Peptide-hla interaction method based on epistatic interaction information on biological sequences and inter-sequence mutual induction mechanism

Abstract

The present application relates to a kind of Peptide-HLA interaction method based on biological sequence upper interaction information and intersequence mutual induction mechanism.It solves the problem of low prediction accuracy of the method for predicting Peptide and HLA in the prior art.It includes S1,Peptide,HLA data collection;S2, biological sequence feature encoding operation;S3, pre-processing and encoder feature extraction;S4, bidirectional gated attention network feature fusion;S5, final fusion and classification module;S6, analysis of the influence of different encoding methods on model performance;S7, verify the performance of specific level model.The advantages of the present application are that the biological sequence upper interaction information and intersequence mutual induction mechanism are introduced, which can capture the unified inductive bias from multiple sequence representations, significantly improve the stability and generalization ability of the model;Reduce the difficulty of model construction and expansion, improve the applicability in different immune prediction tasks;The prediction success rate and AUC value are significantly improved, which verifies the effectiveness and superiority of the model.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, specifically to a Peptide-HLA interaction method based on epistatic interaction information of biological sequences and the mutual induction mechanism between sequences. Background Technology

[0002] Existing methods for predicting peptide and HLA mainly include: methods based on traditional wet experiments, methods based on empirical scoring functions, and methods based on deep learning. Among them:

[0003] Traditional wet experimental methods are time-consuming and costly, making them unsuitable for large-scale screening. Existing methods rely on MS mass spectrometry and X-ray diffraction, which, compared to computational methods, limit the study of unverified HLA-Peptide pairs.

[0004] Empirical scoring function-based methods mainly generate sequence scoring functions based on position weight matrices (PWM) or blosum62 matrices. However, these methods are based on linear fitting and are difficult to accurately capture the interaction information between HLA-Peptides.

[0005] Deep learning-based methods are further subdivided into sequence-based, pseudo-image-based, and graph neural network-based methods. Traditional sequence-based methods mainly use models such as convolutional neural networks (CNN), recurrent neural networks (LSTM, GRU), and attention mechanisms, or combinations thereof, to model biological sequences. However, these methods struggle to characterize the long-range and short-range dependencies between peptides and HLA, and are unable to effectively capture the epistatic effect differences present in biological sequences. Pseudo-image-based methods forcibly represent HLA and peptides as images using a two-dimensional pixel array. However, this method brute-forces the sequence into an image, and this early fusion strategy, which characterizes sequence relationships during the sequence feature processing stage, is prone to information loss and fails to simulate the true information between sequences. Graph neural network-based methods typically trade complex architectures for minor improvements, are time-consuming, and cannot meet the needs of filtering massive amounts of data in real-world scenarios. Summary of the Invention

[0006] The purpose of this invention is to address the aforementioned technical problems by providing a Peptide-HLA interaction method based on biological sequence epistatic interaction information and sequence-to-sequence induction mechanisms, aiming to improve the prediction accuracy of Peptide-HLA interactions and enhance the generalization ability of the model.

[0007] To achieve the above objectives, the present invention employs the following technical solution: a Peptide-HLA interaction method based on epistatic interaction information of biological sequences and the mutual induction mechanism between sequences, comprising the following steps:

[0008] Data collection of S1, Peptide, and HLA;

[0009] S2, biological sequence feature encoding operation;

[0010] S3. Preprocessing and encoder feature extraction;

[0011] S4, bidirectional gating attention network feature fusion;

[0012] S5, final fusion and classification module;

[0013] S6. Analyze the impact of different encoding methods on model performance;

[0014] S7. Verify the performance of the specificity level model.

[0015] In step S1, the datasets are derived from the Anthem dataset and the datasets of Niu et al.; the Anthem dataset is integrated from the IEDB, EPIMHC, MHCBN and SYFPEITHI immunology databases.

[0016] In step S1, the dataset includes a training set and a test set. The training set contains 539,019 HLA-peptide pairing samples, and the test set contains 172,580 samples. The pMHC dataset is obtained from the work of Niu et al., whose training set contains 193,952 HLA-peptide pairing samples, including 120,694 positive samples and 73,258 negative samples, and the test set contains 9,407 positive samples and 10,097 negative samples.

[0017] In step S2, a feature-independent encoding form is constructed, and a unified mapping is performed. The characteristics of, among which:

[0018] d represents the feature dimension;

[0019] l represents the maximum length of the set sequence;

[0020] Before this operation, a unified encoding strategy is used for the input sequences Peptide and HLA, setting the maximum length of variable-length peptides to 15. For sequences that are not long enough, zero padding is used for alignment during the encoding process.

[0021] Step S3 specifically includes the following steps:

[0022] S31. Perform a linear mapping on the sequence features and project them into a high-dimensional latent space.

[0023] S32. Combining batch normalization with a nonlinear activation function.

[0024] S33. Use the Lyra encoder to extract preprocessed features to extract epistatic interaction information of biological sequences.

[0025] In step S4, attention calculation is performed on HLA sequence features using Peptide sequence features as query vectors, and simultaneously on Peptide sequence features using HLA sequence features as query vectors. This simulates the bidirectional induction effect between peptides and HLA molecules during binding and recognition from two directions. Furthermore, a gating network is introduced to adaptively regulate the information flow between Peptide and HLA, thereby suppressing redundant or irrelevant features and strengthening key interaction signals, thus obtaining a stable and biologically interpretable fusion representation.

[0026] Step S5 specifically includes the following steps:

[0027] S51. Take the immune sequence features fused by the bidirectional gating cross-attention mechanism in step S4 as input.

[0028] S52. Perform average pooling and max pooling operations on the fused features respectively to simultaneously extract global contextual information and key local discrimination clues at the sequence level.

[0029] S53. The two types of pooling results are concatenated along the feature dimension to form a refined global feature representation that includes both coarse-grained and fine-grained information.

[0030] S54. The fused features are input into a discriminant network composed of multiple linear mapping units, where a nonlinear decision boundary is introduced through a nonlinear activation function, and the predicted probability of Peptide–HLA interaction is output in the final layer by combining the Sigmoid function, thereby completing the discrimination of pMHC binding relationship.

[0031] In step S6, the method is compared with relevant methods in the pMHC prediction field at the general specificity level, and compared with nine methods on the Anthem dataset: MixMHCpred-2.0.2, NetMHCpan-4.1, NetMHCcons-1.1, NetMHCstabpan-1.0, ACME, MHCNetSeq, DeepSeqPan, Anthem, HLAB, and CapsNet-MHC.

[0032] In step S6, the coding performance is analyzed and verified to obtain the performance improvement compared to existing methods.

[0033] In step S7, the characteristic indicators of five alleles, HLA-A*02:01, HLA-A*01:01, HLA-B*08:01, HLA-B*18:01 and HLA-A*24:02, are compared to obtain performance data.

[0034] Compared with the prior art, the advantages of the present invention are as follows:

[0035] (1) For the first time, the information on the epistatic interaction between Peptide and HLA and the mechanism of mutual induction between sequences are used as key factors to predict their interaction. This is the first time that the information on the epistatic interaction between Peptide and HLA and the mechanism of mutual induction between sequences have been used in the Peptide and HLA interaction prediction method.

[0036] (2) It does not rely on feature engineering or specific encoding methods, supports multiple optional immune sequence encoding methods as input, and achieves feature space alignment through a unified Lyra modeling strategy. It captures a unified inductive bias from the distribution differences between different sequence representations, which significantly improves the stability and generalization ability of the model.

[0037] (3) The Lyra encoder based on the SSM structure and the bidirectional gating cross-attention mechanism were used. The Lyra encoder can capture local and long-range relationships in the sequence and fit the epistatic interactions of biological sequences. The fusion module can explicitly model the mutual induction mechanism between peptides and the Peptide-HLA complex from two directions, and more realistically depict the bidirectional physical constraints and information dependence in the process of immune molecule recognition.

[0038] (4) It adopts an equal-dimensional, modular structure design to avoid complex dimension matching and transformation operations, realizes plug-and-play functionality of each functional module, reduces the difficulty of model construction and expansion, and improves the applicability in different immune prediction tasks. At the same time, it has a low number of parameters and computational cost.

[0039] (5) Compared with existing methods TranspMHC and Capsnet-MHC, the present invention has significantly improved prediction success rate and AUC value, verifying the effectiveness and superiority of the model. Attached Figure Description

[0040] Figure 1 This is a flowchart of the present invention;

[0041] Figure 2 This is a schematic diagram comparing the overall ROC curves of the LyraMHC method in this invention on the Anthem dataset of the pMHC task;

[0042] Figure 3This is a schematic diagram comparing various metrics of the LyraMHC method in this invention on the Niu et al. pMHC task dataset.

[0043] Figure 4 This is a schematic diagram comparing various metrics of the LyraMHC method in this invention on the dataset of Niu et al.'s pMHC task using different encoding methods.

[0044] Figure 5 This is a schematic diagram comparing the allele-specificity levels of different encoding methods and the TranspMHC method in this invention; Detailed Implementation

[0045] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0046] like Figure 1-5 As shown, the Peptide-HLA interaction method, based on epistatic interaction information of biological sequences and the mutual induction mechanism between sequences, includes the following steps:

[0047] Data collection of S1, Peptide, and HLA;

[0048] S2, biological sequence feature encoding operation;

[0049] S3. Preprocessing and encoder feature extraction;

[0050] S4, bidirectional gating attention network feature fusion;

[0051] S5, final fusion and classification module;

[0052] S6. Analyze the impact of different encoding methods on model performance;

[0053] S7. Verify the performance of the specificity level model.

[0054] In step S1, the datasets are derived from the Anthem dataset and the datasets of Niu et al.; the Anthem dataset is integrated from the IEDB, EPIMHC, MHCBN and SYFPEITHI immunology databases.

[0055] In step S1, the dataset contains a training set and a test set.

[0056] The training set of the aforementioned dataset contains 539,019 HLA-peptide pairing samples, and the test set contains 172,580 samples. The pMHC dataset was obtained from the work of Niu et al., whose training set contains 193,952 HLA-peptide pairing samples, including 120,694 positive samples and 73,258 negative samples, and the test set contains 9,407 positive samples and 10,097 negative samples.

[0057] The scheme utilizes epistatic interaction information of biological sequences and the Peptide-HLA interaction mechanism, which is a mechanism for mutual induction between sequences, to predict potential interactions. It is the first method in the prediction of Peptide and HLA interactions to explicitly introduce epistatic interaction information and a bidirectional interaction modeling mechanism to predict the interaction between the two.

[0058] The theoretical basis is that the binding and recognition process between Peptide and HLA molecules is not determined solely by local amino acid matching, but also depends on higher-order epistatic interaction modes within the sequence and the bidirectional structural and energy-induced relationship between Peptide and HLA.

[0059] In step S2, a feature-independent encoding form is constructed, and a unified mapping is performed. The characteristics of, among which:

[0060] d represents the feature dimension;

[0061] l represents the maximum length of the set sequence;

[0062] Before this operation, a unified encoding strategy is used for the input sequences Peptide and HLA, setting the maximum length of variable-length peptides to 15. For sequences that are not long enough, zero padding is used for alignment during the encoding process.

[0063] Subsequent experiments demonstrated that the model can still capture the inductive bias of biological sequences and maintain good robustness under different feature encoding conditions.

[0064] A multi-source biological sequence encoding method that does not rely on feature engineering is constructed, supporting a variety of optional input features, including but not limited to: Blosum62, Blosum80, N_blosum, EDSSMat62, physical, one-hot, and splicing features of one-hot and Blosum62. This method can fully preserve the evolutionary and physicochemical information of immune sequences without relying on a single encoding paradigm.

[0065] Step S3 specifically includes the following steps:

[0066] S31. Perform a linear mapping on the sequence features and project them into a high-dimensional latent space.

[0067] S32. Combining batch normalization with nonlinear activation functions to enhance the stability and expressive power of feature distribution;

[0068] S33. Use the Lyra encoder to extract preprocessed features to extract epistatic interaction information of biological sequences.

[0069] The Lyra encoder comprises a gated convolution module (PGC) and a time-state model-based S4D module. The PGC, as the main module for core feature computation and information flow modulation, processes the received tensor through two parallel paths: one information flow performs one-dimensional convolution to extract local features, and the other multiplies it with the information flow used for convolution feature extraction, achieving information modulation. The final result is normalized by an RMSnorm layer. The features are then fed into the S4D model, which converts the time domain to the frequency domain using Fourier transform and performs calculations in the frequency domain, ultimately fitting the epistatic interaction relationships of the biological sequence through polynomials.

[0070] This solution introduces a Lyra encoder based on epistatic interaction information from biological sequences for feature extraction from peptide and HLA sequences. This encoder integrates a projection-gated convolution (PGC) module for efficient local modeling and a recurrent convolution structure based on a diagonalized state-space model (SSM) to capture long-range dependencies in the sequence. By organically combining these two types of modules, the Lyra encoder can establish a correlation between local key site features and global sequence semantics while maintaining computational efficiency, thereby effectively characterizing epistatic interaction patterns in immune sequences.

[0071] In step S4, attention is calculated on HLA sequence features using peptide sequence features as the query vector, and simultaneously, attention is calculated on peptide sequence features using HLA sequence features as the query vector. The binding relationship between peptide and HLA is essentially a combination of hydrophobic interactions and hydrogen bonds, resulting in a bidirectional induction. This approach simulates the bidirectional induction between peptides and HLA molecules during binding and recognition from two directions, and further introduces a gating network to adaptively regulate the information flow between peptide and HLA, suppressing redundant or irrelevant features and strengthening key interaction signals, thereby obtaining a stable and biologically interpretable fusion representation.

[0072] Based on the biological hypothesis of the interaction between Peptide and HLA molecules, a bidirectional attention mapping mechanism is explicitly established between the two types of immune sequence features to model their interdependence.

[0073] After obtaining the sequence representations of peptide and HLA, this scheme further introduces a bidirectional gated cross-attention mechanism to fuse the features of the two types of immune molecules. This mechanism establishes a bidirectional attention mapping relationship between the peptide and HLA sequences simultaneously and dynamically modulates the interaction information using a gating mechanism, thereby explicitly modeling the bidirectional dependency and mutual induction relationship between them. This approach more realistically reflects the bidirectional physical constraints experienced by peptide and HLA molecules during binding and recognition, providing more comprehensive interaction feature information for subsequent classification tasks.

[0074] Step S5 specifically includes the following steps:

[0075] S51. Take the immune sequence features fused by the bidirectional gating cross-attention mechanism in step S4 as input.

[0076] S52. Perform average pooling and max pooling operations on the fused features respectively to simultaneously extract global contextual information and key local discrimination clues at the sequence level.

[0077] S53. The two types of pooling results are concatenated along the feature dimension to form a refined global feature representation that includes both coarse-grained and fine-grained information.

[0078] S54. The fused features are input into a discriminant network composed of multiple linear mapping units, where a nonlinear decision boundary is introduced through a nonlinear activation function, and the predicted probability of Peptide–HLA interaction is output in the final layer by combining the Sigmoid function, thereby completing the discrimination of pMHC binding relationship.

[0079] After feature extraction and interaction modeling, this scheme inputs the obtained immune sequence interaction features into the prediction module for discrimination. The final prediction module consists of a multi-layer linear mapping unit and a sigmoid function, used to output the corresponding binding probability or recognition probability.

[0080] In step S6, the model is compared with relevant methods in the pMHC prediction field at the general specificity level. To evaluate the prediction accuracy of the constructed model, it is compared with nine methods on the Anthem dataset: MixMHCpred-2.0.2, NetMHCpan-4.1, NetMHCcons-1.1, NetMHCstabpan-1.0, ACME, MHCNetSeq, DeepSeqPan, Anthem, HLAB, and CapsNet-MHC.

[0081] In the Anthem dataset, LyraMHC consistently outperforms other methods in terms of average AUC across all datasets in the 8-12mer length range. Furthermore, in Niu et al.'s pMHC prediction task dataset, LyraMHC is compared to TranspMHC, NetMHCpan-4.0, ACME, BigMHC, pMTnet, CapsNet-MHC, TranspHLA, and STMHCpan. Compared to all these methods, LyraMHC achieves at least a 2.3% improvement on AOURC, a 2.8% improvement on AUPRC, a 2.3% improvement on ACC, and a 4.6% improvement on MCC.

[0082] In step S6, the coding performance is analyzed and verified to obtain the performance improvement compared to existing methods.

[0083] Ultimately, it was demonstrated that even the feature encoding one-hot encoding with minimal prior knowledge achieved at least a 1.3% improvement in AUROC, a 3.9% improvement in MCC, and a 2.0% improvement in ACC compared to all other methods.

[0084] In step S7, the characteristic indicators of five alleles, HLA-A*02:01, HLA-A*01:01, HLA-B*08:01, HLA-B*18:01 and HLA-A*24:02, are compared to obtain performance data.

[0085] LyraMHC outperforms all eight methods mentioned above in multiple metrics, including AUROC, AUPRC, Accuracy, Balanced Accuracy, F1, and MCC.

[0086] In summary, the principle of this embodiment lies in addressing the inherent characteristics of the coexistence of local sequence patterns and long-range dependencies, and the high degree of information coupling during the immune molecule recognition process. By introducing multiple immune sequence encoding methods and employing a Lyra encoder based on a time-state-space model to uniformly model different selectable features, a feature extraction process based on the epistatic interaction information of biological sequences is achieved. Furthermore, a bidirectional gated cross-attention mechanism is designed to explicitly model the interaction between peptides and HLA molecules, characterizing the bidirectional physical constraints and information dependencies between immune molecules at the structural level, providing a feature fusion method more consistent with biological mechanisms for the immune recognition process. Combining equal-dimensional modeling and modular structural design,

[0087] This approach achieves efficient integration between sequence encoding, feature extraction, interactive modeling, and prediction, reducing the complexity of model construction and expansion. Extensive experimental results demonstrate that, in generalized allele-specific scenarios, this approach outperforms existing methods on multiple publicly available Peptide-HLA datasets and maintains good robustness under different sequence encoding conditions and allele-specific scenarios. Furthermore, compared to TranspMHC's 123w parameters and Capsnet-MHC's 37w parameters, this invention has only 26w parameters, resulting in lower model complexity. Moreover, the model itself, based on the Lyra architecture of SSM, exhibits O(N) time complexity with sequence growth compared to the transformer. 2 The complexity of Lyra is O(N log N) as the sequence length increases. Therefore, it can achieve high-precision prediction of Peptide-HLA binding relationship identification while reducing computational cost and training complexity. It can be widely used in neoantigen screening and computation-aided design related to immunotherapy, and has important theoretical significance and broad application prospects.

[0088] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A Peptide-HLA interaction method based on epistatic interaction information of biological sequences and the mutual induction mechanism between sequences, characterized in that, Includes the following steps: Data collection of S1, Peptide, and HLA; S2, biological sequence feature encoding operation; S3. Preprocessing and encoder feature extraction; S4, bidirectional gating attention network feature fusion; S5, final fusion and classification module; S6. Analyze the impact of different encoding methods on model performance; S7. Verify the performance of the specificity level model.

2. The Peptide-HLA interaction method based on epistatic interaction information and inter-sequence induction mechanisms according to claim 1, characterized in that, In step S1, the dataset is derived from the Anthem dataset and the dataset of Niu et al.; wherein, the Anthem dataset is integrated from the IEDB, EPIMHC, MHCBN and SYFPEITHI immunology databases.

3. The Peptide-HLA interaction method based on epistatic interaction information and inter-sequence mutual induction mechanism according to claim 2, characterized in that, In step S1, the dataset includes a training set and a test set. The training set contains 539,019 HLA-peptide pairing samples, and the test set contains 172,580 samples. The pMHC dataset is obtained from the work of Niu et al., whose training set contains 193,952 HLA-peptide pairing samples, including 120,694 positive samples and 73,258 negative samples, and the test set contains 9,407 positive samples and 10,097 negative samples.

4. The Peptide-HLA interaction method based on epistatic interaction information and inter-sequence mutual induction mechanism according to claim 1, characterized in that, In step S2, a coding form independent of feature engineering (i.e., a specific encoding method) is constructed and uniformly mapped to... The characteristics of, among which: d represents the feature dimension; l represents the maximum length of the set sequence; Before this operation, a unified encoding strategy is used for the input sequences Peptide and HLA, setting the maximum length of variable-length peptides to 15. For sequences that are not long enough, zero padding is used for alignment during the encoding process.

5. The Peptide-HLA interaction method based on epistatic interaction information and inter-sequence induction mechanisms according to claim 4, characterized in that, Step S3 specifically includes the following steps: S31. Perform a linear mapping on the sequence features and project them into a high-dimensional latent space. S32. Combining batch normalization with a nonlinear activation function. S33. Use the Lyra encoder to extract preprocessed features to extract epistatic interaction information of biological sequences.

6. The Peptide-HLA interaction method based on epistatic interaction information and inter-sequence mutual induction mechanism according to claim 5, characterized in that, In step S4, attention is calculated on HLA sequence features using Peptide sequence features as query vectors, and simultaneously on Peptide sequence features using HLA sequence features as query vectors. This simulates the bidirectional induction effect between peptides and HLA molecules during binding and recognition from two directions. Furthermore, a gating network is introduced to adaptively regulate the information flow between Peptide and HLA, thereby suppressing redundant or irrelevant features and strengthening key interaction signals, thus obtaining a stable and biologically interpretable fusion representation.

7. The Peptide-HLA interaction method based on epistatic interaction information and inter-sequence mutual induction mechanism according to claim 6, characterized in that, Step S5 specifically includes the following steps: S51. Take the immune sequence features fused by the bidirectional gating cross-attention mechanism in step S4 as input. S52. Perform average pooling and max pooling operations on the fused features respectively to simultaneously extract global contextual information and key local discrimination clues at the sequence level. S53. The two pooling results are concatenated along the feature dimension to form a refined global feature representation that includes both coarse-grained and fine-grained information; S54. The fused features are input into a discriminant network composed of multiple linear mapping units, where a nonlinear decision boundary is introduced through a nonlinear activation function, and the predicted probability of Peptide–HLA interaction is output in the final layer by combining the Sigmoid function, thereby completing the discrimination of pMHC binding relationship.

8. The Peptide-HLA interaction method based on epistatic interaction information and inter-sequence mutual induction mechanism according to claim 7, characterized in that, In step S6, the method is compared with relevant methods in the pMHC prediction field at the general specificity level, and compared with nine methods on the Anthem dataset: MixMHCpred-2.0.2, NetMHCpan-4.1, NetMHCcons-1.1, NetMHCstabpan-1.0, ACME, MHCNetSeq, DeepSeqPan, Anthem, HLAB, and CapsNet-MHC.

9. The Peptide-HLA interaction method based on epistatic interaction information and inter-sequence mutual induction mechanism according to claim 8, characterized in that, In step S6, the coding performance is analyzed and verified to obtain the performance improvement compared to existing methods.

10. The Peptide-HLA interaction method based on epistatic interaction information and inter-sequence induction mechanisms according to claim 8, characterized in that, In step S7, the characteristic indicators of five alleles, HLA-A*02:01, HLA-A*01:01, HLA-B*08:01, HLA-B*18:01 and HLA-A*24:02, are compared to obtain performance data.

Images