A TCR-pMHC binding affinity prediction method, system, electronic device and medium based on a graph neural network

Abstract

The application discloses a TCR-pMHC binding affinity prediction method and system based on a graph neural network, an electronic device and a medium, relates to the technical field of computational structural biology, and comprises the following steps: constructing a TCR-pMHC complex structure dataset, performing graph modeling by taking amino acid residues as nodes and spatial neighbor residues as edges, updating zero initialization node features through network iteration, combining spatial distance, relative position and chain identification information to construct edge features, fusing features by using an encoder-decoder graph neural network, outputting interface interaction features and residue amino acid probability distribution, taking TCR sequences with gradient affinity in the same experimental system as a reference benchmark, scoring each TCR sequence in the system, and verifying the linear correlation between the model prediction score and the real affinity negative logarithm value. According to the application, the relative evaluation of the TCR-pMHC binding affinity and the sorting and screening of mutants can be realized by using the model prediction score, and the TCR-pMHC binding affinity is high in precision and generalization.

Description

Technical Field

[0001] This invention relates to the field of computational structural biology, specifically to a method, system, electronic device, and medium for predicting TCR-pMHC binding affinity based on graph neural networks. Background Technology

[0002] T-cell receptor (TCR)-related targeted technologies are one of the core research directions in current tumor immunotherapy. TCR-T cell therapy and TCR bispecific antibodies, among other drug formulations, have significant advantages and broad clinical application prospects in the field of precision targeted therapy for solid tumors. However, the binding affinity of the natural TCR to the tumor antigen peptide-major histocompatibility complex (pMHC) is generally low. In complex scenarios such as low tumor antigen expression and recognition of autologous antigen epitopes, it is difficult to efficiently complete antigen recognition and tumor cell killing, greatly limiting the therapeutic efficacy of TCR-based drugs. Therefore, while fully preserving antigen recognition specificity, directionally enhancing the binding affinity of TCR to pMHC is a core technological breakthrough for improving the efficacy of TCR drugs and expanding their clinical applicability.

[0003] The binding affinity between TCRs and pMHC is a core parameter determining the intensity of T cell immune activation and drug targeting efficacy, and it is also a key indicator that must be precisely controlled during TCR molecular modification and drug screening. Currently, the industry mainly relies on traditional wet experimental screening techniques such as yeast display and phage display to improve TCR affinity. These techniques suffer from cumbersome experimental procedures, long screening cycles, and high reagent and labor costs. Furthermore, the experimental results are highly susceptible to changes in operating methods, experimental environment, and batch variations, resulting in poor stability and reproducibility.

[0004] To overcome the limitations of traditional wet experiments, current research has begun to introduce artificial intelligence (AI) computational methods to assist in TCR affinity prediction and mutation design. However, specialized computational tools for TCR-pMHC specific interaction systems are currently scarce. Existing AI prediction schemes generally suffer from technical shortcomings such as a lack of real affinity annotation data, insufficient accuracy in protein complex structure modeling, and weak generalization ability across experimental systems. These limitations prevent them from achieving high-precision quantitative prediction and support efficient and universal TCR site-directed mutagenesis and library screening.

[0005] Artificial intelligence computational modeling possesses core advantages such as high throughput, low cost, and high efficiency. It can complete the pre-screening and preliminary evaluation of massive numbers of candidate molecules before traditional wet experimental screening, significantly reducing the cycle of molecular modification and experimental verification. This addresses industry problems such as insufficient prediction accuracy, poor generalization of mutation optimization, heavy reliance on large-scale wet experiments, and low R&D efficiency inherent in existing technologies. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a method and system, electronic device, storage medium, and high-affinity single-stranded TCR product based on graph neural networks for predicting and optimizing TCR-pMHC binding affinity. This invention utilizes a customized graph neural network architecture, combined with structural biology simulations and differentiated mutation strategies, to achieve high-precision relative assessment of TCR binding affinity and efficient mutant screening, solving the problems of low efficiency, high cost, poor accuracy, and weak generalization of traditional screening techniques.

[0007] To achieve the above objectives, the technical solution provided by the present invention is as follows: The first aspect of this application provides a method for predicting TCR-pMHC binding affinity based on graph neural networks, comprising the following steps: After deduplication and cleaning of multi-source TCR-pMHC amino acid pairing data, a TCR-pMHC complex structure dataset was constructed through three-dimensional structure deduction and prediction and molecular dynamics conformation relaxation optimization. Graph structure modeling was performed on the three-dimensional structure of the TCR-pMHC complex. The three-dimensional coordinates of the protein backbone and side chain atoms were extracted. Amino acid residues were used as nodes and the spatial nearest neighbor relationship of residues was used as edges. Node features were obtained by zero initialization of nodes and subsequent network iteration updates. Corresponding edge features were constructed by combining residue spatial distance, relative position and chain identification information. The constructed node features and edge features are input into a graph neural network. Based on the encoder-decoder architecture, the features are iteratively fused to output the interface structural interaction features between TCR and pMHC molecules and the amino acid probability distribution at each residue position. The interface structural interaction features are used to characterize the binding mode and binding capacity of TCR and pMHC. TCR sequences with gradient true affinity in the same experimental system were selected as reference benchmarks. Based on the interaction characteristics, each TCR sequence in the experimental system was scored to verify the linear correlation between the model prediction score and the negative logarithm of the true affinity. Based on the model prediction score, the relative evaluation of TCR-pMHC binding affinity and the ranking and screening of mutants were realized.

[0008] Furthermore, the construction of the TCR-pMHC complex structure dataset includes: The amino acid sequences of the TCR α chain, β chain, and corresponding pMHC were obtained. Multi-source TCR-pMHC pairing data were integrated, and deduplication and cleaning were performed using the CDR3 sequence of the TCR α chain, the CDR3 sequence of the β chain, the MHC sequence, and the peptide sequence as co-bonds. The initial three-dimensional structure of the TCR-pMHC complex was predicted based on the amino acid sequence deduction. Then, conformational relaxation optimization was performed through molecular dynamics simulation to obtain a stable conformation, thus constructing a TCR-pMHC complex structure dataset.

[0009] Furthermore, when constructing the graph structure and features, corresponding residues are selected to establish edge connections based on the preset number of nearest neighbors, and node features and edge features are uniformly configured according to the preset feature dimensions; the encoding layer realizes bidirectional updates of node and edge features, and the decoding layer fixes the edge features and iteratively optimizes the node features to complete feature fusion; Furthermore, the graph neural network adopts an encoder-decoder architecture, where the encoding layer updates node and edge features bidirectionally, the decoding layer fixes edge features and iteratively optimizes node features, and outputs a structure context-aware feature representation through multi-layer message passing and feature aggregation.

[0010] A second aspect of this application provides a method for optimizing the affinity of TCR molecules, comprising the following steps: Based on amino acid sequence deduction, the initial three-dimensional structure of the complex of TCR and pMHC to be optimized was predicted, and molecular dynamics conformation relaxation optimization was performed. Residues whose shortest spatial distance between the CDR region and the pMHC heavy atoms is no greater than a preset spatial distance threshold were identified as interface contact residues. Combined with the residue decomposition results of MM-GBSA binding free energy, hotspot residues with large absolute values ​​of free energy contribution were screened as mutation sites. Mutant sequences are generated based on the difference in the number of mutation sites: when the number of sites is less than a preset threshold, saturation mutation is used to exhaustively search for all amino acid substitution combinations; when the number of sites is greater than or equal to the preset threshold, the amino acid probability distribution based on the graph neural network is used to generate mutant sequences by adjusting the temperature parameter T, which can adaptively adjust the concentration and smoothness of the amino acid sampling probability distribution. The above-mentioned graph neural network-based TCR-pMHC affinity prediction method was used to score all mutant TCR sequences, and mutants with model prediction scores better than wild-type TCR were selected as candidate mutants. Candidate mutant single-stranded TCR genes were synthesized and expression plasmids were constructed. The target protein was obtained through host expression, renaturation, and purification. Binding affinity was measured by SPR or BLI technology, and high-affinity TCRs with significantly enhanced affinity were screened.

[0011] To optimize the above technical solution, the specific measures also include: The random sampling described herein involves adaptively adjusting the sampling preference of the amino acid probability distribution using a temperature parameter T, and constructing a controllable mutant library based on the structural stability and sequence diversity of the mutant sequences. For the amino acid scores output by the graph neural network decoder at the target residue position, the scores are first scaled using the temperature parameter T, and then the corresponding amino acid sampling probabilities are obtained through the Softmax function. The calculation formula is as follows:

[0012] in, is the amino acid sampling probability, is the model prediction score of the target amino acid at the current residue site, are the model prediction scores of all candidate amino acids at the current residue site, j is the index traversing all candidate amino acids; the temperature parameter T>0, with T=1.0 as the boundary, low-temperature sampling is performed when 0<T<1.0, sampling is performed according to the original probability distribution output by the decoder when T=1.0, and high-temperature sampling is performed when T>1.0. The sampling rule is: as the temperature parameter T increases, the probability distribution of amino acid sampling gradually becomes gentler, and the sampling probability of low-probability amino acids increases.

[0013] The high-affinity single-chain TCR is prepared by the above-mentioned TCR molecule affinity optimization method, the CDR region of the single-chain TCR contains mutation sites, the binding affinity is significantly improved compared with the wild-type TCR, and the amino acid sequence of the single-chain TCR is SEQ ID NO: 2 or SEQ ID NO: 3.

[0014] A third aspect of the present application provides a graph neural network-based TCR-pMHC binding affinity prediction and optimization system, which is configured to implement the above-mentioned graph neural network-based TCR-pMHC binding affinity prediction method or TCR molecule affinity optimization method, comprising: a data acquisition module, configured to acquire multi-source TCR α chain, β chain, pMHC amino acid pairing data and experimental affinity data, and perform deduplication and cleaning processing on the multi-source original pairing data; a structure construction module, configured to construct a TCR-pMHC complex structure dataset through three-dimensional structure deduction prediction and molecular dynamics conformational relaxation optimization; a graph feature extraction module, configured to abstract the optimized three-dimensional structure of the TCR-pMHC complex into a graph structure, taking amino acid residues as nodes and the spatial proximity relationship of residues as edges, adopting node zero initialization and obtaining node features through subsequent network iteration update, and combining the spatial distance, relative position and chain identification information of residues to construct node features and corresponding edge features of the graph neural network; a feature fusion and scoring module, configured to input the constructed node features and edge features into the graph neural network, complete iterative feature fusion based on an encoder-decoder architecture, and output the interface structure interaction features between TCR and pMHC molecules and the amino acid probability distribution at each residue position; scoring the TCR sequence based on the interaction features to obtain a model prediction score; a mutation design module, configured to screen hotspot mutation sites in the CDR region of TCR based on the relaxed and optimized three-dimensional structure of the complex, and generate mutant TCR sequences in a high-throughput manner according to the difference in the number of sites; The candidate screening module is used to score the affinity of all mutant TCR sequences, compare the model prediction scores of mutant and wild-type TCRs, and screen candidate mutants with better model prediction scores. The validation output module is used to obtain experimental affinity validation data for candidate mutants and output candidate mutant sequences, prediction results, and validation results.

[0015] Furthermore, the data acquisition module uses the CDR3 sequence of the TCR α chain, the CDR3 sequence of the β chain, the MHC sequence, and the peptide sequence as joint bonds to complete the deduplication and cleaning of multi-source data; the graph feature extraction module extracts the three-dimensional coordinates of key atoms in the protein backbone and side chains, selects corresponding residues to establish edge connections according to a preset number of nearest neighbors, and uniformly configures the feature dimensions according to a preset feature dimension; the feature fusion and scoring module verifies the linear correlation between the model prediction score and the negative logarithm of the true affinity -logKd; the verification output module is specifically used to summarize the candidate mutant sequences, affinity model prediction scores and ranking results, collect and organize SPR or BLI measured affinity data from in vitro experimental systems, compare the binding performance with wild-type TCRs to complete the validity verification of candidate mutants, and output the result data of high-affinity TCR candidate molecules.

[0016] A fourth aspect of this application provides an electronic device including a processor and a memory, the memory storing an executable computer program that, when executed by the processor, implements the above-described graph neural network-based TCR-pMHC binding affinity prediction method or TCR molecule affinity optimization method.

[0017] The fifth aspect of this application provides a computer-readable storage medium, characterized in that it stores a computer program thereon, which, when executed by a processor of an electronic device, implements the above-described TCR-pMHC binding affinity prediction method or TCR molecule affinity optimization method based on graph neural networks.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a structural modeling system that closely matches real-world binding scenarios by acquiring high-quality TCR-pMHC experimental datasets from multiple sources, predicting complex structures with high precision, and optimizing molecular dynamics conformational relaxation. Simultaneously, based on residue-level graph topology modeling and a multi-dimensional feature encoding strategy, combined with an encoder-decoder message-passing graph neural network, it achieves deep fusion learning of structural context features, overcoming the shortcomings of traditional methods that heavily rely on manual prior features and suffer from low structural modeling accuracy. The model employs a self-supervised pre-training method without affinity labels. The linear correlation between the model score and the negative logarithmic value of experimental affinity is verified through gradient affinity samples within the same system. This demonstrates that the model score accurately reflects the relative affinity variation trend of TCR mutants within the same experimental system and can be stably used for ranking mutant affinity and high-throughput pre-screening, achieving high-precision and generalizable relative evaluation of TCR-pMHC binding affinity. This invention integrates structural interface contact analysis with the MM-GBSA free energy decomposition mechanism to screen for CDR region binding hotspot mutation sites. It also employs a differentiated mutation generation strategy that adaptively utilizes saturation mutations and temperature-controlled random sampling based on the number of mutation sites. The concentration and smoothness of the amino acid sampling probability distribution are adaptively controlled by the temperature parameter T, balancing mutation library diversity and screening efficiency while ensuring TCR structural stability. This effectively avoids the problems of strong blindness, combinatorial explosion, and high screening costs associated with traditional mutation screening. Wet experiments have verified that the single-chain TCR mutants optimized by this invention exhibit significantly improved affinity compared to wild-type mutants, while retaining the original antigen recognition specificity and protein structural stability. This invention allows for the pre-screening of a massive number of candidate TCR sequences before wet experiments, significantly reducing the experimental cost of TCR molecule modification, shortening the development cycle, and improving optimization accuracy and reproducibility. It effectively solves the problems of low screening efficiency, large batch-to-batch variations, and limited modification effects in traditional TCR affinity modification, and can efficiently support the targeted development of high-affinity TCR candidate molecules. It can be applied to fields such as TCR-T cell therapy and the development of tumor immunotherapy drugs such as TCR bispecific antibodies. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall technical solution of the present invention.

[0020] Figure 2 This is a schematic diagram of the complete process of the TCR-pMHC binding affinity prediction method of the present invention.

[0021] Figure 3 This is a schematic diagram of the complete process of the TCR molecular affinity optimization method of the present invention.

[0022] Figure 4 This is a schematic diagram of the module structure of the TCR-pMHC binding affinity prediction and optimization system of the present invention. Detailed Implementation

[0023] The present invention will be further described in detail below through specific embodiments, but it should not be construed as limiting the scope of the subject matter of the present invention to the following embodiments. All technologies implemented based on the above content of the present invention fall within the scope of the present invention.

[0024] Unless otherwise specified, the experimental methods used in the following examples are conventional methods, and the reagents, methods and equipment used are conventional reagents, methods and equipment in this technical field.

[0025] This invention provides a method for predicting TCR-pMHC binding affinity based on graph neural networks. Figure 1 This is a schematic diagram of the overall process of the method of the present invention, and the specific steps are as follows: First, raw TCR-pMHC amino acid pairing data from multiple sources were collected, and the raw data underwent deduplication and standardization cleaning. Based on high-quality paired amino acid sequences, the initial three-dimensional structure of the TCR-pMHC complex was predicted through three-dimensional structure deduction. Further conformational relaxation optimization was performed using molecular dynamics simulations to eliminate unreasonable conformations and atomic collisions, obtaining a stable equilibrium complex conformation. Finally, a high-quality TCR-pMHC complex structure dataset suitable for model training was constructed. Specifically, data deduplication and cleaning used the TCR α-chain CDR3 sequence, β-chain CDR3 sequence, MHC sequence, and antigenic peptide sequence as unique joints to remove redundant, incomplete, and invalid data.

[0026] Secondly, graph structure modeling and feature construction were performed on the optimized TCR-pMHC complex's three-dimensional structure. Specifically, the three-dimensional spatial coordinates of the protein backbone Cα, C, N, and O atoms, as well as key side chain atoms, were extracted. Each amino acid residue in the complex was used as a node in a graph neural network, and edge connections were constructed based on the spatial proximity relationships between residues. Nearest neighbor residues were selected according to a preset number of nearest neighbors to establish edge connections, and the feature dimensions of nodes and edges were uniformly configured according to a preset feature dimension. This invention employs a zero-initialization strategy for node features, meaning that node features do not depend on prior residue attributes and can be automatically updated and optimized through subsequent network iterations. Simultaneously, spatial distance encoding between residues, relative sequence position encoding, and protein chain identification information are integrated to construct edge features with spatial invariance and structure specificity.

[0027] Subsequently, the standardized node and edge features are input into a message-passing graph neural network with an encoder-decoder architecture for iterative feature fusion. The network encoding layer is responsible for bidirectionally updating node and edge features, enabling bidirectional interactive learning of node and edge information. The decoding layer fixes the edge feature parameters and iteratively optimizes only the node features. Through multi-layer message passing and global feature aggregation, it outputs the TCR-pMHC complex interface structure interaction features with global structural context awareness, while simultaneously predicting the probability distribution of 20 standard amino acids corresponding to each residue position. The complex interaction features characterize the spatial binding mode, interface interaction strength, and overall binding capacity of TCR and pMHC.

[0028] Finally, TCR-pMHC sample data with continuous gradient true affinity were selected from the same experimental system, and TCR sequences with low, medium and high affinity gradients were used as a unified reference. The trained graph neural network was used to uniformly score all TCR sequences in the experimental system. The correlation between the predicted score output by the model and the negative logarithm (-logKd) of the experimentally measured affinity was analyzed to verify the linear correlation between the two. It was confirmed that the model score can accurately distinguish the affinity differences of different TCR mutants in the same system and can be used for relative ranking of mutant affinity and high-throughput screening.

[0029] The same experimental system in this invention refers to a set of TCR mutants with the same parental origin and their corresponding experimental affinity data, obtained by performing single-point, double-point, or multi-point amino acid mutations in the CDR region of the same pMHC target site and the same wild-type TCR.

[0030] An experimental system may include dozens of wild-type or mutant TCR sequences. Because the CDR regions of each TCR sequence have different mutation sites or combinations, their binding affinity to the same pMHC target exhibits a gradient difference, which can reach hundreds to tens of thousands of times. This invention compares the model's affinity predictions for each TCR mutant with the corresponding experimental results to evaluate the model's predictive ability for differences in TCR-pMHC binding affinity.

[0031] This invention further provides a method for optimizing the affinity of TCR molecules. Based on the aforementioned high-prediction model, it achieves targeted mutational modification and affinity enhancement of TCRs, taking into account both the diversity and structural stability of the mutant library, and efficiently screening high-affinity mutants. Specifically, it includes the following steps: The first step is complex structure modeling and optimization. Based on the amino acid sequences of the wild-type TCR and its corresponding pMHC, the initial three-dimensional structure of the TCR-pMHC complex is obtained through three-dimensional structure deduction and prediction. Then, conformational relaxation optimization is completed through molecular dynamics simulation to obtain a thermodynamically stable, spatially conflict-free equilibrium complex conformation, providing a precise structural basis for subsequent mutation site screening.

[0032] The second step involves precise screening of hotspot mutation sites in the CDR region. Combining three-dimensional structural visualization analysis, residue interface contact analysis, and MM-GBSA binding free energy residue decomposition technology, mutation site screening is completed: residues in the TCR CDR region whose shortest spatial distance to pMHC heavy atoms is no greater than a preset spatial distance threshold are identified as directly contacting residues. Simultaneously, hotspot residues with a large absolute value contributing to the binding free energy are screened. By combining these two types of indicators, the residues that contribute the most to the TCR-pMHC binding interaction are determined as core mutation sites, ensuring the effectiveness and targeting of the mutation modification.

[0033] The third step is the generation of differentiated high-throughput mutant sequences. An adaptive mutation strategy is adopted based on the number of mutation sites selected: when the number of mutation sites is less than a preset threshold, a saturation mutation strategy is used to exhaustively explore all 20 amino acid combinations at each site, maximizing the optimization potential of single-point mutations; when the number of mutation sites is greater than or equal to the preset threshold, the combinatorial explosion problem is avoided. Based on the amino acid probability distribution at each residue position output by the graph neural network, a random sampling strategy is used to regulate the temperature parameter T to generate mutant sequences. For the amino acid scores output by the graph neural network decoder at the target residue position, the scores are first scaled using the temperature parameter T, and then the corresponding amino acid sampling probability is obtained through the Softmax function. The calculation formula is as follows:

[0034] in, This represents the amino acid sampling probability. The model prediction score for the target amino acid at the current residue site. The model prediction score for all candidate amino acids at the current residue site. jis an index for traversing all candidate amino acids; the temperature parameter T>0 is a dimensionless regulation parameter, which can adaptively regulate the concentration and flatness of the amino acid sampling probability distribution; taking T=1.0 as the boundary, low-temperature sampling is performed when 0<T<1.0, sampling is performed according to the original probability distribution output by the decoder when T=1.0, and high-temperature sampling is performed when T>1.0. The sampling rule is: as the temperature parameter T increases, the probability distribution of amino acid sampling gradually becomes flatter, and the sampling probability of low-probability amino acids increases. During low-temperature sampling, the probability distribution is concentrated, and highly adapted conservative amino acid substitutions are preferentially selected to ensure the structural stability of the TCR; during high-temperature sampling, the probability distribution tends to be flat, which increases the sampling probability of low-probability amino acids and effectively expands the diversity of mutant sequences. By adjusting the value of T, the balance between the structural compatibility of the mutation library and the sequence diversity is achieved.

[0035] Step 4: Intelligent screening of mutants. Using the aforementioned TCR-pMHC binding affinity prediction method of the present invention, batch affinity scoring is performed on all mutant TCR sequences, the model prediction score of the wild-type TCR is taken as the baseline, and mutants whose model prediction scores are significantly better than that of the wild type are screened as candidate optimized mutants to complete high-throughput pre-screening before wet experiments.

[0036] Step 5: Wet experiment verification and acquisition of high-affinity TCR. The screened candidate mutants are subjected to single-chain TCR gene synthesis and expression plasmid construction, and high-purity single-chain TCR target proteins are obtained through host expression, protein renaturation and purification processes, and the corresponding pMHC proteins are prepared synchronously; surface plasmon resonance (SPR) or bio-layer interferometry (BLI) technology is used to accurately measure the actual binding affinity of each mutant TCR to pMHC, and compared with the affinity data of wild-type TCR, finally, high-activity TCR mutant sequences with significantly improved affinity are obtained by screening.

[0037] The present invention also provides a high-affinity single-chain TCR targeting a specific antigen, which is prepared by the aforementioned TCR molecule affinity optimization method. The CDR key binding region of the single-chain TCR carries artificial mutation sites, and on the premise of completely retaining the original antigen recognition specificity, compared with the wild-type TCR, its binding affinity to the corresponding specific pMHC complex is significantly improved.

[0038] Specifically, the amino acid sequence of the high-affinity single-chain TCR optimized by the present invention is selected from any one of SEQ ID NO: 2 and SEQ ID NO: 3 in the sequence listing. Verified by SPR experiments, the affinity of the two optimized sequences is more than 40 times higher than that of the wild-type TCR (SEQ ID NO: 1), with stable structure and strong specificity, and can be applied to the research and development of tumor immunotherapy products such as TCR-T cell therapy and TCR bispecific drugs.

[0039] This invention also provides a TCR-pMHC binding affinity prediction and optimization system based on graph neural networks, used to implement the aforementioned TCR-pMHC binding affinity prediction method and TCR molecule affinity optimization method. The system has a modular design and can be deployed on local devices, servers, or cloud computing platforms, specifically including the following functional modules: Data acquisition module: Used to acquire amino acid pairing data of multi-source TCR α chain, β chain, and pMHC, as well as raw experimental affinity data, and to integrate, deduplicate, and standardize the multi-source raw pairing data. Specifically, it uses the TCR α chain CDR3 sequence, β chain CDR3 sequence, MHC sequence, and antigen peptide sequence as joint unique bonds to complete data deduplication, remove incomplete, redundant, and invalid data, and output the paired dataset.

[0040] The structure building module receives cleaned, high-quality amino acid pairing data, generates the initial three-dimensional structure of the TCR-pMHC complex using a three-dimensional structure deduction and prediction tool, performs conformational relaxation optimization through molecular dynamics simulation to eliminate atomic collisions and unreasonable conformations, outputs a stable equilibrium three-dimensional structure of the complex, and constructs a standardized complex structure dataset.

[0041] The graph feature extraction module is used to abstract the optimized three-dimensional structure of the TCR-pMHC complex into a graph structure model at the residue level, with amino acid residues as network nodes and spatial nearest neighbors as edge connections. It adopts a zero-initialization strategy for node features, and the node features are obtained autonomously through subsequent network iterations. At the same time, it extracts the three-dimensional coordinates of key atoms in the protein backbone and side chains, establishes edge connections based on the preset number of nearest neighbors, uniformly configures the feature size according to the preset feature dimensions, and integrates residue spatial distance encoding, relative position encoding, and chain identification information to construct edge features.

[0042] Feature fusion and scoring module: It is used to input standardized node features and edge features into the encoder-decoder architecture graph neural network, complete the iterative fusion of features through multi-layer message passing, and output the TCR-pMHC interface interaction features and the probability distribution of amino acids of each residue; Based on the gradient affinity reference sample of the same experimental system, it verifies the linear correlation between the model's predicted score and the negative logarithmic value of the actual affinity.

[0043] Mutation design module: Based on the three-dimensional structure of the stable complex after relaxation optimization, it is used to accurately screen hotspot mutation sites in the TCRCDR region; and selects a sampling strategy according to the number of mutation sites to generate mutant TCR sequence libraries in a high-throughput and controllable manner.

[0044] Candidate screening module: Used to call the feature fusion and scoring module in batches to complete the affinity prediction and scoring of all mutant TCR sequences, compare the model prediction scores of mutant and wild-type TCR, and screen mutants with better predicted affinity as candidate optimization.

[0045] The validation output module is used to summarize candidate mutant sequences, affinity model prediction scores and ranking results, collect and organize SPR or BLI measured affinity data from in vitro experimental systems, compare the binding performance with wild-type TCRs, complete the validity verification of candidate mutants, and output the result data of high affinity TCR candidate molecules.

[0046] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores an executable computer program, and when the computer program is executed by the processor, it is able to implement the TCR-pMHC binding affinity prediction method described above.

[0047] The present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor of an electronic device, it can implement the TCR-pMHC binding affinity prediction method described above.

[0048] The technical solution of the present invention will be further described in detail below with reference to specific embodiments: Example 1: Construction of TCR-pMHC complex structure dataset and feature extraction using graph neural network This embodiment illustrates the process of constructing a high-quality complex structure dataset in the TCR-pMHC binding affinity prediction method, as well as the extraction of node and edge features of graph neural networks and the graph structure modeling method.

[0049] Construction of the TCR-pMHC complex structure dataset: This invention employs a multi-source public database data integration approach to construct the training dataset, ensuring data volume, authenticity, and diversity. Specific data sources in this embodiment include the 10xgenomics single-cell TCR sequencing dataset and the authoritative VDJdb public database. The 10xgenomics dataset uses high-quality TCR-pMHC pairing data processed, verified, and validated using the DeepAIR algorithm, and its reliability has been verified through multiple rounds of experiments. The VDJdb database only selects and retains valid TCR-pMHC pairing entries that have been verified in vitro and contain complete TCR α-chain and β-chain amino acid sequences, discarding incomplete or invalid data.

[0050] After merging and summarizing the above two types of multi-source raw data, redundancy removal and standardization cleaning are carried out. This invention adopts a multi-dimensional joint deduplication mechanism, using the TCR α chain CDR3 sequence, β chain CDR3 sequence, MHC sequence and antigenic peptide sequence as the joint unique deduplication bond, which can distinguish different pairing combinations, completely eliminate duplicate samples, and finally obtain a redundancy-free, high-fidelity TCR-pMHC amino acid pairing dataset.

[0051] For each valid TCR-pMHC paired sample in the dataset, three-dimensional complex structure prediction was performed. This embodiment uses the TCRdock structure prediction tool, which is based on a specially fine-tuned AlphaFold2 architecture. The TCR α chain amino acid sequence, TCR β chain amino acid sequence, MHC-I heavy chain sequence, and corresponding antigen peptide sequence are used as joint input parameters. The TCRdock prediction process is run in batches to quickly output the initial three-dimensional spatial structure of all TCR-pMHC complexes.

[0052] To address issues such as atomic collisions, unreasonable local conformations, and low confidence in the CDR binding region structure in the initial predicted structures, molecular dynamics simulation conformation relaxation optimization was uniformly performed on all initial 3D structures. The specific simulation system setup was as follows: the complete TCR-pMHC complex was placed in a water box constructed using the TIP3P water molecule model, with the minimum distance between the protein molecule surface and the boundary of the periodic water box limited to 8 Å to ensure sufficient space for molecular motion; 0.15M NaCl salt ions were added to the system to simulate the ion strength of the human physiological environment, closely resembling the actual in vivo binding scenario; by redistributing the hydrogen atom mass, the molecular dynamics simulation time step was effectively increased, improving simulation efficiency without affecting the structure optimization accuracy; throughout the process, the RSFF2C force field was used to perform energy minimization calculations on the system, gradually relaxing unreasonable conformations and eliminating atomic collisions. Finally, the structurally stable and energy-optimal final frame conformation was extracted from the dynamic simulation trajectory as the equilibrium complex structure for subsequent model training, feature extraction, and analysis, thus completing the construction of a high-quality TCR-pMHC complex structure dataset.

[0053] Extraction of node and edge features and construction of graph structure in graph neural networks: This invention abstracts the optimized three-dimensional spatial structure of the TCR-pMHC complex into a standardized topological graph structure G=(V,E), adapting it to the input requirements of a graph neural network. In this graph, each node vi uniquely corresponds to an amino acid residue in the complex, completely covering all amino acid residues of TCR and pMHC; the edges eij in the graph are used to characterize the spatial nearest neighbor interactions between different residues i and j, achieving two-dimensional topological modeling of the three-dimensional structure.

[0054] In terms of edge connection construction and edge feature extraction, this embodiment uses the three-dimensional spatial coordinates of the Cα atom of each amino acid residue backbone as a reference to calculate the Euclidean distance between any two residues. For each residue, several nearest neighboring residues are selected to establish stable edge connections. The initial edge features are obtained by encoding the spatial distance between backbone atoms of residues using a continuous function. This encoding method relies only on atomic spatial distance, naturally possessing translational and rotational invariance, and is unaffected by structural orientation, exhibiting extremely strong robustness. Simultaneously, the edge features further integrate the relative sequence position encoding of residue pairs with protein chain identification information, accurately distinguishing the interactions of residues from different molecular sources such as TCR α chain, TCR β chain, MHC chain, and antigenic peptide chain, thus obtaining the cross-chain binding characteristics within the complex.

[0055] Regarding node feature setting, this invention adopts a zero-vector initialization strategy, where the initial features of all amino acid residues are uniformly set to zero vectors. This avoids introducing any prior knowledge such as residue physicochemical properties or type preferences, thus mitigating the biases and limitations introduced by manually defined prior features. The effective feature representation of a node is obtained through subsequent multi-layer message passing processes, by aggregating the spatial structure information of neighboring edges and iteratively updating the learning process, allowing the model to autonomously discover the structural patterns of TCR-pMHC binding.

[0056] After completing the initial graph structure and basic feature construction, node and edge features are iteratively updated using a multi-layer message-passing neural network. Each round of message passing consists of two steps: first, based on the currently updated node features and residue distance features, edge feature representation is jointly optimized; then, each residue node aggregates the message information of all neighboring edges and autonomously updates its own node features. After iterative processing by the multi-layer encoder, each residue node can obtain perceptual features that integrate the global structural context. Among them, the CDR loop region residues located in the core region of the TCR and pMHC binding interface can encode key structural information that determines the binding affinity between the two. In this embodiment, the preferred number of residue neighbors is 48, and the unified dimension of node and edge features is set to 128, balancing model representation capability and computational efficiency.

[0057] like Figure 2 The diagram shown is a flowchart of the TCR-pMHC binding affinity prediction method of the present invention, illustrating the complete prediction process from constructing a complex structure dataset to sequence scoring and experimental correlation verification.

[0058] Example 2: Message Passing Graph Neural Network Architecture, Feature Fusion, and Mapping Relationship Establishment This embodiment details the core structure of the message-passing graph neural network with encoder-decoder architecture, the multi-layer feature fusion method, the sequence cross-entropy scoring method based on interaction features, and the correlation verification process between model prediction scores and experimental real affinity.

[0059] Architecture and feature fusion mechanism of message-passing graph neural networks: The message-passing graph neural network of the present invention adopts a layered stacked architecture of encoder and decoder. Both the encoder and decoder are composed of multiple message-passing layers. Each layer of the network independently completes the dual operations of edge feature update and node feature update. At the same time, a normalization processing and residual connection mechanism are added between layers to solve the gradient vanishing problem in deep network training and ensure the stability and convergence effect of model training.

[0060] The encoder module in this embodiment uses the standardized graph structure, initial node features, and edge features constructed in Embodiment 1 as input. It achieves bidirectional iterative updates of node and edge features through a multi-layer message passing mechanism. During each layer of encoding, edge features are first optimized and updated by combining real-time node features with residue spatial distance features. Then, each residue node aggregates the latest messages from all neighboring edges to iteratively optimize its own node features, achieving bidirectional interaction and deep fusion of node and edge information. After processing by the multi-layer encoder, each residue node in the complex learns a high-dimensional contextual feature representation that integrates global spatial structure, local interactions, and inter-chain associations.

[0061] The decoder module receives the global structural features from the encoder output and uses them to solve the TCR residue amino acid adaptation inverse folding problem, accurately predicting the optimal amino acid type for pMHC binding at each residue position. During the decoding phase, all edge feature parameters are fixed, and only the node features are iteratively optimized and updated to continuously explore the structural adaptation patterns at each residue site. The decoder output layer maps the optimized node features to the probability distribution of 20 standard amino acids, characterizing the binding adaptation degree of different amino acids at each residue position.

[0062] During the model training phase, the difference between the amino acid probability distribution predicted by the network and the actual natural TCR sequence is used as the optimization objective. The model parameters are iteratively converged by minimizing the cross-entropy loss function. Preferably, this invention also applies small random perturbations to the protein backbone atomic coordinates to simulate the unavoidable small structural biases during the initial structure prediction and molecular dynamics relaxation processes, significantly improving the model's generalization performance and robustness under different structural samples and experimental systems. After complete encoding and decoding training, the network can output the amino acid preference probability distribution at each residue position of the TCR, with the amino acid distribution in the CDR binding loop region being the most strongly constrained by the pMHC interface structure and exhibiting the highest prediction accuracy.

[0063] Validation of the correlation between sequence scoring and experimental affinity data: This invention constructs a dedicated cross-entropy scoring function based on the amino acid probability distribution of the network output, using it as a quantitative proxy index for predicting TCR-pMHC binding affinity. For any TCR sequence to be detected, the model extracts the predicted log probability of the actual amino acid corresponding to each residue position, and obtains the comprehensive score value of the entire TCR sequence through weighted aggregation. This score value can directly characterize the binding ability of TCR to the corresponding pMHC; the higher the score value, the stronger the binding affinity.

[0064] To verify the effectiveness of the model scores and accurately reflect the relative differences in affinity among TCR mutants, this invention employs a same-system gradient validation strategy. Multiple TCR sequences covering low, medium, and high affinity gradients within the same experimental measurement system are selected as a unified reference benchmark, ensuring that the validation data covers the entire affinity range. Using a trained graph neural network, all TCR samples within the experimental system are scored with uniform cross-entropy. Subsequently, a linear correlation analysis is performed between the model's predicted scores and the negative logarithm (-logKd) of the experimentally measured affinity to verify the degree of linear correlation between the two.

[0065] Repeatability correlation analysis results show that the predicted scores of the model in this invention exhibit a high positive correlation with the experimentally measured -logKd values. Validation using TCR-pMHC structures with different affinity gradients as reference frameworks shows that the Pearson correlation coefficient consistently reaches above 0.80, accurately distinguishing the relative affinity of TCR mutants within the same system and meeting the requirements for high-throughput mutant ranking and screening. The inclusion of a full-gradient reference sample ensures the reliability of the model scores across the entire affinity range, confirming that the cross-entropy scoring function can accurately and stably characterize the differences in the actual binding affinity of TCR-pMHC, providing a reliable quantitative benchmark for subsequent high-throughput mutant screening and affinity-oriented optimization.

[0066] Example 3: Differential Optimization of TCR Molecular Affinity and Validation of Mutant Screening This embodiment illustrates the entire optimized process for TCR molecular affinity, from structure prediction, mutation site screening, high-throughput mutation sequence generation, intelligent model scoring and screening to wet experimental verification, to achieve targeted enhancement of TCR affinity.

[0067] Accurate prediction and relaxation optimization of the structure of the TCR-pMHC complex to be optimized: This embodiment utilizes the AlphaFold3 multi-chain complex prediction tool to model the target architecture. Leveraging its advanced diffusion architecture complex modeling capabilities, it captures the matching characteristics between the flexible conformation of the TCRCDR loop region and the pMHC antigen-binding groove. The modeling input parameters include the full-length amino acid sequences of the α-chain and β-chain of the wild-type TCR to be optimized, the pMHC heavy chain sequence, the β2 microglobulin sequence, and the specific antigen peptide sequence. This comprehensive input of multiple components ensures the accuracy of the complex assembly structure. After predicting the initial complex structure, further molecular dynamics simulation relaxation optimization is performed to eliminate local atomic stacking conflicts, unreasonable bond angles, and distorted conformations, extracting the stable conformation in thermodynamic equilibrium.

[0068] Screening using TCRCDR regions combined with hotspot mutation sites: This invention combines spatial structure contact analysis and energy decomposition analysis to screen high-value mutation sites, ensuring the effectiveness of mutation modification. First, the relaxed and optimized stable complex structure is imported into the PyMOL molecular visualization tool to visually observe the spatial orientation of the side chains of residues in each CDR loop region (CDR1, CDR2, CDR3) of the TCR, as well as the adhesion relationship between the residues and the pMHC protein surface. The shortest spatial distance between the heavy atoms of the side chains of all residues in the CDR region and the heavy atoms of the pMHC residues is measured. Residues with a spatial distance ≤4.5 Å are identified as contact residues that directly participate in interfacial binding, forming a preliminary set of candidate mutation sites.

[0069] Building upon this foundation, the MM-GBSA method for quantitative decomposition of binding free energy residues was introduced to accurately screen hotspot residues. Representative conformations from multiple frames were uniformly extracted from molecular dynamics equilibrium simulation trajectories, and the binding free energy contribution (ΔG_residue) of residues in each CDR region was calculated. Residues with negative free energy contributions and large absolute values ​​were screened; these residues are core hotspot residues that maintain TCR-pMHC binding stability and contribute to binding force. Finally, by combining spatial contact relationships and energy contributions, residues possessing both spatial binding advantages and energy contribution advantages were screened as core mutation sites to ensure the feasibility of affinity enhancement after mutation.

[0070] Differentiation strategy for high-throughput generation of mutant TCR sequence libraries: This invention employs an adaptive differentiated mutation strategy for different numbers of mutation sites, balancing the integrity, diversity, and structural stability of the mutant library while avoiding combinatorial explosion. When the number of mutation sites is less than a preset threshold, a saturation mutation strategy is used to exhaustively search for all 20 amino acid substitution combinations at each site. When the number of mutation sites is greater than or equal to the preset threshold, a mutant sequence is generated by random sampling based on the amino acid probability distribution of a graph neural network and controlled by the temperature parameter T.

[0071] The temperature parameter T is a dimensionless regulatory parameter, which is used to adaptively regulate the concentration and flatness of the amino acid sampling probability distribution; the parameter value satisfies T>0, with T=1.0 as the defining threshold: 0<T<1.0 corresponds to low-temperature sampling, which compresses the probability distribution interval to make the probability distribution more concentrated, preferentially samples conservative amino acids with high confidence and high adaptability, and ensures the overall structural stability of the mutated TCR; when T=1.0, sampling is performed using the original probability distribution output by the graph neural network decoder; T>1.0 corresponds to high-temperature sampling, which stretches the probability distribution interval to make the probability distribution more gentle, effectively increasing the sampling probability of low-probability amino acids and significantly expanding the diversity of mutant sequences; as the value of T continues to increase, the flatness of the probability distribution continues to rise, which can balance the structural compatibility and sequence diversity of the mutant library as needed, and construct a high-quality mutant library with a controllable scale.

[0072] Intelligent scoring of mutant sequences and screening of candidate mutants: Taking the relaxed equilibrium structure of wild-type TCR-pMHC as a unified reference framework, the mutation residue information of all mutant TCR sequences is mapped to the standard structure framework to avoid the interference of structural differences on the scoring results. The trained graph neural network model and the cross-entropy scoring function are called to perform independent affinity prediction and scoring for each mutant TCR sequence, and the reference scoring value of the wild-type TCR under the same reference framework is calculated synchronously. By comparing the scoring difference between the mutant and the wild type, mutants whose model prediction scores are significantly higher than the wild-type baseline are screened out and determined as candidate mutants with potentially improved affinity, completing high-throughput computational pre-screening before wet experiments and greatly reducing the subsequent experimental sample size.

[0073] Wet experimental verification of candidate mutants and acquisition of high-affinity TCRs: The high-quality candidate mutant TCR sequences obtained by computational screening are subjected to artificial gene synthesis to construct prokaryotic expression plasmids, and the target single-chain TCR protein with high purity and high activity is obtained through bacterial culture, protein induced expression, in vitro refolding, and multi-step purification processes; the corresponding specific pMHC protein is prepared synchronously for in vitro affinity detection. Surface plasmon resonance (SPR) or biolayer interferometry (BLI) technology is used to determine the equilibrium dissociation constant Kd of each mutant single-chain TCR and pMHC, and the Kd value of the wild-type single-chain TCR is used as a control to screen out high-quality mutant sequences with significantly improved binding affinity, completing the directed optimization of high-affinity TCRs.

[0074] Figure 3 This is a flowchart of the TCR molecule affinity optimization method of the present invention, showing the complete optimization process from complex structure prediction, mutation site screening, mutant sequence generation, model scoring to wet experimental verification.

[0075] Example 4: Effect verification of high-affinity optimized TCR sequences This embodiment verifies the actual binding performance of the mutant TCR sequences optimized in this invention through an in vitro SPR affinity assay. Following the aforementioned full-process calculation, screening, and experimental verification, two novel single-stranded TCR mutant sequences with significantly enhanced affinity and stable antigen specificity were finally obtained, namely SEQ ID NO:2 and SEQ ID NO:3.

[0076] SPR detection results under uniform experimental conditions showed that the equilibrium dissociation constant Kd between the wild-type single-chain TCR (SEQ ID NO:1) and the target pMHC was 3.40 μM; the optimized single-chain TCR-1 (SEQ ID NO:2) had a Kd value as low as 37.5 nM, representing an approximately 90-fold increase in affinity compared to the wild-type; and the optimized single-chain TCR-2 (SEQ ID NO:3) had a Kd value of 85.0 nM, representing an approximately 40-fold increase in affinity compared to the wild-type. Neither optimized sequence lost its original antigen recognition specificity, exhibiting stable protein structure and excellent binding performance. This verifies the accuracy and efficiency of the computational prediction and mutation optimization method of this invention, providing high-quality candidate molecular templates for TCR-T cell therapy and the development of TCR bispecific drugs.

[0077] Example 5: Predictive optimization system, electronic device, and computer-readable storage medium This embodiment details the modular architecture, functional implementation, and deployment application scenarios of the prediction and optimization system, electronic device, and computer-readable storage medium of the present invention.

[0078] Figure 4 This diagram illustrates the modular structure of the prediction and optimization system of this invention, showing functional modules such as data acquisition, structure construction, graph feature extraction, feature fusion scoring, mutation design, candidate selection, and validation output. Each module operates independently yet is interconnected, fully covering the entire process from raw data processing to experimental result output.

[0079] The data acquisition module is responsible for batch collecting the amino acid sequences and experimental affinity raw data of multi-source TCR α chain, β chain, MHC heavy chain, β2 microglobulin, and antigenic peptide, and completing data integration, deduplication, and cleaning. The structure construction module can autonomously call high-precision structure prediction tools such as TCRdock and AlphaFold3 to complete the three-dimensional structure prediction and molecular dynamics relaxation optimization of the complex, and output a stable equilibrium structure model.

[0080] The graph feature extraction module is responsible for transforming the 3D complex structure into a standardized topological graph structure. Using residues as nodes and spatial nearest neighbors as edges, it extracts and constructs multi-dimensional edge features containing distance encoding, relative position encoding, and chain identification information. Combined with zero-initialized node features, this forms the standard input for the model. The feature fusion and scoring module completes multiple rounds of iterative feature fusion based on an encoder-decoder graph neural network architecture. It outputs the interface structural interaction features of the TCR-pMHC complex and the probability distribution of amino acids at each residue position. Based on the structural features, it outputs a standardized model prediction score. Through correlation verification with gradient affinity samples in the same system, it confirms the linear correlation between the model prediction score and the negative logarithmic value of the true affinity. Based on this score, it completes the relative affinity assessment of the TCR sequence and the ranking of mutants.

[0081] The mutation design module identifies hotspot mutation sites in the CDR region based on structural analysis results, and adaptively switches between saturation mutation and temperature-controlled sampling strategies according to the number of sites to construct a high-throughput mutation sequence library; the candidate screening module scores mutants in batches and screens high-quality candidate mutants by comparing them with wild-type baselines; the validation output module summarizes candidate sequences, model prediction scores, ranking results, and SPR / BLI experimental validation data in a unified manner, providing complete data support for subsequent drug development.

[0082] Sequence List: SEQ ID NO: 1 (1) Sequence type: amino acid (2) Molecular type: protein (3) Length: 245 aa (4) Source: Artificial Sequence (5) Name: Wild-type single-stranded TCR-1 SQQGEEDPQALSIQEGENVTINCSYKTSINNLQWYRQDPGRGLVHLILIRSNEREKHSGRLRVTLDTSKKSSSLEITAVRPADTASYFCGYDENARLMFGDGTQLTVNPGGGSEGGGSEGGG SEGGGSEGGTGDAGVTQSPRHLTVPQGQEVTLRCKPISGHDYLFWYRQDPGRGLELLIYFNNNVPIDDSGMPEDRFSAKMPNASHSTLKIQPVEPRDSAVYFCASSLGPGELFFGEGSRLTVN SEQ ID NO: 2 (1) Sequence type: amino acid (2) Molecular type: protein (3) Length: 245 aa (4) Source: Artificial Sequence (5) Name: Optimized single-chain TCR-1 SQQGEEDPQALSIQEGENVTINCSYKTYIQNLQWYRQDPGRGLVHLILIRSNEREKHSGRLRVTLDTSKKSSSLEITAVRPADTASYFCGYDENARFTFGDGTQLTVNPGGGSEGGGSEGGG SEGGGSEGGTGDAGVTQSPRHLTVPQGQEVTLRCKPISGHDYLFWYRQDPGRGLELLIYFNNNVPIDDSGMPEDRFSAKMPNASHSTLKIQPVEPRDSAVYFCASSLGATELFFGEGSRLTVN SEQ ID NO: 3 (1) Sequence type: amino acid (2) Molecular type: protein (3) Length: 245 aa (4) Source: Artificial Sequence (5) Name: Optimized single-chain TCR-2 SQQGEEDPQALSIQEGENVTINCSYKTYIQNLQWYRQDPGRGLVHLILIRSNEREKHSGRLRVTLDTSKKSSSLEITAVRPADTASYFCGYDENARFMFGDGTQLTVNPGGGSEGGGSEGGG SEGGGSEGGTGDAGVTQSPRHLTVPQGQEVTLRCKPISGHDYLFWYRQDPGRGLELLIYFNNNVPIDDSGMPEDRFSAKMPNASHSTLKIQPVEPRDSAVYFCASSLGATELFFGEGSRLTVN The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent substitutions, and improvements made by those skilled in the art to the above embodiments without departing from the scope of the technical solution of the present invention, based on the technical essence of the present invention, shall still fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for predicting TCR-pMHC binding affinity based on graph neural networks, characterized in that, Includes the following steps: After deduplication and cleaning of multi-source TCR-pMHC amino acid pairing data, a TCR-pMHC complex structure dataset was constructed through three-dimensional structure deduction and prediction and molecular dynamics conformation relaxation optimization. Graph structure modeling was performed on the three-dimensional structure of the TCR-pMHC complex. The three-dimensional coordinates of the protein backbone and side chain atoms were extracted. Amino acid residues were used as nodes and the spatial nearest neighbor relationship of residues was used as edges. Node features were obtained by zero initialization of nodes and subsequent network iteration updates. Corresponding edge features were constructed by combining residue spatial distance, relative position and chain identification information. The constructed node features and edge features are input into a graph neural network. Based on the encoder-decoder architecture, the features are iteratively fused to output the interface structural interaction features between TCR and pMHC molecules and the amino acid probability distribution at each residue position. The interface structural interaction features are used to characterize the binding mode and binding capacity of TCR and pMHC. TCR sequences with gradient true affinity in the same experimental system were selected as reference benchmarks. Based on the interaction characteristics, each TCR sequence in the experimental system was scored to verify the linear correlation between the model prediction score and the negative logarithm of the true affinity. Based on the model prediction score, the relative evaluation of TCR-pMHC binding affinity and the ranking and screening of mutants were realized.

2. The TCR-pMHC binding affinity prediction method based on graph neural networks according to claim 1, characterized in that, The construction of the TCR-pMHC complex structure dataset includes: The amino acid sequences of the TCR α chain, β chain, and corresponding pMHC were obtained. Multi-source TCR-pMHC pairing data were integrated, and deduplication and cleaning were performed using the CDR3 sequence of the TCR α chain, the CDR3 sequence of the β chain, the MHC sequence, and the peptide sequence as co-bonds. The initial three-dimensional structure of the TCR-pMHC complex was predicted based on the amino acid sequence deduction. Then, conformational relaxation optimization was performed through molecular dynamics simulation to obtain a stable conformation, thus constructing a TCR-pMHC complex structure dataset.

3. The TCR-pMHC binding affinity prediction method based on graph neural networks according to claim 1, characterized in that, When constructing the graph structure and features, corresponding residues are selected to establish edge connections based on the preset number of nearest neighbors, and node features and edge features are uniformly configured according to the preset feature dimensions. The encoding layer implements bidirectional updates of node and edge features, while the decoding layer fixes the edge features and iteratively optimizes the node features to complete feature fusion. The graph neural network adopts an encoder-decoder architecture. The encoder layer updates node and edge features bidirectionally, while the decoder layer fixes edge features and iteratively optimizes node features. Through multi-layer message passing and feature aggregation, it outputs a structure context-aware feature representation.

4. A method for optimizing the molecular affinity of TCRs, characterized in that, Includes the following steps: Based on amino acid sequence deduction, the initial three-dimensional structure of the complex of TCR and pMHC to be optimized was predicted, and molecular dynamics conformation relaxation optimization was performed. Residues whose shortest spatial distance between the CDR region and the pMHC heavy atoms is no greater than a preset spatial distance threshold were identified as interface contact residues. Combined with the residue decomposition results of MM-GBSA binding free energy, hotspot residues with large absolute values ​​of free energy contribution were screened as mutation sites. Mutant sequences are generated based on the difference in the number of mutation sites: when the number of sites is less than a preset threshold, saturation mutation is used to exhaustively search for all amino acid substitution combinations; when the number of sites is greater than or equal to the preset threshold, the amino acid probability distribution based on the graph neural network is used to generate mutant sequences by adjusting the temperature parameter T, which can adaptively adjust the concentration and smoothness of the amino acid sampling probability distribution. The TCR-pMHC binding affinity prediction method based on graph neural networks as described in any one of claims 1 to 3 is used to score all mutant TCR sequences, and mutants with model prediction scores that are better than wild-type TCRs are selected as candidate mutants. Candidate mutant single-stranded TCR genes were synthesized and expression plasmids were constructed. The target protein was obtained through host expression, renaturation, and purification. Binding affinity was measured by SPR or BLI technology, and high-affinity TCRs with enhanced affinity were screened.

5. The method for optimizing TCR molecular affinity according to claim 4, characterized in that, The random sampling described herein involves adaptively adjusting the sampling preference of the amino acid probability distribution using a temperature parameter T, and constructing a controllable mutant library based on the structural stability and sequence diversity of the mutant sequences. For the amino acid scores output by the graph neural network decoder at the target residue position, the scores are first scaled using the temperature parameter T, and then the corresponding amino acid sampling probabilities are obtained through the Softmax function. The calculation formula is as follows: in which, is the amino acid sampling probability, is the model prediction score of the target amino acid at the current residue site, is the model prediction scores of all candidate amino acids at the current residue site, j is the index traversing all candidate amino acids; the temperature parameter T is greater than 0, taking T=1.0 as the boundary, low-temperature sampling is performed when 0<T<1.0, sampling is performed according to the original probability distribution output by the decoder when T=1.0, and high-temperature sampling is performed when T>1.

0. The sampling rule is: as the temperature parameter T increases, the probability distribution of amino acid sampling gradually becomes smoother, and the sampling probability of low-probability amino acids increases.

6. A high-affinity single-chain TCR targeting a specific antigen, characterized in that, The high-affinity single-chain TCR is prepared by the TCR molecular affinity optimization method according to claim 4 or 5. The CDR region of the single-chain TCR contains a mutation site, and the binding affinity is significantly improved compared with wild-type TCR. The amino acid sequence of the single-chain TCR is SEQ ID NO: 2 or SEQ ID NO:

3.

7. A TCR-pMHC binding affinity prediction and optimization system based on graph neural networks, characterized in that, To implement the TCR-pMHC binding affinity prediction method based on graph neural networks as described in any one of claims 1 to 3, or the TCR molecule affinity optimization method as described in claim 4 or 5, the method comprises: The data acquisition module is used to acquire amino acid pairing data and experimental affinity data of multi-source TCR α chain, β chain, and pMHC, and to perform deduplication and cleaning processing on the original multi-source pairing data. The structure building module is used to construct a TCR-pMHC complex structure dataset through three-dimensional structure deduction and prediction and molecular dynamics conformation relaxation optimization. The graph feature extraction module is used to abstract the three-dimensional structure of the optimized TCR-pMHC complex into a graph structure. It uses amino acid residues as nodes and the spatial nearest neighbor relationship of residues as edges. The node features are obtained by zero initialization of nodes and subsequent network iteration updates. The node features and corresponding edge features of the graph neural network are constructed by combining the spatial distance, relative position and chain identification information of residues. The feature fusion and scoring module is used to input the constructed node features and edge features into the graph neural network, perform iterative feature fusion based on the encoder-decoder architecture, and output the interface structural interaction features between TCR and pMHC molecules and the amino acid probability distribution at each residue position; the TCR sequence is scored based on the interaction features to obtain the model prediction score; The mutation design module is used to screen hotspot mutation sites in the CDR region of TCRs based on the relaxed optimized three-dimensional structure of the complex, and generate mutant TCR sequences in high throughput according to the difference in the number of sites. The candidate screening module is used to score the affinity of all mutant TCR sequences, compare the model prediction scores of mutant and wild-type TCRs, and screen candidate mutants with better model prediction scores. The validation output module is used to obtain experimental affinity validation data for candidate mutants and output candidate mutant sequences, prediction results, and validation results.

8. The TCR-pMHC binding affinity prediction and optimization system based on graph neural networks according to claim 7, characterized in that, The data acquisition module uses the CDR3 sequence of the TCR α chain, CDR3 sequence of the β chain, MHC sequence, and peptide sequence as co-bonds to complete the deduplication and cleaning of multi-source data; the graph feature extraction module extracts the three-dimensional coordinates of key atoms in the protein backbone and side chains, selects corresponding residues to establish edge connections according to a preset number of nearest neighbors, and uniformly configures the feature dimensions according to a preset feature dimension; the feature fusion and scoring module verifies the linear correlation between the model prediction score and the negative logarithm of the true affinity -logKd; the verification output module is specifically used to summarize the candidate mutant sequences, affinity model prediction scores and ranking results, collect and organize SPR or BLI measured affinity data from in vitro experimental systems, compare the binding performance with wild-type TCRs to complete the validity verification of candidate mutants, and output the result data of high affinity TCR candidate molecules.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores an executable computer program, which, when executed by the processor, implements the TCR-pMHC binding affinity prediction method based on graph neural networks as described in any one of claims 1 to 3 or the TCR molecule affinity optimization method as described in claim 4 or 5.

10. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by the processor of an electronic device, implements the TCR-pMHC binding affinity prediction method based on graph neural networks as described in any one of claims 1 to 3, or the TCR molecule affinity optimization method as described in claim 4 or 5.

Images