Peptide-HLA class I allele affinity prediction system based on protein language model and dual multi-instance attention mechanism

Abstract

The invention relates to the technical field of biological information and computational immunity, and discloses a peptide-HLA class I allele affinity prediction system based on a protein language model and a dual multi-instance attention mechanism. The prediction system comprises a peptide and human leukocyte antigen HLA class I molecule sequence coding and distributed expression module, a dynamic selection mechanism long short-term memory network LSTM multi-level feature extraction module, a dual multi-instance attention module and a feature fusion and affinity prediction module. According to the invention, for a molecular sequence containing HLA class I molecules and polypeptides, a self-trained epitope peptide protein language model and an HLA class I molecular protein language model are utilized to obtain a feature coding vector of an amino acid sequence so as to better characterize the molecular sequence; and for the dependence problem of variable lengths among amino acids, an improved dynamic selection mechanism LSTM is utilized to learn dependence characteristics among amino acids with different distances in a sequence, so that the beneficial effect of improving the prediction effect of HLA antigen presentation is achieved.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics and computational immunology, and in particular to a peptide-HLA class I allele affinity prediction system based on a protein language model and a dual multi-instance attention mechanism. Background Art

[0002] Major histocompatibility complex (MHC) molecules are key components of the immune system, responsible for detecting and presenting antigenic peptides to the cell surface, triggering T cell-mediated immune responses. Human leukocyte antigens (HLA) are HLA genes and their encoded products in humans. HLA molecules are divided into class I and class II, which interact with intracellular and extracellular antigens, respectively. It is of great significance to deeply understand the interaction between HLA molecules and antigenic peptides, especially the peptide-HLA binding process. However, traditional methods for experimental studies of peptide-HLA binding, such as affinity determination and liquid chromatography-mass spectrometry, are time-consuming, costly, and labor-intensive, so a variety of computational prediction methods have been developed. Among them, NetMHCpan-4.1 and MHCflurry-2.0 are the latest algorithms widely used in the prediction of HLA class I peptide binding. NetMHCpan-4.1 consists of several shallow neural networks, which integrate multiple shallow neural networks for training and prediction. However, since the data comes from limited HLA molecules and peptides, it may not be able to fully cover all HLA molecule subtypes and their interaction with peptides, resulting in poor prediction performance for some subtypes. MHCflurry-2.0 uses in vitro binding affinity data or single HLA allele (SA) mass spectrometry data as training data, and cannot use multiple HLA allele (MA) mass spectrometry data for model training, resulting in a large amount of multi-HLA allele mass spectrometry data not being effectively utilized.

[0003] These problems lead to a high false positive rate in existing tools, and it is necessary to develop new methods for the prediction of peptide-HLA binding affinity. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] In view of the shortcomings of the prior art, the present invention provides a peptide-HLA class I allele affinity prediction system based on a protein language model and a dual multi-instance attention mechanism, which has the advantages of improving the prediction effect of HLA antigen presentation and solves the above-mentioned technical problems.

[0006] (II) Technical solution

[0007] To achieve the above object, the present invention provides the following technical solutions: a peptide-HLA class I allele affinity prediction system based on a protein language model and a dual multi-instance attention mechanism, the prediction system comprising: a sequence encoding and distributed representation module for peptides and HLA class I molecules, a multi-level feature extraction module using a dynamic selection mechanism LSTM, a dual multi-instance attention module, and a feature fusion and affinity prediction module;

[0008] The peptide and HLA class I molecule sequence encoding and distributed representation module includes: an epitope peptide protein language model and an HLA class I molecule protein language model;

[0009] The dynamic selection mechanism LSTM performs a multi-level feature extraction module as follows: adding a dynamic selection mechanism to LSTM;

[0010] The dual multi-instance attention module includes two submodules: Module 1 is Seq-Attention, which is used to learn the attention weights between different amino acids in a single sequence; Module 2 is Bag-Attention, which is used to dynamically calculate the relative weight of each allele in multiple alleles for affinity contribution;

[0011] The feature fusion and affinity prediction module is used to fuse the allele features processed by the dual multi-instance attention module, and perform nonlinear transformation and feature compression on them.

[0012] Preferably, the epitope peptide protein language model and the HLA class I molecule protein language model are based on ProLLaMa and incorporate the concept of transfer learning, and are specifically optimized for epitope peptide sequence datasets and HLA class I molecule sequence datasets. Through the collected peptide-HLA I binding datasets, random masked language modeling tasks (Masked Language Modeling, MLM) are respectively implemented on the peptide sequences and HLA molecule sequences therein, and the sequence features of peptides and HLA I molecules are learned from the data in an unsupervised manner, which not only inherits the functions of ProLLaMa, but also further refines the understanding of sequence patterns and contextual relationships unique to epitope peptides and HLA I molecules. The epitope peptide protein language model and the HLA class I molecule protein language model convert the original amino acid sequence into a high-dimensional, continuous vector representation, extract the key structural information of the amino acid sequence, and also contain potential functional characteristics.

[0013] Preferably, the dynamic selection mechanism LSTM performs a multi-level feature extraction module as follows: in the historical hidden state collection before the amino acid sequence is input at the current moment, an optimal hidden state is taken and connected to the current moment input. The optimal hidden state at the historical moment is selected by adding a linear neural network layer, and then using SoftMax normalization to obtain the score of each state. The highest score is considered to be the best. The dynamic selection mechanism replaces the original LSTM model in which the current moment input is directly connected to the hidden state at the previous moment, and the long-distance useful information in the sequence is learned to solve the variable length dependency problem between amino acids in the current affinity prediction scenario.

[0014] Preferably, the dual multi-instance attention module receives peptide sequence feature vectors related to different alleles from the dynamically selected LSTM layer, and learns the importance of different features through Dual Multi-instance Attention, DMI-Attention training, with two granularity levels of the model, amino acid level and allele level, to adaptively highlight the allele features that have a greater impact on affinity, while weakening minor influencing factors.

[0015] Preferably, the feature fusion and affinity prediction module is used to integrate the allele features processed by the Dual Multi-instanceAttention and DMI-Attention modules into a global feature vector, which is then subjected to nonlinear transformation and feature compression through a series of fully connected layers (FCN). The function of the fully connected layer is to further refine the key features and map them to the actual affinity value domain. The last fully connected layer generates the binding probability for each HLA class I allele with a given peptide sequence.

[0016] Preferably, the sequence encoding and distributed representation module of the peptide and HLA class I molecule is expressed as:

[0017] Let the epitope peptide sequence be p=(p1,p2,…,p n ), where p i represents the code of the i-th amino acid;

[0018] Let the sequence of HLA class I molecules be h = (h1, h2, ..., h m ), where h j represents the code of the jth amino acid;

[0019] Epitope-peptide-protein language model LM p Map the epitope peptide sequence p into a high-dimensional continuous vector representation:

[0020] p=LM p (p)

[0021] HLA class I protein language model LM h Map the HLA class I molecule sequence h into a high-dimensional continuous vector representation:

[0022] h=LM h (h)

[0023] Where: indicates that vectors p and h capture the sequence features of epitope peptides and HLA class I molecules, including their structural information and potential functional characteristics.

[0024] Preferably, the dynamic selection mechanism LSTM performs multi-level feature extraction module expression as follows:

[0025] Assume that the input amino acid sequence is x=(x1,x2,…,x T ), where x t Represents the input feature vector of the sequence at time step t;

[0026] The hidden state h of LSTM t It is defined by the following recursive relation:

[0027] h t =LSTM(x t ,h t-1 )

[0028] Among them: LSTM is the calculation function of the LSTM unit, which is used to update and output the hidden state h of the current time step t , according to the current input x t and the hidden state h at the previous time step t-1 ;

[0029] The dynamic selection mechanism selects the optimal state of the historical hidden state through a linear neural network layer: a fully connected layer and a SoftMax normalization. As the optimal hidden state for the current time step:

[0030] The optimal state h among the hidden states h at time t-1 k ,

[0031] h k =Max(Softmax(f(h1,…,h t-1 ))

[0032] Where: f is used to calculate the historical hidden state h k Function of importance at the current time step t.

[0033] Preferably, the dual multi-instance attention module includes a Seq-Attention module expression:

[0034] For the molecular sequence (p i ,h i )i, let (p i ,h i )i has n amino acids, and each amino acid vector represents a dimension of m. Then Seq-Attention learns the weights α between different amino acids through the attention mechanism. i =(α i1 ,α i2 ,…,α in ):

[0035] α ik =Softmax(f att (p ik ))

[0036] Among them, f att is the function used to calculate the attention weights.

[0037] Preferably, the dual multi-instance attention module also includes a Bag-Attention module expression:

[0038] For a set of peptide-HLA class I molecule pairs (p i ,h il )i, assuming that each p i The corresponding number of alleles is l, and Bag-Attention dynamically calculates each peptide sequence p i Corresponding multiple alleles h il Contribution weight β for affinity i =(β i1 ,β i2 ,…,β il ):

[0039] β ik =Softmax(g att (h ik ))

[0040] Among them, g att is the function used to calculate the attention weights.

[0041] Preferably, the feature fusion and affinity prediction module expression is:

[0042] Suppose the peptide-HLA class I molecule sequence feature vector after being processed by the dual multi-instance attention module is (p i ,h ij ), where p i represents the i-th peptide sequence molecule, h ij Represents the j-th allele feature vector corresponding to the i-th peptide sequence molecule;

[0043] Feature Fusion:

[0044] The allele features processed by the dual multi-instance attention module are fused using weighted summation:

[0045]

[0046] Where: β ij is the multi-allele weight after being processed by the dual multi-instance attention module, and l represents each peptide p i the corresponding number of multi-allelic genes;

[0047] Affinity prediction:

[0048] For the fused feature vector h merged , through a series of fully connected layers (FCN) for nonlinear transformation and feature compression, and finally mapped to the actual affinity value domain. These fully connected layers are expressed as:

[0049]

[0050] in: is the final affinity prediction score. FCN is a fully connected neural network consisting of several layers, which is used to transform the feature vector h merged Mapped to the output space.

[0051] Compared with the prior art, the present invention provides a peptide-HLA class I allele affinity prediction system based on a protein language model and a dual multi-instance attention mechanism, which has the following beneficial effects:

[0052] 1. The present invention uses a self-trained epitope peptide protein language model and an HLA class I molecule protein language model to obtain a feature encoding vector of an amino acid sequence for a molecular sequence containing HLA class I molecules and polypeptides, thereby better characterizing the molecular sequence; for the variable length dependency problem between amino acids, an improved dynamic selection mechanism LSTM is used to learn the dependency characteristics between amino acids at different distances in the sequence. For the multi-specific properties existing in mass spectrometry data, a dual multi-instance attention module is used to learn the importance of different granularities of alleles (amino acid level and allele level), thereby achieving the beneficial effect of improving the prediction effect of HLA antigen presentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of the system of the present invention;

[0054] Figure 2 It is a schematic diagram of the test results of the F1 index on the single allele test data set of the present invention;

[0055] Figure 3It is a schematic diagram of the test results of the F1 index on the Anthem data set of the present invention;

[0056] Figure 4 It is a schematic diagram of the test results of the F1 value index on the multi-allele test data set of the present invention;

[0057] Figure 5 It is a schematic diagram of the test results of the MCC index on the multi-allele test data set of the present invention;

[0058] Figure 6 This is a schematic diagram of the test results of the top 30 multi-allelic MCC indicators of the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] See also Figure 1 , a peptide-HLA class I allele affinity prediction system based on a protein language model and a dual multi-instance attention mechanism, the prediction system comprising: a sequence encoding and distributed representation module for peptides and HLA class I molecules, a multi-level feature extraction module using a dynamic selection mechanism LSTM, a dual multi-instance attention module, and a feature fusion and affinity prediction module;

[0061] The peptide and HLA class I molecule sequence encoding and distributed representation module includes: an epitope peptide protein language model and an HLA class I molecule protein language model;

[0062] The dynamic selection mechanism LSTM performs a multi-level feature extraction module as follows: adding a dynamic selection mechanism to LSTM;

[0063] The dual multi-instance attention module includes two submodules: Module 1 is Seq-Attention, which is used to learn the attention weights between different amino acids in a single sequence; Module 2 is Bag-Attention, which is used to dynamically calculate the relative weight of each allele in multiple alleles for affinity contribution;

[0064] The feature fusion and affinity prediction module is used to fuse the allele features processed by the dual multi-instance attention module, and perform nonlinear transformation and feature compression on them.

[0065] Preferably, the sequence encoding and distributed representation module of the peptide and HLA class I molecule is expressed as:

[0066] Let the epitope peptide sequence be p=(p1,p2,…,p n ), where p i represents the code of the i-th amino acid;

[0067] Let the sequence of HLA class I molecules be h = (h1, h2, ..., h m ), where h j represents the code of the jth amino acid;

[0068] Epitope-peptide-protein language model LM p Map the epitope peptide sequence p into a high-dimensional continuous vector representation:

[0069] p=LM p (p)

[0070] HLA class I protein language model LM h Map the HLA class I molecule sequence h into a high-dimensional continuous vector representation:

[0071] h=LM h (h)

[0072] Where: indicates that vectors p and h capture the sequence features of epitope peptides and HLA class I molecules, including their structural information and potential functional characteristics.

[0073] The protein language model is used to map peptide and HLA class I molecule sequences into high-dimensional continuous vectors p and h. These vectors capture rich feature information of the sequence and retain the structural and functional characteristics of the sequence. The module can handle peptide and HLA class I molecule sequences of various lengths and different features, so it is suitable for diverse bioinformatics data analysis needs. Through sequence encoding and distributed representation, the module can fully capture the sequence features of epitope peptides and HLA class I molecules, including but not limited to amino acid composition, structure, and function-related information. This enables the module to provide more accurate and comprehensive feature descriptions when predicting peptide-HLA class I molecule binding affinity.

[0074] Preferably, the dynamic selection mechanism LSTM performs multi-level feature extraction module expression as follows:

[0075] Assume that the input amino acid sequence is x=(x1,x2,…,x T ), where x t Represents the input feature vector of the sequence at time step t;

[0076] The hidden state h of LSTM t It is defined by the following recursive relation:

[0077] h t=LSTM(x t ,h t-1 )

[0078] Among them: LSTM is the calculation function of the LSTM unit, which is used to update and output the hidden state h of the current time step t , according to the current input x t and the hidden state h at the previous time step t-1 ;

[0079] The dynamic selection mechanism selects the optimal state of the historical hidden state through a linear neural network layer: a fully connected layer and a SoftMax normalization. As the optimal hidden state for the current time step:

[0080] The optimal state h among the hidden states h at time t-1 k ,

[0081] h k =Max(Softmax(f(h1,…,h t-1 ))

[0082] Where: f is used to calculate the historical hidden state h k Function of importance at the current time step t.

[0083] The recursive relationship makes the module have significant advantages in processing biological sequence data with strong time correlation, such as protein or DNA sequence analysis. The module uses LSTM units to extract multi-level features, and the hidden state h of each time step is t It includes the accumulation and integration of past information, which can gradually improve the representation ability of features and help to more accurately capture the complex structure and pattern of the sequence. It introduces a dynamic selection mechanism to weight the historical hidden state h through the fully connected layer and SoftMax normalization. k , thereby selecting the optimal hidden state for the current time step t This mechanism enables the module to adaptively focus on and integrate information that is particularly important to the current task, improving the model's predictive performance and generalization capabilities. Due to the clear definition of mathematical expressions, the module has good interpretability and flexibility, and is easy to understand and adjust.

[0084] Preferably, the dual multi-instance attention module includes a Seq-Attention module expression:

[0085] For the molecular sequence (p i ,h i )i, let (p i ,h i)i has n amino acids, and each amino acid vector represents a dimension of m. Then Seq-Attention learns the weights α between different amino acids through the attention mechanism. i =(α i1 ,α i2 ,…,α in ):

[0086] α ik =Softmax(f att (p ik ))

[0087] Among them, f att is the function used to calculate the attention weights.

[0088] Preferably, the dual multi-instance attention module also includes a Bag-Attention module expression:

[0089] For a set of peptide-HLA class I molecule pairs (p i ,h il )i, assuming that each p i The corresponding number of alleles is l, and Bag-Attention dynamically calculates each peptide sequence p i Corresponding multiple alleles h il Contribution weight β for affinity i =(β i1 ,β i2 ,…,β il ):

[0090] β ik =Softmax(g att (h ik ))

[0091] Among them, g att is the function used to calculate the attention weights.

[0092] Precise attention mechanism: The Seq-Attention module uses the attention mechanism α i =(α i1 ,α i2 ,…,α in ) learns the weights between amino acids in different peptide sequences, which can accurately capture the important features within the sequence and help distinguish and strengthen the contribution of key amino acids to affinity. The Bag-Attention module dynamically calculates the h j Contribution weight β for affinity j =(β j1 ,β j2 ,…,β jm), which can effectively integrate the information of multiple peptide-HLA class I molecule pairs. This method takes into account the complex relationship between each allele and all related peptide sequences, and improves the adaptability of the prediction model to different pairings. The attention weight calculation in the module adopts the Softmax function, which has good interpretability and flexibility. The weight distribution can be adjusted according to actual needs to optimize model performance. This flexibility makes the module widely applicable in different data sets and application scenarios. Through the combined use of Seq-Attention and Bag-Attention, the module can learn and integrate feature importance at different granularities, thereby improving the prediction accuracy and stability of peptide-HLA class I molecule affinity.

[0093] Preferably, the feature fusion and affinity prediction module expression is:

[0094] Suppose the peptide-HLA class I molecule sequence feature vector after being processed by the dual multi-instance attention module is (p i ,h ij ), where p i represents the i-th peptide sequence molecule, h ij Represents the j-th allele feature vector corresponding to the i-th peptide sequence molecule;

[0095] Feature Fusion:

[0096] The allele features processed by the dual multi-instance attention module are fused using weighted summation:

[0097]

[0098] Where: β ij is the multi-allele weight after being processed by the dual multi-instance attention module, and l represents each peptide p i the corresponding number of multi-allelic genes;

[0099] Affinity prediction:

[0100] For the fused feature vector h merged , through a series of fully connected layers (FCN) for nonlinear transformation and feature compression, and finally mapped to the actual affinity value domain. These fully connected layers are represented as:

[0101]

[0102] in: is the final affinity prediction score. FCN is a fully connected neural network consisting of several layers, which is used to transform the feature vector h merged Mapped to the output space.

[0103] Figure 2This is the test result of the F1 index on the single allele test dataset. The dataset is an evaluation dataset provided by the NetMHCpan online website, involving 15 different single alleles. The results show that compared with two widely used software NetMHCpan-4.1 and MHC Flurry-2.0, the F1 value of the model in this application is slightly higher than that of the other two software. Figure 2 , 0.763vs 0.732-0.760. In another public dataset, Anthem dataset, the model of this application achieved a higher F1 value on peptides with a length of 11-14 amino acids. Figure 3 The mean F1 values ​​of different lengths were 0.788 vs. 0.729-0.741.

[0104] In the multi-allele data evaluation, the proposed model was applied to a test set containing 20,000 MA data points split from the original data set, and the proposed model F1 value was significantly higher than NetMHCpan-4.1EL for eluted ligand, 0.800 vs 0.575, p = 3.992e-8, NetMHCpan-4.1BA for binding affinity, 0.800 vs 0.558, p = 1.335e-8, and MHCflurry-2.0, 0.800 vs. 0.586, p = 9.107e-8, as shown in Figure 2 . Figure 4 , Matthews correlation coefficient MCC, 0.791vs0.580-0.603 also showed a similar performance improvement, which can provide a reliable measure even in the case of imbalanced data. Figure 5 The MCC of the top 30 alleles in the MA test set is as follows: Figure 6 shown.

[0105] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and alterations may be made to the embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A peptide-HLA class I allele affinity prediction system based on a protein language model and dual multi-instance attention mechanism, characterized in that The prediction system includes: a sequence encoding and distributed representation module for peptides and HLA class I molecules, a multi-level feature extraction module using a dynamic selection mechanism LSTM, a dual multi-instance attention module, and a feature fusion and affinity prediction module; The peptide and HLA class I molecule sequence encoding and distributed representation module includes: an epitope peptide protein language model and an HLA class I molecule protein language model; The dynamic selection mechanism LSTM performs a multi-level feature extraction module as follows: adding a dynamic selection mechanism to LSTM; The dual multi-instance attention module includes two submodules: Module 1 is Seq-Attention, which is used to learn the attention weights between different amino acids in a single sequence; Module 2 is Bag-Attention, which is used to dynamically calculate the relative weight of each allele in multiple alleles for affinity contribution; The feature fusion and affinity prediction module is used to fuse the allele features processed by the dual multi-instance attention module, and perform nonlinear transformation and feature compression on them.

2. The peptide-HLA class I allele affinity prediction system based on protein language model and dual multi-instance attention mechanism according to claim 1, characterized in that: The epitope peptide protein language model and the HLA class I molecule protein language model are based on ProLLaMa and incorporate the concept of transfer learning. They are specifically optimized for epitope peptide sequence datasets and HLA class I molecule sequence datasets. Through the collected peptide-HLA I binding datasets, random masked language modeling tasks (Masked Language Modeling, MLM) are respectively implemented on the peptide sequences and HLA molecule sequences therein, and the sequence features of peptides and HLA I molecules are learned from the data in an unsupervised manner. They not only inherit the functions of ProLLaMa, but also further refine the understanding of sequence patterns and contextual relationships unique to epitope peptides and HLA I molecules. The epitope peptide protein language model and the HLA class I molecule protein language model convert the original amino acid sequence into a high-dimensional, continuous vector representation, extract the key structural information of the amino acid sequence, and also contain potential functional characteristics.

3. The peptide-HLA class I allele affinity prediction system based on protein language model and dual multi-instance attention mechanism according to claim 1, characterized in that: The dynamic selection mechanism LSTM performs a multi-level feature extraction module as follows: in the historical hidden state collection before the current moment input of the amino acid sequence, an optimal hidden state is taken and connected with the current moment input. The optimal hidden state at the historical moment is selected by adding a linear neural network layer, and then using SoftMax normalization to obtain the score of each state. The highest score is considered to be the best. The dynamic selection mechanism replaces the original LSTM model in which the current moment input is directly connected to the hidden state at the previous moment, and the long-distance useful information in the sequence is learned to solve the variable length dependency problem between amino acids in the current affinity prediction scenario.

4. The peptide-HLA class I allele affinity prediction system based on protein language model and dual multi-instance attention mechanism according to claim 1, characterized in that: The dual multi-instance attention module receives peptide sequence feature vectors related to different alleles from the dynamically selected LSTM layer, and learns the importance of different features through Dual Multi-instance Attention, DMI-Attention dual multi-instance attention module training. The model has two granularities, amino acid level and allele level, to adaptively highlight the allele features that have a greater impact on affinity, while weakening minor influencing factors.

5. The peptide-HLA class I allele affinity prediction system based on protein language model and dual multi-instance attention mechanism according to claim 4, characterized in that: The feature fusion and affinity prediction module is used to integrate the allele features processed by the Dual Multi-instance Attention and DMI-Attention modules into a global feature vector, which is then subjected to nonlinear transformation and feature compression through a series of fully connected layers (FCN). The function of the fully connected layer is to further refine the key features, and the last fully connected layer generates the binding probability of each HLA class I allele to a given peptide sequence.

6. The peptide-HLA class I allele affinity prediction system based on protein language model and dual multi-instance attention mechanism according to claim 2, characterized in that: The sequence encoding and distributed representation module expressions of the peptide and HLA class I molecule are as follows: Let the epitope peptide sequence be p=(p1,p2,…,p n ), where p i represents the code of the i-th amino acid; Let the sequence of HLA class I molecules be h = (h1, h2, ..., h m ), where h j represents the code of the jth amino acid; Epitope-peptide-protein language model LM p Map the epitope peptide sequence p into a high-dimensional continuous vector representation: p=LM p (p) HLA class I protein language model LM h Map the HLA class I molecule sequence h into a high-dimensional continuous vector representation: h=LM h (h) Where: indicates that vectors p and h capture the sequence features of epitope peptides and HLA class I molecules, including their structural information and potential functional characteristics.

7. The peptide-HLA class I allele affinity prediction system based on protein language model and dual multi-instance attention mechanism according to claim 3, characterized in that: The dynamic selection mechanism LSTM performs multi-level feature extraction module expression as follows: Assume that the input amino acid sequence is x=(x1,x2,…,x T ), where x t Represents the input feature vector of the sequence at time step t; The hidden state h of LSTM t It is defined by the following recursive relation: h t =LSTM(x t ,h t-1 ) Among them: LSTM is the calculation function of the LSTM unit, which is used to update and output the hidden state h of the current time step t , according to the current input x t and the hidden state h at the previous time step t-1 ; The dynamic selection mechanism selects the optimal state of the historical hidden state through a linear neural network layer: a fully connected layer and a SoftMax normalization. As the optimal hidden state for the current time step: The optimal state h among the hidden states h at time t-1 k , h k =Max(Softmax(f(h1,…,h t-1 )) Where: f is used to calculate the historical hidden state h k Function of importance at the current time step t.

8. The peptide-HLA class I allele affinity prediction system based on protein language model and dual multi-instance attention mechanism according to claim 4, characterized in that: The dual multi-instance attention module includes the Seq-Attention module expression: For the molecular sequence (p i ,h i )i, let (p i ,h i )i has n amino acids, and each amino acid vector represents a dimension of m. Then Seq-Attention learns the weights α between different amino acids through the attention mechanism. i =(α i1 ,α i2 ,…,α in ): α ik =Softmax(f att (p ik )) Among them, f att is the function used to calculate the attention weights.

9. The peptide-HLA class I allele affinity prediction system based on protein language model and dual multi-instance attention mechanism according to claim 8, characterized in that: The dual multi-instance attention module also includes a Bag-Attention module expression: For a set of peptide-HLA class I molecule pairs (p i ,h il )i, assuming that each p i The corresponding number of alleles is l, and Bag-Attention dynamically calculates each peptide sequence p i Corresponding multiple alleles h il Contribution weight β for affinity i =(β i1 ,β i2 ,…,β il ): b ik =Softmax(g att (h ik )) Among them, g att is the function used to calculate the attention weights.

10. The peptide-HLA class I allele affinity prediction system based on protein language model and dual multi-instance attention mechanism according to claim 1, characterized in that: The expression of the feature fusion and affinity prediction module is: Suppose the peptide-HLA class I molecule sequence feature vector after being processed by the dual multi-instance attention module is (p i ,h ij ), where p i represents the i-th peptide sequence molecule, h ij Represents the j-th allele feature vector corresponding to the i-th peptide sequence molecule; Feature Fusion: The allele features processed by the dual multi-instance attention module are fused using weighted summation: Where: β ij is the multi-allele weight after being processed by the dual multi-instance attention module, and l represents each peptide p i the corresponding number of multi-allelic genes; Affinity prediction: For the fused feature vector h merged , through a series of fully connected layers (FCN) for nonlinear transformation and feature compression, and finally mapped to the actual affinity value domain. These fully connected layers are represented as: in: is the final affinity prediction score. FCN is a fully connected neural network consisting of several layers, which is used to transform the feature vector h merged Mapped to the output space.

Images