A method for predicting the binding relationship between peptides and TCR molecules based on a large-scale language model

Abstract

The present invention discloses a method for predicting the binding relationship between a polypeptide and a TCR molecule based on a large-scale language model, which includes the following steps: S100, obtaining a polypeptide sequence of a binding relationship to be tested and a CDR3 fragment sequence on a TCR molecule, and preprocessing the sequences to be tested to obtain a target input sequence; S200, feature encoding each amino acid residue in the target input sequence, and outputting an encoding matrix for each sequence; S300, inputting the encoding matrices into a bidirectional GRU module in a trained binding relationship prediction model, and outputting a global feature matrix for each sequence; S400, merging the global feature matrices of each pair of sequences, and inputting them into a large-scale language model RetNet in the trained binding relationship prediction model, and outputting a relationship matrix for each pair of sequences; S500, inputting the relationship matrix of each pair of sequences into the binding relationship prediction model, and outputting the binding probability of the polypeptide sequence of the binding relationship to be tested and the CDR3 fragment sequence on the TCR molecule.

Description

Technical Field

[0001] The present invention relates to the field of immunity and tumor treatment technology, and in particular to a method for predicting the binding relationship between polypeptides and TCR molecules based on a large-scale language model. Background Art

[0002] TCRs are membrane proteins on the surface of T lymphocytes that direct T cell immune responses by recognizing antigen fragments (typically presented by antigen-presenting cells). TCR molecules play a crucial role in clinical cancer treatment, vaccine development, and the treatment of autoimmune diseases. Studying the binding mechanism between antigens (peptides) and TCR molecules can advance personalized immunotherapy and innovative vaccine design, offering new hope for the treatment of some intractable diseases. TCRs are encoded by several gene segments, including, for most T cells, α and β chains. These gene segments are combined during T cell development through genetic recombination to form the complete TCR gene. This genetic recombination process results in enormous genetic diversity. Using deep learning methods to predict peptide-TCR binding can help identify cancer therapeutic targets, aid in diagnosis, clinical immunotherapy, and the development of new vaccines. This can significantly shorten treatment cycles and open up new possibilities for immunotherapy for cancer patients. The increasing amount of data currently presents more opportunities and possibilities for binding prediction tasks.

[0003] Several methods currently exist for predicting peptide binding to TCR molecules. Based on the type of data input they can predict, these tools can be categorized as specific and pan-specific. Specific tools can only establish individual prediction models for peptides for which binding data already exists; such models cannot predict the binding of novel peptides. Pan-specific tools can predict the binding of novel peptides not found in the training data. These tools characterize the binding of known and unknown peptides to TCRs by modeling the binding of a single peptide to a TCR or the binding of a pan-peptide to a TCR. However, these methods are currently less than ideal for predicting the binding of known peptides, making their expansion into clinical applications difficult. While pan-specific tools can predict the binding of unknown peptides, their effectiveness is poor. Therefore, new modeling methods are needed to improve the prediction of peptide-TCR binding and to aid immunotherapy.

[0004] Current methods for predicting peptide-TCR binding remain a challenge, mainly manifested in the following aspects: (1) The model cannot take into account the integrity and relevance of the sequence during the training process; the most commonly used modules in current tools are fully connected, convolutional neural networks (CNN), autoencoders, self-attention, etc., but the structure of these models also largely destroys the sequence integrity and relevance of the two input sequences. (2) Although there are many different methods for specificity and pan-specificity, the prediction results are still not ideal for existing epitope data. (3) The current model does not mine and visualize the binding preferences of each peptide, which is not conducive to clinical applications (such as modified TCR, mutant peptide vaccines, etc.) and further improvements. Summary of the Invention

[0005] (1) Technical issues to be resolved

[0006] Based on the above problems, the present invention provides a method for predicting the binding relationship between polypeptides and TCR molecules based on a large-scale language model. By encoding the features of polypeptides and TCR molecules and modeling them by fusion of bidirectional GRU and RetNet modules, the accuracy and efficiency of their binding relationship prediction are improved.

[0007] (2) Technical solution

[0008] In response to the above technical problems, the present invention provides a method for predicting the binding relationship between a polypeptide and a TCR molecule based on a large-scale language model, comprising the following steps:

[0009] S100, obtaining a polypeptide sequence of a binding relationship to be tested and a CDR3 fragment sequence on a TCR molecule, and preprocessing the sequences to be tested to obtain a target input sequence;

[0010] S200, performing feature encoding on each amino acid residue in the target input sequence, and outputting an encoding matrix for each sequence;

[0011] S300, inputting the encoding matrices into the bidirectional GRU modules in the trained binding relationship prediction model, and outputting the global feature matrix of each sequence;

[0012] S400, after merging the corresponding global feature matrices of each sequence, inputting them into the trained large-scale language model RetNet in the combined relationship prediction model, and outputting the relationship matrix of each pair of sequences;

[0013] S500: Input the relationship matrix of each pair of sequences into a binding relationship prediction model, and output the binding probability of the polypeptide sequence to be tested for binding relationship and the CDR3 fragment sequence on the TCR molecule.

[0014] Furthermore, step S200 includes:

[0015] According to 28 features extracted from the AAIndex database that have been successfully applied in the field of protein binding, the physical and chemical characteristics of each amino acid residue in the target input sequence are encoded.

[0016] Among them, the 28 features that have been successfully applied in the field of protein binding are CHOP780202, CIDH920103, CIDH920105, FAUJ880109, GEIM800106, KANM800102, KLEP840101, KRIW710101, PALJ810107, QIAN880123, RACS770103, RADA880108, ZIMJ68010 4. AURR980120, MUNV940103, NADH010104, FAUJ880111, FINA910104, GEIM800104, NADH010106, LIFS 790101, MEEJ800101, OOBM770102, GUYH850105, ROSM880102, SWER830101, ZIMJ680102, MIYS990104.

[0017] Furthermore, step S300 includes:

[0018] Input the encoding matrix of the peptide into the bidirectional GRU module to obtain the global feature matrix of the peptide;

[0019] The encoding matrix of the CDR3 segment on the TCR molecule is input into the bidirectional GRU module to obtain the global feature matrix of CDR3.

[0020] Furthermore, step S400 is specifically as follows:

[0021] The global feature matrices of the peptide and CDR3 segments were spliced ​​together;

[0022] The concatenated matrix is ​​input into a 24-layer RetNet module in series, and the relationship matrix of each pair of sequences is output.

[0023] Furthermore, step S500 is specifically as follows:

[0024] The relationship matrix of each pair of sequences is input into the binding relationship prediction model, and the binding probability of the polypeptide sequence to be tested for binding relationship and the CDR3 fragment sequence on the TCR molecule is output, specifically:

[0025] Input the relationship matrix into two layers of linear layer and ReLU layer connected in series;

[0026] After another linear layer and a Sigmoid layer, the value is mapped to [0, 1], indicating the binding probability between the input peptide and the TCR molecule.

[0027] The present invention also discloses a device for predicting the binding relationship between polypeptides and TCR molecules based on a large-scale language model, comprising at least one processor; and at least one memory in communication with the processor, wherein:

[0028] The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the above-mentioned method for predicting the binding relationship between polypeptides and TCR molecules based on a large-scale language model, including the following functional modules:

[0029] A data acquisition and preprocessing module, configured to execute step S100;

[0030] A feature encoding module, configured to execute step S200;

[0031] A feature representation module, configured to execute step S300;

[0032] Combined with the relationship information mining module, used to perform the step S400;

[0033] The prediction module is used to execute the step S500.

[0034] Furthermore, the feature representation module includes a bidirectional GRU module, which can independently learn the contextual information of each sequence, including the context of subsequences. The bidirectional GRU module comprehensively learns the information of different subsequences along the sequence direction, ensuring the integrity and sequentiality of information during model training.

[0035] Furthermore, the combined relational information mining module includes a RetNet module, which has 24 layers connected in series. The crucial recursive representation module in the RetNet module is a variant of self-attention, which introduces a D matrix and a GN module. The exponential decay factor Y in the D matrix enables RetNet to retain the semantic influence of sequence proximity during the learning process. In addition, the GN module introduces nonlinearity to facilitate parallel training of modules. Essentially, the addition of the RetNet module significantly enhances the model's understanding of sequences and increases training speed. Compared with traditional Transformer positional encoding, this matrix is ​​more consistent with sequence processing in terms of position perception and is more suitable for the task of predicting peptide-TCR binding.

[0036] Furthermore, after the RetNet module outputs the relationship matrix for each pair of sequences, it enters the prediction module in sequence through three linear layers, two ReLU activation layers, and one random downsampling. Finally, the result passes through a Sigmoid activation layer, mapping the value to [0, 1], representing the binding probability between the input peptide and the TCR molecule.

[0037] The present invention also discloses a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions enable the computer to execute the above-mentioned method for predicting the binding relationship between polypeptides and TCR molecules based on a large-scale language model.

[0038] (3) Beneficial effects

[0039] The above technical solution of the present invention has the following advantages:

[0040] The prediction method provided by the present invention integrates insights from previous tools and adopts a fusion of bidirectional GRU and RetNet modules for modeling, becoming an excellent tool for predicting the binding of peptides to TCR molecules. A large number of experiments have shown that this method not only demonstrates its advantages in three different environments, but also shows its strong predictive performance in tasks such as characterizing peptide binding preferences, distinguishing true binding compounds from false binding compounds, characterizing T cell clone expansion rates, evaluating the impact of alanine mutations on prediction results, and performance on large external databases. This method can be used to assist in the screening of therapeutic targets in clinical immunotherapy, saving target screening time, shortening the treatment cycle, and increasing the success rate of tumor treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The features and advantages of the present invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the present invention in any way. In the accompanying drawings:

[0042] Figure 1 is a flow chart of the prediction method of the present invention;

[0043] Figure 2 The results of the present invention and the comparative method under three different test conditions are compared;

[0044] Figure 3 This is an example of the comparison of the results of the present invention and the comparative method in characterizing the binding preference of the polypeptide;

[0045] Figure 4 The ability of the present invention to reproduce CDR3 binding peptides was demonstrated. DETAILED DESCRIPTION

[0046] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0047] The embodiment of the present invention is a method for predicting the binding relationship between a polypeptide and a TCR molecule based on a large-scale language model, comprising the following steps:

[0048] S100, obtaining a polypeptide sequence of a binding relationship to be tested and a CDR3 fragment sequence on a TCR molecule, and preprocessing the sequences to be tested to obtain a target input sequence;

[0049] S200, performing feature encoding on each amino acid residue in the target input sequence, and outputting an encoding matrix for each sequence;

[0050] S300, inputting the encoding matrices into the bidirectional GRU modules in the trained binding relationship prediction model, and outputting the global feature matrix of each sequence;

[0051] S400, after merging the corresponding global feature matrices of each sequence, inputting them into the trained large-scale language model RetNet in the combined relationship prediction model, and outputting the relationship matrix of each pair of sequences;

[0052] S500: Input the relationship matrix of each pair of sequences into a binding relationship prediction model, and output the binding probability of the polypeptide sequence to be tested for binding relationship and the CDR3 fragment sequence on the TCR molecule.

[0053] Wherein, step S200 includes:

[0054] According to 28 features extracted from the AAIndex database that have been successfully applied in the field of protein binding, the physical and chemical characteristics of each amino acid residue in the target input sequence are encoded.

[0055] Among them, the 28 features that have been successfully applied in the field of protein binding are CHOP780202, CIDH920103, CIDH920105, FAUJ880109, GEIM800106, KANM800102, KLEP840101, KRIW710101, PALJ810107, QIAN880123, RACS770103, RADA880108, ZIMJ68010 4. AURR980120, MUNV940103, NADH010104, FAUJ880111, FINA910104, GEIM800104, NADH010106, LIFS790101, MEEJ800101, OOBM770102, GUYH850105, ROSM880102, SWER830101, ZIMJ680102, and MIYS990104. See Table 1 for details.

[0056] Table 1

[0057]

[0058] Wherein, step S300 includes:

[0059] Input the encoding matrix of the peptide into the bidirectional GRU module to obtain the global feature matrix of the peptide;

[0060] The encoding matrix of the CDR3 segment on the TCR molecule is input into the bidirectional GRU module to obtain the global feature matrix of CDR3.

[0061] Step S400 is specifically as follows:

[0062] The global feature matrices of the peptide and CDR3 segments were spliced ​​together;

[0063] The concatenated matrix is ​​input into a 24-layer RetNet module in series, and the relationship matrix of each pair of sequences is output.

[0064] Furthermore, step S500 is specifically as follows:

[0065] The relationship matrix of each pair of sequences is input into the binding relationship prediction model, and the binding probability of the polypeptide sequence to be tested for binding relationship and the CDR3 fragment sequence on the TCR molecule is output, specifically:

[0066] Input the relationship matrix into two layers of linear layer and ReLU layer connected in series;

[0067] After another linear layer and a Sigmoid layer, the value is mapped to [0, 1], indicating the binding probability between the input peptide and the TCR molecule.

[0068] An embodiment of the present invention is a device for predicting the binding relationship between polypeptides and TCR molecules based on a large-scale language model, comprising at least one processor; and at least one memory in communication with the processor, wherein:

[0069] The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the above-mentioned method for predicting the binding relationship between polypeptides and TCR molecules based on a large-scale language model, including the following functional modules:

[0070] A data acquisition and preprocessing module, configured to execute step S100;

[0071] A feature encoding module, configured to execute step S200;

[0072] A feature representation module, configured to execute step S300;

[0073] Combined with the relationship information mining module, used to perform the step S400;

[0074] The prediction module is used to execute the step S500.

[0075] Furthermore, the feature representation module includes a bidirectional GRU module, which can independently learn the contextual information of each sequence, including the context of subsequences. The bidirectional GRU module comprehensively learns the information of different subsequences along the sequence direction, ensuring the integrity and sequentiality of information during model training.

[0076] The combined relational information mining module includes a RetNet module, which consists of 24 layers connected in series. The crucial recursive representation module in the RetNet module is a variant of self-attention, which introduces a D matrix and a GN module. The exponential decay factor Y in the D matrix enables RetNet to retain the semantic influence of sequence proximity during the learning process. In addition, the GN module introduces nonlinearity to facilitate parallel training of modules. Essentially, the addition of the RetNet module significantly enhances the model's understanding of sequences and increases training speed. Compared with traditional Transformer positional encoding, this matrix is ​​more consistent with sequence processing in terms of positional awareness and is more suitable for the task of predicting peptide-TCR binding.

[0077] Among them, after the RetNet module outputs the relationship matrix of each pair of sequences, it enters the prediction module in sequence, including 3 linear layers, 2 ReLU activation layers, and 1 random downsampling. Finally, the result passes through a Sigmoid activation layer, mapping the value to [0, 1], indicating the binding probability between the input peptide and the TCR molecule.

[0078] The above-mentioned combination relationship prediction method can be converted into software program instructions, which can be implemented by using a combination relationship prediction system including a processor and a memory, or can be implemented by computer instructions stored in a non-transitory computer-readable storage medium. The above-mentioned integrated unit implemented in the form of a software functional unit can be stored in a computer-readable storage medium. The above-mentioned software functional unit is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to perform some steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0079] Example 1

[0080] A method for predicting the binding relationship between peptides and TCR molecules based on a large-scale language model comprises the following steps:

[0081] S100, obtaining a polypeptide sequence RAKFKQLL of a binding relationship to be tested and a CDR3 fragment sequence ASSFTGGTGELF on a TCR molecule, and preprocessing the sequences to be tested to obtain target input sequences RAKFXXXXXXKQLL and ASSFTGGTGELFXXXXXXXX;

[0082] S200, performing feature encoding on each amino acid residue in the target input sequence, and outputting an encoding matrix for each sequence;

[0083] S300, inputting the encoding matrices into the bidirectional GRU modules in the trained binding relationship prediction model, and outputting the global feature matrix of each sequence;

[0084] S400, after merging the global feature matrices of each pair of sequences, the obtained matrix is ​​input into the trained large-scale language model RetNet in the combined relationship prediction model, and the relationship matrix of each pair of sequences is output;

[0085] S500: Input the relationship matrix of each pair of sequences into a binding relationship prediction model, and output the binding probability of the polypeptide sequence to be tested for binding relationship and the CDR3 fragment sequence on the TCR molecule as 0.8197.

[0086] Experimental verification

[0087] In order to verify the effectiveness of the prediction method of the present invention, the present invention evaluated the AUC and AUPR of several prediction tools (TEIM, TEINet, PanPep) newly released in 2023 under three different conditions in the test set. Figure 2 As shown in Figure A, the four tools were run directly on the test set. The results showed that the method of the present invention exhibited the best prediction performance, with an AUC of 0.8164 and an AUPR of 0.8100, followed by TEINet. Secondly, to eliminate bias in the test set that may appear in the training set of the method of the present invention but not in the reference tool, the present invention extracted a total of 14,551 binding pairs containing 25 peptides based on the training sets of the four tools. Figure 2 B shows that even in the cross-test set, the method of the present invention maintains its leading position, with AUC / AUPR of 0.8671 and 0.8665 respectively. This shows that the predictive advantage of the method of the present invention is still reliable on a completely fair test set. Finally, considering that PanPep performed well in its unseen set in the paper, the present invention conducted a further "fair" testing scheme. PanPep is the only tool among the three reference tools that provides all positive and negative samples during training, and the data is divided into visible and unseen sets. The present invention retrained and tested the method of the present invention using the training and test data of PanPep. Figure 2 The results in Figure C show that the method of the present invention achieves excellent prediction performance in both the visible and invisible sets. The excellent results in the invisible set may be attributed to PanPep's use of a different negative sample filling strategy, confirming the superior predictive ability of the method of the present invention in peptide-TCR molecule binding.

[0088] In order to explore the binding characteristics of each peptide, the present invention conducted a detailed analysis of the binding preferences of the peptides based on the existing binding data. Specifically, the present invention extracted a set of peptides with more than 100 real binding CDR3 sequences from all collected binding pairs. The probability of amino acids appearing at each site was visualized using the "ggseqlogo" package. A total of 99 peptides that met the specific requirements were extracted for further analysis. After visualizing the frequency of each binding site, as shown in the figure below: Figure 3 Figure 4 shows a frequency bias due to insufficient length at the end of most CDR3 data, but notable observations include a trend for peptides to exhibit strong binding preferences in the head of CDR3. In contrast, a broader range of binding preferences was observed in the middle and tail of CDR3.

[0089] To evaluate the ability of the tools to characterize peptide binding preferences, we extracted the top N predicted binding scores of CDR3 sequences from the test set and generated a ggseqlogo graph for each tool (N represents the number of actual binding compounds in the test set). As an example, Figure 3 B illustrates the case of TPRVTGGGAM. Figure 3 In the first row of Figure B, the first image shows the ggseqlogo graphs of all collected TPRVTGGGAM complexes, while the second image shows the ggseqlogo graphs of all positive samples in the test set. By comparing the trends of CDR3 binding to amino acids at positions 1 / 2 / 10 / 11 / 12 / 13, it is clear that the method of the present invention shows the best reproducibility in reproducing the original binding preferences of the polypeptide.

[0090] In order to further evaluate the anti-interference ability of the method of the present invention, the present invention extracted a validation data set, including 5 polypeptide sets with the largest number of binding pairs in the test set. From these 5 polypeptides, 100 non-repetitive CDR3 sequences were randomly extracted from the test set, respectively, to obtain a total of 500 CDR3 sequences. The method of the present invention was then used to predict the binding of each CDR3 sequence to the 5 peptides. The statistical ranking of each CDR3 on the 5 polypeptides was obtained. Ideally, the original binding sequence of CDR3 should be ranked 1st. The predicted ranking of the true binders of the 5 polypeptides is as follows: Figure 4 As shown in Figure 2, it is clear that all CDR3s can still recognize their original binders and rank the prediction value first among all binding pairs. Figure 4 It can be observed that many CDR3s also rank other peptides higher. This observation is not surprising, as these are the top five peptides with the most bound compounds, and their biological significance suggests that they are highly toxic and likely to trigger responses from other CDR3s.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations shall fall within the scope defined by the appended claims.

Claims

1. A method for predicting the binding relationship between peptides and TCR molecules based on a large-scale language model, characterized in that: The following steps are involved: S100, obtaining a polypeptide sequence of a binding relationship to be tested and a CDR3 fragment sequence on a TCR molecule, and preprocessing the sequences to be tested to obtain a target input sequence; S200, based on 28 features extracted from the AAIndex database that have been successfully applied in the field of protein binding, as the encoding of the physical and chemical characteristics of each amino acid residue in the target input sequence; Perform feature encoding on each amino acid residue in the target input sequence and output an encoding matrix for each sequence; Among them, the 28 features that have been successfully applied in the field of protein binding are CHOP780202, CIDH920103, CIDH920105, FAUJ880109, GEIM800106, KANM800102, KLEP840101, KRIW710101, PALJ810107, QIAN880123, RACS770103, RADA880108, ZIMJ68010 4. AURR980120, MUNV940103, NADH010104, FAUJ880111, FINA910104, GEIM800104, NADH010106, LIFS 790101, MEEJ800101, OOBM770102, GUYH850105, ROSM880102, SWER830101, ZIMJ680102, MIYS990104; S300, inputting the encoding matrices into the bidirectional GRU modules in the trained binding relationship prediction model, and outputting the global feature matrix of each sequence; S400, after merging the corresponding global feature matrices of each sequence, inputting them into the trained large-scale language model RetNet in the combined relationship prediction model, and outputting the relationship matrix of each pair of sequences; S500: Input the relationship matrix of each pair of sequences into a binding relationship prediction model, and output the binding probability of the polypeptide sequence to be tested for binding relationship and the CDR3 fragment sequence on the TCR molecule.

2. The method for predicting the binding relationship between polypeptides and TCR molecules based on a large-scale language model according to claim 1, characterized in that: In step S100, the preprocessing specifically includes: padding the sequence to be tested with "X" residues to align the lengths of the sequence to be tested.

3. The method for predicting the binding relationship between polypeptides and TCR molecules based on a large-scale language model according to claim 1, characterized in that: Step S300 includes: Input the encoding matrix of the peptide into the bidirectional GRU module to obtain the global feature matrix of the peptide; The encoding matrix of the CDR3 segment on the TCR molecule is input into the bidirectional GRU module to obtain the global feature matrix of CDR3.

4. The method for predicting the binding relationship between polypeptides and TCR molecules based on a large-scale language model according to claim 1, characterized in that: Step S400 is specifically as follows: The global feature matrices of the peptide and CDR3 segments were spliced ​​together; The concatenated matrix is ​​input into a 24-layer RetNet module in series, and the relationship matrix of each pair of sequences is output.

5. The method for predicting the binding relationship between polypeptides and TCR molecules based on a large-scale language model according to claim 1, characterized in that: Step S500 is specifically as follows: Input the relationship matrix of each pair of sequences into two serially connected linear layers and ReLU layers in the combined relationship prediction model; After another linear layer and a Sigmoid layer, the value is mapped to [0, 1], indicating the binding probability between the input peptide and the TCR molecule.

6. A device for predicting the binding relationship between peptides and TCR molecules based on a large-scale language model, characterized in that: comprising at least one processor; and at least one memory communicatively connected to the processor, wherein: The memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the method for predicting the binding relationship between polypeptides and TCR molecules based on a large-scale language model according to any one of claims 1 to 5, comprising the following functional modules: A data acquisition and preprocessing module, configured to execute step S100; A feature encoding module, configured to execute step S200; A feature representation module, configured to execute step S300; Combined with the relationship information mining module, used to perform the step S400; The prediction module is used to execute the step S500.

7. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium stores computer instructions, which enable the computer to execute the method for predicting the binding relationship between polypeptides and TCR molecules based on a large-scale language model as described in any one of claims 1 to 5.

Images