A pHLA-TCR recognition prediction method based on hierarchical fusion of protein sequence and molecular surface layer

Abstract

The application discloses a pHLA-TCR recognition prediction method based on protein sequence and molecular surface hierarchical fusion, and belongs to the technical field of computational immunology and artificial intelligence assisted biomolecular recognition. The method respectively utilizes HLA and antigen peptide corresponding to pHLA sequence, TCR alpha chain sequence and TCR beta chain sequence corresponding to paired TCR alpha beta sequence, and independent pHLA and TCR monomer surface information to obtain sequence recognition score and surface recognition score. Based on the sequence recognition score, a gating coefficient and a correction coefficient are generated according to the prediction uncertainty, model prediction difference, sequence similarity and structure reliability of each input, the difference between the surface recognition score and the sequence recognition score is scaled, and then the sequence recognition score is adaptively corrected. When the monomer conformation set is provided, the conformation variance can be used for variance type high-order correction. The application does not need to construct a pHLA-TCR ternary complex in advance, and can be more widely used for pHLA-TCR recognition determination, candidate sorting, antigen screening and TCR engineering.

Description

Technical Field

[0001] This invention belongs to the fields of computational immunology, artificial intelligence-assisted biomolecular recognition, and bioinformatics. It relates to a pHLA-TCR recognition prediction method based on hierarchical fusion of protein sequence and molecular surface. More specifically, it relates to a method that combines HLA, antigenic peptide, and paired TCRαβ sequence information with mutually independent pHLA and TCR monomer molecular surface information, and realizes pHLA-TCR recognition probability prediction through hierarchical adaptive correction and optional conformational variance correction. Background Technology

[0002] The recognition of peptide-human leukocyte antigen (pHLA) complexes by T cell receptors (TCRs) determines antigen-specific T cell immune responses. Accurately determining whether a given pHLA can effectively recognize a paired TCRαβ can be used for tumor antigen screening, therapeutic TCR screening, TCR engineering optimization, and prioritizing candidate vaccine antigens.

[0003] However, the potential pHLA-TCR combinatorial space is vast, while the amount of experimentally supported identification data is limited and unevenly distributed. Existing sequence prediction methods typically rely primarily on the sequence information of the antigen peptide, HLA, and TCR. While these methods are easy to apply on a large scale, they struggle to directly utilize the local geometry and physicochemical properties of molecular surfaces. On the other hand, structure prediction methods and some multimodal methods depend on the structure of the pHLA-TCR ternary complex or pre-constructed binding interfaces. However, the number of experimentally resolved ternary complexes is limited, and the predicted protein complex binding orientation and interfacial contact introduce additional uncertainties.

[0004] The aforementioned problems are particularly prominent in negative sample construction. For pHLA-TCR combinations that do not actually undergo recognition, there is usually no unique, true binding conformation. If a ternary complex is first artificially constructed and then interface features are extracted, artificial interfaces that do not correspond to real biological processes may be generated. In addition, existing multimodal models often directly fuse sequences and structures as parallel inputs, without fully considering the differences in availability, stability, and sample-level reliability of different information sources. Moreover, structural information may actually reduce prediction stability on some samples.

[0005] At the same time, a single structure cannot fully characterize the conformational dynamics of biomolecules and cannot cover the local surface geometry and physicochemical features. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a pHLA-TCR recognition and prediction method based on hierarchical fusion of protein sequence and molecular surface. The core of this invention is not to directly fuse sequence and structural information as parallel modalities of equal importance, but rather to first establish the foundation for recognition based on sequence information prediction, and then adaptively correct the incremental information provided by the monomer molecule surface based on model prediction uncertainty, model prediction differences, similarity of input sequences, and structural reliability. Furthermore, when multiple monomer conformational states exist, the statistical variance of the incremental information of molecular structures in different monomer conformations can be used as an additional correction.

[0007] This invention is achieved using the following technical solution: 1) Obtain the pHLA sequence and paired TCRαβ chain sequence of the sample to be predicted; if the sample to be predicted has the experimentally resolved pHLA-TCR complex structure, then separate pHLA and TCR from the resolved protein structure; otherwise, construct the protein structures of pHLA and paired TCRαβ chain sequences respectively by calculation method to obtain the pHLA monomer structure and TCR monomer structure. The two structures remain independent of each other in the subsequent surface generation and reasoning process. 2) The HLA sequence and antigenic peptide sequence corresponding to the pHLA sequence, and the TCRα and TCRβ chain sequences corresponding to the paired TCRαβ chain sequences are input into the sequence model. The sequence model uses a pre-trained protein language model to obtain the residue-level representations of the HLA sequence, antigenic peptide sequence, TCRα chain sequence, and TCRβ chain sequence. The residue-level representations of the HLA sequence and antigenic peptide sequence are encoded by pHLA, and the residue-level representations of the TCRα chain sequence and TCRβ chain sequence are specifically encoded by TCR chain, and then cross-molecular context encoding is used to obtain the global sequence representation. The sequence recognition score is output by a multilayer perceptron. ; 3) Molecular surfaces were constructed based on the pHLA and TCR monomer structures, respectively. The molecular surfaces were divided into local patches, and five types of surface features were extracted: hydrogen bonding features, charge, hydrophobicity, curvature, and shape index. Each surface feature was input into the surface model, and after geometric convolution, cross-surface spatial encoding, and cross-surface feature property encoding, the surface recognition score was output by a multilayer perceptron. ; 4) Based on , In addition to sample characteristics such as model prediction uncertainty, model prediction difference, sequence similarity and structural reliability, the adaptive gating coefficient g and the correction coefficient α are calculated. 5) with As a basic recognition score, for and The difference (incremental information) is adaptively scaled to form a hierarchical adaptive correction, resulting in a hierarchical fusion recognition score. The recognition probability after hierarchical fusion is obtained through the Sigmoid function. ; 6) Based on the recognition probability after hierarchical fusion, complete the identification and determination of pHLA-TCR, or prioritize candidate antigens, candidate TCRs and their combinations.

[0008] Furthermore, in step 1), the pHLA monomer structure and the TCR monomer structure can be constructed using the pHLA structural model and the TCR structural model, respectively.

[0009] Further, step 2) specifically involves: the sequence model using a pre-trained protein language model to characterize the sequences of different molecular components (the HLA sequence and antigenic peptide sequence corresponding to the pHLA sequence, and the TCRα chain sequence and TCRβ chain sequence corresponding to the paired TCRαβ chain sequence). For molecular component m, a residue-level representation is obtained after sequence characterization using the pre-trained protein language model. : in, Φ represents any input sequence from HLA, antigenic peptide, TCRα chain, or TCRβ chain, and Φ represents a pre-trained protein language model.

[0010] The residue-level representations of the TCRα and TCRβ chains were obtained by chain-specific encoding and then pooling. and Furthermore, the global characterization of TCR can be obtained based on the following formula. : Where || denotes vector concatenation. Indicates the concatenation weight. This indicates splicing deviation.

[0011] The residue-level representations of HLA and antigenic peptides are first vectorized and then pHLA encoded to obtain a global pHLA characterization. , and After vector concatenation, a global sequence representation is obtained through cross-molecular context encoding, and then a multilayer perceptron outputs the sequence recognition score. And obtain the sequence recognition probability. Furthermore, in step 3), the surface model independently performs geometric convolution for each surface feature, specifically as follows: For the polar coordinates of vertex j in local patch p The k-th learnable Gaussian kernel can be represented as: in, and Let $\mathbf{j}$ represent the radial distance from vertex $j$ to the center of the patch, and $\mathbf{j}$ represent the angle of vertex $j$ relative to the reference direction, respectively. This represents the radial center of the k-th Gaussian kernel within the patch. These represent the angle from the k-th Gaussian kernel to the center of the patch, respectively. This represents the radial coverage width of the k-th Gaussian kernel in the patch. This represents the angular coverage width of the k-th Gaussian kernel in the patch. This represents the angular distance function, describing and The shortest angular distance.

[0012] For surface feature c, local patches The The output vector corresponding to each Gaussian kernel for: in, Localized plaques The number of vertices within, Let be the surface feature value corresponding to vertex j in local patch p.

[0013] After the surface features of pHLA and TCR are processed by the geometric convolution process described above, the pHLA patch output matrix and TCR patch output matrix are obtained. The patch output matrices of pHLA and TCR are merged and then processed by cross-surface spatial encoding to obtain the representation vector of each surface feature. The representation vectors of multiple surface features are merged and then processed by cross-surface feature property encoding fusion. Finally, the surface recognition score is output by the multilayer perceptron. And obtain the surface recognition probability. .

[0014] Furthermore, all encoding methods used in this application are implemented based on the Transformer encoder.

[0015] Further, in step 4), the features of each sample are combined into a 27-dimensional feature vector, wherein the sample features include two model prediction features ( and ), 2 model prediction uncertainty characteristics ( and ), 4 structural reliability features (DOPE, plddt, pTM, ipTM), 4 model prediction difference features ( , , and ) and 15 sequence similarity features.

[0016] The uncertainty in model predictions can be represented using binary entropy: Represents the probability of sequence recognition Or surface recognition probability .

[0017] Differences in model predictions include: probability differences Uncertainty difference , and These represent the prediction uncertainties of the sequence model and the surface model, respectively; probability product features. ,in This represents Z-score standardization; and signed logit product features. ,in For symbolic functions, Structural reliability is used to characterize the credibility of pHLA and TCR structural models, and can be represented by the DOPE score of pHLA structural models (such as PANDORA) and the pLDDT, pTM, and ipTM confidence scores of TCR structural models (AlphaFold3, tFold, etc.). Specifically, DOPE represents discrete optimized protein energy, and the DOPE score is used to assess whether the model conforms to natural laws physically / statistically (lower is better); pLDDT represents the predicted local distance difference test, and the pLDDT confidence score is used to assess the accuracy of predictions for individual amino acid residues (local) (higher is better); pTM represents the predicted template matching score, and the pTM confidence score is used to assess the accuracy of predictions for the entire protein (global) topology (higher is better); ipTM represents the predicted interface template matching score, and the ipTM confidence score is used to assess the accuracy of binding between multimeric complexes (interface) (higher is better).

[0018] Sequence similarity can be calculated in five sequence spaces: antigen peptide, TCRα CDR3, TCRβ CDR3, TCRα variable region and TCRβ variable region. The maximum similarity (5) is extracted from positive training samples of the same HLA, the maximum similarity (5) is extracted from positive training samples of different HLAs, and the mean (5) of highly similar samples, for a total of 15 sequence similarity features.

[0019] The 27-dimensional feature vector constructed from the sample features As a regression model Input, and obtain the model's relative loss. The predicted value.

[0020] Again The predicted values ​​are standardized and then gating coefficients are obtained using the Sigmoid function. .

[0021] When the model relative loss When the predicted value is positive, it indicates that the surface model has a lower loss than the sequence model, and the gating coefficient is correspondingly increased; when the predicted value is negative, the incremental information provided by the molecular surface is suppressed. Incremental information is defined. : Furthermore, a vector composed of sequence recognition score, incremental information, and selected sample features (based on the prediction results of the regression model, analyzing the importance of different sample features and selecting several sample features with high importance) is used as the modified model. The input yields the correction coefficient: in, This indicates the characteristics of the filtered samples. This represents the correction coefficient for sample i. This represents the sequence recognition score of sample i. Indicates incremental information. This represents the surface recognition score of sample i.

[0022] Among them, regression model The training method is to use feature vectors As input, the relative loss of the model, representing the prediction accuracy of the sequence model and the surface model, is used. The regression model is trained to produce the output. Among them, the relative loss value of the prediction accuracy of the sequence model and the surface model. The calculation formula is: in, For sequence model prediction error, For surface model prediction error, and Let represent the sequence model prediction probability and the surface model prediction probability of sample i, respectively. This represents the binary label that the experimental test can identify for sample i. An output of 1 indicates that it can be identified, and an output of 0 indicates that it cannot be identified.

[0023] In step 5), the surface residual correction amount for sample i Recognition score after hierarchical fusion and the recognition probability after hierarchical fusion They are respectively: Furthermore, the sequence model and the surface model are trained using cross-validation, and the models are corrected using samples that were not used in the training. and regression model Training is then performed. This method ensures that the features received by both the regression model and the correction model come from samples that have not participated in model training, thereby reducing information leakage caused by using in-train fitting results.

[0024] Furthermore, when the dynamic trajectories of pHLA and TCR monomers are obtained, multiple monomer conformations can be derived, and conformational variance correction can be performed based on incremental information. The specific method is as follows: Clustering was performed on the dynamic trajectories of pHLA and TCR monomers respectively to obtain the conformational cluster sets of pHLA and TCR monomers. Let the occupancy rate of pHLA conformational cluster m be... The occupancy rate of TCR conformation cluster n is Since the two types of monomer conformations are sampled independently, their joint occupancy rate It can be represented as: For each conformation combination (m,n), maintain the sequence recognition score. The surface model predictions are regenerated based on the conformation combination (m,n), without changing the original values. Gating coefficient and correction factor The surface residual correction for the conformational combination (m,n) is obtained as follows: The weighted mean is calculated from the surface residual corrections of all conformational combinations. and weighted variance : Then, conformational variance correction is performed based on incremental information, specifically as follows: This represents the recognition score after hierarchical fusion with conformational variance correction. Adjust the weights for the conformational variance (which can be adjusted as needed).

[0025] The beneficial effects of this invention are as follows: 1. This invention directly constructs molecular surfaces from the independent pHLA and TCR monomer structures, without requiring prior acquisition or calculation of the pHLA-TCR ternary binding conformation. This avoids the circular dependence of "assuming the binding structure before prediction and recognition" and reduces geometric errors caused by protein complex binding direction and interfacial contact, as well as the influence of negative sample artificial interfaces on the prediction results.

[0026] 2. This invention adopts a hierarchical adaptive fusion method, which uses sequence recognition score as the basis and dynamically adjusts the surface residual correction amount by utilizing model prediction uncertainty, structural reliability, model prediction difference and sequence similarity. Compared with the direct fusion method that uses sequence and structure as equivalent input, it can reduce the interference of unreliable structural information on the prediction results.

[0027] 3. This invention provides a computable form for gating coefficients, correction coefficients, and hierarchical fusion. Furthermore, it utilizes samples not involved in training in the cross-validation method to train the correction model and regression model, thereby isolating the regression model, correction model, surface model, and sequence model from each other, thus improving the feasibility and training stability of the technical solution.

[0028] 4. When obtaining the dynamic trajectory of a single entity, the present invention can further perform conformational variance correction through the incremental information brought by the conformational combination, so that the model can obtain multi-conformation rationality information not contained in a single conformation.

[0029] 5. This invention can sort candidate pHLA-TCR combinations according to recognition probability, and is applicable to applications such as antigen screening, TCR screening, TCR engineering, and priority ranking of vaccine candidate antigens. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the overall process of the pHLA-TCR identification and prediction method of the present invention.

[0031] Figure 2 This is a schematic diagram illustrating the working principle of the hierarchical adaptive correction module of the present invention. Detailed Implementation

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments. The following embodiments are used to illustrate possible implementations of the present invention and are not intended to limit the scope of protection of the present invention; without departing from the core technical concept of the present invention, equivalent substitutions can be made for pre-trained protein language models, structural modeling software, surface representation forms, regression models, neural network structures, conformation sampling methods or clustering methods to implement the present invention.

[0033] Example 1: Training and Independent Testing of a Hierarchical Fusion Model like Figure 1This application discloses a pHLA-TCR recognition and prediction method based on hierarchical fusion of protein sequence and molecular surface, which relies on a hierarchical fusion model and specifically includes the following steps: 1) Obtain the pHLA sequence and paired TCRαβ chain sequence of the sample to be predicted, and obtain the pHLA monomer structure and TCR monomer structure respectively; 2) Input the HLA sequence and antigen peptide sequence corresponding to the pHLA sequence, and the TCRα chain sequence and TCRβ chain sequence corresponding to the paired TCRαβ sequence into the sequence model to obtain the sequence recognition score; 3) Construct molecular surfaces based on the pHLA monomer structure and TCR monomer structure respectively, and extract one or more surface features including at least hydrogen bond features, charge, hydrophobicity, curvature and shape index, input the surface features into the surface model to obtain the surface recognition score; 4) Based on the sequence recognition score and 5) Based on the surface recognition score, and sample features such as model prediction uncertainty, model prediction difference, sequence similarity, and structural reliability, determine the adaptive gating coefficient and correction coefficient; 6) Based on the adaptive gating coefficient and correction coefficient, adaptively correct the difference between the surface recognition score and the sequence recognition score to obtain the surface residual correction amount; based on the surface residual correction amount and the sequence recognition score, obtain the hierarchical fusion recognition score; and based on the hierarchical fusion recognition score, obtain the hierarchical fusion recognition probability; 7) Based on the hierarchical fusion recognition probability, identify and determine pHLA-TCR, or prioritize candidate antigens, candidate TCRs, and their combinations.

[0034] This embodiment collects HLA class I pHLA-TCR positive recognition records, with data sources including TRAIT, VDJdb, McPAS-TCR, and TCR3d. Each record contains at least the antigen peptide sequence, HLA subtype, paired TCRαβ CDR3 sequence, and corresponding V / J gene information. For records that only provide CDR3 and V / J gene information, the complete TCRαβ variable region amino acid sequence can be obtained using a TCR sequence reconstruction tool; HLA sequences are obtained from standard immunogenetic databases with the signal peptide removed. After preprocessing, the records are unified into HLA sequences, antigen peptide sequences, TCRα variable region sequences, TCRβ variable region sequences, and binary tags.

[0035] In a specific data implementation, the training set contained 28,511 positive records, and the independent test set contained 840 positive records. A 1:10 ratio of positive to negative samples was used during training, while a 1:1 balanced ratio was used during testing. Negative training samples were obtained by cross-matching antigenic peptides, HLA, or paired TCRs from different samples, avoiding the construction of non-existent ternary complex interfaces in the surface model.

[0036] For structures with experimentally resolved pHLA-TCR complexes, the pHLA and paired TCR portions are separated; for structures without experimentally resolved structures, monomer structures are independently constructed using both pHLA and TCR structural models. In a preferred implementation, the pHLA monomer structure is obtained by modeling the pHLA sequence using PANDORA, and the TCR monomer structure is obtained by modeling the paired TCR αβ chain sequence using tFold; pHLA-TCR docking is not performed during the entire surface generation and inference process.

[0037] The sequence model takes four inputs: HLA sequence, antigen peptide sequence, TCRα chain sequence, and TCRβ chain sequence. In a preferred implementation, an ESMC 600M pre-trained protein language model is used to generate residue-level representations, with its parameters frozen. The residue-level representations of the TCRα and TCRβ chains are TCR-specific encoded using a TCR chain-specific encoder, while the residue-level representations of the HLA and antigen peptide sequences are pHLA encoded using a pHLA encoder. The outputs of both encoders are then cross-molecular context-coded to obtain a global sequence representation. This global representation is then processed by a multilayer perceptron to output a sequence recognition score. .

[0038] Molecular surfaces are constructed for the pHLA and TCR monomer structures respectively, and the molecular surfaces are divided into local patches. A preferred implementation uses a MaSIF-like processing flow, setting the pHLA surface as up to 13 local patches centered on the antigen peptide region, and the TCR surface as up to 16 local patches centered on the paired TCR CDR3 region; each grid vertex contains local polar coordinates and five types of surface features: hydrogen bonds, charge, hydrophobicity, curvature, and shape index. Each surface feature is input into the surface model, and after geometric convolution, the statistical correlation between pHLA and TCR patches is learned by a cross-surface spatial encoding module. Then, the five types of surface features are fused by a cross-surface property encoding module, and finally, a multilayer perceptron outputs the surface recognition score. .

[0039] This embodiment constructs 27-dimensional sample features, including 2 model prediction features ( and ), 2 binary entropy models predict uncertainty features ( and ), 4 structural reliability features (DOPE, plddt, pTM, ipTM), 4 model prediction difference features ( , , and The study included 15 sequence similarity features. These 15 features were obtained by calculating the mean of the maximum similarity among positive training samples of the same HLA, the maximum similarity among positive training samples of different HLA, and the top five high similarities among different HLA sequences in the five sequence spaces.

[0040] The sequence model and the surface model are trained independently using five-fold hierarchical cross-validation. Each fold model generates predictions for its left-out fold, and the results are combined to obtain the out-of-fold model output before training an adaptive correction module (including regression and correction models). Figure 2 In a preferred implementation, an XGBoost regressor is used as the regression model to estimate the relative loss of the two models. After standardization, the gating coefficients are obtained by passing the Sigmoid function. Then select the 10 most important sample features from XGBoost and... and incremental information A common input correction model (AlphaNet, a multilayer perceptron with 32 hidden layers) is used to obtain the correction coefficient α. The final recognition score is calculated as follows: Calculate and recognize probability .

[0041] In the independent test set, the hierarchical fusion model of this embodiment outperforms the sequence single expert, surface single expert, and direct fusion models in terms of ROC-AUC, PR-AUC, and PPVn, as shown in Table 1. The direct fusion model directly concatenates and jointly optimizes the sequence representation and surface representation, while this invention controls the contribution of surface information through sample-level adaptive residuals.

[0042] Table 1 Performance comparison of different fusion methods on independent test sets In the out-of-distribution test set consisting of unseen antigens and unseen paired TCRs, the hierarchical fusion model of this invention also maintained good predictive performance, as shown in Table 2. This result demonstrates that, with limited training data coverage, basing the model on sequence recognition scores and incorporating monomer surface information based on sample characteristics can improve the model's stability on unseen recognition combinations.

[0043] Table 2 Performance comparison of different fusion methods on out-of-distribution test sets Furthermore, in an external screening dataset consisting of 95 designed antigenic peptides, 32 antigenic peptides were experimentally determined to be responsive. After sorting according to the output score of this invention, 21 of the top 32 candidates were experimentally responsive peptides, with a PPVn of 0.656; 4 of the top 5 candidates were responsive peptides. Therefore, the output of this invention can be directly used for priority ranking of candidate antigens or pHLA-TCR combinations.

[0044] Example 2: Variance-based "higher-order" correction based on monomer conformation sets When further utilization of conformational dynamics information is needed, molecular dynamics sampling can be performed on pHLA and TCR monomers separately, and the two types of trajectories can be clustered independently. In one specific implementation, 175 pHLA-TCR systems with experimental structures were decomposed into 350 monomers, and each monomer was subjected to 5 independent simulations of 100 ns each. pHLA was clustered according to the RMSD of the heavy atoms of the antigen peptide, and TCR was clustered according to the RMSD of the heavy atoms of the CDR1-CDR3 regions of the two chains. The largest cluster represents the conformation used to obtain the dominant conformation prediction. .

[0045] For the retained pHLA conformational cluster m and TCR conformational cluster n, let the occupancy rate of pHLA conformational cluster m be... The occupancy rate of TCR conformation cluster n is Their combined market share Represented as: Maintain sequence recognition score For each conformation combination, the surface recognition score is recalculated. Gating coefficient and correction factor Thus, the surface residual correction amount of the conformational combination (m,n) is obtained. : Then, the weighted mean of the conformational corrections was calculated. and weighted variance : and .

[0046] In the general form of a set of conformations, it is possible to utilize The recognition score is obtained, which includes the conformational mean and variance term. This represents the recognition score after hierarchical fusion with conformational variance correction. Adjust the weights for conformational variance. If only a single conformation exists, then... The method automatically degenerates into .

[0047] In the structural analysis benchmark of this embodiment, the ROC-AUC of the dominant conformation prediction is 0.683; the conformation variance correction weights are exploratoryly set. Subsequently, the ROC-AUC of the variance-based "higher-order" correction reached 0.715, and the PR-AUC and PPVn also improved simultaneously. This result indicates that the variance of the structural correction in the conformation set can provide additional information representing conformations or means aggregations that were not fully preserved.

[0048] The above embodiments demonstrate that the present invention can hierarchically integrate stable sequence recognition information with the physicochemical and geometric information provided by the surface of monomer molecules without pre-constructing the pHLA-TCR ternary complex. The surface correction magnitude can be dynamically adjusted according to the model prediction uncertainty, structural reliability, model prediction difference, and sequence similarity of specific samples. When the conformation set is available, additional corrections can be achieved by utilizing the statistical variance of the conformation correction amount, thereby completing the pHLA-TCR recognition probability prediction and candidate priority ranking.

Claims

1. A pHLA-TCR recognition and prediction method based on hierarchical fusion of protein sequence and molecular surface, characterized in that, Includes the following steps: 1) Obtain the pHLA sequence and paired TCRαβ chain sequence of the sample to be predicted, and obtain the pHLA monomer structure and TCR monomer structure respectively; 2) Input the HLA sequence and antigen peptide sequence corresponding to the pHLA sequence, and the TCRα chain sequence and TCRβ chain sequence corresponding to the paired TCRαβ sequence into the sequence model to obtain the sequence recognition score; 3) Construct molecular surfaces based on the pHLA monomer structure and TCR monomer structure respectively, and extract one or more surface features including at least hydrogen bond features, charge, hydrophobicity, curvature and shape index. Input the surface features into the surface model to obtain the surface recognition score; 4) Determine the adaptive gating coefficient and correction coefficient based on the sample features including the sequence recognition score and surface recognition score, as well as model prediction uncertainty, model prediction difference, sequence similarity and structural reliability. 5) Based on the adaptive gating coefficient and correction coefficient, the difference between the surface recognition score and the sequence recognition score is adaptively corrected to obtain the surface residual correction amount. Based on the surface residual correction amount and the sequence recognition score, the hierarchical fusion recognition score is obtained, and the hierarchical fusion recognition probability is obtained based on the hierarchical fusion recognition score; 6) The pHLA-TCR is identified and determined according to the hierarchical fusion recognition probability, or the candidate antigen, candidate TCR and their combination are prioritized.

2. The pHLA-TCR recognition and prediction method based on hierarchical fusion of protein sequence and molecular surface as described in claim 1, characterized in that, The pHLA monomer structure and TCR monomer structure are obtained as follows: if the sample to be predicted has an experimentally resolved pHLA-TCR complex structure, the pHLA part is separated from the paired TCR part. Otherwise, protein structures of pHLA and paired TCRαβ chain sequences are constructed separately to obtain pHLA monomer structures and TCR monomer structures.

3. The pHLA-TCR recognition and prediction method based on hierarchical fusion of protein sequence and molecular surface as described in claim 1, characterized in that, Step 2) specifically involves the following steps: The sequence model uses a pre-trained protein language model to obtain residue-level representations of the HLA sequence, antigen peptide sequence, TCRα chain sequence, and TCRβ chain sequence. The residue-level representations of the HLA sequence and antigen peptide sequence are encoded by pHLA, and the residue-level representations of the TCRα chain sequence and TCRβ chain sequence are specifically encoded by the TCR chain. Then, cross-molecular context encoding is used to obtain a global sequence representation, and a sequence recognition score is output by a multilayer perceptron.

4. The pHLA-TCR recognition and prediction method based on hierarchical fusion of protein sequence and molecular surface according to claim 1, characterized in that, Step 3) specifically involves: constructing molecular surfaces based on the pHLA monomer structure and the TCR monomer structure, dividing the molecular surfaces into local patches, and extracting five types of surface features: hydrogen bond characteristics, charge, hydrophobicity, curvature, and shape index. After geometric convolution, cross-surface spatial encoding, and cross-surface feature property encoding, each surface feature is output by a multilayer perceptron to determine the surface recognition score.

5. The pHLA-TCR recognition and prediction method based on hierarchical fusion of protein sequence and molecular surface according to claim 1, characterized in that, In step 4), the uncertainty of the model prediction is represented by binary entropy: ； These represent sequence recognition probability or surface recognition probability, where the sequence recognition probability is obtained based on the sequence recognition score and the surface recognition probability is obtained based on the surface recognition score. The structural reliability includes the pHLA structural model DOPE score, and the TCR structural model pLDDT, pTM, and ipTM confidence scores. The model prediction discrepancies include: probability differences. , and Represent the sequence recognition probability and surface recognition probability, respectively; uncertainty difference , and These represent the prediction uncertainties of the sequence model and the surface model, respectively; probability product features. ,in This represents Z-score standardization; and signed logit product features. ,in For symbolic functions, ； The sequence similarity was calculated in the sequence spaces of the antigen peptide, TCRα chain CDR3, TCRβ chain CDR3, TCRα chain variable region, and TCRβ chain variable region, respectively. The maximum similarity was extracted from positive training samples of the same HLA, and the maximum similarity and the mean of highly similar samples were extracted from positive training samples of different HLAs.

6. The pHLA-TCR recognition and prediction method based on hierarchical fusion of protein sequence and molecular surface according to claim 1, characterized in that, The gating coefficient is calculated as follows: a 27-dimensional feature vector consisting of the sequence recognition score, surface recognition score, model prediction uncertainty, model prediction difference, sequence similarity, and structural reliability of sample i is generated. As a regression model The input is used to obtain the predicted values ​​of the relative model loss between the sequence model and the surface model, which are the prediction accuracy of the sequence model and the surface model. Then, the predicted values ​​of the relative model loss between the sequence model and the surface model are standardized, and the gating coefficients for sample i are obtained using the Sigmoid function. .

7. The pHLA-TCR recognition and prediction method based on hierarchical fusion of protein sequence and molecular surface according to claim 6, characterized in that, The correction coefficient is calculated as follows: First, the importance of different sample features is analyzed based on the prediction results of the regression model, and several sample features with the highest importance are selected; then, the vector composed of the sequence recognition score, incremental information, and the selected sample features is used as the correction model. The input yields the correction coefficient: ； ； in, This represents the correction coefficient for sample i. This represents the sequence recognition score of sample i. This represents the feature vector of the filtered samples. This represents the incremental information of sample i. This represents the surface recognition score of sample i.

8. The pHLA-TCR recognition and prediction method based on hierarchical fusion of protein sequence and molecular surface according to claim 7, characterized in that, The method for obtaining the surface residual correction amount is as follows: ； The method for calculating the recognition score after hierarchical fusion is as follows: ； The specific recognition probability after hierarchical fusion is as follows: 。 9. The pHLA-TCR recognition and prediction method based on hierarchical fusion of protein sequence and molecular surface according to claim 1, characterized in that, The sequence model and the surface model were trained using cross-validation, and the modified model and the regression model were trained using samples that were not used in the training.

10. The pHLA-TCR recognition and prediction method based on hierarchical fusion of protein sequence and molecular surface according to claim 1, characterized in that, When the dynamic trajectories of pHLA and TCR monomers are available, the dynamic trajectories of pHLA and TCR monomers are independently clustered to obtain the conformational cluster sets of pHLA and TCR monomers. For pHLA conformational cluster m and TCR conformational cluster n, the joint occupancy rate is determined based on the occupancy rate of each conformational cluster. : ; in, and The occupancy rates of pHLA conformation cluster m and TCR conformation cluster n, respectively; And for each conformation combination (m,n), the surface model prediction value is regenerated. Gating coefficient and correction factor And calculate the surface residual correction amount. : ； in, and These are the surface recognition score and sequence recognition score corresponding to the conformation combination (m,n), respectively; The weighted average is calculated from the surface residual corrections and joint occupancy of all conformational combinations. and weighted variance : ； ； Based on incremental information, conformational variance correction is performed, and combined with the sequence recognition score, a hierarchical fusion recognition score after conformational variance correction is obtained. Specifically, it is expressed as: ； in, Adjust the weights for the conformational variance.

Images