Genetic engineering of antigen-binding proteins

Abstract

Methods for identifying antigen binding proteins that contain chain pairs are described. The method includes providing a query sequence including a first chain sequence, and identifying a corresponding chain sequence by providing the query sequence to a deep learning model configured to take the query first chain sequence as input and generate at least one corresponding chain sequence as output, thereby identifying the corresponding chain sequence of the query sequence, the deep learning model being trained using training first chain sequences and corresponding chain sequences from known chain pairs. The first chain sequence may be a heavy / light chain of an antibody or B cell receptor, or a β / α / δ / γ chain of a T cell receptor, and the corresponding chain may be a light / heavy chain of an antibody or B cell receptor, or a β / α / δ / γ chain of a T cell receptor. The method is utilized in any situation in which it is desirable to identify chain pairs of antigen-binding molecules, such as, for example, in situations in which antigen-binding molecules having desired (e.g., therapeutic or functional) properties are identified. Related methods, systems, and products are also described.

Description

[Technical field]

[0001] FIELD OF THEINVENTION The present invention relates to methods of engineering antigen binding proteins such as B cell receptors, antibodies and T cell receptors by identifying variable chain pairs from a single input variable chain, such as heavy-light chain pairs from an input heavy or light chain or α-β chain pairs from an input α or β chain.The present invention also relates to methods of providing antigen binding proteins such as therapeutic antibodies derived from input variable chains, e.g. B cell receptor / antibody heavy or light chains. [Background technology]

[0002] 2. Background of the Invention Effective humoral immunity requires a large variety of B cells capable of binding different antigens through the B cell receptor (BCR). The theoretical overall size of the BCR repertoire in humans is up to about 10 15 It is estimated that there are about 10 variants of 9 At any given time, there are 100,000 BCRs circulating within an individual [Rees, 2020]. The BCR is composed of two pairs of two proteins: two heavy chains and two light chains. Each B cell expresses a pair of heavy and light chains (which are likely unique) to form the BCR, which is expressed on the surface of the B cell or secreted as an antibody. Over 600 million different human heavy chain sequences and approximately 70 million light chain sequences are currently cataloged in the Observed Antibody Space [Kovaltsuk et al., 2018]. Characterization of an individual's BCR ensemble (also called the individual's BCR repertoire) has proven to be a valuable tool for understanding the biology of various diseases [Vander Heiden et al., 2017; Bashford-Rogers et al., 2019; Nielsen et al., 2020; Simonich et al., 2019] and discovering novel therapeutic antibody drugs [Krawczyk et al., 2019; Galson et al., 2020].

[0003] There are two main approaches to characterize an individual's BCR repertoire: single B cell sequencing and sequencing of bulk B cell populations. Single cell sequencing is more commonly adopted for antibody discovery applications because it preserves the pairing information between heavy and light chains. However, single cell sequencing has limited throughput and different platforms and protocols provide variable coverage of the BCR repertoire present within a single sample. Even the most advanced microfluidic systems typically only capture approximately 10 B cells per sample. 4 Only a small number of B cell sequences can be recovered [King et al., 2021;Eccles et al., 2020;Setliff et al., 2019]. Humans typically produce approximately 10 B cell sequences per milliliter of blood. 6 B cells [Mora and Walczak, 2019], meaning that single-cell approaches cannot characterize the complete B cell diversity even for small samples. Furthermore, single-cell sequencing has very specific sample requirements (e.g., cells typically need to remain viable until processed, thus requiring fresh samples to be processed on the day of collection or samples frozen according to specific protocols), has a very high cost per sample compared to bulk sequencing (single-cell sequencing is at least an order of magnitude more expensive than bulk sequencing), and requires dedicated laboratory equipment.

[0004] Sequencing of bulk B cell populations is performed at approximately 10 per sample. 7B cell sequences can be more easily recovered [Briney et al., 2019], which is much closer to the diversity expected in an individual. However, because B cells are lysed during library preparation, heavy-light chain pairing information is not preserved. Typically, these bulk BCR sequencing methods focus only on heavy chains, since heavy chains play a dominant role in antigen binding and are much more diverse than the light chain repertoire [Kovaltsuk et al., 2018]. However, for antibody discovery, it is necessary to have both the heavy and light chains of an antibody to synthesize and characterize its function. Gaps in light chain pairing information have prompted the development of computational pairing methods [Reddy et al., 2020, Zhu et al., 2013, Raybould et al., 2021, Rakocevic et al., 2021]. However, these are limited to specific datasets and a small number of specific sequences within these datasets.

[0005] Similarly, cell-mediated immunity requires a large variety of T cells capable of binding different antigens through their T cell receptors (TCRs). The total size of the TCR repertoire in humans is up to about 10 15 It has been estimated that the TCR repertoire is composed of 1 unique αβ T cell receptor (TCR) pair [Carter et al., 2019]. Experimental methods for paired αβ TCR sequencing have been developed (including single-cell methods [Zheng et al., 2017] and multi-cell deconvolution-based methods [Howie et al., 2015]), but these remain dedicated and have limited throughput. Thus, most of the available TCR repertoire knowledge is based on bulk sequencing of single-chain repertoires, mostly β-chain repertoires. This is inherently limited, especially since both α and β TCR chains have been shown to be involved in alloreactivity and antigen specificity [Carter et al., 2019]. Summary of the Invention [Problem to be solved by the invention]

[0006] Thus, there remains a need for improved methods to identify chain pairs, such as BCR heavy-light chain pairs or TCR αβ chain pairs, from data that do not contain this pairing information. [Means for solving the problem]

[0007] Summary of the Invention The problem of identifying BCR heavy-light chain pairs is far from trivial. Indeed, the diversity of the BCR repertoire makes the search space large. Moreover, although some heavy-light chain combinations result in stable BCRs (observations that led to speculation that pairing may be random [Glanville et al., 2009; Jayaram et al., 2012; DeKosky et al., 2016]), only a limited number of pairings generate functional BCRs capable of binding to target antigens [Teplyakov et al., 2016; Ling et al., 2018]. This indicates that functional pairings are nonrandom and that the determination of functional pairings is obscured by the number of pairings that may be nonfunctional but nevertheless stable. Indeed, even when stable pairs can be predicted, this means that finding the right light chain is difficult, as these generate a significant number of solutions that require experimental validation and are expected to validate poorly when selected primarily on the basis of stability.

[0008] Several different computational approaches have been suggested, each of which has some major drawbacks. The first approach was based on matching the relative frequency of BCR heavy and light chains when sequenced independently [Reddy et al., 2010]. In this study, mice were first immunized to generate a strong immune response, and then the top 4-5 most frequent heavy and light chains were selected for pairing. For these top 4-5 sequences, pairing based on relative frequency was not possible. More recently, Rakocevic et al.

[2021] showed that this approach only works when the sample is dominated by a small number of high-frequency B cells. Zhu et al.

[2013] proposed a method called phylogenetic pairing, which involves comparing the architecture of phylogenetic trees generated from heavy and light chain sequence data. This method is limited to investigating a specific clonal expansion; in this case, known antiviral antibody lineages, rather than the entire BCR repertoire. Raybould et al.

[2021] proposed a method based on pairing heavy and light chain structural models in silico. The method is inherently limited by the limited and highly skewed availability of high-quality structural templates, and at best can only identify features associated with stability that do not necessarily translate to functionality. Furthermore, the method can only pair families of similar sequences, not specific sequences (limiting practical applicability - not experimentally validated). Thus, the inventors have identified that current methods of computational heavy-light chain pairing are limited in that they only apply to specific datasets and sequences within these datasets. Indeed, existing validated methods are only applicable to datasets where both heavy and light chain sequences are available from the sample and the data are largely clonal expansion dominated, and only facilitate pairing of a limited number of sequences within these datasets.

[0009] The inventors further identified that in generalized applications to antibody discovery, it would be desirable to be able to generate a viable light chain for any given heavy chain. It would be even more desirable to be able to generate this using only heavy chain information, as BCR repertoire bulk sequencing efforts often focus limited resources on sequencing heavy chains, which are believed to play a more important functional role than light chains. To address these issues, the inventors assumed that deep learning methods inspired by recent advances in natural language processing (NLP) could be used. In particular, the problem of identifying native light chain sequences was expressed as a neural machine translation (NMT) problem using the Transformer [Vaswani et al., 2017] and other alternative deep learning architectures. The Transformer has shown state-of-the-art results on a wide range of NLP tasks [Vaswani et al., 2017;Devlin et al., 2019;Liu et al., 2019;Rothe et al., 2020]. Therefore, we devised a method that uses a transformer model (called "Matchmaker") to generate light chains using only BCR heavy chains as input. In many blind tests on single-cell datasets with known pairings, they showed that Matchmaker provides more accurate light chain gene annotations than the other four methods. The predictions were then validated on a set of nine therapeutic antibodies with known targets and 18 heavy chains from COVID-19 patients commonly found in affected individuals. All Matchmaker-generated antibodies were successfully expressed and stable; in addition, three of the nine predictions for therapeutic antibodies and nine of the 18 predictions for COVID-19 antibodies showed signs of binding to the target antigen in vitro. We further showed that this approach can be extended by using tandem transformers to "learn the language" of single-chain sequences in a pre-training step. This approach provides a novel solution to light chain pairing and a path to fill the gap in bulk heavy chain sequencing. The inventors have further determined that the same approach can be used to solve the αβ TCR chain pairing problem.

[0010] Thus, according to a first aspect, there is provided a method of identifying an antigen binding protein comprising a chain pair, the method comprising providing a query sequence comprising a first chain sequence, and identifying a corresponding chain sequence by providing the query sequence to a deep learning model configured to take the query first chain sequence as an input and generate at least one corresponding chain sequence as an output, thereby identifying the corresponding chain sequence of the query sequence, the deep learning model having been trained using training first chain sequences and corresponding chain sequences from known chain pairs. The method may have one or more of the following features.

[0011] The chain pair may be referred to as "variable chains". The phrase "known chain pair" refers to a pair of variable chain sequences known to be present in an antigen binding protein that exhibits a desired antigen binding function or that forms part of the B-cell or T-cell repertoire of at least one subject. The latter may also be referred to as a "native" chain pair.

[0012] The antigen binding protein may comprise a heavy-light chain pair, the first chain sequence being a heavy or light chain sequence and the counterpart chain sequence being a light or heavy chain sequence. The first chain sequence may be a heavy chain sequence and the counterpart sequence may be a light chain sequence. The antigen binding protein may be a B cell receptor or an antibody or a protein derived therefrom. Thus, the antigen binding protein may comprise a heavy-light chain pair. The query sequence may comprise a heavy or light chain sequence. The counterpart chain sequence may be a light or heavy chain sequence.

[0013] The antigen binding protein may comprise an αβ chain pair, the first chain sequence being a β chain sequence or an α chain sequence, and the corresponding chain sequence being an α chain sequence or a β chain sequence. The first chain sequence may be a β chain sequence, and the corresponding sequence may be an α chain sequence. The antigen binding protein may comprise a γδ chain pair, the first chain sequence being a δ chain sequence or a γ chain sequence, and the corresponding chain sequence being a γ chain sequence or a δ chain sequence. The first chain sequence may be a δ chain sequence, and the corresponding sequence may be a γ chain sequence. The antigen binding protein may be a T cell receptor or a protein derived therefrom. Thus, the antigen binding protein may comprise an αβ chain pair or a γδ chain pair. Thus, the query sequence may comprise a β or δ chain sequence, or an α or γ chain sequence. The corresponding chain sequence may be an α or γ chain sequence, or a β or δ chain sequence.

[0014] The deep learning model may be an array-to-array model. The deep learning model may be a recurrent neural network or a transformer. The deep learning model may be an array-to-array transformer-based model. The recurrent neural network may be a gated recurrent unit (GRU)-based model or a long short-term memory (LSTM) model. For example, the GRU-based model may include a GRU-based encoder and a GRU-based decoder. The encoder may be, for example, a 4-layer bidirectional GRU with a hidden dimension of 1024. The decoder may be, for example, a 4-layer forward-only GRU with a hidden dimension of 1024. The transformer is a deep learning model that uses an attention mechanism. The transformer-based model may be a transformer model with an architecture that uses both self-attention and pointwise fully connected layer encoder and decoder. The encoder and / or decoder may be composed of a stack of four identical layers. Each layer of the encoder may have two sublayers: a multi-head self-attention layer and a positionwise fully connected feedforward network layer. Each layer of the decoder may have three sublayers: a self-attention sublayer, a layer that performs multi-head attention on the output of the encoder stack, and a feedforward network layer. The model may have a feedforward dimension of 1024.

[0015] The deep learning model may be configured to generate one or more corresponding strand sequences as output. Each corresponding strand sequence may be associated with a confidence metric, such as a probability. The corresponding strand sequence of the query sequence may be identified as the sequence of the corresponding strand sequence or sequences associated with the highest confidence metric among the corresponding strand sequences. The deep learning model may be configured to generate a single corresponding strand sequence as output. The deep learning model may be configured to predict each strand sequentially. In other words, the deep learning model may be configured to provide predictions in a greedy manner. The deep learning model may be configured to predict the strands using a beam search technique or related techniques such as beam stack search [Zhou & Hansen, 2005] and depth-first beam search [Furcy & Koenig, 2005]. The deep learning model may be configured to generate multiple corresponding strand sequences as output. Each of the multiple corresponding strand sequences may be associated with a confidence metric, such as a probability. The single corresponding strand sequence associated with the highest confidence metric may be reported. Thus, the method may include identifying the corresponding strand sequence as one of a plurality of corresponding strand sequences associated with the highest confidence metric among one or more corresponding strand sequences. Alternatively, all predicted corresponding strand sequences may be reported, advantageously with associated confidence metrics. Alternatively, all predicted corresponding strand sequences that meet one or more additional criteria may be reported. For example, any corresponding strand sequence that is predicted and associated with a confidence metric within a predetermined range from the corresponding strand sequence associated with the highest confidence metric may be reported.

[0016] The training first chain sequence and the corresponding chain sequence from the known chain pair may include paired training heavy and light chain sequences from single B cell sequencing data. The training data may include one or more data sets previously obtained by single B cell sequencing of a sample obtained from a subject or sequencing of a library derived therefrom. The training data may further include paired training heavy and light chain sequences from a known antibody / B cell repertoire. For example, the training data may include paired training heavy and light chain sequences from one or more antibody / BCR databases, from one or more known therapeutic antibodies / BCRs, and / or from one or more antibodies / BCRs known to have a desired binding function.

[0017] The training data may include paired training heavy and light chain sequences from a naive B cell receptor library. The training data may include paired training heavy and light chain sequences from an antigen-experienced B cell receptor library. Thus, the training data may include paired training heavy and light chain sequences obtained from subjects that have been exposed to one or more particular antigens.

[0018] The training first chain sequences and the corresponding chain sequences from the known chain pairs may include paired training α and β chain sequences from single T cell sequencing data. The training data may include one or more data sets previously obtained by single T cell sequencing of a sample obtained from a subject or sequencing of a library derived therefrom, respectively. The training data may further include paired training first chain sequences and the corresponding chain sequences from a known T cell repertoire. For example, the training data may include paired training α and β chain sequences from one or more T cell receptor databases, from one or more known therapeutic TCRs, and / or from one or more TCRs known to have a desired binding function. The training data may include paired training α and β chain sequences (or δ and γ) from a naive T cell receptor library. The training data may include paired training α and β chain sequences (or δ and γ chain sequences) from an antigen-experienced T cell receptor library. Thus, the training data may include paired training alpha and beta chain sequences (or delta and gamma chain sequences) obtained from subjects who have been exposed to one or more particular antigens.

[0019] The training first and corresponding chain sequences from known chain pairs may comprise paired training chain sequences, each pair comprising a chain sequence comprising or consisting of a V gene sequence or identifier, a J gene sequence or identifier, and a junction sequence, and optionally a D gene sequence or identifier. The training first and corresponding chain sequences from known chain pairs may comprise paired training chain sequences, each pair comprising a chain sequence comprising or consisting of a V gene sequence or identifier, a J gene sequence or identifier, and a junction sequence. The training data may comprise at least 80,000, at least 100,000, at least 120,000, or at least 150,000 pairs of training sequences, e.g., training heavy chain sequences and computational sequences. Advantageously, the training data may comprise at least 150,000 pairs of training heavy and light chain sequences. The training data may comprise chain sequence pairs from a mammal, e.g., a human. The training data may comprise heavy and / or light chain sequences from a mammal. The training data may include human heavy and / or light chain sequences. The training data may include training sequence pairs from the same species as the query sequence. The training data may include at least 50%, at least 70%, at least 80%, at least 90%, or at least 95% sequences from the same species as the query sequence. The query sequence may be a sequence not present in the training data. The query sequence may be a sequence obtained from a sample from a subject having a desired characteristic, such as a desired phenotype. For example, the subject may have a particular clinical characteristic.

[0020] The training data may further include unpaired training first and / or counterpart sequences, as further described below. The unpaired training first and / or counterpart sequences may have any of the sequence characteristics described in connection with the paired sequences. In particular, the unpaired strand sequences may be of the same type of sequence as the paired sequences (e.g., if the paired training sequences are heavy and light chain pairs, the unpaired training first / counter strand sequences may include unpaired heavy and / or light chains), may include sequences from the same organism (e.g., may include mammalian sequences and / or human sequences, may include sequences from one or more organisms, may include sequences from naive libraries and / or challenge libraries, etc.), and may include the same information (e.g., gene segment identifiers, sequences, and combinations thereof, etc.). The unpaired training sequences may include some or all of the first and / or counterpart sequences present in the paired training sequences. Advantageously, the unpaired training sequences may include more first strand sequences and / or more counter strand sequences than the paired training strand sequences. The use of additional unpaired training sequences may be particularly advantageous in embodiments where the deep learning model is a transformer-based model that includes an encoder pre-trained using unpaired training first sequences and / or light counter sequences, and a decoder pre-trained using unpaired training counter strand sequences and / or first strand sequences.

[0021] The query chain sequence may comprise or consist of a V gene sequence or identifier, a J gene sequence or identifier, and a junction sequence, and optionally a D gene sequence or identifier. The corresponding chain sequence may comprise or consist of a V gene sequence or identifier, a J gene sequence or identifier, and a junction sequence. The query chain sequence may comprise or consist of a V gene sequence or identifier, a J gene sequence or identifier, and a junction sequence. The corresponding chain sequence comprises or consists of a V gene sequence or identifier, a J gene sequence or identifier, and a junction sequence, and optionally a D gene sequence or identifier. The format of the query chain sequence and the corresponding chain sequence are related to the format of the training chain sequence. Thus, a deep learning model trained using training chain sequences comprising or consisting of a V gene sequence or identifier, a J gene sequence or identifier, and a junction sequence, and optionally a D gene sequence or identifier, may accept as input or generate as output chain sequences comprising or consisting of these components. Similarly, a deep learning model trained using training strand sequences comprising or consisting of V gene sequences or identifiers, J gene sequences or identifiers, and junction sequences may accept as input or generate as output strand sequences comprising or consisting of these components. The query sequence may comprise or consist of one or more first strand CDR sequences. The corresponding sequence may comprise or consist of one or more corresponding strand CDR sequences, and optionally the query / corresponding sequence comprises or consists of a CDR3 sequence.

[0022] All sequences may be amino acid sequences. Providing the query sequence to the deep learning model may include encoding the query sequence using an encoding scheme, where each gene sequence identifier corresponds to an individual token. Providing the query sequence to the deep learning model may include encoding the query sequence using an encoding scheme, where each amino acid corresponds to an individual token. Providing the query sequence to the deep learning model may include encoding the query sequence using an encoding scheme, where the sequence (i.e., the sequence available as a complete sequence rather than a gene identifier) ​​is encoded using tokens each corresponding to an individual k-mer or using a byte pair encoding. Each sequence may be encoded using overlapping k-mers. The k-mers may be of length 1-5. The k-mers may be of a fixed length. For example, a fixed k-mer length of 1, 2, 3, 4, or 5 may be used. A k-mer of length 1 is equivalent to encoding each character (e.g., each amino acid) individually. A k-mer of length k>2 (e.g., 3, etc.) may be used as part of an encoding scheme using overlapping or non-overlapping k-mers. Overlapping k-mers may overlap to different degrees. For example, k-mers of length 3 may overlap by 1 or 2 characters. In a scheme using k=3, each token corresponds to a unique set of 3 characters (e.g., a motif of 3 amino acids). Identifying the corresponding strand sequences may include decoding the corresponding sequences output by the deep learning model using an encoding scheme in which each gene sequence identifier corresponds to an individual token. Identifying the corresponding strand sequences may include decoding the corresponding sequences output by the deep learning model using an encoding scheme in which each amino acid corresponds to an individual token. Identifying the corresponding strand sequences may include decoding the corresponding sequences output by the deep learning model using an encoding scheme in which the sequences are encoded using tokens that each correspond to an individual k-mer or using a byte pair encoding. Each sequence may be encoded using overlapping k-mers. The encoding scheme may be previously defined based on the content of the training strand sequences.The encoding scheme may have been previously defined based on the content of the training strand sequences, and a token is excluded from the vocabulary constructed based on the content of the training strand sequences if it is used (separately or collectively in the first strand and corresponding strand sequences in the paired training data) less than a predetermined threshold number of times in the training data (e.g., 2). The encoding scheme may have been previously defined based on the content of the training strand sequences by constructing training first strand and training corresponding strand vocabularies separately in the training data.

[0023] The training data may be filtered to exclude any pairs that include a junction sequence (in the first and / or corresponding strand) outside a predetermined range of lengths. In other words, the training data may not include any pairs that include a first (e.g., heavy) strand junction outside a predetermined range of lengths and / or a corresponding (e.g., light) strand junction outside a predetermined range of lengths. For example, pairs that include heavy chain junction sequences below a predetermined length, such as 3, 4, 5, 6, 7, 8, 9, or 10 amino acids, may be excluded. As another example, pairs that include heavy chain junction sequences above a predetermined length, such as 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, or 35 amino acids, may be excluded. As another example, pairs that include light chain junction sequences below a predetermined length, such as 3, 4, 5, 6, 7, 8, 9, or 10 amino acids, may be excluded. As another example, pairs containing light chain junction sequences exceeding a predetermined length, such as, for example, 15, 16, 17, 18, 19, 20, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, or 30 amino acids, may be excluded. The predetermined length may be the same or different for the junction sequences in the corresponding (e.g., light) and first (e.g., heavy) chains of the pair. In a specific example, pairs containing heavy chain junction sequences fewer than 7 amino acids may be excluded and / or pairs containing heavy chain junction sequences more than 30 amino acids may be excluded. Alternatively or additionally, pairs containing light chain junction sequences fewer than 7 amino acids may be excluded and / or pairs containing light chain junction sequences more than 20 amino acids may be excluded.

[0024] The query sequence and / or the corresponding sequence may comprise one or more genetic sequence identifiers, and the method may further comprise replacing the one or more genetic sequence identifiers with the corresponding germline sequence.

[0025] The deep learning model may be a transformer-based model including an encoder pre-trained using unpaired training first strand sequence and / or corresponding strand sequence, and a decoder pre-trained using unpaired training corresponding strand sequence and / or first strand sequence. The encoder and / or decoder may include an autoregressive transformer model such as a BERT model, or a variant thereof, such as, for example, BERT, RoBERTa, or DistilBERT, and / or GPT-2. The encoder and / or decoder may include a RoBERTa model, a BERT model, and / or a GPT-2 model. The encoder and decoder may include the same model. The encoder and decoder may both include models trained using unpaired training corresponding strand sequence and first strand sequence. For example, the encoder and decoder may both include models trained using random pairs each including the first strand sequence and the corresponding strand sequence. Alternatively, the encoder may include a model trained using a trained first (e.g., heavy or light) strand sequence, and the decoder may include a model trained using a corresponding (e.g., light or heavy) strand sequence. When the encoder and decoder both include models trained using unpaired training first strand sequences and corresponding strand sequences, the encoder and decoder may both include the same pre-trained model. Thus, the encoder and decoder may be initialized using a pre-trained model having the same architecture with the same parameters. The unpaired training strand sequences may include full-length sequences in the variable region of the corresponding strand. The unpaired training corresponding strand sequences may include full-length sequences in the variable region of the first strand. The transformer-based model may have been trained using paired first strand sequences and corresponding (e.g., heavy and light) strand sequences from a known strand pair, which do not include full-length sequences in the variable region of the corresponding strand and / or the first strand.In such embodiments, the transformer-based model may be trained by obtaining paired training sequences that include full-length sequences in the variable regions of the corresponding strand and / or the first strand by inferring missing sequence information.

[0026] Inputting the missing sequence information may include replacing the gene identifier with the corresponding germline sequence. Inputting the missing sequence information may include predicting the full-length sequence from each of the paired training first (e.g., heavy) chain sequences and / or corresponding (e.g., light) chain sequences using a pre-trained encoder and / or a pre-trained decoder (e.g., depending on which chain the missing sequence information is predicted for). Alternatively, the unpaired training corresponding (e.g., light) chain sequences and / or the unpaired training first (e.g., heavy) chain sequences may have been converted to a format that matches the format of each paired training sequence prior to training the encoder and / or decoder.

[0027] Providing the query sequence may include obtaining the query sequence from a user through a user interface, from a computing device, from a sequence obtaining means or a computing device associated with the sequence obtaining means, from a database or other computer readable medium. Providing the query sequence may include sequencing a sample including genetic material encoding an antigen binding molecule including the query sequence. Obtaining the query sequence may include performing B cell bulk sequencing of a sample including B cells, T cell bulk sequencing of a sample including T cells, or bulk sequencing of a sample including any other cells expressing an antigen binding molecule including the query sequence or genetic material derived therefrom, such as a B cell receptor library or a T cell receptor library. Providing the query sequence may include obtaining a sample including B cells, T cells, other cells expressing an antigen binding molecule including the query sequence, or genetic material derived therefrom, such as a B cell receptor library or a T cell receptor library. Providing the query sequence may include sequencing a sample containing genetic material encoding an antigen binding molecule comprising the query sequence, for example by performing B cell bulk sequencing of a sample containing B cells (or any other cells expressing an antigen binding molecule comprising the query sequence or genetic material derived therefrom, such as a B cell receptor library). Providing the query sequence may include obtaining a sample containing B cells or other cells expressing an antigen binding molecule comprising the query sequence or genetic material derived therefrom, such as a B cell receptor library.

[0028] The method may further include providing the identified corresponding sequences, portions thereof, or information derived therefrom to a user through a user interface.

[0029] According to a second aspect, there is provided a method of providing antigen binding protein chain pairs for a plurality of query sequences comprising a first chain sequence, the method comprising performing the method of any embodiment of the first aspect for each query sequence. The plurality of query sequences may be heavy or light chain sequences obtained by bulk B cell repertoire sequencing. The plurality of query sequences may comprise at least 100, at least 1000, at least 10,000 or at least 100,000 sequences. The plurality of query sequences may have been obtained by bulk B cell sequencing of a heavy or light chain repertoire in a sample, such as a sample from a subject. The plurality of sequences may be a subset of the set of sequences obtained by bulk B cell sequencing of a heavy or light chain repertoire in a sample. This method according to this aspect of the invention may have any feature as described in relation to the first aspect.

[0030] According to a third aspect, there is provided a method of providing an antigen binding protein with desired properties, the method comprising providing one or more query sequences comprising a first chain sequence, at least one of the one or more query sequences being likely to have the desired property, and identifying a corresponding chain sequence for each of the one or more query sequences using the method of any embodiment of the first aspect. The method may have one or more of the following features:

[0031] The method may further include obtaining one or more antigen-binding protein candidates, each of which comprises one of the query sequence and one or more corresponding sequences derived therefrom. The method may further include assaying the one or more antigen-binding protein candidates for a desired property. The method of this embodiment may have any of the features described in relation to the first or second embodiment. The one or more antigen-binding protein candidates may be antibodies or fragments thereof. The sequences derived from the identified chain pairs may include sequences that comprise the same CDRs but differ in framework regions, sequences that comprise one or more mutations compared to the identified chain pairs, and sequences that comprise one or more fragments of the identified chain pairs. Obtaining the antigen-binding protein candidates may include identifying coding sequences for the antigen-binding protein candidates and expressing the sequences in a suitable expression system (e.g., a suitable host cell, etc.).

[0032] The desired property may be a desired binding property (e.g., ability to bind one or more targets, ability to bind one or more targets with affinity above one or more respective thresholds, etc.), a desired expression property (e.g., expression level higher than standard in one or more expression systems, expression level above a predetermined level in one or more expression systems, yield above a predetermined level in one or more expression systems, etc.), a desired stability quality (e.g., stability above a particular threshold under one or more conditions, etc.), or a combination thereof. The desired property may include ability to bind to a predetermined target. Assaying the one or more antigen binding protein candidates for the desired property may include identifying one or more antigens to which the one or more antigen binding protein candidates bind, e.g., by assaying for binding to the one or more antigen candidates. Assaying the one or more antigen binding protein candidates for the desired property may include identifying one or more antigens to which the one or more antigen binding protein candidates bind, e.g., by comparison to one or more antibodies with known targets. The antigen binding protein may be a therapeutic antibody and the desired property may include binding of a therapeutic target. Antigen binding proteins are sometimes referred to herein as "immunoproteins."

[0033] Assaying one or more candidate antigen binding proteins for a desired property may include identifying the presence or absence of a desired phenotype in an organism (e.g., an animal model, etc.) or the presence or absence of cells expressing one or more candidate antigen binding proteins. Identifying the presence of a desired phenotype may include expressing one or more candidate antigen binding proteins in one or more model cells (e.g., one or more cell lines) or organisms (e.g., one or more animal models, etc.).

[0034] The method may further include optimizing at least one sequence of one or more antigen-binding protein candidates. Optimizing the sequence of the antigen-binding protein candidates may be performed, for example, using any antibody optimization technique known in the art. Optimizing the sequence of the antigen-binding protein candidates may be performed, for example, using information from sequence data in which chain pairs were identified, for example, by analyzing sequences similar to the input sequence in which chain pairs were identified. Methods for optimizing antigen-binding proteins are known in the art and include, among others, the methods described in Mason et al.

[2021] , Seeliger et al.,

[2015] , Warszawski et al.

[2019] , Hsiao et al.

[2019] , and Richardson et al.

[2021] . Any of these methods may be used within the context of the present invention.

[0035] The query sequence may comprise a heavy chain sequence (or a portion of a heavy chain sequence) of a known antibody. Thus, the first chain may be a heavy chain sequence or a portion of a heavy chain sequence of a known antibody.

[0036] The query sequence may have been obtained by bulk BCR sequencing of the heavy chain repertoire in one or more samples. The method may include obtaining the query sequence by bulk BCR sequencing of the heavy chain repertoire in one or more samples. The one or more samples may be from one or more subjects. The one or more subjects may be identified as having a desired characteristic, such as a clinically relevant characteristic, such as a particular clinical phenotype or biomarker profile. For example, the one or more subjects may have resilience to a particular disease or condition. The disease or condition may be selected from cancer (e.g., breast cancer), neurodegenerative disease (e.g., amyotrophic lateral sclerosis), and infectious disease (e.g., COVID-19).

[0037] The method may include identifying chain pairs (e.g., heavy-light pairs) of a plurality of query chain sequences (e.g., heavy chain sequences) selected from a first (e.g., heavy) chain sequence identified in one or more samples, thereby obtaining a set of chain pairs (e.g., heavy-light chain pairs).

[0038] The method may further comprise identifying one or more targets by screening antibodies from the same source as the one or more samples against a plurality of peptide candidates. The plurality of peptide candidates may be selected based on the species from which the one or more samples are derived. For example, the source of the one or more samples may be one or more human subjects, and an antibody repertoire from the same source as the one or more samples may be screened against a set of peptide candidates representing the human peptidome to select a plurality of peptide candidates.

[0039] Identifying antigens to which the one or more antigen binding protein candidates bind may include using one or more targets identified by screening antibodies from the same source as the one or more samples against a plurality of peptide candidates. The method may further include filtering the set of identified chain pairs based on one or more criteria. The one or more criteria may be applied to identify an antigen or set of antigens to which the antigen binding protein candidate binds or is predicted to bind.

[0040] Providing one or more query sequences may include providing a first query (e.g., heavy) chain sequence and a second query (e.g., heavy) chain sequence, and identifying the corresponding (e.g., light) chain sequence for each of the one or more query sequences may include identifying one or more first corresponding (e.g., light) chain sequences and one or more second corresponding (e.g., light) chain sequences. The method may further include comparing the first and second corresponding chain sequences to identify one or more light chains that may be suitable for use as a common corresponding (e.g., light) chain for a bispecific antibody that includes both the first (e.g., heavy) chains.

[0041] According to a fourth aspect, there is provided a method of providing a tool for identifying antigen binding proteins comprising chain pairs, the method comprising: providing training data comprising training first sequences and corresponding sequences from known first and corresponding chain pairs; and using the training data to train a deep learning model to take a query first chain sequence as input and generate at least one corresponding chain sequence as output. The method of this aspect may have any of the features described in relation to the first aspect. The method may have any one or more of the following features. The method may further comprise obtaining a vocabulary for encoding the training first chain sequences and a vocabulary for encoding the training corresponding chain sequences. The vocabulary may be obtained using an encoding scheme, where any genetic sequence identifier corresponds to an individual token. The vocabulary may be obtained using an encoding scheme in which any sequence is at least partially encoded using tokens each corresponding to an individual amino acid. The vocabulary may be obtained using an encoding scheme in which any sequence is encoded at least partially using tokens each corresponding to an individual k-mer, or encoded using a byte pair encoding.

[0042] Providing training data may include providing unpaired first strand sequences and corresponding strand sequences. The unpaired training first strand sequences and corresponding strand sequences may be referred to as pre-training data. The method may further include training a first Transformer-based model using the unpaired training first and / or corresponding (e.g., heavy and / or light) strand sequences, training a second Transformer-based model using the unpaired training corresponding and / or first (e.g., light and / or heavy) strand sequences, and initializing an encoder and a decoder of the deep learning model using the pre-trained first and second Transformer models, respectively.

[0043] The first and second transformer-based models may each include a BERT model or a variant thereof, such as, for example, BERT, RoBERTa, or DistilBERT, or an autoregressive transformer model, such as GPT-2. The first and second transformer-based models may each include a RoBERTa model, a BERT model, or a GPT-2 model.

[0044] The method may further include providing the trained deep learning model to a user.

[0045] The methods described herein are performed on a computer unless the context indicates otherwise, such as when samples are obtained, processed or analyzed, or molecules or compositions are used for production, assay, or any other purpose.

[0046] According to a fifth aspect, there is provided a system comprising a processor and a computer readable medium comprising instructions that, when executed by the processor, cause the processor to perform the method steps of any embodiment of any preceding aspect. The instructions may cause the processor to perform the method steps of any embodiment of the first and / or fourth aspect.

[0047] According to a sixth aspect, there is provided one or more computer readable media comprising instructions which, when executed by one or more processors, cause the one or more processors to perform the method steps of any embodiment of any preceding aspect. The instructions may cause the processor to perform the method steps of any embodiment of the first and / or fourth aspect. [Brief description of the drawings]

[0048] BRIEF DESCRIPTION OF THE DRAWINGS [Figure 1] 1 is a flow chart outlining a method for identifying strand pairs according to the present disclosure. [Diagram 2]1 illustrates one embodiment of a system for identifying strand pairs according to the present disclosure. [Figure 3A] Schematic of several light chain prediction methods using only heavy chains as input. A: Transformer architecture with heavy chain tokenization and light chain token-to-sequence conversion. [Figure 3B] A: GRU model with heavy chain input and light chain output. B: GRU model with heavy chain input and light chain output. C: GRU model with heavy chain input and light chain output. D ... C: GRU model with heavy chain input and light chain output. D: GRU model with heavy chain input and light chain output. D [Figure 3C] Schematic diagram of several light chain prediction methods using only heavy chains as input. C: Database search method workflow. [Figure 3D] Schematic diagram of multiple light chain prediction methods using only heavy chains as input.D: Frequency search matches the ranked distribution of heavy and light chain read counts and then pairs chains with similar ranks. Random search is a variation of the database search method and is not shown. [Figure 4] Schematic diagram of heavy and light chain architecture and the organization of data used herein in relation to this architecture. A. Ig heavy chain architecture. The rough boundaries of the V, D, and J genes are marked along with the junction boundaries. The segment of the V gene towards the N-terminus is within the dotted boundary because the read length from many NGS methods is too short to cover this region and / or some primers used for NGS are slightly inserted within the V region. However, it is still possible to infer the V gene using sequences within the solid boundary. Same as BA but for the light chain. C. Re-creation of the paired read architecture design from DeKosky et al. (2015). [Diagram 5] Figure 1 shows an attention heatmap of heavy chain input and light chain predictions from the Transformer model described herein. Each column corresponds to an input heavy chain token and each row represents an output light chain token. Four of the eight attention heads are shown, showing how each head focuses on different tokens of the heavy chain. [Figure 6A]Prediction performance on the held-out test set and single-cell blind testing of the method depicted in Figure 3. A: Percentage of predictions with the correct light chain V gene. [Figure 6B] Figure 3 shows the prediction performance on the held-out test set and single-cell blind test of the method shown in Figure 3. B: Levenshtein distance distribution of predicted light chain junction sequences in the King et al.

[2021] dataset. Levenshtein distance is a metric that quantifies how different the predicted light chain junction amino acid sequence is from the original, with a longer distance indicating a poorer prediction. [Figure 7] A schematic representation of the contacts between the native pertuzumab light chain (yellow) and heavy chain (white); PDB:1s78. In the center of the image, an alanine is shown in ball-and-stick model, and an asparagine with a larger side chain could have clashed and tethered without expression, as predicted by Matchmaker. [Figure 8] ELISA traces of atezolizumab binding are shown. Filled dots represent binding curves of different antibodies to the target, while open dots represent binding to an irrelevant antigen. [Figure 9] We use two linked AntiBERTa models to illustrate the training procedure for the Tandem Transformer model described herein. A. Creating a set of training, validation, and testing masked language model tasks. B. Setting up the pre-training procedure and how to "warm up" the model feed for the following steps. C. An overview of how the warmed-up model can be used as part of a Seq2Seq model for NMT. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0049] Detailed Description of the Invention In describing the present invention, the following terminology will be employed and are intended to be defined as indicated below.

[0050] The B cell receptor is a transmembrane protein expressed on the surface of B cells. It contains a binding portion (also called the "antigen-binding subunit" or "membrane immunoglobulin", "mlg") that comprises a membrane-bound immunoglobulin molecule (also called an antibody) that recognizes the cognate antigen, and a signaling portion. The membrane-bound immunoglobulin molecule contains two immunoglobulin light chains and two immunoglobulin heavy chains and is identical to the corresponding secreted antibody except for the integral membrane domain. The signaling portion is a heterodimer called Ig-α / Ig-β (CD79) that is bound together and attached to the immunoglobulin by a disulfide bridge. An antibody (Ab) or immunoglobulin (Ig) is an immune protein that contains an antigen-binding site and a constant region that belongs to one of a limited set of isotypes (IgA, IgD, IgE, IgG, or IgM) and that mediates interactions with other components of the immune system. In humans and most mammals, antibodies contain four polypeptide chains: two identical heavy chains and two identical light chains connected by disulfide bonds. The light chains typically contain one variable domain, V L and one constant domain C L whereas the heavy chain typically consists of one variable domain, V H and 3-4 constant domains C H 1. C H The variable domains form the antigen-binding region and consist of F VEach variable domain contains three hypervariable regions called complementarity determining regions (CDRs) that together form the antigen-binding site. The variable region of each immunoglobulin heavy or light chain is encoded in several individual pieces - known as gene segments (minigenes): Ig heavy chains contain a variable (V), diversity (D), and joining (J) segment, and Ig light chains contain a V and a J segment. Multiple copies of the V, D, and J gene segments are present in the genome, and B cell precursors assemble Ig variable regions by (almost) randomly selecting and combining one V, one D, and one J gene segment (or one V and one V segment in light chains) in a process called V(D)J gene rearrangement. This process involves the formation of a two-stranded break between the required segments that forms a hairpin loop, which is then joined together. The joining process is imprecise, resulting in variable additions or deletions of nucleotides between the V and J segments (light chains) or between the V, DJ, D and J segments (heavy chains), creating a great diversity in the sequences at the junctions between the segments (called "junction sequences"). The V(D)J gene rearrangement process produces novel amino acid sequences in the antigen-binding region of Ig, creating an enormous diversity in antigen-recognition capabilities. As a result of the process, each Ig heavy chain variable region contains a V segment, a D segment, and a J segment, along with junction sequences that span the junctions between these segments (as shown in FIG. 4A). Similarly, each Ig light chain hypervariable region contains a V segment and a J segment, as well as junction sequences that span the junctions between these segments (as shown in FIG. 4B). Within the variable region, CDR1 and CDR2 are found in the V segment, and CDR3 contains some of the V segment, all of the D segment (in the heavy chain), and some of the J segment.

[0051] T cell receptors are membrane anchored proteins expressed on the surface of T cells. They contain a pair of protein chains that together form a binding moiety that recognizes cognate antigens. In mammals, they are expressed in a complex with the constant T cell coreceptor chain CD3, which includes the CD3γ chain, the CD3δ chain, and two CD3ε chains. The constant chains associate with the T cell receptor and the constant ζ chain to form the TCR complex, which together can generate a signal when antigen is bound to the T cell receptor. The TCR is a heterodimeric protein that contains two highly variable chains, the α and β chains (in the majority of T cells) or alternatively the γ and δ chains (in a minority of T cells). Each chain contains two extracellular domains: a variable region (or variable domain) and a constant region (or constant domain, close to the cell membrane), a transmembrane region, and a short cytoplasmic tail. In the case of αβ TCRs, the variable regions come together to bind peptides (antigens) within the context of an MHC (major histocompatibility complex) molecule. Each variable domain contains three hypervariable regions called complementarity determining regions (CDRs, called CDR1, CDR2, and CDR3 in each chain, respectively), which come together to form the antigen-binding site. TCRs are members of the immunoglobulin superfamily, which includes BCRs and antibodies. In a process similar to that described above, the variable regions of each TCR chain are encoded in several individual pieces - known as gene segments (minigenes): β and δ chains contain a variable (V), diversity (D), and joining (J) segment, and α and γ chains contain a V and a J segment. Multiple copies of the V, D, and J gene segments are present in the genome, and T cell progenitors assemble the variable regions of TCR chains by (almost) randomly selecting and combining one V, one D, and one J gene segment (or one V and one J segment in the α / γ chains) in a process called V(D)J gene rearrangement, which involves the formation of a two-strand break between the required segments that forms a hairpin loop, which is then joined together.The joining process is imprecise, resulting in variable additions or deletions of nucleotides between the V and J segments (α / γ chains) or between the V, DJ, D and J segments (β / δ chains), creating a great diversity in the sequences at the junctions between the segments (called "junction sequences"). The V(D)J gene rearrangement process produces novel amino acid sequences in the antigen-binding region of the TCR, creating an enormous diversity in antigen recognition capabilities. As a result of the process, the variable region of each β / δ chain contains a V segment, a D segment, and a J segment, along with junction sequences that span the junctions between these segments. Similarly, each α / γ chain hypervariable region contains a V segment and a J segment, as well as junction sequences that span the junctions between these segments. Within the variable region, CDR1 and CDR2 are found in the V segment, and CDR3 contains some of the V segment, all of the D segment (in the heavy chain), and some of the J segment.

[0052] As used herein, a "variable chain" of an antigen binding protein (also referred to herein simply as a "chain") refers to the chain of an antigen binding protein that is involved in antigen recognition or a portion of the antigen binding protein chain that contains at least a portion of the variable region of that chain. A variable chain contains the variable region that is responsible for the diverse repertoire of antigen recognition within an antigen binding protein. A variable chain can be a BCR heavy or light chain, an antibody heavy or light chain, a TCR alpha or beta chain, a TCR gamma or delta chain, or any portion of such a chain that contains at least a portion of one or more variable regions within these chains.

[0053] The B cell receptor repertoire (or corresponding antibody repertoire) present in a sample can be investigated using sequencing techniques. As explained above, two main sequencing techniques are used: single B cell sequencing and bulk B cell population sequencing. Because the transmembrane domains of the BCR signaling and antigen binding moieties are not variable, these techniques focus on the common parts between the B cell repertoire and the corresponding antibody repertoire. Thus, in this disclosure, references to BCR sequences, BCR repertoires, BCR heavy chain sequences, BCR light chain sequences, and any parts thereof are used interchangeably with the corresponding antibody sequences, antibody repertoires, antibody heavy chain sequences, antibody light chain sequences, and any parts thereof. For example, references to sequencing the BCR heavy chain variable region are equivalent to sequencing the corresponding antibody heavy chain variable region, and these two terms may be used interchangeably. The term "antigen binding protein" is used herein to refer to a BCR protein, a TCR protein, an antigen binding portion of a BCR protein, an antibody, or the original BCR protein. It is used to refer to the TCR protein, or any part thereof that maintains the antigen-binding properties of an antibody. It should be noted that the repertoire of antibodies circulating in the blood of an individual may not match the B cell receptor repertoire present in the sample at the same time, since antibodies produced by B cells that are no longer present in the individual (e.g. because they have died) may be present in the sample. Thus, the term "corresponding antibody repertoire" refers to the repertoire of antibodies expressed by B cells present in the sample, rather than the repertoire of antibodies (proteins) actually present in the sample.

[0054] Single B cell sequencing can maintain the correspondence between heavy and light chain sequences. Two main approaches can be used to do this. The first approach is physical linkage of VH and VL [DeKosky et al., 2016]. The second approach is cell barcoding (e.g., as provided by 10x Genomics) [King et al., 2021]. Physical linkage approaches have a higher throughput than cell barcoding approaches, but are more difficult to recover the entire sequence. Conversely, cell barcoding has a lower throughput, but allows easier recovery of the entire sequence. Whatever the approach, single B cell sequencing is limited (to various degrees) in terms of throughput, as explained above. Some single B cell sequencing techniques are further limited in terms of the length of the sequences recovered. As a result, the BCR / antibody sequences identified using some single B cell sequencing methods may be limited to investigating a single CDR region, e.g., CDR3, in both the heavy and light chains (in other words, adjacent V and J segments may be identified, but may not be fully sequenced to obtain the sequences of CDR1 and CDR2 in the V segment). In other words, the data set from a single B cell sequencing method may vary in the extent to which heavy and light chain sequences are identified. It may not be practical to sequence (or record) every single base of the V(D)J segment within the sequenced region, and therefore such sequencing efforts may focus on obtaining junction sequences and sufficient information to identify the V, D, and J genes. As a result, such methods may provide information including: the identity of the V, D, and J segments of the heavy chain (e.g., in the form of a V / D / J gene segment identifier), the sequence of the joining segment in the heavy chain, the identity of the V and J segments of the light chain (e.g., in the form of a V / J gene identifier), and the sequence of the joining segment in the light chain. The identity of each segment can be used to retrieve the corresponding germline alignment from the database. However, if the data only includes the identity of each segment, any mutations (e.g., somatic mutations) that may be present in a particular chain compared to a reference germline sequence may not be captured.Conversely, sequencing of bulk B cell populations does not maintain the pairing between heavy and light chain sequences, but is less restrictive in terms of the sequencing capacity within the heavy and light chain repertoires, respectively (especially the depth of sequencing of the BCR repertoire). Sequencing of bulk B cell populations may include sequencing of the heavy chain repertoire, the light chain repertoire, or both of the B cell population. However, as mentioned above, due to the bulkiness of the process, even when both the light and heavy chain repertoires are sequenced, it is not possible to maintain pairing information during the sequencing process. Such sequencing may generate information as sparse as that obtained using single cell B sequencing, or more detailed information, including, for example, complete CDR sequences, sequences of multiple CDRs, complete variable region sequences, or complete variable region sequences and enough constant regions to identify the isotype of the sequence. Similar considerations apply to sequencing of T cell repertoires. In particular, many of the processes and limitations discussed above in relation to the study of B cell repertoires and antibodies (and in particular in relation to the sequencing of these repertoires) also apply to T cell repertoires.

[0055] As used herein, the term "variable chain sequence" encompasses the terms "heavy chain sequence", "light chain sequence", "α chain sequence", "β chain sequence", "γ chain sequence", and "δ chain sequence" and refers to any information that can be obtained from B-cell or T-cell sequencing technology, ranging from a combination of one or more gene segment identifiers and / or a junction sequence at one end to a complete chain sequence at the other end. In particular, the terms "heavy chain sequence" and "light chain sequence" refer to any information that can be obtained from B-cell sequencing technology, ranging from a combination of one or more gene segment identifiers and / or a junction sequence at one end to a complete chain sequence at the other end. Furthermore, the terms "variable chain sequence", "heavy chain sequence", "light chain sequence", "α chain sequence", "β chain sequence", "γ chain sequence", and "δ chain sequence" refer interchangeably to amino acid sequences or corresponding nucleic acid coding sequences. Similarly, variable chain pairing or pairing (heavy-light chain pairing or pairing), as defined herein, refers to a combination of heavy and light chain sequences, α and β chain sequences, or γ and δ chain sequences, each of which spans one or more gene segment identifier combinations and / or junction sequences at one end to the complete chain sequence at the other end.

[0056] Within the context of providing a desired antibody or antigen-binding protein, such as, for example, a therapeutic antibody, the term "antibody" (Ab) includes monoclonal antibodies, polyclonal antibodies, multispecific antibodies (e.g., bispecific antibodies), and antibody fragments that exhibit the desired biological activity and comprise a heavy-light chain pairing identified as described herein or a heavy-light chain pairing derived from a heavy-light chain pairing identified as described herein (e.g., by further optimization, affinity maturation, etc.).

[0057] A "sample", as used herein, may be a cell or tissue sample, biological fluid, extract (e.g., DNA or RNA extract obtained from a subject) from which B cell genomic material (e.g., RNA or DNA) can be obtained for genomic analysis, such as by sequencing (e.g., whole genome sequencing, whole exome sequencing, target / capture sequencing, RNAseq, etc.). A sample may be a cell, tissue, or biological fluid sample obtained (e.g., biopsy) from a subject. Such a sample may be referred to as a "subject sample". In particular, a sample may be a blood sample, a lymph node sample, a spleen sample, a tumor sample, or a sample derived therefrom (e.g., B cell purification, T cell purification, RNA extraction, etc.). As used herein, terms such as "genomic material", "genomic sequencing", and the like encompass both material / sequences present in the genome and the transcriptome of a sample, unless the context indicates otherwise. A sample may be obtained fresh from a subject or may be processed and / or stored (e.g., frozen, fixed, or subjected to one or more purification, enrichment, or extraction steps) prior to genomic analysis. The sample may be a cell or tissue culture sample. Thus, a sample as described herein may refer to any type of sample that includes B cells or genomic material derived therefrom, whether it is a biological sample obtained from a subject or from a sample obtained, for example, from a cell line. The sample is preferably from a mammal (e.g., a mammalian cell sample or a sample from a mammalian subject such as a cat, dog, horse, donkey, sheep, pig, goat, cow, mouse, rat, rabbit, or guinea pig), preferably a human (e.g., a human cell sample or a sample from a human subject). Furthermore, the sample may be transported and / or stored, collection may occur at a location remote from the site of sequence data acquisition (e.g., sequencing), and / or any computer-implemented method steps described herein may occur at a location remote from the site of sample collection and / or the site of genomic data acquisition (e.g., sequencing) (e.g., computer-implemented method steps may be performed by a networked computer such as a "cloud" provider).

[0058] The term "sequence data" refers to information that indicates the genomic (DNA or RNA) or proteomic material in a sample with a particular sequence. Thus, sequence data may include one or more nucleotide sequences and / or one or more amino acid sequences. Such information may be obtained using sequencing technologies, such as next generation sequencing (NGS), such as whole exome sequencing (WES), whole genome sequencing (WGS), whole transcriptome sequencing (RNAseq), or sequencing of captured genomic loci (targeted or panel sequencing). When NGS technologies are used, the sequence data may include a count of the number of sequencing reads with a particular sequence. The sequence data may be mapped to a reference sequence, e.g., a reference genome, using methods known in the art, such as, for example, Bowtie (Langmead et al., 2009). Thus, a particular location or locus may be associated with the count of sequencing reads or an equivalent non-digital signal ("location" refers to the location within the reference genome or transcriptome to which the sequence data is mapped). Additionally, a location may contain mutations, in which case each of the possible variants (also called "alleles") at that particular location may be associated with a count of sequencing reads or an equivalent non-digital signal. The process of identifying the presence of mutations at a particular location in a sample is called "variant calling" and can be performed using methods known in the art (e.g., a general-purpose NGS variant caller such as GATK HaplotypeCaller (https: / / gatk.broadinstitute.org / hc / en-us / articles / 360037225632-HaplotypeCaller) or a tool specifically designed for immune sequences such as IgBLAST (https: / / www.ncbi.nlm.nih.gov / igblast / ) [Ye et al., 2013], etc.). Genomic sequence data may be converted to amino acid sequences by in silico translation of coding regions (either directly from the mRNA sequence or from identified coding regions in the genomic sequence), as known in the art.

[0059] As used herein, "treatment" refers to the reduction, alleviation, or absence of one or more symptoms of the disease being treated compared to symptoms prior to treatment. "Prevention" (or prophylaxis) refers to delaying or preventing the onset of symptoms of the disease. Prevention may be absolute (such that the disease does not occur) or may be effective only in some individuals or for a limited period of time.

[0060] The compositions described herein may be pharmaceutical compositions that further comprise a pharma- ceutically acceptable carrier, diluent, or excipient.The pharmaceutical compositions may optionally comprise one or more additional pharma- ceutically active polypeptides and / or compounds.Such formulations may be, for example, in a form suitable for intravenous infusion.

[0061] As used herein, the term "computer system" includes hardware, software, and data storage for implementing the system or performing the method according to the above embodiments. For example, a computer system may include a central processing unit (CPU), a graphical processing unit (GPU), input means, output means, and data storage, which may be embodied as one or more connected computing devices. Preferably, the computer system comprises a computing device having a display or providing a visual output display (e.g., of a business process design). The data storage may include RAM, disk drives, or other computer readable media. A computer system may include multiple computing devices connected by a network and capable of communicating with each other via the network. It is expressly contemplated that a computer system may consist of or include a cloud computer. The term "processor" encompasses any processing unit or combination of processing units, including in particular a CPU and a GPU.

[0062] As used herein, the term "computer-readable medium" includes, without limitation, any one or more non-transitory media that can be directly read and accessed by a computer or computer system. Media can include, but are not limited to, magnetic storage media, such as floppy disks, hard disk storage media, and magnetic tape; optical storage media, such as optical disks or CD-ROMs; electrical storage media, such as memory, including RAM, ROM, and flash memory; and hybrids and combinations of the above, such as magnetic / optical storage media.

[0063] Identification of variable chain pairs The present disclosure provides a method for identifying variable chain pairs from variable chain sequences. An exemplary method is described with reference to FIG. 1. FIG. 1 illustrates an embodiment in which heavy or light chain sequences of a B cell receptor or antibody are used to identify heavy and light chain pairs. In other words, FIG. 1 illustrates an embodiment in which the variable chain sequences are BCR / antibody heavy and light chains from a BCR. However, the method described with reference to FIG. 1 is applicable to embodiments in which TCR alpha, beta, gamma, or delta chain sequences are used to identify alpha beta (if the input chain pair is an alpha or beta chain) or gamma delta (if the input chain pair is a gamma or delta chain) chain pairs. In optional step 10, a sample may be obtained from a subject that includes B cell genomic material (typically in the form of RNA, such that the RNA encoding for the BCR expressed by the cells from which the B cell genomic material is derived can be extracted and sequenced). Similarly, a sample including T cell genomic material may be used in embodiments in which TCR chain pairs are identified. In optional step 12, the BCR repertoire in the sample may be sequenced using bulk BCR sequencing. This may include sequencing the heavy chain BCR repertoire in the sample. Similarly, the TCR repertoire in the sample may be sequenced using bulk TCR sequencing. This may include sequencing the β chain repertoire in the sample. In step 14, a query chain sequence is provided. In the illustrated embodiment, the query sequence is a heavy chain sequence. In other embodiments, the query chain sequence may be a light chain sequence. Providing a query sequence may include, in step 14A, selecting one of the heavy chain sequences sequenced in step 12 as a query sequence. Providing a query sequence may include, in step 14B, providing a sequence including a V gene sequence or identifier, a J gene sequence or identifier, and a junction sequence. For example, step 14B may include extracting the V gene sequence or identifier, the J gene sequence or identifier, and the junction sequence for the selected sequence from the bulk BCR sequencing dataset. Similarly, steps may be performed in the context of TCR pairing, for example using a query beta chain sequence.In optional step 16, a deep learning model is provided, the deep learning model configured to take a query variable chain sequence as input and generate at least one corresponding variable chain sequence as output. In the illustrated embodiment, the query sequence is a heavy chain sequence, and thus the deep learning model is configured to take a query heavy chain sequence as input and generate at least one corresponding light chain sequence as output. In other embodiments, the query chain sequence may be a light chain sequence, and thus the deep learning model may be configured to take a query light chain sequence as input and generate at least one corresponding heavy chain sequence as output. In yet other embodiments, the query chain sequence may be a β chain sequence (or an α, δ, or γ chain sequence), and thus the deep learning model may be configured to take a query β chain sequence (or an α, δ, or γ chain sequence) as input and generate at least one corresponding α chain sequence (or at least one corresponding β, γ, or δ chain sequence) as output. The deep learning model may have been previously trained using training variable chain sequences from known variable chain pairs, such as training heavy and light chain sequences from known heavy-light chain pairs in the illustrated embodiment. Thus, providing the deep learning model may simply involve retrieving or otherwise receiving the trained deep learning model from a computer-readable medium, such as a memory associated with the processor executing the method. Training of the deep learning model is described in more detail below. Alternatively, the deep learning model may be trained as part of the method using training variable chain sequences from known heavy-light chain pairs, such as in the illustrated embodiment training heavy and light chain sequences from known heavy-light chain pairs.

[0064] In step 18, the query chain sequence is provided to the deep learning model. Step 18 may include an optional step 18A of encoding the query sequence using a predefined encoding scheme. Step 18 may include an optional step 18B of decoding each corresponding sequence output by the deep learning model using a predefined encoding scheme. The encoding scheme and the decoding scheme used for encoding may be previously defined based on the content of the (e.g. heavy and light) chain sequences of the training variable chain sequences used to train the deep learning model. Step 18 may include an optional step 18C of selecting a sequence of one or more corresponding variable chain sequences (light chain sequences in the illustrated embodiment) with which is associated the highest confidence metric among the one or more corresponding variable chain sequences, and the deep learning model is configured to generate as output one or more corresponding chain sequences each associated with a confidence metric such as a probability. Step 18C may be performed before or after step 18B. In optional step 20, one or more gene sequence identifiers in at least one counter chain sequence (light chain in the illustrated embodiment) may be replaced with the corresponding germline sequence. In optional step 22, the results of any of the previous steps (particularly steps 18 and / or 20) may be provided to a user, for example through a user interface. These results may be used, for example, to provide a therapeutic antibody, as described further below. The method may be repeated for multiple query sequences. This may include repeating steps 14-18.

[0065] Training of the deep learning model will now be described with reference to optional steps 10'-16'. In step 10', training data is provided that includes training variable chain sequences from known variable chain pairs. In the illustrated embodiment, the training data includes heavy and light chain sequences from known heavy-light chain pairs. The training data may include at least 20,000 training chain pairs, at least 30,000, at least 40,000, at least 50,000, at least 60,000, at least 70,000, at least 80,000, at least 90,000, at least 100,000, at least 120,000, or at least 150,000 training chain pairs. In B cell receptor / antibody related embodiments, the training data may include at least 800,000, at least 100,000, at least 120,000, or at least 150,000 pairs of training heavy and light chain sequences. The training data may further include unpaired training sequences, which in the illustrated embodiment are heavy and light chain sequences. The unpaired training strand sequences may be referred to as "pre-training data." Thus, the training data may include the training data itself (which includes paired strand sequences, particularly, in the illustrated embodiment, paired heavy and light chain sequences) and the pre-training data (which includes unpaired strand sequences, particularly, in the illustrated embodiment, unpaired heavy and light chain sequences). The training data can include at least 100,000, at least 200,000, at least 300,000, at least 400,000, at least 500,000, at least 600,000, at least 700,000, at least 800,000, at least 900,000, at least 1 million (or at least 5 million, 10 million, 15 million, 20 million, 25 million, 30 million, 35 million, or 40 million) unpaired training strand sequences of a first type and / or corresponding type.The training data may include at least 1 million (or at least 5 million, 10 million, 15 million, 20 million, 25 million, 30 million, 35 million, or 40 million) unpaired training heavy chain sequences and at least 1 million (or at least 5 million, 10 million, or 15 million) unpaired training light chain sequences. As one of skill in the art will appreciate, the amount of training data and / or pre-training data may be limited by the amount of suitable data available and may change as more data becomes available. Furthermore, the amount of data available may depend on the particular use case, e.g., the identification of the first and matching chain sequences (e.g., there may be more data available for the αβ TCR than for the rarer γδ TCR), the criteria used when filtering the data (see step 12'), etc. The numbers provided may be applied to the data before and / or after any filtering is applied. In step 12', the training data is filtered. For example, the training data may be filtered to exclude any pairs that include junction sequences (e.g., in the heavy and / or light chains) outside a predetermined range of lengths. As another example, the training data may be filtered based on any characteristic of the data, including, for example, the type of cell or organism from which the data is derived, whether the data is from a naive library, whether the data is from a subject immunized to a particular antigen, etc. In other words, the training data may be filtered to ensure that the training data includes only data having one or more characteristics of interest (inclusion filter) or does not include such data (exclusion filter). In step 14', one or more encodings are defined in the training data by obtaining a vocabulary for encoding training chain sequences, and in particular, in the illustrated embodiment, a vocabulary for encoding training heavy chain sequences and a vocabulary for encoding training light chain sequences.Defining the encoding scheme may include excluding any tokens whose number of occurrences in the training data is below a predetermined threshold (e.g., 2) from a vocabulary built based on the content of the training strand sequences. In step 16', the training data is used to train a deep learning model using the training data to take a query heavy chain sequence (in the illustrated embodiment) as an input and generate at least one corresponding light chain sequence (in the illustrated embodiment) as an output. Training the deep learning model may include first training a transformer-based model using the unpaired training strand sequences and initializing an encoder and a decoder of the deep learning model using the pre-trained transformer model. Alternatively, training the deep learning model may include training a first and a second transformer-based model using respective unpaired training strand sequences of a first type and a second type (in the illustrated embodiment, the first type is a heavy chain and the second type is a light chain) and initializing an encoder and a decoder of the deep learning model, respectively, using the pre-trained transformer model. Training the deep learning model may include obtaining training sequences that include full-length sequences in the variable regions of the second type chain and / or the first type chain (e.g., the light chain and / or the heavy chain in the illustrated embodiment) by inferring missing sequence information when the chain sequences from the known chain pair do not include full-length sequences in said variable regions.

[0066] system FIG. 2 shows an embodiment of a system for identifying variable chain pairs from input variable chains according to the present disclosure. The system comprises a computing device 1, which comprises a processor 101 and a computer readable memory 102. In the illustrated embodiment, the computing device 1 also comprises a user interface 103, which is shown as a screen, but may include any other means of conveying information to a user, such as through an audible or visual signal. The computing device 1 is communicatively connected, such as through a network 6, to a sequence data acquisition means 3, such as a sequencing machine, and / or one or more databases 2 storing sequence data. The one or more databases may further store other types of information, such as reference sequences, parameters, etc., that may be used by the computing device 1. The computing device may be a smartphone, a server, a tablet, a personal computer, or other computing device. The computing device is configured to perform a method for identifying variable chain pairs from input variable chains (heavy or beta chains as appropriate, preferably heavy chains) as described herein. In an alternative embodiment, the computing device 1 is configured to communicate with a remote computing device (not shown) that is itself configured to perform a method for identifying variable chain pairs from input variable chains as described herein. In such a case, the remote computing device can also be configured to transmit the results of the method to the computing device. The communication between the computing device 1 and the remote computing device can be through a wired or wireless connection and can be performed via a local or public network, such as via the public Internet or WiFi. The sequence data acquisition means 3 can be connected to the computing device 1 by wire or through a wireless connection, such as through the network 6 as shown. The connection between the computing device 1 and the sequence data acquisition means 3 can be direct or indirect (e.g., through a remote computer, etc.).The sequence data acquisition means 3 is configured to acquire sequence data from a nucleic acid sample, for example a genomic DNA sample or an RNA sample extracted from B cells or T cells purified from a fluid and / or tissue sample (such as peripheral blood, spleen, lymph nodes, tumor tissue, or any other type of sample containing B cells or T cells). In some embodiments, the sample may have undergone one or more pre-processing steps such as DNA / RNA purification, fragmentation, library preparation, target sequence capture (such as exon capture and / or panel sequence capture). Any sample preparation process suitable for use in identifying B cell receptor sequences or repertoires can be used within the context of the present invention. The sequence data acquisition means 3 is preferably a next generation sequencer. The sequence data acquisition means 3 may be directly or indirectly connected to one or more databases 2 in which sequence data (raw or partially processed) may be stored.

[0067] Purpose The above methods are applied in any situation where it is desirable to identify antibodies or BCRs that are likely to bind to a target from limited information on heavy chains, light chains, or portions thereof (e.g., V genes, J genes, and junction sequences, etc.). This is often the case in the context of the antibody therapeutic discovery process. Antibody therapeutics have proven to be a successful approach for a wide range of diseases, from neurodegenerative diseases to cancer. Thus, the methods described herein are applied in the context of providing treatment in each of these clinical situations. Furthermore, the methods described herein can be used to identify potentially functional antibodies or BCRs from any input heavy / light chains or portions thereof, regardless of whether the input information is generated de novo for a specific purpose (e.g., from a patient or sample identified as having a desired phenotype) or is from an existing / historic dataset (e.g., mining or re-mining an existing dataset to discover new treatments or identify immune proteins that may explain why a particular clinical phenotype persists).

[0068] Thus, the present invention also provides a method of providing an antibody therapy, the method comprising identifying a heavy-light chain pairing using any of the methods described herein or identifying a heavy-light chain pairing derived (e.g., by further optimization, mutation, etc.) from a heavy-light chain pairing identified using any of the methods described herein. The heavy-light chain pairing may be obtained for an input heavy chain sequence obtained by bulk BCR sequencing of the heavy chain repertoire in one or more samples. The one or more samples may be from one or more subjects. The one or more subjects may have been identified as having a desired characteristic, such as a clinically relevant characteristic, such as a particular clinical phenotype or biomarker profile. For example, the one or more subjects may have resilience against a particular disease or condition. The disease or condition may be selected from cancer (e.g., breast cancer), a neurodegenerative disease (e.g., amyotrophic lateral sclerosis), or an infectious disease (e.g., COVID-19). The method may include identifying heavy-light chain pairings of a plurality of input heavy chain sequences selected from the heavy chain sequences identified in the one or more samples, thereby obtaining a set of heavy-light chain pairings. The method may further include identifying a target (or a plurality of putative targets or a set of targets) of each heavy-light chain pairing within the heavy-light chain pairing or set of heavy-light chain pairings. The method may further include identifying the one or more targets by screening antibodies from the same source as the one or more samples against a plurality of peptide candidates. The plurality of peptide candidates may be selected based on the species from which the one or more samples are derived. For example, the source of the one or more samples may be one or more human subjects, and an antibody repertoire from the same source as the one or more samples may be screened against a set of peptide candidates representing the human peptidome. Identifying the target (or putative target or set of targets) of the heavy-light chain pairing, or of each heavy-light chain pairing within a heavy-light chain pairing, may include using one or more targets identified by screening antibodies from the same source as the one or more samples, against a plurality of peptide candidates.The method may further include filtering the set of heavy-light chain pairings based on one or more criteria. The one or more criteria may be applied to the identification of the set of putative targets or the identified heavy-light chain pairings. The method may further include obtaining an antibody or a fragment thereof comprising the identified heavy-light chain pairing or a heavy-light chain pairing derived from the identified heavy-light chain pairing. Obtaining the antibody or fragment thereof may include identifying a coding sequence for the antibody or fragment thereof and expressing the sequence in a suitable expression system (e.g., a suitable host cell, etc.). The method may further include identifying one or more antigens to which the antibody or fragment thereof binds, for example, by assaying for binding to one or more candidate antigens. The method may further include optimizing the sequence of the antibody or fragment thereof. Optimizing the sequence of the antibody or fragment thereof may be performed using any antibody optimization technique known in the art. Optimizing the sequence of an antibody or fragment thereof can be performed, for example, using information from sequence data in which heavy-light chain pairings have been identified by analyzing the sequence as well as the input sequence in which heavy-light chain pairings have been identified.

[0069] The invention also provides a method of providing an immunotherapeutic composition, the method comprising identifying a heavy-light chain pairing as described herein and generating an immunotherapeutic composition comprising an antibody comprising the heavy-light chain pairing or an antibody derived from the heavy-light chain pairing (e.g., by further optimization, mutation, etc.).

[0070] The methods described herein can also be utilized in the context of providing bispecific antibodies. For example, the methods described herein can be used to identify a light chain suitable for pairing with two different heavy chains of interest. Thus, the present invention also provides a method of providing a bispecific antibody, comprising identifying a light chain pairing common to each of the two heavy chains using any of the methods described herein, or a combination of two heavy chains with a common light chain derived (e.g., by further optimization, mutation, etc.) from a heavy-light chain pairing identified using any of the methods described herein. In such an embodiment, it may be advantageous for the deep learning model to output a plurality of corresponding light or heavy chain sequences. For example, the deep learning model may be used to predict a first plurality of corresponding light chain sequences for a first heavy chain sequence, and a second plurality of corresponding light chain sequences for a second heavy chain sequence. The predicted first and second plurality of light chain sequences may then be compared to identify one or more light chains that may be suitable for use as a common light chain for a bispecific antibody comprising both heavy chains.

[0071] The methods described herein can also be utilized in the context of antibody optimization. For example, the methods described herein can be used to identify a light chain suitable for pairing with a heavy chain, the pairing having one or more advantageous properties (e.g., improved functionality or developability, etc.) compared to the original pairing of the heavy chain. Thus, the present invention also provides a method of providing an improved antibody, the method comprising identifying a heavy-light chain pairing from an input heavy chain of the original antibody using any method described herein, or identifying a heavy-light chain pairing derived (e.g., by further optimization, mutation, etc.) from a heavy-light chain pairing identified using any method described herein. The methods described herein can also be utilized in any situation where it is desirable to identify a TCR likely to bind to a target from limited information on the β, α (or less commonly, γ or δ) chain, or parts thereof (e.g., V gene sequences, J gene sequences, and joining sequences, etc.). This is often the case in the context of the discovery process of cell therapy, such as genetically engineered T cells. Thus, the present invention also provides methods of providing TCR-based therapy, such as engineered T cells expressing a particular TCR, comprising identifying an αβ or γδ chain pairing using any of the methods described herein, or identifying an αβ or γδ chain pairing derived (e.g., by further optimization, mutation, etc.) from an αβ or γδ chain pairing identified using any of the methods described herein. Thus, the methods described herein can also be utilized in the context of T cell receptor optimization, similar to those described above for antibodies.

[0072] The following are presented as examples and should not be construed as limitations on the claims. EXAMPLES

[0073] Working Example These examples illustrate how heavy-light chain pairings are identified according to the present invention and validated using single-cell data sets with known pairings and newly acquired experimental data.

[0074] method Dataset Summary statistics for all datasets are provided below in Table 1. The sources and hierarchies of the training, validation, validation, and blind test sets are described below.

[0075] [Table 1]

[0076] Training, validation, and test sets: Paired heavy-light chain sequences from three donors in DeKosky et al.,

[2015] were combined with three naive BCR libraries from DeKosky et al.

[2016] . These data sets contain entries that each contain: heavy chain V gene identifiers, heavy chain junction sequences (nucleotides and amino acids), heavy chain J gene identifiers, light chain V gene identifiers, light chain junction sequences (nucleotides and amino acids), and light chain J gene identifiers. This training set was chosen primarily due to its public availability and size. Note that the data entries also contained heavy chain D gene identifiers, but this information was not used. This is because annotation of D genes is considered less accurate than annotation of V and J genes. Sequences were filtered to retain those with heavy chain junctions between 7 and 30 amino acids in length and light chain junctions between 7 and 20 amino acids in length. This is because these out-of-bounds junction sequences are considered rare, and increasing the length comes with a cost in terms of computational power that is unlikely to be balanced by a gain in information for longer sequences that are rare. This filter removed 84 pairs that had heavy chain junctions with out-of-bounds sizes and 168 pairs that had light chain junctions with out-of-bounds sizes. To maintain a more concise vocabulary, the data was filtered for sequences in which the heavy and light chain V genes were observed in at least two sequences. In fact, increasing the vocabulary size comes with a cost of training additional model parameters that is unlikely to be balanced by a benefit in including very rare V genes. This filter removed only a small number of entries. In fact, only one heavy chain V gene was observed once among the starting pairs. All light chains were observed in at least two sequences. Sequences with IMGT pseudogenes (entries containing sequences labeled as pseudogenes in the Immunogenetics Database (http: / / www.imgt.org) which are sequences whose coding regions contain stop codons and / or frameshift mutations and / or mutations affecting the start codon) were also removed.Finally, any identical heavy-light chain pairs were removed, which is functionally equivalent to a 99% redundancy cutoff (in other words, any heavy-light chain pair across the different sets is at most 99% identical, ensuring that pairs in the training dataset are indeed unique). The length and number of entries that the filters removed together gave a total of 253 pairs (84+168+1) from the 190,240 starting pairs that passed the pseudogene and redundancy filters, leaving a set of 189,987 pairs (see Table 1).

[0077] The 189,987 sequences were divided into training, validation, and test sets of 153,889, 17,099, and 18,999 sequences (corresponding to a split of approximately 80% training / 10% validation / 10% testing). Although the same heavy chain sequence can be present across the three sets, none of these heavy chains have an identical light chain counterpart.

[0078] Training data survey: χ for heavy-light chain pairs with at least five observations in the training set 2 The training data were characterized using a χ 2 =8102.9, p-value <10 -9 ). However, contingency tables of heavy-light chain pairs were inconsistent between studies, and inferences regarding the randomness of pairing were inconclusive. At the sequence level, a single heavy chain sequence typically had one unique partner light chain sequence, but there can be up to seven different light chain partners per heavy chain. Light chains proved to be more promiscuous, with a single light chain sequence found to pair with up to 1042 different heavy chains. Nevertheless, 58123 out of 73549 light chains had one unique heavy chain partner.

[0079] Blind test set: A single-cell dataset was obtained from the Paired Observed Antibody Space [Kovaltsuk et al., 2018]. Three studies of the human BCR repertoire were pooled from King et al.,

[2021] , Eccles et al.,

[2020] , and Setliff et al.,

[2019] . There were a total of 30,332, 741, and 4,944 unique heavy-light chain pairs, respectively.

[0080] Array Tokenization The training set data did not contain full-length heavy and light chain sequences because the single-cell sequencing method used to generate this data was unable to recover the complete amino acid sequences of the light / heavy chains. Instead, the training dataset included: (i) for each heavy chain: V gene identifier, junction sequence, J gene identifier, and D gene identifier (but the D gene identifier was not used), and (ii) for each light chain: V gene identifier, junction sequence, and J gene identifier. Thus, each entry included a combination of gene identifier and sequence, such as, for example, IGHV3-23 / CAR...DYW / IGHJ6-IGKV3-20 / CQQ... / IGKJ2. The model was trained to take as input the tokenized heavy chain sequences corresponding to the V gene identifier, junction sequence, and J gene identifier, and generate as output the tokenized light chain sequences corresponding to the V gene identifier, junction sequence, and J gene identifier. To address this data format, a custom encoding method was designed for tokenization. Each V gene consisted of a single token, each J gene consisted of a single token, and junction amino acid sequences were tokenized as overlapping 3-mers. The junction sequence is the most diverse region of the sequence and is believed to mediate the majority of the binding functions, therefore we increased the granularity of tokenization of this sequence. Tokens were used if they occurred at least twice in the training set (as already mentioned above in connection with filtering the data).

[0081] Other possible ways to tokenize the junction amino acid sequence (or the complete sequence) include, for example, byte-pair encoding or a token for each amino acid. Schemes such as byte-pair encoding can be particularly useful when more complete sequence data, such as single-cell data, is available, for example when on the order of hundreds of thousands or even millions of sequences are used to train the model.

[0082] In total, 7986 tokens were constructed in the heavy chain vocabulary, while the light chain vocabulary had 6452 tokens. In this example, a "sentence" is a tokenized representation of a heavy or light chain sequence. Each sentence is a unique token. <sos>followed by a token representing a V gene of a heavy or light chain, a duplicated 3-mer token, a J gene token, and then a special token <eos>In any sentence with fewer tokens than the maximum lengths of 34 and 24 for the heavy and light chains, respectively, the sequence is split into a special <pad>We padded with tokens. Matchmaker makes no predictions if the sequence contains unseen V genes, J genes, novel 3-mers, or if the junction is longer than 30 amino acids. In fact, J gene diversity is much lower than V gene diversity, and therefore there were no instances of unseen J genes.

[0083] Light Chain Pairing Method Five strategies for light chain prediction using only the heavy chain as input were investigated, which are shown in Figure 3 and further detailed below. The strategies included: - a Transformer architecture that uses heavy chain tokenization and light chain tokenization transformation to sequence (referred to as "Matchmaker" in FIG. 3A and below); - a GRU (Gated Recurrent Unit) model with heavy chain inputs, using the same vocabulary as the Transformer, and also providing a light chain output (Figure 3B); GRUs, and more broadly other Recurrent Neural Network (RNN) architectures such as LSTM (Long Short-Term Memory Networks) with Attention, were the "state of the art" for neural machine translation, before Transformers became popular for this task; - Database search workflow (Figure 3C); - a variation of the database search method called "random search"; and - A frequency search method that matches ranked distributions of heavy and light chain read counts and then pairs chains with similar ranks (Figure 3D).

[0084] Four of these methods (Transformer, GRU, Database Search, Random Search) were newly developed as part of this work. The frequency search method is based on the method previously described in Reddy et al., 2010.

[0085] Matchmaker architecture and inference: Matchmaker was built using PyTorch (version 1.6.0). Matchmaker's hyperparameters and optimization procedure are similar to the sequence-to-sequence (Seq2Seq) transformer from Vaswani et al.

[2017] . Differences from the original transformer in Vaswani et al.

[2017] are described below. In summary, the model was made slightly smaller (few layers) for better constraining and trained using a different optimization technique to improve training. Matchmaker has 4 encoder layers and 4 decoder layers with feedforward dimension 1024 and dropout 0.2. Layer normalization was applied within the residual block [Child et al., 2018; Xiong et al., 2020]. The model was optimized using AdamW with weight decay of 0.1. Gradient clipping was performed using an L2 norm of 1.0. Matchmaker has a total of 31.7M learnable parameters. Training was stopped when the validation loss did not improve over three epochs, and the model with the best validation loss was used.

[0086] Gated recurrent unit (GRU) neural network model: An alternative deep learning Seq2Seq model with attention mechanism [Bahdanau et al., 2015] was trained using two GRU networks [Cho et al., 2014]. Here, the encoder is a 4-layer bidirectional GRU with hidden dimension 1024, and the decoder is a 4-layer forward-only GRU with hidden dimension 1024. Other hyperparameters of the model, such as the dimension of the embedding layer, were matched as close as possible to Matchmaker. In total, this model has 131.8M learnable parameters. The encoder-decoder GRU model was trained in an identical manner to Matchmaker. For brevity, we refer to this architecture as the "GRU model".

[0087] GRU networks (and by extension recurrent neural networks) have a completely different mechanism for how they process sequences compared to Transformers. In short, Transformers use a set of "self-attention" mechanisms that make them not only faster but also more accurate, whereas recurrent nets do not have this at all.

[0088] Both the Matchmaker and GRU models used in this study provided a single prediction for each input strand. In particular, the models predicted each strand sequentially by outputting at each subsequent position the probability of all possible tokens at the current position, selecting the token with the highest probability at the current position, and then moving on to the next position. Other implementations are possible and contemplated. For example, multiple tokens can be considered at each position simultaneously, and a combination of tokens across multiple positions can be selected that optimizes the global probability across multiple positions (e.g., the entire sequence, i.e., all positions, etc.). This can be performed using a beam search approach or related approaches such as beam stack search [Zhou & Hansen, 2005] and depth-first beam search [Furcy & Koenig, 2005]. Beam search is a heuristic search algorithm that explores a graph by expanding the most promising nodes in a limited set. After reaching a predefined maximum depth, the solution with the highest probability can be output. Database search method: Heavy chains were paired with sequences homologous to known heavy-light chain pairs in the Matchmaker training set. For a query heavy chain sequence, only heavy-light chain pairs with a matching heavy chain V gene to the query are selected. From this subset, two pairs are selected. The first pair is the "closest" light chain, which is from the pair with the closest heavy chain junction amino acid sequence to the query. The second pair is the "top" light chain, which is from the pair with the light chain V gene most frequently associated with the V gene of the query heavy chain. The closest light chain is used if there is 65% or more identity to the query junction amino acid sequence or 75% or more germline V gene identity between the closest light chain and the top light chain. Otherwise, the top light chain is used. The rationale for distinguishing between these two cases is that if there is a sufficiently similar junction sequence in the search, the VL sequence from this can be used. If there is not one that is sufficiently similar, a cruder approach is taken in which the most common VL for that V gene is used. Different cutoffs may be used or even no cutoffs can be used (i.e., the closest light chain is used in all cases). In either case, this strategy is called the "database search" method.

[0089] Random search method: With this method, one light chain is randomly selected from the database without considering any characteristics of the heavy chain.

[0090] Frequency-based search method: As a baseline, the search for light chain pairs was performed using read counts. Since all datasets used in this study were pre-paired, the situation of having two bulk sequencing libraries was first emulated by splitting the paired sequences. Then, the read counts were aggregated on the heavy or light chain sequences. For example, considering the sequence HeavyA:LightA with 4 reads and the sequence HeavyB:LightA with 5 reads, the splitting and aggregation produces HeavyA with 4 reads, HeavyB with 5 reads, and LightA with 9 reads. The heavy and light chains are then ranked based on the total read counts (see Figure 3D). The heavy chains are then paired with the light chains with matching ranks.

[0091] Construction of complete light chain sequences: For all methods described above, the output is the light chain V gene identifier, the joining amino acid sequence, and the light chain J gene identifier (Figure 3). If the training set did not have the full-length sequence, the light chain V and J genes were replaced with their germline amino acid sequences.

[0092] Characterization of monoclonal antibody binding to cognate target antigens A calculated pairing of each therapeutic antibody was characterized for binding to each cognate antigen by indirect ELISA. Briefly, recombinant versions from commercial sources of antigens (see Table 2 below) were immobilized and antibodies were added in solution as a 10-point titration using a top concentration of 500 nM and a 3-fold dilution series. Anti-human IgG secondary antibody conjugated to HRP (horseradish peroxidase) was used to detect antibodies still bound after washing. After developing with TMB-ELISA substrate and stopping with sulfuric acid, the absorbance was read at 450 nm. Absorbance was then plotted against concentration using GraphPad Prism 9 (www.graphpad.com).

[0093] [Table 2]

[0094] Thermal stability of monoclonal antibodies The thermal stability of the monoclonal antibodies was measured in a thermal denaturation assay. In triplicate, each antibody was heated from 25°C to 95°C in the presence of SYPRO™ orange and the fluorescence was measured. The melting curve derivative was then plotted as the average of the three replicates. The temperature at which the fluorescence increased most rapidly was noted for each antibody.

[0095] Tandem Transformer Model In addition to the five light chain pairing methods mentioned above, we designed another method that can utilize additional training data. In particular, we built a tandem transformer model that includes two RoBERTa models [Liu et al., 2019; Rothe et al., 2020] that were pre-trained on (full-length) BCR heavy and light chain sequences (roughly 42.3 million heavy chain sequences and 15.3 million light chain sequences), respectively. Unlike the Matchmaker and GRU models, this model was trained using each amino acid as a token, thus allowing for complete coverage of the sequences. This pre-trained RoBERTa model trained on unpaired antibody sequences is referred to herein as "AntiBERTa" and the training procedure is described in Figure 9. The AntiBERTa model was trained in a similar style to the RoBERTa base (Liu et al., 2019), but with a smaller batch size of 768 and a peak learning rate of 10 -4 , as well as over 225,000 pre-training steps and 10,000 warm-up steps. For prediction, we used beam search (Sustkever et al., 2014) with a beamwidth of 3.

[0096] The two AntiBERTa models were combined as a sequence-sequence model, with the encoder and decoder each initialized as a copy of the AntiBERTa model using the Huggingface transformer library [Wolf et al., 2019]. The AntiBERTa-AntiBERTa model was then fine-tuned using a slightly larger dataset of paired sequences than that used in the MatchMaker and GRU models described above. In particular, the training data was expanded by introducing more sequences from the antigen empirical library (from the same sources as described previously, i.e., DeKosky et al. 2016 and DeKosky et al., 2015). In total, there were 171984 paired sequences. However, as before, the dataset here does not contain the complete heavy chain sequence and the complete light chain sequence. Therefore, the germline V and J gene annotations were used to infer the complete sequence. In the fine-tuning step, the model was trained for 20 epochs with a peak learning rate of 3 × 10 -5 and 5% warm-up. Parameters were shared between the encoder and decoder AntiBERTa.

[0097] The use of pre-trained so-called "checkpoints" has proven to be a powerful strategy in the context of NLP [Rothe et al., 2020]. In particular, the BERT-BERT architecture was shown to perform well for NMT in Rothe et al.

[2020] . Having shown that a transformer model can be used to tackle the heavy-light chain pairing problem as an NMT problem (see below), we hypothesize that the BERT-BERT architecture is also likely to perform very well, and that the RoBERTa-RoBERTa architecture may improve on this even further, as the RoBERTa model has also been shown to improve on the BERT model on a variety of NLP tasks.

[0098] Because the paired training data used to train the full model (including the pre-trained AntiBERTa model) does not include full-length BCR heavy and light chain sequences (as described above), an equivalent paired data set including full-length sequences is generated using one of two alternative approaches. In the first approach, full-length sequences are obtained by replacing the V and J gene identifiers with the corresponding germline sequences. In the second approach, the pre-trained AntiBERTa model (or any other such "checkpoint" model, such as the GPT-2 model) is used to predict full-length sequences of the training set independently (using each model) for heavy and light chains based on known chain pairs. Predictions from the "checkpoint" models may be obtained using some or all of the known chain pairs (e.g., gene segment identifiers, partial sequences, etc.) in combination with some information obtained from the germline sequence of any segment for which the complete sequence is not available (e.g., some of the amino acids of the segment, such as the identification of the first k amino acids of the segment, where k can be, for example, 1, 2, 3, 5, 10, etc.). In other words, an AntiBERTa model trained on full-length unpaired heavy and light chain data can be used to predict the full-length sequence of a heavy chain in the training data from the V gene identifier, J gene identifier, and junction sequence provided in the data. Similarly, an AntiBERTa model trained on full-length unpaired heavy and light chain data can be used to predict the full-length sequence of a light chain in the training data from the V gene identifier, J gene identifier, and junction sequence provided in the data. The same two approaches can be used to map any limited, paired training data to a more expanded form of data that may have been available for training the "checkout" model. Alternatively, the data used to train the "checkpoint" model can be converted to a limited form that matches the form of the paired training data. This can still benefit from the potential additional information gleaned by the pre-trained model from the vast number of unpaired sequences available.However, such unpaired sequence data may not fully utilize the breadth of information available.

[0099] result An NLP-inspired model generates light chains using only heavy chains as input The heavy-light chain pairing problem is framed as an NMT task to predict the light chain sequence given only the heavy chain sequence as input. In particular, a Seq2Seq transformer similar to [Vaswani et al., 2017] was implemented. The model is trained with a tokenized representation of the heavy chain sequence as input, as described in Methods (Figure 3A), and returns a tokenized representation of the light chain sequence as output. For the input heavy chain sequence, the model encoder layer computes self-attention between V genes, overlap junction k-mers, and J genes. The encoder's self-attention scores on the heavy chain sequence are then used by the decoder to autoregressively predict the light chain sequence. An example of decoder attention is shown in Figure 5, where each attention head focuses on a different subset of the heavy chain tokens to determine the output light chain tokens.

[0100] Figure 6 shows the prediction performance of all methods for light chain V gene prediction on the held-out test set and single cell blind test. Because the V gene represents the largest portion of the chain that determines antigen binding and therefore has a large influence on stability and to some extent on binding (which is also greatly influenced by the junction sequence), the V gene results are examined separately from the full prediction results. The light chain V gene prediction results are further discussed below. The number of correct complete light chains consisting of V genes, junction amino acid sequences, and J genes was much lower than the number of correct V genes across all methods. The full set of prediction results is in Tables 3 and 4.

[0101] [Table 3]

[0102] [Table 4]

[0103] Of the 18999 sequences in the test set, 18803 were predicted by all five prediction methods. The Transformer-based model predicted the correct light chain V genes in 2431 heavy chain sequences (12.9%), while the GRU model had 1786 correct predictions (9.5%). The use of a self-attention mechanism in Transformer means that the Transformer model should be better able to utilize information about the heavy chain sequences compared to the GRU model. This likely explains, at least in part, the higher performance of the Transformer-based model compared to the GRU model. The frequency-rank-based prediction method had the lowest predictions (658 correct V genes; 3.5%), which was worse than randomly picking light chains from the database (915 correct V genes; 4.9%).

[0104] The database search method was the most accurate, with 4940 correct V genes (26.3%, Figure 6A). However, as can be seen in Figure 6A, the database search method did not perform as well in the single-cell blind tests as it did in the test set. In fact, the transformer-based model outperformed the database search method in all single-cell blind tests. The high performance of the database search method in the test set is likely due to the presence of clonal relatives in the training and test sets, which are not the case in the blind set. In fact, the members of a single B-cell clone were split across the training and test sets (see Table 5). For example, of the 18803 heavy-light chain pairs for which predictions were made, 897 had the same heavy chain sequence as the training set. When allowing up to one mismatch in the junction amino acid sequence, 3178 sequences were found in the test set that were related to the heavy chains in the training set. Of these, 729 (23%) were paired with the same light chain sequence. These are more likely to be correctly paired by database search methods, while redundancy in the heavy chain repertoire in the training set could complicate the task of NLP-inspired models (similar to a translation model having to learn translations from language A to language B, where multiple sentences in language A are translated into a single sentence in language B). This bias in evaluations using test data can be mitigated by not splitting clones across training and test sets, i.e., by ensuring that members of the same B cell clone (cluster) are not split between training and test sets.

[0105] [Table 5]

[0106] Thus, evaluation on a test set gives a distorted view of the performance of a database searching method that is only realistic if the amount of paired sequence data available to perform such searches truly represents the expected diversity of the BCR repertoire (far from being the case in reality). In other words, evaluation on blind tests gives a much more realistic view of the performance comparison of methods.

[0107] To exclude the influence of clonal relatedness across training and test datasets, we validated Matchmaker in silico using known paired sequences from single-cell datasets from the Paired OAS database [Eccles et al., 2020; King et al., 2021; Setliff et al., 2019]. Since none of these blind test datasets shared individuals with the training set, any overlapping heavy chain sequences were likely public sequences rather than members of the same B-cell clone. None of the heavy chain sequences in the datasets from Eccles et al.

[2020] and Setliff et al.

[2019] matched the training set, while 16 heavy chain sequences in the King et al.

[2021] dataset were identical to the training set. The 0.05% overlap in heavy chain sequences was broadly consistent with previous observations of heavy chain convergence [Briney et al., 2019].

[0108] Across the three single-cell datasets, Matchmaker performed best, predicting up to 9.8% of heavy chain sequences with the correct light chain V genes (Figure 6A). Matchmaker was also able to predict the correct light chain V genes, junction sequences, and J genes for 105, 7, and 23 heavy chains in the King, Eccles, and Setliff datasets, respectively (Table 4). In comparison, the database search method had 59, 4, and 6 correct predictions. Although the GRU method was not as accurate as Matchmaker, it still outperformed the database search method in the two single-cell datasets. These results indicate that deep learning models can be more useful than light chain prediction, especially in situations where novel heavy chain sequences exist. The results further show that the advantage of deep learning models over other approaches is much more striking when looking at full-length sequences than when looking at only V gene sequences.

[0109] Thus, the data in Figure 6A and Tables 3 and 4 show that the two machine learning based methods (and the custom designed database search method) all outperformed the prior art (frequency based search)-like method by a large margin. Of these, the transformer based method performed best across all blind test datasets, and showed an even more striking performance advantage when looking at full length sequences. Both machine learning based methods also outperformed the database search method on two of the three blind test datasets.

[0110] Figure 6B shows the Levenshtein distance distribution of predicted light chain junction sequences in the King et al.

[2021] dataset. This shows that the transformer-based method (Matchmaker) has more predictions with shorter distances than any other method, and even when the predicted light chain junction amino acid sequences are not entirely correct, they tend to be closer than with any other method.

[0111] Interestingly, in the King et al.

[2021] dataset, Matchmaker generated 29 light chain sequence predictions that were not in the training set, and two of these had the correct light chain V genes. However, in these novel light chains, the prediction of the light chain junction sequence was less successful. The average Levenshtein distance between the 29 predicted light chain junction amino acid sequences and the true amino acid sequence was 8 amino acids.

[0112] Experimental Verification Due to practical limitations, only the best performing machine learning based method (Matchmaker) and database search methods mentioned above were experimentally validated. As a final validation of the methodology, these methods were used to predict the light chains in nine therapeutic antibodies (Table 6). These nine were chosen because their targets are well established and the heavy chains have different levels of influence on antigen binding. This was quantified by calculating the buried surface area of ​​the heavy chain by examining the molecular structures of native antibodies available in the Protein Data Bank. This information was only available for 8 / 9 antibodies in the table. For cetuximab, which is of murine, not human, origin, ANARCI [Dunbar and Deane., 2016] was used to infer the closest human V and J genes.

[0113] [Table 6]

[0114] All nine light chain predictions from the database search resulted in stable antibodies, while one of the nine Matchmaker pairings failed and one was not tested because it could not be expressed initially (see Table 7). With the exception of the Matchmaker paired form of Nivolumab, all other antibodies from either method produced thermostable antibodies. Pertuzumab with the Matchmaker light chain (IGKV1-39) was expressed but the pairing did not bind to the target, despite its similarity to the native light chain (IGKV1-33). However, there are a few positions at the heavy chain interface that differ between the predicted and native light chain sequences. For example, the native Pertuzumab light chain has an alanine at position L40, while the Matchmaker predicted light chain has an asparagine. Thus, both methods (database search prediction and transformer-based prediction) were able to produce stable pairs in most or all cases. Based on the relatively poor performance of prior art inspired frequency-based methods (which, as noted above, performed worse than random database searches), this is likely to be a better result than could have been obtained using frequency-based methods.

[0115] [Table 7]

[0116] Each of the 27 antibodies (9 native, 9 Matchmaker, 9 database) was assayed by indirect ELISA for binding to each antigen. None of the predictions had sequences identical to the true sequence, and therefore none of them recovered the full binding affinity of the original antibody. This was expected, since the sequences of these therapeutic antibodies are the result of intensive genetic engineering towards affinity optimization (sometimes not based on sequences of human origin), whereas the prediction methods are based on naturally occurring sequences. However, the antibodies from Matchmaker pairing showed signs of binding in 3 / 9 cases, whereas the antibodies generated by database searching showed binding in 1 / 9 (Table 8, Figure 8). Thus, the data show that the Matchmaker method is capable of generating a large proportion of pairs that show sufficient binding affinity to form a good basis for further genetic engineering, a step that, until now, was a major bottleneck in the process of providing functional antibodies from bulk sequencing of the BCR repertoire.

[0117] Furthermore, it is expected that the performance of this method can be further enhanced using additional training data that reflects how mutated and / or engineered antibody sequences, such as the therapeutic antibodies used herein, pair. In other words, expanding the training data to allow the model to learn from more data and / or from data that includes optimized sequences is expected to further improve the model's ability to predict the functional pairing of native and engineered antibodies.

[0118] [Table 8]

[0119] Interestingly, Matchmaker predicted the correct light chain V gene for panitumumab, but the synthesized antibody did not show binding. This highlights that even predicting the correct V gene does not necessarily rescue the function of the BCR. Indeed, none of the pairings predicted by Matchmaker or database searching methods that showed binding predicted the correct V gene. The determinants of functional antibody pairs are poorly understood. As mentioned above, the V genes form the majority of the binding region, so the identification of the correct V gene provides an indication that the pairing is likely to be stable and at least a promising starting point for identifying binders. However, incorrect junctions can greatly affect the final binding affinity of the pairing, leading to binding even when slightly different V genes are present, or conversely, to no binding even when the correct V gene is present. These results, combined with the results in Figures 6A and 6B and Table 4, show that the machine learning-based approach described herein is capable of learning the complex features underlying stability and binding involved in the complex interactions between different parts of the light chain sequence (where a strict V gene sequence match may not lead to as good a binding affinity as a non-strict match combined with a more suitable junction sequence). This capability, although not complete, clearly outperforms the newly described non-deep learning-based method (database search) when evaluated experimentally. Furthermore, results from in-silico studies show that the machine learning-based approach described herein outperforms the current state-of-the-art frequency-based approach by an even larger margin.

[0120] The strongest binder in the Matchmaker predictions was the Pertuzumab heavy chain paired with the database light chain. However, the Pertuzumab interaction was driven primarily by the heavy chain, which may indicate that any stable light chain may be capable of binding. Conversely, the Matchmaker light chains of Nivolumab and Cetuximab were linked to binders, despite the relatively low importance of the heavy chain.

[0121] As further validation and evidence of the clinical utility of the methods described herein, the Matchmaker model described above was used to identify heavy-light chain pairings of heavy chain sequences identified in COVID-19 patients as likely to participate in immune responses and bind to coronavirus spike proteins. In particular, 18 heavy chain sequences identified from COVID-19 patients with these properties, selected from the data of Galson et al. (2020), were provided as input to the method for pairing with light chains. The pairings were expressed and assayed for binding to the Wuhan strain spike antigen using homogeneous time-resolved fluorometry (HTRF). All 18 antibodies were successfully expressed, and 9 of the 18 antibodies were confirmed as binding to the spike protein (binding defined as delta F%>100, where delta F represents the percentage increase in the fluorescent readout compared to the negative control). Thus, 50% of Matchmaker predictions lead to functional antibodies.

[0122] Tandem Transformer Model The structure of the GRU-based and transformer-based models described herein imposes some limitations on the amount of available training data, since paired heavy-light chain data is required for training. This type of data is available in relatively limited amounts, and its content is further restricted by providing V and J gene identifiers instead of complete sequences (thereby effectively restricting predictions to germline sequences of these segments). To circumvent these limitations, a model was constructed that includes two transformers, each trained on a much larger data set of unpaired heavy and light chain sequences (42.3 million and 15.3 million sequences, respectively). The results are shown in Table 9. Although the V gene prediction results were slightly lower than Matchmaker, the number of correct predictions with the correct V genes, junction amino acid sequences, and J genes was higher (compare Table 9 with Tables 5, 6). The data show that this machine learning-based method also outperforms a method similar to the prior art (frequency-based search). The data further indicate that the tandem transformer approach has the potential to further improve upon transformer-based methods, at least when looking at performance over full-length sequences, where the transformer-based approach has already shown a striking performance advantage over other methods. It is likely that the performance of this approach will be further increased with additional training data and training time (limited here for practical reasons).

[0123] [Table 9]

[0124] Consideration This example describes a machine learning, NLP-inspired approach to the BCR heavy-light chain pairing problem. In particular, we describe two architectures that consider the BCR heavy-light chain pairing problem as an NMT task: Matchmaker, the Seq2Seq transformer model, and the Seq2Seq GRU-based model. To the best of our knowledge, this work is the first application of a deep artificial neural network (ANN), specifically the Seq2Seq model, to BCR heavy chain pairing, which requires only heavy chain sequences as input. Matchmaker is the first application of the transformer model for this purpose. The deep learning-based approach described herein offers the benefit of covering the BCR repertoire as deeply as possible while also eliminating the need for bulk light chain sequencing. Furthermore, the approach is able to learn general features of pairings from a set of training data and use this learning to predict pairings of previously unseen chains. Conversely, approaches such as database search approaches tend to stall quickly when they see query chains that are not present in known pairings or that are more distant from chains present in known pairings. This is likely to be advantageous in many cases, given the enormous diversity of the BCR repertoire, but particularly in the context of applications such as identification of specific or other rare antibodies that may underlie a desired phenotype in an individual.

[0125] The method was benchmarked against several alternative strategies. On the held-out test set of heavy-light chain pairs, both deep ANNs (especially Matchmaker) performed reasonably well, but the database search method had the highest accuracy. Given that associates of the same B cell clone were found across the training and held-out test sets, it is not surprising that the database search method had the highest accuracy. Indeed, if obtaining a sufficiently large knowledge base of heavy-light chain pairs was a realistic prospect, it would aid in the identification of the correct light chain counterpart. However, blind tests showed that the accuracy of the database search was poor when no relevant heavy chain sequence was available. The frequency search was found to perform the worst across all test situations, with even random selection being slightly better. This is lower than expected, confirming the limitations of using light chain frequency as a means of pairing.

[0126] An interesting observation in the blind test was that Matchmaker was able to predict completely novel light chains, i.e., the predicted light chain sequence was not in the training data. This was also true for the GRU model, which predicted one novel light chain with a fairly low Levenshtein distance to the true sequence (distance = 4). Only 29 / 29785 heavy chains had novel light chain predictions by Matchmaker, suggesting that the deep learning model learned some rules of heavy-light chain pairing by an attention mechanism as opposed to memorizing the training set. With higher quality data consisting of complete heavy and light chain sequences, it is expected that deep ANN models such as Matchmaker will learn improved identification of heavy-light chain pairings. Taking those rules into account, Matchmaker should generate other novel light chains with higher accuracy than those observed in this study.

[0127] In a set of nine experimentally validated positive control therapeutic antibodies, Matchmaker generated more binders than database search methods. None of the predicted antibodies from either strategy were as strong as the native antibodies, and the majority of predictions did not lead to binders, highlighting the difficulty of the problem. Variation in binding strength appeared to reflect heavy chain contributions to binding, but Matchmaker did not appear to be affected by these differences. Although the precise relationship between light chains and binding remains unclear, the fact that Matchmaker generates binding antibodies provides a starting point for further antibody engineering opportunities. In other words, the method is capable of predicting binders for a percentage of heavy chains and represents a significant improvement over the prior art, at least because: (a) it requires only heavy chain sequences as input, (b) it is not limited in terms of the type of sequences and datasets that can be used as input (i.e., it is expected to be able to provide useful predictions for any type of sequence, not just highly over-represented sequences in clonal dominated samples), and (c) it has a higher binder prediction hit rate than other methods (even the newly described method herein shown to have a higher predictive performance than the current state of the art). Any such binders that are so successfully obtained represent an improvement over the prior art and can be used as a promising starting point for further affinity improvements. Further validation using heavy chain sequences from COVID-19 patients indicates that Matchmaker's performance may be even higher than that shown based on this data, since 50% (9 / 18) of the pairings generated with the limited set of therapeutic antibodies tested showed some binding to the target. The improvements observed over comparable methods are not only significant in themselves, but are particularly significant when one considers that, given the diversity of the BCR repertoire, identification of functional heavy-light chain pairings is an extremely complex problem such that any improvement that leads to sufficiently viable candidates for further development is of paramount practical importance.This may be especially true in the context of applications such as therapeutic antibody discovery from samples from subjects, where even small improvements in search efficiency (and this is by no means a small improvement) can lead to faster results and / or less wasted effort and resources in therapeutic antibody discovery, testing, optimization, and development. Furthermore, the present model was trained using a training set of 153889 heavy-light chain pairs, which is quite small compared to the datasets typically used for training transformer-based models (usually on the order of millions of sequences).

[0128] There are several possible improvements to the Matchmaker implementation. For example, the model described above predicts tokens greedily (i.e., one token at a time, i.e., the most likely token at each individual position is predicted). Using a strategy that considers multiple (e.g., three) positions simultaneously, such as beam search [Sustkever et al., 2014], should help increase prediction accuracy, as the model explores more solutions and is therefore less likely to be "stuck" on a suboptimal solution. This is the approach used in the Tandem Transformer model described above. Additionally, the Matchmaker and GRU models exemplified above use a combination of overlapping fixed-size k-mers and gene identifiers to encode the BCR sequence. Using single amino acids or byte pairs of amino acids as tokens is expected to further improve the model. For example, the Tandem Transformer model described above used a single token for each amino acid. With byte pair encoding, several non-overlapping amino acids are encoded as tokens using a dictionary that is automatically defined in a data-driven manner.

[0129] A limitation of the training set is the lack of full-length heavy and light chain sequences. Indeed, such data is currently only available in limited amounts, and therefore a larger dataset was used that includes only J gene identifiers, V gene identifiers, and junction sequences. An alternative tokenization scheme was designed to address this gap, but effectively predicts only germline light chain sequences of V and J genes (as that is all of the information available in the training data). Because CDRL1 and CDRL2 loops can form contacts with CDRH3 [Leem et al., 2016], the predicted light chains may miss any mutant amino acids that form crucial interactions with CDRH3 in vivo. This can be overcome by using different paired heavy-light chain data for training, providing full-length sequences, such as data based on cell barcoding, such as from the 10x Genomics platform. This can also be addressed using the tandem transformer approach described above, where a combination of a larger training dataset of unpaired full-length sequences (to train the checkpoint model) and a smaller paired dataset with potentially narrower sequence coverage can be used to train the model.

[0130] To apply bulk heavy chain repertoire analysis to antibody discovery, the light chain pairing problem remains applicable. The deep learning-based approach described herein is unique in relying only on heavy chains as input. In particular, the Matchmaker (transformer-based) model has the highest in silico accuracy based on multiple metrics and has been validated in vitro for the generation of functional antibodies. This allows Matchmaker to predict the light chain on its own to predict light chains when paired light chain information is not available. Thus, this approach has the potential to fill the gap in light chain pairing information and enable therapeutic antibody discovery and a better understanding of the immune system.

[0131] Finally, although the method has been described in the context of identifying light chain pairings for a heavy chain query, it is also applicable to the inverse problem of identifying heavy chain pairings for a light chain query, which is a less frequent problem since heavy chain sequencing is more common and the heavy chain is likely to play a more significant role in determining specificity and affinity.

[0132] References Vander Heiden et al., 2017. Dysregulation of B Cell Repertoire Formation in Myasthenia Gravis Patients Revealed through Deep Sequencing. J Immunol. 2017 Feb 15;198(4):1460-1473. Bashford-Rogers et al, 2019. Analysis of the B cell receptor repertoire in six immune-mediated diseases. Nature volume 574, pages122-126(2019). Nielsen et al., 2020. Human B Cell Clonal Expansion and Convergent Antibody Responses to SARS-CoV-2. bioRxiv. Preprint. 2020 Jul 9. doi: 10.1101 / 2020.07.08.194456. Simonich et al., 2019. Kappa chain maturation helps drive rapid development of an infant HIV-1 broadly neutralizing antibody lineage. Nature Communications volume 10, Article number: 2190 (2019). Krawczyk et al., 2019. Looking for therapeutic antibodies in next-generation sequencing repositories. mAbs. Volume 11, 2019 - Issue 7, Pages 1197-1205. Galson et al., 2020. Deep Sequencing of B Cell Receptor Repertoires From COVID-19 Patients Reveals Strong Convergent Immune Signatures. Front. Immunol., 15 December 2020. doi.org / 10.3389 / fimmu.2020.605170. Mora and Walczak, 2019. How many different clonotypes do immune repertoires contain? Current Opinion in Systems Biology. Volume 18, December 2019, Pages 104-110 Kovaltsuk et al., 2018. Observed Antibody Space: A Resource for Data Mining Next-Generation Sequencing of Antibody Repertoires. J Immunol October 15, 2018, 201 (8) 2502-2509. Tiller et al., 2013. A fully synthetic human Fab antibody library based on fixed VH / VL framework pairings with favorable biophysical properties. MAbs. 2013 May 1; 5(3): 445-470. Teplyakov et al., 2016. Structural diversity in a human antibody germline library. MAbs. Aug-Sep 2016;8(6):1045-63. Glanville et al., 2009. Precise determination of the diversity of a combinatorial antibody library gives insight into the human immunoglobulin repertoire. PNAS December 1, 2009 106 (48) 20216-20221. Jayaram et al., 2012.Germline VH / VL pairing in antibodies. Protein Engineering, Design and Selection, Volume 25, Issue 10, October 2012, Pages 523-530. Ling et al., 2018. Effect of VH-VL Families in Pertuzumab and Trastuzumab Recombinant Production, Her2 and FcγIIA Binding. Front. Immunol., 12 March 2018. doi.org / 10.3389 / fimmu.2018.00469 DeKosky et al., 2013. High-throughput sequencing of the paired human immunoglobulin heavy and light chain repertoire. Nature Biotechnology volume 31, pages166-169(2013). DeKosky et al., 2015. In-depth determination and analysis of the human paired heavy- and light-chain antibody repertoire. Nature Medicine volume 21, pages86-91(2015). DeKosky et al., 2016. Large-scale sequence and structural comparisons of human naive and antigen-experienced antibody repertoires. PNAS May 10, 2016 113 (19) E2636-E2645. King et al., 2021. Single-cell analysis of human B cell maturation predicts how antibody class switching shapes selection dynamics. Science Immunology 12 Feb 2021. Vol. 6, Issue 56, eabe6291 Eccles et al., 2020. T-bet+ Memory B Cells Link to Local Cross-Reactive IgG upon Human Rhinovirus Infection. Cell Reports Volume 30, Issue 2, 14 January 2020, Pages 351-366.e7 Setliff et al., 2019.High-Throughput Mapping of B Cell Receptor Sequences to Antigen Specificity. Cell Volume 179, Issue 7, 12 December 2019, Pages 1636-1646.e15 Reddy et al., 2010. Monoclonal antibodies isolated without screening by analyzing the variable-gene repertoire of plasma cells. Nature Biotechnology volume 28, pages965-969(2010). Zhu et al., 2013. Mining the antibodyome for HIV-1-neutralizing antibodies with next-generation sequencing and phylogenetic pairing of heavy / light chains. PNAS. 2013 Apr 16;110(16):6470-5. Raybould et al., 2021. Public Baseline and shared response structures support the theory of antibody repertoire functional commonality. PLoS Comput Biol 17(3): e1008781. Rakocevic et al., 2021.The landscape of high-affinity human antibodies against intratumoral antigens. bioRxiv. 8 Feb 2021. doi.org / 10.1101 / 2021.02.06.430058 Vaswani et al., 2017. Attention Is All You Need. arXiv:1706.03762 Devlin et al., 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv:1810.04805 Radford et al., 2019.Language Models are Unsupervised Multitask Learners. https: / / openai.com / blog / better-language-models / Liu et al., 2019.RoBERTa: A Robustly Optimized BERT Pretraining Approach. arXiv:1907.11692 Rothe et al., 2020.Leveraging Pre-trained Checkpoints for Sequence Generation Tasks. arXiv:1907.12461 Bahdanau et al., 2015. Neural machine translation by jointly learning to align and translate. arXiv:1409.0473 Cho et al., 2014. Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling. arXiv:1412.3555 Child et al., 2019. Generating Long Sequences with Sparse Transformers. arXiv:1904.10509 Xiong et al., 2020. On Layer Normalization in the Transformer Architecture. arXiv:2002.04745 Dunbar and Deane, 2016. ANARCI: antigen receptor numbering and receptor classification. Bioinformatics. 2016 Jan 15;32(2):298-300. Sutskever et al., 2014. Sequence to Sequence Learning with Neural Networks.arXiv:1409.3215 Rees, 2020. Understanding the human antibody repertoire. MAbs. Jan-Dec 2020;12(1):1729683. Leem et al., 2016. ABodyBuilder: Automated antibody structure prediction with data-driven accuracy estimation. MAbs. 2016 Oct;8(7):1259-1268. Ye et al., 2013. IgBLAST: an immunoglobulin variable domain sequence analysis tool. Nucleic Acids Res. 2013 Jul;41(Web Server issue):W34-40. Furcy, David. Koenig, Sven."Limited Discrepancy Beam Search”. IJCAI’05: Proceedings of the 19th international joint conference on Artificial intelligence. July 2005 Pages 125-131. Zhou, Rong. Hansen, Eric."Beam-Stack Search: Integrating Backtracking with Beam Search”. Conference: Proceedings of the Fifteenth International Conference on Automated Planning and Scheduling (ICAPS 2005), June 5-10 2005. Wolf et al., 2019. HuggingFace’s Transformers: State-of-the-art Natural Language Processing. arXiv:1910.03771 Carter Jason A., Preall Jonathan B., Grigaityte Kristina, Goldfless Stephen J., Jeffery Eric, Briggs Adrian W., Vigneault Francois, Atwal Gurinder S."Single T Cell Sequencing Demonstrates the Functional Role of αβ TCR Pairing in Cell Lineage and Antigen Specificity .”Frontiers in Immunology. Vol. 10. 2019, p. 1516.Zheng GXY, Terry JM, Belgrader P, Ryvkin P, Bent ZW, Wilson R, et al. Massively parallel digital transcriptional profiling of single cells. Nat Commun. (2017) 8:14049. Howie B, Sherwood AM, Berkebile AD, Berka J, Emerson RO, Williamson DW, et al. High-throughput pairing of T cell receptor α and β sequences. Sci Transl Med. (2015) 7:301ra131. Eve Richardson, Jacob D. Galson, Paul Kellam, Dominic F. Kelly, Sarah E. Smith, Anne Palser, Simon Watson & Charlotte M. Deane (2021) A computational method for immune repertoire mining that identifies novel binders from different clonotypes, demonstrated by identifying anti-pertussis toxoid antibodies, mAbs, 13:1. Yi-Chun Hsiao, Yonglei Shang, Danielle M. DiCara, Angie Yee, Joyce Lai, Si Hyun Kim, Diego Ellerman, Racquel Corpuz, Yongmei Chen, Sharmila Rajan, Hao Cai, Yan Wu, Dhaya Seshasayee & Isidro Hoetzel (2019) Immune repertoire mining for rapid affinity optimization of mouse monoclonal antibodies, mAbs, 11:4, 735-746. Warszawski S, Borenstein Katz A, Lipsh R, Khmelnitsky L, Ben Nissan G, Javitt G, et al. (2019) Optimizing antibody affinity and stability by the automated design of the variable light-heavy chain interfaces. PLoS Comput Biol 15(8): e1007207. Seeliger D, Schulz P, Litzenburger T, Spitz J, Hoerer S, Blech M, Enenkel B, Studts JM, Garidel P, Karow AR. Boosting antibody developability through rational sequence optimization. MAbs. 2015;7(3):505-15. doi: 10.1080 / 19420862.2015.1017695. Mason, D.M., Friedensohn, S., Weber, C.R. et al. Optimization of therapeutic antibodies by predicting antigen specificity from antibody sequence via deep learning. Nat Biomed Eng (2021). Galson Jacob D., Schaetzle Sebastian, Bashford-Rogers Rachael J. M., Raybould Matthew I. J., Kovaltsuk Aleksandr, Kilpatrick Gavin J., Minter Ralph, Finch Donna K., Dias Jorge, James Louisa K., Thomas Gavin, Lee Wing-Yiu Jason, Betley Jason, Cavlan Olivia, Leech Alex, Deane Charlotte M., Seoane Joan, Caldas Carlos, Pennington Daniel J., Pfeffer Paul, Osbourn Jane. Deep Sequencing of B Cell Receptor Repertoires From COVID-19 Patients Reveals Strong Convergent Immune Signatures. Frontiers in Immunology. Vol. 11, 2020.

[0133] All references cited in this specification are herein incorporated by reference in their entirety to the same extent as if each individual published patent, patent, or patent application was specifically and individually indicated to be incorporated by reference in its entirety.

[0134] The specific embodiments described herein are provided by way of example and not by way of limitation.Various modifications and variations of the compositions, methods, and uses of the techniques described will be apparent to those skilled in the art without departing from the scope and spirit of the techniques described.Any subtitles herein are included for convenience only and should not be construed as limiting the present disclosure in any way.

[0135] The methods of any embodiment described herein may be provided as a computer program, or as a computer program product, or as a computer readable medium having a computer program configured to perform the above-mentioned method when executed on a computer.

[0136] Unless otherwise indicated by context, the above feature descriptions and definitions are not limited to any particular aspect or embodiment of the present invention, but apply equally to all aspects and embodiments described.

[0137] Throughout this specification and the claims, the following terms shall take the meanings specifically associated therewith, unless the context clearly indicates otherwise. As used herein, the phrase "in one embodiment" may refer to the same embodiment, but not necessarily to the same embodiment. Additionally, as used herein, the phrase "in another embodiment" may refer to different embodiments, but not necessarily to different embodiments. Thus, as described below, various embodiments of the invention can be readily combined without departing from the scope or spirit of the invention.

[0138] It must be noted that, as used in this specification and the appended claims, the singular forms "a," "an," and "the" include the plural forms unless the context clearly dictates otherwise. Ranges may be expressed herein as from "about" one particular value and / or to "about" another particular value. When such a range is expressed, another embodiment includes from the one particular value and / or to the other particular value. Similarly, when values ​​are expressed as approximations, by use of the preceding "about," it will be understood that the particular value forms another embodiment. The term "about" in connection with numerical values ​​is optional and may mean, for example, + / - 10%.

[0139] Throughout this specification, including the claims which follow, unless the context otherwise requires, the words "comprise", "including", and variations such as "comprises", "comprising", and "including", will be understood to imply the inclusion of stated integers or steps or groups of integers or steps, but not the exclusion of any other integers or steps or groups of integers or steps.

[0140] Other aspects and embodiments of the invention provide the above-mentioned aspects and embodiments in which the term "comprising" is replaced with the term "consisting of" or "consisting essentially of", unless the context indicates otherwise.

[0141] "And / or," as used herein, should be construed as a specific disclosure of each of the two specified features or components with or without the other. For example, "A and / or B" should be construed as a specific disclosure of (i) A, (ii) B, and (iii) each of A and B, as if each were individually described herein.

[0142] The features disclosed in the above description, or in the following claims, or the accompanying drawings, which may be expressed in a particular form or in terms of means for performing a disclosed function or a method or process for obtaining a disclosed result, may be used separately or in any combination of such features to realize the invention in its various forms.< / pad> < / eos> < / sos>

Claims

1. 1. A computer-implemented method for identifying an antigen binding protein comprising a pair of a first chain and a second chain, comprising: providing a query sequence comprising a first strand sequence; providing the query sequence to a deep learning model configured to take as an input a query first strand sequence and generate as an output at least one corresponding second strand sequence, thereby identifying corresponding second strand sequences of the query sequence. and identifying a corresponding second strand sequence by:

2. The antigen-binding protein comprises (i) a heavy-light chain pair, wherein the first chain sequence is a heavy chain sequence or a light chain sequence and the corresponding second chain sequence is a light chain sequence or a heavy chain sequence, optionally wherein the first chain sequence is a heavy chain sequence and the corresponding second chain sequence is a light chain sequence; or (ii) an αβ chain pair, wherein the first chain sequence is a β chain sequence or an α chain sequence and the corresponding second chain sequence is an α chain sequence or a β chain sequence, optionally wherein the first chain sequence is a β chain sequence and the corresponding second chain sequence is an α chain sequence; or (iii) a γδ chain pair, wherein the first chain sequence is a δ chain sequence or a γ chain sequence and the corresponding second chain sequence is a γ chain sequence or a δ chain sequence, optionally wherein the first chain sequence is a δ chain sequence and the corresponding second chain sequence is a γ chain sequence; The method of claim 1 , comprising:

3. The deep learning model is a sequence-sequence transformer based model, and / or the deep learning model includes a recurrent neural network or a transformer; and / or the deep learning model is configured to generate as output one or more corresponding second strand sequences, each corresponding second strand sequence being associated with a confidence metric, optionally the confidence metric being a probability.

4. 3. The method of claim 1 or 2, wherein the training first chain sequences and the corresponding second chain sequences from a known chain pair comprise paired training heavy and light chain sequences from single B-cell sequencing data, or the training first chain sequences and the corresponding second chain sequences from a known chain pair comprise paired training alpha and beta chain sequences from single T-cell sequencing data.

5. the query chain sequence comprises or consists of a V gene sequence or identifier, a J gene sequence or identifier, and a junction sequence, and optionally a D gene sequence or identifier; 3. The method of claim 1 or 2, wherein the corresponding second chain sequence comprises or consists of a V gene sequence or identifier, a J gene sequence or identifier, and a joining sequence.

6. the query chain sequence comprises or consists of a V gene sequence or identifier, a J gene sequence or identifier, and a junction sequence; 3. The method of claim 1 or 2, wherein the corresponding second chain sequence comprises or consists of a V gene sequence or identifier, a J gene sequence or identifier, and a joining sequence, and optionally a D gene sequence or identifier.

7. 3. The method of claim 1 or 2, wherein the query sequence comprises or consists of one or more first chain CDR sequences and / or the corresponding sequence comprises or consists of one or more corresponding chain CDR sequences, optionally wherein the query / corresponding sequence comprises or consists of a CDR3 sequence.

8. The method of claim 1 or 2, wherein all sequences are amino acid sequences.

9. providing the query sequence to the deep learning model includes encoding the query sequence using an encoding scheme; Each genetic sequence identifier corresponds to an individual token, each amino acid corresponds to an individual token, and / or the sequences are encoded using tokens, each corresponding to an individual k-mer, or using a byte pair encoding, optionally each sequence being encoded using overlapping k-mers; and / or wherein identifying the corresponding second strand sequence includes decoding the corresponding sequence output by the deep learning model using an encoding scheme; Each genetic sequence identifier corresponds to an individual token, each amino acid corresponds to an individual token, and / or 3. The method of claim 1 or 2, wherein the sequences are encoded using tokens each corresponding to an individual k-mer or using a byte pair encoding, optionally each sequence is encoded using overlapping k-mers.

10. 10. The method of claim 9, wherein the encoding scheme is predefined based on the content of the training strand sequences.

11. 3. The method of claim 1 or 2, wherein the query sequence and / or the corresponding sequence comprises one or more genetic sequence identifiers, the method further comprising replacing one or more genetic sequence identifiers with corresponding germline sequences.

12. 3. The method of claim 1 or 2, wherein the deep learning model is a transformer-based model including an encoder pre-trained using unpaired training first strand sequences and / or corresponding strand sequences, and a decoder pre-trained using unpaired training corresponding strand sequences and / or first strand sequences.

13. 13. The method of claim 12, wherein the unpaired training strand sequence comprises a full-length sequence in the variable region of the corresponding strand and / or the unpaired training strand sequence comprises a full-length sequence in the variable region of the first strand.

14. Providing a query sequence is Retrieving said query sequence from a user through a user interface, from a computing device, from a sequence retrieval means or a computing device associated with a sequence retrieval means, from a database or other computer readable medium; the query sequence is obtained by sequencing a sample comprising genetic material encoding an antigen binding molecule comprising the query sequence, optionally wherein obtaining the query sequence comprises performing B cell bulk sequencing of a sample comprising B cells, T cell bulk sequencing of a sample comprising T cells, or bulk sequencing of a sample comprising any other cells expressing an antigen binding molecule comprising the query sequence, or genetic material derived therefrom, such as a B cell receptor library or a T cell receptor library; The method according to any one of claims 1 to 13, comprising:

15. 13. A method for providing antigen-binding protein chain pairs for a plurality of query sequences, the method comprising: performing the method of claim 1 for each of the query sequences.

16. Optionally, the plurality of query sequences are heavy or light chain sequences obtained by bulk B cell repertoire sequencing.

17. 1. A method for providing an antigen binding protein having a desired property, comprising: providing one or more query sequences comprising first strand sequences, at least one of the one or more query sequences having a high likelihood of having the desired repertoire; 2. Identifying a corresponding strand sequence for each of the one or more query sequences using the method of claim 1; A method comprising:

18. 20. The method of claim 17, further comprising obtaining one or more candidate antigen binding proteins, each of which comprises one of the query sequence and one or more of the corresponding sequences derived therefrom, and optionally testing the one or more candidate antigen binding proteins for the desired property.

19. 1. A method for providing a tool for identifying an antigen binding protein comprising a pair of a first chain and a second chain, comprising: providing training data comprising training first sequences and corresponding second strand sequences from known first strand and corresponding strand pairs; training a sequence-sequence deep learning model using the training data to take a query first strand sequence as an input and generate at least one corresponding second strand sequence as an output; Includes Optionally, the method further comprises obtaining a vocabulary for encoding the training first strand sequences and a vocabulary for encoding the training corresponding second strand sequences, optionally wherein the vocabulary is obtained using an encoding scheme, where any gene sequence identifier corresponds to an individual token, and / or any sequence is encoded, at least in part, using tokens each corresponding to an individual amino acid, and / or any sequence is encoded, at least in part, using tokens each corresponding to an individual k-mer, or using a byte pair encoding. method.

20. A processor; A computer readable medium containing instructions which, when executed by the processor, cause the processor to perform the steps of the method of claims 1, 15, 17, and / or 19. Including, the system.

21. One or more computer readable media containing instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the method of claims 1, 15, 17, and / or 19.

Images