B-cell epitope prediction

Abstract

Predicting whether a protein contains B cell epitopes that are likely to trigger antibody binding events. [Solution] (a) accessing the amino acid sequence of a protein; (b) accessing one or more structural and / or surface properties of a protein; (c) generating one or more candidate codes for each of one or more candidate B cell epitopes on the protein, where each candidate code represents the B cell epitope as multiple data elements corresponding to amino acids in the amino acid sequence; and (d) implementing a computer to predict, using a first trained machine learning model, which is likely to be a true B cell epitope on the protein.

Description

[Technical Field]

[0001] background Lymphocytes are a type of white blood cell and are a central component of the immune system in most vertebrates. They include B cells, T cells, and natural killer (NK) cells. Their function is to identify pathogens such as bacteria, viruses, and malignant cancer cells with mutated protein antigens, and to protect against them. Both B cells and T cells express surface receptors that identify specific molecular components on pathogens called antigens (Ag). Successful recognition of Ag triggers a broader adaptive immune response, which can ultimately eliminate the potential pathogenic threat.

[0002] Antibodies (Ab), also known as immunoglobulins (Ig), are glycosylated protein molecules (surface immunoglobulins) or B cell-eluting Ig molecules present on the surface of B cells that act as antigen-specific B cell receptors (BCRs). Upon successful Ag recognition, Ig or membrane-bound BCRs bind to specific Ag amino acid contact sites, collectively referred herein as B cell epitopes (BCEs). This binding event can trigger the activation of an immune response, including clonal B cell proliferation that produces large quantities of effector B cells possessing Ag-specific Ig serum-eluting molecules or Ag-specific BCRs. The secreted soluble Ig and effector BCRs, also called antibodies (Ab), have the same binding specificity as the original Ig-Ag binding interaction event. The Abs then circulate in the serum and bind to the same BCEs, signaling the elimination of pathogenic or cellularly stressed cells corresponding to the identified Ag.

[0003] Ab-based vaccines have the potential to prepare or develop each individual's immune system against potentially harmful Ag, resulting from infectious pathogenic threats or stressed malignant cancer cells. Vaccine design aims to modify Ag in a way that bypasses harmful malignant, pathogenic, or infectious properties while maintaining a potential protective immune response driven by immunogenic epitopes. Therefore, accurate identification or prediction of B-cell epitopes on antigens is crucial not only for vaccine design but also in the areas of molecular diagnostics, immune monitoring, and immunotherapy.

[0004] BCEs can be classified into two main categories: linear B cell epitopes (LBCEs) and conformational B cell epitopes (CBCEs). Both LBCEs and CBCEs can bind to Abe and therefore trigger an immune response. It is estimated that over 90% of BCEs are conformational, while the remainder are linear.

[0005] LBCEs consist of sequential amino acid sequences. These BCEs can be discovered by identifying the position on an unfolded Ag sequence to which Ab can bind. LBCEs are often cataloged in immunoepitope databases (IEDBs) derived from a wide variety of experimental assays. Each assay provides information about the tested protein sequence (i.e., Ag), the LBCE sequence and its coordinates on Ag, and the Ab to which the current LBCE was tested. Furthermore, this provides an indicator of whether the LBCE was bound by a given Ab. Binding information may also be described in the IEDB in the form of human-readable labels, i.e., positive, highly positive, moderately positive, lowly positive, and negative. Because LBCEs are convenient as linear molecular entities and are characterized using well-established experimental methods, a relatively large amount of available LBCE data is available compared to CBCE data. Nevertheless, annotation errors frequently occur in this type of data because many amino acids in the identified BCEs do not actually interact with Ab under the tertiary conformation of Ag.

[0006] CBCEs are composed of short segments of amino acids with a variety of potential cis, trans, short, or distant coordinate distances between each of the amino acid contact points constituting the BCE. These segments are brought into close proximity when Ag folds during binding to Ab. CBCEs can be experimentally characterized using X-ray crystallography of Ag-Ab structural complexes. This data is typically cataloged in the Protein Databank (PDB) in the form of 3D structures, and usually consists of one or more Ag structures bound by at least one Ab structure. PDB structural data provides information about the 3D coordinates of each molecule in the structure. However, because this method is complex, CBCE data is somewhat limited compared to LBCE data. Moreover, 3D structures may not accurately reflect the true CBCE. For example, Ab on the complex may be missing its heavy or light chain, or the 3D structure may only provide Ag-Ab contacts outside the CDR region. Furthermore, the extremely sparse available information regarding undiscovered CBCEs on any Ag, and ground truth-negative epitopes on Ag, makes predicting true CBCEs even more difficult.

[0007] Various algorithms exist for predicting either LBCE and / or CBCE. However, while many of these algorithms perform well on test data, they do not yield the same satisfactory results on new, independent test data when compared to other algorithms of a similar type.

[0008] In conventional CBCE prediction algorithms, the expected input is either the linear protein sequence of Ag or the 3D protein structure file of Ag. Algorithms that require the 3D protein structure as input generally produce better results than those that only require the sequence. However, due to the labor-intensive and expensive scale of these laboratory experiments, the 3D protein structure of Ag is almost certainly unavailable for prediction in most use cases, which limits the use of 3D structure input algorithms. In the case of algorithms that require the sequence as input, it is difficult to capture the relationship between spatially separated amino acids and distinguish between potentially different CBCEs present on the same antigen. Summary of the Invention Problems to be Solved by the Invention

[0009] Therefore, it is desirable to improve the prediction of B cell epitopes, particularly CBCE. Means for Solving the Problems

[0010] Summary of the Invention According to a first aspect of the present invention, there is provided a computer-implemented method for predicting whether a protein contains a B cell epitope likely to induce a binding event with an antibody, the method comprising: (a) accessing the amino acid sequence of the protein; and (b) accessing one or more structural and / or surface characteristics of the protein; and (c) generating one or more candidate codes for each of the one or more candidate B cell epitopes on the protein, each candidate code representing a respective candidate B cell epitope as a plurality of data elements corresponding to the amino acids of the amino acid sequence. (d) Using the trained first machine learning model to predict whether any of the candidate B cell epitopes are likely to be true B cell epitopes on the protein, wherein the input to the trained first machine learning model includes the amino acid sequence of the protein, one or more structures and / or surface characteristics of the protein, and one or more candidate codes, and the prediction A method is provided that includes.

[0011] The inventors recognized that a common problem with many conventional sequence-based techniques used to predict the presence of B cell epitopes on proteins is that the output predictions are typically in the form of binary values per amino acid indicating whether each particular amino acid in the protein sequence is part of a B cell epitope. This conventional approach does not allow the user to distinguish potentially different distinguishable BCEs present on a protein having several overlapping amino acid residues, and moreover does not necessarily consider the relationship between widely separated amino acids across the protein sequence being analyzed in 3D conformational BCEs.

[0012] To improve these prior arts for BCE prediction, the present invention applies a trained machine learning model to a combination of the input amino acid sequence of the protein under investigation and one or more candidate codes for each of the one or more candidate B cell epitopes, where each candidate code represents a respective candidate B cell epitope.

[0013] Thus, the predictions provided by the present invention are made "epitope-wise" based on the query protein (by considering each candidate code corresponding to a candidate epitope) rather than "amino acid-wise" approaches. Accordingly, the present invention considers whether a candidate B cell epitope encoded on the amino acid sequence under test is (or is not) a true BCE on the protein. This approach used by the present invention is particularly advantageous because it eliminates the need to infer the presence of an epitope (or multiple overlapping epitopes) based on individual amino acid scores, and as a whole, leads to increased confidence in the actual presence (or absence) of an epitope on the protein. Providing "epitope-wise" predictions also leads to improved accuracy and reliability in predicting multiple (and potentially overlapping) epitopes on the same protein.

[0014] Typically, the output of the first trained machine learning model is the probability that each candidate code represents a true B cell epitope on the protein.

[0015] In this specification, the term “true” B cell epitope is used to mean a B cell epitope that triggers an antibody binding event (e.g., one that triggers and / or is triggered by an immunogenic response).

[0016] A B cell epitope can be a conformational B cell epitope. A B cell epitope can be a linear B cell epitope. The first trained machine learning model of the present invention can typically predict the presence of both linear and conformational B cell epitopes, but in some embodiments, different models may be used to predict linear and conformational B cell epitopes, respectively, based on the reference data used for training.

[0017] Generally, a protein can be a protein, a protein domain, or a protein subunit. The term “protein” can include any protein subsequence that may have a viable 3D or functional structure. Typically, a protein contains an antigen. A protein may contain a nascent antigen. The amino acid sequence of a protein can be obtained using techniques known to those skilled in the art, for example, by one of the following: oligonucleotide hybridization, nucleic acid amplification-based methods (including, but not limited to, polymerase chain reaction-based methods), DNA or RNA sequencing-based automated prediction, de novo peptide sequencing, Edman sequencing, or any peptide-related sequencing, including, but not limited to, mass spectrometry. Amino acid sequences can be downloaded from bioinformatics depositories such as UniProt (www.uniprot.org).

[0018] In some embodiments, one or more candidate codes represent all possible combinations of amino acids in an amino acid sequence. This can include both linear B-cell epitopes (where each encoded amino acid is located contiguously in the amino acid sequence) and conformational B-cell epitopes (where at least some of the encoded amino acids are spaced apart in the amino acid sequence). However, in some embodiments, prior knowledge of the protein and / or B-cell epitope configuration can be used to reduce the search space for codes. For example, if the protein contains a nascent antigen, the codes can be fixed at the locations of mutation sites (e.g., each code will have identical data elements at the locations of mutation sites in the amino acid sequence).

[0019] One or more candidate codes representing amino acid combinations within an amino acid sequence can be queried via a deep reinforcement learning model. One or more candidate codes representing amino acid combinations within an amino acid sequence can be supplied from a generative model. In some embodiments, deep reinforcement learning (DRL) techniques can be used to efficiently reduce or query a large search space.

[0020] As outlined above, candidate codes for B cell epitopes represent each candidate B cell epitope as multiple data elements corresponding to amino acids in an amino acid sequence. Typically, each data element of a candidate code (or at least a subset of multiple data elements) corresponds to an amino acid in the amino acid sequence of a protein. Preferably, each of one or more codes is in the form of a binary vector. In such embodiments, the binary vector is typically in the form of a sequence of "1"s and "0"s, where each "1" corresponds to an amino acid in the protein that is part of the candidate epitope (e.g., a predicted Ag-Ab contact point), and each "0" corresponds to an amino acid in the protein that does not form part of the candidate epitope (e.g., not a predicted Ag-Ab contact point). In some embodiments, different coding protocols may be used to represent candidate epitopes (e.g., one-hot coding).

[0021] The first machine learning model can be (or may include) any neural network, such as a convolutional neural network or a feedforward neural network. Preferably, the first machine learning model is (or includes) a recurrent neural network. Typically, the first machine learning model is (or includes) a long-short-term memory network (LSTM), preferably a bidirectional long-short-term memory network (BLSTM). Typically, each amino acid in a query protein sequence is treated as a time step in the network, and its properties may be influenced by both preceding and succeeding amino acids. Thus, an entire protein amino acid sequence (e.g., paired with other input features) can be treated as a "single data point," in contrast to conventional approaches that require segmenting the protein amino acid sequence into smaller amino acid subsequences, e.g., "k-mers." In this way, the present invention advantageously enables the discovery of relationships between spatially distant amino acids in a protein sequence that may be part of the same (e.g., conformational) BCE, relationships that cannot be discovered by the "k-mer" approach. The use of BLSTMs is particularly advantageous because it allows processing of protein amino acid sequences from both directions.

[0022] As described above, in step (b) of this method, one or more (e.g., three-dimensional) structural and / or surface (e.g., surface exposure) properties of the protein are accessed. Since Ab is highly specific to its binding BCE, the structural and / or surface properties of the protein under investigation are important for predicting B cell epitopes (particularly CBCEs). One or more structural and / or surface properties of the protein may be properties at the whole protein level (e.g., secondary structure) or at the amino acid level (e.g., features related to individual amino acids of the protein). One or more structural and / or surface properties may be in the form of continuous features (such as relative solvent-accessible RSA or hemispheric-exposed HSA) and / or categorical features such as secondary structure.

[0023] In embodiments, one or more structural and / or surface properties include one or more of the following: secondary structure of the protein; relative solvent accessibility (RSA); and hemispheric exposure (HSA). The HSA value may be divided into upper hemispheric exposure and lower hemispheric exposure values. SS may be considered a structural property, while RSA and HSE may be considered surface properties. A particular advantage of the present invention is that it does not require the fully folded three-dimensional structure of the input protein (which may be difficult to measure or, if predicted, may contain a large number of errors). Instead, the method accesses one or more structural and / or surface properties of the protein that are more advantageously easier to obtain than the fully folded structure and typically contain fewer errors. The inventors have found that RSA, HSE, and secondary structure properties provide reliable indicators of the three-dimensional structure of the protein and therefore provide particularly good results for prediction. Further structural and / or surface properties may also be used as input to the model. Thus, the method of the present invention does not require obtaining the complete three-dimensional structure of the protein under investigation, and can provide good BCE prediction results by using the amino acid sequence of the input protein and a limited number of its structural and / or surface properties. In some embodiments, as further described herein, the protein sequence may be used in combination with one or more further machine learning models to obtain the structure and / or surface properties.

[0024] Typically, the structure and / or surface properties accessed are in the form of values ​​and / or classes assigned to each amino acid in the protein amino acid sequence that represent the structure and / or surface properties of the protein. Thus, the input to the first machine learning model may be a set of amino acid-by-amino acid features of the protein (e.g., data elements from binary code along with one or more structure and / or surface properties of amino acids).

[0025] The method may further include accessing one or more biochemical properties of a protein and / or the amino acids that make up the protein, and inputting one or more biochemical properties (in addition to one or more structural and / or surface properties) into a first machine learning model. Examples of such biochemical properties include bulkiness, mole fraction of buried residues, mean flexibility index, normalized frequency for alpha helices, known codon number encoding each amino acid of the universal genetic code, polarity, conformational parameters for beta turns, normalized frequency for beta sheets, side chain class, side chain polarity, retention coefficient, hydrophobicity constant, normalized frequency of beta turns, PK1 α-COOH, membrane-embedded helix parameters, and antigenicity. Different sets of biochemical properties may be used to predict CBCE and LBCE. One or more biochemical properties may be calculated or measured using techniques known in the art (e.g., from the amino acid sequence of a protein). Biochemical properties may be determined for each amino acid of the protein.

[0026] Typically, each amino acid (or side chain) constituting a protein sequence has physiological and chemical properties that contribute in some way to its SS and 3D properties. Thus, embodiments that include the use of one or more physiological and / or surface properties in addition to the physiological and chemical properties can further enhance the performance of the first machine learning model.

[0027] In general, each of the structural and / or surface properties, as well as the physiological and chemical properties (if used), may contribute separately or in combination to the training and subsequent use of a first machine learning model for predicting whether a protein contains a B cell epitope.

[0028] Typically, the first machine learning model is (i) Multiple amino acid sequences of each first reference protein, each first reference protein comprising multiple amino acid sequences containing at least one true B cell epitope, (ii) Multiple reference codes for each true B cell epitope, each reference code representing each true B cell epitope as multiple data elements corresponding to amino acids in an amino acid sequence, (iii) one or more structural and / or surface properties of each first reference protein and It is trained using the first reference dataset, which includes [the specified data].

[0029] In this way, the first machine learning model can be trained to learn the relationships between reference codes, the structure and / or surface properties of the first reference protein, and the corresponding true B cell epitopes (e.g., those classified as true B cell epitopes on the first reference protein). In some embodiments, the first reference dataset may further include one or more physiological and chemical properties of the first reference protein and / or the amino acids that constitute them.

[0030] The reference codes for each true B cell epitope used in the first reference dataset are typically generated in the same manner as described above.

[0031] In a particularly preferred embodiment, one or more structural and / or surface properties of each first reference protein are one or more structural and / or surface properties of each first reference protein in its unbound state. The three-dimensional structure of the BCE can change during binding between the B cell epitope and the antibody. Therefore, a key objective of the present invention is to predict the true B cell epitope in its unbound state (e.g., the unbound state of the protein antigen). In other words, it is the state of the epitope that is unaffected by other proteins, structures, or interacting molecules (e.g., before the binding event occurs). Importantly, this unbound state is not the state in which Ab has already "recognized" and bound, but the state in which Ab "recognizes" and initiates the binding event. This aspect of the present invention offers a significant advantage over state-of-the-art models that can be inaccurate, as the prediction is based on structural data obtained after the binding event has occurred and trained on the Ag-Ab complex (i.e., the bound protein). For example, the 3D structure of an epitope in a bound Ag-Ab complex may differ from the 3D structure of its unbound ("native") epitope.

[0032] Thus, the first reference dataset used to train the first machine learning model is preferably generated to predict the true B cell epitopes before the binding event occurs. As will be discussed in more detail herein, in some preferred embodiments, the structure and / or surface properties of each first reference protein may be accessed by predicting the structure and / or surface properties using one or more second trained machine learning models trained on the unbound reference protein.

[0033] Referring again to step (b) of the method of the present invention, the structure and / or surface properties of the protein (e.g., in its unbound state) can be obtained using known techniques. For example, the features of a protein can be predicted or measured using known algorithms (including computational approaches for predicting a complete 3D protein structure) or X-ray crystallography techniques. For example, RSA, HSE, and secondary structure (and other) parameters can be accessed from the complete 3D structure of the protein by prediction or experiment, and the complete structure itself is not required as input to the model.

[0034] However, in a particularly preferred embodiment, one or more structural and / or surface properties accessed in step (b) are predicted by applying one or more trained second machine learning models to the amino acid sequences of the proteins accessed in step (a), and the one or more trained second machine learning models are trained on a second reference dataset containing multiple amino acid sequences of each protein and their corresponding structural and / or surface properties. Preferably, one or more structural and / or surface properties of the proteins accessed (e.g., predicted) in step (b) are one or more structural and / or surface properties of the unbound proteins.

[0035] In this way, it may be advantageous to calculate the structure and / or surface properties of the protein under investigation directly from the amino acid sequence without requiring any additional information supplied or input by the user. Different learning models may be used to predict continuous and categorical features. Typically, the output of the trained second machine learning model is a value and / or class assigned to each amino acid in the protein amino acid sequence that represents the structure and / or surface properties of the protein.

[0036] In a preferred embodiment, one or more structural and / or surface properties of each first reference protein are predicted by applying one or more trained second machine learning models to the corresponding amino acid sequences, and the one or more trained second machine learning models are trained on a second reference dataset containing multiple amino acid sequences of each protein and their corresponding structural and / or surface properties.

[0037] Therefore, the same second machine learning model can be used to predict the structure and / or surface properties of the query protein, as well as the structure and / or surface properties of the reference protein used to train the first machine learning algorithm. In some embodiments, different second machine learning models can be used to predict different properties. For example, in one embodiment, a trained second machine learning model can be used to predict the secondary structure of a protein, and different trained second machine learning models can be used to predict RSA and HSE properties.

[0038] In a particularly preferred embodiment, one or more second machine learning models are trained on a second reference dataset containing multiple amino acid sequences of each second reference protein, as well as their corresponding (e.g., three-dimensional) structures and / or surface properties in an unbound state. This allows the structural and / or surface property predictions made by the second machine learning model to be predictions of these features in the unbound state of the protein (e.g., before binding to Ab). This offers particular advantages to the present invention, as outlined above, especially when generating the first reference dataset used to train the first machine learning model.

[0039] Typically, in addition to the amino acid sequence, the input to one or more second machine learning models includes several properties (e.g., features) of the first reference protein. The input properties are typically physiological and chemical properties, but in some embodiments, they may include structural and / or surface properties. These features are typically calculated or measured amino acid-wise using techniques known in the art. Examples of such (e.g., physiological and chemical) properties include one or more of hydrophobicity, side-chain class, side-chain polarity (e.g., for predicting HSE and RSA), and one or more of the coil's conformational parameters and the alpha-helix's conformational parameters (e.g., for predicting secondary structure). Further features may be used. Thus, one or more second machine learning models can be trained on the second reference dataset to learn the relationship between one or more (e.g., physiological and chemical) properties of the amino acid sequence of each second reference protein and the corresponding structural and / or surface properties (e.g., one or more of RSA, HSE, and secondary structure) in the unbound state. Typically, the second machine learning algorithm utilizes a different set of protein properties than the first machine learning algorithm.

[0040] The second machine learning model can be (or may include) any neural network, such as a convolutional neural network or a feedforward neural network. Typically, one or more trained second machine learning models are (or include) recurrent neural networks, preferably long-short-term memory networks (LSTMs), and more preferably bidirectional long-short-term memory networks (BLSTMs).

[0041] One problem with conventional approaches to predicting B cell epitopes, particularly CBCEs, is the limited amount of data available to define true CBCEs. Therefore, embodiments of the present invention provide a method for obtaining a first reference protein, each containing at least one true B cell epitope.

[0042] Therefore, in some embodiments, at least some of the first reference proteins are (i) Accessing multiple (preferably non-absolute) protein complexes containing at least three different protein chains; (ii) Filter multiple protein complexes to determine the effective weight of the antibody (V H ) Protein chain mapped as a chain, effective light (V) of the antibody L ) Only those containing protein chains mapped as chains, and protein chains mapped as antigens, should be retained; (iii)V H Chain and V L Pairing the chains, V H and V L To form a chain pair; (iv) Each V H and V L For each chain pair, define the true B cell epitope on the corresponding antigen; It can be obtained by, however, At least some of the first reference proteins in the first reference dataset correspond to the antigen mapped in step (ii) that includes at least one true B cell epitope defined in step (iv).

[0043] The first reference protein obtained or identified by the method described above is typically mapped onto its corresponding amino acid sequence. In this way, the structure and / or surface properties of the reference protein in its unbound state can be obtained, for example, using one or more trained second machine learning models discussed above.

[0044] Typically, the protein complex accessed in step (i) could belong to any organism or entity that has been assigned a classification ID by an established institution such as NCBI. This advantageously allows the predictions made by the first trained machine learning model to be non-ab specific.

[0045] In an embodiment, each true B cell epitope is defined as the amino acid of the corresponding antigen that contacts the chain pair of V H and V L and preferably, the amino acid contains any atom within a predetermined distance of the complementarity determining region CDR of the chain pair of V H and V L and is defined as contacting the corresponding VH and the chain pair of V L and preferably, the predetermined distance is 4 angstroms or less.

[0046] As discussed above, the present invention advantageously generates the code of one or more candidate B cell epitopes to be tested in order to model and "test" each candidate epitope as a whole. Due to this approach, the first reference dataset used to train the first machine learning model may contain only observed true positive BCEs (i.e., only "positive" training data). Thus, in some preferred embodiments, the method may further include generating pseudo-random data based on the plurality of reference codes of each true B cell epitope of the first reference dataset, and the first reference dataset includes the pseudo-random data as negative data. In this way, the performance of the first machine learning model can be improved because the model can be trained even when there is no validated negative data.

[0047] The pseudo-random data is based on the encoded true B cell epitope. Preferably, the pseudo-random data includes a plurality of permutations of a plurality of data elements representing the true B cell epitope. In some embodiments, the pseudo-random data can be a random permutation of the data elements. In some embodiments, the pseudo-random data can share at least one characteristic with the code of the true B cell epitope. The characteristic can be a physical characteristic of the true B cell epitope, for example, the position on the amino acid sequence of one or more amino acids (e.g., the first and last amino acids of the epitope). In another example, the characteristic can be the linear distance between specific amino acids.

[0048] In some embodiments, this method To identify a first B cell epitope that is likely to be a true B cell epitope on a protein; The process involves substituting one or more amino acids that form the identified first B cell epitope with different amino acids, thereby generating a modified amino acid sequence for the protein; Using a first trained machine learning model, predict whether a candidate code representing the identified first B cell epitope is likely to correspond to a true B cell epitope on the modified amino acid sequence of the protein. It may further include the following.

[0049] In this way, by applying one or more amino acid substitutions to the identified B cell epitopes, the first machine learning model can be used to predict B cell epitopes that do not naturally exist on the protein under analysis. This may be advantageous for identifying potential future viral variants and vaccine candidates that offer protection across a wide range of viral or pathogen species, as described in further detail herein.

[0050] In the substitution application step, one or more amino acids in the first B cell epitope (typically a wild-type B cell epitope) are substituted with any other amino acids present. This can be called an amino acid "mutation." Using a trained first machine learning model, the likelihood of each possible amino acid substitution in the identified first B cell epitope forming a true B cell epitope on the protein can be predicted. Further analysis can be performed on B cell epitopes predicted to be positive for the modified amino acid sequence. For example, structural stability analysis can be performed using techniques known in the art (e.g., computational).

[0051] Typically, if the candidate code is in the form of a binary vector, only the amino acids corresponding to the "1" in the binary vector (e.g., representing direct amino acid contact points on the 3D protein sequence with the antibody) are substituted ("mutated").

[0052] In the step of using a trained first machine learning model to predict whether a candidate code representing an identified first B cell epitope is likely to correspond to a true B cell epitope on a modified amino acid sequence of a protein, the input to the model includes the modified amino acid sequence, one or more structural and / or surface properties of the modified protein, and a candidate code representing the identified first B cell epitope. The structural and / or surface properties can be obtained, for example, by applying a second machine learning model to the modified amino acid sequence using the techniques described above.

[0053] A further aspect of the present invention is a method for training a machine learning model, This involves generating a reference dataset, and the reference dataset is, Multiple amino acid sequences of each reference protein, wherein each reference protein contains at least one true B cell epitope; Multiple reference codes for true B cell epitopes, where each reference code represents a true B cell epitope as multiple data elements corresponding to amino acids in an amino acid sequence; One or more structural and / or surface properties of each reference protein, Including generating; The machine learning model is trained using a reference dataset to learn the relationships between candidate codes, structures and / or surface properties, and their corresponding true B cell epitopes (e.g., those classified as true B cell epitopes on proteins). A method for training machine learning models is provided, including [specific examples of this method].

[0054] Therefore, the method for training a machine learning model offers the advantages described above and can be used in combination with any of the examples described in the present invention. Advantageously, once trained, the machine learning model can be used to predict whether a query protein contains a B cell epitope that is likely to trigger an antibody binding event.

[0055] Typically, one or more structural and / or surface properties of each reference protein are one or more structural and / or surface properties of each first reference protein in its unbound state.

[0056] As described above, the reference dataset may further include one or more physiological and chemical properties of each protein and / or the amino acids that make up the reference protein.

[0057] The machine learning model may be (or include) any neural network, such as a convolutional neural network or a feedforward neural network. Preferably, the machine learning model is (or includes) a recurrent neural network. Typically, the first machine learning model is (or includes) a long-short-term memory network (LSTM), preferably a bidirectional long-short-term memory network (BLSTM).

[0058] The present invention also provides the use of a machine learning model trained using the methods described above (for example, to predict the presence of BCE on a query protein). The present invention also provides a machine learning model trained using the methods described above.

[0059] The present invention may further include synthesizing one or more proteins predicted to contain a B cell epitope and / or a predicted B cell epitope.

[0060] The proteins predicted by this invention to contain B cell epitopes may be used to treat or prevent infectious diseases or other diseases, such as autoimmune or immune-related diseases and cancer. This invention is used in one or more of the following applications: • Therapeutic antibody mapping • Vaccine development and production (both for prevention and treatment) • Immunodiagnosis and immunomonitoring.

[0061] In some embodiments, the method may further include encoding one or more proteins predicted to contain a B cell epitope, and / or the predicted B cell epitope, and / or B cell epitope variants, and / or predicted or simulated variants of the B cell epitope, into corresponding protein, peptide, DNA, or RNA sequences. The DNA or RNA sequences may be incorporated into a delivery system for use in a vaccine (e.g., using naked or encapsulated DNA or encapsulated RNA). The method may also include incorporating the protein, peptide, DNA, or RNA sequences into the genome of a bacterial or viral delivery system to produce a vaccine.

[0062] Therefore, according to a further aspect of the present invention, a method for producing a vaccine, the following: To predict that a protein contains a B cell epitope, follow any of the examples discussed above; Synthesizing a protein and / or a predicted B cell epitope, or encoding at least one of a protein, a predicted B cell epitope, a B cell epitope variant, or a predicted or simulated variant of a B cell epitope in a corresponding protein, peptide, DNA, or RNA sequence. A method including this is provided.

[0063] Such proteins, peptides, DNA, or RNA sequences can be delivered in a naked or encapsulated form, or incorporated into the genome or cells of a bacterial or viral delivery system to produce a vaccine. In addition, DNA can be delivered to vaccinated host cells using bacterial vectors. In the case of peptide vaccines, the identified proteins can typically be synthesized as amino acid sequences or "chains."

[0064] Such vaccines may be preventive or therapeutic vaccines. For example, this method can be used to produce personalized vaccines for individuals, such as cancer treatment vaccines if the protein contains a newly synthesized antigen.

[0065] A further aspect of the present invention provides a method for creating (e.g., and / or designing) a diagnostic assay for determining whether a patient has or has had a past infection with cancer or a pathogen, the diagnostic assay being performed on a biological sample obtained from a subject and comprising identifying at least one protein of a pathogen or tumor that is predicted to contain a B cell epitope using a method according to any of the examples discussed above, the diagnostic assay comprising utilization or identification of the identified at least one protein and / or B cell epitope in the biological sample.

[0066] Thus, the present invention can be advantageously used to create diagnostic tests or assays by rapid and / or automated molecular target discovery. Query proteins identified as likely to contain true B-cell epitopes can be further analyzed in laboratory tests to create such diagnostic tests or assays, thereby significantly reducing the time required for test development compared to conventional laboratory methods.

[0067] The term “utilization,” as used herein, is intended to mean that at least one protein and / or a B-cell epitope on it is used in an assay for identifying an immune response (e.g., protective) in a patient. In this context, the identified protein and / or epitope within is a component of the assay, not a target of the assay.

[0068] An in vitro diagnostic assay may involve the identification of an immune system component in a biological sample that recognizes at least one epitope within an identified protein. Thus, the diagnostic assay may utilize at least one identified protein and / or at least one predicted epitope. Typically, the diagnostic assay contains at least one identified protein, protein subunit, and / or predicted epitope (e.g., a synthesized one). In a preferred embodiment, the immune system component may be a B cell. For example, the assay may involve the identification of an antibody or B cell that recognizes a predicted B cell epitope within the identified protein.

[0069] As an example of such a diagnostic application, a sample isolated from a patient, preferably a blood sample, may be analyzed for the presence of B cells or antibodies in the biological sample that recognize and bind to epitopes in identified proteins, which are identified as part of the present invention and included in the assay.

[0070] Suitable diagnostic assays, as will be recognized by those skilled in the art, may include enzyme-coupled immunosorbent spot (ELISPOT) assays, enzyme-coupled immunosorbent assays (ELISA), cytokine capture assays, intracellular staining assays, tetramer staining assays, microfluidic devices, lab-on-a-chips, microarrays, flow cytometry, CyTOF, proteomics, molecular sequencing, or limiting dilution culture assays.

[0071] In a method for creating a diagnostic test, the amino acid sequences of one or more proteins to be tested for the presence of BCE may be selected based on the desired response to be tested. For example, one or more source proteins may be one or more source proteins of any pathogen and / or virus (or fragment thereof), such as the SARS-CoV-2 virus. In such cases, the present invention may be used to create a diagnostic test for determining whether a patient has or has had a past infection with the SARS-CoV-2 virus and / or its variants and / or related viral species. However, as will be recognized by those skilled in the art, one or more source proteins may originate from any pathogen (e.g., any virus, parasite, bacterium, or cancer indication).

[0072] Further disclosed herein are diagnostic assays for determining whether a patient has or has had a history of infection and / or cancer caused by a pathogen, wherein the diagnostic assay is performed on a biological sample obtained from a subject, and the diagnostic assay comprises the use or identification in the biological sample of at least one protein of a pathogen or tumor predicted to contain a B-cell epitope, identified using any of the methods discussed above. The diagnostic assay may also comprise the identification of an immune system component (e.g., a B cell) in the biological sample that recognizes at least one identified protein and / or at least one predicted epitope.

[0073] A further aspect of the present invention provides a computer program product that, when executed by a computer, includes instructions causing the computer to perform any of the methods described above. The present invention also provides a non-temporary computer-readable medium that, when executed by a computer, includes executable instructions causing the computer to perform any of the methods described above.

[0074] A further aspect of the present invention provides a system for predicting whether a protein contains a B cell epitope that is likely to trigger an antibody binding event, the system comprising at least one processor communicating with at least one memory device, the at least one memory device storing instructions for causing the at least one processor to perform a method according to any of the examples of the present invention considered above.

[0075] A further aspect of the present invention provides a method for synthesizing a protein, comprising predicting that the protein contains a B cell epitope that is likely to induce an antibody binding event by any of the examples discussed above, and synthesizing the predicted protein. The synthesis of the predicted protein can be carried out using techniques known in the art. A further aspect of the present invention provides a protein synthesized using the method.

[0076] Embodiments of the present invention may be advantageously used to identify one or more B cell epitopes that are predicted to induce a protective immunogenic response across multiple species of pathogens or viruses. Such “broad-spectrum protective” B cell epitopes may be used, for example, as vaccine elements or to provide diagnostic tests or assays.

[0077] Therefore, further disclosed herein is a method (e.g., computer implementation) for identifying one or more B cell epitopes that are predicted to induce a protective immunogenic response across multiple species of pathogens or viruses (e.g., and / or variants thereof), (i) For each of several different proteins of a pathogen or virus, perform a method according to any of the examples described above in relation to a first aspect of the present invention to identify several first B cell epitopes, each of which is predicted to be likely to be a true B cell epitope on at least one of the several proteins; (ii) For each identified first B cell epitope, determine the number of different species of pathogens or viruses for which the first B cell epitope is predicted to be a true B cell epitope; (iii) Classifying one or more of the first B cell epitopes as broad-spectrum protective B cell epitopes based on the number of different species or the number of variants of any one given species in which the first B cell epitope is predicted to be a true B cell epitope. This method includes [something].

[0078] Thus, the present invention can be used to classify one or more B cell epitopes as those expected to induce a broad immunogenic response (e.g., cross-reactivity) across multiple different species or variants of pathogens or viruses. The classification is based on determining the number of different species of pathogens or viruses on which the B cell epitope is expected to be a true B cell epitope (e.g., a viable B cell epitope). For example, a first B cell epitope expected to be a viable B cell epitope on the maximum number of different species or variants tested may be classified as a broad-protective B cell epitope (e.g., may be classified as a vaccine candidate). In another example, if the number of species on which the B cell epitope is expected to be a true B cell epitope is greater than a predetermined threshold, the identified first B cell epitope may be classified as broad-protective.

[0079] Preferably, to reduce the computational load, one or more candidate codes may be mapped only to the receptor-binding domain (RBD) of the amino acid sequence of each protein being tested.

[0080] In some embodiments, the method applies to at least one (preferably each) of the first B cell epitopes classified as broad-area protective B cell epitopes. The method involves substituting one or more amino acids that form the first B cell epitope with different amino acids, thereby generating a modified first B cell epitope; For each of the multiple proteins of each of the multiple species of viral pathogens, To generate a modified amino acid sequence of a protein according to a modified first B cell epitope; Using a first trained machine learning model, predict whether a candidate code representing a first B cell epitope, classified as a broad-spectrum protective B cell epitope, is likely to correspond to a true B cell epitope on the modified amino acid sequence of the protein. It may further include the following.

[0081] Thus, the method of the present invention can be advantageously used to predict viable B cell epitopes that may not naturally exist on proteins under analysis. This can be advantageously used to identify broad-protective vaccine candidates that provide broad-protective responses across multiple different species (and / or variations thereof) of pathogens or viruses and may also be protective against future viral variants or pathogens that have not yet spread pathogenically in humans. Such identified B cell epitopes can form vaccine candidates or be used to create diagnostic tests or assays as outlined above.

[0082] The inventors have found that some of the modified B cell epitopes identified in this manner are predicted to generate broader immunogenic responses across a greater number of different virus or pathogen species compared to the most cross-reactive wild-type B cell epitopes.

[0083] Further analysis may be performed on identified modified B cell epitopes that do not exist naturally in viruses or pathogens to assess their structural stability. For example, any "mutant" B cell epitopes predicted to have low structural stability may be discarded from future analysis. Such structural analysis may be performed using techniques known in the art (e.g., computational).

[0084] Brief explanation of the drawing Here, embodiments of the present invention will be described with reference to the attached figures. [Brief explanation of the drawing]

[0085] [Figure 1] This is a flowchart illustrating the steps of a method according to one embodiment of the present invention. [Figure 2] This flowchart illustrates the steps of a method for training a first machine learning algorithm according to one embodiment of the present invention. [Figure 3] This flowchart illustrates the steps of a method for training a second machine learning algorithm according to one embodiment of the present invention. [Figure 4] An example of a system suitable for realizing an embodiment of this method is shown. [Figure 5] This is an example of a suitable server. [Figures 6a-6f] The correlation results between feature pairs are shown. [Figure 7] (a) Spearman correlation coefficients for continuous features of RSA, (b) UHSE, and (c) LHSE are shown. [Figure 8] (a) Kendall correlation coefficients for categorical features of RSA, (b) UHSE, and (c) LHSE are shown. [Figure 9] (a) Kendall correlation coefficients for continuous features of SS, (b) CBCE, and (c) LBCE are shown. [Figure 10] (a) Mutual information criteria for categorical features of SS, (b) CBCE, and (c) LBCE are shown. [Figure 11] The structure of a predictive model for RSA and HSE according to one embodiment of the present invention is schematically shown. [Figure 12a-f] The cross-validation results for an RSA / HSE model according to one embodiment of the present invention are shown. [Figure 13a-f] The test results of an RSA / HSE model according to one embodiment of the present invention are shown. [Figure 14a-f]The final training results of an RSA / HSE model according to one embodiment of the present invention are shown. [Figure 15a-f] The cross-validation (CV) results for the secondary structure (SS) prediction model according to the present invention are illustrated. [Figure 16a-b] The test results for the SS model are shown below. [Figure 17a-d] The final training results for an SS model according to one embodiment of the present invention are shown. [Figure 18a-b] A paratope analysis of CBCE is presented. [Figure 19a] A schematic example of the architecture of a first machine learning algorithm for predicting CBCE according to one embodiment of the present invention is provided. [Figure 19b] A schematic representation of the architecture of a first machine learning algorithm for predicting CBCE according to a further embodiment of the present invention is shown. [Figure 20a-d] This section presents an outlier analysis of CBCE. [Figure 21a-f] Further embodiments of the present invention show the results of the precision-recall metric for a first machine learning algorithm for predicting CBCE. [Figure 22a-f] The results of the precision-recall metric for a first machine learning algorithm for predicting CBCE according to one embodiment of the present invention are shown. [Figure 23a-b] The results of the precision-recall metric for a first machine learning algorithm for predicting CBCE according to one embodiment of the present invention are shown. [Figure 24a-b] The results of the precision-recall metric for a first machine learning algorithm for predicting CBCE according to one embodiment of the present invention are shown. [Figure 25a-c] The final training results for a first machine learning algorithm for predicting CBCE according to one embodiment of the present invention are shown. [Figure 26a-b] The final training results for a first machine learning algorithm for predicting CBCE according to one embodiment of the present invention are shown. [Figure 27a-d] This example illustrates outlier analysis using LBCE. [Figure 28] A pipeline according to one embodiment of the present invention is schematically illustrated. [Figure 29a-c] This example illustrates the successful mapping of receptor-binding domains (RBDs) to betacoronavirus species. [Figure 30] This example illustrates the prediction of a structural epitope that is predicted to be broadly immunogenic across multiple virus species. [Figures 31a-31b] This section illustrates the changes in predicted immunogenicity after the implementation of a mutation policy. [Modes for carrying out the invention]

[0086] Detailed explanation 1. Overview Embodiments of the present invention provide a B-cell epitope (BCE) prediction algorithm ("First Machine Learning Algorithm" or "First Machine Learning Model") trained using BLSTM. The model accurately predicts BCEs on a given Ag from its primary sequence in combination with one or more of its 3D structures and / or surface properties, without any requirements regarding the complete 3D structure of Ag as input to the predictor. The model is trained using BLSTM, which allows processing of the entire Ag sequence as a single data point processed from both directions without requiring k-mer segmentation. Each amino acid of Ag is treated as a time step in the network, and its properties can be influenced by both preceding and succeeding amino acids flanking to the amino acid in question, thus allowing for the discovery of sequential relationships between spatially separated amino acids that may be an important part of conformational BCEs (CBCEs).

[0087] The first machine learning model has the ability to discover distinct or independent CBCEs that may exist on a single Ag protein sequence. A particularly advantageous feature of this model is that it takes an Ag protein sequence, a binary "permutation vector" or "code" representing each single candidate CBCE (or LCBE), and one or more 3D structures and / or surface properties of Ag as input. The model's output is a single probability indicating the likelihood that the input permutation vector is a bone fide CBCE. In other words, the user asks the model whether a particular query CBCE, coded as a binary permutation vector on a given Ag sequence, is a true CBCE. However, the model may also be used to provide a second output in the form of a probability vector having a probability for each amino acid in the Ag sequence. The second output can be considered the contribution of each amino acid to the CBCE in question and can be used to compare the current approach with conventional models that predict "amino acid by amino acid". The first machine learning algorithm is preferably trained on the unbound 3D Ag protein structure before the Ag-Ab binding event to learn the properties of BCEs on Ag before the Ag binding event in order to capture bone fide ("true") BCEs on the query Ag protein sequence.

[0088] The structural properties of the Ag sequence are important for predicting CBCE. However, embodiments of the present invention negate the need for the complete 3D structure of the protein sequence, which is experimentally measured or predicted and then used as input to the algorithm. The first machine learning algorithm does not require the coordinates of each atom of each amino acid in the 3D protein sequence (i.e., does not require the complete 3D structure), but instead can predict BCE using one or more structural and / or surface properties of the query protein, such as RSA, HSE, SS.

[0089] The most important features of "top-level" proteins such as RSA, HSE, and SS can typically be predicted using one or more trained machine learning algorithms ("second machine learning algorithms" or "second machine learning models"), taking the protein's amino acid sequence as input. This way, advantageously, the structure and / or surface properties of the protein under investigation can be calculated directly from the amino acid sequence without requiring any additional information supplied or input by the user. These second algorithms also utilize BLSTM networks and have been shown to have high performance.

[0090] In this way, the first and second machine learning models can work together to form a "pipeline" that requires only the amino acid sequence of the query protein as input to predict whether the query protein contains a true B cell epitope. A schematic diagram of such a pipeline constituting the first and second machine learning models according to one embodiment of the present invention is shown in Figure 28.

[0091] Figure 1 is a flowchart 100 outlining the steps of a preferred embodiment of the present invention. In step S101, the amino acid sequence of the protein to be analyzed for the presence of a B cell epitope is accessed. Typically, the protein under investigation includes an antigen. The amino acid sequence can be accessed using various techniques known to those skilled in the art, and for example, the sequence can be downloaded from a bioinformatics repository such as UniProt.

[0092] In step S103, one or more (e.g., three-dimensional) structural and / or surface properties of the unbound protein are accessed. Typically, these properties of the protein are predicted using one or more trained second machine learning models (schematically illustrated in 300 and described below) that take the amino acid sequence accessed in step S101 as input. The structural and / or surface properties typically include the secondary structure and relative solvent accessibility of the query protein. Hemispheric exposure (both upper and lower hemispheric exposure) may also be used. A value or category for each of these properties is typically assigned to each amino acid in the sequence. However, other techniques may be used to access the three-dimensional structure and surface properties of the query protein, such as database entries, X-ray crystallography, and computational approaches to predict the complete 3D protein folding structure (from which the structural and / or surface properties can be obtained).

[0093] In step S105, candidate codes for one or more candidate BCEs on the query protein are generated. Each candidate code is typically in the form of a binary vector, where each data element in the binary vector corresponds to an amino acid in the amino acid sequence that forms each candidate BCE. In such a binary “permutation vector” code, each amino acid in the sequence that forms part of the candidate BCE is assigned a “1”, and each amino acid in the sequence that does not form part of the candidate BCE is assigned a “0”. For example, in the case of the exemplary amino acid sequence “ABCDEFG”, a candidate conformational B cell epitope consisting of amino acids B, D, F, and G may be coded as the binary vector (0101011). In some embodiments, a candidate code may be generated for each possible candidate BCE on the protein. However, to reduce the computational load, the search space may be limited based on physical prior knowledge (e.g., knowledge of mutation sites on the nascent antigen of the input protein or the receptor-binding domain on the SARS-CoV-2 spike protein) or techniques such as deep reinforcement learning models or association-generative models.

[0094] In step S107, the amino acid sequence accessed in step 101, one or more structural and / or surface properties accessed in step 103, and the candidate codes generated in step S105 are input to a trained first machine learning model being trained using BLSTM. One or more biochemical properties assigned to each amino acid in the amino acid sequence may also be used as input to the trained first machine learning model. The output of the trained first machine learning model is the probability that each candidate code represents a true B cell epitope on the query protein.

[0095] Figure 2 is a flowchart 200 outlining the steps of a preferred embodiment for training a first machine learning model. In step S201, multiple protein complexes are accessed. These can be accessed from publicly available sources, for example, the Protein Data Bank (PDB) for CBCE or the Immune Epitope Database (IEDB) for LBCE.

[0096] In step S203, the protein complex accessed in step S201 is analyzed to define the true epitope and antigen. A linear BCE may be defined in the IEDB complex. Further analysis may be required to define the true conformational epitope from the protein complex accessed from the PDB, and further details of these techniques are described herein.

[0097] In step S205, each antigen defined in step S203 as having a true B cell epitope is mapped onto its amino acid sequence. In step S207, each true epitope is encoded onto the amino acid sequence of the corresponding antigen. In this way, multiple reference epitope codes corresponding to true epitopes are generated. The reference codes are typically encoded as binary vectors in the same manner as described above in step S105.

[0098] In step S209 of this method, the three-dimensional structure and / or surface properties of the unbound antigen are accessed, for example, by prediction from the amino acid sequence of the antigen obtained in step S205. Typically, the structure and / or surface properties of the antigen are predicted using a trained second machine learning model (schematically illustrated in 300). However, other techniques may be used to access the structure and / or surface properties of the unbound antigen, such as using known computational approaches for predicting 3D protein structures. Typically, these properties include secondary structure, hemispheric exposure (both UHSE and LHSE), and relative solvent accessibility.

[0099] In step S211, a first reference dataset is generated. The first reference dataset includes the amino acid sequence of the reference antigen (step S205), the encoding of the true epitope on the corresponding reference antigen (step S207), and the accessed (e.g., predicted) structure and / or surface properties of the unbound reference antigen. In some embodiments, the first reference dataset may further include one or more physiological and chemical properties assigned to each amino acid of the reference antigen.

[0100] In step S213, the first machine learning model is trained using the first reference dataset generated in step S211. Here, the first machine learning model is or includes BLSTM and may be trained using techniques known in the art. By training the model using the first reference dataset generated by performing steps S201 to S211, the first machine learning model is trained to learn the reference codes, structures and / or surface properties of true epitopes and the relationships between them and the corresponding true B cell epitopes.

[0101] As discussed above with reference to Figures 1 and 2, in a preferred embodiment, a second machine learning model (schematically shown in 300) is used to predict the structure and / or surface properties both when training the first machine learning model (Figure 2) and when using the trained first machine learning model to predict the presence of B cell epitopes on the protein of interest (Figure 1). A particularly advantageous feature of using the second machine learning model is that it is trained to predict the structure and / or surface properties of the protein in its unbound state. Thus, the first machine learning model is configured to predict the BCE in its unbound state, i.e., to predict a BCE having a three-dimensional form and properties that Ab "recognizes" and initiates binding.

[0102] Figure 3 is a flowchart summarizing the steps of a preferred embodiment for training a second machine learning model.

[0103] In step S301, multiple second reference proteins (i.e., single protein structures) in an unbound state are accessed. These second reference proteins can be accessed from publicly available databases such as the PDB.

[0104] In step S303, the amino acid sequence of the second reference protein is accessed.

[0105] In step S305, the three-dimensional structure and / or surface properties of the second reference protein are accessed. Typically, this includes secondary structure, relative solvent accessibility, and upper and lower hemisphere exposure. The secondary structure and RSA properties are accessed using the DSSP algorithm (Wolfgang Kabsch and Christian Sander. Dictionary of protein secondary structure: Pattern recognition of hydrogen-bonded and geometrical features. Biopolymers, 22(12):2577-2637, 2018 / 11 / 30 December 1983), and the UHSE and LHSE are calculated using the BioPython package (Peter JA Cock, Tiago Antao, Jeffrey T Chang, Brad A Chapman, Cymon J Cox, Andrew Dalke, Iddo Friedberg, Thomas Hamelryck, Frank Kauff, Bartek Wilczynski, et al. Biopython: freely available python tools for computational molecular biology and bioinformatics. Bioinformatics, 25(11):1422-1423, 2009). DSSP and BioPython algorithms can be applied to a second reference protein structure file (e.g., obtained from the Protein Data Bank, PDB).

[0106] In step S307, a second reference dataset is generated using the amino acid sequence accessed in step S303 and the structure and / or surface properties of the second reference protein accessed in step S305. In some embodiments, the second reference dataset may further include one or more biochemical properties assigned to each amino acid of the second reference protein.

[0107] In step S309, a second machine learning model is trained using a second reference dataset (for example, to learn the relationship between one or more features of the amino acid sequence of a second reference protein and the corresponding structure and / or surface properties). Here, the second machine learning model is or includes BLSTM and may be trained using techniques known in the art. It will be recognized that different second machine learning models may be trained to predict different respective structures and / or surface properties. For example, one second machine learning model may be trained to predict RSA and HSE metrics, and a further second machine learning model may be trained to predict secondary structure.

[0108] The various steps outlined in Figures 1-3 are described in more detail in the following sections. Section 2 describes the pre-training techniques used in all models. Section 3 presents some preliminary analyses performed to select the input features of the models. Section 4 presents the RSA, HSE, and SS structure algorithms used in the current BCE algorithm. Section 5 presents the RSA, HSE, and SS structure algorithms used as features in the BCE model. Section 6 presents an example of the first machine learning algorithm according to the present invention used to predict CBCE, while Section 7 presents an overview of the first model used to predict linear BCE (LBCE). Section 8 presents the applications of the present invention.

[0109] 2. Pre-training techniques An autoencoder (aE) was used to ensure reproducible results and to avoid the need for random seeds in the network used in the model. An aE is a type of neural network trained on compressed data that reflects the input of the main model. More specifically, the AE receives the input at layer number i=0 and processes it through any number of layers that make up the encoder section, e.g., i=1, ..., N. It then processes it through a mirror structure of the encoder section called the decoder section. Finally, it returns an output identical to the input. Hidden layers can reduce or expand the dimensionality of the input.

[0110] Each layer of the main network is converted to an AE and trained with some compressed data. The pre-trained weights of the outer layers are used as constants for subsequent aEs. This is repeated until all main network layers are pre-trained. The pre-trained weights can then be used as the starting point for the main training procedure.

[0111] The following procedure was performed once, and the pre-trained weights were used in all cross-validation splits, testing of unknown data, and the final model.

[0112] First, 10,000 protein sequences, ranging in length from 50 to 1,000 amino acids, were randomly generated. The actual amino acids used were also randomly selected. Next, each layer of the neural network was pre-trained as an individual AE (Enhanced Aspect) on the entire generated data. Pre-training was performed for a maximum of 1,000 epochs for each AE, with the possibility of stopping early if the loss did not decrease after 10 epochs. The weights of the outer layers were copied to deeper AEs after pre-training and were kept constant during their pre-training as well. The final layer of the model (output) was not pre-trained at all.

[0113] This procedure was performed once, and the same pre-training weights were used for all validation steps. This procedure was faster, used significantly more data for pre-training, and used the same initial weights for all downstream analyses.

[0114] 3. Preliminary Analysis and Feature Selection Techniques A preliminary analysis was conducted to investigate how various features correlate with and with the target we want to predict. This analysis had two objectives. First, it aimed to identify highly relevant feature-target correlations that would help determine which features to use to predict the target. Second, feature-feature correlations help to subset the feature-target correlation space by preserving features that are not correlated with each other, thus reducing the input space for the model. However, it should be noted that everything analyzed using this selection method captures, in most cases, linear relationships between features and targets and between features.

[0115] 3.1. Method 1 Correlation analysis was performed using the features in Table 0 to determine which features to use as input for predicting different targets of the models. The analysis was then performed using the entire dataset for these models; that is, the datasets used for these analyses were those described in Section 4.1. This first analysis method was run for the 3D structure prediction model but not for the CBCE or LBCE models.

[0116] 3.1.1 Preprocessing of Continuous Features Continuous features provide a single value for each amino acid in the alphabet. To model the effect of neighboring amino acids in the protein sequence, the values ​​of each feature were averaged by scanning the protein sequence using a sliding window of 9 amino acid peptide lengths. The average feature value was assigned to the middle amino acid in the peptide. This was done for all features on each protein sequence across the entire training and test datasets.

[0117] 3.1.2 Correlation between continuous features and continuous targets These analyses were performed for RSA (relative solvent accessibility), UHSE (upper hemisphere exposure), and LHSE (lower hemisphere exposure). First, continuous features were preprocessed as described in Section 3.1.1. The values ​​of the target variables were used as is; that is, they were not averaged as they already represent proximity values. Spearman's correlation coefficient was calculated for all window-mean features and all target variables in each protein sequence. Then, the correlations from all proteins were averaged for each feature-target pair. Next, for all three targets, the 30 most correlated features (positive or negative) were selected, and features with high correlation to all three targets were chosen.

[0118] 3.1.3 Correlation between categorical features and continuous targets These analyses were performed for RSA, UHSE, and LHSE. Categorical features provide a single value for each amino acid. These features were used as is for each amino acid in any given sequence. Point bilinear correlation coefficients between features and targets were calculated for each class. Targets were also used as is; i.e., there was no window for the same reasons as in Section 3.1.2. The correlations from all proteins were then averaged for each feature-target pair. The features that correlated most strongly (positively or negatively) with all three targets were selected as inputs to the model.

[0119] 3.1.4 Correlation between continuous features and categorical targets This analysis was performed on SS. The same procedure as in Section 3.1.2 was used for the feature window. The only difference was that the target was categorical. Point bilinear correlation coefficients were used to assess the correlation with continuous features. The correlations from all proteins were then averaged for each feature-target pair. Top-level correlated features (positive or negative) were used in the model.

[0120] 3.1.5 Correlation between categorical features and categorical targets This analysis was performed on SS. Chi-squared values ​​were used for each protein to assess the categorical feature-target dependency. Multiple hypothesis correction of the resulting p-values ​​was then performed using the Benjamin-Hochberg procedure. Finally, the harmonic mean was used to average the corrected p-values ​​from all sequences for each feature-response pair. Top-class significant feature-target correlations were used in the model.

[0121] 3.2 Second Method Correlation analysis was performed using the features in Table 0 to determine which features to use as input for predicting different targets of the models. Further information on each feature (such as traits) is provided in the appendix. The analysis was performed using all the data for these models. That is, the datasets used for these analyses were those described in Section 4.1 (for RSA, UHSE, LHSE, and SS) and Sections 5.1 and 6.1 (for CBCE and LBCE). Therefore, which data was used depends on the model being tested. Furthermore, for the CBCE and LBCE analyses, these two datasets were merged.

[0122] [Table 1]

[0123] [Table 2]

[0124] 3.2.1 Preprocessing of Continuous Features The same procedure as in Section 3.1.1.

[0125] 3.2.2 Feature-to-Feature Correlation First, feature-feature correlation analysis was performed on all features in Table 0. This was done on two different datasets: one from the dataset in Section 5.1, and the other from the merged dataset from Sections 6.1 and 7.1. The purpose of these analyses was to identify highly correlated feature pairs so that if both features from a correlated pair correlated with the response in a later analysis, only one of them would be used as the input feature for the corresponding model. Thus, the input feature space was reduced.

[0126] First, continuous features were preprocessed as described in Section 3.2.1. Categorical features were used as is; that is, a single value for each amino acid was used for all sequences. Next, correlation metrics were calculated for all feature pairs on all sequences. Spearman's correlation coefficient was used for continuous-continuous feature pairs. Kendall's rank correlation coefficient was used for continuous-categorical features, and the mutual information criterion was used for categorical-categorical features. Outliers were removed from all pairwise feature coefficient populations using the IQR (box plot) method. The median of the entire population was then considered the final coefficient for all pairwise feature-feature correlations. Continuous-continuous feature pairs were defined as having a median Spearman correlation coefficient > 0.7, continuous-categorical feature pairs with a Kendall rank correlation coefficient > 0.7, and categorical-categorical feature pairs with mutual information > 2.

[0127] Figure 6 shows the feature-feature pair correlation results for the RSA / HSE model. Figures 6(a) to (c) show feature-feature correlations using data from 6.1 and 7.1. Figures 6(d) to (f) show feature-feature correlations using data from 5.1. Black dots correspond to correlated pairs. The heatmap is symmetrical. Note that not all feature names are listed on the x and y axes. For the datasets from 6.1 and 7.1, the current model was used to predict RSA, UHSE, LHSE, and SS features (see 4). For the dataset from 5.1, these figures were observed from the PDB. As can be seen, for all datasets, continuous-continuous feature pairs yielded many correlated features. There was very little correlation between continuous-categorical feature pairs, but no correlation was observed between categorical-categorical feature pairs.

[0128] 3.2.3 Characteristics-Response Correlation The same procedure was followed for each of the following feature-response correlation analyses. For continuous features, the same preprocessing as described in Section 3.1.1 was performed, but using windows of 0, 3, 7, 9, and 11. For each feature-response pair in all sequences of the corresponding data, correlations were calculated based on all windows. Furthermore, outliers were removed using the IQR (box plot) method, and then the median of each correlation group was calculated.

[0129] All correlations from all windows for each feature were merged into a single population, outliers were removed using the same method, and finally the median of the global population was recalculated. The final "pooled" median was used to rank the features correlated with the corresponding responses from highest to lowest. Categorical responses and predicted RSA, UHSE, and LHSE were used as is, i.e., without windows.

[0130] Each of the following analyses provided an index of highly correlated feature-response pairs. When two features correlated with the response and with each other (using the results from Section 3.2.3), we selected the top-level feature with the highest correlation to the response so that only the single feature with the highest correlation to the response was used.

[0131] 3.2.3.1 Correlation between continuous features and continuous targets These analyses were performed for RSA, UHSE, and LHSE. Data from 5.1 was used. The metric used here was Spearman's correlation coefficient. See Figure 7 for the results and Section 5 for the models and features used.

[0132] Figure 7 shows Spearman correlation coefficients for continuous features for (a) RSA, (b) UHSE, and (c) LHSE. In each figure, the x-axis represents the correlation, and the y-axis represents the feature names sorted from highest to lowest absolute correlation of the median of the population. Each plot is from all windows.

[0133] 3.2.3.2 Correlation between categorical features and continuous targets These analyses were performed for RSA, UHSE, and LHSE. Data from 5.1 was used. The metric used here was the Kendall rank correlation coefficient. See Figure 8 for the results and Section 5 for the models and features used.

[0134] Figure 8 shows Kendall correlation coefficients for categorical features for (a) RSA, (b) UHSE, and (c) LHSE. In each figure, the x-axis represents the correlation, and the y-axis represents the feature names sorted from highest to lowest absolute correlation of the median of the population. Each plot is from all windows.

[0135] 3.2.3.3 Correlation between continuous features and categorical targets These analyses were performed for SS, CBCE, and LBCE. Data from section 5.1 (for SS) and from sections 6.1 and 7.1 (for CBCE and LBCE) were used. The metric used here was the Kendall rank correlation coefficient. See Figure 9 for the results, and sections 5, 6, and 7 for the models and features used.

[0136] Figure 9 shows Kendall correlation coefficients for continuous features for (a) SS, (b) CBCE, and (c) LBCE. In each figure, the x-axis represents the correlation, and the y-axis represents the feature names sorted from highest to lowest absolute correlation of the median of the population. Each plot is from all windows.

[0137] 3.2.3.4 Correlation between categorical features and categorical targets These analyses were performed for SS, CBCE, and LBCE. Data from section 5.1 (for SS) and from sections 6.1 and 7.1 (for CBCE and LBCE) were used. The metric used here was the mutual information criterion. See Figure 10 for the results, and sections 5, 6, and 7 for the models and features used.

[0138] Figure 10 shows the mutual information criteria for categorical features for (a) SS, (b) CBCE, and (c) LBCE. In each figure, the x-axis is correlation, and the y-axis is the feature name sorted from highest to lowest absolute correlation of the median of the population. Each plot is from all windows.

[0139] 4. Prediction of 3D features ("Second Machine Learning Algorithm") Two BLSTM models were created to predict the structure and surface properties of proteins from their natural sequences. These models were later used in the first machine learning model to predict the structural and surface properties of Ag before the binding event with Ab occurs. For structural properties, we chose to create a model for SS, and for surface properties, we chose RSA, UHSE, and LHSE.

[0140] 4.1 Data preparation for 3D features To accurately predict SS, RSA, UHSE, and LHSE from native protein sequences, we used all available high-quality 3D protein structures from all available biological PDBs. The goal of these models was to predict the most appropriate surface and structural properties of a single protein, unaffected by binding interactions to any other protein, including Ab-to-B binding. Therefore, only structures not bound to other structures or molecules were retained. Structures containing two or more copies of the same molecule, but with slightly different conformations, were retained in the data. By filtering the structures and retaining only those with a resolution of ≤3 angstroms, we ensured that every atom of each amino acid was reliably mapped using coordinates and using protein chains longer than 200 amino acids.

[0141] After filtering, the database consisted of a total of 41,592 structures. These structures contained 70,489 protein sequences, resulting in 53,524 unique sequences. Subsequences of longer sequences were retained as separate data point entries because their shorter lengths may result in different 3D conformal folding characteristics.

[0142] The DSSP algorithm was used to calculate SS and RSA for each molecule in each structure file. DSSP calculates the following secondary structure classes for protein sequences: α-helix, 310-helix, π-helix, isolated β-bridge, β-strand, turn, bend, and coil. These classes were merged into three superclasses: helix (α-helix, 310-helix, and π-helix), strand (isolated β-bridge and β-strand), and coil (turn, bend, and coil). Finally, the BioPython package was used to calculate UHSE and LHSE.

[0143] At the end of filtering, each amino acid in the database was assigned RSA, UHSE, LHSE values, and SS class. To create a unique database, the RSA, UHSE, and LHSE values ​​for each amino acid were averaged across identical sequences. Finally, amino acids with identical sequences but different SS classes were assigned to a coil class.

[0144] 4.2 Predictive models for RSA and HSE (struct 3d rsa hse m2 model) We chose to create a single BLSTM model to predict all surface features. More specifically, this model predicts RSA, UHSE, and LHSE from the primary protein sequence. While LHSE may not provide useful information about the surface location of amino acids, both together form a stochastic metric around each amino acid, which can help predict UHSE more accurately.

[0145] This is a BLSTM model that takes a batch of features calculated for each amino acid from each input sequence as input and predicts a ternary output. For each protein sequence given as input, the values ​​of each RSA, UHSE, and LHSE are predicted for each amino acid.

[0146] For all three outputs, the individual losses were calculated based on the mean squared error (MSE), although the losses are not limited to this function in principle. The global model loss was the weighted sum of the individual losses, weighted 50, 100, and 125 for RSA, UHSE, and LHSE, respectively. The weights were determined by first training a model without them and then examining the difference in magnitude of the three losses. The weights were designed so that the three losses gave equal contributions to the global model loss.

[0147] Figure 11 shows a schematic, simplified architecture of the model. As shown, the first layer is the input layer. The input layer may have one or more input nodes to receive the input features, for example, as considered below. Each input node is coupled to its respective masking node in the masking layer. Similarly, each masking node is coupled to its respective BLSTM node in the BLSTM layer. Each BLSTM node is coupled to a concatenation node in the concatenation layer. Thus, multiple outputs from the nodes of the preceding BLSTM layers are concentrated and integrated in the concatenation layer. The concatenation layer may be coupled to one or more further BLSTM layers, each containing a corresponding BLSTM node. The final nodes of these BLSTM nodes are coupled to the output layer, which in this case contains multiple output nodes corresponding to the RSA, UHSE, and LHSE output predictions, respectively.

[0148] Note: The input features were calculated amino acid by amino acid without being averaged by a window.

[0149] 4.2.1 Preliminary analysis and characteristics used For the procedure, please refer to the procedure in Section 3.1. The 14 features (including the binary representation of protein sequences) selected for use in the model can be seen in Table 1.

[0150] [Table 3]

[0151] 4.2.2 Pre-training See section 2.1.

[0152] 4.2.3 Cross-Validation (CV) Results Approximately 80% of the data was used in a 5-fold cross-validation to evaluate the model's performance. See Figure 12 for the Spearman correlation coefficient metric. The model performed well without overfitting and exhibited high performance across all evaluation metrics. Furthermore, the model demonstrated high stability, with only slight variability observed between different cross-validation runs.

[0153] Figure 12 shows the results for the RSA / HSE model. Figures 12(a) to (c) show Spearman correlation coefficients, and 12(d) to (f) show mean squared error (MSE) loss. The x-axis is the number of epochs, and the y-axis is the metric value. The plots are the average of 5-fold cross-validation, including Gaussian CI. The blue line represents the training case, and the red line (dashed line) represents the validation set case.

[0154] 4.2.4 Test Results The remaining approximately 20% of the data was used as test data to evaluate the model's performance on completely unknown data. See Figure 13 for the Spearman correlation coefficient metric. The model performed well without overfitting and exhibited high performance across all evaluation metrics.

[0155] Figure 13 shows the results of the RSA / HSE model. (a)-(c): Spearman correlation coefficients. (d)-(f): MSE loss. The x-axis is the number of epochs, and the y-axis is the metric value. The solid line represents the training case, and the dotted line represents the validation set case.

[0156] 4.2.5 Final Training Final training was performed on the entire dataset, using the optimal number of epochs determined by the cross-validation procedure. See Figure 14 for the Spearman correlation coefficient metric. The model appears to perform well across all metrics. The model was trained for 70 epochs selected based on cross-validation and test set analysis.

[0157] Figure 14 shows the results for the RSA / HSE model. (a)-(c): Spearman correlation coefficients. (d)-(f): MSE loss. The x-axis is the number of epochs, and the y-axis is the metric value.

[0158] 4.3 Predictive model for SS (struct 3d ss m3 model) We chose to construct a BLSTM model to predict three-layer class outputs for SS. The model uses categorical cross-entropy loss to assign each amino acid in its input sequence to one of the following classes: helix, strand, or coil. The model architecture is substantially identical to that illustrated in Figure 11, which has a single output node in the output layer. Note: Input features were calculated amino acid by amino acid because they are not averaged across the sliding window.

[0159] 4.3.1 Preliminary analysis and characteristics used Refer to the procedure in Section 3.1. The eight features chosen for use in the model (including the binary representation of protein sequences) can be seen in Table 2.

[0160] [Table 4]

[0161] 4.3.2 CV Results Approximately 80% of the data was used in a 5-fold cross-validation to evaluate the model's performance. See Figure 15 for the high-accuracy recall metric. The model appears to perform well without significant overfitting, giving high performance across all metrics. Furthermore, only slight variations in loss are observed between different cross-validation runs. Finally, for demonstration purposes, the model was trained for 70 epochs, exhibiting slight overfitting from epoch 60 onward.

[0162] Figure 15 illustrates the results of the SS model. Figures 15(a) to (e) show the PR curves. The x-axis is TPR and the y-axis is PPV. Figure 15(f) shows the categorical cross-entropy loss. The x-axis is the number of epochs and the y-axis is the loss value. The plot is the average of 5-fold cross-validation, including Gaussian CI. The blue line represents the training case and the red line (dashed line) represents the validation set case.

[0163] 4.3.3 Test Results The remaining approximately 20% of the data was used as test data to evaluate performance on completely unknown data. See Figure 16 for the high-precision recall metric. The model performed well without significant overfitting and showed high performance for all evaluation metrics. The model was trained for 100 epochs, selected based on the CV analysis, so that it required slightly more epochs than the CV analysis.

[0164] Figure 16 shows the results of the SS model. Figure 16(a): PR curve. The x-axis is TPR and the y-axis is PPV. Figure 16(b): Categorical cross-entropy loss. The x-axis is the number of epochs and the y-axis is the loss value. The solid line represents the training case and the dotted line represents the validation set case.

[0165] 4.3.4 Final Training Final training was performed on the entire dataset, using the optimal number of epochs determined by the cross-validation procedure. See Figure 17 for the F1 metric. The trained final model performed well without significant overfitting and demonstrated high performance across all metrics. The model was trained for 100 epochs, selected based on the cross-validation analysis, with a slightly greater number of epochs than the CV and test set analysis.

[0166] Figure 17 shows the results of the SS model. Figures 17(a)-(c): F1 metric. The x-axis is the number of epochs, and the y-axis is the F1 value. Figure 17(d): Categorical cross-entropy loss. The x-axis is the number of epochs, and the y-axis is the loss value.

[0167] 5. Prediction of 3D features ("Second Machine Learning Algorithm") (Version 2) 5.1 Data preparation for 3D features (version 2) This is identical to Section 4.1. The only difference is that there was one incorrect assignment of classes to the SS classes. Specifically, the isolated β-bridge class was assigned to the coil superclass, but here it is correctly moved to the strand superclass. This minor detail only affects a portion of the SS model data.

[0168] 5.2 RSA and HSE Prediction Model (Version 2) (RSA HSE M4 V1 Model) This section is identical to Section 4.2, with only the preliminary analysis and input features being modified. The model is also deeper. The model architecture is substantially the same as that illustrated in Figure 11. Note: Input features were calculated amino acid by amino acid without averaging across the sliding window.

[0169] The global model loss was a weighted sum of the individual models, as in the previous model. However, this time the weights were 1, 10, and 125 for RSA, UHSE, and LHSE, respectively. The weights were determined by first training a model without it and observing the difference in the magnitude of the three losses.

[0170] 5.2.1 Preliminary analysis and characteristics used For the procedure, please refer to Section 3.2. The 19 features (including binary representations of protein sequences) selected for use in the model can be seen in Table 3.

[0171] [Table 5]

[0172] 5.2.2 Pre-training See section 2.2.

[0173] 5.2.3 CV Results Approximately 80% of the data was used with 5-fold cross-validation to evaluate the model's performance.

[0174] 5.2.5 Test Results The remaining approximately 20% of the data was used as test data to verify performance with completely unknown data.

[0175] 5.2.6 Final Training The final training was performed on the entire dataset, using the optimal number of epochs determined by the cross-validation procedure.

[0176] 5.3 SS Prediction Model (Version 2) (SS3 M6 V1 Model) The results are similar to those described in Section 4.3. The model architecture is substantially identical to that illustrated in Figure 11. Note: Input features were calculated amino acid by amino acid without being averaged by a window.

[0177] 5.3.1 Outlier This is identical to Section 4.3.

[0178] 5.3.2 Preliminary analysis and characteristics used For the procedure, please refer to Section 3.2. The 19 features (including binary representations of protein sequences) selected for use in the model can be seen in Table 4.

[0179] [Table 6]

[0180] 5.3.3 Pre-training See section 2.2.

[0181] 5.3.4 CV Results Approximately 80% of the data was used with 5-fold cross-validation to evaluate the model's performance.

[0182] 5.3.5 Test Results The remaining approximately 20% of the data was used as test data to verify performance with completely unknown data.

[0183] 5.3.6 Final Training The final training was performed on the entire dataset, using the optimal number of epochs determined by the cross-validation procedure.

[0184] 6. The First Machine Learning Model (CBCE) A BLSTM-based model was used to predict CBCE. This allowed for bidirectional modeling of the entire protein sequence as a single observation without the need for segmentation. Thus, remote and anterior-posterior amino acid relationships were captured by the model. Furthermore, the use of BLSTM enabled simultaneous training of different protein lengths. The primary objective of the model was to create a non-ab-specific CBCE predictor. This was achieved by the training data preparation method described in Section 6.1.

[0185] 6.1 Data Preparation To model CBCE, we downloaded non-old protein complexes from the PDB. We allowed any resolution and biological structure as long as they had at least three different protein chains. This was because two of the chains were Ab's V H and V L This is because it can be a chain, and the third chain can be Ag. Naturally, multiple Ab or Ag atoms can exist in a single PDB structure.

[0186] A local database was created using all the V, D, and J region genes of immunoglobulins from the IMGT (International ImMunoGeneTics Information System) database. V genes were collected from all available organisms, namely humans, mice, rhesus monkeys, rabbits, and rats. H and V L I downloaded both genes. IMGT is the V of Ab. H and V L This provides information about the chains and their paratope regions (the CDR [antibody complementarity determination region] and FR [antibody framework region] of each chain).

[0187] To identify Ab in this local database, IgBlast (Jian Ye, Ning Ma, Thomas L Madden, and James M Ostell. Igblast: an immunoglobulin variable domain sequence analysis tool. Nucleic Acids Res, 41 (Web Server issue): W34-40, Jul 2013) was used for proteins in the IMGT database. To identify paratopes in the chain, the Kabat system provided by IgBlast was used (G Johnson and TT Wu. Kabat database and its applications: 30 years after the first variability plot. Nucleic acids research, 28(1):214-218, 01 2000). V was only identified if at least CDR1 and CDR2 were found by IgBlast. H and V L The chains were considered valid. Since CDR3 is more difficult to map, we allowed for its absence in the search. Then, all protein chains of each structure were aligned to their database. At least one chain was considered valid. H As and one valid V L If it was not mapped as such, the structure was filtered and removed. VH Even if V L All other chains that were not mapped were assumed to be Ag chains. They were longer than 100 amino acids and V H Chains and V L A chain that is not bound by any other chains was considered an effective Ag chain. Finally, at least one V H chain, V L Only the chains and structures containing the Ag chain were used for further analysis.

[0188] Next, in each structure, V H Chain and V L The effective Ab was reconstructed by pairing the chains. Multiple V molecules were found in a single structure. H Chain and V L If a chain exists, each V H From all atoms on the chain, each V L The average distance to all atoms on the chain was measured. V has the minimum average atomic distance. H Chain and V L We assigned the chains as pairs and assumed they belonged to a single Ab. Standalone V H Chain or V L The chain was not considered in downstream analysis because a complete Ab could not be defined. Paratope analysis provided information on contact formation with each Ag. Two amino acids were assumed to be in contact if either of their atoms was located within a certain probe distance from each other. Using probe distances of 4, 6, and 8 angstroms, the total number of Ag amino acids forming contact with each paratope moiety was calculated. Figure 18(a) shows the normalized mean of those contacts (x axis represents amino acids, y axis represents paratope moieties). The majority of contacts were V H CDR1 and CDR2 and V L V performed on CDR1 and CDR3 HThe absence of CDR3 above may be due to the difficulty in identifying it using IgBlast. Furthermore, the FR region does not appear to form as many contacts as the CDR region, as expected. In addition, the relatively low standard deviation of the total number of contacts per amino acid and paratope in Figure 18(b) suggests that a similar number of contacts occur between different Ag and Ab. This suggests the possibility of predicting CBCE without any further information about the actual Ab. Figure 18(b) shows the standard deviation of the number of contacts. The x-axis represents amino acids and the y-axis represents paratopes.

[0189] Each V H and V L For a chain pair of V, a single CBCE was defined. This is for any V H and V L This was done by first identifying any Ag atom whose distance from the CDR region of the chain pair was ≤4 angstroms. The amino acids to which those atoms belonged were defined as contacts between Ag and Ab, i.e., as CBCEs. Multiple CBCEs from different Abs may be defined on the same Ag. Finally, structures that did not define any CBCEs within a 4-angstrom distance were discarded.

[0190] Observed CBCEs were also mapped to similar Ag. Undiscovered CBCEs may increase the false negative rate in the predictive model. To reduce the likelihood of assigning undiscovered CBCEs as negative data, observed CBCEs were copied to similar Ag. First, clusters of similar Ag were identified using BlastPlus with >90% similarity and a maximum gap of 2. CBCEs were copied from Ag within a cluster to all other Ag within the same cluster, as long as the corresponding location and distance were exactly identical based on the mapping from BlastPlus.

[0191] Finally, a unique Ag local database was created. Duplicate Ag sequences were completely removed from the data. On the other hand, Ag sequences that are subsequences of longer Ag sequences were retained in the data and treated as different observations. The resulting database consisted of 1003 Ag sequences and 12968 CBCEs in FASTA format, of which 6986 were unique. Each Ag sequence was associated with at least one CBCE.

[0192] 6.2 Summary of the CBCE Model We created two main CBCE predictors, namely model 13 conf v1 and model 17 conf v1. The models have two main differences. First, the negative data generators are different; see Section 6.3 for details. Second, model 13 conf v1 gives two outputs, while model 17 conf v1 gives one output. Both models take permutation vectors as input features. Every sequence in the data is associated with at least one true CBCE, and these CBCEs are converted into binary 2D vectors and given as input to the models. The common output of the two models is a probability, as both models essentially ask the question, "Is this permutation vector a true CBCE on this particular sequence?" "Is this permutation vector a true CBCE on this particular sequence, or not?" The second output of model 13 conf v1 is the permutation vector itself. This part of the model acts as an AE, returning a probability for each amino acid. This output can be seen as the contribution of each amino acid in the sequence to the specific CBCE in question.

[0193] The objective of these models was to predict the CBCE on the protein Ag sequence before the Ab binding event occurred. The dataset used, described in Section 6.1, contains linear protein sequences and true CBCEs. Therefore, there is no information regarding the structural or surface properties of these protein sequences. However, information regarding the secondary structure and surface appearance of each protein sequence before the binding event is needed. Accordingly, predictive models for RSA, UHSE, and LHSE (model struct 3d rsa hse m2 res), as well as SS (model struct 3d ss m3 res1), described in Section 4, were used to predict their features for all proteins in the dataset.

[0194] The model differs from existing publicly available CBCE algorithms in three main ways. Firstly, by using BLSTM, it becomes possible to capture the relationships between distant and preceding / preceding amino acids that can constitute CBCE on the natural folded 3D protein structure. Other algorithms, on the other hand, segment the protein sequence into k-mers and therefore do not capture these important relationships. Secondly, the training dataset is constructed to predict true CBCE before the binding event occurs. This difference conceptually brings the model closer to modeling the reality of Ab binding. The 3D folded protein structure is, strictly speaking, what Ab recognizes and triggers binding, not what Ab has already recognized and bound to.

[0195] Thirdly, existing publicly available CBCE algorithms require experimentally measured or potentially predicted 3D protein Ag structures to predict epitopes. The first machine learning model does not require the 3D coordinates of all atoms of each amino acid on Ag or Ab. Finally, the model addresses the epitope question in a conceptually very different way, but in a way that is more finely tuned to capture bone fide ("true") BCEs. Because it predicts the probability of standalone CBCEs on each Ag as pairs or independent classes, multiple CBCEs on the same protein sequence can be discovered and evaluated independently. The other algorithms, on the other hand, provide amino acid-by-amino acid probabilities without the ability to separate different CBCEs.

[0196] Figure 19(a) schematically shows the architecture of the model 13 conf v1 model. The model architecture is substantially identical to that shown in Figure 11. Here, the output layer has two output nodes. The BCE amino acid output (BCE_aa) is the actual permutation output using binary cross-entropy loss. The BCE perm output is a probability vector indicating whether the input permutation is a true CBCE (second position on the vector) or not (first position on the vector). Figure 19(b) shows the architecture of the model 17 conf v1 model. BCE perm represents the same output as the previous model. Note: Input features were calculated amino acid by amino acid without averaging by a window.

[0197] For both models, the BCE perm output is calculated from amino acid values. The final layer of the output is a dense layer containing sigmoid activation. As a result, for each protein sequence, a value ∈[0,1] is returned for each amino acid. The two-class probability vector is then calculated for that sequence as follows:

number

[0198] The model 13 conf v1 model had a weighted global loss. The global model loss is the weighted sum of the individual model losses, i.e., the weighted sum of BCE perm and BCE aa, using weights 1 and 1, respectively. Similarly, it is possible to test various weights.

[0199] 6.3 Negative Data When attempting to model CBCE amino acids individually, positive data would be all amino acids in a sequence that are part of any true CBCE, and negative data would be any other amino acids on the same sequence. However, since we model the entire CBCE as a single permutation vector of a given protein sequence, the response to this vector is single probability (this is the BCE perm output from Section 6.2). Since we observed only true positive CBCEs from the data in Section 6.1, there are no permutations corresponding to negative data.

[0200] We developed a solution to improve the model's performance. In doing so, we generated negative permutations of data to resample as close to ground truth as possible. To cover a wide range of negative data and make the model more robust, we generated negative permutations for all sequences in the data at every epoch during training. The following sections describe two different generation schemes.

[0201] 6.3.1 CBCE Model (model 13 conf v1) This model generated completely random permutations. For each sequence in the training and validation data, 10 completely random permutations were generated, and it was assumed that they were not true CBCEs. Both the total number of amino acids in a given sequence and their arrangement were random. A new permutation was generated in each epoch.

[0202] The advantage of this method is that, because it is completely random, the likelihood of any generated CBCE permutation being false negative is quite low. However, the disadvantage is that randomly generated CBCE permutations may differ significantly from true CBCE permutations.

[0203] 6.3.2 CBCE Model (model 17 conf v1) This model was subjected to three different methods for generating negative data. For each sequence in the training and validation data, 30 completely random permutations were generated, 10 from each method, and assumed not to be true CBCEs. The first method is similar to that described in 6.3.1. The second and third methods were applied to all true CBCEs in a given protein sequence. The second method kept the first and last amino acids of a given true CBCE in the correct positions, while randomly shuffling the internal CBCE amino acid indices within the region. The third method kept the total amount of amino acids in the true CBCE and their linear distances constant. Then, the entire true CBCE on other parts of the protein was randomly shifted. A new permutation was generated from all methods in each epoch.

[0204] The advantage of this method is that the randomly generated CBCE permutations cover both very similar and very different true CBCE permutations. This likely leads to the algorithm being more robust to both positive and negative data. On the other hand, the disadvantage of this method is that it may increase false negatives, leading to decreased performance with new data.

[0205] 6.4 Outlier Outliers were removed before training any model. This was done by applying an isolation forest (IF) to the total amino acid of each observed CBCE using data from Section 5.1. The IF algorithm from the sklearn Python package was run for all observed total amino acids as the max samples argument (F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830, 2011.). The most common minimum and maximum boundaries from all runs were considered the global boundaries of the inliers.

[0206] Figure 20 shows the total amount of amino acids from all observed CBCEs. Figures 20(a) and 20(b) show the density and box plots of the observed total amount of amino acids in the true CBCE, respectively. Figures 20(c) and 20(d) show the logarithmic scales of Figures 20(a) and 20(b), respectively. Here, it can be seen that the median is 10 amino acids. After outlier analysis, it was found that the effective CBCE should contain a total amount of amino acids within a range smaller than that discussed in Section 1.3. However, these studies did not detect or remove outliers from the observed CBCEs. As can be seen from Figure 20, the maximum amount of total amino acids before removing outliers was approximately 30 to 35, which is consistent with what was reported in previous studies. Finally, outliers were removed from the data and the model was trained without them.

[0207] 6.5 Preliminary analysis and characteristics used See Section 3.2 for the procedure. The 14 features selected for use in the model (including binary representations of protein sequences and permutation vectors) are shown in Table 5. Note that RSA and SS were predicted by the models struct 3D rsa hse m2 res and struct 3D ss m3 res1, respectively, as described in Section 4. The same features were used in both model 13 conf v1 and model 17 conf v1.

[0208] [Table 7]

[0209] 6.6 Pre-training See section 2.2.

[0210] 6.7 CV results The "age" of protein Ag in the data was defined as the date the most recent CBCE was registered in the protein sequence. To evaluate the model's performance, approximately 94% of the "oldest class" proteins in the data were used in a 5-fold cross-validation. Custom callbacks were used in TensorFlow to allow for early termination if the loss did not decrease.

[0211] 6.7.1 Model model 13 conf v1 Figure 21 shows the results for the precision-recall metric. The model performed well without significant overfitting and exhibited high performance for all evaluation metrics. Furthermore, only slight variability in loss was observed between different CV runs. The model was trained for approximately 466 epochs, and the apparent bump in loss may suggest that the random generator produces negative CBCE permutations similar to positive ones. However, this negative impact diminished as the epochs progressed.

[0212] Figure 21 illustrates the results of the CBCE model 13 conf v1. Figures 21(a)-(e): PR curves. The x-axis is the TPR, and the y-axis is the PPV. Figure 21(f): Weighted cross-entropy loss. The x-axis is the number of epochs, and the y-axis is the loss value. The plots are the average of 5-fold CV including Gaussian CI. The blue line is for the training case, and the red line (dashed) is for the validation set case.

[0213] 6.7.2 Model model 17 conf v1 Figure 22 shows the results of the precision-recall metric. The model appears to perform well without significant overfitting before 250 epochs, giving high performance on all metrics. Moreover, slight variation is observed between different CV runs of the loss. The model was trained for 455 epochs.

[0214] Figure 22 illustrates the results of the CBCE model 17 conf v1. Figures 22 (a)-(e): PR curves. The x-axis is the TPR, and the y-axis is the PPV. Figure 22(f): Weighted cross-entropy loss. The x-axis is the number of epochs, and the y-axis is the loss value. The plots are the average of 5-fold CV including Gaussian CI. The blue line is for the training case, and the red line (dashed) is for the validation set case.

[0215] 6.8 Test Results To examine the performance on completely unknown data, approximately 6% of the "latest" proteins were used in the data. The reason for holding the latest data as the test set is to use the same data for comparison with other algorithms, thus ensuring that those algorithms were not trained on that data. Training in the following models was done on approximately 94% of the "oldest class" of the data.

[0216] 6.8.1 Model model 13 conf v1 Figure 23 shows the results for the precision-recall metric. The model performed well without significant overfitting and showed high performance for all evaluation metrics. The model was trained for 466 epochs, which was selected based on the CV analysis to be slightly more epochs than the CV analysis.

[0217] Figure 23 illustrates the results for CBCE model 13 conf v1. Figure 23(a): PR curve. The x-axis is TPR and the y-axis is PPV. Figure 23(b): Overall model loss. The x-axis is the number of epochs and the y-axis is the loss value. The solid line represents the training case and the dotted line represents the validation set case.

[0218] 6.8.2 Model 17 conf v1 Figure 24 shows the results for the precision-recall metrics. The model appears to perform well without significant overfitting and gives high performance across all metrics. The model was trained for 455 epochs, which was chosen based on the CV analysis to be slightly more epochs than the CV analysis.

[0219] Figure 24 illustrates the results for CBCE model 17 conf v1. Figure 24(a): PR curve. The x-axis is TPR and the y-axis is PPV. Figure 24(b): Overall model loss. The x-axis is the number of epochs and the y-axis is the loss value. The solid line represents the training case and the dotted line represents the validation set case.

[0220] 6.9 Final Training The final training was performed on the entire dataset, using the optimal number of epochs determined by the cross-validation procedure.

[0221] 6.9.1 Model model 13 conf v1 Figure 25 shows the results for the F1 metric. The model appears to perform well without significant overfitting, giving high performance across all metrics. The model was trained over 500 epochs selected based on CV analysis.

[0222] Figure 25 shows the results for CBCE model 13 conf v1. Figures 25(a)-(b): F1 metric. The x-axis is the number of epochs, and the y-axis is the F1 value. Figure 25(c): Overall model loss. The x-axis is the number of epochs, and the y-axis is the loss value.

[0223] 6.9.2 Model 17 conf v1 Figure 26 shows the results for the F1 metric. The model appears to perform well without significant overfitting, giving high performance across all metrics. The model was trained over 700 epochs selected based on CV analysis.

[0224] Figure 26 shows the results for CBCE model 17 conf v1. Figure 26(a): F1 metric. The x-axis is the number of epochs, and the y-axis is the F1 value. Figure 26(b): Overall model loss. The x-axis is the number of epochs, and the y-axis is the loss value.

[0225] 7. The First Machine Learning Model (LBCE) To predict LBCE, we similarly use the BLSTM model. This allows us to model the entire protein sequence as a single observation without the need for segmentation. Therefore, distant amino acid relationships should be captured by the model. Furthermore, the use of BLSTM should allow us to train different protein lengths simultaneously. Thus, the model type here is identical to the CBCE model type in Section 6. The primary objective of the model was to create a non-Ab specific LBCE predictor. This is achieved by the method of preparing the training data as described in Section 7.1.

[0226] 7.1 Data Preparation For LBCE modeling, I downloaded LBCE data from IEDB. I downloaded all non-old assays from all organisms from IEDB.

[0227] The primary objective of the model was to create a non-ab-specific LBCE predictor, so all positive LBCEs were considered for the database. That is, an LBCE was included in the database if it tested positive for at least one Ab. Next, each Ag ID was mapped to the Universal Protein Resource (UniProt) to obtain the protein sequence in fasta format. Assays were discarded if the ID was unmappingable, or if the ID was mappingable but the coordinates of the corresponding LBCE did not give the same sequence on the searched Ag sequence.

[0228] Observed LBCEs were mapped to similar Ags, as was done for CBCEs in Section 6.1. Undiscovered LBCEs may increase the false negative rate in the predictive model. To reduce the likelihood of assigning undiscovered LBCEs as negative data, observed LBCEs were copied to similar Ags. First, clusters of similar Ags were identified using BlastPlus with >90% similarity and a maximum gap of 2. LBCEs were copied from Ags within a cluster to all other Ags within the same cluster, as long as the corresponding linear coordinates were exactly identical based on the mapping from BlastPlus.

[0229] Finally, a unique Ag database was created. Duplicate Ag sequences were completely removed from the data. On the other hand, Ag sequences that are subsequences of longer Ag sequences were retained in the data and treated as different observations. The resulting database consisted of 4695 Ag sequences and 177782 CBCEs in fasta format, of which 63363 were unique. Each Ag sequence was associated with at least one LBCE.

[0230] 7.2 Overview of the LBCE Model Every sequence of data is associated with at least one true LBCE. The model takes permutation vectors as input features, which are LBCEs, converted to binary 2D vectors, and given as input to the model. The model's output is a probability, as it essentially asks "is this permutation vector a true LBCE on this particular sequence?". Essentially, the permutation vectors are identical to those described in Section 6.2, the only difference being that they represent LBCEs instead of CBCEs.

[0231] The goal of these models was to predict LBCEs before the binding event occurred. The dataset used is described in Section 7.1 and contains linear protein sequences and true LBCEs. Therefore, there is no information about the structural or surface properties of those protein sequences. However, information about what the surface of each protein sequence looks like before the binding event is needed. Therefore, to predict their features for all proteins in the dataset, we used the RSA predictive model (model struct 3d rsa hse m2 res) described in Section 4.

[0232] The model differs from publicly available LBCE algorithms in three main ways. Firstly, by using BLSTM, it is possible to capture the relationships between distant amino acids that can constitute an LBCE. Other algorithms, on the other hand, segment the protein sequence and thus cleave such bonds. Secondly, the training dataset is designed to predict a true LBCE before the binding event occurs—one that Ab recognizes and triggers binding, not one that Ab has already recognized and bound to. Finally, to predict the probability of a standalone LBCE, it can discover and separate multiple LBCEs on the same protein sequence. Other algorithms, on the other hand, provide amino acid-by-amino acid probabilities without the ability to separate different LBCEs.

[0233] The architecture of the model 12 linear v1 model has a structure that is substantially the same as the model depicted in Figure 19(b). The bce output is a probability vector indicating whether the input permutation is the true LBCE (the second position on the vector) or not (the first position on the vector). The bce output is calculated from the amino acid values as described in Section 6.2. Note: The input features were calculated for each amino acid as is, without being averaged by a window.

[0234] 7.3 Negative Data The procedure similar to Section 6.3.1 was executed. The only difference here was ensuring that the randomly generated LBCE was not truncated. Let N be the length of a given protein sequence. First, generate the total amount X of amino acids that the random LBCE has from a uniform distribution of [0,N]. Then, generate the starting position i of this LBCE from a uniform distribution of [0,N-X]. Finally, assign all the amino acids within [i,i+X] as corresponding to this random LBCE. The randomly generated LBCE with a total amount of 10 was the pre-sequence generated for each epoch.

[0235] 7.4 Outliers Outliers were removed before any training of the model 12 linear v1 model. Following the exact same procedure as Section 6.4, the data used was the total amount of amino acids of each observed LBCE in the data from Section 7.1.

[0236] Figure 27 shows the total amount of amino acids from all observed LBCEs. Here, it can be seen that the median is 17 amino acids. After outlier analysis, it was found that the valid LBCEs should contain the total amount of amino acids in the range [5,30]. The outliers were finally removed from the data, and the model was trained without them.

[0237] Figure 27 shows the outlier analysis of LBCE. Figures 27(a) and (b) show box plots of the density and observed amounts of total amino acids in true LBCE, respectively. Figures 27(c) and (d) show the logarithmic scales of Figures 27(a) and 27(b), respectively.

[0238] 7.5 Preliminary analysis and characteristics used See Section 3.2 for the procedure. The 13 features chosen for use in the model (including the binary representation of the protein sequence and permutation vectors) can be seen in Table 6. Note that RSA was predicted by the model struct 3d rsa hse m2 res described in Section 4.

[0239] [Table 8]

[0240] 8. Applications Precise prediction of B cell epitopes (particularly 3D conformational B cell epitopes) will lead to significant improvements in diagnosis, drug design, and vaccine development in the fields of autoimmunity, infectious diseases, and cancer. Examples are provided below.

[0241] 8.1 Therapeutic Antibody Mapping Antibody binding events to antigens are crucial in activating humoral B-cell immune responses to any pathogenic threat, including cancer antigens in malignant tumor cells, as well as viral and bacterial infections. B-cell epitope mapping is essential for therapeutic antibody development. The prediction of true B-cell epitopes according to this invention not only provides a functional understanding of critical residues involved in antibody-antigen binding, but also aids in the selection of therapeutic antibodies against predicted epitopes that are preferred for further development as therapeutic antibodies.

[0242] 8.2 Vaccine Development In antibody-driven vaccine design, in silico prediction is performed to predict BCEs that can trigger a humoral immune response. Previous in silico-guided vaccine design initiatives tend to selectively apply BCE prediction by focusing on surface exposure sites within the antigen. However, by using the approach of the present invention (particularly by training on unbound protein structures and using candidate BCE codes ("permutation vectors") mapped to antigen protein sequences), BCEs can be identified with greater accuracy and guide Ab-based vaccine design with much higher fidelity. In addition to potentially generating long-lasting and potent universal immunity, BCE vaccines guided by the approach of the present invention can advantageously identify specific antibody-epitope interactions on a large scale that have the ability to evade undesirable immune responses. Moreover, the method described herein avoids the need for years of target discovery and laborious, costly, and time-consuming work with minimal cost and time.

[0243] 8.3 Immunological diagnosis and immunological monitoring Predicting B cell epitopes directly contributes to the design of therapeutic strategies and immunodiagnostic and immunomonitoring reagents. Diagnosis can be significantly accelerated by integrating predictive approaches for accurate and comprehensive detection of BCEs. One example is the rapid and accurate detection of BCE epitopes on the spike protein of betacoronavirus, enabling COVID diagnosis in a way that can distinguish between infections caused by common cold viruses and those caused by the SARS-CoV-2 virus that causes COVID. Following the same approach, BCE predictive techniques can be used to identify immunodominant or humorally reactive epitopes in tumor-associated or nascent antigens in immunodiagnostic and immunomonitoring clinical applications.

[0244] 8.4 System Example Figure 4 schematically illustrates an example of a system suitable for realizing an embodiment of the present method. The system 1100 comprises at least one server 1110 that communicates with a reference data store 1120. The server may also communicate with an automated peptide synthesis device 1130, for example, via a communication network 1140.

[0245] In a particular embodiment, the server may first use, for example, a reference data store to obtain the amino acid sequences of one or more query proteins. The server may then predict whether the query protein contains BCEs that are likely to trigger an antibody binding event (e.g., an immunogenic response).

[0246] A query protein predicted to contain one or more BCEs may be sent to an automated peptide synthesis device 1130 for the synthesis of the query protein or a portion thereof. Techniques for automated peptide synthesis are well known in the art, and it will be understood that any known technique may be used. Typically, the query protein or epitope is synthesized using standard solid-phase peptide synthesis chemistry, purified using reverse-phase high-performance liquid chromatography, and then formulated in aqueous solution. When used for vaccination, the peptide solution is usually mixed with an adjuvant before being administered to the patient.

[0247] Peptide synthesis technology has existed for over 20 years, but it has improved rapidly in recent years, and now, with commercial machines, synthesis can be completed in just a few minutes. For brevity, I will not describe such machines in detail, but their operation will be understood by those skilled in the art, and such conventional machines can be adapted to receive candidate regions or epitopes from a server.

[0248] The server may include the aforementioned functionality for identifying query proteins that are predicted to contain one or more viable BCEs. Naturally, it will be understood that such functionality can be subdivided across different processing entities and different processing modules communicating with each other within a computer network.

[0249] Techniques for identifying BCE-containing proteins can be integrated into a broader ecosystem for customized vaccine development. Exemplary vaccine development ecosystems are well-known and have been described at a high level in the art; however, for brevity, these ecosystems will not be described in detail here.

[0250] In an exemplary ecosystem, the first sampling step might involve isolating DNA from tumor biopsies and matching healthy tissue controls. In the second sequencing step, the data is sequenced and variants, i.e., mutations, are identified. In the immunoprofiler step, the relevant mutant peptides may be generated "in silico".

[0251] Using the relevant mutant peptides and the techniques described herein, protein candidates can be predicted, selected, and target epitopes identified for vaccine design. Specifically, candidate peptide sequences are selected based on their predicted binding affinity determined using the techniques described herein.

[0252] Next, the target epitope is synthesized using the conventional techniques described above. Before administration, the peptide solution is usually mixed with an adjuvant before being administered to the patient (vaccinated). Alternatively, the target epitope can be incorporated into DNA or RNA, or into the genome of bacteria or viruses, by engineering, as in the case of any conventional vaccine.

[0253] The proteins predicted by the methods described herein may also be used to produce other types of vaccines besides peptide-based vaccines. For example, the protein (or the predicted epitope therein) may be encoded in a corresponding DNA or RNA sequence and used to vaccinate a patient. Note that DNA is typically inserted into a plasmid construct. Alternatively, the DNA can be incorporated into the genome of a bacterial or viral delivery system (which may also be RNA, depending on the viral delivery system) that can be used to vaccinate a patient, thus yielding a vaccine produced in a genetically engineered virus or bacterium that produces the target after immunization in the patient, i.e., in vivo.

[0254] A suitable example of a server 1110 is shown in Figure 5. In this example, the server includes at least one microprocessor 1200, memory 1201, an optional input / output device 1202, e.g., a keyboard and / or display, and an external interface 1203, all interconnected via a bus 1204 as shown. In this example, the external interface 1203 can be used to connect the server 1110 to peripheral devices, e.g., a communication network 1140, a reference data store 1120, other storage devices, etc. Although a single external interface 1203 is shown, this is for illustrative purposes only, and in practice, multiple interfaces using various methods (e.g., Ethernet, serial, USB, wireless, etc.) may be provided.

[0255] During use, the microprocessor 1200 enables the execution of necessary processing, including communicating with the reference data store 1120 to receive and process input data, and / or communicating with a client device to receive sequence data of one or more query proteins and generating immunogenicity potential predictions in accordance with the method described above, by executing instructions in the form of application software stored in memory 1201. The application software may include one or more software modules and may be executed in a suitable execution environment, such as an operating system environment.

[0256] Therefore, it will be recognized that server 1200 may be formed from any suitable processing system, such as a suitably programmed client device, PC, web server, network server, etc. In one particular example, server 1200 may be a standard processing system, such as an Intel architecture-based processing system that runs software applications stored on non-volatile (e.g., hard disk) storage, but this is not required. However, it will also be understood that the processing system may be any electronic processing device, such as firmware optionally associated with a logic implementation such as a microprocessor, microchip processor, logic gate configuration, FPGA (Field Programmable Gate Array), or any other electronic device, system, or configuration. Thus, the term “server” is used for illustrative purposes only and is not intended to be limiting.

[0257] Although server 1200 is presented as a single entity, it will be recognized that it can be distributed across multiple geographically dispersed locations, for example, by using a processing system and / or database 1201 provided as part of a cloud-based environment. Therefore, the configuration described above is not mandatory, and other suitable configurations can be used.

[0258] 9. Methodological Case Studies of Mutation Policy Generation Herein, embodiments of the present invention may be used to identify one or more B-cell epitopes that are predicted to induce a protective immunogenic response across multiple pathogens or viruses. For this analysis, unique permutations were randomly generated from previous distributions for each sequence and RBD combination. SARS, SARS-CoV-2, and MERS receptor-binding domains (RBDs) were aligned onto available spike proteins from 304 different betacoronavirus species.

[0259] Figures 29(a)-(c) show successful examples of RBD mapping to 304 betacoronavirus species. Species are divided by subgenus group on the X-axis, and the Y-axis plots the total number of species for which RBD alignment was successful.

[0260] Next, we selected each permutation vector from a given RBD region of each spike protein sequence. For each RBD type (SARS, SARS-CoV-2, MERS), we performed one random generation, using the same permutation for each sequence. From the random permutation generation of these alignments, the following results were revealed: 1. 450,000 permutation vectors were created for the SARS RBD. 2. 450,000 permutation vectors were created for the SARS-CoV-2 RBD. 3. A permutation vector of 450,000 was created for the MERS RBD. These permutations were used in various cases where the corresponding RBDs were aligned.

[0261] 9.1. Statistics of prediction results on all sequences (all RBDs of all types): Total predictions made for each RBD (note that there will be differences in the totals because not all species have all RBDs, and some RBDs will generate fewer permutations): Permutations are generated from prior knowledge about the RBD region.

[0262] [Table 9]

[0263] 9.2. Immunodominance permutation vectors and epitopes are revealed from data analysis. Based on the above data for positive results (confirmed BCE score > 0.9), we were able to capture robust structural motifs or B-cell immunodominant structural epitopes (broad immunogenicity across most virus species tested). These are plotted in Figure 30. Here, the BCE predictor is clearly demonstrated to have the ability to capture latent structural motifs shared across numerous RBD species, i.e., immunodominant structural motifs. In Figure 30, each panel is an RBD. The filtering of conformational epitopes used in this plot was BCE >= 0.9. The y-axis is the total number of unique species, and the x-axis represents a single permutation vector (structural motif). Thus, each column can be considered a single permutation vector shared among betacoronavirus species counted on the Y-axis. The X-axis is sorted in descending order of total unique species. The coloring of the bars is based on subgenus.

[0264] The permutation vectors shown in the columns of Figure 30 are structural representations of epitopes predicted by the BCEP predictor. Next, amino acid presentations were created for these immunodominant (broadly immunogenic) epitopes, and the sequence logos of the top 1 permutations (including ties) were calculated for each RBD. The logos were created by taking the highest-permutation column (permutation) from Figure 30 above, and the sequence logos for those epitopes were created based on a positive BCE prob > 0.9.

[0265] This was created by extracting the highest column (permutation vector) from Figure 30 and generating a sequence logo for its epitope based on a positive BCE prob > 0.9. It was also found that the shared permutation vectors consist of shared repeat amino acid sequences that capture broadly immunogenic B cell epitope amino acid sequences that are dominant across most species.

[0266] 9.3. Realization of mutation policies using a B cell epitope predictor Next, we proceeded to integrate the BCEP predictor into an in silico mutation policy in order to 1) identify more potent BCE epitopes and 2) identify B cell epitopes predicted by the model to be even more immunogenic across more species than those analyzed. We demonstrated the mutation policy method here by mutating the most common motif in each RBD and then carried out the following steps: 1. In silico mutation is performed on each amino acid at each position of the motif, replacing it with any other known acid. 2. For each in silico mutation, the permutation vector (candidate epitope) is re-predicted for all spike proteins from 304 betacoronavirus species using the new mutation motif, and the total number of species that are positive (BCE>0.9) for that mutation motif is counted.

[0267] The essential mutation is designed to be broadly immunogenic against a panel of 304 betacoronavirus species. The original, unmutated counterpart was chosen because it was shared across most species, and the mutation policy, as seen in Figure 31(a), captures an additional 27 species (with respect to predicted immunogenicity) in the most successful mutation, with the highlighted bar 310 representing the mutation being found to be positive in more species compared to the original counterpart 311. When this mutated spike protein is applied under vaccination conditions, it can be inferred that the immunodominant epitope within it will generate an antibody response that may be potentially broadly immunogenic and protective. The mutation policy results illustrated in Figures 31(a) and (b) are for demonstration purposes only and can be applied to the broadly shared permutation vector and all three pathogenic RBD motifs across 304 species.

[0268] 10. Addendum [1] T P Hopp and K R Woods. Prediction of protein antigenic determinants from amino acid sequences. Proceedings of the National Academy of Sciences of the United States of America, 78(6):3824-3828, 06 1981. [2] R. Grantham. Amino acid difference formula to help explain protein evolution. Science, 185(4154):862, 09 1974. [3] JOEL JANIN. Surface and inside volumes in globular proteins. Nature, 277:491 EP--, 02 1979. [4] H R Guy. Amino acid side-chain partition energies and distribution of residues in soluble proteins. Biophys J, 47(1):61-70, Jan 1985. [5] J. K. Mohana Rao and Patrick Argos. A conformational preference parameter to predict helices in integral membrane proteins. Biochimica et Biophysica Acta (BBA) - Protein Structure and Molecular Enzymology, 869(2):197-214, 1986. [6] Sanzo Miyazawa and Robert L. Jernigan. Estimation of effective interresidue contact energies from protein crystal structures: quasi-chemical approximation. Macromolecules, 18(3):534-552, 03 1985. [7] Gang Zhao and Erwin London. An amino acid ”transmembrane tendency” scale that approaches the theoretical limit to accuracy for prediction of transmembrane helices: relationship to biological hydrophobicity. Protein science : a publication of the Protein Society, 15(8):1987-2001, 08 2006. [8] Cyrus Chothia. The nature of the accessible and buried surfaces in proteins. Journal of Molecular Biology, 105(1):1-12, 1976. [9] GD Rose, AR Geselowitz, GJ Lesser, RH Lee, and MH Zehfus. Hydrophobicity of amino acid residues in globular proteins. Science, 229(4716):834, 08 1985.

[10] P Y Chou and G D Fasman. Prediction of the secondary structure of proteins from their amino acid sequence. Adv Enzymol Relat Areas Mol Biol, 47:45-148, 1978.

[11] Shneior Lifson and Christian Sander. Antiparallel and parallel -strands differ in amino acid residue preferences. Nature, 282(5734):109-111, 1979.

[12] M. Cooper and Robert E. Hausman. The cell : a molecular approach, volume 4. Washington, DC ASM Press, 2007.

[13] E A Emini, J V Hughes, D S Perlow, and J Boger. Induction of hepatitis a virus-neutralizing antibody by a virus-specific synthetic peptide. Journal of virology, 55(3):836-839, 09 1985.

[14] G. Deleage and B. Roux. An algorithm for protein secondary structure prediction based on class prediction. Protein Engineering, Design and Selection, 1(4):289-294, 08 1987.

[15] P. MANAVALAN and P. K. PONNUSWAMY. Hydrophobic character of amino acid residues in globular proteins. Nature, 275(5681):673-674, 1978.

[16] Michael Levitt. Conformational preferences of amino acids in globular proteins. Biochemistry, 17(20):4277-4285, 10 1978.

[17] Serafin Fraga. Theoretical prediction of protein antigenic determinants from amino acid sequences. Canadian Journal of Chemistry, 60(20):2606-2610, 2019 / 03 / 01 1982.

[18] Gjalt W. Welling, Wicher J. Weijer, Ruurd van der Zee, and Sytske Welling-Wester. Prediction of sequential antigenic regions in proteins. FEBS Letters, 188(2):215-218, 2019 / 02 / 28 1985.

[19] Henry B. Bull and Keith Breese. Surface tension of amino acid solutions: A hydrophobicity scale of the amino acid residues. Archives of Biochemistry and Biophysics, 161(2):665-670, 1974.

[20] C. A. Browne, H. P. J. Bennett, and S. Solomon. The isolation of peptides by high-performance liquid chromatography using predicted elution positions. Analytical Biochemistry, 124(1):201-208, 1982.

[21] R. Wolfenden, L. Andersson, P. M. Cullis, and C. C. B. Southgate. Affinities of amino acid side chains for solvent water. Biochemistry, 20(4):849-855, 02 1981.

[22] J L Meek. Prediction of peptide retention times in high-pressure liquid chromatography on the basis of amino acid composition. Proceedings of the National Academy of Sciences of the United States of America, 77(3):1632-1636, 03 1980.

[23] R. BHASKARAN and P. K. PONNUSWAMY. Positional flexibilities of amino acid residues in globular proteins. International Journal of Peptide and Protein Research, 32(4):241-255, 2019 / 02 / 28 1988.

[24] Schwartz R.M. Dayhoff, M.O. and B.C. Orcutt. A model of evolutionary change in proteins, volume 5. Atlas of Protein Sequence and Structure. Natl. Biomed. Res. Found., Washington DC, 1978.

[25] David L. (David Lee) Nelson and Michael M. Cox. Lehninger principles of biochemistry. Principles of biochemistry. W.H.Freeman, New York, seventh edition. edition, 2017.

[26] K JWilson, A Honegger, R P Stotzel, and G J Hughes. The behaviour of peptides on reverse-phase supports during high-pressure liquid chromatography. The Biochemical journal, 199(1):31-41, 10 1981.

[27] A. S. Kolaskar and Prasad C. Tongaonkar. A semi-empirical method for prediction of antigenic determinants on protein antigens. FEBS Letters, 276(1-2):172-174, 2018 / 12 / 07 December 10, 1990.

[28] Peter McCaldon and Patrick Argos. Oligopeptide biases in protein sequences and their use in predicting protein coding regions in nucleotide sequences. Proteins: Structure, Function, and Bioinformatics, 4(2):99-122, 2019 / 02 / 28 1988.

[29] J.M. Zimmerman, Naomi Eliezer, and R. Simha. The characterization of amino acid sequences in proteins by statistical methods. Journal of Theoretical Biology, 21(2):170-201, 1968.

[30] Daniel D. Jones. Amino acid properties and side-chain orientation in proteins: A cross correlation approach, volume 50. 04 1975.

[31] D. Eisenberg, E. Schwarz, M. Komaromy, and R.Wall. Analysis of membrane and surface protein sequences with the hydrophobic moment plot. Journal of Molecular Biology, 179(1):125-142, 1984.

[32] J. M. R. Parker, D. Guo, and R. S. Hodges. New hydrophilicity scale derived from high-performance liquid chromatography peptide retention data: correlation of predicted surface residues with antigenicity and x-ray-derived accessible sites. Biochemistry, 25(19):5425-5432, 09 1986.

[33] Shaun D. Black and Diane R. Mould. Development of hydrophobicity parameters to analyze proteins which bear post- or cotranslational modifications. Analytical Biochemistry, 193(1):72-82, 1991.

[34] J L. Fauchere and V Pliska. Hydrophobic parameters II of amino acid side-chains from the partitioning of N-acetyl-amino acid amides, volume 18. 01 1983.

[35] Robert M. Sweet and David Eisenberg. Correlation of sequence hydrophobicities measures similarity in three-dimensional protein structure. Journal of Molecular Biology, 171(4):479-488, 1983.

[36] Jack Kyte and Russell F. Doolittle. A simple method for displaying the hydropathic character of a protein. Journal of Molecular Biology, 157(1):105-132, 1982.

[37] Charles. Tanford. Contribution of hydrophobic interactions to the stability of the globular conformation of proteins. Journal of the American Chemical Society, 84(22):4240-4247, 11 1962.

[38] Donald J. Abraham and Albert J. Leo. Extension of the fragment method to calculate amino acid zwitterion and side chain partition coefficients. Proteins: Structure, Function, and Bioinformatics, 2(2):130-152, 2019 / 02 / 28 1987.

[39] Richard Cowan and R GWhittaker. Hydrophobicity indices for amino acid residues as determined by HPLC, volume 3. 01 1990.

[40] Mark A. Roseman. Hydrophilicity of polar amino acid side-chains is markedly reduced by flanking peptide bonds. Journal of Molecular Biology, 200(3):513-522, 1988.

[41] A. A. Aboderin. An empirical hydrophobicity scale for alpha-amino-acids and some of its applications. International Journal of Biochemistry, 2:537-544, 1971.

Claims

1. A computer implementation method for predicting whether a protein contains a B cell epitope that is likely to trigger an antibody binding event, wherein the method is: (a) Accessing the amino acid sequence of the protein; (b) Accessing one or more structural and / or surface properties of the protein; (c) To generate one or more candidate codes for each of the one or more candidate B cell epitopes on the protein, wherein each candidate code represents each of the candidate B cell epitopes as a plurality of data elements corresponding to amino acids in the amino acid sequence; (d) Using a first trained machine learning model, predict whether any of the candidate B cell epitopes are likely to be true B cell epitopes on the protein, wherein the input to the first trained machine learning model includes the amino acid sequence of the protein, one or more structures and / or surface properties of the protein, and one or more candidate codes. Methods that include...

2. The method according to claim 1, wherein the output of the trained first machine learning model is the probability that each of the candidate codes represents a true B cell epitope on the protein.

3. The method according to claim 1 or 2, wherein the one or more candidate codes represent all possible combinations of amino acids in the amino acid sequence.

4. The method according to any one of claims 1 to 3, wherein the first machine learning model includes a long-short-term memory network (LSTM), preferably a bidirectional long-short-term memory network (BLSTM).

5. The method according to any one of claims 1 to 4, wherein the one or more structures and / or surface properties are the secondary structure of the protein; relative solvent accessibility RSA; and hemispheric exposure HSA.

6. The first machine learning model described above, (i) Multiple amino acid sequences of each first reference protein, each first reference protein comprising multiple amino acid sequences containing at least one true B cell epitope, (ii) A plurality of reference codes for each true B cell epitope, each reference code representing the respective true B cell epitope as a plurality of data elements corresponding to amino acids in the amino acid sequence; (iii) One or more structural and / or surface properties of each first reference protein and The method according to any one of claims 1 to 5, which is trained using a first reference dataset including .

7. The method according to claim 6, wherein the one or more structural and / or surface properties of each first reference protein are the one or more structural and / or surface properties of each first reference protein in an unbound state.

8. The one or more structural and / or surface properties of each first reference protein are predicted by applying one or more trained second machine learning models to the corresponding amino acid sequences. The method according to claim 6 or 7, wherein the one or more trained second machine learning models are trained on a second reference dataset comprising multiple amino acid sequences of each protein and their corresponding structures and / or surface properties.

9. One or more structural and / or surface properties of the protein accessed in step (b) are predicted by applying one or more trained second machine learning models to the corresponding amino acid sequences. The method according to any one of claims 1 to 8, wherein the one or more trained second machine learning models are trained on a second reference dataset comprising multiple amino acid sequences of each protein and their corresponding structures and / or surface properties.

10. The method according to claim 8 or 9, wherein one or more second machine learning models are trained on a second reference dataset comprising multiple amino acid sequences of each second reference protein and their corresponding structures and / or surface properties in an unbound state.

11. At least some of the first reference proteins are: (i) Accessing multiple protein complexes containing at least three different protein chains; (ii) Filter the plurality of protein complexes to obtain the effective weight of the antibody (V Ｈ ) Protein chain mapped as a chain, effective light (V) of the antibody Ｌ ) Only those containing protein chains mapped as chains, and protein chains mapped as antigens, should be retained; (iii) The above V Ｈ Chain and V Ｌ Pairing the chains, V Ｈ and V Ｌ To form a chain pair; (iv) Each V Ｈ and V Ｌ With respect to the chain pair, define the true B cell epitope on the corresponding antigen; Obtained by, The method according to any one of claims 7 to 10, wherein at least some of the first reference proteins in the first reference dataset correspond to an antigen mapped in step (ii) that comprises at least one true B cell epitope defined in step (iv).

12. The method according to any one of claims 7 to 11, further comprising generating pseudorandom data based on a plurality of reference codes for each true B cell epitope in the first reference dataset, wherein the first reference dataset includes the pseudorandom data as negative data.

13. The method according to any one of claims 1 to 12, wherein the B cell epitope is a conformational B cell epitope.

14. The method according to any one of claims 1 to 13, wherein the protein comprises an antigen.

15. To identify a first B cell epitope that is predicted to be highly likely to be a true B cell epitope on the aforementioned protein; Substituting one or more amino acids that form the identified first B cell epitope with different amino acids, thereby generating a modified amino acid sequence of the protein; Using the trained first machine learning model, predict whether the candidate code representing the identified first B cell epitope is likely to correspond to the true B cell epitope on the modified amino acid sequence of the protein. The method according to any one of claims 1 to 14, further comprising:

16. A method for training machine learning models, The process involves generating a reference dataset, wherein the reference dataset is: Each reference protein has multiple amino acid sequences, each of which contains at least one true B cell epitope; A plurality of reference codes for the true B cell epitopes, each of which represents the respective true B cell epitope as a plurality of data elements corresponding to the amino acids of the amino acid sequence; One or more structural and / or surface properties of each reference protein; Including generating; The machine learning model is trained using the aforementioned reference dataset to learn the relationships between the candidate codes, the structure and / or surface properties, and the corresponding true B cell epitopes. Methods that include...

17. The method according to claim 16, wherein one or more structural and / or surface properties of each reference protein are one or more structural and / or surface properties of each first reference protein in an unbound state.

18. A method for producing a vaccine, Predicting that a protein contains a B cell epitope by the method according to any one of claims 1 to 15; Synthesize the protein and / or the predicted B cell epitope, or encode at least one of the protein, the predicted B cell epitope, a B cell epitope variant, or a predicted or simulated variant of a B cell epitope into a corresponding protein, peptide, DNA, or RNA sequence. A method that includes this.

19. A method for developing a diagnostic assay to determine whether a patient has or has had a history of cancer or a pathogen infection, the diagnostic assay being performed on a biological sample obtained from a subject and comprising identifying at least one protein of a pathogen or tumor predicted to contain a B cell epitope using the method according to any one of claims 1 to 15; The diagnostic assay is a method comprising the use or identification of the identified at least one protein and / or B cell epitope within the biological sample.

20. A system for predicting whether a protein contains a B cell epitope that is likely to trigger an antibody binding event, comprising at least one processor communicating with at least one memory device, wherein the at least one memory device stores instructions for causing the at least one processor to perform the method according to any one of claims 1 to 14.

21. A computer program product that, when executed by a computer, includes an instruction causing the computer to perform the method according to any one of claims 1 to 20.

22. Use of a machine learning model trained using the method described in claim 16 or 17.

23. A machine learning model trained using the method described in claim 16 or 17.

24. A method for synthesizing a protein, wherein the method according to any one of claims 1 to 15 predicts that the protein contains a B cell epitope that is likely to trigger an antibody binding event; To synthesize the aforementioned protein and A method that includes this.

25. A protein synthesized using the method of claim 24.

26. A method for identifying one or more B cell epitopes that are predicted to induce a protective immunogenic response across multiple species of pathogens or viruses, (i) performing the method according to any one of claims 1 to 14 on each of a plurality of proteins of a plurality of different species of pathogen or virus, thereby identifying a plurality of first B cell epitopes on at least one of the plurality of proteins that are each predicted to be likely to be true B cell epitopes; (ii) For each identified first B cell epitope, determine the number of different species of pathogens or viruses for which the first B cell epitope is predicted to be a true B cell epitope; (iii) Classifying one or more of the first B cell epitopes as broad-spectrum protective B cell epitopes based on the number of different species or the number of variants of any one given species in which the first B cell epitope is predicted to be a true B cell epitope. A method that includes this.

27. For at least one of the first B cell epitopes classified as a broad-area protective B cell epitope, The first B cell epitope is formed by substituting one or more amino acids with different amino acids, thereby generating a modified first B cell epitope; For each of the multiple proteins of each of the multiple species of viral pathogens, To generate the modified amino acid sequence of the protein according to the modified first B cell epitope; Using the trained first machine learning model, predict whether a candidate code representing the first B cell epitope, which has been classified as a broad-spectrum protective B cell epitope, is likely to correspond to a true B cell epitope on the modified amino acid sequence of the protein. The method according to claim 26, further comprising:

Images