually randomly generated) population is evaluated. Second, the fittest individuals are then selected preferentially. Finally, the chromosomes of these individuals are recombined and mutated to generate a new population. The procedure then repeats until some stopping criterion is met, e.g. the fitness of the best solution ceases to improve for a set number of iterations. Because the fittest individuals are preferentially selected at every iteration, the overall fitness population tends to increase until a maximum fitness value is reached, at which point the fitness stops improving. In the example illustrated here, the individuals being evolved are sequences of 8 integers between 1 and 9 which could, for example, represent a sequence of 8 actions to be performed by a robot.

[0037] Fig. 2 illustrates the mutation and crossover genetic operators. (A) The mutation operator changes the elements in peptide sequence at random. The probability that a given element in a peptide sequence is mutated follows a mutation rate. (B) Crossover generates new individuals by recombining parts of parental peptide sequences. Three common crossover operators are the one-point, two-points crossover, and uniform crossover. (C) Selection creates a new population from candidate solutions by selecting fitter individuals preferentially. Individuals may be selected more than once to make up the next generation. The selection operator is analogous to maturation to reproductive age. Some make it, but some don’t.

[0038] Fig. 3 illustrates the evidence tetrad underpinning similarity-based prediction methods. Similarity-based PPI prediction methods rely on the idea that a query protein pair which shares similarity windows with a known interacting pair is more likely to interact. In the example illustrated here, query protein PA“looks like”because windows wAandare similar, and PB“looks like” I2because the windows wBand w2are similar. This provides evidence that PAand PBmay engage in an interaction in a similar fashion to and I2.

[0039] Fig.4 illustrates extraction of s-mers from the proteome. To extract s-mers, a space seed of length 12 is slid along the length of every protein, and at every “care” position (1, in green), a 5-bit code for the amino acid is retrieved from a code table to encode the amino acid at that position. The s-mer is a unsigned 64-bit integer that encodes the amino acid at every position with 5-bits (the first 5 bits encode the first amino acid in the sequence, the next 5 the second amino acid, and so on). The s-mers extracted are added to a table that keeps track of their locations in the protein set. The data structure holding these HSPs is a hash table withthe s-mer numerical value as the key and a vector of locations (pair of proteins and positions) as values.

[0040] Fig. 5 illustrates identification of HSPs. SPRINT identifies HSPs by comparing window pairs that contain similar s-mers. It scores window pairs of length k that encompass the s-mers occurences and attempts to extend them in both directions if the score exceeds a threshold THSP by considering all residues in the windows.

[0041] Fig. 6 provides a visual representation of a HSP and its components. HSPs are fully described with tuples of the form (PAi:j, PBk:l, ssim), where PAi:jis the amino acid window comprising residues at positions i to j in protein PA, and ssimis the similarity score of the window, as assessed by a substitution matrix such as BLOSUM62. In the example illustrated here, the HSP is fully described by the tuple (PA1:31, PB436:466, ssim).

[0042] Fig. 7 illustrates processing of HSPs to correct for frequent subsequences. SPRINT splits HSPs where residues are involved in τcount(e.g. 5) or more HSPs. In the example illustrated here, the residue T (on pink background) is involved in 5 HSPs ≥ τcount= 5. Therefore, all HSPs encompassing that residue are split, leading to the formation of newer, smaller HSPs.

[0043] Fig. 8 illustrates scoring protein pairs with known protein-protein interactions. SPRINT iterates through known, validated PPIs and increments the scores for pairs of query proteins that share HSPs with proteins in the interacting pair. The score increment is weighed by the similarity of the query proteins to the known interactors.

[0044] Fig. 9 illustrates optimizing peptides against a fitness model. Since the domain of the fitness landscape (dark) is too vast for it to be fully defined, maximization is performed against a model of the fitness landscape (light).

[0045] Fig. 10 illustrates the limitation of the fitness function used by InSiPS. By merely considering the highest scoring off-target protein for a given peptide as a proxy for specificity, the InSiPS fitness score fails to account for all other off-target interactions and loses useful context.

[0046] Fig. 11 illustrates exposomes for intra- and extra-cellular targets. The training set of interactor pairs is used to score the putative interactions between a peptide and its exposome. (A) In scenario A, the target is an intracellular protein. Assuming that the peptide is delivered to an organism through injection, ingestion, or nasally to inhibit a methyltransferase (intracellular) and can penetrate the cell, then, the peptide will have the opportunity to come in contact with proteins in the extracellular space (not shown), on the cell surface, and inside thecell. The exposome, in that case, corresponds to all of those proteins. The putative interactions between the peptide binder and all of those proteins must be scored. (B) In scenario B, the target is a receptor expressed on the surface of the cell. Assuming that the peptide cannot penetrate the cell, then, the exposome does not include intracellular proteins (shaded).

[0047] Fig. 12 illustrates the structure of an evidence tetrad for peptide-exposon interaction scoring. Two types of HSPs within an evidence tetrad are distinguished: a peptide-interactor HSP where the peptide shares a similarity region with a known interactor (left), and another where an exposon shares a similarity region with the other interactor (right).

[0048] Fig. 13 illustrates identification of individual residue contributions to peptide-exposon interaction scores.

[0049] Fig. 14 illustrates adaptation of the interaction score matrix for peptide scoring. (A) Darwin’s peptide scoring module stores the scores only once (half-matrix in light grey) to prevent redundant memory usage. (B) For peptide scoring with Darwin’s peptide scoring module, the scores are only stored and only stored once if they correspond to a peptide-exposon pair. Scores for irrelevant pairs are not stored. Using SPRINT’S original score matrix for peptide scoring in a context where a population of thousands or tens of thousands of peptides are to be scored at a time would be tremendously memory inefficient, as only a fraction of the scores in the matrix are unique and relevant.

[0050] Fig. 15 illustrates data race causing some stochasticity in predictions and solution. The SPRINT algorithm (when adapted for peptide-protein interaction scoring) increments scores for peptide-protein pairs contained in a Npeptidex Nproteinscore matrix. Each individual known interacting protein pairs is processed by a single thread. This thread searches for HSPs involving proteins in the pair in a HSP table optimized and increments the score for a protein-peptide pairs that share HSPs with the known interacting pair. (A) Occasionally, as shown in pane A, two threads attempt to modify the same location in the array, leading to a data race and stochasticity in the scores across runs. (B) This issue is addressed by scoring interactions in a 3-D tensor where threads can only increment scores along their own slice along the threads axis. Since two threads operate along different slices, they cannot modify the same tensor entry, even for the same protein-peptide pair (e.g. pale and dark grey cells correspond to the same protein-peptide pair).

[0051] Fig. 16 provides a high-level overview of the proposed method employed by Darwin. (A) An initial pool of parent peptide candidates is generated, prior to being subjected to mutation and crossover operations to generate new peptide candidates. The interaction score of these peptides against the exposome are then evaluated with the peptide scoring module to computetheir fitness. The fittest peptides are selected with higher probability to generate a new peptide pool. The process is repeated until the fitness ceases to improve. (B) The peptide scoring module quantifies the evidence that a peptide P might interact with the target T by counting the contributions of evidence tetrads (in this illustration, p -- / 2- T and p - 13- / 4- T).

[0052] Fig. 17 illustrates the runtime difference between the original implementation of SPRINT and the proposed optimized peptide scoring module, showing a significant runtime improvement as compared to the alternative.

[0053] Fig. 18 illustrates the impact of population size and training set size on runtime.

[0054] Fig. 19 illustrates the space requirement of the PPI matrix for peptide scoring.

[0055] Fig. 20 (A) evolution curves for three different fitness functions and targets, (B) a comparison of three performance metrics for all three fitness functions and targets, and (C) clusters generated by CD-HIT at 70% identity on peptides explored with all three fitness functions and targets. The number of clusters required to cluster an equal number of peptides is reflective of the sequence diversity generated under the various fitness functions. The run parameters are as follows: Population size=5000; Tournament size=3; Mutation probability=0.5; Mutation rate=0.2; Crossover probability=0.6; and Crossover type=Two-point.

[0056] Fig. 21 illustrates the evolution curve of a Darwin run against NPR1.

[0057] Fig. 22 illustrates evolving bioactive anti-SMYD3 peptides with the method of an embodiment of the present invention. (A) Evolution curve for a Darwin run against SMYD3 showing how fitness evolves with each iteration. Plotted are the fitness values of the fittest peptide produced at each iteration, as computed with the rank-based fitness function and the rank-based fitness function for peptide selection. The parameters used for the run. (B) Plot showing the for 96 peptides (top, bottom, and mid-scoring) selected from the peptides evolved in the run. (C) Characteristics for the two peptides with the highest anti-SMYD3 inhibitory activity. Note that inhibition >100% may be explained by the fact that SMYD3 automethylates, and the decrease in methylation signal includes reduced methylation of the substrate and SMYD3.

[0058] Fig. 23 illustrates evolving an anti-SMYD3 peptide with the method of an embodiment of the present invention. (A) One-to-all curve showing the C1 peptide-protein SPRINT interaction scores for all proteins. The target protein, SMYD3, is plotted as a solid circle. (B) Ten most thermodynamically favoured conformations of the fittest peptide produced in the Darwin run against SMYD3. The peptide was docked in a SMYD3 crystal structure (PDB: 7o2c; ligand removed) using the HPEPDOCK server (Zhou et al. 2018). (C) The residues withthe highest contributions to the SPRINT interaction score between C1 and SMYD3 are highlighted in gray. The contribution of each residue along the length of SMYD3 is also shown below, with residues contributing most to the score being lighter in color. (D) Methyltransferase activity of SMYD3 on MAP3K2 substrate with different truncations of the C1 peptide, assessed with the MTase glow assay. The first control corresponds to measurements in absence of C1. The second control corresponds to measurements in absence of SMYD3. The remaining bars show the methyltransferase activity of SMYD3 towards MAP3K2 for different truncations of the C1 peptide. Each bar corresponds to a truncation with 2 additional residues removed from the C-terminus, starting from the full 20 a. a. sequence.

[0059] Fig. 24 illustrates runs of the method of an embodiment of the present invention against the SMYD1 lysine methyltransferase. Evolution curves for 10 runs against KMT SMYD1, a protein sharing high homology with SMYD3 for which a peptide binder was successfully designed. All 10 runs converged towards low fitness values, when compared to the values obtained with the successful run against SMYD3.

[0060] Fig. 25 illustrates a potential issue with the fitness function of an embodiment of the method of the present invention. In the scenario on the left, the peptide is predicted to interact with the target and several other off-target proteins. In this formulation, this produces a nonzero fitness value, because the maximum off-target protein scores below the threshold T. In the scenario on the right, the peptide is predicted to interact with the target with the same probability and only one off-target protein, but is assigned a fitness value of zero, because the off-target protein’s score is too high. In practice, the peptide that yields the one-to-all curve on the right is more desirable, even if it has very few (only one) off-target interaction.

[0061] Fig. 26 provides a schematic representation of the search strategy in a latent space. In this strategy which builds on top of the method of an embodiment of the present invention, new peptides will be optimized in a latent space spanned by models trained to capture physicochemical, functional and structural information encoded in protein / peptide sequences and scored in the sequence space with our peptide scoring module.

[0062] Fig. 27 is an example computing device utilized by the proposed system for peptide evolution and selection.

[0063] New peptides are optimized in a latent space spanned by models trained to capture physicochemical, functional and structural information encoded in protein / peptide sequences and scored in the sequence space with a peptide scoring module.DETAILED DESCRIPTION

[0064] The following disclosure relies on the phrases, terms, symbols, and notations as defined below:• application programming interface (API): software interface that exposes services (e.g.functions, classes, etc.) to other pieces of software;• area under the precision-recall curve (AUPRC): performance metric used to assess binary classification models that captures the trade-off between precision and recall over the range of cut-off values;• cell-penetrating peptide (CPP): short sequence of residues (peptide) that is conjugated to a molecule to facilitate its penetration into the cell;• class imbalance: challenge in classification tasks caused by the predominance of certain classes of instance over others;• command line interface (CLI): interfacethat allows a user to interact with software using the command line;• convolutional neural network (CNN): popular neural network architecture widely used for computer vision-related tasks that relies heavily on convolution operations and parameter sharing;• de novo process by which a model trained on one distribution for one task is bootstrapped and adapted to complete a related task on a different, but related distribution;• domain adaptation: process by which a model trained on one distribution for one task is bootstrapped and adapted to complete a related task on a different, but related distribution;• evidence tetrad: set of 4 proteins comprising two proteins known to interact, and two query proteins where one of the query proteins shares a high-scoring segment pair with one of the interactors and the other query protein shares a high-scoring segment pair with the other interactor (evidence supporting a potential interaction between the query proteins);• evolution curve: curve displaying the fitness of the fittest peptide in the peptide pool as a function of the generation;• evolutionary algorithm (EA): family of optimization algorithms that are inspired by evolution principles or replicate biological or ecological phenomena;• exposon: protein that could potentially engage in a physical interaction with a peptide;• exposome: set of all exposons;• fitness: a function of predicted interaction scores with the target protein and off-target proteins;• i score: metric used to assess binary classifiers which combines recall and precision by taking their harmonic mean;• genetic algorithm (GA): evolutionary algorithm inspired by the principles of evolution that iteratively generates new candidate solutions with mutation, crossover and selection operations with the aim of maximizing the quality (“fitness”) fitness of solutions;• ground truth: the correct label or target value associated with an instance in a dataset;• high-scoring segment pair (HSP): a pair of protein subsequences whose similarity score computed with a substitution matrix meets a pre-defined threshold;• In Silica Peptide Synthesizer (InSiPS): a target-conditioned, sequence-based peptide design algorithm built on top of PIPE;• interactome: the set of all interactions in a given proteome or across proteomes;• interactor: protein in the training set that is known to engage in a physical interaction;• large language model (LLM): large neural networks built from transformer layers and heavily reliant on the attention mechanism;• long short-term memory (LSTM): recurrent neural network architecture used to process sequential data such as time series, text or biological sequences;• mass spectrometry (MS): technique used in biochemistry to compute the mass-to- charge ratio of ions generated from a sample (e.g. cell lysate, mixture of proteins) to analyze its molecular contents methyllysine: mono-, di-, ortri-methylated lysine residue in a protein;• methyllysome: set of all methyllysines in a proteome (also referred to as “methyllysine proteome”);• multilayer perceptron (MLP): simplest form of neural network consisting of sequential fully connected layers;• operating threshold: in the context of binary classification, corresponds to the score or probability threshold above which an instance is classified as positive and below which an instance is classified as negative;• Peptide Binder Design Algorithm via Masked Language Modeling (PepMLM):sequence-based, protein language model developed by fine-tuning the ESM-2 protein language model with a dataset of known peptide-protein interacting pairs and a masked language modeling objective for target-conditioned peptide design;• Peptide Optimization with Optimal Learning (POOL): Bayesian Optimization-based hybrid computational- / ?) vitro method for specific / orthogonal peptide design;• Peptide Prioritization via CLIP (PepPrepCLIP): target-conditioned, sequence-based peptide prioritization algorithm that identifies potential peptide binders with protein language model trained with a contrastive learning loss;• peptide-protein interaction (PepPI): interaction between a peptide and a protein;• Peptide scoring module (PSM): Darwin component inspired by SPRINT and optimized for sequence-based peptide binder fitness assessment;• post-translational modifications (PTM): chemical modification to the sidechain of a residue within a protein that occurs after translation;• pre-training: deep learning technique where large models are trained once on large datasets to learn effective representations, which allows subsequent users to bootstrap these representations and finetune them fortheir own problem (knowledge transfer);• precision: metric used to report how often a binary classifier is correct when reporting an instance as belonging to the positive class;• precision-recall curve (PRC): plot displaying the precision and recall of a binary classifier as a function of the probability decision threshold used to assign an instance to one of the two classes;• protein language model (pLM): large language model neural network architecture trained specifically to learn representations for proteins and solve various downstream protein-related problems (function prediction, property prediction, etc.);• protein-protein interaction (PPI): (physical) interaction between two proteins;• Protein-Protein Interaction Prediction Engine (PIPE): first sequence protein-protein interaction prediction algorithm capable of computing the entire human interactome; developed at Carleton University proteome: set of all proteins in an organism;• recall: metric used to report how many of the actual positive instances are classified as such by a binary classifier;• recurrent neural network (RNN): family of neural network architectures designed to process sequential data such as time series, text, or biological sequences;• Scoring Protein INTeractions (SPRINT): fast sequence-based protein-protein interaction predictor;• self-supervised learning: machine learning paradigm where the supervisory signal used during training is derived from the data itself rather than from external or humangenerated labels;• single instruction, multiple data (SIMD): parallel processing technique that leverages the ability of certain processors to act on multiple data points with a single instruction dispatched to multiple processing elements; not to be confused with concurrency where the instructions need not be identical;• specificity: metric used to assess binary classifiers that represents the fraction of actually negative instances classified as such;• state of the art (SOTA): in the context of this proposed thesis, an algorithm is said to achieve state-of-the-art performance if its accuracy and / or computational efficiencysurpass that of other algorithms supervised learning: machine learning paradigm where a function mapping inputs to outputs is learned from a dataset where the desired output associated with each input is known;• training set: set of protein pairs that have been experimentally demonstrated to engage in physical interactions; used for scoring protein-protein and peptide-protein interactions; and• unsupervised learning: machine learning paradigm where models learn about the data’s internal structure without labels.

[0065] Symbols and notations used throughout the disclosure have the following meaning: General notation(■)aReference to an object a(■ja bReference to a pair consisting of objects a and b)kElement at position k in an object (sequence)0m:nElements m to n, inclusively, of a sequence object| ■ | Size of a set or length of a sequence, depending on the operand (■)mxnGeneric matrix of dimensions m x nProtein-related objectsP A proteinp A peptide (sequence of < 50 amino acids)e Exposon; a protein or peptide which may come in physical contact with the protein or peptide under considerationI Interactor; i.e. a protein or peptide known to engage in an interaction with itself or another protein or peptideh A high-scoring segment pair; a pair of similar amino acid windows t A 2-tuple of interactorsw An amino acid window within a protein i.e. a subsequence)04 - Ii ~ I2~ OB Evidence tetrad providing evidence for an interaction between proteins / peptides A and B; A and interactorare in a high-scoring segment pair, while / 2, which is known to interact with / 2, is in a high- scoring segment pair with BSetsP Set of proteinsSet of peptides£ Exposome; set of all exposons7 Set of interactorstK Set of high-scoring segment pairsT Set of training pairsS Set of interaction scoresMatricesBLOSUM62 BLOSUM62 substitution matrixPAM120 PAM 120 substitution matrixS Interaction score matrixOtherss Interaction scoressimSubtitution matrix-derived similarity score of two sequences of amino acidsc Contribution of an evidence tetrad to the interaction score of a pair T Threshold

[0066] Embodiments disclosed herein provide a method and system for in silico sequencebased design of peptides for binding to a target protein. The method of the present disclosure utilizes a genetic algorithm-based optimization strategy to “evolve” peptide binders while taking into account the probability that designed peptides will bind to off-target proteins, which is one. In certain embodiments, the method outputs potential peptide binders along with on-target and off-target interaction scores. In certain embodiments, the method requires substantially less computing resources than alternatives, such as InSiPS.

[0067] In certain embodiments of the present invention, there is provided an in silico method of designing peptides that bind a target, the method comprising:(a) providing an initial in silico pool of parent peptides;(b) subjecting the in silico pool of parent peptides to in silico mutation and crossover to produce an in silico pool of offspring peptides;(c) determining the interaction (fitness) score of each of the offspring peptides with a target; (d) selecting the offspring peptides with the highest interaction (fitness) scores to generate a new in silico pool of parent peptides;(e) repeating steps (b) to (d) until the interaction (fitness) scores cease to increase; and (f) selecting peptides with the highest interaction (fitness) scores as potential binders of the target.

[0068] In certain embodiments, there is provided an in silico method of designing peptides that bind to a target protein. In certain embodiments, the method comprises: a) providing a pool of parent peptides each having an amino acid sequence; b) subjecting the amino acidsequences to in silico mutation and crossover to produce an in silico pool of offspring peptides; c) determining an interaction score for each offspring peptide based on their sequence similarity to known interactors of the target protein or homologs thereof; d) determining a calculated fitness value based on the interaction score that quantifies the likelihood of interaction between each offspring peptide and the target protein relative to interaction between each offspring peptide and off-target proteins, and selecting offspring peptides with the highest calculated fitness value to generate a new in silico pool of parent peptides; e) repeating steps (b) to (d) until the interaction scores cease to increase; and f) selecting one or more offspring peptides with the highest fitness values as candidate peptides that bind the target protein.

[0069] In specific embodiments, the target is a protein. In specific embodiments, a pool of in silica parent peptides are provided. In other specific embodiments, a sample pool of peptides are obtained (for example, from in vitro or in vivo) and their amino acid sequences determined, which are then used as parent peptide sequences. The pool of parent peptides may be derived from a subset of a larger sample.

[0070] In some embodiments, these designed proteins are further expressed in a host cell, or chemically synthesized (for example, by solid-phase peptide synthesis (SPPS) and / or solution-phase peptide synthesis (SPS)). In some embodiments, these designed proteins are screened for activity (for example, in vivo, in vitro, or in silica assays). In some embodiments, these designed peptides are administered to patients, used in treatment in a subject in need thereof, or used in the manufacture of a medicament. In some embodiment, the designed peptides are visually presented in off / on-target plots or charts, or one-to-all curves (2D visual representation of algorithm) to illustrate their interaction with the target protein relative to other proteins.

[0071] In some embodiments, determining the calculated fitness value comprises calculating a numerical value with a rank-based fitness function configured to consider the rank of the target protein relative to a given set of possible protein interactions each offspring peptide may engage in. In one embodiment, the rank is determined based on the interaction score. The rank is dependent on the defined exposome, or the set of all proteins with which the designed peptide may interact. The exposome may be the set of all proteins in the target organism or a subset thereof. The rank may be incorporated in the measure of fitness through any computation alone or in combination.

[0072] In some embodiments, the calculated fitness value of a given offspring peptide incorporates the interaction score for said offspring peptide. In some embodiment, the numerical value of the calculated fitness reflects or is derived from interaction score. In certainembodiments of the rank-based fitness function, the on-target score for the peptide considered is incorporated in the calculation of fitness.

[0073] In some embodiments, the rank-based fitness function has a rank term for quantifying the likelihood of interaction between one offspring peptide and the target protein relative to off-target proteins. In one embodiment, the rank term is normalized.

[0074] In some embodiments, determining the calculated fitness value comprises applying a fitness penalty to offspring peptides having a rank term below a given threshold. In certain embodiments of a rank-based fitness function, a fitness penalty is applied to peptides for which the target ranks too poorly relative to off-target proteins.

[0075] Systems are also provided to carry out the methods and processes described herein. In some embodiments, a system is provided comprising a processor configured to design peptides that bind a target protein and a memory storage device. The system may also include assay kits and systems (for example for in vitro counter screening against homologues), display device, printer, protein sequencer (i.e. mass spectrometry sequencer systems, Edman degradation sequencer systems), spot synthesis protein sequencer, SPPS and / or SPS.

[0076] In some embodiments, systems are provided comprising a processor for in silico design of peptides and a peptide sequencer to integrate parent peptide sequence identification and in silico peptide design. Such systems allow for automation and / or integration of in silico peptide design from initial parent peptide sampling to peptide design output. In some embodiments, systems are provided comprising a processor for in silico design of peptides and peptide synthesizer systems (such as solid-phase peptide synthesis (SPPS) and / or solutionphase peptide synthesis (SPS), or host cell expression systems) to integrate in silico peptide design and expression or manufacture of the designed peptide. Such systems allow for automation and / or integration of in silico peptide design from initial parent peptide sampling to peptide design output. In some embodiments, systems are provided comprising a processor for in silico design of peptides and protein assay screening apparatuses (i.e. in vivo, in vitro, or in silico assays) to integrate in silico peptide design and validation, including in vivo and / or in vitro validation. Such systems allow for automation and / or integration of in silico peptide design from initial parent peptide sampling to peptide activity validation. In some embodiments, systems are provided comprising a processor for in silico design of peptides and display device for visually representing interactions between candidate proteins and the target protein to integrate in silico peptide design and result output. Such systems allow for real time and / or interactive representation of candidate proteins and the target protein interactions, level of fitness, and / or contribution of candidate protein sequence modification to target protein interactions.

[0077] The following section is guided by a sequence of steps ((a) to (f)) making up an in silico method of designing peptides that bind to a target protein, according to some embodiments.(a) providing an initial in silico pool of parent peptides

[0078] Parent peptides may be randomly generated. Parent peptides may be generated based known interactors of the protein target or homologs thereof. Parent peptides may be generated by first mapping to clinical outcomes. Accordingly, in certain embodiments of the present invention, providing an initial pool of parent peptides comprises randomly generating peptides. In some embodiments, parent peptide sequences are generated in silico. In other embodiments, parent peptide sequences are obtained by protein sequencing by mass spectrometry, Edman degradation, de novo peptide sequencing, for example.

[0079] In certain embodiments, the pool is optimized to reduce the likelihood that the pool comprises peptides that are nonfunctional, have completely undesirable properties or to be inconsistent with the “grammar” of protein sequences. Exemplary approaches to build an optimized pool of peptides include but are not limited to:1. sampling a variety of naturalistic peptides: Generative models are publicly available that were trained to model the sequence distribution of large protein databases, such as ProtGPT2, which can generate an arbitrary number of random protein sequences that follow the “grammar” of proteins and produce naturalistic peptides. Accordingly, in certain embodiments, the initial parent pool is generated using a model, such as ProtGPT2, that generates naturalistic peptides; or2. using subsequences of known interacting proteins: PPI prediction methods such as SPRINT and PIPE rely on the fundamental assumption that protein-protein interactions are mediated by windows of contiguous amino acids. Accordingly, a person skilled in the art would regard building an initial population from subsequences of known interacting proteins as reasonable. Such an approach is used in the SaLT& PepPr peptide prioritization approach to binder design (Brixi et al. 2023). Accordingly, in certain embodiments, the initial parent pool is generated using subsequences of proteins known to interact with the target of interest.

[0080] In certain embodiments, the generated peptides are subject to cluster analysis. Clustering peptides based on sequence identity is known in the art. For example, CD-HIT is a publicly available program for clustering and comparing protein or nucleotide sequences. In certain embodiments, the peptides are clustered into Npopclusters and a cluster representative from each cluster is used to form the initial pool of parent peptides. In certain embodiment, theclustered sequences have an identity threshold of 50%, 60%, 70%, 80%, 90% or 95% to other sequences in the cluster.

[0081] Accordingly, in certain embodiments of the present invention, providing an initial in silica pool of parent peptides comprises (i) generating a population of peptides; (ii) clustering the peptides into Npopclusters based on sequence identity; and selecting a representative peptide from each cluster to produce the initial pool of parent peptides.(b) subjecting the in silica pool of parent peptides to in silico mutation and / or crossover to produce an in silico pool of offspring peptides

[0082] New offspring peptides may be generated in silico with operators which mimic genetic events in biological populations: mutation and / or crossover. Computational manipulation of strings representative of organic material is utilized to simulate the effects of evolutionary events that would otherwise take up an impractically long time under uncontrollable circumstances.

[0083] Mutation replaces one or more discrete residues within a peptide sequence with one or more other discrete residues randomly. The peptides in the pool to be mutated may be selected randomly with a probability set using a mutation probability. Residues within peptides selected for mutation may then be subsequently mutated with a probability given by the mutation rate. In certain embodiments of the method, residues are mutated randomly, i.e. by sampling the new residue from a uniform distribution over all 20 amino acids. In such cases, there is therefore a 5% chance that a residue remains unchanged even if it selected for mutation.

[0084] With crossover, the sequences from two parent peptides are recombined to produce a new peptide sequence. The peptides in the pool will recombine may be selected randomly with a probability set using the crossover probability. Two common crossover operators are the one-point and two-points crossover. In certain embodiments of the method, two-point crossover is used.(c) determining the interaction (fitness) score of each of the offspring peptides with a target;

[0085] The majority of modern PPI predictors fall under the categories of sequence-based and structure- based methods. In certain embodiments of the present method, the interaction (fitness) score is determined using sequence-based methods. Sequence-based methods may include but are not limited to protein-protein interaction prediction with deep learning or with similarity-based methods.

[0086] Exemplary known similarity-based methods include PIPE and SPRINT. PIPE and SPRINT require protein sequences only and rely on the fundamental assumption that interactions between proteins are mediated by windows (subsequences) of consecutive amino acids. They consider the presence of similar window pairs in query protein pairs and known interacting pairs as evidence for an interaction in query protein pairs. The more similarities between a query pair and known interactors, the more evidence there is to support a potential physical interaction. Similarity of windows (subsequences) maybe scored using a substitution matrix such as PAM matrix or BLOSUM matrix).

[0087] Described herein is an optimized process for peptide design to produce the peptide scoring module while tracking the contribution of individual residues within the peptide / protein to contribute to interaction score. This has been implemented in the peptide scoring module.

[0088] Accordingly, in certain embodiments of the proposed methods, a similarity-based sequence-based method of determining interaction with a target is utilized. In more specific embodiments, evidence of peptide interaction with a target is determined by using a similaritybased sequence-based method comprising the steps of extraction of all s-mers; identification of HSPs; correction for frequent windows; and interaction scoring. In more specific embodiments, the similarity-based sequence-based method is SPRINT or a proposed peptide scoring module Darwin.

[0089] In some embodiments, determining the interaction score comprises: i) extracting all s-mers occurring in the sequence of each offspring peptide, based one or more space seeds of a given length; ii) identifying high-scoring segment pairs (HSPs) based on a predetermined threshold; iii) correcting for frequent windows by removing residues that are involved in more than a given number of HSPs; and iv) scoring the interaction of each offspring peptide and all possible targets with a training set of known protein-protein interactions. In one embodiment, the method further comprises identifying a potential binding region of the target protein based on the contributions of the HSPs to the interaction score, by calculating an individual residue contribution for each amino acid residue of the candidate peptides using a set of four proteins comprising two proteins known to interact and two query proteins, where one of the query proteins shares a HSP with one of the known interacting proteins and the other query protein shares a HSP with the other known interacting protein, and wherein amino acid residues having higher individual residue contribution values are identified as having higher likelihood of comprising the binding region.

[0090] In some embodiments, calculating the individual residue contribution comprises first calculating contribution of the set of four proteins (c(he, hp)) based on formula:She(|hp| - k + l)sh(| / te| - k + 1)C\ h-e> J -|hp||he|where the set of four proteins has one exposon-interactor HSP (he) and one peptide-interactor HSP (hp),shis a substitution matrix-derived similarity score of the HSP, and |h| is the length of the HSP.

[0091] In some embodiments, calculating the individual residue contribution comprises deriving a mean per-residue contribution (c(he, hp)) from the c(he, hp) based on formula:(4)and adding said mean per-residue contribution to the individual residue contribution.In some embodiments, the rank-based fitness function ( / flFF)has the formula:ifmax(p r) < TfRFF (P' T, Sp, 8') — (1)otherwisewhere spe Spis a predicted peptide protein interaction score determined between the peptide (p) and the target protein (T), £ is the exposome considered, S£ Tis a set of all predicted interaction scores between p and all non-target proteins, hyperparameter T controls a penalty term used to constrain the search space, and rPiTis the rank of sPiTrelative to all other interactions on the exposome.

[0092] In some embodiments, electing the offspring peptides with the highest interaction scores comprise calculating the interaction score with a selection fitness function (fRFF (selection)) having the formula:fRFF (selection) (P> $p, T (2)where spe Spis the interaction score determined between the peptide (p) and the target protein (T) £ is the exposome considered, and rPiTis the rank of sPiTrelative to all other interactions on the exposome.(d) selecting the offspring peptides with the highest interaction (fitness) scores to generate a new in silico pool of parent peptides;

[0093] A variety of techniques are known in the art to preferentially select fitter individuals in genetic algorithms. Such techniques include but are not limited to Fitness ProportionateSelection; Roulette Wheel Selection; Stochastic Universal Sampling (SUS) and Tournament Selection.

[0094] In certain embodiments of the method, tournament selection is utilized. In this approach, a new pool of peptides is populated by selecting peptides of known fitness with N tournaments, where N is the population size. Each tournament involves selecting k peptides and adding the fittest of those peptides to the new pool. The parameter k is referred to as the tournament size.(e) repeating steps (b) to (d) until the interaction (fitness) scores cease to increase; and (f) selecting peptides with the highest interaction (fitness) scores as a potential binder of the target.

[0095] Table 1 below lists the genetic algorithm hyperparameters used by the proposed method, according to some embodiments.Parameter Value DescriptionPopulation size 5000 Number of peptides being evolvedTournament size 3 Number of peptides being considered when using the tournament strategy for selectionMutation probability 0.5 Probability that a peptide is selected for mutation prior to being scoredMutation rate 0.2 Probability that a residue in a peptide selected for mutation will be mutated Crossover probability 0.6 Probability that a peptide will recombine with another peptide prior to being scored Crossover type Two-point Number of points around which the subsequences are “swapped”

[0096] Certain embodiments of the proposed methods further comprise synthesizing the designed peptides. Methods of synthesizing peptides include chemical synthesis and biological methods. Chemical synthesis methods include but are not limited to liquid-phase peptide synthesis and solid-phase peptide synthesis. In addition to chemical synthesis, the designedpeptides may be prepared by various biological methods including but not limited to enzymatic synthesis and recombinant technologies, which may include both in vitro and in vivo methods.

[0097] Methods of peptide purification include but are not limited to reverse-phase chromatography (RPC); size-exclusion chromatography, ion exchange chromatography (IEC); partition chromatography and high-performance liquid chromatography (HPLC).

[0098] The in silico methods of the invention identify potential binders of targets. Accordingly, in certain embodiments of the invention, the method further comprises screening the peptides identified as potential binders for binding to target.

[0099] The methods of the present invention may be used in the development of targeting molecules and / or leader sequences and / or signal peptides and / or peptide antagonists or peptide agonist and for other uses.

[0100] Accordingly, in some embodiments, the methods further comprise screening and / or testing the peptide for activity, which includes targeting activity, ability to act as an agonist, and ability to act as an antagonist.

[0101] The present invention further comprises a system comprising a computing device configured to perform the method steps set forth above. In certain embodiments, the computing device and any databases are cloud-based.

[0102] Also provided is a computer readable memory having recorded thereon statements and instructions for execution by a computer the method steps detailed above.

[0103] FIG. 27 is an example computing device, according to various embodiments. The device 2800, such as a computer server, can be utilized to implement the proposed system for designing peptides that bind to a target protein, or the underlying methods to be executed by the system. The device 2800 includes one or more processors 2802, which operate in conjunction with computer memory 2804.

[0104] The computer memory 2804 can be adapted to store various non-transitory computer readable instruction sets, which can be executed on the processors 2802 to cause the performance of various methods described herein.

[0105] The computer memory 2804 can maintain a memory hierarchy including on-chip caches, scratchpad memory for accelerator kernels, volatile system memory for scoring matrices and peptide pools, and non-volatile storage for training sets, exposome definitions, configuration profiles, and audit logs.

[0106] An input / output interface 2806 can be utilized to receive various input commands, for example, through an attached keyboard, mouse, or touch input, and generate various outputs, such as screen renderings for display on a computer display of a computer or a mobile device. The input / output interface 2806 receives sequences and metadata via network interfaces and / or storage, including target protein sequences, curated training PPIs, exposome lists (e.g., surfaceome), and prior binders (e.g., phage display libraries). For designed peptides, the device 2800 further generates machine-readable manufacturing packages to be transmitted via input / output interface 2806 for downstream synthesis and recombinant expression.

[0107] A networking interface 2808 can be utilized to communicate one or more data sets, for example, through a message bus or other enterprise communication linkage. Networking interface 2808 can be adapted to obtain machine learning models or inputs thereof, for example, as the model validation and auditing process takes place.

[0108] The device 2800, in some embodiments, can be a special purpose machine, such as a standalone appliance, a rack-mounted unit in a data centre, or a distributed assembly of interoperable nodes communicating via a deterministic fabric, purpose-built to run peptide scoring and algorithmic genetic evolution loops.

[0109] The proposed methods can be implemented as machine-readable instructions stored on a non-transitory medium and executed by processing elements of the special purpose machine to perform PepPI scoring and Darwin optimisation. When executed, the instructions cause the machine to ingest sequences and training data; normalise and encode s-mers / embeddings; execute PepPI scoring and residue attribution; compute fitness scores and select peptides; emit candidate sets and artefacts; persist states; and maintain security / audit guarantees. The instructions may be partitioned into firmware resident in immutable memory and updatable software images cryptographically verified at load time.

[0110] To gain a better understanding of the embodiments described herein, the following examples are set forth. It will be understood that the examples are intended to describe illustrative embodiments of the proposed systems and methods in a non-limiting fashion.EXAMPLES

[0111] The proposed peptide scoring module Darwin relies on a set of PPIs to score the interactions between the peptide candidates and the exposome and evaluate their fitness. This “training set” is carefully assembled so as to detect PepPls interactions while minimizing false positives.

[0112] Using PIPE to score PPIs, Dickefa / . showed that filtering for high-quality interaction data is beneficial for prediction accuracy of PPIs in humans, mice, yeast, and arabidopsis, even if it results in a smaller training set. In fact, meaningful improvements in Re@0.5Pr (up to 22%) were observed for dataset sizes up to 14 times as small after careful curation.

[0113] Building on top of these observations, a training set for Darwin is designed by applying filters to the comprehensive set of pairs of interacting human proteins in the BioGRID database (Oughtred et al. 2021). More specifically, only physical interactions whose detection methods were annotated were retained as one of the following detection methods:• “Two-hybrid”;• “Affinity Capture-MS”;• “Affinity Capture- Western”;• “Reconstituted Complex”;• “Affinity Capture Luminescence”;• “Co-crystal Structure”;• “Far Western”;• “FRET”;• “Protein-peptide”;• “Colocalization”;• “Affinity Capture-RNA”;• “Copurification”.

[0114] Finally, interactions without multiple lines of evidence were discarded, i.e. interactions that have been reported more than once.

[0115] Starting with an initial dataset of 705,341 human PPIs extracted from BioGRID, filters as described previously were applied to generate a final training set of 86,970 PPIs involving 12,962 unique human proteins i.e. |J| = 12,962).Peptide scoring module

[0116] A peptide scoring module (PSM) is proposed in the Rust programming language and a number of optimizations was made in order to enhance its performance and suitability for use in a peptide design where tens of millions of PepPls may need to be scored and computational efficiency is critical. Rust was selected because it is a modern statically typed programming language that compiles to highly efficient machine code that achieves speeds similar to good C / C++ code.

[0117] One key feature that Rust possesses that other fast languages such as C / C++ don’t is its unique approach to memory safety with an ownership model that detects a host of bugsincluding data races, dereferencing freed memory, dangling pointers, and more at compiletime. The ownership model is implemented through a borrow checker that ensures that variables have a sole owner with the ability to mutate them at any given time, and that tracks the lifetime of variables. The ability of the compiler to detect bugs at compile-time makes developing and maintaining Rust programs significantly easier, faster, and less error-prone.

[0118] The optimizations implemented in the PSM and their technical effects are described below in Table 2, showing optimizations and new features in Darwin’s peptide scoring module, according to some embodiments.Useful for Useful forNew feature / interactome peptide Benefit(s)change prediction scoringOptimization of s- y y • Significantly reduces the runtime of the HSP mer storage extraction stepQuantification of • Provides evidence to support the role of residue individual residues which in mediating the contribution y y predicted interaction of two predicted interactors towards interactionscoreReorganization of • Far less memory is required to store the relevant the PPI score interaction scores; allows the incurrence of a matrix y memory cost elsewhere for enhanced reproducibility of the predictionsReduction of • Predicted interaction scores are reproducible stochastic noise in y y and sufficiently reliable, with no runtime cost PPI scoresSeparation of the• Runtime improvement (elimination of redundant HSP processing computation)and PPI / PepPI yscoring steps • Enables scoring of PepPlsRemoval of y • Runtime improvement (cuts out superfluous irrelevant HSPs computation)Python API • Enables calls from Python scripts and interoperability with the Python scientific y ycomputing ecosystem(plotting, statistics, optimization, etc.) s-mer storage optimization

[0119] The authors of the original SPRINT implementation opted to store the s-mers extracted in Step 1 in a hash table data structure that they implemented manually (including the hashing function, collision handling, etc.).

[0120] In the peptide scoring module, the s-mers are stored in a highly optimized hash table data structure mapping the unsigned 64-bit integer value of an s-mer to a vector of (protein, position) tuples corresponding to occurrences of the s-mer in the provided protein sequences. The implementation of the hash table data structure in Rust uses single instruction, multiple data (SIMD) lookups, which improves the runtime cost of value access in the hash table.Identification of residue contributions to PPI scores

[0121] In contrast with PIPE, SPRINT does not track the contribution of individual residues within the proteins to the PPI / PepPI score. In fact, the authors of the SPRINT acknowledged that sequence-based PPI prediction method are especially well-suited for interaction site prediction, but did not implement this feature.

[0122] However, the implementation of this feature was sought within the peptide scoring module, positing that it may provide clues as to which residues within the target protein engage in physical interactions with an exposon (typically the target protein), i.e., suggest a potential binding site. This calculation is performed post hoc with good peptide binder candidates.

[0123] Let hpbe short form for / i'1and hebe short form for h1*. We can adapt the scoring equation in for the contribution of an evidence tetrad to the peptide binder design scenario:she(IM ~k+ l)^p(]he| - k + 1)c(hefhpj| 11 h-e Iwhere c( / ie, / ip) is the contribution c of the evidence tetrad p - l - 12- e with one exposon-interactor HSP, h1*, and one peptide-interactor HSP, h1^. The symbols s and | / t| denote the similarity score and the length of the HSP, respectively.

[0124] Once the contribution of the tetrad to the score has been calculated, its mean per-residue contribution, c( / ie, / ip), can be calculated by:and added to the contribution of the residues in e that are comprised in he.

[0125] Some embodiments of the calculation of residue contributions is further illustrated in Fig. 13.Reorganization of the PPI score matrix

[0126] During the interactome scoring step, the original implementation of SPRINT stores the interaction scores in a score matrix, SNxN, where N = \£ u J|, is the number of proteins that are in the exposome or the training set.

[0127] The way SPRINT stores scores is inefficient memory wise, because the matrix is symmetric; i.e. the interaction score between p and p2is equal to the interaction score between p2and p±(as shown in Fig. 14, pane A). The data structure was therefore constructed to be a1-D array with scores. To facilitate indexing in this 1-D array, a function was implemented to convert artificial row-column 2-D index to a scalar integer index.

[0128] For peptide binder scoring, the values of interest are the scores between peptide candidates and exposons. The binder scoring process was therefore further modified to only store scores between the peptide-exposon pairs, not between peptide pairs, exposon pairs or interactors that are not exposons (as shown in Fig. 14, pane B), as those interactions contribute no useful information. These changes, when implemented in combination, reduce space complexity from quadratic complexity O(|T’|2+ |£ uJ|2) to complexity O(|T’| |£|), therefore significantly reducing the memory and processing power required in processing, storing, and querying the interaction scores.

[0129] These optimizations are critical when operating at scale, as the algorithm requires the creation of one such matrix for every thread used for scoring in practice.Reduction of stochastic noise in predictions

[0130] In practice, the scoring step in the proposed algorithm is executed in parallel with a multithreading approach.

[0131] The original SPRINT algorithm uses a shared memory model where the score matrix is a simple 2-D array of 32-bit floats that is accessible to independent threads managed with the OpenMP library. Notably, the implementation contains no locking mechanism to prevent data races (e.g. mutexes). The model was replicated to circumvent the very high cost associated with the use of locking mechanisms in this context.

[0132] While the score matrix is large in practice, which does decrease the odds of data races occurring, some scores varied significantly between runs, meaning data races still occurred.

[0133] To address this issue, a strategy is employed which reduces stochasticity at the cost of increased memory requirements. An additional dimension is added to the score matrix, turning it into a IJ’I x \p\ x Nthreatensor. Individual threads are assigned an index, i e {0,1, —, Nthreads- 1}, and increment the scores along a single slice (corresponding to their index) along the thread axis (as shown in Fig. 15).

[0134] At the end of the scoring procedure, the scores along the thread axis are summed to collapse the tensor back into a 2-D matrix. The final score matrix is obtained as a fixed-order sum, ensuring a reproducible summation order and thereby reducing run-to-run variability induced by the non-associativity of floating-point addition.

[0135] This approach introduces a linear increase in memory requirements proportional to ^threads’asthe tensor holds Nthreacopies of the score plane. However, the memory overhead is offset by improved determinism and a simplified write path without synchronisation primitives.

[0136] To compare the stochastic noise in scores before and after implementing this strategy, the entire human interactome (20,423 proteins) was predicted 6 times using 55,788 training pairs. The first run was used as a reference and the scores generated in subsequent runs were compared with the reference. The experiments were run on a workstation equiped with a 64-core (128 threads) AMD Ryzen Threadripper PRO 3995WX and 128GB RAM.Separation of HSP processing and scoring steps

[0137] In the original SPRINT implementation, the HSP processing step aiming to correct for residues that occur within too many HSPs (Step 3 described in the background) is completed prior to, but jointly with the scoring step (Step 4 described in the background), i.e. within the same program. For peptide engineering, that is not ideal, because this HSP processing step would need to be performed every time a peptide or population of peptides is input to SPRINT for scoring.

[0138] To understand the rationale behind this modification, it is important to mention that three types of HSPs arise within the peptide engineering context: peptide-protein HSPs, protein-protein HSPs, and peptide-peptide HSPs (irrelevant for the task). The relevant types of HSPs and the relationships between the peptide / proteins involved are shown in Fig. 12.

[0139] The HSP processing step is separated from the scoring step in the proposed embodiments for two reasons:1. the processing of exposon-interactor (£u3) HSPs only needs to be completed once, because the set of HSPs extracted from £ u 7 does not change throughout the design process. Only HSPs from the sets P u J need to be computed at each iteration, as the peptides change throughout the design process. Separating the two steps addresses the aforementioned issue of redundant computation and allows the extraction and processing of all exposon-interactor HSPs once as a pre-processing step. The processed HSPs are loaded in memory at the start the optimization routine and reused throughout; and2. it is desirable to process HSPs between pairs of proteins, but not HSPs involving a peptide and a protein, because doing so drives peptide-exposon scores to 0. This likely occurs because short peptides are less likely to share HSPs with interactors because of their length. “Chopping” these few peptide-interactor HSPs dramatically reduce the number and length of peptide-interactor HSPs. Separating the HSP process for the scoring process allows for the processing HSPs involving two proteins while leaving HSPs involving peptides intact.Removal of HSPs that do not contribute towards PepPI scores

[0140] Only a subset of the pre-processed HSPs is useful in computing peptide-exposon scores. An HSP is guaranteed to not contribute towards any peptide-exposon scores if both partners do not appear in the training set or if the two partners are peptides. In fact, to be useful for scoring, the HSP must be a peptide-interactor pair or an exposon-interactor pair ( / .e. be one of the two types of HSPs in the evidence tetrad, Fig. 12). Therefore all protein-protein HSPs are discarded where none of the two proteins are in the training set. Peptide-peptide HSPs do not contribute towards peptide-protein interaction scores, and therefore, they are discarded as well. This modification to SPRINT further accelerates the scoring, because the program no longer needs to iterate through HSPs that have no bearing on PepPI scores for each training pair.Development of an API for calls from Python

[0141] The fitness evaluation of peptide candidates with the PSM is the bottleneck in the peptide design process and justifies the use of a very fast, low-level programming language. Other steps, such as generation of new peptide candidates or file I / O, are not computationally intensive. The runtime benefits of implementing these operations in a statically typed, compiled language like Rust over a scripting language such as Python are negligible and do not outweigh the superior ease of development of Python programs. Furthermore, enabling interoperability between Python and Darwin’s PSM allows utilization of the vast ecosystem of libraries available in Python for GAs / EAs (e.g. DEAP), data manipulation and statistical analyses (e.g. Pandas), and plotting (e.g. matplotlib).

[0142] For those reasons, it was elected to implement an application programming interface (API) that allows the PSM functions to be invoked from Python programs in addition to the standalone version of the PSM steps via a command line interface (CLI). This API was implemented with PyO3 which produces a shared object (.so) library file that can be imported in a Python program like any other Python package. The principal functions exposed by this PSM Python API are documented below:

[0143] The first function defined in the API allows users to extract HSPs (similarity windows) from proteins provided as an input:de f extract hsps (# Li st o f prot e ins ( I D, sequence and whethe r this s equence is new) prote ins: List [ ( P id: s t r, sequence: st r, new: bool ) ],# Score thre shold that 2 s -mers must meet to find HSPs around their occurence s t sme r: int,# Score thre shold that 2 windows must meet t o be st ored as a HSPt hit: int,# Minimum length o f an HSPkme r s ize: int,# Whethe r to only sto re the HSPs involving a new prote in s equencenew prote ins only: bool = Fals e,# Whethe r to s tore HS Ps that involve one prote in ( a prote in i s obvious ly an HSP with itse l f )include t rivial hsps: bool = True ) -> Set [ ( Pl id: st r, P2 id: s tr,P l pos: int, P2 pos: int, length: int ) ]

[0144] The second function exposed by the API processes the HSPs and splits HSPs at residues that are disproportionately frequent among HSPs. It returns a set of fewer, “chopped” HSPs.de f proce s s hsps (# Li st o f prot e ins ( I D, sequence and whethe r this s equence is new) prote ins: List [ ( P id: s t r, sequence: st r, new: bool ) ],# HS Ps to proce s s ( ext racted with ' ext ract hsps ' funct ion )hsps: Set [ ( Pl id: st r, P2 id: s t r, Pl pos: int, P2 pos: int, length: int ) ], # Occurrence s o f a re s idue within HSPs requi red to remove that res idue f rom HSPs t count: int,# Minimum length o f an HSP kme r si ze: int, ) -> Set [ ( Pl id: st r,P2 id: st r, Pl pos: int, P2 po s: int, length: int ) ]

[0145] The third relevant function exposed to Python programs is a function which scores PepPls. Its output is a dictionary where each key corresponds to the ID of one of the peptides scored, and the value is another dictionary mapping the ID of a protein partner to the SPRINT2 score of the PepPI.de f score pept ide s ( # List o f prote ins ( I D, sequence and whethe r this s equence i s new) prote ins: List [ ( P id: st r, s equence: s t r, new: bool ) ], # new = F alse # Li st o f pept ide s ( s ame form as prote ins ) pept ide s: List [ ( P id: st r, s equence: s t r, new: bool ) ], # new = T rue # Se t o f HSPs ext ract ed from prote ins only hsps: Set [ ( Pl id: st r, P2 id: s t r, Pl pos: int, P2 pos: int, length: int ) ], # Training pai rs ( known prot ein-prote in inte ract ions ) t raining pairs: List [ ( P l id: s tr, P2 id: st r ) ], # Expos ome, or prote ins against which the pept ide s should be s co red exposome: Set [ P id: s tr ], # Score thre shold that 2 s -mers must meet to find HSPs around their occurence s t sme r: int, # Score thre shold that 2 windows must meet t o be st ored as a HSP t hit: int, # Minimum length o f an HSP kme r s ize: int, # Numbe r o f thre ads t o be us ed for sco ring the pept ide -protein int e ract ions n thre ads: int, ) -> Diet [ pept ide id: st r, Diet [ prote in id:s tr, sprint score: float ] ]

[0146] Normally, new = False for proteins and new = True for peptides, because the HSPs between proteins have typically already been extracted prior to this function call and are passed to the function via the parameter.Modification to the Selection Operator to Handle Stochastic Noise in Predictions

[0147] We observed that while implementing a change to the data structure used by SPRINT to store scores drastically reduced the stochastic noise caused by the absence of a locking mechanism to prevent data races, very small variation in scores still occurred.

[0148] To address this scenario, we modified the tournament selection operator to reduce its sensitivity to small variation in PepPI scores. More specifically, when the fitness values of peptides selected in a tournament are too close, i.e. the difference in value is smaller than some threshold T (e.g. 5%), they are considered to have equal fitness and the peptide which comes first alphabetically is selected.

[0149] We found that Darwin runs became fully reproducible after implementing this additional change, using a T value of 5%, a value slightly larger than the maximum percent difference of 3.3% in PepPI scores in the previous section. This ensures that the fittest peptide is not selected because of noise.Implementation of the Method

[0150] The method of the present invention, like InSiPS uses a genetic algorithm-based optimization strategy to “evolve” peptide binders. It frames the peptide design problem as an optimization problem where the peptide space is probed so as to maximize fitness. The method of the present invention computes fitness using interaction scores output by SPRINT as opposed to those output by MP-PIPE, because the SPRINT PPI prediction algorithm produces more accurate results with much lower requirements in terms of computational resources. A high-level overview of the method is shown in Fig. 16.Results- reproducibility of PSM-computed scores

[0151] Table 3 shows the reproducibility of SPRINT and PSM-computed scores over 5 runs:SPRINT PSMMaxMax Runtime Max score percent Runtime Max score Run difference (s) difference difference (s) difference(%) (%) 1 148 35.7 337 123 2.6 3.3 2 148 40.7 1563 125 1.6 3.3 3 144 12.7 432 125 1.0 3.3 4 144 12.7 8370 125 1.2 3.3 5 144 12.7 274 125 1.2 3.3

[0152] In addition to a significant reduction in stochastic noise (as shown by the decreased score differences under PSM in Table 3), benefits on runtime were observed as well. Because of memory constraints, experiments on PSM were conducted with 80 threads as opposed to all 128 available threads, as done for the 2-D score matrix implementation, yet a -13% reduction in runtime was still observed.Scalability

[0153] An embodiment of the proposed methods was executed with various population sizes and different training set sizes. While it was observed that the method’s runtime scales linearly with the number of peptides in the population of candidate peptides subject to optimization (as shown in Fig. 18, pane A), the number of training PPI pairs only marginally impacts runtime (as shown in Fig. 18, pane B).Fitness functions

[0154] Several experiments on three protein targets (NSD1, SMYD3, METTL13) were conducted to characterize three different fitness functions.

[0155] The first is a “naive” fitness function where the objective is simply to maximize the interaction score as predicted by the PSM:f(P. T'Sp) = SP, Twhere sPiTe Spis the predicted interaction score between the peptide p and the target t.

[0156] A technical shortcoming of the above fitness function is that it does not account for off-target interactions, which must be considered for a peptide in practice. Schoenrock et al. defined the InSiPS fitness function which combines the objective of maximizing the interaction score predicted with MP-PIPE while penalizing off-target interactions:f(P. T' Sp) = Sp, TJS’maxwhere S£\Tis the set of all predicted interaction scores between peptide p and all non-target proteins. This fitness function was developed for PIPE scores which are bounded between 0 and 1, so the maximum off-target score can be normalized using the maximal theoretical fitness Smax. Thismaximum score was calculated using the proposed method where the goal was to generate a peptide with the highest possible interaction score (with any human protein). A value of 6292.6 was calculated for the Q20(polyglutamine) peptide.

[0157] To address the issue of contextual information loss due to the InSiPS fitness function only considering the maximum off-target interaction to evaluate specificity, another fitness function, the rank-based fitness function (RFF), was devised:ifmax(Sp T) < Tffi. FF (P' T, Sp, 8') —otherwisewhere T is a score cut-off hyperparameter used to constrain the search space, and rPiTis the rank of sPiTon the one-to-all curve ( / .e. when ranking all scores in Spin decreasing order), and £ is the exposome, or the set of all the proteins the peptide could physically interact with.

[0158] This RFF captures more contextual information, because it considers the rank (rp,t) of the target protein on the one-to-all curve, rather than considering the highest predicted interaction score of all non-target proteins only. It also includes a term whose aim is to constrain the design space to regions where off-target interactions are not too likely to occur. The parameter T is set to control the off-target interactions term to 140 for initial rounds of experimentation, based on the observation that the peptide Carperitide interacted with its target NPR1 with a PepPI score of roughly 150. It is reasonable to infer that a score of 150 between a peptide candidate and an off-target protein is likely to result in an actual off-target interaction, therefore a value slightly lower than 150 (r = 140) was set as a starting point for the experiment.

[0159] Evolution curves summarizing how the fitness changes with each generation are shown in Fig. 20 for the different fitness functions previously described. It was observed that the relative difficulty of the targets is the same for all three fitness functions, with NSD2 leading to higher scoring peptides, followed by SMYD3 and METTL13.

[0160] The use of RFF additionally lead to the exploration of a wider diversity of peptides. As shown in Fig. 20, pane C, for generated peptides at clustered at 70% sequence identity with CD-HIT, the peptides resulting from the RFF consistently yielded more unique clusters than those generated using the InSiPS-style fitness function and the naive fitness function. These experiments will be repeated with different initializations to determine whether this observation is robust and statistically significant. In fact, the naive fitness function converged towards Q20(peptide consisting of 20 glutamine (“Q”) residues) for all three targets.Validation- reproducing an NPR1-binding peptide

[0161] As a means of validating the proposed method, an experiment was conducted to determine whether an embodiment of the method can reproduce a known peptide binder, given a specific target. The Natriuretic peptide receptor A (NPR1; Uniprot ID P16066) was selected as a target, which is a receptor for atrial and brain natriuretic peptides. A previously FDA-approved peptide targeting the Natriuretic peptide receptor (Nesiritide) was used to treat congestive heart failure. A parameter T = 140 (This parameter was set arbitrarily based on intuition developed during the completion of other work. A score of approximately 150 was obtained for the peptide Carperitide and its target, NPR1.) and a peptide length of 32 (like Nesiritide) were used in execution. Notably, neither NPR1 nor Nesiritide were included the list of PPIs used for scoring interactions. However, the training PPI interaction list did include interactions between unprocessed B-type natriuretic peptide, of which Nesiritide is a fragment,and the NPR2 and NPR3, which are related to NPR1 (61.6% and 32.6% identity, respectively; computed with Clustal Omega).

[0162] As shown in Fig. 21, after 53 iterations, the run converged towards a peptide that shared a region of 21 amino acids with high similarity with Nesiritide. The results validated that the proposed method (or variant embodiments thereof) can generate realistic peptide binders in execution.Validation- design of a peptide binding to the lysine methyltransferase inhibitor SMYD3

[0163] SMYD3 is a KMT whose upregulation has been found to drive to the carcinogenesis of several solid tumours including breast, liver, lung, ovarian and colon cancers. SMYD3 is a prime target for inhibition, and several commercial inhibitors targeting it are available. However, none of the commercial inhibitors are peptides. Given that peptides tend to have higher specificity than small molecules, it is desirable to determine whether a peptide competitor with similar or better affinity could be developed using an embodiment of the proposed method.

[0164] Iterations of the proposed method (Darwin) was executed against SMYD3 until there were no improvements in fitness for 20 consecutive runs. The evolution parameters and resulting curve are shown in Fig. 23, pane A. The run was completed in 10.7 hours on a 10-node cluster with 402.4GHz Intel(R) Xeon(R) CPUs (E5-4650) each.

[0165] Visual inspection of the peptides generated in the final generation (after no improvement in fitness for 20 iterations) revealed that the target SMYD3 ranked first on the one-to-all curve for none of the peptides. However, several peptides with high on-target scores relative to scores observed when assessing therapeutic peptides in prior experiments and relatively high rank for SMYD3 on the one-to-all curve were assigned a fitness of 0 because of the maximum off-target score being larger than the T value set, leading to the conjecture that the parameter T may have steered the design in a good region of the peptide space, but that it may have led to an overly harsh fitness assessment.

[0166] The peptide fitness was reassessed with a revised formula for fitness which excludes the T specificity term:[rp, T]fpFF (selection) (P> T, $p, T I CIThe revised assessment sampled 87 peptides from the peptides produced in the last iteration of the previous run: the 48 peptides in the top 1 %, 3 peptides in the bottom 1 %, and 36 peptides taken at random in-between.

[0167] In vitro experiments were performed to determine whether Darwin-generated peptides targeting SMYD3 can 1) bind to it and 2) inhibit its methyltransferase activity. More specifically, methylation assays were performed with SMYD3 and MAP3K2, a known substrate of SMYD3, to determine how the addition of the peptide affects methylation levels. The inhibitory activities of the peptides selected are shown in Fig. 22, pane B. The two peptides with the highest inhibitory activity are tabulated in Fig. 22, pane C.

[0168] Examination of the one-to-all curve for the lead peptide (referred to as “C1”) revealed that it ranks highly (11th) on the one-to-all curve (as shown in Fig. 23, pane A), but it is closely followed by several non-target proteins on the curve. This suggests that it is not predicted to be highly specific and that further optimization may be necessary to enhance specificity.

[0169] The interaction between C1 and SMYD3 is predicted to occur in the substratebinding pocket of SMYD3 by the peptide docking software HPEPDOCK (as shown in Fig. 23, pane B). Interestingly, it was observed that the residues in SMYD3 that contribute most to the SPRINT interaction score (128-145, 280-297) are located close to the predicted binding site (as shown in Fig. 23, pane C), but not quite within the substrate-binding pocket.

[0170] According to the HPEPDOCK-predicted SMYD3-C1 complex, the N-terminus portion of the peptide should be responsible for the inhibitory activity (in particular, residues 1-10). It is possible that specificity of the peptide could be enhanced by modifying the portion of the peptide that is not predicted to be involved in the interaction (the C-terminus end), either through mutations or truncation. This could prevent off-target interactions which involve the C-terminus end of the peptide which is not important for SMYD3 binding. It was determined with methylation assays involving truncated versions of the C1 peptide that the peptide starts losing its inhibitory activity once more than 6 amino acids are removed from its C-terminus (as shown in Fig. 23, pane D).Difficult-to-Target Proteins

[0171] Using embodiments of the proposed method (Darwin), a high fitness peptide which binds to SMYD3 was identified. For SMYD1, a protein that belongs to the same family as SMYD3, the method converged toward low fitness values relative to SMYD3 (as shown in Fig.24).

[0172] The space of linear peptides spanned by the set of all 20 canonical amino acids grows exponentially with peptide length, and the number of synthesizable linear 20-mer peptides is on the order of 1026. Assuming that a typical run of the proposed method does not re-evaluate a peptide more than once in 500 iterations with a population size of 5,000 peptides, a run would evaluate only ~10-18% of all 20 synthesizable amino acid long peptides. Given this fact, it is difficult to ascertain whether the method failed to generate high fitness peptides for a given target because useful parts of the peptide space were left unprobed or because none of the training protein interactions contains useful information regarding how the target protein may interact. One thing that is known is that, for certain targets, the method can produce peptides that differ in sequence, but that are of similar fitness for different initializations. This suggests that the number of optimal solutions may vary from one target to another, with “easy” targets possibly having more high-scoring local optima than “difficult” proteins.

[0173] Unsurprisingly, it was also observed that the fitness of the peptides that the method converges towards is sensitive to how the initial population is initialized. This is also illustrated in Fig. 24 where the maximum fitness values of the anti-SMYD1 peptides produced from 10 different randomly generated initial peptide pools varies between 9.15 and 24.03 (62%).

[0174] In short, it was shown that Darwin can produce peptide binders for SMYD3 (and other proteins which are not discussed as part of this disclosure), but there may be ways in which the method may be improved to probe the peptide space more effectively.The rank-based fitness function is defined to have the following form:ifmax(Sp,£\T) < Tfft. FF (P' T, Sp, 8') —otherwisewhere T is a threshold on the maximum allowable off-target SPRINT score.

[0175] The constraint enforced by the parameter T = 140 had the positive effect that it prevented convergence of the algorithm towards highly promiscuous peptides such as the polyglutamine peptide. In addition, based on previous runs designing an anti-SMYD3 peptide using this fitness function, it appears that this fitness function steered the search towards desirable portions of the peptide space. On the other hand, one lesson learned in the process of designing an anti-SMYD3 peptide is that this formulation of fitness is overly sensitive to a single off-target score ( / .e. that of the highest-scoring off-target protein max(SPi£V)). Several anti-SMYD3 peptides were produced that were high-scoring and high-ranking but had a fitness of 0, because the maximum off-target predicted interaction score was above the specifiedthreshold T = 140. Setting T correctly is likely to be protein-dependent and require tuning. This issue is illustrated in Fig. 25.Optimizing peptides sequences in a latent space

[0176] With the advent of LLMs, protein researchers have increasingly been turning to transformer-based models to generate rich protein representations and optimize proteins in continuous latent spaces. A handful of approaches have been developed to design peptide binders by sampling solutions from a space generated by protein language models (pLMs). One such example is PepPrCLIP, where peptide candidates for a binder target are sampled from the latent space spanned by Facebook’s ESM-2, a pLM trained to predict protein structures from sequence. PepPrCLIP applies perturbations to known interacting peptides embedded in the ESM-2 latent space to generate similar, but non-identical peptides and combines those with a contrastive learning loss that jointly trains protein and peptide encoders to generate latent representations wherein interacting protein-peptide pairs are close and nonbinding pairs are distant. PepMLM, developed by the same group, also uses the ESM-2 pLM, but with the intent of creating peptides binders entirely by appending a masked region for the peptide at the end of the target’s sequence and training ESM-2 to re-generate the masked residues of the binder, conditioned on the target’s sequence. One limitation of these methods is that, like almost all other peptide binder design methods currently under investigation, they do not take off-target interactions into account at all.

[0177] A search of the embedded peptide space as opposed to searching the sequence space directly is proposed. Searching the embedded space with Darwin (as shown in Fig. 26) could lead to a more efficient and potentially successful searches of the peptide space. By searching the embedded space, the search would be inherently guided by the physicochemical properties of the peptides, whereas the search strategy in the sequence space that Darwin followed at this point only view peptides as sequence of letters. Darwin currently considers AAAA as equidistant to GAAG as to KAAK when the first two peptides share much higher similarity than they are to the third, since G is much more similar to A than K is to A in terms of physicochemical properties.

[0178] Optimization was investigated in the embedding spaces spanned by three different pLMs which have been open-sourced and can all handle 20-mer peptides: ESM-2; ProtT5 and PeptideBERT.

[0179] There are several ways to generate new peptide embeddings from existing ones. Two examples are detailed below:

[0180] Perturbation with Gaussian noise (mutation): For every column in the embedding, a perturbation vector of Gaussian noise following the distribution N (0, ka^'), where o- is the rowwise variance along a dimension d, is added to each column. Higher k values are associated with a higher likelihood that the perturbation will lead to an amino acid mutation. The authors of PepPrepCLIP showed that k values between 5 and 22 were associated with an edit distance of 1 (a single amino acid change) and an edit distance of 20 (completely new sequence) when applied to 20-mers embedded with ESM-2. The parameter k could be set to be inversely proportional to the fitness of the peptide, since it is unlikely that an unfit peptide will produce a fit peptide with few mutations, while it is quite possible to observe an improvement in fitness with few mutations to a fit peptide.

[0181] Swapping rows between embeddings (crossover): Each dimension (row) is believed to captures some physicochemical and evolutionary information about the sequence. Therefore, swapping blocks of rows between two embeddings could be used as a 2-D form of crossover operation. The fraction of rows being swapped is a hyperparameter associated with this operation.Exploring peptide pool initialization

[0182] In some embodiments of the Darwin implementation, the initial pool of peptide candidates is generated completely at random. It is generally believed that how a search is initialized can affect the quality of the solutions produced. While simple, the naive peptide pool initialization approach may be suboptimal. Many of the randomly generated peptides are likely to be completely nonfunctional, have completely undesirable properties or to be inconsistent with the “grammar” of protein sequences. This may cause the optimization procedures to converge towards low fitness maxima. Proper initialization is challenging, because of the vastness of the peptide space, the majority of it being of no interest.

[0183] A clustering-based strategy for initialization is proposed where the initial peptide pool is selected so as to capture as much as the diversity of the chemically-relevant peptide space as possible.

[0184] To build an initial library of ancestral peptide populations, clustered peptides are generated using two different approaches:1. Sampling a variety of naturalistic peptides: generative models can be trained to model the sequence distribution of large protein databases to generate “realistic” sequences. One example of such models is the decoder-only ProtGPT2 model which can generate sequences that are “similar yet distant to natural ones” (Ferruz et al. 2022). ProtGPT2 canproduce an arbitrary number of random protein sequences that follow the “grammar” of proteins. ProtGPT2 works with tokens that correspond to fragments of ~4 residues. Given a target sequence length and a temperature parameter, Tk, which specifies how many token options (fragments of ~4 amino acids) should be considered when autoregressively sampling the next token in the sequence, an arbitrary number of random, but realistic sequences can be generated.2. Subsequences of known interacting proteins: PPI prediction methods such as SPRINT and PIPE rely on the fundamental assumption that PPIs are mediated by windows of contiguous amino acids. It is reasonable to expect, if this assumption is taken to be true, that proteins known to engage in physical interactions are likely to contain windows of amino acids involved in interactions. For this reason, “chopping” fragments out of interacting proteins seems like a reasonable approach to building an initial population of potential peptide binders. The clustering can be conducted in the embedded spaces described previously.

[0185] The overall strategy to assemble initial populations consists of 3 main steps:1. generate a large number of peptides (e.g. one million) using the methods described above; 2. clustering the peptides into Npopclusters: achieving this may require a reduction in the dimensionality of the peptides, possibly using a combination of linear (e.g. principal component analysis) and non-linear methods (e.g. UMAP or t-SNE). Clustering large numbers of 1 024-dimensional vectors may be computationally intractable; and3. selection of cluster representatives: the peptide closest to the cluster’s centroid will be selected as a cluster representative.REFERENCESA. Schoenrock et al., “Engineering Inhibitory Proteins with InSiPS: The in-Silico Protein Synthesizer,” in Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis on - SC '15, Austin, Texas: ACM Press, 2015, pp. 1-11. doi: 10.1145 / 2807591.2807630.D. Burnside et al., “In Silico Engineering of Synthetic Binding Proteins from Random Amino Acid Sequences,” iScience, vol. 11, pp. 375-387, Jan. 2019, doi: 10.1016 / j.isci.2018.11.038.S. Pitre et al., “PIPE: A Protein-Protein Interaction Prediction Engine Based on the Re-Occurring Short Polypeptide Sequences between Known Interacting Protein Pairs,” BMC bioinformatics, vol. 7, p. 365, Jul. 2006, doi: 10.1186 / 1471-2105-7-365.S. Pitre et al., “Global Investigation of Protein-Protein Interactions in Yeast Saccharomyces Cerevisiae Using Re-Occurring Short Polypeptide Sequences,” Nucleic Acids Research, vol. 36, no. 13, pp. 4286-4294, Aug. 2008, doi: 10.1093 / nar / gkn390.K. Dick et al., “PIPE4: Fast PPI Predictor for Comprehensive Inter- and Cross-Species Interactomes,” Scientific Reports, vol. 10, no. 1, p. 1390, Dec. 2020, doi: 10.1038 / S41598-019-56895-w.Y. Li and L. Hie, “SPRINT: Ultrafast Protein-Protein Interaction Prediction of the Entire Human Interactome,” BMC Bioinformatics, vol. 18, no. 1, p. 485, Nov. 2017, doi: 10.1186 / s12859-017-1871x.R. Oughtred et al., “The BioGRID Database: A Comprehensive Biomedical Resource of Curated Protein, Genetic, and Chemical Interactions,” Protein Science, vol. 30, no.1, pp. 187-200, 2021, doi: 10.1002 / pro.3978.L. Wang et al., “Therapeutic Peptides: Current Applications and Future Directions,” Signal Transduction and Targeted Therapy, vol. 7, no. 1, p. 48, Feb. 2022, doi: 10.1038 / S41392-022-00904-4.W. Jaroszewicz, J. Morcinek-Orlowska, K. Pierzynowska, L. Gaffke, and G. Węgrzyn, “Phage Display and Other Peptide Display Technologies,” FEMS Microbiology Reviews, vol. 46, no. 2, p.fuab52, Mar. 2022, doi: 10.1093 / femsre / fuab052.P. Molek, B. Strukelj, and T. Bratkovic, “Peptide Phage Display as a Tool for Drug Discovery: Targeting Membrane Receptors,” Molecules, vol. 16, no. 1, pp. 857-887, Jan. 2011, doi: 10.3390 / molecules16010857.J. Pande, M. M. Szewczyk, and A. K. Grover, “Phage Display: Concept, Innovations, Applications and Future,” Biotechnology Advances, vol. 28, no. 6, pp. 849-858, Nov.2010, doi: 10.1016 / j.biotechadv.2010.07.004.A. Schoenrock, F. Dehne, J. R. Green, A. Golshani, and S. Pitre, “MP-PIPE: A Massively Parallel Protein-Protein Interaction Prediction Engine,” in Proceedings of the International Conference on Supercomputing - ICS '11, Tucson, Arizona, USA: ACM Press, 2011, p. 327. doi: 10.1145 / 1995896.1995946.M. Hajikarimlou et al., “A Computational Approach to Rapidly Design Peptides That Detect SARS-CoV-2 Surface Protein S,” NAR Genomics and Bioinformatics, vol. 4, no.3, p. lqac58, Jul. 2022, doi: 10.1093 / nargab / lqac058.L. Fu, et al., “CD-HIT: Accelerated for Clustering the next-Generation Sequencing Data,” Bioinformatics, vol. 28, no. 23, pp. 3150-3152, Dec. 2012, doi: 10.1093 / bioinformatics / bts565.P. Zhou, et al., “HPEPDOCK: A Web Server for Blind Peptide-Protein Docking Based on a Hierarchical Algorithm,” Nucleic Acids Research, vol. 46, no. W1, pp. W443-W450, Jul. 2018, doi: 10.1093 / nar / gky357.N. Ferruz, S. Schmidt, and B. Hocker, “ProtGPT2 Is a Deep Unsupervised Language Model for Protein Design,” Nature Communications, vol. 13, no. 1, p. 4348, Jul. 2022, doi: 10.1038 / s41467-022-32007-7.G. Brixi etal., “SaLT& PepPr Is an Interface-Predicting Language Model for Designing PeptideGuided Protein Degraders,” Communications Biology, vol. 6, no. 1, p. 1081, Oct. 2023, doi: 10.1038 / S42003-023-05464-Z.Although the invention has been described with reference to certain specific embodiments, various modifications thereof will be apparent to those skilled in the art without departing from the spirit and scope of the invention. All such modifications as would be apparent to one skilled in the art are intended to be included within the scope of the following claims.

Claims

1. WE CLAIM:

1. An in silico method of designing peptides that bind a target protein, the method comprises:a) providing a pool of parent peptides each having an amino acid sequence;b) subjecting the parent peptides to in silico mutation and crossover to produce an in silico pool of offspring peptides;c) determining an interaction score for each offspring peptide based on their sequence similarity to known interactors of the target protein or homologs thereof;d) determining a calculated fitness value based on the interaction score that quantifies the likelihood of interaction between each offspring peptide and the target protein relative to interaction between each offspring peptide and off-target proteins, and selecting offspring peptides with the highest calculated fitness value to generate a new in silico pool of parent peptides;e) repeating steps (b) to (d) until the interaction scores cease to increase;andf) selecting one or more offspring peptides with the highest fitness values as candidate peptides that bind the target protein.

2. The method of claim 1, wherein determining the calculated fitness value comprises calculating a numerical value with a rank-based fitness function configured to consider the rank of the target protein relative to a given set of possible protein interactions each offspring peptide may engage in.

3. The method of claim 2, wherein the rank is determined based on the interaction score.

4. The method of claim 2, wherein the calculated fitness value of a given offspring peptide incorporates the interaction score for said offspring peptide.

5. The method of claim 2, wherein the rank-based fitness function has a rank term for quantifying the likelihood of interaction between one offspring peptide and the target protein relative to off-target proteins.

6. The method of claim 5, wherein determining the calculated fitness value comprises applying a fitness penalty to offspring peptides having a rank term below a given threshold.

7. The method of claim 2, wherein the rank-based fitness function (fRFF) has the formula:ifmax(^g\T) < T fRFF (p> T • $p> — (1)otherwise wherein:sp Te Spis a predicted peptide protein interaction score determined between the peptide (p) and the target protein (T),8 is the exposome considered,S£\Tis a set of all predicted interaction scores between p and all non-target proteins,hyperparameter T controls a penalty term used to constrain the search space, andrp Tis the rank of sPiTrelative to all other interactions on the exposome.

8. The method of claim 1, wherein selecting the offspring peptides with the highest interaction scores comprise calculating the interaction score with a selection fitness function (fRFF^selection)) having the formula:fRFF (selection)CP > T, 8) = Sp T[1 — (2)wherein:sPiTe Spis the interaction score determined between the peptide (p) and the target protein (T),8 is the exposome considered, andrp Tis the rank of sP: Trelative to all other interactions on the exposome.

9. The method of claim 1, wherein determining the interaction score comprises:i) extracting all s-mers occurring in the sequence of each offspring peptide, based one or more space seeds of a given length;ii) identifying high-scoring segment pairs (HSPs) based on a predetermined threshold;iii) correcting for frequent windows by removing residues that are involved in more than a given number of HSPs; andiv) scoring the interaction of each offspring peptide and all possible targets with a training set of known protein-protein interactions; andwherein the method further comprises identifying a potential binding region of the target protein based on the contributions of the HSPs to the interaction score, by calculating an individual residue contribution (c) for each amino acid residue of the candidate peptides using a set of four proteins comprising two proteins known to interact and two query proteins, where one of the query proteins shares a HSP with one of the known interacting proteins and the other query protein shares a HSP with the other known interacting protein, and wherein amino acid residues having higher individual residue contribution values are identified as having higher likelihood of comprising the binding region.

10. The method of claim 9, wherein calculating the individual residue contribution comprises first calculating contribution of the set of four proteins (c(he, hp)) based on formula:she(M -fc + l)sh(|he| - k + 1) c(hefhpj -|hp||he| where the set of four proteins has one exposon-interactor HSP (he) and one peptide-interactor HSP (hp), shis a substitution matrix-derived similarity score of the HSP, and | / t| is the length of the HSP.

11. The method of claim 9, wherein calculating the individual residue contribution comprises deriving a mean per-residue contribution (c(he, hp)) from the c(he, hp) based on formula:= (4)and adding said mean per-residue contribution to the individual residue contribution.

12. A system comprising one or more processors operating in conjunction with a memory storage device, the one or more processors configured to perform the method of any one of claims 1 to 11.

13. The system of claim 12, wherein the memory storage device stores data representing s-mers as unsigned 64-bit integers in a hash table optimized for single instruction multiple data (SIMD) operations.

14. The system of claim 13, wherein the data representing s-mers further comprise corresponding location data in the exposome and / or known interactors.

15. The system of claim 12, wherein only peptide interaction scores between offspring peptide and protein target pairs are stored in the memory storage device.

16. The system of claim 12, wherein the memory storage device comprises a shared memory module configured to store the interaction scores as a score matrix, the score matrix comprising a 1-D array of floats and a second dimension of a plurality of threads tensor.

17. The system of claim 16, wherein the plurality of threads are assigned an index and increment the interaction scores along a single slice along the threads axis, and wherein the shared memory module is configured to collapse the tensor back into a 2-D matrix by summing the interaction scores along the threads axis.

18. A non-transitory computer readable medium storing machine interpretable instructions, which when executed by a processor, cause the processor to perform the method of any one of claims 1 to 11.

19. The system of any one of claims 12-17, further comprising one or more of: a) a display device for visually representing interactions between candidate proteins and the target protein, b) a peptide synthesizer, c) a protein assay screening apparatus, and d) a peptide sequencer.

20. The system of claim 19, wherein the peptide synthesizer is a solid-phase peptide synthesis (SPPS) apparatus and / or solution-phase peptide synthesis (SPS) apparatus, and wherein the display device displays the interactions with off / on-target plots and / or or one-to-all curves.

Images