Method for identifying peptide fragment bound to class MHC-i / HLA-i and recognized by TCR, device, and storage medium

Abstract

A method for identifying a peptide fragment bound to a class MHC-I / HLA-I and recognized by TCR, a device, and a storage medium. The method comprises: acquiring a protein sequence, an MHC-I / HLA-I sequence, and a peptide fragment sequence; using the protein sequence to train a protein sequence tokenizer; using the protein sequence tokenized by the protein sequence tokenizer to train a protein characterization model, and obtaining a trained protein characterization pre-trained model; using the MHC-I / HLA-I sequence and the peptide fragment sequence tokenized by the protein sequence tokenizer to fine-tune the trained protein characterization pre-trained model, and obtaining a binary classification model; using the binary classification model to perform identification processing on a target wild-type / mutant peptide fragment, classifying the mutant peptide fragment, and acquiring a candidate neoantigen mutant peptide positively correlated with TCR recognition.

Description

A method, device and storage medium for identifying peptides that bind to MHC-I / HLA-I and are recognized by TCR Technical Field

[0001] The present invention relates to the field of vaccines, and in particular to a method, device and storage medium for identifying peptide segments that bind to MHC-I / HLA-I class and are recognized by TCR. Background Art

[0002] Antigen screening technology primarily aims to identify antigens that can generate the strongest immune response during vaccine development. This process involves screening antigens from various sources to identify the optimal antigen combination. Commonly used antigen screening techniques include biochip-based screening and protein analysis-based screening. Tumor vaccines are currently a hot topic in the vaccine field. The earliest tumor vaccine was described in 1994. Recent technological advances have enabled targeted therapy with patient-specific mutant antigens. These vaccines activate the patient's immune system by injecting peptides or mRNA encoding tumor-associated or tumor-specific antigens into the patient. Breakthroughs in personalized tumor vaccine technology lie in the development of novel antigen detection and prediction technologies to facilitate the preparation of novel antigen vaccines. Predicting the binding potential of specific mutant peptides to the histocompatibility complex (MHC) is a crucial component of neoantigen detection. Tumor-specific antigens must be recognized and presented on the cell surface by the MHC to be recognized by T cell receptors and ultimately kill tumor cells. Therefore, predicting peptide binding to MHC-I / HLA-I, as well as whether the resulting pMHC complex remains stably bound and can be presented on the cell surface and recognized by the TCR, is crucial.

[0003] In recent years, the field of neoantigens has seen significant development, generating extensive and extensive databases for training peptide affinity prediction models for MHC-I / HLA-I. A variety of algorithms have been applied to predict peptide binding potential for MHC-I / HLA-I, including traditional machine learning methods and the rapidly developing deep learning approach. Artificial neural networks (ANNs), convolutional neural networks (CNNs), Bayesian least squares methods (SMMs), and random forests have all been used to construct affinity models. While these algorithms have been reported to perform well when tested on a separate dataset, their performance has been more limited when used on completely independent datasets. Peptide affinity prediction for MHC-I / HLA-I still suffers from high error rates, particularly false positive rates, which can have serious implications for clinical application. The reasons for this problem are as follows: First, it is impossible to obtain a comprehensive and large dataset for model training, and it is difficult to learn more features from a limited dataset; second, conventional algorithms used in prediction models rarely consider the characteristics of the protein sequence itself and the sequence features that may have an important impact on protein sequence interactions; third, whether it is the representation of the MHC-I / HLA-I full-length sequence or the representation of the MHC-I / HLA-I pseudo-sequence, the model cannot learn the characteristics of important sites in peptide binding in a targeted manner, affecting model performance; fourth, the current common methods only focus on pMHC presentation, the characteristics of mutant peptides such as binding position and amino acid residue type, and do not further consider the important influence of pMHC binding stability, pMHC presentation ability and central tolerance mechanism on TCR recognition, which affects the further selection of more accurate neoantigen mutant peptides.

[0004] Summary of the Invention

[0005] To address the above problems, the present invention provides a method, device, and storage medium for identifying peptides that bind to MHC-I / HLA-I and are recognized by TCRs. This method provides a more targeted method for obtaining the key features of tumor-specific peptide identification, transport, presentation, and TCR recognition. In response to the current situation of insufficient data for model construction, new modeling ideas are applied to model training, feature extraction, and grouping, thereby improving the accuracy of identifying peptides and neoantigens that bind to MHC-I / HLA-I.

[0006] The present invention provides a method for identifying peptides that bind to MHC-I / HLA-I class and are recognized by TCR, comprising:

[0007] Obtain protein sequences, MHC-I / HLA-I sequences, and peptide sequences;

[0008] training a protein sequence word segmenter using the protein sequence;

[0009] Using the protein sequence segmented by the protein sequence segmenter to train a protein representation model, to obtain a trained protein representation pre-training model;

[0010] Using the MHC-I / HLA-I sequence and the peptide sequence after word segmentation processing by the protein sequence word segmenter to perform splicing processing, the trained protein representation pre-training model is fine-tuned to obtain a binary classification model;

[0011] The binary classification model is used to identify and process target wild-type / mutant peptide segments, classify mutant peptide segments, and obtain candidate neoantigen mutant peptides that are positively correlated with TCR recognition;

[0012] The binary classification model can be applied to the prediction of any peptide that can bind to MHC-I / HLA-I. Candidate neoantigens can be predicted by applying the binary classification model and the location of the mutation.

[0013] The present invention provides a method for identifying peptides that bind to MHC-I / HLA-I class and are recognized by TCR, comprising:

[0014] Obtain protein sequences, MHC-I / HLA-I sequences, and peptide sequences;

[0015] training a protein sequence word segmenter using the protein sequence;

[0016] Using the protein sequence segmented by the protein sequence segmenter to train a protein representation model, to obtain a trained protein representation pre-training model;

[0017] Using the MHC-I / HLA-I sequence and the peptide sequence after word segmentation processing by the protein sequence word segmenter to perform splicing processing, the trained protein representation pre-training model is fine-tuned to obtain a binary classification model;

[0018] The two-class model is used to identify the peptides to be identified, and peptides that bind to MHC-I / HLA-I are obtained.

[0019] Wherein, the MHC-I refers to major histocompatibility complex class I; the HLA-I refers to major histocompatibility complex class I in humans; and the TCR refers to T cell receptor.

[0020] Preferably, the peptide sequence includes a positive peptide sequence and a negative peptide sequence, wherein the positive peptide sequence is a peptide that binds to MHC-I / HLA-I, and the label is 1; the negative peptide sequence is a peptide sequence randomly generated in the protein sequence according to the characteristics of the positive peptide sequence, and the label is 0.

[0021] Preferably, the protein sequence includes protein sequences collected from a protein database; accordingly, the collected protein sequences are used to perform word segmentation training on a high-frequency repeated amino acid residue combination motif to obtain a protein sequence word segmenter.

[0022] Preferably, the protein sequence segmented by the protein sequence segmenter is used to train a protein representation model to obtain a trained protein representation pre-training model, comprising:

[0023] The unlabeled protein sequence is input into the protein sequence tokenizer for token processing to obtain the protein sequence after segmentation of the amino acid functional residue combination motif;

[0024] Converting the protein sequence segmented by the amino acid functional residue combination motif into a digitally represented protein sequence;

[0025] The protein sequence represented by the digits is used to train a protein representation model to obtain a trained protein representation pre-training model.

[0026] Preferably, the method of fine-tuning the trained protein representation pre-training model using the MHC-I / HLA-I sequence and peptide sequence segmented by the protein sequence segmenter to obtain a binary classification model includes:

[0027] Inputting the labeled MHC-I / HLA-I sequence into a protein sequence tokenizer for token processing to obtain an MHC-I / HLA-I sequence segmented by amino acid functional residue combination motifs, and converting the MHC-I / HLA-I sequence segmented by amino acid functional residue combination motifs into a digitally represented MHC-I / HLA-I sequence;

[0028] Convert the labeled peptide sequence into a digital peptide sequence;

[0029] The binary classification model is obtained by converting to obtain a spliced ​​sequence and using the spliced ​​sequence to fine-tune the trained protein characterization pre-training model.

[0030] In a specific embodiment, fine-tuning may include inheriting the original pre-trained model and adding a classification layer; in addition, a fully connected layer, an attention mechanism layer, or other suitable task layers may be further added on the basis of the added classification layer.

[0031] In some embodiments, the splicing strategy is direct splicing, that is, directly splicing the digitally represented MHC-I / HLA-I sequence and the digitally represented peptide sequence.

[0032] In other embodiments, the splicing strategy involves splicing digitally represented MHC-I / HLA-I sequences, digitally represented peptide sequences, and other digitally represented features related to peptide binding to MHC-I / HLA-I. Those skilled in the art can select an appropriate splicing method based on the splicing target. Exemplary splicing methods include directly adding, subtracting, multiplying, and dividing the feature tensor using torch.cat() or using a custom splicing method.

[0033] In a specific embodiment, the other characteristics associated with the binding of the peptide to MHC-I / HLA-I are selected from one or more of the cleavage probability of the protein sequence, the binding stability of MHC-I / HLA-I to the peptide, and the presentation probability of the complex of MHC-I / HLA-I and the peptide.

[0034] The cleavage probability of a protein sequence can be determined by inputting a peptide sequence using appropriate software. Those skilled in the art will appreciate that any suitable software and its appropriate version can be used, as long as it can determine the cleavage probability of a protein sequence based on the peptide sequence. For example, software such as pepsickle and iPCPS can achieve this purpose.

[0035] The binding stability of MHC-I / HLA-I peptides can be determined using appropriate software by inputting the peptide sequence and MHC-I / HLA-I sequence. Those skilled in the art will appreciate that any suitable software and its appropriate version can be used, as long as it can determine the binding stability of MHC-I / HLA-I peptides based on the peptide sequence and MHC-I / HLA-I sequence. For example, software such as LKMHCstablepan and TLStab can achieve this purpose.

[0036] The presentation probability of a complex of MHC-I / HLA-I and a peptide can be obtained by inputting a peptide sequence using appropriate software. Those skilled in the art will appreciate that any suitable software and its appropriate version can be used, as long as it can determine the presentation probability of a complex of MHC-I / HLA-I and a peptide based on the peptide sequence and the MHC-I / HLA-I sequence (optional). For example, DeepTAP, TAPPred, TAPREG, and the like can achieve this purpose.

[0037] Preferably, before the step of fine-tuning, the method of the present invention further comprises calculating other features related to the binding of peptides to MHC-I / HLA-I using the protein sequence, MHC-I / HLA-I sequence and peptide sequence.

[0038] Preferably, the identification process of the target wild-type / mutant peptide using the binary classification model includes:

[0039] The target wild-type / mutant peptide is input into the binary classification model, and the attention score of each site of the mutant peptide predicted to be positive and the attention score of the amino acid functional residue combination motif of the MHC-I / HLA-I sequence are output;

[0040] According to the attention score of each site of the positive mutant peptide and the attention score of the MHC-I / HLA-I sequence motif, the anchor position when the MHC-I / HLA-I sequence binds to the positive peptide and the key characterization site of the MHC-I / HLA-I sequence are determined, and the positive mutant peptides are classified using the anchor point when the MHC-I / HLA-I sequence binds to the peptide and the mutation site of the peptide.

[0041] Preferably, the classification process using the anchor position when the MHC-I / HLA-I sequence binds to the positive mutant peptide segment and the mutation position relative to the wild-type peptide segment includes:

[0042] Positive mutant peptides are classified based on the anchor point when the MHC-I / HLA-I sequence binds to the peptide and the location of the mutation, and the suitability of being selected as a candidate neoantigen mutant peptide positively correlated with TCR recognition is determined based on the classification results.

[0043] Preferably, judging whether a candidate neoantigen mutant peptide is suitable for selection as a positively correlated candidate for TCR recognition according to the classification result comprises:

[0044] When the wild-type peptide has weak MHC-I / HLA-I binding and weak TCR recognition (generally, wild-type peptides have weak TCR recognition), and the mutant peptide has a mutation at the MHC-I / HLA-I anchor site, resulting in strong MHC-I / HLA-I and strong TCR recognition, it can be selected as a neoantigen mutant peptide;

[0045] When the wild-type peptide has weak MHC-I / HLA-I binding and weak TCR recognition (generally, wild-type peptides have weak TCR recognition), and the mutant peptide has a non-MHC-I / HLA-I anchor site mutation, which turns to strong MHC-I / HLA-I and strong TCR recognition, it can be selected as a neoantigen mutant peptide;

[0046] When the wild-type peptide has strong MHC-I / HLA-I binding and weak TCR recognition (generally, wild-type peptides have weak TCR recognition), and the mutant peptide has a non-MHC-I / HLA-I anchor site mutation, which transforms into strong MHC-I / HLA-I and strong TCR recognition (strong TCR recognition here refers to a change in the sequence exposed to TCR), it can be judged as a neoantigen mutant peptide;

[0047] When the wild-type peptide has strong MHC-I / HLA-I binding and weak TCR recognition (generally, wild-type peptides have weak TCR recognition), and the mutant peptide has a mutation in the MHC-I / HLA-I anchor site, which turns to strong MHC-I / HLA-I and weak TCR recognition (weak TCR recognition here means that the sequence exposed to TCR has not changed), it is judged that it cannot be selected as a neoantigen mutant peptide.

[0048] The beneficial effects of the present invention are that, by applying a deep learning algorithm of natural language processing, the recognition, transport and presentation of wild-type / mutant peptides and the tissue compatibility molecule complex MHC-I / HLA-I are identified, the biological characteristics of protein interactions are recognized, the anchoring sites of peptides and MHC-I / HLA-I are determined, and a more representative sequence representation method of MHC-I / HLA-I is constructed. At the same time, the influence of the central tolerance mechanism on TCR recognition is introduced, thereby further promoting the recognition of tumor neoantigens. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] FIG1 is a flow chart of a method for identifying peptides that bind to MHC-I / HLA-I class and are recognized by TCR, provided by the present invention;

[0050] FIG2 is a flow chart of an algorithm for identifying peptides that bind to MHC-I / HLA-I class and are recognized by TCRs, provided by the present invention;

[0051] FIG3 is a flow chart of an algorithm for identifying peptides that bind to MHC-I / HLA-I class and are recognized by TCRs, provided by the present invention, and having other features. DETAILED DESCRIPTION

[0052] It should be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. In the subsequent description, suffixes such as "module," "component," or "unit" used to represent elements are used solely to facilitate the description of the present invention and have no specific meaning. Therefore, "module," "component," or "unit" may be used interchangeably.

[0053] As used herein, the term "splicing process" refers to the process of splicing multiple features converted into digital representations to obtain a spliced ​​sequence that can be used to fine-tune a pre-trained protein representation model.

[0054] As used herein, the term "pMHC" refers to a complex of a peptide and MHC-I / HLA-I.

[0055] As used herein, the term "other features" is used interchangeably with "further features" to refer to features associated with peptide-MHC-I / HLA-I binding that are calculated using protein, MHC-I / HLA-I, and peptide sequences. Exemplary features include protein sequence cleavage potential, pMHC binding stability, and pMHC presentation ability.

[0056] The present invention provides a deep learning algorithm for identifying wild-type and mutant peptides for their binding to major histocompatibility complex class I (MHC-I / HLA-I), their transport, presentation to the cell surface, and TCR recognition. This algorithm further facilitates the identification of MHC-I / HLA-I class I-binding peptides and the selection of tumor-specific neoantigens. Specifically, the method applies natural language processing techniques to protein sequence feature discovery and extraction to construct a pre-trained protein representation model. Using protein sequence and protein interaction datasets, the algorithm extracts amino acid residue combination motifs and key site features that are positively correlated with peptide binding, transport, and presentation. Based on the pre-trained protein representation model, the model is fine-tuned by collecting positive peptide sequence data and constructing negative peptide sequence data to obtain an accurate model for inferring the binding probability of wild-type and mutant peptides to MHC-I / HLA-I. The model obtained by this method can be applied to predict binding of pan-type peptides or MHC-I / HLA-I classes, inspiring the biological characteristics of peptide presentation. This model can be applied to model development, resulting in a more interpretable and accurate prediction model than conventional algorithms. At the same time, the recognition characteristics of pMHC and TCR after the peptide binds to MHC-I / HLA-I can be further identified, so as to further predict whether the peptide has the ability to activate TCR, rather than just binding to MHC, which can ultimately be used to select more promising candidate neoantigens.

[0057] FIG1 is a flow chart of a method for identifying peptides that bind to MHC class I and are recognized by TCR, provided by the present invention. As shown in FIG1 , the method comprises:

[0058] Step S101: Obtain protein sequence, MHC-I / HLA-I sequence and peptide sequence;

[0059] Step S102: using the protein sequence to train a protein sequence word segmenter;

[0060] Step S103: using the protein sequence segmented by the protein sequence segmenter to train a protein representation model, to obtain a trained protein representation pre-training model;

[0061] Step S104: fine-tuning the trained protein representation pre-training model using the MHC-I / HLA-I sequence and peptide sequence segmented by the protein sequence segmenter to obtain a binary classification model;

[0062] Step S105: using the binary classification model to identify and process the target wild-type / mutant peptide segments, classify the mutant peptide segments, and obtain candidate neoantigen mutant peptides that are positively correlated with TCR recognition.

[0063] Specifically, the peptide sequence includes a positive peptide sequence and a negative peptide sequence, wherein the positive peptide sequence is a peptide that binds to MHC-I / HLA-I, and the label is 1; the negative peptide sequence is a peptide sequence randomly generated in the protein sequence based on the characteristics of the positive peptide sequence, and the label is 0.

[0064] The protein sequence includes collecting protein sequences from a protein database; accordingly, the collected protein sequences are used to train a word segmenter for a high-frequency repeated amino acid residue combination motif to obtain a protein sequence word segmenter.

[0065] Furthermore, the method of using the protein sequence after word segmentation processing by the protein sequence word segmenter to train the protein representation model to obtain a trained protein representation pre-trained model includes: inputting the unlabeled protein sequence into the protein sequence word segmenter for word segmentation processing to obtain a protein sequence after amino acid functional residue combination motif segmentation; converting the protein sequence after amino acid functional residue combination motif segmentation into a protein sequence represented by a number; and using the protein sequence represented by the number to train the protein representation model to obtain a trained protein representation pre-trained model.

[0066] More specifically, the method of fine-tuning the trained protein representation pre-training model using the MHC-I / HLA-I sequence and peptide sequence after word segmentation processing by the protein sequence word segmenter to obtain a binary classification model includes: inputting the labeled MHC-I / HLA-I sequence into the protein sequence word segmenter for word segmentation processing to obtain the MHC-I / HLA-I sequence after amino acid functional residue combination motif segmentation, and converting the MHC-I / HLA-I sequence after amino acid functional residue combination motif segmentation into a digitally represented MHC-I / HLA-I sequence; The peptide sequence of the tag is converted into a peptide sequence represented by a number; the MHC-I / HLA-I sequence converted into a digital representation and the peptide sequence converted into a digital representation are spliced ​​to obtain a spliced ​​sequence, wherein the splicing strategy includes: direct splicing, or adding other digitally represented features, including one or more of the cleavage possibility of the protein sequence, the binding stability of MHC-I / HLA-I and the peptide, and the presentation possibility of the MHC-I / HLA-I and peptide complex; and the spliced ​​sequence is used to fine-tune the trained protein characterization pre-trained model to obtain a binary classification model.

[0067] Furthermore, the use of the binary classification model to identify and process the target wild-type / mutant peptide segment includes: inputting the target wild-type / mutant peptide segment into the binary classification model, outputting the attention score of each site of the mutant peptide segment predicted to be positive, and the attention score of the amino acid functional residue combination motif of the MHC-I / HLA-I sequence; based on the attention score of each site of the positive mutant peptide segment and the attention score of the MHC-I / HLA-I sequence motif, judging the anchor position when the MHC-I / HLA-I sequence binds to the positive peptide segment and the key characterization site of the MHC-I / HLA-I sequence, and using the anchor position point when the MHC-I / HLA-I sequence binds to the peptide segment and the mutation site of the peptide segment to classify the positive mutant peptide segment.

[0068] Furthermore, the classification processing using the anchor position when the MHC-I / HLA-I sequence binds to the positive mutant peptide segment and the mutation position relative to the wild-type peptide segment includes: classifying the positive mutant peptide segment according to the anchor point when the MHC-I / HLA-I sequence binds to the peptide segment and the location of the mutation, and judging whether it is suitable to be selected as a candidate neoantigen mutant peptide positively correlated with TCR recognition based on the classification result.

[0069] More specifically, the method of judging whether a candidate neoantigen mutant peptide is suitable for selection as a positively correlated candidate for TCR recognition according to the classification results includes: when the wild-type peptide segment has weak MHC-I / HLA-I binding and weak TCR recognition, and the mutant peptide segment has a mutation at the MHC-I / HLA-I anchor site, which is converted to strong MHC-I / HLA-I and strong TCR recognition, then it is judged that it can be selected as a neoantigen mutant peptide; when the wild-type peptide segment has weak MHC-I / HLA-I binding and weak TCR recognition, and the mutant peptide segment has a mutation at the non-MHC-I / HLA-I anchor site, which is converted to strong MHC-I / HLA-I and strong TCR recognition, then it is judged that it can be selected as a neoantigen mutant peptide. When the wild-type peptide has strong MHC-I / HLA-I binding and weak TCR recognition, it is judged to be selected as a neoantigen mutant peptide; when the wild-type peptide has strong MHC-I / HLA-I binding and weak TCR recognition, and the mutant peptide has a non-MHC-I / HLA-I anchor site mutation, which is converted to strong MHC-I / HLA-I and strong TCR recognition, it is judged to be selected as a neoantigen mutant peptide; when the wild-type peptide has strong MHC-I / HLA-I binding and weak TCR recognition, and the mutant peptide has a MHC-I / HLA-I anchor site mutation, which is converted to strong MHC-I / HLA-I and weak TCR recognition, it is judged not to be selected as a neoantigen mutant peptide.

[0070] An electronic device provided in an embodiment of the present application includes: a memory; a processor; and a computer program; wherein the computer program is stored in the memory and is configured to be executed by the processor to implement a method, device and storage medium for identifying peptide segments that bind to MHC-I / HLA-I class and are recognized by TCR.

[0071] An embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon; the computer program is executed by a processor to implement a method for identifying peptides that bind to MHC-I / HLA-I class and are recognized by TCR.

[0072] In summary, the present invention provides a method for inferring the possibility of peptide binding to MHC-I / HLA-I. The method learns biological features from protein sequences and, based on a pre-trained model, is applied to obtain an interpretable and accurate peptide and MHC-I / HLA-I binding prediction model. The method specifically comprises: applying a large amount of unlabeled protein data to construct a protein sequence segmenter and a protein representation pre-trained model; constructing a labeled training set, a validation set, and a test set comprising positive / negative data sets; wherein the positive data set comprises positive peptide sequences obtained from a database by mass spectrometry or in vitro binding experiments and their corresponding MHC-I / HLA-I subtypes; wherein the negative sequences are negative peptide sequences randomly generated in the protein sequence based on the length characteristics of the positive peptide sequences, and the negative peptides need to exclude the peptides present in the positive peptides. The MHC-I / HLA-I subtypes were segmented using a protein sequence segmenter, and the MHC-I / HLA-I sequences and peptide sequences were converted into digital representations. A binary classification model was constructed and fine-tuned based on the protein representation pre-training model. The generated digital representation data was used for supervised learning to further improve the model's performance on this task. The binary classification model was further applied to obtain the MHC-I / HLA-I representation method and the inference method based on the attention score of the predicted positive peptides and the MHC-I / HLA-I sequence motif. The binding anchor point of the peptide segment to MHC-I / HLA-I is broken; at a deeper level, a method is introduced to infer whether the peptide segment can bind to MHC-I / HLA-I and activate TCR response and activate immune response. The anchor position of the peptide segment binding to MHC-I / HLA-I is used to classify the wild-type / mutant peptide segment binding to MHC-I / HLA-I and TCR recognition type according to the location of the mutated peptide segment and the influence of central tolerance. The neoantigen peptide segment is further eliminated according to the classification results to obtain more accurate neoantigens.

[0073] The present invention relates to a novel method for predicting wild-type / mutant peptide binding to MHC-I / HLA-I, transport, presentation, and TCR activation. This method extracts protein residue combinations and key sites that play an important role in the binding process, better characterizes MHC-I / HLA-I sequences, and applies transfer learning technology to construct a more accurate and biologically interpretable model predictor. The constructed model is used to explore key MHC-I / HLA-I binding sites and further classify neoantigens based on the location of the mutant peptide, thereby improving the accuracy of neoantigen recognition. Compared with previous classifiers, this method is more efficient and accurate and has important application value in tumor immunotherapy.

[0074] As demonstrated in the examples below, compared to other existing analytical methods, this method has the following advantages:

[0075] (1) Apply the biological characteristics of the protein sequence itself to model construction. By learning the combination of amino acid residues in the protein sequence, we can find high-frequency amino acid residue combinations and key sites with significant characteristics, build a word segmenter, and more reasonably represent the amino acid sequence.

[0076] (2) Apply natural language processing algorithms to bioinformatics language - protein sequences, treat protein sequences as sentences, use natural language processing methods to build prediction models, and learn the features and contextual relationships in the sequences.

[0077] (3) A pre-trained model is constructed using a large amount of existing unlabeled data, and the pre-trained model is fine-tuned using labeled data. A mask model is constructed to infer masked amino acid residue combinations to learn the contextual information and expression rules of protein sequences, helping the pre-trained model to learn rich language knowledge. The pre-trained model is then fine-tuned using limited labeled peptides and MHC-I / HLA-I binding data to construct a classification model that is more suitable for affinity prediction.

[0078] (4) With the help of natural language processing models, the attention score of each site / motif in the protein sequence can be output to determine the key sites when MHC-I / HLA-I binds to peptides, the key binding sites of peptides, and a more representative and significant representation method of MHC-I / HLA-I sequences.

[0079] (5) It not only predicts the binding of peptides to MHC-I / HLA-I, but also covers the binding stability, transport and presentation of peptides to MHC-I / HLA-I through splicing. Starting from the actual biological process, the features that play an important role in the biological pathway are converted into digital form and combined with other features for model training and prediction, thereby improving the practical application ability of the model.

[0080] (6) By determining the location of the peptide where the tumor-specific mutation is located, different pMHC binding and TCR recognition modes are grouped to further extract more accurate bioantigen peptides. According to literature reports, the antigen sequence binds to MHC on one side and is recognized by TCR on the other side. The location of the tumor-specific antigen mutation will affect both MHC binding and TCR recognition. Therefore, we obtain the key MHC binding site based on the attention score of each antigen peptide position, and further classify the new antigens according to the mutation position, eliminating peptides with high binding potential to MHC-I / HLA-I but cannot activate TCR, thereby improving the accuracy of antigen peptide screening. The improved algorithm can more accurately predict whether wild-type / mutant peptides bind to the major histocompatibility complex, transport and presentation, and TCR recognition potential, promoting the identification of tumor new antigens and their clinical application.

[0081] The following is a detailed description of the content of this application with reference to Figures 2 and 3.

[0082] Example 1

[0083] 1.1. Datasets required for model training and screening

[0084] 1.1.1. Obtain protein sequences from databases such as UniProt;

[0085] 1.1.2. Obtain positive peptide sequences from databases such as IEDB. Positive peptides must meet the following conditions:

[0086] Positive peptides with a peptide length of 8-15 bp;

[0087] The MHC / HLA type that recognizes the positive peptide is type I;

[0088] The corresponding wild-type peptides have specific UniProt numbers.

[0089] 1.1.3. Obtain HLA sequences from databases such as IPD-IMGT / HLA, and obtain HLA-I sequences after multiple sequence alignment for characterization of HLA sequences.

[0090] 1.2. Constructing positive and negative datasets

[0091] The positive dataset consists of protein sequences / positive peptide sequences. Protein sequences are used to build pre-training models to learn the expression characteristics and contextual connections of protein sequences. Positive peptide sequences are labeled peptides that can bind to MHC-I / HLA-I. Negative sequences are randomly constructed based on positive sequences. That is, peptides that do not bind to MHC-I / HLA-I are inferred from positive peptides. The negative peptide construction method is as follows:

[0092] Negative peptides with a peptide length of 8-15 bp;

[0093] The positive peptide is generated by randomly changing the amino acids on the basis of the same protein sequence;

[0094] Peptides that appeared in the positive peptides were eliminated.

[0095] 1.3. Build and apply protein sequence tokenizer

[0096] 1.3.1. Using machine learning algorithms or statistical inference models, we analyze amino acid residue combinations across a large number of proteins and infer the most frequently occurring combinations within the protein sequence. These frequently occurring combinations are likely functional motifs, or in other words, play a crucial role in the protein's function. Furthermore, we can infer the locations of important motifs. Rather than simply focusing on individual amino acid residues, tokenizers are more effective in exploring key features of protein function.

[0097] 1.3.2 Protein Sequence Segmentation: All protein sequences are segmented by a tokenizer, grouping amino acid residues that may act together to facilitate subsequent model building.

[0098] An embodiment of the present invention can learn the combination pattern of amino acid residues in protein sequences in a protein database, group the high-frequency combination residues in the protein sequence into a group, and construct a protein sequence tokenizer, including: downloading a protein data set from the protein database and extracting protein sequences therefrom; training the protein sequence, and training an interpretable protein sequence amino acid residue tokenizer based on the frequency of occurrence of amino acid residues / residue combinations in known protein sequences, left-right connectivity, and other features.

[0099] 1.4. Construction of protein characterization pre-training model

[0100] A large number of segmented protein sequences hide features related to protein function, protein expression patterns, and the relationships between motifs in the sequences. Therefore, we use the masked language model in natural language processing to randomly mask the motifs in the protein sequence, predict them, and learn the features hidden in the protein sequence. The masked language model is a self-supervised learning task that aims to allow the model to learn the contextual information and language rules of the motif during pre-training. The model receives the input sequence, then randomly selects a certain proportion of motifs in the input text sequence to mask, and requires the model to predict these masked motifs. For example, given the following sentence:

[0101] Input sequence: MAF SAE DVL KE YD RRRR ME ALLL

[0102] Mask sequence: MAF SAE [mask] KE YD RRRR ME ALLL

[0103] The present invention learns a wealth of protein motif-related information by performing unsupervised learning on a large amount of unlabeled protein data, and then applies transfer learning technology to apply these models to specific labeled classification tasks, including: inputting a large number of unlabeled protein sequences into a protein sequence tokenizer, tokenizing the protein sequences, and characterizing the protein sequences in the form of amino acid combination residues; constructing a masked language pre-training model (MLM), masking random amino acid residue combinations in the input tokenized protein sequence, and learning the contextual information and expression rules of the protein sequence when inferring the masked amino acid residue combinations, helping the model learn key residue features.

[0104] 1.5. Fine-tuning the protein representation pre-training model

[0105] 1.5.1. Label the generated positive peptides that can bind to MHC-I / HLA-I with labels 1 / 0, as well as the negative peptides randomly generated from the positive peptides. It is worth noting that these positive data are obtained through experiments, which means that only limited labeled training data can be obtained. This is why a large amount of unlabeled protein data is used to pre-train the model.

[0106] 1.5.2 Divide the labeled data in 1.5.1 into training set / validation set / test set according to a certain ratio.

[0107] 1.5.3. Fine-tune the pre-trained model generated using the data from 1.5.2 and use it as input to build a binary classification model.

[0108] 1.5.4. The model structure can be seen in Figure 2.

[0109] 1.6. Application of binary classification model

[0110] 1.6.1. Identify peptides with potential to bind to MHC-I / HLA-I.

[0111] 1.6.2. Extract important features from protein sequences, including protein functional amino acid residue combination motifs, important functions / binding sites, etc.

[0112] Through the above method, we can construct peptides that can predict the potential of peptides to bind to MHC-I / HLA-I based on biological characteristics and identify their important characteristics.

[0113] 1.7. Identify the impact of mutation positions on peptide binding to MHC-I / HLA-I and TCR recognition potential

[0114] Based on the obtained peptides and MHC-I / HLA-I binding sites, combined with the mutation location of the wild-type / mutant peptides, and considering the impact of central tolerance mechanisms on neoantigen screening, pMHC binding and TCR recognition types are divided into four categories, including the first three categories with neoantigen potential and the last category that needs to be excluded from neoantigen screening:

[0115] Category 1: Wild-type peptides have weak MHC-I / HLA-I binding and weak TCR recognition, while mutant peptides have mutations in the MHC-I / HLA-I binding site, resulting in strong MHC-I / HLA-I and strong TCR recognition. In the absence of central tolerance, mutations at the anchor site can also activate TCR.

[0116] Category II: Wild-type peptides have weak MHC-I / HLA-I binding and weak TCR recognition, while mutant peptides have mutations at non-MHC-I / HLA-I binding sites, resulting in strong MHC-I / HLA-I and TCR recognition. In the absence of central tolerance, the mutation occurs at a non-anchor site and can activate TCR.

[0117] Category III: Wild-type peptides have strong MHC-I / HLA-I binding and weak TCR recognition, while mutant peptides have mutations at non-MHC-I / HLA-I binding sites, resulting in strong MHC-I / HLA-I and strong TCR recognition. In the presence of central tolerance, mutations at non-anchor sites can activate TCR.

[0118] Category 4: The wild-type peptide has strong MHC-I / HLA-I binding and weak TCR recognition, while the mutant peptide has a mutation in the MHC-I / HLA-I binding site, resulting in strong MHC-I / HLA-I and weak TCR recognition. In the presence of central immune tolerance, the mutation occurs at a non-anchored site and cannot activate the TCR. This is further explained as follows: In this combination, the mutation site occurs at the MHC-I / HLA-I binding site, and the sequence exposed to the TCR recognition end has not changed. Furthermore, this peptide segment is still affected by central immune tolerance and is not a suitable neoantigen peptide.

[0119] Through the above steps, the present invention infers features that play an important role in peptide and MHC-I / HLA-I recognition based on biological features and attention scores, including: characterizing MHC-I / HLA-I sequences, obtaining key amino acid residue combinations and their locations; predicting the role of amino acid residue combination types and their different positions in peptide recognition, binding, and presentation in a binary classification model; obtaining important amino acid residue combinations at key sites based on the attention score; inferring key peptide binding sites, obtaining key amino acids and their locations. Utilizing the obtained peptide binding site information, wild-type / mutant peptides are classified into MHC-I / HLA-I binding and TCR recognition types based on the location of the mutation in the peptide, introducing the impact of central tolerance on neoantigen screening, and further eliminating neoantigen peptides based on the classification results to obtain more accurate neoantigens.

[0120] Example 2

[0121] The present invention obtains protein sequences from the UniProt protein database to construct a word segmenter; obtains positive peptides and MHC-I / HLA-I binding sequences from the IEDB, and generates negative peptide segments for the positive sequences; constructs a protein characterization pre-training model for the protein sequence after word segmentation by the word segmenter; converts the MHC-I / HLA-I sequence and peptide segment after word segmentation by the word segmenter into a digital representation format and splices them together, inputs the protein characterization pre-training model for fine-tuning, and finally obtains a binary classification model that can predict the binding potential of peptide segments and MHC-I / HLA-I; according to the model generated by training and the characteristics of the algorithm, extracts the key action sites and motifs in the protein sequence to characterize the MHC-I / HLA-I sequence; uses the position of tumor-specific mutations on the antigen peptide and the binding potential with MHC to further classify the antigen peptide, eliminates the antigen peptides that cannot be recognized by TCR due to the influence of central immune tolerance, and finally obtains more accurate and comprehensive neoantigens. Specifically comprising the following steps:

[0122] 2.1、Word segmenter construction and use

[0123] Use it for training, set the vocabulary size to 10,000, and the word segmentation effect is as follows:

[0124] Before participle: MIINPTSDPEVSALEKKNTGRIAQIIGPVLDVTFPPGKMP……

[0125] After participle: MII NP TTSD PE VS ALEK KN TGRI AQI IGP VLD VT FP PG KMP……

[0126] 2.2 Pre-training model construction

[0127] A natural language model was used to construct a pre-training model for protein representation. Unsupervised training was performed on the segmented protein sequences to learn the language characteristics and contextual connections of the protein sequences. The mask ratio was set to 15%, and 568,745 protein sequences were used with a total sentence piece length of 206M.

[0128] 2.3 Fine-tuning the protein representation pre-training model

[0129] 2.3.1. Generate corresponding negative peptides based on the positive peptides that bind to MHC-I / HLA-I. Segment the MHC-I / HLA-I sequences. Label the positive and negative data entries as 0 / 1, respectively. The distribution of the positive data set is as follows:

[0130] Table 1

[0131] As shown in Table 1, there are only 121 HLA types, so it is very difficult to perform protein feature recognition using such a limited dataset. This is why a protein characterization pre-training model is first constructed.

[0132] 2.3.2. Example of the raw data format input to the protein characterization pre-training model:

[0133] EIEICDGF,M RVT AP RTLL LLL WG AV ALT ET WAG SH SM RY FH TS VS RPG RG EP RFI TVG YVD DT…….AGL AV------L AVV-VIG AV VAAVM-CRR KS SGG KGG SY SQ AAC SD S AQG SD VSL TA------------,0

[0134] 2.4 Application of binary classification model

[0135] 2.4.1. Input the binary classification model according to the data format in 2.3.2 above to predict the binding potential of peptides to MHC-I / HLA-I. The input results are shown in Table 2:

[0136] Table 2

[0137] Among them, Label is the actual label; prediction is the model prediction label;

[0138] probability is the model prediction result [A, B], A is the probability of the result being negative, and B is the probability of the result being positive

[0139] 2.4.2. Evaluate the binary classification model.

[0140] The evaluation dataset is a validation set independent of the dataset used for model training. The composition of the validation set is shown in Table 3:

[0141] Table 3

[0142] The model performance is calculated using the following formula:

[0143] TP: true positive; TN: true negative; FP: false positive; FN: false negative;

[0144] AUC is a plot of the true positive rate (TPR) versus the false positive rate (FPR).

[0145] The model performance was evaluated by comparing it with the commonly used NetMHCpan4.1 software. The results are shown in Table 4:

[0146] Table 4

[0147] 2.4.3. Extract MHC-I / HLA-I binding sites and obtain the locations of key amino acid residues. Explain the model's prediction performance from a biological perspective and plot a heat map based on the attention scores of different sites. The brighter the color, the more likely it is to be a key binding site between the peptide and MHC. This indirectly demonstrates the good performance of the model and can be used for subsequent key site feature extraction of MHC-I / HLA-I sequences.

[0148] 2.4.4. Characterize protein sequences, obtain the combination and location of key amino acid residues, and explain the model prediction performance from a biological perspective. Draw a heat map of the attention score of each position on the HLA-A / B / C sequence, infer the total key binding sites of the MHC-I / HLA-I sequence, and help better characterize the MHC-I / HLA-I sequence. Compared with the common hot spots provided by LKMHCpan or MHCflurry (31,33,48,69,83,86,87,90,91,93,94,97,98,100,101,104,1 05,108,115,119,121,123,126,138,140,142,167,171,174,176,180,182,183,187,191,195,223) is more comprehensive and representative, focusing on the functional motif sequences of hot spots, and extracting more key sites beyond the above sites, such as sites 39,235, etc., which are more in line with the functional significance of the protein sequence.

[0149] 2.5. Add classification information of peptide binding to MHC-I / HLA-I and TCR recognition to promote the identification of new antigen sequences in a more detailed and comprehensive manner.

[0150] Based on the above analysis, the peptides were analyzed to obtain the binding anchor positions of the peptides to MHC-I / HLA-I. The peptide types can be divided into four groups according to the binding potential of wild-type peptides / mutant peptides and the tumor-specific mutation positions, as shown in Table 5 below:

[0151] Table 5

[0152] In the neoantigen screening, the antigenic peptides generated by category 4 were eliminated to improve the accuracy of neoantigen prediction.

[0153] Example 3

[0154] Steps 3.1 to 3.4 are the same as steps 1.1 to 1.4 in Example 1.

[0155] 3.5 Fine-tuning the protein representation pre-training model

[0156] 3.5.1. Label the generated positive peptides that can bind to MHC-I / HLA-I with labels 1 / 0, as well as the negative peptides randomly generated from the positive peptides. It is worth noting that these positive data are obtained through experiments, which means that only limited labeled training data can be obtained. This is why a large amount of unlabeled protein data is used to pre-train the model.

[0157] 3.5.2 Digitally represent and concatenate the MHC-I / HLA-I sequences and peptide sequences segmented by the labeled protein segmenter in 3.5.1. The concatenation strategy is as follows:

[0158] Calculate the protein sequence cleavage probability through the peptide segment and the protein sequence in which it is located;

[0159] Calculate the binding stability and presentation ability of peptides and MHC-I / HLA-I;

[0160] The digitally represented protein sequence cleavage probability, peptide and MHC-I / HLA-I binding stability, and peptide and MHC-I / HLA-I presentation ability are spliced ​​with the digitally represented MHC-I / HLA-I sequence and peptide sequence segmented by the protein segmenter.

[0161] 3.5.3. Divide the labeled data in 3.5.2 into training set / validation set / test set according to a certain ratio.

[0162] 3.5.4. Fine-tune the pre-trained model generated using the data from 3.5.3 and use it as input to build a binary classification model.

[0163] 3.5.5. The model structure can be seen in Figure 3.

[0164] Steps 3.6 to 3.7 are the same as steps 1.6 to 1.7 in Example 1.

[0165] The present invention infers important features of peptide and MHC-I / HLA-I recognition based on biological features and attention scores. The method includes: characterizing MHC-I / HLA-I sequences, obtaining key amino acid residue combinations and their locations; predicting the role of amino acid residue combination types and their different positions in peptide recognition, binding, and presentation based on a binary classification model; obtaining important amino acid residue combinations at key sites based on attention scores; inferring key peptide binding sites, obtaining key amino acids and their locations. Using the obtained peptide binding site information, the method classifies wild-type / mutant peptides into MHC-I / HLA-I binding and TCR recognition types based on the location of the mutation in the peptide. The method also considers the effects of pMHC binding stability, pMHC presentation ability, and central nervous system tolerance on neoantigen screening. Based on the classification results, neoantigen peptides are further eliminated to obtain more accurate neoantigens.

[0166] Example 4

[0167] The present invention obtains protein sequences from the UniProt protein database to construct a word segmenter; obtains positive peptides and MHC-I / HLA-I binding sequences from the IEDB, and generates negative peptide segments for the positive sequences; constructs a protein representation pre-training model for the protein sequence after word segmentation by the word segmenter; converts the MHC-I / HLA-I sequences and peptide segments after word segmentation by the word segmenter into a digital representation format; calculates the protein sequence cleavage possibility based on the peptide segments and the protein sequences in which they are located; calculates the binding stability and the presentation possibility after binding based on the peptide segments and MHC-I / HLA-I; and converts the above digital representations of the MHC-I / HLA-I sequences and peptide segments and their corresponding peptide cleavage possibilities, pMHC binding stability and pM The HC presentation possibilities are spliced ​​together using torch.cat(). The pre-trained protein representation model is input and fine-tuned. The original pre-trained model is inherited and a fully connected layer and a classification layer are added to finally obtain a binary classification model that can predict the binding potential of peptides and MHC-I / HLA-I. Based on the trained model and algorithm characteristics, key sites and motifs in the protein sequence are extracted to characterize the MHC-I / HLA-I sequence. The location of tumor-specific mutations on the antigen peptide and the binding potential with MHC are used to further classify the antigen peptides, eliminating antigen peptides that cannot be recognized by TCR due to central immune tolerance, and ultimately obtaining more accurate and comprehensive neoantigens. The specific steps include:

[0168] Steps 4.1 to 4.2 are the same as 2.1 to 2.2 in Example 2

[0169] 4.3 Fine-tuning the protein representation pre-training model

[0170] 4.3.1. Generate corresponding negative peptides based on the positive peptides that bind to MHC-I / HLA-I. Segment the MHC-I / HLA-I sequence. Label the positive and negative data entries as 0 / 1, respectively. The distribution of the positive data set is shown in Table 6 below:

[0171] Table 6

[0172] 4.3.2. Calculate the characteristics that affect peptide binding to MHC-I / HLA-I, including the peptide's cleavage probability, pMHC stability, and pMHC presentation probability. For example, this can be calculated as follows:

[0173] Use the software pepsickle-v0-2-2 to input the peptide sequence and output the cleavage probability score of each site. The cleavage probability of the peptide is the average cleavage score of the C- / N-terminal amino acid of the peptide, ranging from [0,1]. The larger the score, the higher the cleavage probability.

[0174] Use the software LKMHCstablepan-v1 to input the peptide sequence and MHC-I / HLA-I sequence, calculate the pMHC binding stability, and output the pMHC binding stability score. The higher the score, the more stable it is.

[0175] The software DeepTAP-v1.0.1 was used to input the peptide sequence, calculate the possibility of presentation on the cell surface after pMHC binding, and output the probability score of presentation on the cell surface after pMHC binding, ranging from [0,1]. The larger the score, the higher the possibility of presentation.

[0176] 4.3.3. Example of the raw data format input into the protein characterization pre-training model:

[0177] EIEICDGF,M RVT AP RTLL LLL WG AV ALT ET WAG SH SM RY FH TS VS RPG RG EP RFI TVG YVD DT…….AGL AV------L AVV-VIG AV VAAVM-CRR KS SGG KGG SY SQ AAC SD S AQG SD VSL TA------------, 0.057, 0.20, 0.12, 0 (corresponding to: peptide sequence, MHC-I / HLA-I sequence, peptide cleavage possibility, pMHC binding stability score, pMHC binding presentation possibility score, and positive / negative)

[0178] After the MHC-I / HLA-I sequences and peptides segmented by the tokenizer are converted into sequences in digital format, they are spliced ​​along the specified dimensions with the peptide cleavage probability score, pMHC binding stability score, and pMHC presentation probability score, and input into the protein representation pre-training model for model training.

[0179] 4.4 Application of binary classification model

[0180] 4.4.1. Input the binary classification model according to the data format in 4.3.3 to predict the binding potential of peptides to MHC-I / HLA-I. The input results are shown in Table 7:

[0181] Table 7

[0182] Among them, Label is the actual label; prediction is the model predicted label; probability is the model prediction result [A, B], A is the probability of the result being negative, and B is the probability of the result being positive

[0183] 4.4.2. Evaluating the binary classification model

[0184] The evaluation dataset is the same as that in 2.4.2 of Example 2:

[0185] The formula for calculating model performance is the same as 2.4.2 in Example 2:

[0186] Compared with the common NetMHCpan4.1 software and the binary classification model constructed in Example 2, the model performance evaluation was performed, as shown in Table 8:

[0187] Table 8

[0188] Steps 4.4.3 to 4.4.4 are the same as 2.4.3 to 2.4.4 in Example 2.

[0189] Step 4.5 is the same as step 2.5 in Example 2.

[0190] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and essence of the present invention shall fall within the scope of the present invention.

Claims

1. A method for identifying peptides that bind to MHC-I / HLA-I class and are recognized by TCR, characterized in that: include: Obtain protein sequences, MHC-I / HLA-I sequences and peptide sequences; Using the protein sequence to train a protein sequence segmenter; Using the protein sequence segmented by the protein sequence segmenter to train a protein representation model, to obtain a trained protein representation pre-training model; Using the MHC-I / HLA-I sequence and peptide sequence that have been segmented by the protein sequence segmenter, fine-tuning the trained protein representation pre-training model to obtain a binary classification model; as well as The binary classification model is used to identify and process the target wild-type / mutant peptide segments, classify the mutant peptide segments, and obtain candidate neoantigen mutant peptides that are positively correlated with TCR recognition; Wherein, the MHC-I refers to major histocompatibility complex class I; the HLA-I refers to major histocompatibility complex class I in humans; and the TCR refers to T cell receptor.

2. The method according to claim 1, characterized in that The peptide sequence includes a positive peptide sequence and a negative peptide sequence, wherein the positive peptide sequence is a peptide that binds to MHC-I / HLA-I, and the label is 1; the negative peptide sequence is a peptide sequence randomly generated in the protein sequence according to the characteristics of the positive peptide sequence, and the label is 0.

3. The method according to claim 2, characterized in that The protein sequence includes collecting protein sequences from a protein database; accordingly, the collected protein sequences are used to perform word segmentation training on high-frequency repeated amino acid residue combination motifs to obtain a protein sequence word segmentation device.

4. The method according to claim 3, characterized in that The method of training a protein representation model using the protein sequence after word segmentation by the protein sequence word segmenter to obtain a trained protein representation pre-training model comprises: The unlabeled protein sequence is input into the protein sequence tokenizer for tokenization, and the protein sequence after the amino acid functional residue combination motif segmentation is obtained; Converting the protein sequence after segmentation of the amino acid functional residue combination motif into a protein sequence represented by a number; and The protein sequence represented by the digits is used to train a protein characterization model to obtain a trained protein characterization pre-training model.

5. The method according to claim 4, characterized in that The method of fine-tuning the trained protein representation pre-training model using the MHC-I / HLA-I sequence and peptide sequence after word segmentation processing by the protein sequence word segmenter to obtain a binary classification model includes: Inputting the labeled MHC-I / HLA-I sequence into a protein sequence tokenizer for tokenization processing to obtain an MHC-I / HLA-I sequence segmented by amino acid functional residue combination motifs, and converting the MHC-I / HLA-I sequence segmented by amino acid functional residue combination motifs into a digitally represented MHC-I / HLA-I sequence; Converting the labeled peptide sequence into a digital peptide sequence; and The MHC-I / HLA-I sequence converted into digital representation and the peptide sequence converted into digital representation are spliced ​​to obtain a spliced ​​sequence, and the trained protein characterization pre-trained model is fine-tuned using the spliced ​​sequence to obtain a binary classification model.

6. The method according to claim 5, characterized in that The splicing process further includes converting the MHC-I / HLA-I sequence into a digital representation, converting the peptide sequence into a digital representation, and converting other features related to the binding of the peptide and MHC-I / HLA-I into a digital representation, wherein the other features are selected from one or more of the cleavage possibility of the protein sequence, the binding stability of MHC-I / HLA-I and the peptide, and the presentation possibility of the MHC-I / HLA-I and peptide complex.

7. The method according to claim 6, characterized in that Prior to the fine-tuning step, the method further comprises calculating other features associated with the binding of peptides to MHC-I / HLA-I using the protein sequence, MHC-I / HLA-I sequence and peptide sequence.

8. The method according to claim 5, characterized in that The identification process of the target wild-type / mutant peptide segment using the binary classification model includes: Inputting the target wild-type / mutant peptide into the binary classification model, outputting the attention score of each site of the mutant peptide predicted to be positive, and the attention score of the amino acid functional residue combination motif of the MHC-I / HLA-I sequence; and According to the attention score of each site of the positive mutant peptide and the attention score of the MHC-I / HLA-I sequence motif, the anchor position when the MHC-I / HLA-I sequence binds to the positive peptide and the key characterization site of the MHC-I / HLA-I sequence are determined, and the positive mutant peptide is classified using the anchor point when the MHC-I / HLA-I sequence binds to the peptide and the mutation site of the peptide.

9. The method according to claim 8, characterized in that The classification process using the anchor position when the MHC-I / HLA-I sequence is combined with the positive mutant peptide segment and the mutation position relative to the wild-type peptide segment includes: According to the anchor point when the MHC-I / HLA-I sequence binds to the peptide segment, and in combination with the location of the mutation, the positive mutant peptide segment is classified and processed, and according to the classification result, it is determined whether it is suitable to be selected as a candidate neoantigen mutant peptide positively correlated with TCR recognition.

10. The method according to claim 9, characterized in that The method of judging whether a candidate neoantigen mutant peptide is suitable for selection as a positively correlated candidate for TCR recognition according to the classification result comprises: When the wild-type peptide has weak MHC-I / HLA-I binding and weak TCR recognition, and the mutant peptide has a mutation at the MHC-I / HLA-I anchor site, which turns into strong MHC-I / HLA-I and strong TCR recognition, it can be selected as a neoantigen mutant peptide; When the wild-type peptide has weak MHC-I / HLA-I binding and weak TCR recognition, and the mutant peptide has a non-MHC-I / HLA-I anchor site mutation and turns to strong MHC-I / HLA-I and strong TCR recognition, it can be selected as a neoantigen mutant peptide; When the wild-type peptide has strong MHC-I / HLA-I binding and weak TCR recognition, and the mutant peptide has a non-MHC-I / HLA-I anchor site mutation, which is transformed into strong MHC-I / HLA-I and strong TCR recognition, it can be selected as a neoantigen mutant peptide; When the wild-type peptide has strong MHC-I / HLA-I binding and weak TCR recognition, and the mutant peptide has a mutation at the MHC-I / HLA-I anchor site, which turns to strong MHC-I / HLA-I and weak TCR recognition, it is judged that it cannot be selected as a neoantigen mutant peptide.

11. An electronic device, characterized in that: include: Memory; processor; and a computer program; wherein the computer program is stored in the memory and is configured to be executed by the processor to implement the method according to any one of claims 1 to 10.

12. A computer-readable storage medium, characterized in that: A computer program is stored thereon; the computer program is executed by a processor to implement the method according to any one of claims 1 to 10.

Images