An immune fingerprint-based prediction method for TCR-pMHC binding specificity

Abstract

The application discloses a TCR-pMHC binding specificity prediction method based on immune fingerprints. * 02-pMHC interface geometry and physicochemical characteristics (i.e. "immune fingerprints") as the core reference information, combined with a geometric deep learning network architecture based on geodesic convolution to construct a TCR-pMHC binding specificity prediction model based on immune fingerprints, for realizing TCR-pMHC binding preference prediction. The screened HLA-A * 27 system is further verified, showing good generalization ability in complex immune interaction scenarios. The method has the advantages of strong scalability, good interpretability, and suitability for various HLA subtypes, and can be widely applied in the fields of antigen design, vaccine development, TCR discovery, and immunotherapy.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of immunology and computational biology, and specifically relates to a method and device for predicting TCR-pMHC binding specificity based on immune fingerprinting. Background Technology

[0002] T cell-mediated immune responses begin with the recognition of pMHC (peptide-MHC, also known as peptide-HLA, pHLA) molecules by the TCR (T cell receptor). This interaction is a key determinant of T cell activation and subsequent immune responses. However, due to the high specificity and cross-reactivity of TCR binding, its binding behavior is complex and exhibits a certain degree of polymorphism, resulting in the TCR–pMHC interaction mechanism remaining incompletely understood.

[0003] Currently, the number of experimentally resolved 3D structures of TCR–pMHC complexes is limited, with only a few hundred cases resolved, while the number of TCR clones within an individual can reach tens of millions. This scarcity of structural data severely restricts systematic research on TCR recognition patterns. Existing studies mostly rely on qualitative analysis or visual comparison of the local interaction environment of a small number of structures to explore recognition rules, or analyze the overall footprint of TCR binding on the pMHC surface, focusing on spatial regions or physicochemical properties. However, these methods generally suffer from insufficient generalization, heavy reliance on high-quality complex structures and quantitative binding affinity data, and a lack of a learnable and interpretable modeling framework that can simultaneously integrate spatial, geometric, and chemical properties. Summary of the Invention

[0004] To address the deficiencies in the aforementioned background technology, this invention provides a TCR-pMHC binding specificity prediction method based on immune fingerprinting. This method aims to learn and apply the structural patterns recognized by TCR-pMHC.

[0005] This invention is achieved using the following technical solution: A method for predicting TCR-pMHC binding specificity based on immune fingerprinting includes the following steps: (i) Obtain the three-dimensional structural data of the target pMHC complex and construct the surface of the target pMHC complex; (ii) Triangulate the surface of the target pMHC complex to obtain a discretized pMHC surface grid. Extract radial surface patches based on the geodesic radius of each pMHC surface grid vertex. Identify the surface patches whose centers are located in the vicinity of the peptide (the center distance from any atom of the peptide is less than a predetermined threshold) as interface patches. (iii) Calculate the multimodal geometric and physicochemical features of each pMHC surface grid vertex on the discretized pMHC surface grid, construct an immune fingerprint based on the geometric and physicochemical features, and spatially locate the immune fingerprint in each interface patch using geodesic polar coordinates. (iv) Randomly sample a predetermined number of interface patches from the interface patches and input them into the TCR-pMHC binding specificity prediction model based on immune fingerprints. The TCR-pMHC binding specificity prediction model based on immune fingerprints encodes the sampled interface patches into a fingerprint matrix according to the immune fingerprints and their spatial positioning and makes predictions based on the fingerprint matrix, thereby obtaining the target pMHC complex binding preference prediction vector for different TCR types. (v) Repeat step (iv) to obtain multiple prediction vectors, and integrate the multiple prediction vectors to output the final TCR–pMHC combined preference discrimination result; (vi) Associate multiple prediction vectors with the contribution of each interface patch to generate an importance map at the patch level and an importance map at the residue level, and perform interpretability analysis based on the importance maps at the patch level and the importance maps at the residue level.

[0006] In the above technical solution, further, in step (ii), firstly, the pMHC complex is protonated using Reduce, and then the structure of the protonated complex is calculated using MSMS and pymesh. Solvent removal is performed on the surface and the result is calculated according to 1.0. Triangulation is performed at a resolution of [resolution value] to obtain a discretized pMHC surface mesh; then, with each pMHC surface mesh vertex as the center, according to [relevant data]... The geodesic radius (the geodesic distance is the shortest path length connecting two points when "walking" on a surface; the geodesic radius refers to a measure of radius using the geodesic distance) is used to extract radial surface patches, and among all surface patches, those centered near any atom on the peptide are selected. Surface patches within the range are identified as interface patches.

[0007] Further, the geometric features described in step (iii) include: shape index (calculated by Gaussian curvature and mean curvature) and distance-dependent curvature (calculated by the distance between the mesh vertices within the interface patch and the angle between the vertices and the normals); physicochemical features include: Poisson-Boltzmann continuous electrostatics (calculated by PDB2PQR and APBS); hydrophobicity index (calculated using the Kyte-Doolittle scale); and the positions of free electrons and proton donors (calculated by the positions of surface lattice points relative to their nearest neighbor polar atoms).

[0008] Further, in step (iv), the training process of the TCR-pMHC binding-specific prediction model based on immune fingerprinting includes the following steps: (1) Obtain TCR-pMHC complexes containing different TCR types, extract TCR types as binding-specific tags for pMHC complexes, extract three-dimensional structural data of pMHC complexes and construct the surface of pMHC complexes. (2) Triangulate the surface of the pMHC complex to obtain a discretized pMHC surface grid. Extract radial surface patches based on geodesic radius of each pMHC surface grid vertex and identify the surface patches whose center is located in the peptide neighborhood as interface patches. (3) Calculate geometric and physicochemical features on the discretized pMHC surface mesh, construct an immune fingerprint based on the geometric and physicochemical features, and spatially locate the immune fingerprint in each interface patch using geodesic polar coordinates; the geometric features include shape index and distance-related curvature; the physicochemical features include Poisson-Boltzmann continuous electrostatics, hydrophobicity index, and the positions of free electrons and proton donors; (4) Randomly sample a predetermined number of interface patches from the interface patches and input them into a geodesic convolution-based geometric deep learning network. The geodesic convolution-based geometric deep learning network encodes the sampled interface patches into a fingerprint matrix based on the immune fingerprint and its spatial localization, and generates a pMHC binding preference prediction vector for different TCR types based on the fingerprint matrix. (5) Repeat step (4) to obtain multiple pMHC binding preference prediction vectors for different TCR types; use the TCR type labels corresponding to the pMHC complex to perform supervised learning on the geodesic convolution-based geometric deep learning network to obtain the TCR-pMHC binding specificity prediction model based on immune fingerprint.

[0009] Furthermore, in step (4), the number of interface patches for the geodesic convolution-based geometric deep learning network is set to 32. The geodesic convolution-based geometric deep learning network adopts the MaSIF (molecular surface interaction fingerprinting) architecture, as detailed in the article "Deciphering interaction fingerprints from protein molecular surfaces using geometric deep learning".

[0010] Furthermore, in step (5), 100 independent prediction vectors are obtained by randomly sampling 100 times from the 32 interface patches. An ensemble strategy, including vector averaging or majority voting, is adopted for all prediction vectors to improve the robustness of the TCR-pMHC combined preference discrimination result. The majority voting ensemble strategy is used to integrate the 100 prediction vectors. Specifically, each prediction vector corresponds to a TCR type. The TCR types corresponding to the 100 prediction vectors are statistically analyzed, and the TCR type with the most votes is the final TCR-pMHC combined preference discrimination result. The sampled interface patches (32 patches) are encoded into a fingerprint matrix. Specifically, each interface patch is first encoded into an 80-dimensional patch descriptor, and then the 32 80-dimensional patch descriptors are combined and processed into an 80*80 fingerprint matrix.

[0011] Further, in step (vi), the method of associating multiple prediction vectors with the contribution of each interface patch to generate an importance map at the patch level and an importance map at the residue level is as follows: the patch contribution is obtained by normalizing the interface patches sampled in the high confidence (confidence score not less than 90%) predictions, and an importance map at the patch level is generated; and the patch contribution of adjacent specific peptide residues is further regionally aggregated to generate an importance map at the residue level.

[0012] The present invention also provides a TCR-pMHC binding specificity prediction device based on immune fingerprinting, which consists of the following modules: The first module is used to obtain the three-dimensional structural data of the target pMHC complex and construct the surface of the target pMHC complex. The three-dimensional structural data of the target pMHC complex can be obtained independently or locally extracted from the modeled structure of the corresponding experimental analysis or prediction software. The second module is used to triangulate the surface of the target pMHC complex to obtain a discretized pMHC surface grid. Based on the vertex of each pMHC surface grid, radial surface patches are extracted according to the geodesic radius. Among all surface patches, the surface patch whose center is located in the vicinity of the peptide is identified as an interface patch. The geodesic distance between any two pMHC surface grid vertices is approximately calculated using the Dijkstra algorithm. The third module is used to calculate the geometric and physicochemical features of each pMHC surface mesh vertex on the discretized pMHC surface mesh, construct an immune fingerprint based on the geometric and physicochemical features, and spatially locate the immune fingerprint within each interface patch using geodesic polar coordinates; the geometric features include shape index and distance-dependent curvature; the physicochemical features include Poisson-Boltzmann continuous electrostatics, hydrophobicity index, and the positions of free electrons and proton donors; The fourth module: Randomly and repeatedly sample a predetermined number of interface patches from the interface patches, and input the predetermined number of interface patches sampled each time into the TCR-pMHC binding specificity prediction model based on immune fingerprints. The TCR-pMHC binding specificity prediction model based on immune fingerprints encodes each sampled interface patch into a fingerprint matrix according to the immune fingerprint and its spatial location, and makes predictions based on the fingerprint matrix, thereby obtaining the binding preference prediction vector of the target pMHC complex for different TCR types; the predetermined number of interface patches sampled multiple times correspond to multiple prediction vectors. The fifth module is used to integrate multiple prediction vectors to output the final TCR–pMHC combined preference discrimination result; Module 6: Correlate multiple prediction vectors with the contribution of each interface patch to generate importance maps at the patch level and residue level, and perform interpretability analysis based on the importance maps at the patch level and residue level.

[0013] The beneficial effects of this invention are as follows: This invention provides a TCR-pMHC binding specificity prediction method based on immune fingerprinting. Compared with existing molecular surface modeling methods represented by MaSIF, this invention, while inheriting the technical route of "using molecular surfaces as characterization carriers," has made targeted extensions and innovations around the complex immune recognition system of TCR–pMHC, which has both high specificity and cross-immunity. Specifically, it is reflected in: (1) System validation and expansion in complex immune recognition systems.

[0014] Existing methods are mainly geared towards general protein-protein interaction tasks. This invention applies a molecular surface multimodal modeling framework system to the TCR–pMHC recognition system, validating and expanding its discrimination ability and application boundaries in scenarios with greater biological complexity.

[0015] (2) Construct an immune fingerprint of “peptide focusing” by combining domain mechanism knowledge.

[0016] By combining the peptide focusing characteristics of pMHC identified by TCR, this invention defines the surface region adjacent to the peptide as a functional interface and extracts an immunofinite fingerprint, forming an interface definition method with biological and physical rationality. Even with only high-quality pMHC monomer structures and lacking complex structures, binding preference prediction can still be achieved, improving the practical applicability of the method.

[0017] (3) Multi-scale interpretability tracing mechanism.

[0018] By statistically analyzing the importance of interface patches and mapping them to the corresponding peptide residues, combined with structural detail analysis, key interaction mechanisms are identified, thus providing a molecular-level explanation of the TCR-pMHC binding specificity and offering a theoretical basis for antigen peptide optimization and rational design.

[0019] (4) Generalization inference ability across HLA alleles.

[0020] Obtaining predictive results and mechanistic explanations consistent with experimental trends in allele systems not involved in training verifies the underlying stability and robustness of the constructed immune fingerprint framework and expands its cross-system application potential. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall framework for TCR-pMHC binding specificity prediction based on interface immunofingerprinting of the present invention. (a) The process of TCR “scanning” and recognizing the surface of pMHC molecules; (b) Extracting interface immunofingerprints from the three-dimensional structure of the pMHC complex; (c) Sampling interface patches and inputting them into the TCR-pMHC binding specificity prediction model based on immunofingerprinting to obtain TCR binding preference prediction results. Figure 2 (a) pMHC surface triangulation, radial surface patch extraction and geometric and physicochemical feature calculation process; (b) pMHC binding preference prediction vectors for different TCR types and the integration of multiple prediction vectors are obtained based on random sampling interface patches and TCR-pMHC binding specificity prediction model based on immune fingerprint. Figure 3 The diagram illustrates cross-validation on the HLA-A*02 dataset: (a) Sample distribution and training / test set partitioning for different TCR categories; (b) Discrimination accuracy and confusion for each category; (c) Discrimination confidence distribution for a single structure. Figure 4 (a) Importance distribution of different peptide sites; (b) Conserved motifs of peptide sequences; (c) and (d) Importance distribution at the pMHC interface and details of peptide-TCR residue interactions in key regions, respectively. Figure 5 The following is a schematic diagram of the generalization application of the model in the HLA-B*27 allele system: (a) structural differences under different alleles and peptide conditions; (b) combined with preference discrimination results; (c) and (d) are the key interface region of pMHC and peptide-MHC structural details near the key region, respectively. Detailed Implementation

[0022] The present invention will be further described below with reference to the embodiments.

[0023] Example 1: Overall Framework and Implementation of a TCR-pMHC Binding Specific Prediction Method Based on Immune Fingerprinting This embodiment, in conjunction with the accompanying drawings, describes the overall technical solution and implementation process of the immune specificity prediction method based on interface feature fingerprints described in this invention.

[0024] like Figure 1 As shown in Figure a, this invention formalizes the T-cell receptor (TCR) recognition process of peptide-major histocompatibility complex (pMHC) into a recognition and evaluation problem based on interfacial surface features. Specifically, by systematically extracting and modeling the geometric and physicochemical features of the peptide-adjacent interfacial region in the pMHC structure, an "immunofaminifera" characterizing the properties of the pMHC interface is constructed. Figure 1 (b) and based on the immune fingerprint, discriminative prediction of the binding preference of pMHC for different TCR types is performed. Figure 1 (c) In this example, the prediction result is a seven-dimensional normalized probability vector, where each dimension represents a TCR type. The TCR type corresponding to the dimension with the highest prediction probability is the combined preference prediction result.

[0025] In this embodiment, the construction process of the immune fingerprint is as follows: Figure 2 As shown in Figure a, the steps include: First, the three-dimensional structural data of the target pMHC complex is acquired, and the surface of the target pMHC complex is constructed. The surface of the target pMHC complex is then triangulated to obtain a discretized pMHC surface mesh. Subsequently, radial surface patches are extracted based on geodesic distances, centered on the vertices of the pMHC surface mesh. Further, based on the spatial distance between the center point of the surface patch and any atom in the peptide, surface patches located in the vicinity of the peptide (i.e., the spatial distance is less than a predetermined threshold) are identified as interface patches. These interface patches constitute the surface representation basis of the pMHC interface region. All identified interface patches are then aggregated to obtain an interface patch set.

[0026] Multimodal interface features, including geometric features (shape index and distance-dependent curvature) and physicochemical features (Poisson-Boltzmann continuity electrostatics, hydrophobicity index, and the positions of free electrons and proton donors), are calculated for each vertex of the discretized pMHC surface mesh. An immune fingerprint is constructed based on these geometric and physicochemical features. The immune fingerprint is spatially located within each interface patch using geodesic polar coordinates. The feature information carried by multiple interface patches collectively constitutes a description of the pMHC interface properties.

[0027] like Figure 2As shown in Figure b, this embodiment further employs an immunofingerprint-based TCR-pMHC binding specificity prediction model. The sampled interface patches are encoded based on the immunofingerprint and its spatial location to obtain a fingerprint matrix. Prediction is then performed based on the fingerprint matrix to obtain a pMHC binding preference prediction vector for different TCR types. Specifically, for a target pMHC complex, a predetermined number of interface patches (referred to as interface patch combinations) are randomly sampled from its interface patch set. These sampled interface patch combinations are then input into the immunofingerprint-based TCR-pMHC binding specificity prediction model for encoding to obtain a fingerprint matrix. Prediction is then performed based on the fingerprint matrix to output the corresponding TCR binding preference prediction vector.

[0028] To improve the stability and robustness of the prediction results, this embodiment employs a repeated prediction mechanism based on random interface patch sampling. Specifically, for the same pMHC structure, multiple interface patch combination sampling and predictions are performed, and the prediction vectors obtained from these multiple predictions are integrated to obtain the final TCR–pMHC combined preference discrimination result. The integration method may include vector averaging or majority voting.

[0029] Using the above method, this embodiment achieves effective identification and evaluation of pMHC's binding preference for different TCR types based solely on the structure and interface surface features of pMHC without introducing TCR structural information, and provides a unified technical framework for subsequent interpretability analysis and cross-system applications.

[0030] Multiple prediction vectors are correlated with the contribution of each interface patch to generate importance maps at the patch level and residue level. Interpretability analysis is then performed based on the importance maps at the patch level and residue level.

[0031] Example 2: Validation of discrimination performance on a specific dataset This embodiment illustrates the discriminative performance and stability of the TCR-pMHC binding specific prediction method based on immune fingerprinting proposed in this invention on a specific human pMHC structure dataset.

[0032] In this embodiment, the human leukocyte antigen HLA-A*02 system was selected as the training data source for the TCR-pMHC binding specificity prediction model based on immunofingerprinting. HLA-A*02 is the most widely distributed and structurally rich MHC allele in the population, and has high representativeness. A supervised training dataset was constructed based on the resolved TCR–pMHC complex structures in publicly available protein structure databases. Specifically, only TCR types with at least three independent structural instances in the database were retained, ensuring that each category had the most basic statistical learning conditions. The final dataset contained 40 HLA-A*02–peptide–TCR complex structures, covering 7 different TCR receptor types.

[0033] In the evaluation of the TCR-pMHC binding specificity prediction model based on immune fingerprinting, an iterative cross-validation strategy was used for validation. Figure 3 (a) For each iteration, the 40 structural samples are randomly divided into a training set and a test set, while maintaining a consistent proportion of different TCR categories in both sets to avoid class distribution bias. In each iteration, the training set contains 27 structural samples, and the test set contains 13 structural samples. Based on the above strategy, multiple independent TCR-pMHC-based specific prediction models (with different parameters) based on immune fingerprints are trained, and each model is used to predict and evaluate the corresponding test set structures.

[0034] For each pMHC complex structure in the test set, its immunofingerprint was first extracted according to the method described in Example 1. Multiple sets of interface patch combinations were then generated through random sampling and input into the discriminant model to obtain the corresponding prediction vectors. By integrating multiple prediction vectors, the final discrimination result of the pMHC structure on different TCR categories was obtained.

[0035] Experimental results show that, in the 40 HLA-A*02–pMHC structure samples, the method of this invention can effectively distinguish the binding preference of the pMHC complex to different TCR receptors, and its average discrimination accuracy is significantly higher than the theoretical baseline level corresponding to random guessing. Further analysis of the discrimination results for different TCR categories revealed ( Figure 3 (b) The TCR category with a relatively large number of samples has a higher discrimination accuracy, while the category with a small number of samples has a relatively lower discrimination performance, indicating that the model performance is reasonably consistent with the richness of structural information in the training data.

[0036] Furthermore, this embodiment also performs statistical analysis on the discrimination confidence of each structural sample ( Figure 3(c). The results show that most pMHC structures have high prediction confidence in their corresponding TCR categories, indicating that their immune fingerprints contain relatively stable and discriminative interface features.

[0037] In summary, this embodiment verifies that the immunofingerprint-based discrimination method proposed in this invention can stably and effectively identify the TCR binding preference of pMHC complexes under limited structural sample conditions, providing a reliable foundation for subsequent interpretability analysis and cross-system generalization applications.

[0038] Example 3: Interpretability Analysis Based on Interface Patch Contribution This embodiment illustrates how, after predicting the TCR binding preference of the pMHC complex, the method of the present invention can further perform interpretability analysis on the prediction results, thereby identifying the interface structure regions that make key contributions to the discrimination results.

[0039] In this embodiment, a patch-level importance analysis workflow is constructed for the 40 HLA-A*02–peptide–TCR complex structure samples described in Example 2. Specifically, for each pMHC complex, based on the multiple prediction vectors obtained in Example 2, the interface patches sampled in the high-confidence predictions are normalized and statistically analyzed to obtain the patch contribution, generating a patch-level importance map.

[0040] The method for generating the importance map at the patch level is to normalize the patch contribution and map it back to the corresponding pMHC complex surface position. Specifically, the method is to take the center point of each interface patch as the mapping position, project the normalized patch contribution onto the pMHC complex surface, thereby forming the importance map at the patch level.

[0041] Furthermore, to transform the importance results at the patch level into a representation with greater biological and engineering interpretability, this embodiment also introduces a clustering strategy based on peptide chain residue positions to generate an importance map at the residue level. Specifically, based on the spatial distance relationship between the interface patch and each residue of the peptide chain, the interface patches are grouped and clustered, and the contributions of interface patches belonging to the same peptide chain residue neighborhood are statistically summarized and averaged, thereby assigning a corresponding importance value to each peptide chain residue to characterize the relative contribution of that residue in the TCR recognition process.

[0042] Figure 4 This illustrates an application example of the above-described interpretable analysis method in a specific TCR–pMHC complex. As an example, a TCR receptor type with high identification accuracy in Example 2 was selected for analysis. After statistical analysis of the importance of residues in multiple pMHC structural samples within this category, it was found that ( Figure 4In section a), several consecutive residues in the middle of the peptide chain showed high and stable importance values ​​in different structural samples, indicating that this region plays a key role in the discrimination process; while sequence analysis showed that the peptide sequence segment corresponding to this region is a conserved motif in different structural samples ( Figure 4 (b)

[0043] Further, a representative pMHC–TCR complex structure from the above categories was selected, and its patch-level importance map was visualized and analyzed. The results showed that highly important regions were mainly concentrated in the spatially protruding regions formed in the middle of the peptide chain, where a tight contact was formed with the TCR molecule in the three-dimensional structure. Figure 4 (c) Through analysis of the spatial configuration of this region ( Figure 4 (d) It can be observed that the local topological features formed by the residues in the middle of the peptide chain form a complementary contact relationship with specific residues in the TCR molecule, thus providing a clear structural basis for explaining the TCR-pMHC binding specificity at the molecular level.

[0044] In summary, this embodiment demonstrates that the method proposed in this invention can automatically identify interface regions that play a key role in the binding specificity of TCR–pMHC from the prediction results without relying on manual prior rules, thereby achieving a molecular mechanism-level explanation of TCR-pMHC binding specificity. This interpretable analysis method not only enhances the reliability of the prediction results but also provides an effective tool for subsequently revealing the structural basis of different immune recognition patterns.

[0045] Example 4: Application of Generalization Inference Across HLA Allele Systems This embodiment illustrates that the TCR-pMHC binding specificity prediction model based on immune fingerprinting proposed in this invention can still effectively infer TCR–pMHC binding preference and distinguish interface differences caused by allele polymorphism in different HLA allele systems that have not participated in training the geodesic convolution-based geometric deep learning network.

[0046] In this embodiment, a set of HLA-B*27 allele-related pMHC–TCR complexes were selected as input. The pMHC–TCR complexes include four HLA-B*27–peptide–TCR structural combinations, where specific alleles (HLA-B*27:05 and HLA-B*27:09) differ only at one MHC residue position: residue 116 of HLA-B*27:05 is ASP (D116), and residue 116 of HLA-B*27:09 is HIS (H116). This difference is located at the bottom of the MHC molecule's peptide binding groove, adjacent to the terminal residues of the peptide chain, and has a potential impact on peptide chain conformation and the interfacial physicochemical environment. Figure 5 (a) Figure 5 The peptides in the left and right images of 'a' are the autoantigen peptide (GQVMVVAPR, abbreviated as GQV) and the bacterial antigen peptide (LRVMLAPF, abbreviated as LRV), respectively.

[0047] Without retraining the model specifically, the above-mentioned HLA-B*27 pMHC complex structure is input into the model trained based on HLA-A*02 data, and the discrimination results corresponding to each input structure are obtained according to the method described in Example 1.

[0048] The inference results show that ( Figure 5 (b) For different peptide presentation scenarios under the same allele, the discrimination results show high consistency, indicating that the model can identify relatively stable and discriminative interface features under this allele system. However, when the input pMHC complex changes only at a single MHC residue site, the discrimination results show a significant shift, indicating a different interface feature category from the original pMHC complex.

[0049] To further analyze the sources of the discriminant differences mentioned above, this embodiment performed the same interface patch importance analysis procedure as in Example 3 on the relevant pMHC structures. The results showed that in each complex, the interface region (P9) located at the end of the peptide chain adjacent to the MHC polymorphic residues occupied a high weight in the discriminant results. Figure 5 (c)

[0050] Further analysis of interface features revealed that ( Figure 5 In the key regions mentioned above (d), the substitution of MHC residues corresponding to different alleles causes changes in local charge distribution and spatial matching relationships, thereby altering the interaction pattern between peptide terminal residues and MHC molecules. This change is ultimately reflected in the significant differences in the immunofingerprint in the model feature space, which are successfully captured by the model and reflected in the discrimination results.

[0051] Therefore, this embodiment demonstrates that the model proposed in this invention can not only accurately predict in known HLA allele systems, but also sensitively respond to changes in interface features caused by minor structural differences in different allele systems that were not trained, achieving cross-allele generalization inference. This capability enables the method of this invention to serve as a structural-level analysis tool for studying TCR recognition differences under different HLA alleles and different peptide combinations, exhibiting good scalability and application potential.

[0052] Example 5: A TCR-pMHC binding specificity prediction device based on immune fingerprinting This embodiment provides a TCR-pMHC binding specificity prediction device based on immune fingerprinting, which consists of the following modules: The first module is used to obtain the three-dimensional structural data of the target pMHC complex and construct the surface of the target pMHC complex. The three-dimensional structural data of the target pMHC complex can be obtained independently or locally extracted from the modeled structure of the corresponding experimental analysis or prediction software. The second module is used to triangulate the surface of the target pMHC complex to obtain a discretized pMHC surface grid. Based on the vertex of each pMHC surface grid, radial surface patches are extracted according to the geodesic radius. Among all surface patches, the surface patch whose center is located in the vicinity of the peptide is identified as an interface patch. The geodesic distance between any two pMHC surface grid vertices is approximately calculated using the Dijkstra algorithm. The third module is used to calculate the geometric and physicochemical features of each pMHC surface mesh vertex on the discretized pMHC surface mesh, construct an immune fingerprint based on the geometric and physicochemical features, and spatially locate the immune fingerprint within each interface patch using geodesic polar coordinates; the geometric features include shape index and distance-dependent curvature; the physicochemical features include Poisson-Boltzmann continuous electrostatics, hydrophobicity index, and the positions of free electrons and proton donors; The fourth module: Randomly and repeatedly sample a predetermined number of interface patches from the interface patches, and input the predetermined number of interface patches sampled each time into the TCR-pMHC binding specificity prediction model based on immune fingerprints. The TCR-pMHC binding specificity prediction model based on immune fingerprints encodes each sampled interface patch into a fingerprint matrix according to the immune fingerprint and its spatial location, and makes predictions based on the fingerprint matrix, thereby obtaining the binding preference prediction vector of the target pMHC complex for different TCR types; the predetermined number of interface patches sampled multiple times correspond to multiple prediction vectors. The fifth module is used to integrate multiple prediction vectors to output the final TCR–pMHC combined preference discrimination result; Module 6: Correlate multiple prediction vectors with the contribution of each interface patch to generate importance maps at the patch level and residue level, and perform interpretability analysis based on the importance maps at the patch level and residue level.

Claims

1. A method for predicting TCR-pMHC binding specificity based on immune fingerprinting, characterized in that, Includes the following steps: (i) Obtain the three-dimensional structural data of the target pMHC complex and construct the surface of the target pMHC complex; (ii) Triangulate the surface of the target pMHC complex to obtain a discretized pMHC surface grid. Extract radial surface patches based on the geodesic radius of each pMHC surface grid vertex, and identify the surface patches whose center is located in the peptide neighborhood as interface patches. (iii) Calculate the geometric and physicochemical features of each pMHC surface grid vertex on the discretized pMHC surface grid, construct an immune fingerprint based on the geometric and physicochemical features, and spatially locate the immune fingerprint in geodesic polar coordinates within each interface patch; the geometric features include shape index and distance-dependent curvature; the physicochemical features include Poisson-Boltzmann continuous electrostatics, hydrophobicity index, and the positions of free electrons and proton donors; (iv) Randomly sample a predetermined number of interface patches from the interface patches and input them into the TCR-pMHC binding specificity prediction model based on immune fingerprints. The TCR-pMHC binding specificity prediction model based on immune fingerprints encodes the sampled interface patches into a fingerprint matrix according to the immune fingerprints and their spatial positioning and makes predictions based on the fingerprint matrix, thereby obtaining the target pMHC complex binding preference prediction vector for different TCR types. (v) Repeat step (iv) to obtain multiple prediction vectors, and integrate the multiple prediction vectors to output the final TCR–pMHC combined preference discrimination result; (vi) Associate multiple prediction vectors with the contribution of each interface patch to generate an importance map at the patch level and an importance map at the residue level, and perform interpretability analysis based on the importance maps at the patch level and the importance maps at the residue level.

2. The method according to claim 1, characterized in that, Step (ii) is as follows: First, the surface of the target pMHC complex is triangulated using a preset resolution to obtain a discretized pMHC surface grid; then, radial surface patches are extracted from the vertices of each pMHC surface grid according to the set geodesic radius, and radial surface patches whose center distance from any atom on the peptide is less than a predetermined threshold are identified as interface patches.

3. The method according to claim 1, characterized in that, In step (iii), the shape index is calculated using Gaussian curvature and average curvature, the distance-related curvature is calculated using the distance between the mesh vertex and the center of the interface patch and the normal angle, the Poisson-Boltzmann continuous electrostatics are calculated using PDB2PQR and APBS, the hydrophobicity index is calculated using the Kyte-Doolittle scale, and the positions of the free electrons and proton donors are calculated using the position of the mesh vertex relative to its nearest neighbor polar atom.

4. The method according to claim 1, characterized in that, In step (iii), the geometric and physicochemical features are spatially located within the interface patch using a geodesic polar coordinate system. Specifically, the radial coordinate is the geodesic distance between the grid vertex and the center of the interface patch, and the angular coordinate is the angle between the line connecting the grid vertex and the center of the interface patch and a randomly specified direction.

5. The method according to claim 1, characterized in that, In step (iv), the training process of the TCR-pMHC binding-specific prediction model based on immune fingerprinting includes the following steps: (1) Obtain TCR-pMHC complexes containing different TCR types, extract TCR types as binding-specific tags for pMHC complexes, extract three-dimensional structural data of pMHC complexes and construct the surface of pMHC complexes. (2) Triangulate the surface of the pMHC complex to obtain a discretized pMHC surface grid. Extract radial surface patches based on geodesic radius of each pMHC surface grid vertex and identify the surface patches whose center is located in the peptide neighborhood as interface patches. (3) Calculate geometric and physicochemical features on the discretized pMHC surface mesh, construct an immune fingerprint based on the geometric and physicochemical features, and spatially locate the immune fingerprint in each interface patch using geodesic polar coordinates; the geometric features include shape index and distance-related curvature; the physicochemical features include Poisson-Boltzmann continuous electrostatics, hydrophobicity index, and the positions of free electrons and proton donors; (4) Randomly sample a predetermined number of interface patches from the interface patches and input them into a geodesic convolution-based geometric deep learning network. The geodesic convolution-based geometric deep learning network encodes the sampled interface patches into a fingerprint matrix based on the immune fingerprint and its spatial localization, and generates a pMHC binding preference prediction vector for different TCR types based on the fingerprint matrix. (5) Repeat step (4) to obtain multiple pMHC binding preference prediction vectors for different TCR types; use the TCR type labels corresponding to the pMHC complex to perform supervised learning on the geodesic convolution-based geometric deep learning network to obtain the TCR-pMHC binding specificity prediction model based on immune fingerprint.

6. The method according to claim 1, characterized in that, Step (v) involves integrating multiple prediction vectors, specifically by using vector averaging or majority voting.

7. The method according to claim 1, characterized in that, In step (vi), the method of associating multiple prediction vectors with the contribution of each interface patch to generate an importance map at the patch level and an importance map at the residue level is as follows: the patch contribution is obtained by normalizing the interface patches sampled in the high confidence prediction, and an importance map at the patch level is generated; and the patch contribution of adjacent specific peptide residues is further regionally aggregated to generate an importance map at the residue level.

8. A TCR-pMHC binding specificity prediction device based on immune fingerprinting, characterized in that, include: The first module is used to obtain the three-dimensional structural data of the target pMHC complex and construct the surface of the target pMHC complex. The three-dimensional structural data of the target pMHC complex can be obtained independently or locally extracted from the modeled structure of the corresponding experimental analysis or prediction software. The second module is used to triangulate the surface of the target pMHC complex to obtain a discretized pMHC surface grid. Based on the vertex of each pMHC surface grid, radial surface patches are extracted according to the geodesic radius. Among all surface patches, the surface patch whose center is located in the vicinity of the peptide is identified as an interface patch. The geodesic distance between any two pMHC surface grid vertices is approximately calculated using the Dijkstra algorithm. The third module is used to calculate the geometric and physicochemical features of each pMHC surface mesh vertex on the discretized pMHC surface mesh, construct an immune fingerprint based on the geometric and physicochemical features, and spatially locate the immune fingerprint within each interface patch using geodesic polar coordinates; the geometric features include shape index and distance-dependent curvature; the physicochemical features include Poisson-Boltzmann continuous electrostatics, hydrophobicity index, and the positions of free electrons and proton donors; The fourth module: Randomly and repeatedly sample a predetermined number of interface patches from the interface patches, and input the predetermined number of interface patches sampled each time into the TCR-pMHC binding specificity prediction model based on immune fingerprints. The TCR-pMHC binding specificity prediction model based on immune fingerprints encodes each sampled interface patch into a fingerprint matrix according to the immune fingerprint and its spatial location, and makes predictions based on the fingerprint matrix, thereby obtaining the binding preference prediction vector of the target pMHC complex for different TCR types; the predetermined number of interface patches sampled multiple times correspond to multiple prediction vectors. The fifth module is used to integrate multiple prediction vectors to output the final TCR–pMHC combined preference discrimination result; Module 6: Correlate multiple prediction vectors with the contribution of each interface patch to generate importance maps at the patch level and residue level, and perform interpretability analysis based on the importance maps at the patch level and residue level.

Images