Methods and related aspects for predicting antibody binding

Abstract

Provided herein are computer-implemented methods of predicting a biomolecule sequence to which a target antibody binds. In some embodiments, the methods include passing at least one target antibody data set through a model that relates antibody binding to biomolecule sequence, wherein the model was created using one or more antibody-biomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library, and wherein the target antibody data set is representative of one or more attributes of the target antibody and is sufficient for the model to produce a prediction of one or more biomolecule sequences to which the target antibody binds, and outputting from the model one or more predicted cognate epitope sequences and / or one or more predicted off-target biomolecule sequences to which the target antibody binds. Related systems, computer readable media, and additional methods are also provided.

Description

METHODS AND RELATED ASPECTS FOR PREDICTING ANTIBODY BINDINGCross-Reference to Related Applications

[0001] This application claims priority to, and the benefit of, U.S. Provisional Patent Application Ser. No. 63 / 593,187, filed October 25, 2023, the disclosure of which is incorporated herein by reference.Sequence Listing

[0002] The instant application contains a Sequence Listing which has been submitted electronically in XML format and is hereby incorporated by reference in its entirety. Said XML copy, created on October 17, 2024, is named 0391_0093- PCT SL.xml and is 291 ,081 bytes in size.Field

[0003] This disclosure relates generally to machine learning, e.g., in the context of medical applications, such as diagnostics and pharmacology.Background

[0004] Monoclonal antibodies (mAbs) have broad applications in research, diagnostics and therapeutics. Currently there are more than 150 approved mAbs in use therapeutically or under review and many more involved in diagnostic tests and as research reagents. The key quality of mAbs that distinguish them from other classes of proteins is their ability to simultaneously achieve high affinity and specificity. mAbs often have dissociation constants of tens to hundreds of picomolar with very little off- target binding. Binding typically involves interactions between the mAb and 5-8 amino acids and in a specific spatial arrangement. This can be the particular order of amino acids in a linear sequence (continuous epitopes) or their arrangement at the surface of a protein, potentially from multiple noncontiguous regions (discontinuous epitopes). The multiple interactions at defined positions is what affords the mAb both high affinity and high specificity. The complexity of these interactions has made predictive modeling of mAb binding a challenging problem.

[0005] Many experimental and computational tools exist to explore mAb binding interactions. X-ray crystallography can provide the highest structural resolution ofantibody-antigen binding, but is not suited for exploring weaker binding interactions or very large numbers of mAb-antigen pairs. Cross-linking mass spectrometry can provide insight into off-target binding interactions, but lacks the resolution and sensitivity to quantitate relative binding affinities. Peptide microarrays and bead libraries consisting of tiled protein or whole proteome sequences can be used to determine binding sites of antibodies, but are generally specific to the sequences tiled and, for whole proteomes, may require either very large, or rather specialized, libraries, though generalized approaches have also been developed. Molecular docking is a useful computational tool to estimate antibody-antigen binding sites but can be computationally expensive when screening a mAb against a large library of potential targets and usually does not take advantage of empirical binding data. Epitope binding sites can be predicted by modeling epitope databases or antibodyantigen structures, but these predictions are generally not specific to a given mAb. An important application of these approaches is the application to off-target binding, a significant issue particularly with regard to therapeutic mAbs. While the approaches described above provide important insight into the specific binding interactions mAbs engage in, most face practical limitations in scaling to large numbers of mAbs and large numbers of potential off-targets.

[0006] Accordingly, there is a need for additional methods of determining or predicting antigen and epitope targets of antibody binding.Summary

[0007] The present disclosure provides, in certain aspects, an artificial intelligence (Al) system capable of predicting antigen and epitope targets of antibody binding. Applications of these methods include determining the target of antibody binding in a particular protein, individual proteome, or group of proteomes. Additionally, these methods can be used for characterizing a binding site of an antibody sample and / or a potential off target site. These and other aspects will be apparent upon a complete review of the present disclosure, including the accompanying figures.

[0008] According to various embodiments, a computer-implemented method of predicting a biomolecule sequence to which at least one target antibody binds ispresented. The method includes: passing at least one target antibody data set through a model that relates antibody binding to biomolecule sequence, wherein the model was created using one or more antibody-biomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library, and wherein the target antibody data set is representative of one or more attributes of the target antibody and is sufficient for the model to produce a prediction of one or more biomolecule sequences to which the target antibody binds; and, outputting from the model one or more predicted cognate epitope sequences and / or one or more predicted off-target biomolecule sequences to which the target antibody binds, thereby predicting the biomolecule sequence to which the target antibody binds.

[0009] Various optional features of the above embodiments include the following. The biomolecule sequences in the biomolecule library are substantially random. The biomolecule sequences in the biomolecule library are substantially nonrandom. The target antibody data set and / or the antibody-biomolecule binding data sets comprise structural, chemical, and / or physical property data of the target antibody and / or of the biomolecules in the biomolecule library. The target antibody data set and / or the antibody-biomolecule binding data sets comprise amino acid sequence information from the target antibody and / or from the biomolecules in the biomolecule library. The target antibody comprises a monoclonal antibody. The target antibody comprises a polyclonal antibody. The method comprises passing multiple target antibody data sets through the model. The target antibody is from a sample obtained from a subject. The sample comprises blood or serum. In some embodiments, antibodies (e.g., mAbs, polyclonal Abs, etc.) are produced and isolated or purified. A trained electronic neural network comprises the model. The electronic neural network was trained after weighting one or more biomolecule sequences in the antibody- biomolecule binding data sets. A given amino acid sequence in the biomolecule library is disposed at different positions in a larger frame register of possible amino acid residue positions such that the model was trained to recognize the given amino acid sequence as a sequence motif in multiple frame registers.

[0010] According to various embodiments, a method of generating a model that predicts targets of antibody binding is presented. The method includes determiningbinding affinity values of one or more reference antibodies to biomolecules in a biomolecule library to produce one or more antibody- biomolecule binding data sets; and, creating a model that relates antibody binding to biomolecule sequence using the antibody-biomolecule binding data sets, thereby generating the model that predicts targets of antibody binding.

[0011] Various optional features of the above embodiments include the following. The reference antibodies comprise at least one monoclonal antibody. The reference antibodies comprise at least one polyclonal antibody. At least one of the biomolecules comprises a potential continuous epitope. At least one of the biomolecules comprises a potential discontinuous epitope. The biomolecule library comprises at least about 100 biomolecules, at least about 1000 biomolecules, at least about 10000 biomolecules, at least about 100000 biomolecules, at least about 1000000 biomolecules, at least about 10000000 biomolecules, or more biomolecules. Antigen molecules comprise the biomolecules in the biomolecule library. The biomolecules comprise at least portions of carbohydrate molecules, nucleic acid molecules, protein molecules, or a combination of two or more thereof. The biomolecules comprise amino acid sequences. The amino acid sequences comprise natural amino acids and / or nonnatural amino acids. The amino acid sequences comprise selected sequence lengths and wherein the biomolecule library comprises a selected level of coverage of a combinatorial space that comprises substantially all amino acid sequence permutations of the sequence lengths. The selected level of coverage comprises less than about 5%, less than about 2.5%, less than about 0.5%, less than about 0.25%, less than about 0.05%, less than about 0.025%, or less of the combinatorial space that comprises substantially all of the amino acid sequence permutations of the sequence lengths. The biomolecule library comprises a substantially random sample of the amino acid sequence permutations in the combinatorial space. The selected sequence lengths are in the range of about 3-15 amino acids, about 4-14 amino acids, about 5-13 amino acids, about 6-12 amino acids, about 7-1 1 amino acids, or about 8-10 amino acids. The model relates the binding affinity values of the reference antibodies to amino acid sequences in each of the antibody-biomolecule binding data sets. The method comprises using one or more markers to designate an N-terminus and / or a C-terminus of one or more of the aminoacid sequences. A given amino acid sequence in the biomolecule library is disposed at different positions in a larger frame register of possible amino acid residue positions such that the model is trained to recognize the given amino acid sequence as a sequence motif in multiple frame registers. A number of different frame registers comprises at least about 3, at least about 4, at least about 5, at least about 6, at least about 7, at least about 8, at least about 9, at least about 10, or more different frame registers.

[0012] Various additional optional features of the above embodiments include the following. The determining step comprises ascertaining the binding affinity values of different reference antibodies to the biomolecules in the biomolecule library substantially simultaneously with one another. The determining step comprises ascertaining the binding affinity values of different concentrations of a given reference antibody to the biomolecules in the biomolecule library. The method comprises using a slope of concentration dependence as a metric for determining biomolecule sequences to which a target antibody binds. The determining step comprises contacting the reference antibodies and the biomolecules in the biomolecule library. The reference antibodies comprise fluorescent labels and wherein the determining step comprises detecting levels of fluorescent signals emitted from the fluorescent labels of the reference antibodies that are bound to the biomolecules to determine the binding affinity values of the reference antibodies to the biomolecules in the biomolecule library. The determining step comprises binding secondary antibodies that comprise fluorescent labels to the reference antibodies that are bound to the biomolecules and detecting levels of fluorescent signals emitted from the fluorescent labels of the secondary antibodies that are bound to the reference antibodies to determine the binding affinity values of the reference antibodies to the biomolecules in the biomolecule library. One or more solid supports comprise the biomolecules in the epitope library. The biomolecules in the biomolecule library are arrayed on a substrate. A plurality of beads comprises the biomolecules in the biomolecule library. The model comprises a trained machine learning model. The creating step comprises training a machine learning model to predict a binding affinity value of a target antibody to a target biomolecule using the antibody-biomolecule binding data sets. An electronic neural network comprises the model. A given reference antibody-biomolecule bindingpair that comprises a binding affinity value that is higher than the binding affinity values of one or more other reference antibody-biomolecule binding pairs is weighted higher than the other reference antibody-biomolecule binding pairs when creating the model. The method comprises creating the model using representations of biomolecule sequences in the antibody-biomolecule binding data sets. The representations comprise one-hot vector representations. The representations comprise chemical parameters and / or chemical structural representations of monomers in the biomolecule sequences. The representations comprise concatenated vectors of amino acid physical or chemical or biological properties. The method comprises fitting the antibody-biomolecule binding data sets to another model that is based on biomolecule sequence composition to produce fitted data sets prior to creating the model. The method comprises normalizing the antibody-biomolecule binding data sets using the fitted data sets.

[0013] Various additional optional features of the above embodiments include the following. The method comprises disregarding the binding affinity values or contribution of the reference antibodies to biomolecules in the biomolecule library that depend only on compositions of the biomolecules prior to creating the model. The method comprises using the model to predict levels of binding of a target antibody, if any, to one or more biomolecule sequences of interest. The biomolecule sequences of interest comprise predicted cognate epitope sequences. The biomolecule sequences of interest comprise target antigen sequences. The target antigen sequences are derived from at least one pathogenic agent. The biomolecule sequences of interest comprise at least a portion of a whole proteome of protein sequences or at least portions of multiple whole proteomes of protein sequences. The method comprises using the model to identify potential target cognate epitope sequences within a library of biomolecule sequences, a target antigen or sets of target antigens, or a target proteome or sets of target proteomes. The method comprises using the model to identify potential off-target biomolecule sequences within a library of biomolecule sequences, a target antigen or sets of target antigens, or a target proteome or sets of target proteomes. The method comprises using the model to determine a likely functional significance of binding to enzymatic activity, sensing function, signal transduction, molecular binding interference, molecular bindingincrease, gene activity, transcription activity, translation activity, and / or post translational modification activity. The method comprises using the model to predict a potential autoantigen in a mammalian biological system. The method comprises using the model to predict off-target binding of an antibody based therapeutic agent. The method comprises using the model to predict off-target binding of an antibody used in a diagnostic assay. The method comprises using the model to predict an identity of a pathogen that gave rise to an antibody present in a sample. The method comprises using the model to predict one or more portions of an antigen to which a target antibody binds. The method comprises using the model in conjunction with a protein structure to define an epitope structure or noncontiguous binding site of an off-target binding interaction. The model uses predicted antibody binding as a basis to predict epitope sequences to which antibodies bind.

[0014] According to various embodiments, a system for generating a model that predicts targets of antibody binding is presented. The system comprising: a processor; and a memory communicatively coupled to the processor, the memory storing instructions which, when executed on the processor, perform operations comprising: receiving binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library to produce one or more antibody-biomolecule binding data sets; and, creating a model that relates antibody binding to biomolecule sequence using the antibody-biomolecule binding data sets.

[0015] According to various embodiments, a system for predicting a biomolecule sequence to which at least one target antibody binds is presented. The system comprising: a processor; and a memory communicatively coupled to the processor, the memory storing instructions which, when executed on the processor, perform operations comprising: passing at least one target antibody data set through a model that relates antibody binding to biomolecule sequence, wherein the model was created using one or more antibody-biomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library, and wherein the target antibody data set is representative of one or more attributes of the target antibody and is sufficient for the model to produce a prediction of one or more biomolecule sequences to which the target antibody binds; and, outputting from the model one or more predicted cognateepitope sequences and / or one or more predicted off-target biomolecule sequences to which the target antibody binds.

[0016] According to various embodiments, a computer readable media is presented. The computer readable media comprising non-transitory computer executable instructions which, when executed by at least one electronic processor, perform at least: receiving binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library to produce one or more antibody-biomolecule binding data sets; and creating a model that relates antibody binding to biomolecule sequence using the antibody-biomolecule binding data sets.

[0017] According to various embodiments, a computer readable media is presented. The computer readable media comprising non-transitory computer executable instructions which, when executed by at least one electronic processor, perform at least: passing at least one target antibody data set through a model that relates antibody binding to biomolecule sequence, wherein the model was created using one or more antibody-biomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to epitope molecules in a biomolecule library, and wherein the target antibody data set is representative of one or more attributes of the target antibody and is sufficient for the model to produce a prediction of one or more biomolecule sequences to which the target antibody binds; and, outputting from the model one or more predicted cognate epitope sequences and / or one or more predicted off-target biomolecule sequences to which the target antibody binds.Drawings

[0018] The above and / or other aspects and advantages will become more apparent and more readily appreciated from the following detailed description of examples, taken in conjunction with the accompanying drawings, in which:

[0019] FIG. 1 A is a flow chart that schematically shows exemplary method steps of predicting a biomolecule sequence to which at least one target antibody binds according to some aspects disclosed herein.

[0020] FIG. 1 B is a flow chart that schematically shows exemplary method steps of generating a model that predicts targets of antibody binding according to some aspects disclosed herein.

[0021] Fig. 2 is a schematic diagram of an exemplary system suitable for use with certain aspects disclosed herein.

[0022] FIG. 3 show violin plots of log 10 binding. Unfilled circles: array peptides, Filled circles with lines upward-facing to the right: cognate sequences. Cognate sequences on the array are shown to the left of those filled circles with lines upward- facing to the right, top ten array sequences are to the right in black. Cognate sequences to the left in plots for TF3B5, C3, and 9E10.3 are not on the array. The SeqLogo derived from the top ten sequences is shown above them. Figure discloses SEQ ID NOS 1 -94, respectively, in order of columns within each graph.

[0023] FIG. 4 show single amino acid mutation / substitution matrices for each of the mAb cognate sequences. The y-axis is the cognate sequence and the x-axis is the amino acid used for substitution. The greyscale is proportional to the slope predicted by the neural network (see greyscale bar). In each case, the values were normalized to the predicted or measured slope of the cognate sequence, correspondingly. Note that any sequence with a value of 1 .1 times the slope of the cognate sequence is light grey. This allowed better resolution of values less than 1. Figure discloses SEQ ID NOS 13, 13, 39, 39, 95, 95, 96, and 96, respectively in order of appearance.

[0024] FIG. 5: The sequence of amino acids was tiled into overlapping peptides of 7 residues (6 residue overlap). The concentration dependent slope was predicted and converted to a greyscale, lighter to darker, and mapped onto the Alphafold2 representation of the sequence provided on Uniprot (Uniprot IDs given in Table 1 ).

[0025] FIG. 6: Pearson correlation coefficients between the linear binding values of array data after averaging replicates. For each monoclonal antibody, there are four concentrations (in increasing order, 0.125 nM, 0.5 nM, 2.0 nM, 8.0 nM).

[0026] FIG. 7: Plots showing Logw binding distributions of each monoclonal antibody at four concentrations.

[0027] FIG. 8: Plots showing the average binding of mAbs to peptides on the array vs. concentration for all peptides that do not saturate at any concentration.

[0028] FIG. 9: Scatter plot of logw predicted values of the sequence dependent binding vs. the logw sequence dependent binding determined from the values measured on the array. Filled circles with lines that are upward-facing to the right are the cognate sequences. Only 100,000 randomly picked data points are shown for each mAb of the -500,000 total points across all concentrations.

[0029] FIG. 10: Sequence-dependent binding vs. compositional binding for each of the mAbs at 0.125 nM concentration. Each sequence is plotted as a scatter plot. Also plotted for reference are the cognate sequences that are present on the array (circles filled with lines that are upward-facing to the right). The top ten noncognate sequences from the array are shown as unfilled circles. Both the cognate epitope sequences (if present on the array) and the top 10 binding array sequences are listed for each mAb. Figure discloses SEQ ID NOS 1 -2, 4, 3, 5-24, 28-29, 25, SO- 33, 26, 34-37, 97, 38-39, 41 -42, 40, 43-44, 48-50, 98-100, 51 -54, 56, 58, 60-61 , 101 - 103, 64-72, 104, 78, 75, 74, 77, 80-83, 105-106, and 85-94, respectively, in order of columns within each graph.

[0030] FIG. 11 : Predicted mutation matrix for each of the mAb cognate sequences. The y-axis is the cognate sequence and the x-axis is the amino acid used for substitution. The grayscale is proportional to the relative binding predicted by the neural network (see grayscale bar). In each case, the values were normalized to the predicted binding of the cognate sequence. Note that any sequence with a value of 1 .1 times the binding value of the cognate sequence is light grey. This allowed better resolution of values less than 1. Figure discloses SEQ ID NOS 107, 14, 108, 38, 51 , and 109-11 1 , respectively, in order of appearance.

[0031] FIG. 12: Plots showing Ranks of the 7mer cognate epitope (Table 1 ) in 1 million random sequences using training sets of array peptides in which a random subset of peptides was used. The Y axis is the percent rank, with 0.0001 % signifying rank 1 in a million. The X axis is the fraction of the -122,000 peptide sequences on the array used, randomly chosen. The points and error bars represent the average and standard error of 5 randomly chosen training sets each evaluated by 12 randomly initiated models.Definitions

[0032] In order for the present disclosure to be more readily understood, certain terms are first defined below. Additional definitions for the following terms and other terms may be set forth throughout the specification. If a definition of a term set forth below is inconsistent with a definition in an application or patent that is incorporated by reference, the definition set forth in this application should be used to understand the meaning of the term.

[0033] As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” include plural references unless the context clearly dictates otherwise. Thus, for example, a reference to “a method” includes one or more methods, and / or steps of the type described herein and / or which will become apparent to those persons skilled in the art upon reading this disclosure and so forth.

[0034] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. Further, unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains. In describing and claiming the methods, systems, and computer readable media, the following terminology, and grammatical variants thereof, will be used in accordance with the definitions set forth below.

[0035] Antibody: As used herein, the term “antibody” refers to an immunoglobulin or an antigen-binding domain thereof. The term includes but is not limited to polyclonal, monoclonal, monospecific, polyspecific, non-specific, humanized, human, caninized, canine, felinized, feline, single-chain, chimeric, synthetic, recombinant, hybrid, mutated, grafted, and in vitro generated antibodies. The antibody can include a constant region, or a portion thereof, such as the kappa, lambda, alpha, gamma, delta, epsilon and mu constant region genes. For example, heavy chain constant regions of the various isotypes can be used, including: IgGi , lgG2, IgGs, lgG4, IgM, IgAi, lgA2, IgD, and IgE. By way of example, the light chain constant region can be kappa or lambda. The term “monoclonal antibody” refers to an antibody that displays a single binding specificity and affinity for a particular target, e.g., epitope.

[0036] Binding Affinity: As used herein, the term “binding affinity”, “binding affinity contribution”, or “binding intensity”, typically refers to a strength of non-covalent association between or among two or more entities.

[0037] Biomolecule: As used herein, “biomolecule” refers to the general categories or classes of organic molecules produced by or otherwise found in a living organism. Non-limiting examples of biomolecules, include lipids, carbohydrates or polysaccharides, poly- or oligopeptides (e.g., proteins, peptides, etc.), poly- or oligonucleotides (e.g., DNA, RNA, etc.), as well as the monomer units of such compounds, such as, amino acids, nucleotides, saccharides, and the like. In some embodiments, a given compound may include one or more modified or non-naturally occurring monomer units and yet remain in a general category or class of a biomolecule, such as a polypeptide or a polynucleotide. For example, a given polypeptide can include one or more nonnatural amino acids.

[0038] Data set: As used herein, “data set” refers to a group or collection of information, values, or data points related to or associated with one or more objects, records, and / or variables. In some embodiments, a given data set is organized as, or included as part of, a matrix or tabular data structure. In some embodiments, a data set is encoded as a feature vector corresponding to a given object, record, and / or variable, such as a given test or reference subject. For example, a medical data set for a given subject can include one or more observed values of one or more variables associated with that subject.

[0039] Electronic neural network: As used herein, “electronic neural network” or “neural network” refers to a machine learning algorithm or model that includes layers of at least partially interconnected artificial neurons (e.g., perceptrons or nodes) organized as input and output layers with one or more intervening hidden layers that together form a network that is or can be trained to classify data, such as test subject medical data sets (e.g., peptide sequence and binding value pair data sets or the like).

[0040] Machine Learning Algorithm: As used herein, "machine learning algorithm" generally refers to an algorithm, executed by computer, that automates analytical model building, e.g., for clustering, classification or pattern recognition. Machine learning algorithms may be supervised or unsupervised. Learning algorithms include, for example, artificial or electronic neural networks (e.g., back propagationnetworks), discriminant analyses (e.g., Bayesian classifier or Fisher’s analysis), multiple-instance learning (MIL), support vector machines, decision trees (e.g., recursive partitioning processes such as CART -classification and regression trees, or random forests), linear classifiers (e.g., multiple linear regression (MLR), partial least squares (PLS) regression, and principal components regression), hierarchical clustering, and cluster analysis. A dataset on which a machine learning algorithm learns can be referred to as "training data." A model produced using a machine learning algorithm is generally referred to herein as a “machine learning model.”

[0041] Peptide. As used herein, “peptide” refers to a sequence of 2-50 amino acids attached one to another by a peptide bond. These peptides may or may not be fragments of full proteins. Examples of peptides include KPLEEVLN (SEQ ID NO: 112), FLPFQQK (SEQ ID NO: 1 13) etc.

[0042] Protein. As used herein, “protein” or “polypeptide” refers to a polymer of typically more than 50 amino acids attached to one another by a peptide bond. Examples of proteins include enzymes, hormones, antibodies, peptides, and fragments thereof.

[0043] Sample. As used herein, a “sample,” such as a biological sample, is a sample obtained from a subject. As used herein, biological samples include all clinical samples including, but not limited to, cells, tissues, and bodily fluids, such as saliva, tears, breath, and blood; derivatives and fractions of blood, such as filtrates, dried blood spots, serum, and plasma; extracted galls; biopsied or surgically removed tissue, including tissues that are, for example, unfixed, frozen, fixed in formalin and / or embedded in paraffin; milk; skin scrapes; nails, skin, hair; surface washings; urine; sputum; bile; bronchoalveolar fluid; pleural fluid, peritoneal fluid; cerebrospinal fluid; prostate fluid; pus; or bone marrow. In a particular example, a sample includes blood obtained from a subject, such as whole blood or serum. In another example, a sample includes cells collected using an oral rinse. The sample may be isolated from the subject and then directly utilized in a method for determining the presence or absence of antibodies, or alternatively, the sample may be isolated and then stored (e.g., frozen) for a period of time before being subjected to analysis.

[0044] Subject. As used herein, “subject” or “test subject” refers to an animal, such as a mammalian species (e.g., human) or avian (e.g., bird) species. Morespecifically, a subject can be a vertebrate, e.g., a mammal such as a mouse, a primate, a simian or a human. Animals include farm animals (e.g., production cattle, dairy cattle, poultry, horses, pigs, and the like), sport animals, and companion animals (e.g., pets or support animals). A subject can be a healthy individual, an individual that has or is suspected of having a disease or pathology or a predisposition to the disease or pathology, or an individual that is in need of therapy or suspected of needing therapy. The terms “individual” or “patient” are intended to be interchangeable with “subject.” A “reference subject” refers to a subject known to have or lack specific properties (e.g., a known pathology, such as melanoma and / or the like).

[0045] System: As used herein, "system" in the context of analytical instrumentation refers a group of objects and / or devices that form a network for performing a desired objective.

[0046] Value: As used herein, “value” generally refers to an entry in a data set that can be anything that characterizes the feature to which the value refers. This includes, without limitation, numbers, words or phrases, symbols (e.g., + or -) or degrees.Description of the Embodiments

[0047] Reference will now be made in detail to example implementations. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention and it is to be understood that other embodiments may be utilized and that changes may be made without departing from the scope of the invention. The following description is, therefore, merely exemplary.

[0048] I. Introduction

[0049] There is considerable interest in developing rapid methods for identifying or predicting epitopes that bind to a monoclonal antibody, a polyclonal antibody, or other antibody preparations, whether naturally occurring or non-naturally manufactured. In the case of monoclonal antibodies, specificity is typically of paramount importance. Whether they are being used as therapeutics, as diagnostics, for antigen retrieval and purification or for analytical research purposes, knowing what the precise target is and knowing what the potential off-target binding sites are isimportant to any particular use case. Similarly, polyclonal antibodies are often used for analytical research purposes and specificity information is also of considerable use. Analysis of the specific reactive sites involved in a humoral response to disease, either using antibodies in blood, serum or after purification from other components of the blood in an individual with a disease state is useful in diagnosis, determining therapeutic targets, following the immune status of a population or mechanistic research.

[0050] Accordingly, in some aspects, the present disclosure provides for antibody target identification based on libraries of peptides or other short sequences of amino acids. However, it will be clear to those skilled in the art that other types of molecular libraries could also be used including oligosaccharides, nucleic acid sequences, and the like. In some embodiments, for example, the methods of the present disclosure utilize libraries of peptides, short amino acid sequences of roughly 10 residues. These sequences are chosen to cover all possible sequences in this length range evenly, although sparsely. In the examples provided herein, the number of sequences is about 126,000 while the total possible number of sequences is about 1013. Antibodies or antibody preparations are bound to the peptides and the amount of binding to each peptide is determined. The sequences are put in a form suitable for input into an electronic neural network such as one-hot encoding or substitution with values characteristic of the physical properties of the amino acids or other methods known to those skilled in the art. The electronic neural network is trained to predict the binding values to each sequence (normally the log of the binding values are used for this). This may be one value for a single antibody preparation, or it may be values for many different antibody preparations or concentrations simultaneously analyzed.

[0051] In some embodiments, prior to training, the binding values are fit to a linear relationship between amino acid composition and the binding value(s) of each sequence. The predicted binding due to composition is typically either subtracted from the overall binding or used to normalize it. Antibodies are typically very specific, there are often only a small number of sequences that bind strongly out of a set of near random sequences. Accordingly, these sequences are given additional weight in the training in some embodiments. Further, to avoid the electronic neural network training on a specific position of any given sequence motif within the larger sequence, theamino acid sequence is typically bounded by N- and C- terminal markers and placed within a larger sequence that includes place holder characters and then shifted at multiple positions within that sequence to force the network to learn to recognize a sequence motif regardless of where it occurs in an antigen.

[0052] To illustrate, FIG. 1 A is a flow chart that schematically shows exemplary method steps of predicting a biomolecule sequence to which at least one target antibody binds according to some aspects disclosed herein. As shown, method 100 includes passing at least one target antibody data set through a model that relates antibody binding to biomolecule sequence in which the model was created using one or more antibody-biomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library, and in which the target antibody data set is representative of one or more attributes of the target antibody and is sufficient for the model to produce a prediction of one or more biomolecule sequences to which the target antibody binds (step 102). Method 100 also includes outputting from the model one or more predicted cognate epitope sequences and / or one or more predicted off-target biomolecule sequences to which the target antibody binds to thereby predict the biomolecule sequence to which the target antibody binds (step 104).

[0053] In some embodiments, the target antibody data set and / or the antibody- biomolecule binding data sets comprise structural, chemical, and / or physical property data of the target antibody and / or of the biomolecules in the biomolecule library. In some embodiments, the target antibody data set and / or the antibody-biomolecule binding data sets comprise amino acid sequence information from the target antibody and / or from the biomolecules in the biomolecule library. In some embodiments, the target antibody comprises a monoclonal antibody, whereas in others, the target antibody comprises a polyclonal antibody. In some embodiments, the target antibody is from a sample (e.g., blood, serum, etc.) obtained from a subject.

[0054] In some embodiments, method 100 includes passing multiple target antibody data sets through the model. Typically, a trained electronic neural network comprises the model. In some embodiments, the electronic neural network was trained after weighting one or more biomolecule sequences in the antibody-biomolecule binding data sets. In some embodiments, a given amino acid sequence in thebiomolecule library is disposed at different positions in a larger frame register of possible amino acid residue positions such that the model was trained to recognize the given amino acid sequence as a sequence motif in multiple frame registers.

[0055] To further illustrate aspects of the present disclosure, FIG. 1 B is a flow chart that schematically shows exemplary method steps of generating a model that predicts targets of antibody binding according to some aspects disclosed herein. As shown, method 106 includes determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library to produce one or more antibody-biomolecule binding data sets (step 108). In addition, method 106 also includes creating a model that relates antibody binding to biomolecule sequence using the antibody-biomolecule binding data sets to thereby generate the model that predicts targets of antibody binding (step 1 10). In some embodiments, the model uses predicted antibody binding as a basis to predict biomolecule sequences to which antibodies bind.

[0056] In some embodiments, the reference antibodies comprise at least one monoclonal antibody. In some embodiments, the reference antibodies comprise at least one polyclonal antibody. In some embodiments, at least one of the biomolecules comprises a potential continuous epitope, whereas in others, at least one of the biomolecules comprises a potential discontinuous epitope.

[0057] In some embodiments, the biomolecule library comprises at least about 100 biomolecules, at least about 1000 biomolecules, at least about 10000 biomolecules, at least about 100000 biomolecules, at least about 1000000 biomolecules, at least about 10000000 biomolecules, or more biomolecules. In some embodiments, antigen molecules comprise the biomolecule in the biomolecule library. In some embodiments, the biomolecules comprise at least portions of carbohydrate molecules, nucleic acid molecules, protein molecules, or a combination of two or more thereof.

[0058] In some embodiments, the biomolecules comprise amino acid sequences. In some embodiments, the amino acid sequences comprise natural amino acids and / or nonnatural amino acids. In some embodiments, the amino acid sequences comprise selected sequence lengths and wherein the biomolecule library comprises a selected level of coverage of a combinatorial space that comprisessubstantially all amino acid sequence permutations of the sequence lengths. In some embodiments, the selected level of coverage comprises less than about 5%, less than about 2.5%, less than about 0.5%, less than about 0.25%, less than about 0.05%, less than about 0.025%, or less of the combinatorial space that comprises substantially all of the amino acid sequence permutations of the sequence lengths. In some embodiments, the biomolecule library comprises a substantially random sample of the amino acid sequence permutations in the combinatorial space. In some embodiments, the selected sequence lengths are in the range of about 3-15 amino acids, about 4-14 amino acids, about 5-13 amino acids, about 6-12 amino acids, about 7-1 1 amino acids, or about 8-10 amino acids. In some embodiments, the model relates the binding affinity values of the reference antibodies to amino acid sequences in each of the antibodybiomolecule binding data sets. In some embodiments, the methods of the present disclosure include using one or more markers to designate an N-terminus and / or a C- terminus of one or more of the amino acid sequences. In some embodiments, a given amino acid sequence in the biomolecule library is disposed at different positions in a larger frame register of possible amino acid residue positions such that the model is trained to recognize the given amino acid sequence as a sequence motif in multiple frame registers. In some embodiments, a number of different frame registers comprises at least about 3, at least about 4, at least about 5, at least about 6, at least about 7, at least about 8, at least about 9, at least about 10, or more different frame registers.

[0059] In some embodiments, the determining step of method 106 comprises ascertaining the binding affinity values of different reference antibodies to the biomolecules in the biomolecule library substantially simultaneously with one another. In some embodiments, the determining step of method 106 comprises ascertaining the binding affinity values of different concentrations of a given reference antibody to the biomolecules in the biomolecule library. In some embodiments, method 106 comprises using a slope of concentration dependence as a metric for determining biomolecule sequences to which a target antibody binds.

[0060] In some embodiments, the determining step of method 106 comprises contacting the reference antibodies and the biomolecules in the biomolecule library. In some embodiments, the reference antibodies comprise fluorescent labels andwherein the determining step comprises detecting levels of fluorescent signals emitted from the fluorescent labels of the reference antibodies that are bound to the biomolecules to determine the binding affinity values of the reference antibodies to the biomolecules in the biomolecule library. In some embodiments, the determining step of method 106 comprises binding secondary antibodies that comprise fluorescent labels to the reference antibodies that are bound to the biomolecules and detecting levels of fluorescent signals emitted from the fluorescent labels of the secondary antibodies that are bound to the reference antibodies to determine the binding affinity values of the reference antibodies to the biomolecules in the biomolecule library. In some embodiments, one or more solid supports comprise the biomolecules in the biomolecule library. In some embodiments, the biomolecules in the biomolecule library are arrayed on a substrate. In some embodiments, a plurality of beads comprises the biomolecules in the biomolecule library.

[0061] In some embodiments, the model comprises a trained machine learning model. In some embodiments, the creating step comprises training a machine learning model to predict a binding affinity value of a target antibody to a target biomolecule using the antibody-biomolecule binding data sets. In some embodiments, an electronic neural network comprises the model.

[0062] In some embodiments, a given reference antibody-biomolecule binding pair that comprises a binding affinity value that is higher than the binding affinity values of one or more other reference antibody-biomolecule binding pairs is weighted higher than the other reference antibody- biomolecule binding pairs when creating the model. In some embodiments, method 106 comprises creating the model using representations of biomolecule sequences in the antibody-biomolecule binding data sets. In some embodiments, the representations comprise one-hot vector representations. In some embodiments, the representations comprise chemical parameters and / or chemical structural representations of monomers in the biomolecule sequences. In some embodiments, the representations comprise concatenated vectors of amino acid physical or chemical or biological properties. In some embodiments, method 106 comprises fitting the antibody-biomolecule binding data sets to another model that is based on biomolecule sequence composition to produce fitted data sets prior to creating the model. In some embodiments, method106 comprises normalizing the antibody-biomolecule binding data sets using the fitted data sets. In some embodiments, method 106 comprises disregarding the binding affinity values of the reference antibodies to biomolecules in the biomolecule library that depend only on compositions of the biomolecules prior to creating the model.

[0063] In some embodiments, method 106 comprises using the model to predict levels of binding of a target antibody, if any, to one or more biomolecule sequences of interest. In some embodiments, the biomolecule sequences of interest comprise predicted cognate epitope sequences. In some embodiments, the biomolecule sequences of interest comprise target antigen sequences. In some embodiments, the target antigen sequences are derived from at least one pathogenic agent. In some embodiments, the biomolecule sequences of interest comprise at least a portion of a whole proteome of protein sequences or at least portions of multiple whole proteomes of protein sequences.

[0064] In some embodiments, method 106 comprises using the model to identify potential target cognate epitope sequences within a library of biomolecule sequences, a target antigen or sets of target antigens, or a target proteome or sets of target proteomes. In some embodiments, method 106 comprises using the model to identify potential off-target biomolecule sequences within a library of biomolecule sequences, a target antigen or sets of target antigens, or a target proteome or sets of target proteomes. In some embodiments, method 106 comprises using the model to determine a likely functional significance of binding to enzymatic activity, sensing function, signal transduction, molecular binding interference, molecular binding increase, gene activity, transcription activity, translation activity, and / or post translational modification activity. In some embodiments, method 106 comprises using the model to predict a potential autoantigen in a mammalian biological system. In some embodiments, method 106 comprises using the model to predict off-target binding of an antibody based therapeutic agent. In some embodiments, method 106 comprises using the model to predict off-target binding of an antibody used in a diagnostic assay. In some embodiments, method 106 comprises using the model to predict an identity of a pathogen that gave rise to an antibody present in a sample. In some embodiments, method 106 comprises using the model to predict one or more portions of an antigen to which a target antibody binds. In some embodiments, method 106 comprises usingthe model in conjunction with a protein structure to define an epitope structure or noncontiguous binding site of an off-target binding interaction.

[0065] Fig. 2 is a schematic diagram of a hardware computer system 200 suitable for implementing various embodiments. For example, Fig. 2 illustrates various hardware, software, and other resources that can be used in implementations of any of methods disclosed herein, including method 100 and / or one or more instances of an electronic neural network. System 200 includes training corpus source 202 and computer 201 . T raining corpus source 202 and computer 201 may be communicatively coupled by way of one or more networks 204, e.g., the internet.

[0066] Training corpus source 202 may include an electronic records system, such as an LIS, a database, a compendium of clinical data, or any other source of antibody-biomolecule binding data sets suitable for use as a training corpus as disclosed herein. According to some embodiments, each component is implemented as a vector, such as a feature vector.

[0067] Computer 201 may be implemented as any of a desktop computer, a laptop computer, can be incorporated in one or more servers, clusters, or other computers or hardware resources, or can be implemented using cloud-based resources. Computer 201 includes volatile memory 214 and persistent memory 212, the latter of which can store computer-readable instructions, that, when executed by electronic processor 210, configure computer 201 to perform any of the methods disclosed herein, including method 100, and / or form or store any electronic neural network, and / or perform any classification technique as described herein. Computer 201 further includes network interface 208, which communicatively couples computer 201 to training corpus source 202 via network 204. Other configurations of system 200, associated network connections, and other hardware, software, and service resources are possible.

[0068] Certain embodiments can be performed using a computer program or set of programs. The computer programs can exist in a variety of forms both active and inactive. For example, the computer programs can exist as software program(s) comprised of program instructions in source code, object code, executable code or other formats; firmware program(s), or hardware description language (HDL) files. Any of the above can be embodied on a transitory or non-transitory computer readablemedium, which include storage devices and signals, in compressed or uncompressed form. Exemplary computer readable storage devices include conventional computer system RAM (random access memory), ROM (read-only memory), EPROM (erasable, programmable ROM), EEPROM (electrically erasable, programmable ROM), and magnetic or optical disks or tapes.

[0069] II. Description of Example Embodiments

[0070] Example: Machine Learning Models of Monoclonal AntibodyBinding based on a Broad Sampling of Sequence Space

[0071] Introduction

[0072] In this example, a general approach is described for using mAb binding measurements to a near-random array of peptide sequences in order to generate a comprehensive and quantitative model of mAb binding to any possible sequence of 5- 11 amino acids. First, mAb binding is measured to a sample of roughly 126,000 unique peptide sequences that were selected to sample the trillion or so possible amino acid sequences of around ten residues in length as evenly as possible, with no regard to covering any particular biological sequence information. This data is then modeled with machine learning to create a general sequence binding relationship for the mAb which then can be used to predict the binding of the mAb to any amino acid sequence, making it a general tool for evaluating mAb sequence specificity. Previous work from this lab has demonstrated that it is possible to use binding of target samples to high density arrays in order to generate sequence-binding relationships capable of predicting with high statistical accuracy the binding values of sequences that are not present in the training set. This has been done both for binding of purified proteins and for binding of serum samples from cohorts of individuals with 5 different infectious diseases or uninfected individuals. In the latter case, the sequence-binding relationship not only maintained the ability to accurately classify disease cohorts, but it was also able to remove sequence independent noise from the measured binding data and improve the ability to classify cohorts.

[0073] An objective of the current example was to build an accurate sequencebinding relationship for previously characterized mAbs with continuous epitopes, and to evaluate its accuracy by showing that the model can predict specificity consistentwith the known mAb epitope sequence. Some key advantages of this approach are its generality and scalability; the same array is used to generate binding data for any mAb and the resulting model can be projected onto any sequence or any protein. While it is true that the focus here is on continuous epitopes, due to the linear nature of peptide sequences, the foundational models presented can serve as a generalizable methodology for quantitating mAb specificity and identifying potential off-target risk factors. It may also be that as additional information, such as three dimensional structure, is incorporated, more complex binding interactions can also be predicted.

[0074] Results and Discussion

[0075] Monoclonal Antibodies Used. The eight monoclonal antibodies (mAb) used in this study are listed in Table 1 . All of these have known, linear epitopes of at least 6 contiguous amino acids (for Ab8, 4C1 , Lnkb2 the antigen analysis is published and for others epitope and antigen information is provided by the supplier listed inTable 1 ). All were raised to human targets except for Ab1 , and all are Murine antibodies. In each case, 4 concentrations of the mAb (0.125 nM, 0.5 nM, 2.0 nM, 8.0 nM) were used in binding experiments and the binding measurements at each concentration were replicated 4 times.

[0076] Peptide Array Assay. Monoclonal antibody preparations were incubated with peptide arrays consisting of 126,051 unique peptide sequences that were selected to sample all possible amino acid sequence combinations of approximately 10 amino acids as evenly as possible (a complete description of the size distribution and amino acid distribution as a function of residue position has been published previously). These peptides were built using 16 of the 20 natural amino acids, A D E F G H K L N P Q R S V W and Y. Note that with the exception of 9E10.3, all of the mAbs had cognate epitope binding sequences that only included these amino acids (9E10.3 contains an isoleucine and for the analyses below, this was assumed to be similar to valine). Control sequences distributed across the arrays were used to gauge the uniformity of binding and multiple copies of cognate epitopes and some variants were synthesized on the peptide array for 5 of the 8 mAbs (Table 1 ). The mAb cognate epitope sequences were used to gauge the actual binding level of the mAb to its cognate epitope in the context of an isolated linear sequence attached to substrate. The binding of each mAb to the peptide arrays was performed as described in Methods and binding of a fluorescently labeled secondary antibody was used to determine how much mAb was bound to each sequence by detecting with an array scanner (Methods).

[0077] The four replicates of each mAb at each concentration were averaged as the mean of the logw values of the observed binding. Based on the comparisons of Pearson correlation coefficients between replicates (Fig. 6), no replicate arrays were excluded from the averaging. Based on the standard deviations between replicates for each binding value in each array, no individual binding values were excluded from the averaging; all collected data was used without removing outliers. Except where noted below, the average logw values from the replicates were used for analysis and all modeling of the sequence-binding relationships was performed after removing themAb cognate sequences from the array data to avoid training on those epitope sequences.

[0078] Distributions of binding to array sequences. Figs. 3 and 7 show the distribution of binding values for each mAb to the array sequences in different ways after averaging replicates. Fig. 7 provides a plot of the logw number of peptides at each binding value for each mAb and concentration. Fig. 3 focuses on the lowest concentration, comparing violin plots of the distribution. Note that the y-axis is logw binding. The cognate sequence binding of each mAb (sometimes several versions) are represented as filled circles with lines upward-facing to the right for the first five mAbs and the cognate sequence itself is given in to the left of those filled circles with lines upward-facing to the right for those 5 mAbs and to the left in plots for the final 3 mAbs (TF3B5, C3, and 9E10.3). In all but one case (4C1 ) at least one version of the mAb cognate sequence saturates the detector (65,536 counts or 4.82 on a logw scale). The top ten binding array sequences and the SeqLogo consensus sequence determined from them is shown.

[0079] Several things are evident in Fig. 3. First, as a result of the specificity of mAbs the vast majority of sequences bind at values very close to the background (about 800 counts or 2.9 on the log scale). For most mAbs there are only a handful of points higher than 3.5 and for Ab8 there are none. Second, in most cases, the top ten sequences define some identifiable motif related to the known cognate sequence. While this is less evident for Ab8, one sees that the composition of the top sequences is similar to that of the cognate epitope. Thus, there is antibody-specific information in the binding to even a sample of near-random peptide sequences. Finally, in most cases there is also one or more apparently unrelated sequence in the top ten that may represent a mimotope, i.e., a peptide that, despite apparent dissimilarity in the letter code, can have similar physicochemical properties to the cognate sequence. The binding distributions of the mAbs exemplify the challenge of using binding data from peptides sampling all of the 16 amino acid sequence space (1610= ~1012sequences) without regard for any specific biological relevance to create comprehensive sequence-binding models. While this approach ensures generality as it is not biased towards any particular outcome, the specificity of mAbs means that the bulk of the data from the array is from sequences that bind very weakly.

[0080] Modeling the sequence-binding relationship using machine learning methods. To build a general relationship between sequence and mAb binding from the array data, a fully connected neural network was employed with 2 hidden layers and 250 nodes per layer. This was similar in structure to networks used previously for describing serum IgG binding or binding to isolated proteins. Two additions were made to the sequence representation beyond what has been done previously. Peptide sequences with fewer than the maximum number of residues were padded with a “blank” residue (marked as “X”), so that they were all the same length. Also, the N- and C-termini were marked with end markers “(” andThe sequences, including the three additional characters, were then input into the neural network as one-hot encoded vectors, similar to previous work. Note that only the peptides between 5 and 11 residues in length were used in the training (there were 121715 unique peptides in this length range) because the focus was on identifying continuous epitopes which are generally in that length range and because there is a tendency for the neural network to overestimate the binding of longer peptides, simply due to the larger number of elements to assign value to, potentially biasing the model.

[0081] One other adjustment was made to the dataset prior to running the neural network. At higher concentrations, for some of the mAbs, there were a number of sequences whose binding saturated the detector (Fig. 6), potentially causing the neural network to underestimate high binding values. This was corrected by substituting saturating values at one concentration with the value from the next lower concentration multiplied by the concentration increase. This assumption of linearity with concentration is justified by the fact that binding values of peptides on the array in the linear range, do, in fact, increase linearly with concentration (Figure 8). Each mAb was modeled separately, but all four concentrations for the mAb were used as targets simultaneously. Thus, the final weight matrix of the neural network had four columns, each associated with a different concentration.

[0082] Evaluating the neural network prediction performance. For performing the binding predictions below, the neural network was trained 12 separate times using the binding of all measured array sequences to the mAbs in the training set (except for the cognate sequences which were removed during training), and the results were averaged. (For reference, Figure 10 shows the results of a 10-fold cross validation ofthe sequence binding relationship for each mAb to illustrate the accuracy of predictions within the measured dataset.) After training, for each mAb, the resulting model was used to predict the binding of the mAb at each concentration to one million randomly generated sequences of the same length as the cognate sequence (Table 1 , either 6 or 7 residues). The binding values predicted for the random sequences and cognate sequence were used to fit the slope of the binding vs. concentration dose response relationship for each sequence. Using the dose response slope rather than raw binding assured that only concentration dependent binding was considered. This slope was then used to determine the rank order of the cognate sequence among the million random sequences (Table 2). Ranking was also calculated based directly on predicted binding for comparison (Table 3). The entire process described above was repeated 5 times and the error of the mean of each rank was determined.

[0083] Baseline prediction. Table 2, column 2 gives the rank of the predicted cognate epitope sequence based on the slope of the concentration dependence of the binding as described above using the input sequences with no further modification. Five of the eight mAb cognate sequences (Ab1 , 4C1 , DM1 A, TF3B5, C3) are in the top 10% of the random sequences, and two of them, DM1 A and TF3B5, are in the top 1 %. The prediction for both Ab8 and 9E10.3 is essentially random. While being in the top 10% implies this simple modeling approach predicts at least some specificity of mAb binding, the specificity is clearly not at the level suggested by the binding of the cognate sequence in Fig. 3.

[0084] Weighting the highest value binding sequences. It is clear from Fig. 3 that the number of sequences that bind substantially to the monoclonal antibodies is generally small. Yet it is these sequences that contain some of the most important binding information. It thus makes sense to balance the dataset by weighting high binding values more than the very large number of low binding values in the modeling. To accomplish this, binding distributions of the 8 nM mAbs were used to develop a weighting rule and weighting was implemented by increasing the number of copies of the high binding sequences accordingly (see Methods). When the same neural network model was applied to the weighted dataset, predictions for six of the mAbs improved markedly. In particular Ab1 , DM1 A, Lnkb2, TF3B5, C3, and 9E10.3 improved their ranks in the million random sequences by one to three orders of magnitude. Neither Ab8 nor 4C1 improved with weighting. This is interesting in the case of 4C1 given the similarity in general to the distribution of sequences other mAbs (see below).

[0085] Training on multiple copies of each sequence with shifted registers. One of the weaknesses with the above approaches to creating a predictive binding model from sequence is that the residue positions within the sequence description vector are static relative to the sequences. As a result, the neural network tends to associate the important motifs with one, artificial, position within the sequence vector. In other words, the model “expects” a motif to be located at a particular position in the peptide. To mitigate this effect, the training set was expanded to include each sequence in 6 different registers within an extended vector (see Methods). All shifted registeredversions have the same target binding values associated with them for training. As can be seen in the fourth column of Table 2, this generalization greatly improves the predicted ranks of almost all of the mAbs cognate epitope sequences such that half rank in the top 0.01% of sequences and all are in the top 0.1 %.

[0086] It is worth considering why this works as well as it does. The effect on the prediction for Ab8 is particularly profound, improving its ranking by 4 orders of magnitude. It is likely this is due to the fact that there are so few sequences which contain the needed sequence-specific information. The matrix manipulations in a neural network are register-specific; an amino acid in the first position always “sees” the same set of processes take place as the information moves through the neural network. Thus with sequence shifting, the number of available sequences to train on is effectively increased 6-fold. Further, any bias that may exist in amino acid composition of a particular position or any affects from the fact that the sequences are not all the same length are effectively washed out. This was clearly the most significant means of improving the predictive power of the model, and it is an approach that can easily be applied to other situations, such as whole sera for epitope predictions associated with an immune response to disease.

[0087] Training after normalizing for the level of composition-dependent binding. In any antibody interaction with a target sequence, there will be some level of interaction which is strictly sequence dependent (sequence specific) and some level of binding that depends only on the overall nature of the peptide (charge, hydrophobicity, etc.). One way to separate out the binding that is strictly dependent on sequence order is to determine what fraction of the binding for each peptide can be determined strictly from information about the peptide composition and then subtract this. Figure 10 shows the result of fitting the binding of the mAbs to the peptide array by only considering 16 linear coefficients, one for each amino acid used, that operate on a composition vector representing the peptide (a vector made up of the number of each amino acid present). The part of the logw binding that could be described by composition alone was subtracted from total logw binding and the neural network was then trained on the resulting sequence dependent binding for each mAb. Using this approach, there is very little change in most of the mAbs but both Ab8 and 9E10.3 significantly improve. 4C1 appears to drop in rank by about 2-fold, but the gains in theother two more than make up for this. Apparently, removing composition-dependent binding from Ab8 and 9E10.3 allows the neural network to focus on learning the sequence dependence without simultaneously modeling general compositional binding. The final ranks obtained are all within the top 0.01 % of the 1 ,000,000 random sequences except for 4C1 , which is about tenfold lower in rank than the others. However, this is consistent with the lower specificity indicated by Fig. 3; it was the one mAb for which none of the cognate sequence variants measured was saturating and in fact there were near-randomly chosen array sequences that bound more strongly.

[0088] Predicting which amino acids are important in the cognate sequences. The neural network models can also be used to predict the binding of all possible single amino acid substitutions of the mAb cognate sequences using the 16 amino acids that make up the array peptides. The slope of the binding vs concentration relationship was used as the metric for determining the binding level of each substitution, and the values were normalized to the cognate sequence value. For four of the mAbs, these substituted sequences were also synthesized on a separate array and their values measured for comparison. The predicted and measured values for those four mAbs are shown in Fig. 5. The agreement between predicted and measured variants of the cognate sequences is generally very good, with the model accurately determining which amino acids at which positions are critical in all but two cases. For Ab8, W is predicted to be easily substituted by several other possible amino acids. However, in the measured peptides W can only be substituted by Y, and that with some loss of binding. In TF3B5, E is predicted to be essential, but in the measured sequences, it is clear that it can be replaced by any of five other amino acids with no loss in binding. However, in all other cases, the prediction of which amino acids are most essential is in line with what is measured.

[0089] Predicted substitution matrices for the other four mAbs are given in Fig. 11 . The results for 4C1 are of particular interest. In all other mAbs, there are either 4 or 5 specific binding residues in the cognate sequence. However, for 4C1 , there were only 2 amino acids in the cognate sequence that were highly specific with a third amino acid showing somewhat weaker specificity. This is consistent with the data and discussion from figure 3 and Table 2, in which a lower specificity overall for 4C1 binding is due to the simpler motif required.

[0090] Mapping mAb binding to antigen sequences. The ability to apply the neural network binding models developed here to biological sequences is currently limited by the fact that 1 ) only 16 of the 20 natural amino acids are included in the peptide sequences on the arrays used in this study and 2) three dimensional structures are not currently considered in determining binding sites in the application of the models onto proteins which limits the ability to map noncontinuous epitopes. However, to demonstrate proof of principle with the current model system, Figure 5 shows color maps of the binding of each of the mAbs modeled here to its respective antigen structure. The 4 missing amino acids were dealt with by similarity substitution (l-> V, T->S, M->L, C->S) determined using the PAM250 similarity matrix, based on comparison of various similarity matrices. The sequences of the appropriate antigen subunit are rendered as Alphfold2 predicted structures to provide an entire representation of the sequence, including unstructured regions that are not present in the crystal structures (Table 1 gives the Uniprot ID numbers from which the structures were derived). The sequences of the antigen subunit were tiled as overlapping 7 residue peptides (6 residue overlap). The slope of the binding concentration dependence, calculated as described above, was then predicted and each tiled peptide assigned a greyscale shade between lighter and darker based on the predicted slope.

[0091] Seven of the eight models predict the cognate epitope sites essentially perfectly with little other significant binding. 9E10.3, in contrast to the others, shows three strong binding sites. However, the issue in this case is not the algorithm but rather both similar sequences within the myc antigen and the limitations of the model system used. The three sequences of the myc protein antigen that bind strongly to 9E10.3 are: KLISE (SEQ ID NO: 114) (cognate, only this particular helix was used to select the mAb, see product literature, ThermoFisher, Cat# MA5-12080), KLVSE (SEQ ID NO: 1 15) and MVTE (SEQ ID NO: 116). However, because the model system contains only 16 of the 20 amino acids, it was necessary to replace I and M with V and L, respectively. Thus, the sequences become KLVSEE (SEQ ID NO: 117), KLVSE (SEQ ID NO: 1 15) and LVSE (SEQ ID NO: 1 18). Looking at the substitution matrices of Fig. 4, one can see that the critical part of the cognate sequence is LVSE (SEQ ID NO: 1 18). Development of arrays for more general use in identifying epitopes withinproteins and proteomes will be needed to allow for complete modeling of all amino acids.

[0092] Conclusions

[0093] What is described here is an approach for measuring mAb binding to a set of near-random peptides sequences and creating from that a general model that successfully recognizes the cognate sequence of the mAb as one of the strongest binding sequences in a large number of random sequences (Table 2). Surprisingly, this is even effective when there is very little sequence dependent binding for the model to learn from, such as the case of Ab8. Further, the model allows one to accurately predict which amino acids within the cognate epitope are critical for binding compared to empirical determination (Fig. 4) and to map the binding onto the sequences of the respective antigens, accurately showing strong binding to the cognate sequence (Fig. 5).

[0094] As shown in Table 2, the neural network model in its simplest form was somewhat predictive for most of the mAbs, but seriously limited by the fact that the vast majority of the peptides show essentially baseline binding. Weighting the peptides that bind more strongly was a great help in improving the model for all but Ab8 and 9E10.3. The biggest advance in the model came from forcing the neural network to ignore the absolute position of the sequence in the vector and generalize its learning so that it recognized the sequence regardless of its position within the artificial register of the input vector. It is remarkable in many respects that such a simple neural network model (2 hidden layers of 250 nodes, fully connected) would be able to generalize in this way. The improvement in the rank of Ab8 binding in this case was four orders of magnitude and most other mAbs improved by 1 -3 orders of magnitude in rank. Removing binding due to composition was also helpful in some cases, presumably by focusing the learning of the neural network on the sequence specific aspect of binding (though the rank in one case, 4C1 , did lower).

[0095] Note that because there was so little high binding data available to learn from in some mAb binding experiments, the absolute predicted values of the cognate sequences were underestimated (Fig. 10). Particularly for Ab1 and Ab8, where high binding sequences were extremely scarce, this effect was large. Apparently, it was difficult for the model to learn to predict the absolute binding values of sequences thatwere more than an order of magnitude higher than anything the model was trained on. However, the relative values, and thus the rank of the binding value, was predictable, even when the absolute value was underestimated, and from a practical perspective this is the important outcome.

[0096] As described in the introduction, there are a number of different approaches to predicting epitope sequences, as well as off-target binding sequences, of mAbs. This includes peptide array approaches in which proteins or proteomes of interest have been directly tiled. It is fair to ask, why wouldn’t one simply tile the proteome of interest rather than generating model predictions from sequences broadly sampled from all possible sequences? There are two reasons for exploring the approach described here, one practical and one more fundamental.

[0097] From a practical perspective, the array-based approach employed here and the comprehensive model it produces provide both a universal assay platform that can be applied to any mAb at low cost / high throughput and complete flexibility in the evaluation of mAb binding to any sequence. For example, suppose one is interested in exploring all potential targets in the human proteome for 100 lead mAbs. A single 8 inch silicon wafer substrate used in array generation contains >300 individual arrays, thus the arrays from one wafer could assay all of these mAbs at three concentrations and create models for each (and one photolithographic fabrication station generates 4 wafers at a time). In contrast, tiling the ~11 million overlapping peptides in the human proteome would be very difficult and expensive to do at this throughput. In addition, such a tiled array would only be generated for a particular reference proteome and would not necessarily contain mutations of interest in a particular disease state or for individual proteomes. Using a comprehensive model based on much smaller, nearrandom sequence arrays, one can predict mAb binding to any sequence including the proteome sequences of individuals determined from their personal DNA sequences. Models also allow the prediction of binding to sequences that may be hard to synthesize in a tiled library. One limitation of the approach described here is that the use of linear peptide sequences lends itself to the identification of continuous, rather than discontinuous, epitopes. However, discontinuous epitopes are often made from multiple short binding regions. This approach effectively assembles information from short binding regions in any case, since the peptides on the array generally don’tcontain the whole binding sequence. Thus, one may be able to generalize to at least some noncontinuous epitopes simply by evaluating the proximity of binding segments in the three dimensional structure. In addition, this approach allows one to ask more generally, how specific is a mAb? Does it bind to a relatively large number of sequences strongly, such as 4C1 , or only very few, such as Ab8? Are there particular off-target proteins that it appears to bind most strongly? That type of information can be very useful in evaluating many possible lead mAbs raised against the same target for therapeutic or diagnostics purposes.

[0098] From a more fundamental perspective, the study’s outcome is important for another reason. Almost all of the machine learning work to date that relates amino acid sequence to structure and function has involved taking a set of known functional sequences (folded proteins, antibiotic peptides, epitopes, etc.) and learning from known example. Alphafold2 is perhaps the most ambitious such achievement, taking -140,000 known examples of protein structures and generalizing to predict the structure of any sequence. In such cases, one does not know how much of the primary sequence space the predictions are actually valid for because the training is biased towards known function and there is usually no way to empirically test all the possible sequences to see how significant that bias is. However, in the work described here, the training was done without any known examples, based on a near-random sampling of the possible sequences. Yet this very sparse and near-random sample of a trillionmember sequence space is sufficient to provide a reasonably accurate model of the entire molecular recognition landscape, including predicting the relative binding of something as specific as a mAb cognate epitope. Thus, at least for a relatively modest sized sequence space in the context of molecular recognition, it is possible to create a comprehensive model from sparse sampling where there was no a priori bias in the sequences used to train the model and no reason to expect bias in its predictions.

[0099] Methods

[0100] Peptide arrays. The peptide microarrays used here were produced by Cowper Sciences (Chandler, AZ) using methods previously described. Briefly, photolithographic approaches were used to produce 126,051 unique peptide sequences plus copies of control sequences on 200 mm silicon wafers with a distribution of lengths around 10 amino acids. 16 different amino acids were used inthe synthesis (ADEFGHKLNPQRSVWY (SEQ ID NO: 1 19)) and the arrays were diced into 25x75 mm slides containing 24 arrays each in a 3 x 8 pattern, allowing for high- throughput screening using 96-well plate robotic technologies. The sequences were designed to sparsely sample all of the roughly trillion possible amino acid combinations as evenly as possible.

[0101] Array assays. Samples were diluted 1 :625 in assay buffer (PBST with 0.05% Tween 20, 0.1 % Proclin 950 and 1 % mannitol), and 90 pl was incubated with rehydrated microarrays for 1 hr at 37°C (rehydration involved distilled water for 1 hour (h), PBS for 30 minutes then assay buffer for 1 hour). Dilution and liquid handling steps were performed using a BRAVO robotic system (Agilent, Santa Clara, CA) and mixing during incubation was performed using a TeleShake95 platform mixer (inheco, Marinsried, Germany). Arrays were washed three times in assay buffer using a plate washer (BioTek Instruments, Inc., Winooski, VT). Goat anti-human IgG (H+L) conjugated to AlexaFluor 555 (Invitrogen-Thermo Fisher Scientific, Inc., Carlsbad, CA) or Goat anti-human IgM (H+L) (Novus Biologicals, Centennial, CO), conjugated to DyLight 550 (Invitrogen-Thermo Fisher Scientific, Inc., Carlsbad, CA) was then incubated with the arrays in secondary incubation buffer (0.5% casein in array buffer) at 4.0 nM and 37°C for 1 h with mixing. Slides were then washed three times with array buffer and subsequently by distilled water for residual salt removal. Finally, the slides were sprayed with isopropanol and dried by centrifugation.

[0102] Imaging of the fluorescent signals from the slides was performed on an ImageXpress imaging system (Molecular Devices, San Jose, CA). Gridding of the images was performed with Mapix (version 7.2.1 ; Innopsys, Carbonne, France) using the central 60% of each feature and the intensities read were saved as GenePix results files. Arrays found to have serious image artifacts were re-run.

[0103] Neural network architecture. The basic neural network architecture has been described previously. A fully connected neural network with 2 hidden layers and 250 nodes in each layer was used and the inputs were one-hot representations of sequences. Training was performed relating the sequences to binding values on the array. All of the measured sequence-binding pairs, except the sequences on the array that were cognate epitopes of the 8 mAbs, were used in the training. This did not introduce binding, since neither the cognate sequences or any of the million randomsequences predicted to determine the rank were part of the training. (For completeness, Fig. 10 shows a 10-fold cross validation of predictions of the array data.) Each mAb was separately modeled, but all four concentrations were modeled at once as a four-column target. What is different in the analysis of this data from previous approaches is the way the input sequences and the targets were modified prior to training the neural network.

[0104] Padding and designation of N- and C-termini. The peptides on the array used for this analysis had a distribution of lengths from 5 to 1 1 residues, centered between 9 and 10 residues. To deal with varying lengths, sequences less than 11 residues were padded with a character “X” and the N-terminus was marked with a character “(” and the C-terminus marked with a character “)”. As a result, the sequences that were assigned one-hot representations had 19 different one-hot bits, 16 for the amino acids used in the synthesis and 3 for the “X()” characters. For example:(PWRGPWARV)XX (SEQ ID NO: 120)(LPGVQG)XXXXX (SEQ ID NO: 121 ) (GNFAYQRDG)XX (SEQ ID NO: 122)

[0105] This approach to designating the sequences becomes particularly important below when the position of the sequences within a larger character vector is varied.

[0106] Weighting of high binding data. As described in the text, the number of sequences that bound to mAbs with values substantially greater than background was often small. As a result, the background sequence-binding pairs often completely dominated the loss function in the training and key binding information was not emphasized. To avoid this, multiple copies of higher binding sequence-binding pairs were used. The sequences were binned by their binding values at a concentration of 8 nM (binwidth of 0.2, bining the logw binding values) and any bin with less than 300 values in it was expanded to 300 values by randomly picking sequence-binding pairs and copying them. No one sequence-binding pair was allowed to be copied more than 100 times.

[0107] Creating register shifted versions of the input vectors. One of the most important approaches used to improve the ability of the neural network to predict highbinding of mAb cognate sequences was by forcing the network to consider multiple frames of each sequence as being equivalent. Each sequence-binding pair in the training set was expanded to 6 copies, each in a different frame within the original character vector as follows:(PWRGPWARV)XXXXXXX (SEQ ID NO: 123) X(PWRGPWARV)XXXXXX (SEQ ID NO: 124) XX(PWRGPWARV)XXXXX (SEQ ID NO: 125) XXX(PWRGPWARV)XXXX (SEQ ID NO: 126) XXXX(PWRGPWARV)XXX (SEQ ID NO: 127) XXXXX(PWRGPWARV)XX (SEQ ID NO: 128)

[0108] In the training, the 6 sequences are all assigned the same measured binding value. While this greatly increased the number of sequence-binding pairs in the training set, and thus the computational intensity of training, it greatly improved the fidelity of the model. Evaluation of the model was performed in the same way. In other words, any sequence for which a binding value was to be predicted was also submitted as 6 copies in 6 frames as above and in this case, the highest binding value was then chosen for ranking that sequence. Since all sequences were processed the same way (in this case all million randomly generated sequences), this does not introduce any bias in the overall ranking.

[0109] Compositional binding estimates. Estimates of compositional binding for each sequence were determined by taking all sequences in the training set and assigning a simple vector of 16 composition values (the number of occurrences of each of the 16 amino acids used in that sequence). The values were then used in a linear fit (16 coefficients and a bias term). In the case of sequences to be predicted, the compositional binding was determined by applying the linear model coefficients learned from the fit of the training set.

[0110] Arrays and Data Quality

[0111] Peptide microarrays containing diverse peptides were synthesized at Cowper Sciences, Inc., (Chandler, AZ) as discussed in the main text. After completion of synthesis, each wafer was diced into 13 slides (25x75mm) and 4 MALDI-MS arrays. Slides were further subjected to a standard chemical cocktail to remove side chain protecting groups and stored under nitrogen prior to use. Synthesis verification wasperformed by high-resolution Matrix-Assisted Laser Desorption Ionization Mass Spectrometry (MALDI-MS), using a previously described protocol.1Briefly, N-termini of peptides are labeled with TMPP (tris(2,4,6-trimethoxyphenyl) phosphine) to provide net positive charge, followed by selective chemical cleavage from the surface using gaseous ammonia. Alpha-cyano-4-hydroxycinnamic acid (CHCA) is then applied to the array surface using an automated sprayer (TM-Sprayer, HTX Technologies, LLC). MALDI mass spectra are acquired from each MALDI feature (representing each amino acid monomer and / or peptide) on an Autoflex Speed MALDI MS (Bruker Daltonik). Synthesis is verified when desired peptides are identified on the expected m / z, with a set tolerance of 0.5Da and minimum signal-to-noise (S / N) of 3.

[0112] Fig. 6 shows the Pearson correlation coefficient between the mAb's logw binding values at each concentration and every other mAb’s logw binding values at each concentration. The four replicates of each mAb and each concentration were averaged prior to calculating the correlation coefficient. The data is organized from the lowest concentration (0.125 nM) to the highest concentration (8 nM) for each mAb. As one moves along the diagonal, one can see that different concentrations of any particular mAb are generally similar, and different mAbs are quite dissimilar from each other. No data was excluded from the analysis.

[0113] Distributions. Fig. 7 shows the distributions of each of the datasets (each mAb at each concentration). Note that both the frequency scale (y-axis) and the binding value (x-axis) are logw; binding is given in logw fluorescence counts. The background binding of the labeled secondary antibody to the arrays is -850 counts. There is considerable variation in the number of high binding value sequences on these arrays, with both Ab1 and Ab8 having very few sequences that bind with values above 3,000 counts (~3.5 on the log scale) and TF3B5 having hundreds of sequences with values that saturate the detector at 65,536 counts at the highest concentration (4.82 on a logw scale). The binding values of the mAbs to their cognate sequences are not shown, but with the exception of 4C1 , the signal from the cognate sequence is saturating even at 0.125 nM for those mAbs that had a cognate sequence synthesized on the array (Ab1 , Ab8, 4C1 , DM1 A, Lnkb2).

[0114] The distributions give an immediate indication of the specificity of each mAb. Since the binding data is from peptides sampled nearly randomly over all of the16 amino acid sequence space (1610= 1012sequences) without regard for any specific biological relevance, this platform gives a simple empirical comparison of specificity in an unbiased dataset. In general, the vast majority of the binding for all the mAbs is near the background level and in the case of Ab8, there is almost no binding more than a few times the background level. As can be seen in Fig. 6, the correlation between binding of different mAbs is generally low, presumably due to very different interactions and sequence requirements. Thus, what one learns about binding of one mAb is unlikely to be useful in characterizing the binding of another. For this reason they were modeled separately.Table 3: Number of Peptide Sequences >2 Standard Deviations Above the Median Binding Value

[0115] Number of Sequences Above the Median for Each mAb

[0116] Table 3 shows the number of sequences for each mAb at each concentration that was at least 2 standard deviations above the median value of binding for all sequences for that mAb and concentration. The standard deviation was calculated for each peptide, each mAb and each concentration from the four experimental replicates.

[0117] The numbers of mAbs above background by two standard deviations does not always grow in the way one might expect with concentration. This is for two reasons. First, there is a general increase in the median value as the concentration increases. Thus, the value being compared to is increasing. Second, if there happens to be an array that is uniformly scaled a bit higher or lower than the other replicates, that results in a large standard deviation which also causes the data to not exceed the 2 standard deviation threshold.

[0118] Linearity with concentration

[0119] In order to take into account the fact that at higher concentrations detector saturation occurs for some peptide binding, the points that saturated at one concentration are assumed to increase linearly with concentration at higher concentration. Fig. 8 justifies that assumption looking at the overall linearity of binding as a function of concentration for each of the mAbs. In each case, the average binding of all peptides that do not saturate at any concentration is plotted, showing that there is a linear dependence of binding on concentration at least within the dynamic range of the detector. Thus, the assay itself is very linear with concentration.

[0120] Cross validation study of binding prediction

[0121] Fig. 9 shows a 10-fold cross validation study of sequence-specific binding prediction models for each mAb as a density scatter plot2. Here the Iog10 predicted binding value is plotted against the Iog10 measured binding value. As described in the main text and above, saturating points were extrapolated based on concentration, which is why there are logio binding values >4.82. 90% of the peptide- binding pairs (randomly selected) were used to train the model in each case and 10% were used as the test set. This was done for all ten possible 90:10 combinations so that all peptides were ultimately included in the test set. Then the entire prediction was repeated 10 times and averaged. Fig. 9 plots 100,000 predicted test binding values against the measured values. The displayed values were randomly chosen from all 4 concentrations modeled. It also shows, as a filled circle with lines that are upward-facing to the right, the 8 nM predicted and measured value of each mAb cognate sequence when that value is present on the array. Note that the predicted value of the cognate is generally lower than the measured value, particularly in cases like Ab1 and Ab8 where there are very few peptides that bind strongly. However, even though the value predicted for the cognate is generally lower than its measured value, its rank remains high among the 100,000 points plotted. Note also that the cognate prediction is being compared to sequences of many lengths (the array peptides range from 5 to 1 1 residues), rather than only to sequences of the same length. The length need not be exactly the length of the cognate sequence as long as it contains the cognate sequence, but both the predictions and the measured data tend to show higher binding for longer peptides. Thus, when looking, for example, at binding to aprotein, one would want to tile the protein sequence as even length pieces for an accurate comparison of different regions of the protein.

[0122] Compositionally Dependent Binding

[0123] Fig. 10 shows scatter plots of sequence-dependent binding (y-axis) vs. compositional binding (x-axis) determined as described in Methods of the main text. Basically, a set of coefficients relating the number of each amino acid in the peptide to the binding of that amino acid is determined in a linear fit (there is also a bias term). The amount of binding that can be described this way is subtracted from the total binding (subtraction is performed after taking the logio). In every case but 4C1 , when the cognate epitope sequence in present on the array (shown as circles filled with lines that are upward-facing to the right) it represents the strongest sequence-specific binding (the cognate epitope sequences for TF3B5, C3 and 9E10.3 were not present on the array). Most of the mAbs have a very narrow range of composition-dependent binding. In other words, the only term in the linear fit that was really significant in most cases was the bias term describing the background binding. Only a small fraction of the binding distribution (0.05 on the log scale) could be explained by composition alone. The two exceptions to this rule are TF3B5 and 9E10.3, both of which show a distribution of compositional logw binding values >0.1. Even for these two mAbs, however, compositional binding is still modest compared to sequence specific logw binding of 1.5-2. The overall amount of sequence specific binding among the array peptides varies greatly between mAbs. Ab1 and Ab8 have the least number of sequence-specific binding sequences. Ab8 in particular is devoid of any sequences that bind with sequence-specific binding values above 0.5, while Ab1 has only 6 such sequences. In all other mAbs, at least the top 10 sequence-specific binding sequence values (unfilled circles) are higher than 0.5. TF3B5 is at the opposite end of the spectrum, having about 100 sequences that show substantial sequence specific binding, including a number that saturate the detector (~1.9 on the y-axis). Other mAbs are between these extremes. 4C1 is the one case where none of the sequences synthesized on the array to represent its cognate epitope sequence are the highest binding sequences. This suggests that at least in the context of binding to a short linear sequence attached to an array surface, this mAb may be less specific than the other mAbs; there are apparently multiple different sequences that bind this mAb aboutas strongly as the cognate epitope sequence. Note, however that at least in terms of prediction based on the neural network models, adding additional amino acids from the binding region of the antigen increases binding of 4C1 (see below).

[0124] Also shown in Fig. 10 are the ten highest binding sequences on the array (in black type to the right) as well as the sequences of the cognate epitope representations considered (circles filled with lines that are upward-facing to the right, only present for those mAbs where they were measured on the same array as the rest of the data). As seen in Fig. 3 of the above text, in every case but Ab8, at least several of the top ten sequences are clearly related to the cognate epitope sequence. In the case of Ab8, the relationship is more difficult to discern, though one can see a similarity in the particular amino acids used. The general similarity of some of the stronger sequence-dependent binding sequences to the corresponding cognate epitope suggests that sequence-specific information is indeed available on the array. It is important to remember that in this work, the epitopes are taken out of their folded protein environment. This may well decrease the binding value and specificity.

[0125] Ranking based on binding

[0126] Ranking of predicted cognate binding values is determined based on binding at the four concentrations of mAb. Each concentration was separately ranked and then the lowest rank of the four was assigned to each peptide. This can result with two peptides having the same rank, and thus the peptides were reranked so that the rankings went from 1 to N were N is the total number of peptides. The same algorithm is used for all of the million random peptides and the cognate, so there is no bias towards the cognate sequence. The reason this maximum rank approach was used is that different peptides lie in different parts of the dynamic range of the binding measurement. This was the least biased approach to letting the data decide what concentration to use for each peptide.

[0127] Top Random Sequences Predicted

[0128] The top 20 ranked sequences predicted in the 1 million sequence random library for each of the mAbs is given in Table 4. Note that the mAb cognate sequences are included as reference at the top. Again, one can generally pick out a commonality in sequence among them. In some instances, most notably Ab8, there are also unrelated sequences. For Ab8, those tend to be highly positively chargedsequences. Interestingly, the N-terminus of the P53 antigen that Ab8 was raised against has a very highly positively charged region.Table 4: Top 20 Sequences Predicted from Random Library of 1 Million

[0129] Randomized Sequence Order Control

[0130] Table 5 shows the result of training on a dataset (using the full model) in which the sequence order is randomized relative to the order of the binding values from the array. As expected, after averaging results across multiple training runs, the values approach the rank of 500,000 (50%).

[0131] Table 5 Results from Full Model After Training with Randomized Array Sequence Order

[0132] Varying the length of the antigen sequence used as the cognate sequence

[0133] In order to better understand both the ability of the model to work with different lengths of sequences and to understand to what extent sequence context in the cognate region of the antigen matters in the modeling, model predictions similar to those of the full model in Table 2 of the above text were repeated for sequence regions between 6 and 10 residues, where the longer sequences were taken from the known binding sequence region of the antigen. In each case the full model including weighting, shifting and removal of compositional binding was used. Table 6 shows thesequences used and Table 7 shows the resulting ranks out of 1 million random sequences of that length. While there is some variation in rank with sequence length, it is modest, with some of the predictions being slightly worse with sequences longer than 7 and some slightly better.Table 6: Antigen sequences used for rank prediction at various lengthsTable 7: Ranks out of 1 million random sequences1of the sequences in Table S41The one million random sequences in each case are all the same size as the antigen sequence used.

[0134] Mutation Substitution Matrices

[0135] Fig. 11 is similar to Fig. 4 in the main text, showing the predicted binding values of every single amino acid substitution of the cognate sequence of each of the mAbs. Fig. 4 showed only four of these and compared the predicted and measured results. Here all eight mAb predictions are shown (no measured values). The cognate reference sequences used here are the same as those in Table 2 of the above text, but not always identical to Fig. 4 of the above text.

[0136] Reducing the Total Number of Sequences in the Training Set Randomly

[0137] Fig. 12 shows the change in rank of the 7mer cognate sequence in a million random 7mer sequences as the number of sequences in the training set is reduced. In each case a random subset of sequences was chosen and the full model (weighting, shifting and compositional binding removal) was used (see the final column of Table 2 in the above text for comparison). The results are described in the above text.

[0138] Removing Similar Sequences from the Array Peptide Library Before Training

[0139] An important question is how strongly the models created depend on the similarity of a few sequences to the mAb cognate sequence. To test this, starting with the cognate 7mer sequence (Table 1 in the above text), first only peptides with no more than 6 amino acids in common were used in training (note there were none with all 7 so <=6 is the full set), then only those with no more than 5 in common, etc. down to only those with no more than 2 in common. The definition of “in common” used was having the right sequence and the right relative position, but occurring in any register within the 7mer (including registers where the cognate sequence and the comparator sequences only partially overlap). Thus, for SDLWKLL (SEQ ID NO: 14), the sequence RGRWADLGK (SEQ ID NO: 314) would be considered to have 3 in common with the cognate (XXXXXDLXK) because the sequence order and spacing of three amino acids match, even though the register is shifted. Table 8 shows how many sequences were removed to achieve each reduced training set. Table 9 shows the ranks that resulted using the full model (weighting, shifting and compositional binding) trained on each reduced peptide set.Table 8: Number of Peptides Removed to Reduce SimilarityTable 9: Ranks After Training on Peptide Sequences with Reduced Similarity to the Cognate

[0140] Some further aspects are defined in the following clauses:

[0141] Clause 1 : A computer-implemented method of predicting a biomolecule sequence to which at least one target antibody binds, the method comprising passing at least one target antibody data set through a model that relates antibody binding to biomolecule sequence, wherein the model was created using one or more antibodybiomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library, and wherein the target antibody data set is representative of one or more attributes of the target antibody and is sufficient for the model to produce a prediction of one or more biomolecule sequences to which the target antibody binds; and, outputting from the model one or more predicted cognate epitope sequences and / or one or more predicted off-target biomolecule sequences to which the target antibody binds, thereby predicting the biomolecule sequence to which the target antibody binds.

[0142] Clause 2: The method of Clause 1 , wherein the biomolecule sequences in the biomolecule library are substantially random.

[0143] Clause 3: The method of Clause 1 or Clause 2, wherein the biomolecule sequences in the biomolecule library are substantially non-random.

[0144] Clause 4: The method of any one of the preceding Clauses 1 -3, wherein the target antibody data set and / or the antibody-biomolecule binding data sets comprise structural, chemical, and / or physical property data of the target antibody and / or of the biomolecules in the biomolecule library.

[0145] Clause 5: The method of any one of the preceding Clauses 1 -4, wherein the target antibody data set and / or the antibody-biomolecule binding data sets comprise amino acid sequence information from the target antibody and / or from the biomolecules in the biomolecule library.

[0146] Clause 6: The method of any one of the preceding Clauses 1 -5, wherein the target antibody comprises a monoclonal antibody.

[0147] Clause 7: The method of any one of the preceding Clauses 1 -6, wherein the target antibody comprises a polyclonal antibody.

[0148] Clause 8: The method of any one of the preceding Clauses 1 -7, comprising passing multiple target antibody data sets through the model.

[0149] Clause 9: The method of any one of the preceding Clauses 1 -8, wherein the target antibody is from a sample obtained from a subject.

[0150] Clause 10: The method of any one of the preceding Clauses 1 -9, wherein the sample comprises blood or serum.

[0151] Clause 1 1 : The method of any one of the preceding Clauses 1 -10, wherein a trained electronic neural network comprises the model.

[0152] Clause 12: The method of any one of the preceding Clauses 1 -1 1 , wherein the electronic neural network was trained after weighting one or more biomolecule sequences in the antibody-epitope binding data sets.

[0153] Clause 13: The method of any one of the preceding Clauses 1 -12, wherein a given amino acid sequence in the biomolecule library is disposed at different positions in a larger frame register of possible amino acid residue positions such that the model was trained to recognize the given amino acid sequence as a sequence motif in multiple frame registers.

[0154] Clause 14: A method of generating a model that predicts targets of antibody binding, the method comprising: determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library to produce one or more antibody-biomolecule binding data sets; and, creating a model that relates antibody binding to biomolecule sequence using the antibody-biomolecule binding data sets, thereby generating the model that predicts targets of antibody binding.

[0155] Clause 15: The method of Clause 14, wherein the reference antibodies comprise at least one monoclonal antibody.

[0156] Clause 16: The method of Clause 14 or Clause 15, wherein the reference antibodies comprise at least one polyclonal antibody.

[0157] Clause 17: The method of any one of the preceding Clauses 14-16, wherein at least one of the biomolecules comprises a potential continuous epitope.

[0158] Clause 18: The method of any one of the preceding Clauses 14-17, wherein at least one of the biomolecules comprises a potential discontinuous epitope.

[0159] Clause 19: The method of any one of the preceding Clauses 14-18, wherein the biomolecule library comprises at least about 100 biomolecules, at least about 1000 biomolecules, at least about 10000 biomolecules, at least about 100000 biomolecules, at least about 1000000 biomolecules, at least about 10000000 biomolecules, or more biomolecules.

[0160] Clause 20: The method of any one of the preceding Clauses 14-19, wherein antigen molecules comprise the biomolecules in the biomolecule library.

[0161] Clause 21 : The method of any one of the preceding Clauses 14-20, wherein the biomolecules comprise at least portions of carbohydrate molecules, nucleic acid molecules, protein molecules, or a combination of two or more thereof.

[0162] Clause 22: The method of any one of the preceding Clauses 14-21 , wherein the biomolecules comprise amino acid sequences.

[0163] Clause 23: The method of any one of the preceding Clauses 14-22, wherein the amino acid sequences comprise natural amino acids and / or nonnatural amino acids.

[0164] Clause 24: The method of any one of the preceding Clauses 14-23, wherein the amino acid sequences comprise selected sequence lengths and wherein the biomolecule library comprises a selected level of coverage of a combinatorialspace that comprises substantially all amino acid sequence permutations of the sequence lengths.

[0165] Clause 25: The method of any one of the preceding Clauses 14-24, wherein the selected level of coverage comprises less than about 5%, less than about 2.5%, less than about 0.5%, less than about 0.25%, less than about 0.05%, less than about 0.025%, or less of the combinatorial space that comprises substantially all of the amino acid sequence permutations of the sequence lengths.

[0166] Clause 26: The method of any one of the preceding Clauses 14-25, wherein the biomolecule library comprises a substantially random sample of the amino acid sequence permutations in the combinatorial space.

[0167] Clause 27: The method of any one of the preceding Clauses 14-26, wherein the selected sequence lengths are in the range of about 3-15 amino acids, about 4-14 amino acids, about 5-13 amino acids, about 6-12 amino acids, about 7-11 amino acids, or about 8-10 amino acids.

[0168] Clause 28: The method of any one of the preceding Clauses 14-27, wherein the model relates the binding affinity values of the reference antibodies to amino acid sequences in each of the antibody-biomolecule binding data sets.

[0169] Clause 29: The method of any one of the preceding Clauses 14-28, comprising using one or more markers to designate an N-terminus and / or a C- terminus of one or more of the amino acid sequences.

[0170] Clause 30: The method of any one of the preceding Clauses 14-29, wherein a given amino acid sequence in the biomolecule library is disposed at different positions in a larger frame register of possible amino acid residue positions such that the model is trained to recognize the given amino acid sequence as a sequence motif in multiple frame registers.

[0171] Clause 31 : The method of any one of the preceding Clauses 14-30, wherein a number of different frame registers comprises at least about 3, at least about 4, at least about 5, at least about 6, at least about 7, at least about 8, at least about 9, at least about 10, or more different frame registers.

[0172] Clause 32: The method of any one of the preceding Clauses 14-31 , wherein the determining step comprises ascertaining the binding affinity values ofdifferent reference antibodies to the biomolecules in the biomolecule library substantially simultaneously with one another.

[0173] Clause 33: The method of any one of the preceding Clauses 14-32, wherein the determining step comprises ascertaining the binding affinity values of different concentrations of a given reference antibody to the biomolecules in the epitope library.

[0174] Clause 34: The method of any one of the preceding Clauses 14-33, comprising using a slope of concentration dependence as a metric for determining biomolecule sequences to which a target antibody binds.

[0175] Clause 35: The method of any one of the preceding Clauses 14-34, wherein the determining step comprises contacting the reference antibodies and the biomolecules in the biomolecule library.

[0176] Clause 36: The method of any one of the preceding Clauses 14-35, wherein the reference antibodies comprise fluorescent labels and wherein the determining step comprises detecting levels of fluorescent signals emitted from the fluorescent labels of the reference antibodies that are bound to the biomolecules to determine the binding affinity values of the reference antibodies to the biomolecules in the biomolecule library.

[0177] Clause 37: The method of any one of the preceding Clauses 14-36, wherein the determining step comprises binding secondary antibodies that comprise fluorescent labels to the reference antibodies that are bound to the biomolecules and detecting levels of fluorescent signals emitted from the fluorescent labels of the secondary antibodies that are bound to the reference antibodies to determine the binding affinity values of the reference antibodies to the biomolecules in the biomolecule library.

[0178] Clause 38: The method of any one of the preceding Clauses 14-37, wherein one or more solid supports comprise the biomolecules in the biomolecule library.

[0179] Clause 39: The method of any one of the preceding Clauses 14-38, wherein the biomolecules in the biomolecule library are arrayed on a substrate.

[0180] Clause 40: The method of any one of the preceding Clauses 14-39, wherein a plurality of beads comprises the biomolecules in the biomolecule library.

[0181] Clause 41 : The method of any one of the preceding Clauses 14-40, wherein the model comprises a trained machine learning model.

[0182] Clause 42: The method of any one of the preceding Clauses 14-41 , wherein the creating step comprises training a machine learning model to predict a binding affinity value of a target antibody to a target biomolecule using the antibodybiomolecule binding data sets.

[0183] Clause 43: The method of any one of the preceding Clauses 14-42, wherein an electronic neural network comprises the model.

[0184] Clause 44: The method of any one of the preceding Clauses 14-43, wherein a given reference antibody-biomolecule binding pair that comprises a binding affinity value that is higher than the binding affinity values of one or more other reference antibody-biomolecule binding pairs is weighted higher than the other reference antibody-biomolecule binding pairs when creating the model.

[0185] Clause 45: The method of any one of the preceding Clauses 14-44, comprising creating the model using representations of biomolecule sequences in the antibody-biomolecule binding data sets.

[0186] Clause 46: The method of any one of the preceding Clauses 14-45, wherein the representations comprise one-hot vector representations.

[0187] Clause 47: The method of any one of the preceding Clauses 14-46, wherein the representations comprise chemical parameters and / or chemical structural representations of monomers in the biomolecule sequences.

[0188] Clause 48: The method of any one of the preceding Clauses 14-47, wherein the representations comprise concatenated vectors of amino acid physical or chemical or biological properties.

[0189] Clause 49: The method of any one of the preceding Clauses 14-48, comprising fitting the antibody-biomolecule binding data sets to another model that is based on biomolecule sequence composition to produce fitted data sets prior to creating the model.

[0190] Clause 50: The method of any one of the preceding Clauses 14-49, comprising normalizing the antibody-biomolecule binding data sets using the fitted data sets.

[0191] Clause 51 : The method of any one of the preceding Clauses 14-50, comprising disregarding the binding affinity values of the reference antibodies to biomolecules in the biomolecule library that depend only on compositions of the biomolecules prior to creating the model.

[0192] Clause 52: The method of any one of the preceding Clauses 14-51 , comprising using the model to predict levels of binding of a target antibody, if any, to one or more biomolecule sequences of interest.

[0193] Clause 53: The method of any one of the preceding Clauses 14-52, wherein the biomolecule sequences of interest comprise predicted cognate epitope sequences.

[0194] Clause 54: The method of any one of the preceding Clauses 14-53, wherein the biomolecule sequences of interest comprise target antigen sequences.

[0195] Clause 55: The method of any one of the preceding Clauses 14-54, wherein the target antigen sequences are derived from at least one pathogenic agent.

[0196] Clause 56: The method of any one of the preceding Clauses 14-55, wherein the biomolecule sequences of interest comprise at least a portion of a whole proteome of protein sequences or at least portions of multiple whole proteomes of protein sequences.

[0197] Clause 57: The method of any one of the preceding Clauses 14-56, comprising using the model to identify potential target cognate epitope sequences within a library of biomolecule sequences, a target antigen or sets of target antigens, or a target proteome or sets of target proteomes.

[0198] Clause 58: The method of any one of the preceding Clauses 14-57, comprising using the model to identify potential off-target biomolecule sequences within a library of biomolecule sequences, a target antigen or sets of target antigens, or a target proteome or sets of target proteomes.

[0199] Clause 59: The method of any one of the preceding Clauses 14-58, comprising using the model to determine a likely functional significance of binding to enzymatic activity, sensing function, signal transduction, molecular binding interference, molecular binding increase, gene activity, transcription activity, translation activity, and / or post translational modification activity.

[0200] Clause 60: The method of any one of the preceding Clauses 14-59, comprising using the model to predict a potential autoantigen in a mammalian biological system.

[0201] Clause 61 : The method of any one of the preceding Clauses 14-60, comprising using the model to predict off-target binding of an antibody based therapeutic agent.

[0202] Clause 62: The method of any one of the preceding Clauses 14-61 , comprising using the model to predict off-target binding of an antibody used in a diagnostic assay.

[0203] Clause 63: The method of any one of the preceding Clauses 14-62, comprising using the model to predict an identity of a pathogen that gave rise to an antibody present in a sample.

[0204] Clause 64: The method of any one of the preceding Clauses 14-63, comprising using the model to predict one or more portions of an antigen to which a target antibody binds.

[0205] Clause 65: The method of any one of the preceding Clauses 14-64, comprising using the model in conjunction with a protein structure to define an epitope structure or noncontiguous binding site of an off-target binding interaction.

[0206] Clause 66: The method of any one of the preceding Clauses 14-65, wherein the model uses predicted antibody binding as a basis to predict epitope sequences to which antibodies bind.

[0207] Clause 67: A system for generating a model that predicts targets of antibody binding, the system comprising: a processor; and a memory communicatively coupled to the processor, the memory storing instructions which, when executed on the processor, perform operations comprising: receiving binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library to produce one or more antibody-biomolecule binding data sets; and, creating a model that relates antibody binding to biomolecule sequence using the antibody-biomolecule binding data sets.

[0208] Clause 68: A system for predicting a biomolecule sequence to which at least one target antibody binds, the system comprising: a processor; and a memory communicatively coupled to the processor, the memory storing instructions which,when executed on the processor, perform operations comprising: passing at least one target antibody data set through a model that relates antibody binding to biomolecule sequence, wherein the model was created using one or more antibody-biomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library, and wherein the target antibody data set is representative of one or more attributes of the target antibody and is sufficient for the model to produce a prediction of one or more biomolecule sequences to which the target antibody binds; and, outputting from the model one or more predicted cognate epitope sequences and / or one or more predicted off-target biomolecule sequences to which the target antibody binds.

[0209] Clause 69: A computer readable media comprising non-transitory computer executable instructions which, when executed by at least one electronic processor, perform at least: receiving binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library to produce one or more antibody- biomolecule binding data sets; and creating a model that relates antibody binding to biomolecule sequence using the antibody-biomolecule binding data sets.

[0210] Clause 70: A computer readable media comprising non-transitory computer executable instructions which, when executed by at least one electronic processor, perform at least: passing at least one target antibody data set through a model that relates antibody binding to biomolecule sequence, wherein the model was created using one or more antibody-biomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library, and wherein the target antibody data set is representative of one or more attributes of the target antibody and is sufficient for the model to produce a prediction of one or more biomolecule sequences to which the target antibody binds; and, outputting from the model one or more predicted cognate epitope sequences and / or one or more predicted off-target biomolecule sequences to which the target antibody binds.

[0211] Clause 71 : A computer-implemented method of predicting a biomolecule sequence to which at least one target antibody binds, the method comprising: collecting a set of target antibody data, wherein the target antibody data set is representative of one or more attributes of the target antibody; creating a model usingone or more antibody-biomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library wherein the model was created; applying the model to the target antibody data, wherein the model relates antibody binding to the biomolecule sequence; and predicting one or more target and / or off-target biomolecule sequences to which the target antibody binds using the application of the model to the target antibody data.

[0212] Clause 72: The method of Clause 71 , wherein the data is sufficient for the model to produce a prediction of one or more biomolecule sequences to which the target antibody binds.

[0213] Clause 73: The method of Clause 71 or Clause 72, wherein the biomolecule sequences in the biomolecule library are substantially random.

[0214] Clause 74: The method of any one of the preceding Clauses 71 -73, wherein the biomolecule sequences in the biomolecule library are substantially nonrandom.

[0215] Clause 75: The method of any one of the preceding Clauses 71 -74, wherein the target antibody data set and / or the antibody-biomolecule binding data sets comprise structural, chemical, and / or physical property data of the target antibody and / or of the biomolecules in the biomolecule library.

[0216] Clause 76: The method of any one of the preceding Clauses 71 -75, wherein the target antibody data set and / or the antibody-biomolecule binding data sets comprise amino acid sequence information from the target antibody and / or from the biomolecules in the biomolecule library.

[0217] Clause 77: The method of any one of the preceding Clauses 71 -76, wherein the target antibody comprises a monoclonal antibody.

[0218] Clause 78: The method of any one of the preceding Clauses 71 -77, wherein the target antibody comprises a polyclonal antibody.

[0219] Clause 79: The method of any one of the preceding Clauses 71 -78, comprising passing multiple target antibody data sets through the model.

[0220] Clause 80: The method of any one of the preceding Clauses 71 -79, wherein the target antibody is from a sample obtained from a subject.

[0221] Clause 81 : The method of any one of the preceding Clauses 71 -80, wherein the sample comprises blood or serum.

[0222] Clause 82: The method of any one of the preceding Clauses 71 -81 , wherein the model comprises a trained electronic neural network.

[0223] Clause 83: The method of any one of the preceding Clauses 71 -82, wherein the electronic neural network was trained after weighting one or more biomolecule sequences in the antibody-epitope binding data sets.

[0224] Clause 84: The method of any one of the preceding Clauses 71 -83, wherein a given amino acid sequence in the biomolecule library is disposed at different positions in a larger frame register of possible amino acid residue positions such that the model was trained to recognize the given amino acid sequence as a sequence motif in multiple frame registers.

[0225] Clause 85: A computer-implemented method of training an electronic neural network for predicting a biomolecule sequence to which at least one target antibody binds, the method comprising: collecting a set of target antibody data from a sample, wherein the target antibody data set is representative of one or more attributes of the target antibody; creating a training set comprising one or more antibodybiomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library wherein the electronic neural network was created; applying the electronic neural network to the target antibody data, wherein the electronic neural network relates antibody binding to the biomolecule sequence; and predicting one or more target and / or off-target biomolecule sequences to which the target antibody binds using the application of the electronic neural network to the target antibody data.

[0226] While the invention has been described with reference to the exemplary embodiments thereof, those skilled in the art will be able to make various modifications to the described embodiments without departing from the true spirit and scope. The terms and descriptions used herein are set forth by way of illustration only and are not meant as limitations. In particular, although the method has been described by examples, the steps of the method can be performed in a different order than illustrated or simultaneously. Those skilled in the art will recognize that these and other variations are possible within the spirit and scope as defined in the following claims and their equivalents.

Claims

What is claimed is:1 . A computer-implemented method of predicting a biomolecule sequence to which at least one target antibody binds, the method comprising: passing at least one target antibody data set through a model that relates antibody binding to biomolecule sequence, wherein the model was created using one or more antibody-biomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library, and wherein the target antibody data set is representative of one or more attributes of the target antibody and is sufficient for the model to produce a prediction of one or more biomolecule sequences to which the target antibody binds; and, outputting from the model one or more predicted cognate epitope sequences and / or one or more predicted off-target biomolecule sequences to which the target antibody binds, thereby predicting the biomolecule sequence to which the target antibody binds.

2. The method of claim 1 , wherein the biomolecule sequences in the biomolecule library are substantially random.

3. The method of claim 1 , wherein the biomolecule sequences in the biomolecule library are substantially non-random.

4. The method of claim 1 , wherein the target antibody data set and / or the antibody-biomolecule binding data sets comprise structural, chemical, and / or physical property data of the target antibody and / or of the biomolecules in the biomolecule library.

5. The method of claim 1 , wherein the target antibody data set and / or the antibody-biomolecule binding data sets comprise amino acid sequence information from the target antibody and / or from the biomolecules in the biomolecule library.

6. The method of claim 1 , wherein the target antibody comprises a monoclonal antibody.

7. The method of claim 1 , wherein the target antibody comprises a polyclonal antibody.

8. The method of claim 1 , comprising passing multiple target antibody data sets through the model.

9. The method of claim 1 , wherein the target antibody is from a sample obtained from a subject.

10. The method of claim 9, wherein the sample comprises blood or serum.11 . The method of claim 1 , wherein a trained electronic neural network comprises the model.

12. The method of claim 11 , wherein the electronic neural network was trained after weighting one or more biomolecule sequences in the antibody-epitope binding data sets.

13. The method of claim 1 , wherein a given amino acid sequence in the biomolecule library is disposed at different positions in a larger frame register of possible amino acid residue positions such that the model was trained to recognize the given amino acid sequence as a sequence motif in multiple frame registers.

14. A method of generating a model that predicts targets of antibody binding, the method comprising: determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library to produce one or more antibody-biomolecule binding data sets; and, creating a model that relates antibody binding to biomolecule sequence usingthe antibody-biomolecule binding data sets, thereby generating the model that predicts targets of antibody binding.

15. The method of claim 14, wherein the reference antibodies comprise at least one monoclonal antibody.

16. The method of claim 14, wherein the reference antibodies comprise at least one polyclonal antibody.

17. The method of claim 14, wherein at least one of the biomolecules comprises a potential continuous epitope.

18. The method of claim 14, wherein at least one of the biomolecules comprises a potential discontinuous epitope.

19. The method of claim 14, wherein the biomolecule library comprises at least about 100 biomolecules, at least about 1000 biomolecules, at least about 10000 biomolecules, at least about 100000 biomolecules, at least about 1000000 biomolecules, at least about 10000000 biomolecules, or more biomolecule.

20. The method of claim 14, wherein antigen molecules comprise the biomolecules in the biomolecule library.21 . The method of claim 14, wherein the biomolecules comprise at least portions of carbohydrate molecules, nucleic acid molecules, protein molecules, or a combination of two or more thereof.

22. The method of claim 14, wherein the biomolecules comprise amino acid sequences.

23. The method of claim 22, wherein the amino acid sequences comprise natural amino acids and / or nonnatural amino acids.

24. The method of claim 22, wherein the amino acid sequences comprise selected sequence lengths and wherein the biomolecule library comprises a selected level of coverage of a combinatorial space that comprises substantially all amino acid sequence permutations of the sequence lengths.

25. The method of claim 24, wherein the selected level of coverage comprises less than about 5%, less than about 2.5%, less than about 0.5%, less than about 0.25%, less than about 0.05%, less than about 0.025%, or less of the combinatorial space that comprises substantially all of the amino acid sequence permutations of the sequence lengths.

26. The method of claim 24, wherein the biomolecule library comprises a substantially random sample of the amino acid sequence permutations in the combinatorial space.

27. The method of claim 24, wherein the selected sequence lengths are in the range of about 3-15 amino acids, about 4-14 amino acids, about 5-13 amino acids, about 6-12 amino acids, about 7-11 amino acids, or about 8-10 amino acids.

28. The method of claim 22, wherein the model relates the binding affinity values of the reference antibodies to amino acid sequences in each of the antibodybiomolecule binding data sets.

29. The method of claim 22, comprising using one or more markers to designate an N-terminus and / or a C-terminus of one or more of the amino acid sequences.

30. The method of claim 22, wherein a given amino acid sequence in the biomolecule library is disposed at different positions in a larger frame register of possible amino acid residue positions such that the model is trained to recognize the given amino acid sequence as a sequence motif in multiple frame registers.31 . The method of claim 30, wherein a number of different frame registers comprises at least about 3, at least about 4, at least about 5, at least about 6, at least about 7, at least about 8, at least about 9, at least about 10, or more different frame registers.

32. The method of claim 14, wherein the determining step comprises ascertaining the binding affinity values of different reference antibodies to the biomolecules in the biomolecule library substantially simultaneously with one another.

33. The method of claim 14, wherein the determining step comprises ascertaining the binding affinity values of different concentrations of a given reference antibody to the biomolecules in the biomolecule library.

34. The method of claim 33, comprising using a slope of concentration dependence as a metric for determining biomolecule sequences to which a target antibody binds.

35. The method of claim 14, wherein the determining step comprises contacting the reference antibodies and the biomolecules in the biomolecule library.

36. The method of claim 35, wherein the reference antibodies comprise fluorescent labels and wherein the determining step comprises detecting levels of fluorescent signals emitted from the fluorescent labels of the reference antibodies that are bound to the biomolecules to determine the binding affinity values of the reference antibodies to the biomolecules in the biomolecule library.

37. The method of claim 35, wherein the determining step comprises binding secondary antibodies that comprise fluorescent labels to the reference antibodies that are bound to the biomolecules and detecting levels of fluorescent signals emitted from the fluorescent labels of the secondary antibodies that arebound to the reference antibodies to determine the binding affinity values of the reference antibodies to the biomolecules in the biomolecule library.

38. The method of claim 35, wherein one or more solid supports comprise the biomolecules in the biomolecule library.

39. The method of claim 38, wherein the biomolecules in the biomolecule library are arrayed on a substrate.

40. The method of claim 38, wherein a plurality of beads comprises the biomolecules in the biomolecule library.41 . The method of claim 14, wherein the model comprises a trained machine learning model.

42. The method of claim 14, wherein the creating step comprises training a machine learning model to predict a binding affinity value of a target antibody to a target biomolecule using the antibody-biomolecule binding data sets.

43. The method of claim 42, wherein an electronic neural network comprises the model.

44. The method of claim 14, wherein a given reference antibody- biomolecule binding pair that comprises a binding affinity value that is higher than the binding affinity values of one or more other reference antibody-biomolecule binding pairs is weighted higher than the other reference antibody-biomolecule binding pairs when creating the model.

45. The method of claim 44, comprising creating the model using representations of biomolecule sequences in the antibody-biomolecule binding data sets.

46. The method of claim 45, wherein the representations comprise one-hot vector representations.

47. The method of claim 45, wherein the representations comprise chemical parameters and / or chemical structural representations of monomers in the biomolecule sequences.

48. The method of claim 45, wherein the representations comprise concatenated vectors of amino acid physical or chemical or biological properties.

49. The method of claim 14, comprising fitting the antibody-biomolecule binding data sets to another model that is based on biomolecule sequence composition to produce fitted data sets prior to creating the model.

50. The method of claim 49, comprising normalizing the antibody- biomolecule binding data sets using the fitted data sets.51 . The method of claim 14, comprising disregarding the binding affinity values of the reference antibodies to biomolecules in the biomolecule library that depend only on compositions of the biomolecules prior to creating the model.

52. The method of claim 14, comprising using the model to predict levels of binding of a target antibody, if any, to one or more biomolecule sequences of interest.

53. The method of claim 51 , wherein the biomolecule sequences of interest comprise predicted cognate epitope sequences.

54. The method of claim 52, wherein the biomolecule sequences of interest comprise target antigen sequences.

55. The method of claim 54, wherein the target antigen sequences arederived from at least one pathogenic agent.

56. The method of claim 52, wherein the biomolecule sequences of interest comprise at least a portion of a whole proteome of protein sequences or at least portions of multiple whole proteomes of protein sequences.

57. The method of claim 14, comprising using the model to identify potential target cognate epitope sequences within a library of biomolecule sequences, a target antigen or sets of target antigens, or a target proteome or sets of target proteomes.

58. The method of claim 14, comprising using the model to identify potential off-target biomolecule sequences within a library of biomolecule sequences, a target antigen or sets of target antigens, or a target proteome or sets of target proteomes.

59. The method of claim 14, comprising using the model to determine a likely functional significance of binding to enzymatic activity, sensing function, signal transduction, molecular binding interference, molecular binding increase, gene activity, transcription activity, translation activity, and / or post translational modification activity.

60. The method of claim 14, comprising using the model to predict a potential autoantigen in a mammalian biological system.61 . The method of claim 14, comprising using the model to predict off- target binding of an antibody based therapeutic agent.

62. The method of claim 14, comprising using the model to predict off- target binding of an antibody used in a diagnostic assay.

63. The method of claim 14, comprising using the model to predict anidentity of a pathogen that gave rise to an antibody present in a sample.

64. The method of claim 14, comprising using the model to predict one or more portions of an antigen to which a target antibody binds.

65. The method of claim 14, comprising using the model in conjunction with a protein structure to define an epitope structure or noncontiguous binding site of an off-target binding interaction.

66. The method of claim 14, wherein the model uses predicted antibody binding as a basis to predict epitope sequences to which antibodies bind.

67. A system for generating a model that predicts targets of antibody binding, the system comprising: a processor; and a memory communicatively coupled to the processor, the memory storing instructions which, when executed on the processor, perform operations comprising: receiving binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library to produce one or more antibody-biomolecule binding data sets; and, creating a model that relates antibody binding to biomolecule sequence using the antibody-biomolecule binding data sets.

68. A system for predicting a biomolecule sequence to which at least one target antibody binds, the system comprising: a processor; and a memory communicatively coupled to the processor, the memory storing instructions which, when executed on the processor, perform operations comprising: passing at least one target antibody data set through a model that relates antibody binding to biomolecule sequence, wherein the model was created using one or more antibody-biomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to biomolecules in anepitope library, and wherein the target antibody data set is representative of one or more attributes of the target antibody and is sufficient for the model to produce a prediction of one or more biomolecule sequences to which the target antibody binds; and, outputting from the model one or more predicted cognate epitope sequences and / or one or more predicted off-target biomolecule sequences to which the target antibody binds.

69. A computer readable media comprising non-transitory computer executable instructions which, when executed by at least one electronic processor, perform at least: receiving binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library to produce one or more antibody-biomolecule binding data sets; and, creating a model that relates antibody binding to biomolecule sequence using the antibody-biomolecule binding data sets.

70. A computer readable media comprising non-transitory computer executable instructions which, when executed by at least one electronic processor, perform at least: passing at least one target antibody data set through a model that relates antibody binding to biomolecule sequence, wherein the model was created using one or more antibody-biomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library, and wherein the target antibody data set is representative of one or more attributes of the target antibody and is sufficient for the model to produce a prediction of one or more biomolecule sequences to which the target antibody binds; and, outputting from the model one or more predicted cognate epitope sequences and / or one or more predicted off-target biomolecule sequences to which the target antibody binds.71 . A computer-implemented method of predicting a biomolecule sequence to which at least one target antibody binds, the method comprising: collecting a set of target antibody data, wherein the target antibody data set is representative of one or more attributes of the target antibody; creating a model using one or more antibody-biomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library wherein the model was created; applying the model to the target antibody data, wherein the model relates antibody binding to the biomolecule sequence; and predicting one or more target and / or off-target biomolecule sequences to which the target antibody binds using the application of the model to the target antibody data.

72. The method of claim 71 , wherein the data is sufficient for the model to produce a prediction of one or more biomolecule sequences to which the target antibody binds.

73. The method of claim 71 , wherein the biomolecule sequences in the biomolecule library are substantially random.

74. The method of claim 72, wherein the biomolecule sequences in the biomolecule library are substantially non-random.

75. The method of claim 71 , wherein the target antibody data set and / or the antibody-biomolecule binding data sets comprise structural, chemical, and / or physical property data of the target antibody and / or of the biomolecules in the biomolecule library.

76. The method of claim 71 , wherein the target antibody data set and / or the antibody-biomolecule binding data sets comprise amino acid sequence information from the target antibody and / or from the biomolecules in the biomolecule library.

77. The method of claim 71 , wherein the target antibody comprises a monoclonal antibody.

78. The method of claim 71 , wherein the target antibody comprises a polyclonal antibody.

79. The method of claim 71 , comprising passing multiple target antibody data sets through the model.

80. The method of claim 71 , wherein the target antibody is from a sample obtained from a subject.81 . The method of claim 80, wherein the sample comprises blood or serum.

82. The method of claim 71 , wherein the model comprises a trained electronic neural network.

83. The method of claim 82, wherein the electronic neural network was trained after weighting one or more biomolecule sequences in the antibody-epitope binding data sets.

84. The method of claim 83, wherein a given amino acid sequence in the biomolecule library is disposed at different positions in a larger frame register of possible amino acid residue positions such that the model was trained to recognize the given amino acid sequence as a sequence motif in multiple frame registers.

85. A computer-implemented method of training an electronic neural network for predicting a biomolecule sequence to which at least one target antibody binds, the method comprising: collecting a set of target antibody data from a sample, wherein the targetantibody data set is representative of one or more attributes of the target antibody; creating a training set comprising one or more antibody-biomolecule binding data sets produced by determining binding affinity values of one or more reference antibodies to biomolecules in a biomolecule library wherein the electronic neural network was created; applying the electronic neural network to the target antibody data, wherein the electronic neural network relates antibody binding to the biomolecule sequence; and predicting one or more target and / or off-target biomolecule sequences to which the target antibody binds using the application of the electronic neural network to the target antibody data.

Images