Universal HLA antigen presentation prediction method and system based on protein language model and multi-modal neural network

Abstract

The invention discloses a universal HLA antigen presentation prediction method based on a protein language model and a multi-modal neural network, which comprises the following steps: firstly, extracting verified HLA binding peptide fragment sequences from an immune epitope database, generating equivalent non-epitope peptide fragment sequences in combination with a protein database, and constructing an HLA-I class and HLA-II class balanced data set; then, extracting sequence embedding characteristics and contact graph structure information of the peptide fragment sequence by utilizing a protein language model; the method comprises the following steps: processing a protein map constructed by a contact map through a map neural network to obtain global structure features, and meanwhile, carrying out regional convolution and residual convolution processing on sequence embedding by adopting a one-dimensional convolutional neural network (1DCNN) to extract local context sequence features; the method effectively fuses sequence semantics and space structure information, improves the accuracy and generalization ability of HLA antigen presentation prediction, and is suitable for immune recognition modeling tasks under various HLA subtypes.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics and artificial intelligence technology, and more specifically, relates to a universal HLA antigen presentation prediction method and system based on a protein language model and a multimodal neural network. Background Art

[0002] With the widespread application of artificial intelligence in the biomedical field, deep learning-based methods for predicting protein structure and function are constantly emerging, significantly advancing research in precision medicine, vaccine development, and immunotherapy. In the immune system, human leukocyte antigen (HLA) molecules play a key role in presenting exogenous or endogenous antigenic peptides to T cells, forming the core link in inducing immune responses. Accurately predicting whether a peptide has the potential to be presented by HLA has become a key issue in peptide vaccine screening, antigen recognition, and immune target discovery.

[0003] Existing HLA antigen presentation prediction methods can be mainly divided into two categories: (1) pan-allele modeling methods based on joint embedding, which usually input the peptide sequence and HLA molecule sequence into the model at the same time, and model their interaction mechanism through a shared representation space, thereby improving the model's generalization ability for new HLA subtypes; (2) allele-specific modeling methods, whose core idea is to construct a separate prediction model for each HLA allele, and achieve accurate prediction by learning the common features of its binding peptides.

[0004] However, the two existing HLA antigen presentation prediction methods mentioned above still have some non-negligible defects: (1) The pan-allele modeling method based on joint embedding is highly dependent on HLA sequences as input. In practical applications, if there is a lack of clear HLA subtype information for individuals, it is difficult to carry out universal screening tasks for pan-populations, which limits its adaptability and practicality in unknown HLA backgrounds. (2) Allele-specific modeling methods usually focus on learning the binding characteristics between HLA subtypes and their known peptides, but do not adequately model the presentation ability of the peptides themselves and lack the ability to independently judge the potential presentation of peptides. Therefore, they are limited in scenarios that require rapid generalization predictions, such as high-throughput peptide screening. (3) The above two methods generally use single-modal sequence input as the main basis for modeling, and fail to effectively integrate multimodal features such as the structural conformation of peptides, local amino acid environment, spatial characteristics of HLA binding sites, and contextual semantics. This leads to limitations in the model in capturing the complex biological mechanism of antigen presentation, which in turn affects the accuracy and interpretability of the prediction. Summary of the Invention

[0005] In response to the above-mentioned deficiencies or improvement needs of the prior art, the present invention provides a universal HLA antigen presentation prediction method and system based on a protein language model and a multimodal neural network. The method aims to improve the ability to generalize modeling of peptides from different sources, enhance the accuracy and biological interpretability of antigen presentation potential predictions, and effectively adapt to the needs of large-scale, high-throughput peptide screening tasks, thereby resolving the technical problem that the existing pan-allele modeling method based on joint embedding is highly dependent on HLA sequences as input. In practical applications, if there is a lack of clear individual HLA subtype information, it is difficult to carry out universal screening tasks for a pan-population, which limits its adaptability and practicality in unknown HLA backgrounds. In addition, existing allele-specific modeling methods usually focus on learning the binding characteristics between HLA subtypes and their known peptides, and lack the ability to model the presentation ability of the peptides themselves. They lack the ability to independently judge the potential presentation of peptides, and therefore have limited effectiveness in scenarios such as high-throughput peptide screening that require rapid generalization predictions. In addition, the two existing methods generally use single-modal sequence input as the main modeling basis, and fail to effectively integrate multimodal features such as the structural conformation of the peptides, the local amino acid environment, the spatial characteristics of the HLA binding site, and the contextual semantics. This leads to limitations in the model in capturing the complex biological mechanisms of antigen presentation, which in turn affects the accuracy and interpretability of the predictions.

[0006] To achieve the above objectives, according to one aspect of the present invention, a universal HLA antigen presentation prediction method based on a protein language model and a multimodal neural network is provided, comprising the following steps: (1) Obtaining a peptide sequence to be predicted, and performing fixed-length processing on the peptide sequence to obtain a peptide sequence after fixed-length processing; (2) Inputting the fixed-length peptide sequence obtained in step (1) into the protein language model to obtain the sequence embedding matrix and residue contact probability matrix corresponding to the peptide sequence; (3) The sequence embedding matrix and residue contact probability matrix corresponding to the peptide sequence obtained in step (2) are input into a pre-trained universal HLA antigen presentation prediction model to obtain a presentation probability prediction value for the peptide sequence.

[0007] Preferably, step (1) is specifically as follows: specifically, the HLA-I class peptide sequence is processed into a peptide sequence with a fixed length equal to 14, and the HLA-II class peptide sequence is processed into a peptide sequence with a fixed length equal to 21; for peptide sequences shorter than the fixed length, "1" is used to fill the ends so that all peptide sequences meet the input dimension requirements of the universal HLA antigen presentation prediction model; for peptide sequences longer than the fixed length, the amino acid residues at the tail of the peptide sequence are deleted until its length is equal to the fixed length.

[0008] Preferably, the universal HLA antigen presentation prediction model includes a structural encoder based on a graph isomorphic neural network, a sequence encoder based on a one-dimensional convolutional neural network, a feature fusion module, and a classification prediction module; The structural encoder based on graph isomorphism neural network consists of a graph construction layer, two graph convolution layers, two normalization layers and an average pooling layer; The 1DCNN-based sequence encoder consists of a region convolution initialization layer, two convolution layers, and a maximum pooling layer; The feature fusion module consists of a convolutional feature compression layer and a multimodal feature concatenation layer.

[0009] Preferably, the input of the graph construction layer is a sequence embedding feature of dimension B×N×1280 and a contact probability matrix of dimension B×N×N corresponding to the peptide sequence. The graph construction layer performs graph construction processing on the sequence embedding feature and the contact probability matrix to obtain a node feature representation of the protein structure graph processed by the graph construction layer of dimension L×1280, and performs feature transformation processing on the node feature representation to obtain a node feature representation of the protein structure graph processed by the graph construction layer of dimension L×256, where L=B×N, B represents the batch size; N represents the length of the peptide sequence, for HLA-I class peptide sequence, the length of the peptide sequence is N=14, and for HLA-II class peptide sequence, the length of the peptide sequence is N=21; The input of the first graph convolutional layer is the node feature representation of the protein structure graph processed by the graph construction layer with a dimension of L×256. The graph convolutional layer uses the graph isomorphism network to process the node feature representation to obtain the node feature representation of the protein structure graph processed by the first graph convolutional layer with a dimension of L×256. The input of the first normalization layer is the node feature representation of the protein structure graph with a dimension of L×256 after being processed by the first graph convolutional layer. The normalization layer performs PairNorm normalization on the node feature representation (the normalization method is set to first subtract the node feature mean and then uniformly scale, and the feature scaling coefficient is 1) to obtain the node feature representation of the protein structure graph with a dimension of L×256 after being processed by the first normalization layer.

[0010] Preferably, the input of the second graph convolution layer is a node feature representation of the protein structure graph having a dimension of L×256 and processed by the first normalization layer. The graph convolution layer performs convolution processing on the node feature representation using a graph isomorphism network to obtain a node feature representation of the protein structure graph having a dimension of L×128 and processed by the second graph convolution layer. The input of the second normalization layer is the node feature representation of the protein structure graph after processing by the second graph convolution layer, with a dimension of L×128. The normalization layer performs PairNorm normalization on the node feature representation (where the normalization method is set to first subtract the node feature mean and then uniformly scale, with the feature scaling factor being 1) to obtain the node feature representation of the protein structure graph after processing by the second normalization layer, with a dimension of L×128. The input of the average pooling layer is the node feature representation of the protein structure graph with a dimension of L×128 after being processed by the second normalization layer. The average pooling layer performs global average pooling on the node feature representation to obtain a graph representation vector with a dimension of B×128 after being processed by the average pooling layer.

[0011] Preferably, the input of the regional convolution initialization layer is a sequence embedding feature of dimension B×N×1280 corresponding to the peptide sequence. The regional convolution initialization layer transposes the sequence embedding matrix to obtain a transposed sequence embedding feature of dimension B×1280×N. Then, a one-dimensional convolution operation with a convolution kernel size of 3 is performed on the transposed sequence embedding feature to output an intermediate feature vector of dimension B×256×N-2. Then, the intermediate feature vector of dimension B×256×N-2 is edge-filled to obtain a local region feature of dimension B×256×N after processing by the regional convolution initialization layer. The input of the first convolutional layer is the local area feature with a dimension of B×256×N after being processed by the regional convolution initialization layer. The convolutional layer performs rectified linear unit (ReLU) activation, convolution processing, and Dropout operations on the local area feature to obtain an intermediate feature vector with a dimension of B×256×N after being processed by the first convolutional layer. The input of the second convolutional layer is the intermediate feature vector of dimension B×256×N after being processed by the first convolutional layer. The convolutional layer performs ReLU activation, convolution and Dropout operations on the intermediate feature vector in sequence to obtain a feature vector of dimension B×256×N after being processed by the second convolutional layer. The input of the maximum pooling layer is the local area features with a dimension of B×256×N after being processed by the regional convolution initialization layer and the feature vector with a dimension of B×256×N after being processed by the second convolution layer. The maximum pooling layer performs residual connection processing on the local area features and the feature vector to obtain a feature vector with a dimension of B×256×N after residual processing. Then, the feature vector with a dimension of B×256×N after residual processing is subjected to one-dimensional global maximum pooling processing to obtain a CNN feature vector with a dimension of B×256×1 after being processed by the maximum pooling layer; The input of the convolutional feature compression layer is the CNN feature vector with a dimension of B×256×1 after being processed by the maximum pooling layer. The convolutional feature compression layer compresses the CNN feature vector to obtain an intermediate feature vector with a dimension of B×256. Then, the intermediate feature vector is batch normalized to obtain a convolutional feature vector with a dimension of B×256. The input of the multimodal feature splicing layer is the convolution feature vector with a dimension of B×256 output by the convolution feature compression layer, and the graph representation vector with a dimension of B×128 after processing by the average pooling layer. The multimodal feature splicing layer splices the convolution feature vector and the graph representation vector to obtain a fused representation with a dimension of B×384.

[0012] Preferably, the input of the classification prediction module is the fused representation with a dimension of B×384 output by the multimodal feature splicing layer. The classification prediction module uses a fully connected layer to process the fused representation to obtain a vector with a dimension of B×64, and then performs linear classification processing on the vector to obtain a presentation probability prediction value with a dimension of B×1.

[0013] Preferably, the universal HLA antigen presentation prediction model is obtained by training through the following steps: (3-1) Extract a peptide sequence set consisting of verified HLA-I and HLA-II peptide sequences from the immune epitope database IEDB as a positive sample peptide sequence set, and use the protein database to generate a negative sample peptide sequence set equal to the positive sample peptide sequence set. The positive sample peptide sequence set and the negative sample peptide sequence set together constitute a dataset, and the dataset is divided into a training set and a test set in a ratio of 4:1, where each HLA-I peptide sequence and HLA-II peptide sequence has a true label; (3-2) For each peptide sequence in the training set obtained in step (3-1), the peptide sequence is fixed-length processed to obtain a fixed-length processed peptide sequence; then, the pre-trained evolutionary scaling model ESM is used to obtain the sequence embedding matrix corresponding to the peptide sequence with a dimension of B×N×1280 from the fixed-length processed peptide sequence. The contact probability matrix with dimension B×N×N ; (3-3) For each peptide sequence in the training set obtained in step (3-1), embed the sequence corresponding to the peptide sequence obtained in step (3-2) with a dimension of B×N×1280 into features The contact probability matrix with dimension B×N×N Input the graph construction layer in the structure encoder of the universal HLA antigen presentation prediction model to perform graph construction and transformation processing on the sequence embedding features and contact probability matrix, thereby obtaining the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×256 after being processed by the graph construction layer , where L = B × N; (3-4) For each peptide sequence in the training set obtained in step (3-1), the node feature representation of the protein structure graph corresponding to the peptide sequence obtained in step (3-3) after being processed by the graph construction layer is expressed as Input the first graph convolutional layer in the structure encoder in the universal HLA antigen presentation prediction model to represent the node feature Perform graph isomorphism network convolution processing to obtain the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×256 after the first graph convolution layer processing ; (3-5) For each peptide sequence in the training set obtained in step (3-1), the node feature representation of the protein structure graph with a dimension of L×256 and processed by the first graph convolution layer corresponding to the peptide sequence obtained in step (3-4) is Input to the first normalization layer in the structure encoder of the universal HLA antigen presentation prediction model to represent the node features of the protein structure graph PairNorm normalization is performed (where the normalization method is set to first subtract the node feature mean and then uniformly scale, with the feature scaling factor being 1) to obtain the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×256 after the first normalization layer. ; (3-6) For each peptide sequence in the training set obtained in step (3-1), the node feature representation of the protein structure graph with a dimension of L×256 and processed by the first normalization layer corresponding to the peptide sequence obtained in step (3-5) is Input the second graph convolutional layer in the structure encoder in the universal HLA antigen presentation prediction model to represent the node features of the protein structure graph Perform graph isomorphism network convolution processing to obtain the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×128 after the second graph convolution layer processing ; (3-7) For each peptide sequence in the training set obtained in step (3-1), the node feature representation of the protein structure graph with a dimension of L×128 and processed by the second graph convolution layer corresponding to the peptide sequence obtained in step (3-6) is Input the second normalization layer in the structure encoder in the universal HLA antigen presentation prediction model to represent the node features of the protein structure graph PairNorm normalization is performed (where the normalization method is set to first subtract the node feature mean and then uniformly scale, with the feature scaling factor being 1) to obtain the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×128 after the second normalization layer. ; (3-8) For each peptide sequence in the training set obtained in step (3-1), the node feature representation of the protein structure graph with a dimension of L×128 and processed by the second normalization layer corresponding to the peptide sequence obtained in step (3-7) is Input the average pooling layer in the structure encoder of the universal HLA antigen presentation prediction model to represent the node features of the protein structure graph Perform global average pooling to obtain the graph representation vector corresponding to the peptide sequence with a dimension of B×128 after average pooling layer processing ; (3-9) For each peptide sequence in the training set obtained in step (3-1), embed the sequence corresponding to the peptide sequence obtained in step (3-2) into a sequence embedding matrix with a dimension of B×N×1280 Input the regional convolution initialization layer in the sequence encoder of the universal HLA antigen presentation prediction model to obtain the local region features corresponding to the peptide sequence with a dimension of B×256×N after processing by the regional convolution initialization layer ; (3-10) For each peptide sequence in the training set obtained in step (3-1), the local region feature of the peptide sequence obtained in step (3-9) with a dimension of B×256×N after being processed by the regional convolution initialization layer is Input the first convolutional layer in the sequence encoder of the universal HLA antigen presentation prediction model to perform local region feature ReLU activation, convolution processing and Dropout operation are performed in sequence to obtain the feature intermediate vector corresponding to the peptide sequence with a dimension of B×256×N after the first convolution layer processing. ; (3-11) For each peptide sequence in the training set obtained in step (3-1), the intermediate feature vector corresponding to the peptide sequence obtained in step (3-10) with a dimension of B×256×N after the first convolutional layer processing is Input the second convolutional layer in the sequence encoder of the universal HLA antigen presentation prediction model to ReLU activation, convolution processing and Dropout operation are performed in sequence to obtain the feature vector corresponding to the peptide sequence with a dimension of B×256×N after the second convolution layer processing. ; (3-12) For each peptide sequence in the training set obtained in step (3-1), the local region feature of the peptide sequence obtained in step (3-9) with a dimension of B×256×N after processing by the regional convolution initialization layer is The feature vector corresponding to the peptide sequence obtained in step (3-11) with a dimension of B×256×N after processing by the second convolution layer Input the maximum pooling layer in the sequence encoder of the universal HLA antigen presentation prediction model to obtain the CNN feature vector corresponding to the peptide sequence with a dimension of B×256×1 after the maximum pooling layer processing ; (3-13) For each peptide sequence in the training set obtained in step (3-1), the CNN feature vector corresponding to the peptide sequence obtained in step (3-12) with a dimension of B×256×1 and processed by the maximum pooling layer is Input the convolutional feature compression layer in the feature fusion module in the universal HLA antigen presentation prediction model to Compression processing is performed to obtain the intermediate feature vector corresponding to the peptide sequence with a dimension of B×256 , and the intermediate eigenvector Perform batch normalization to obtain the convolution feature vector corresponding to the peptide sequence with a dimension of B×256 ; (3-14) For each peptide sequence in the training set obtained in step (3-1), the graph representation vector corresponding to the peptide sequence obtained in step (3-8) with a dimension of B×128 after the average pooling layer is processed The convolution feature vector of dimension B×256 corresponding to the peptide sequence obtained in step (3-13) Input the multimodal feature concatenation layer in the feature fusion module of the universal HLA antigen presentation prediction model to convolute the feature vector and graph representation vector Perform splicing to obtain a fusion representation of the peptide sequence with a dimension of B×384 ; (3-15) For each peptide sequence in the training set obtained in step (3-1), the fusion representation of the dimension B×384 corresponding to the peptide sequence obtained in step (3-14) is Input the multimodal feature concatenation layer in the feature fusion module in the universal HLA antigen presentation prediction model to represent the fusion Perform full connection layer processing to obtain a vector of dimension B×64 corresponding to the peptide sequence , and the vector of dimension B×64 corresponding to the peptide sequence Perform linear classification processing to obtain the presentation probability prediction value corresponding to the peptide sequence with a dimension of B×1 ; (3-16) For each peptide sequence in the training set obtained in step (3-1), the predicted value of the presentation probability of dimension B×1 corresponding to the true label corresponding to the peptide sequence obtained in step (3-1) and the peptide sequence obtained in step (3-15) is A binary cross entropy loss function is constructed, and the universal HLA antigen presentation prediction model is iteratively optimized and trained using the back propagation algorithm until the binary cross entropy loss function converges, thereby obtaining a preliminarily trained universal HLA antigen presentation prediction model; the loss function used in this step is: ; Where S represents the total number of peptide sequences in the training set; ∈{0,1} represents the true label of the i-th peptide sequence in the training set, where 1 indicates that the peptide sequence has presentation ability, and 0 indicates that the peptide sequence has no presentation ability; ∈{0,1} represents the probability that the universal HLA antigen presentation prediction model predicts that the i-th sample is presented; (3-17) Using the test set obtained in step (3-1), the preliminarily trained universal HLA antigen presentation prediction model obtained in step (3-16) is verified, and the universal HLA antigen presentation prediction model with the best performance in the verification results is selected as the final trained universal HLA antigen presentation prediction model.

[0014] Preferably, the process of generating a negative sample peptide sequence set equal in number to the positive sample peptide sequence set in step (3-1) is as follows: first, using the Double Index Alignment of Next-generation Sequencing Data (DIAMOND) tool, each positive sample peptide sequence in the positive sample peptide sequence set is aligned to a non-redundant protein database to locate the source protein corresponding to the positive sample peptide sequence. Multiple peptide sequences are randomly intercepted from positions other than the known epitope region in the source protein (HLA-I: peptide sequences with a length of 8 to 14 amino acids are intercepted; HLA-II: peptide sequences with a length of 13 to 21 amino acids are intercepted). CD-HIT, a tool for removing sequence redundancy and clustering, is used to remove highly similar peptide sequences from the multiple peptide sequences obtained, and finally a negative sample peptide sequence set is generated; The fixed-length processing in step (3-2) is specifically to perform "1" end padding on peptide sequences shorter than the fixed length. That is, for HLA-I peptide sequences, the fixed amino acid sequence length is 14; for HLA-II peptide sequences, the fixed amino acid sequence length is 21, to meet the input requirements of the universal HLA antigen presentation prediction model; Steps (3-9) are as follows: First, embed the sequence matrix Perform transposition to obtain a sequence embedding feature with a dimension of B×1280×N Then, the sequence is embedded into the feature Perform regional convolution and edge filling to obtain the sequence embedding feature tensor corresponding to the peptide sequence with a dimension of B×256×N ; Steps (3-12) are as follows: first, the local area features of the peptide sequence with a dimension of B×256×N after the regional convolution initialization layer processing are processed. The feature vector corresponding to the peptide sequence with a dimension of B×256×N after processing by the second convolutional layer After residual connection processing, the feature vector corresponding to the peptide sequence with a dimension of B×256×N after residual processing is obtained. Then, the feature vector corresponding to the peptide sequence with a dimension of B×256×N after residual processing is Perform one-dimensional global maximum pooling to obtain the CNN feature vector corresponding to the peptide sequence with a dimension of B×256×1 after the maximum pooling layer processing ; According to another aspect of the present invention, a universal HLA antigen presentation prediction system based on a protein language model and a multimodal neural network is provided, comprising: The first module is used to obtain a peptide sequence to be predicted and perform fixed-length processing on the peptide sequence to obtain a peptide sequence after fixed-length processing; The second module is used to input the peptide sequence obtained by the first module after fixed-length processing into the protein language model to obtain the sequence embedding matrix and residue contact probability matrix corresponding to the peptide sequence; The third module is used to input the pre-trained universal HLA antigen presentation prediction model into the sequence embedding matrix and residue contact probability matrix corresponding to the peptide sequence obtained in the second module to obtain the presentation probability prediction value of the peptide sequence.

[0015] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art: (1) The present invention adopts steps (1) to (3-13) and adopts a prediction paradigm that does not rely on HLA sequence input, that is, it does not explicitly introduce HLA sequence or subtype characteristics, but makes predictions based on the multimodal characteristics of the peptide itself, thereby solving the technical problem of the existing pan-allele modeling method based on joint embedding that is highly dependent on HLA sequence as input; (2) The present invention adopts steps (3-1) to (3-10) to design a dual-path encoding mechanism that integrates a protein language model and a multimodal neural network. On the one hand, the protein language model is used to extract rich contextual semantic features from the peptide sequence to capture the intrinsic relationship between amino acid residues; on the other hand, a residue graph structure based on contact probability is independently constructed, and its spatial configuration and structural pattern are modeled using a graph neural network. By extracting structural and semantic information through multi-layer GIN graph convolution and sequence CNN modules and fusing them, it is finally possible to accurately judge the presentation potential of a peptide based solely on its own characteristics without the need for clear HLA sequence information, significantly improving the independent prediction ability and generalization performance of the model, thereby solving the technical problem of insufficient modeling of the presentation ability of the peptide itself in existing allele-specific modeling methods; (3) The present invention adopts steps (3-11) to (3-13), which proposes a multimodal feature fusion mechanism to jointly embed the graph structure feature vector extracted by the structure encoder and the convolution feature vector extracted by the sequence encoder. In the fusion module, the CNN features are first compressed and standardized, and then spliced ​​with the GIN structure vector to form a fusion representation, and input into the two-layer fully connected network and the final classifier. During the fusion process, the independence of sequence and structural features is retained, and at the same time, a learnable low-dimensional shared representation is introduced through linear transformation, which effectively enhances the information integration ability of peptides in multimodal space. This design realizes the effective coupling of sequence semantics, spatial structure and local area perception, improves the model's deep understanding and discrimination of HLA presentation patterns, and can thus solve the technical problems of lack of prediction accuracy and interpretability in existing pan-allele modeling based on joint embedding and allele-specific modeling methods; (4) The present invention minimizes the difference between the predicted probability output by the universal HLA antigen presentation prediction model and the true label through the standard binary cross entropy loss function adopted in steps (3-14) to (3-15), thereby obtaining a presentation judgment capability with robustness and generalization ability; the loss function has a symmetrical modeling effect on positive samples and negative samples during the training process, and further combined with the balanced data set constructed in the early stage, it effectively avoids category bias and overfitting problems during the loss function optimization process, which is conducive to improving the stability and adaptability of the universal HLA antigen presentation prediction model in actual unknown samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of the universal HLA antigen presentation prediction method based on protein language model and multimodal neural network of the present invention.

[0017] Figure 2 Schematic diagram of the network structure of the universal HLA antigen presentation prediction model based on the protein language model and multimodal neural network used in the method of the present invention. DETAILED DESCRIPTION

[0018] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0019] This paper proposes a universal HLA antigen presentation prediction method based on a protein language model and a multimodal neural network. Its core design combines a protein language model with a multimodal feature extraction network to accurately identify peptides with HLA presentation potential through sequence semantic understanding and structural graph modeling. The method first uses a protein language model to perform high-dimensional embedding encoding on the input peptide sequence. This protein language model has been self-supervised trained on a large number of protein sequences and is capable of capturing deep semantic and contextual information between amino acid residues, thereby generating a sequence representation with biological function characteristics. The system then extracts multimodal features using two parallel pathways: the first is a structure-aware graph neural network pathway. This pathway uses the residue probabilistic contact matrix simultaneously acquired by the protein language model and combines it with the embedded features to construct a protein graph, where nodes represent amino acid residues and edges represent spatial contact relationships between residues. This constructed graph is then fed into a graph isomorphism network for graph structure modeling. Two layers of GIN convolution are used with PairNorm normalization to avoid over-smoothing. A fixed-dimensional graph structure representation is then obtained through global average pooling. The second pathway is a sequence convolution pathway. This pathway uses a 1D convolutional network to extract regional features from sequence embeddings. It first extracts structural fragment features of continuous residue windows through local convolution, and designs residual connections to enhance gradient stability. After processing with two consecutive layers of convolution blocks, it uses maximum pooling for step-by-step downsampling to extract a compressed representation at the sequence level. Finally, it is normalized by BatchNorm to form a regional feature vector. The outputs of the above two pathways are graph structure features and convolution features, respectively. The system performs feature splicing on the two to form a final fusion representation. The fusion feature is then mapped to a low-dimensional space through a two-layer fully connected network, and finally input into a linear classifier to predict whether the peptide has the potential to present HLA molecules. The method described in the present invention overcomes the limitations of traditional rule-based or fixed template methods in terms of weak pan-allele prediction ability, poor model scalability, and lack of structural perception, achieves accurate prediction of unknown peptides, and significantly improves the versatility, accuracy, and practicality of HLA antigen presentation modeling.

[0020] To make the technical solutions and advantages of the present invention more clearly understood, the entire universal HLA antigen presentation prediction method is further described in detail below with reference to the accompanying drawings.

[0021] like Figure 1 As shown, the present invention provides a universal HLA antigen presentation prediction method based on a protein language model and a multimodal neural network, comprising the following steps: (1) Obtaining a peptide sequence to be predicted, and performing fixed-length processing on the peptide sequence to obtain a peptide sequence after fixed-length processing; Specifically, this step processes the HLA-I class peptide sequence into a peptide sequence with a fixed length of 14, and the HLA-II class peptide sequence into a peptide sequence with a fixed length of 21. For peptide sequences shorter than the fixed length, "1" is used to fill the ends so that all peptide sequences meet the input dimension requirements of the universal HLA antigen presentation prediction model. For peptide sequences longer than the fixed length, the amino acid residues at the tail of the peptide sequence are deleted until its length equals the fixed length.

[0022] (2) Inputting the fixed-length peptide sequence obtained in step (1) into the protein language model to obtain the sequence embedding matrix and residue contact probability matrix corresponding to the peptide sequence; The protein language model in this step is the Evolutionary Scale Modeling (ESM) model.

[0023] (3) The sequence embedding matrix and residue contact probability matrix corresponding to the peptide sequence obtained in step (2) are input into a pre-trained universal HLA antigen presentation prediction model to obtain a presentation probability prediction value for the peptide sequence.

[0024] Specifically, the prediction result obtained in this step is the presentation probability score of the peptide segment, which reflects the possibility of the peptide segment being presented by the HLA molecule and can be used for immunogenicity screening and candidate epitope screening tasks.

[0025] like Figure 2 As shown, the universal HLA antigen presentation prediction model of the present invention includes a structural encoder based on a graph isomorphic neural network, a sequence encoder based on a one-dimensional convolutional neural network (1DCNN), a feature fusion module, and a classification prediction module.

[0026] The structural encoder based on graph isomorphic neural network consists of a graph construction layer, two graph convolution layers, two normalization layers and an average pooling layer.

[0027] The 1DCNN-based sequence encoder consists of a region convolution initialization layer, two convolution layers, and a maximum pooling layer.

[0028] The feature fusion module consists of a convolutional feature compression layer and a multimodal feature concatenation layer.

[0029] The input to the graph construction layer is the sequence embedding features corresponding to the peptide sequence, with a dimension of B × N × 1280 (where B is the batch size and N represents the length of the peptide sequence; for HLA-I peptide sequences, the peptide sequence length is N = 14, and for HLA-II peptide sequences, the peptide sequence length is N = 21) and a contact probability matrix with a dimension of B × N × N. The graph construction layer performs graph construction on the sequence embedding features and the contact probability matrix (a node in the protein structure graph represents each residue in the peptide sequence, and the node feature is the sequence embedding feature corresponding to the node. The edges are composed of residue pairs in the contact probability matrix with a contact probability greater than or equal to a preset threshold, which is 0.5 in this invention) to obtain a node feature representation of the protein structure graph processed by the graph construction layer with a dimension of L × 1280. This node feature representation is then subjected to feature transformation processing (i.e., transformation and activation through a linear layer and a BatchNorm layer) to obtain a node feature representation of the protein structure graph processed by the graph construction layer with a dimension of L × 256, where L = B × N.

[0030] The input of the first graph convolutional layer is the node feature representation of the protein structure graph with a dimension of L×256 after being processed by the graph construction layer. The graph convolutional layer uses a graph isomorphism network to process the node feature representation (the graph isomorphism network uses two linear layers, a BatchNorm layer and a ReLU activation function) to obtain a node feature representation of the protein structure graph with a dimension of L×256 after being processed by the first graph convolutional layer.

[0031] The input of the first normalization layer is the node feature representation of the protein structure graph with a dimension of L×256 after being processed by the first graph convolutional layer. The normalization layer performs PairNorm normalization on the node feature representation (the normalization method is set to first subtract the node feature mean and then uniformly scale, and the feature scaling coefficient is 1) to obtain the node feature representation of the protein structure graph with a dimension of L×256 after being processed by the first normalization layer.

[0032] The input of the second graph convolutional layer is the node feature representation of the protein structure graph with a dimension of L×256 after being processed by the first normalization layer. The graph convolutional layer uses the graph isomorphism network to convolve the node feature representation (the graph isomorphism network internally uses two linear layers, a BatchNorm layer and a ReLU activation function) to obtain the node feature representation of the protein structure graph with a dimension of L×128 after being processed by the second graph convolutional layer.

[0033] The input of the second normalization layer is the node feature representation of the protein structure graph with a dimension of L×128 after being processed by the second graph convolutional layer. The normalization layer performs PairNorm normalization on the node feature representation (the normalization method is set to first subtract the node feature mean and then uniformly scale, and the feature scaling factor is 1) to obtain the node feature representation of the protein structure graph with a dimension of L×128 after being processed by the second normalization layer.

[0034] The input of the average pooling layer is the node feature representation of the protein structure graph with a dimension of L×128 after processing by the second normalization layer. The average pooling layer performs global average pooling on the node feature representation (i.e., the features of all nodes in the protein structure graph are averaged along the dimension, and the global graph-level representation vector of each protein is output) to obtain a graph representation vector with a dimension of B×128 after processing by the average pooling layer.

[0035] The input of the regional convolution initialization layer is the sequence embedding feature of dimension B×N×1280 corresponding to the peptide sequence. The regional convolution initialization layer transposes the sequence embedding matrix to obtain a transposed sequence embedding feature of dimension B×1280×N. Then, a one-dimensional convolution operation with a convolution kernel size of 3 is performed on the transposed sequence embedding feature to output an intermediate feature vector of dimension B×256×N-2. Then, the intermediate feature vector of dimension B×256×N-2 is edge-filled (one position is filled at the beginning and end of the sequence dimension) to obtain a local area feature of dimension B×256×N after processing by the regional convolution initialization layer.

[0036] The input of the first convolutional layer is the local area feature with a dimension of B×256×N after being processed by the regional convolution initialization layer. The convolutional layer performs rectified linear unit (ReLU) activation, convolution processing and Dropout operation on the local area feature in sequence to obtain the intermediate feature vector with a dimension of B×256×N after being processed by the first convolutional layer; The input of the second convolutional layer is the intermediate feature vector of dimension B×256×N after being processed by the first convolutional layer. The convolutional layer performs ReLU activation, convolution and Dropout operations on the intermediate feature vector in sequence to obtain a feature vector of dimension B×256×N after being processed by the second convolutional layer. The input of the maximum pooling layer is the local area features with a dimension of B×256×N after being processed by the regional convolution initialization layer and the feature vector with a dimension of B×256×N after being processed by the second convolution layer. The maximum pooling layer performs residual connection processing on the local area features and the feature vector to obtain a feature vector with a dimension of B×256×N after residual processing. Then, the feature vector with a dimension of B×256×N after residual processing is subjected to one-dimensional global maximum pooling processing (the one-dimensional global maximum pooling operation refers to taking the maximum value of the eigenvalue of each channel along the sequence length dimension N, thereby compressing the residual processed feature vector into a CNN feature vector with a dimension of B×256×1) to obtain a CNN feature vector with a dimension of B×256×1 after being processed by the maximum pooling layer.

[0037] The input of the convolutional feature compression layer is the CNN feature vector with a dimension of B×256×1 after being processed by the maximum pooling layer. The convolutional feature compression layer compresses the CNN feature vector to obtain an intermediate feature vector with a dimension of B×256. Then, the intermediate feature vector is batch normalized to obtain a convolutional feature vector with a dimension of B×256.

[0038] The input of the multimodal feature splicing layer is the convolution feature vector with a dimension of B×256 output by the convolution feature compression layer, and the graph representation vector with a dimension of B×128 after processing by the average pooling layer. The multimodal feature splicing layer splices the convolution feature vector and the graph representation vector to obtain a fused representation with a dimension of B×384.

[0039] The input of the classification prediction module is the fused representation of the dimension B×384 output by the multimodal feature concatenation layer. The classification prediction module processes the fused representation using a fully connected layer (there are two fully connected layers. The first layer uses linear transformation to convert the vector of dimension B×384 into an intermediate feature vector of dimension B×256, and performs nonlinear enhancement and distribution stabilization through batch normalization and ReLU activation function; then the second layer uses linear transformation to convert the intermediate feature vector of dimension B×384 into a vector of dimension B×64) to obtain a vector of dimension B×64. Then, the vector is linearly classified (i.e., using a fully connected linear transformation, the output dimension is B×1) to obtain a presentation probability prediction value of dimension B×1.

[0040] Specifically, the universal HLA antigen presentation prediction model of the present invention is obtained by training through the following steps: (3-1) Extract a peptide sequence set consisting of verified HLA-I and HLA-II peptide sequences from the Immune Epitope Database and Analysis Resource (IEDB) as a positive sample peptide sequence set, and use the protein database to generate a negative sample peptide sequence set equal to the positive sample peptide sequence set. The positive sample peptide sequence set and the negative sample peptide sequence set together constitute a dataset, and the dataset is divided into a training set and a test set in a ratio of 4:1, where each HLA-I peptide sequence and HLA-II peptide sequence has a true label; The process for generating a negative peptide sequence set equal in size to the positive peptide sequence set in this step is as follows: First, each positive peptide sequence in the positive peptide sequence set is aligned to a non-redundant protein database using the Double Index Alignment of Next-generation Sequencing Data (DIAMOND) tool to locate the source protein to which it corresponds. Multiple peptide sequences are randomly extracted from locations within the source protein outside of known epitope regions (HLA-I: peptide sequences of 8 to 14 amino acids in length; HLA-II: peptide sequences of 13 to 21 amino acids in length). CD-HIT, a tool for sequence redundancy removal and clustering, is used to remove highly similar peptide sequences from the resulting peptide sequences (to ensure negative sample diversity), ultimately generating a negative peptide sequence set.

[0041] The advantage of this step (1) is that a peptide sequence set consisting of verified HLA-I and HLA-II peptide sequences is extracted from the IEDB database as the positive sample peptide sequence set. After excluding known epitope regions in the source protein, the DIAMOND alignment and CD-HIT clustering tools are combined to randomly generate non-epitope peptide sequences with lengths consistent with biological characteristics as the negative sample peptide sequence set. This strategy effectively ensures the diversity and biological plausibility of negative samples, alleviates the problem of uneven distribution of positive and negative samples, improves the training stability and generalization ability of the model, and is suitable for antigen screening tasks in unknown HLA backgrounds.

[0042] (3-2) For each peptide sequence in the training set obtained in step (3-1), the peptide sequence is fixed-length processed to obtain a fixed-length processed peptide sequence; then, the pre-trained evolutionary scale model (ESM) is used to obtain the sequence embedding matrix corresponding to the peptide sequence with a dimension of B×N×1280 from the fixed-length processed peptide sequence. The contact probability matrix with dimension B×N×N .

[0043] The fixed-length processing in this step specifically involves padding the ends of peptide sequences shorter than the fixed length with "1" (for HLA-I peptide sequences, the fixed amino acid sequence length is 14; for HLA-II peptide sequences, the fixed amino acid sequence length is 21) to meet the input requirements of the universal HLA antigen presentation prediction model.

[0044] The ESM model in the present invention is esm2_t33_650M_UR50D.

[0045] The advantage of this step (3-2) is that it uses a pre-trained protein language model (ESM) to embed and extract peptide sequences. This not only obtains high-dimensional semantic features rich in biological evolutionary information, but also simultaneously generates a contact probability matrix between residues, providing data support for subsequent structural modeling. By using a "1" end-padding method to unify sequence lengths, even short peptide sequences can adapt to the input requirements of deep neural networks, improving the integrity and compatibility of sequence modeling.

[0046] (3-3) For each peptide sequence in the training set obtained in step (3-1), embed the sequence corresponding to the peptide sequence obtained in step (3-2) with a dimension of B×N×1280 into features The contact probability matrix with dimension B×N×N Input the graph construction layer in the structure encoder of the universal HLA antigen presentation prediction model to perform graph construction and transformation processing on the sequence embedding features and contact probability matrix, thereby obtaining the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×256 after being processed by the graph construction layer , where L=B×N.

[0047] (3-4) For each peptide sequence in the training set obtained in step (3-1), the node feature representation of the protein structure graph corresponding to the peptide sequence obtained in step (3-3) after being processed by the graph construction layer is expressed as Input the first graph convolutional layer in the structure encoder in the universal HLA antigen presentation prediction model to represent the node feature Perform graph isomorphism network convolution processing to obtain the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×256 after the first graph convolution layer processing .

[0048] (3-5) For each peptide sequence in the training set obtained in step (3-1), the node feature representation of the protein structure graph with a dimension of L×256 and processed by the first graph convolution layer corresponding to the peptide sequence obtained in step (3-4) is Input to the first normalization layer in the structure encoder of the universal HLA antigen presentation prediction model to represent the node features of the protein structure graph PairNorm normalization is performed (where the normalization method is set to first subtract the node feature mean and then uniformly scale, with the feature scaling factor being 1) to obtain the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×256 after the first normalization layer. .

[0049] (3-6) For each peptide sequence in the training set obtained in step (3-1), the node feature representation of the protein structure graph with a dimension of L×256 and processed by the first normalization layer corresponding to the peptide sequence obtained in step (3-5) is Input the second graph convolutional layer in the structure encoder in the universal HLA antigen presentation prediction model to represent the node features of the protein structure graph Perform graph isomorphism network convolution processing to obtain the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×128 after the second graph convolution layer processing .

[0050] (3-7) For each peptide sequence in the training set obtained in step (3-1), the node feature representation of the protein structure graph with a dimension of L×128 and processed by the second graph convolution layer corresponding to the peptide sequence obtained in step (3-6) is Input the second normalization layer in the structure encoder in the universal HLA antigen presentation prediction model to represent the node features of the protein structure graph PairNorm normalization is performed (where the normalization method is set to first subtract the node feature mean and then uniformly scale, with the feature scaling factor being 1) to obtain the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×128 after the second normalization layer. .

[0051] (3-8) For each peptide sequence in the training set obtained in step (1), the node feature representation of the protein structure graph with a dimension of L×128 and processed by the second normalization layer corresponding to the peptide sequence obtained in step (3-7) is Input the average pooling layer in the structure encoder of the universal HLA antigen presentation prediction model to represent the node features of the protein structure graph Perform global average pooling to obtain the graph representation vector corresponding to the peptide sequence with a dimension of B×128 after average pooling layer processing .

[0052] The advantage of the above sub-steps (3-1) to (3-8) is that by constructing a graph structure using the sequence embedding and residue contact graph output by the protein language model, and using a multi-layer GIN graph neural network to deeply encode the peptide sequence, the spatial topology and local structural characteristics of the peptide are effectively captured, thereby enhancing the model's ability to understand the three-dimensional structural information of the peptide and improving the expressiveness and prediction accuracy of the structural representation.

[0053] (3-9) For each peptide sequence in the training set obtained in step (3-1), embed the sequence corresponding to the peptide sequence obtained in step (3-2) into a sequence embedding matrix with a dimension of B×N×1280 Input the regional convolution initialization layer in the sequence encoder of the universal HLA antigen presentation prediction model to obtain the local region features corresponding to the peptide sequence with a dimension of B×256×N after processing by the regional convolution initialization layer .

[0054] This step is specifically as follows: first, embed the sequence matrix Perform transposition to obtain a sequence embedding feature with a dimension of B×1280×N Then, the sequence is embedded into the feature Perform regional convolution and edge filling (first embed the sequence features The local area features are extracted by regional convolution initialization processing, and the intermediate feature vector with the output dimension of B×256×N-2 is ; Then the intermediate eigenvector Perform edge filling operation) to obtain the sequence embedding feature tensor corresponding to the peptide sequence with a dimension of B×256×N .

[0055] (3-10) For each peptide sequence in the training set obtained in step (3-1), the local region feature of the peptide sequence obtained in step (3-9) with a dimension of B×256×N after being processed by the regional convolution initialization layer is Input the first convolutional layer in the sequence encoder of the universal HLA antigen presentation prediction model to perform local region feature ReLU activation, convolution processing and Dropout operation are performed in sequence to obtain the feature intermediate vector corresponding to the peptide sequence with a dimension of B×256×N after the first convolution layer processing. .

[0056] (3-11) For each peptide sequence in the training set obtained in step (3-1), the intermediate feature vector corresponding to the peptide sequence obtained in step (3-10) with a dimension of B×256×N after the first convolutional layer processing is Input the second convolutional layer in the sequence encoder of the universal HLA antigen presentation prediction model to ReLU activation, convolution processing and Dropout operation are performed in sequence to obtain the feature vector corresponding to the peptide sequence with a dimension of B×256×N after the second convolution layer processing. .

[0057] (3-12) For each peptide sequence in the training set obtained in step (3-1), the local region feature of the peptide sequence obtained in step (3-9) with a dimension of B×256×N after processing by the regional convolution initialization layer is The feature vector corresponding to the peptide sequence obtained in step (3-11) with a dimension of B×256×N after processing by the second convolution layer Input the maximum pooling layer in the sequence encoder of the universal HLA antigen presentation prediction model to obtain the CNN feature vector corresponding to the peptide sequence with a dimension of B×256×1 after the maximum pooling layer processing .

[0058] Specifically, this step is as follows: first, the local area features corresponding to the peptide sequence with a dimension of B×256×N after being processed by the regional convolution initialization layer are The feature vector corresponding to the peptide sequence with a dimension of B×256×N after processing by the second convolutional layer After residual connection processing, the feature vector corresponding to the peptide sequence with a dimension of B×256×N after residual processing is obtained. Then, the feature vector corresponding to the peptide sequence with a dimension of B×256×N after residual processing is Perform one-dimensional global maximum pooling to obtain the CNN feature vector corresponding to the peptide sequence with a dimension of B×256×1 after the maximum pooling layer processing .

[0059] The advantage of the above sub-steps (3-9) to (3-12) is that by designing regional convolution and multi-layer one-dimensional convolution modules, the local context information and sequence patterns of the peptide sequence are deeply mined, and residual connections and multiple pooling operations are combined to achieve feature compression and enhancement, which effectively improves the robustness and recognition ability of sequence features and enhances the model's ability to capture and represent sequence features.

[0060] (3-13) For each peptide sequence in the training set obtained in step (3-1), the CNN feature vector corresponding to the peptide sequence obtained in step (3-12) with a dimension of B×256×1 and processed by the maximum pooling layer is Input the convolutional feature compression layer in the feature fusion module in the universal HLA antigen presentation prediction model to Compression processing is performed to obtain the intermediate feature vector corresponding to the peptide sequence with a dimension of B×256 , and the intermediate eigenvector Perform batch normalization to obtain the convolution feature vector corresponding to the peptide sequence with a dimension of B×256 .

[0061] (3-14) For each peptide sequence in the training set obtained in step (3-1), the graph representation vector corresponding to the peptide sequence obtained in step (3-8) with a dimension of B×128 after the average pooling layer is processed The convolution feature vector of dimension B×256 corresponding to the peptide sequence obtained in step (3-13) Input the multimodal feature concatenation layer in the feature fusion module of the universal HLA antigen presentation prediction model to convolute the feature vector and graph representation vector Perform splicing to obtain a fusion representation of the peptide sequence with a dimension of B×384 .

[0062] (3-15) For each peptide sequence in the training set obtained in step (3-1), the fusion representation of the dimension B×384 corresponding to the peptide sequence obtained in step (3-14) is Input the multimodal feature concatenation layer in the feature fusion module in the universal HLA antigen presentation prediction model to represent the fusion Perform full connection layer processing to obtain a vector of dimension B×64 corresponding to the peptide sequence , and the vector of dimension B×64 corresponding to the peptide sequence Perform linear classification processing to obtain the presentation probability prediction value corresponding to the peptide sequence with a dimension of B×1 .

[0063] (3-16) For each peptide sequence in the training set obtained in step (3-1), the predicted value of the presentation probability of dimension B×1 corresponding to the true label corresponding to the peptide sequence obtained in step (3-1) and the peptide sequence obtained in step (3-15) is A binary cross entropy loss function is constructed, and the universal HLA antigen presentation prediction model is iteratively optimized and trained using the back propagation algorithm until the binary cross entropy loss function converges, thereby obtaining a preliminarily trained universal HLA antigen presentation prediction model; Specifically, the loss function used in this step is: ; Where S represents the total number of peptide sequences in the training set; ∈{0,1} represents the true label of the i-th peptide sequence in the training set, where 1 indicates that the peptide sequence has presentation ability, and 0 indicates that the peptide sequence has no presentation ability; ∈{0,1} represents the probability that the universal HLA antigen presentation prediction model predicts that the i-th sample is presented.

[0064] (3-17) Using the test set obtained in step (3-), the universal HLA antigen presentation prediction model obtained in step (3-) after preliminary training is verified, and the universal HLA antigen presentation prediction model with the best performance in the verification results is selected as the final trained universal HLA antigen presentation prediction model.

[0065] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A universal HLA antigen presentation prediction method based on protein language model and multimodal neural network, characterized in that: The following steps are involved: (1) Obtaining a peptide sequence to be predicted, and performing fixed-length processing on the peptide sequence to obtain a peptide sequence after fixed-length processing; (2) Inputting the fixed-length peptide sequence obtained in step (1) into the protein language model to obtain the sequence embedding matrix and residue contact probability matrix corresponding to the peptide sequence; (3) The sequence embedding matrix and residue contact probability matrix corresponding to the peptide sequence obtained in step (2) are input into a pre-trained universal HLA antigen presentation prediction model to obtain a presentation probability prediction value for the peptide sequence.

2. The universal HLA antigen presentation prediction method based on protein language model and multimodal neural network according to claim 1, characterized in that: Step (1) is specifically as follows: specifically, the HLA-I class peptide sequence is processed into a peptide sequence with a fixed length of 14, and the HLA-II class peptide sequence is processed into a peptide sequence with a fixed length of 21. For peptide sequences shorter than the fixed length, "1" is used to fill the ends so that all peptide sequences meet the input dimension requirements of the universal HLA antigen presentation prediction model. For peptide sequences longer than the fixed length, the amino acid residues at the end of the peptide sequence are deleted until its length is equal to the fixed length.

3. The universal HLA antigen presentation prediction method based on protein language model and multimodal neural network according to claim 1 or 2, characterized in that: The universal HLA antigen presentation prediction model includes a structural encoder based on a graph isomorphic neural network, a sequence encoder based on a one-dimensional convolutional neural network, a feature fusion module, and a classification prediction module; The structural encoder based on graph isomorphism neural network consists of a graph construction layer, two graph convolution layers, two normalization layers and an average pooling layer; The 1DCNN-based sequence encoder consists of a region convolution initialization layer, two convolution layers, and a maximum pooling layer; The feature fusion module consists of a convolutional feature compression layer and a multimodal feature concatenation layer.

4. The universal HLA antigen presentation prediction method based on a protein language model and a multimodal neural network according to any one of claims 1 to 3, characterized in that: The input of the graph construction layer is the sequence embedding features of dimension B×N×1280 and the contact probability matrix of dimension B×N×N corresponding to the peptide sequence. The graph construction layer performs graph construction processing on the sequence embedding features and the contact probability matrix to obtain a node feature representation of the protein structure graph after the graph construction layer processing with dimension L×1280, and performs feature transformation processing on the node feature representation to obtain a node feature representation of the protein structure graph after the graph construction layer processing with dimension L×256, where L=B×N, B represents the batch size; N represents the length of the peptide sequence. For HLA-I class peptide sequences, the length of the peptide sequence is N=14, and for HLA-II class peptide sequences, the length of the peptide sequence is N=21; The input of the first graph convolutional layer is the node feature representation of the protein structure graph processed by the graph construction layer with a dimension of L×256. The graph convolutional layer uses the graph isomorphism network to process the node feature representation to obtain the node feature representation of the protein structure graph processed by the first graph convolutional layer with a dimension of L×256. The input of the first normalization layer is the node feature representation of the protein structure graph with a dimension of L×256 after being processed by the first graph convolutional layer. The normalization layer performs PairNorm normalization on the node feature representation to obtain the node feature representation of the protein structure graph with a dimension of L×256 after being processed by the first normalization layer.

5. The universal HLA antigen presentation prediction method based on protein language model and multimodal neural network according to claim 4, characterized in that: The input of the second graph convolutional layer is the node feature representation of the protein structure graph after processing by the first normalization layer with a dimension of L×256. The graph convolutional layer uses the graph isomorphism network to convolve the node feature representation to obtain the node feature representation of the protein structure graph after processing by the second graph convolutional layer with a dimension of L×128. The input of the second normalization layer is the node feature representation of the protein structure graph after processing by the second graph convolution layer, with a dimension of L×128. The normalization layer performs PairNorm normalization on the node feature representation (where the normalization method is set to first subtract the node feature mean and then uniformly scale, with the feature scaling factor being 1) to obtain the node feature representation of the protein structure graph after processing by the second normalization layer, with a dimension of L×128. The input of the average pooling layer is the node feature representation of the protein structure graph with a dimension of L×128 after being processed by the second normalization layer. The average pooling layer performs global average pooling on the node feature representation to obtain a graph representation vector with a dimension of B×128 after being processed by the average pooling layer.

6. The universal HLA antigen presentation prediction method based on protein language model and multimodal neural network according to claim 5, characterized in that: The input of the regional convolution initialization layer is the sequence embedding feature of dimension B×N×1280 corresponding to the peptide sequence. The regional convolution initialization layer transposes the sequence embedding matrix to obtain a transposed sequence embedding feature of dimension B×1280×N. Then, a one-dimensional convolution operation with a convolution kernel size of 3 is performed on the transposed sequence embedding feature to output an intermediate feature vector of dimension B×256×N-2. Then, the intermediate feature vector of dimension B×256×N-2 is edge-filled to obtain a local region feature of dimension B×256×N after processing by the regional convolution initialization layer. The input of the first convolutional layer is the local area feature with a dimension of B×256×N after being processed by the regional convolution initialization layer. The convolutional layer performs rectified linear unit (ReLU) activation, convolution processing, and Dropout operations on the local area feature to obtain an intermediate feature vector with a dimension of B×256×N after being processed by the first convolutional layer. The input of the second convolutional layer is the intermediate feature vector of dimension B×256×N after being processed by the first convolutional layer. The convolutional layer performs ReLU activation, convolution and Dropout operations on the intermediate feature vector in sequence to obtain a feature vector of dimension B×256×N after being processed by the second convolutional layer. The input of the maximum pooling layer is the local area features with a dimension of B×256×N after being processed by the regional convolution initialization layer and the feature vector with a dimension of B×256×N after being processed by the second convolution layer. The maximum pooling layer performs residual connection processing on the local area features and the feature vector to obtain a feature vector with a dimension of B×256×N after residual processing. Then, the feature vector with a dimension of B×256×N after residual processing is subjected to one-dimensional global maximum pooling processing to obtain a CNN feature vector with a dimension of B×256×1 after being processed by the maximum pooling layer; The input of the convolutional feature compression layer is the CNN feature vector with a dimension of B×256×1 after being processed by the maximum pooling layer. The convolutional feature compression layer compresses the CNN feature vector to obtain an intermediate feature vector with a dimension of B×256. Then, the intermediate feature vector is batch normalized to obtain a convolutional feature vector with a dimension of B×256. The input of the multimodal feature splicing layer is the convolution feature vector with a dimension of B×256 output by the convolution feature compression layer, and the graph representation vector with a dimension of B×128 after processing by the average pooling layer. The multimodal feature splicing layer splices the convolution feature vector and the graph representation vector to obtain a fused representation with a dimension of B×384.

7. The universal HLA antigen presentation prediction method based on protein language model and multimodal neural network according to claim 6, characterized in that: The input of the classification prediction module is the fused representation of the output of the multimodal feature concatenation layer with a dimension of B×384. The classification prediction module uses the fully connected layer to process the fused representation to obtain a vector of dimension B×64. Then, the vector is linearly classified to obtain a presentation probability prediction value with a dimension of B×1.

8. The universal HLA antigen presentation prediction method based on protein language model and multimodal neural network according to claim 7, characterized in that: The universal HLA antigen presentation prediction model is trained through the following steps: (3-1) Extract a peptide sequence set consisting of verified HLA-I and HLA-II peptide sequences from the immune epitope database IEDB as a positive sample peptide sequence set, and use the protein database to generate a negative sample peptide sequence set equal to the positive sample peptide sequence set. The positive sample peptide sequence set and the negative sample peptide sequence set together constitute a dataset, and the dataset is divided into a training set and a test set in a ratio of 4:1, where each HLA-I peptide sequence and HLA-II peptide sequence has a true label; (3-2) For each peptide sequence in the training set obtained in step (3-1), the peptide sequence is fixed-length processed to obtain a fixed-length processed peptide sequence; then, the pre-trained evolutionary scaling model ESM is used to obtain the sequence embedding matrix corresponding to the peptide sequence with a dimension of B×N×1280 from the fixed-length processed peptide sequence. The contact probability matrix with dimension B×N×N ; (3-3) For each peptide sequence in the training set obtained in step (3-1), embed the sequence corresponding to the peptide sequence obtained in step (3-2) with a dimension of B×N×1280 into features The contact probability matrix with dimension B×N×N Input the graph construction layer in the structure encoder of the universal HLA antigen presentation prediction model to perform graph construction and transformation processing on the sequence embedding features and contact probability matrix, thereby obtaining the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×256 after being processed by the graph construction layer , where L = B × N; (3-4) For each peptide sequence in the training set obtained in step (3-1), the node feature representation of the protein structure graph corresponding to the peptide sequence obtained in step (3-3) after being processed by the graph construction layer is expressed as Input the first graph convolutional layer in the structure encoder in the universal HLA antigen presentation prediction model to represent the node feature Perform graph isomorphism network convolution processing to obtain the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×256 after the first graph convolution layer processing ; (3-5) For each peptide sequence in the training set obtained in step (3-1), the node feature representation of the protein structure graph with a dimension of L×256 and processed by the first graph convolution layer corresponding to the peptide sequence obtained in step (3-4) is Input to the first normalization layer in the structure encoder of the universal HLA antigen presentation prediction model to represent the node features of the protein structure graph PairNorm normalization is performed (where the normalization method is set to first subtract the node feature mean and then uniformly scale, with the feature scaling factor being 1) to obtain the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×256 after the first normalization layer. ; (3-6) For each peptide sequence in the training set obtained in step (3-1), the node feature representation of the protein structure graph with a dimension of L×256 and processed by the first normalization layer corresponding to the peptide sequence obtained in step (3-5) is Input the second graph convolutional layer in the structure encoder in the universal HLA antigen presentation prediction model to represent the node features of the protein structure graph Perform graph isomorphism network convolution processing to obtain the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×128 after the second graph convolution layer processing ; (3-7) For each peptide sequence in the training set obtained in step (3-1), the node feature representation of the protein structure graph with a dimension of L×128 and processed by the second graph convolution layer corresponding to the peptide sequence obtained in step (3-6) is Input the second normalization layer in the structure encoder in the universal HLA antigen presentation prediction model to represent the node features of the protein structure graph PairNorm normalization is performed (where the normalization method is set to first subtract the node feature mean and then uniformly scale, with the feature scaling factor being 1) to obtain the node feature representation of the protein structure graph corresponding to the peptide sequence with a dimension of L×128 after the second normalization layer. ; (3-8) For each peptide sequence in the training set obtained in step (3-1), the node feature representation of the protein structure graph with a dimension of L×128 and processed by the second normalization layer corresponding to the peptide sequence obtained in step (3-7) is Input the average pooling layer in the structure encoder of the universal HLA antigen presentation prediction model to represent the node features of the protein structure graph Perform global average pooling to obtain the graph representation vector corresponding to the peptide sequence with a dimension of B×128 after average pooling layer processing ; (3-9) For each peptide sequence in the training set obtained in step (3-1), embed the sequence corresponding to the peptide sequence obtained in step (3-2) into a sequence embedding matrix with a dimension of B×N×1280 Input the regional convolution initialization layer in the sequence encoder of the universal HLA antigen presentation prediction model to obtain the local region features corresponding to the peptide sequence with a dimension of B×256×N after processing by the regional convolution initialization layer ; (3-10) For each peptide sequence in the training set obtained in step (3-1), the local region feature of the peptide sequence obtained in step (3-9) with a dimension of B×256×N after being processed by the regional convolution initialization layer is Input the first convolutional layer in the sequence encoder of the universal HLA antigen presentation prediction model to perform local region feature ReLU activation, convolution processing and Dropout operation are performed in sequence to obtain the feature intermediate vector corresponding to the peptide sequence with a dimension of B×256×N after the first convolution layer processing. ; (3-11) For each peptide sequence in the training set obtained in step (3-1), the intermediate feature vector corresponding to the peptide sequence obtained in step (3-10) with a dimension of B×256×N after the first convolutional layer processing is Input the second convolutional layer in the sequence encoder of the universal HLA antigen presentation prediction model to ReLU activation, convolution processing and Dropout operation are performed in sequence to obtain the feature vector corresponding to the peptide sequence with a dimension of B×256×N after the second convolution layer processing. ; (3-12) For each peptide sequence in the training set obtained in step (3-1), the local region feature of the peptide sequence obtained in step (3-9) with a dimension of B×256×N after processing by the regional convolution initialization layer is The feature vector corresponding to the peptide sequence obtained in step (3-11) with a dimension of B×256×N after processing by the second convolution layer Input the maximum pooling layer in the sequence encoder of the universal HLA antigen presentation prediction model to obtain the CNN feature vector corresponding to the peptide sequence with a dimension of B×256×1 after the maximum pooling layer processing ; (3-13) For each peptide sequence in the training set obtained in step (3-1), the CNN feature vector corresponding to the peptide sequence obtained in step (3-12) with a dimension of B×256×1 and processed by the maximum pooling layer is Input the convolutional feature compression layer in the feature fusion module in the universal HLA antigen presentation prediction model to Compression processing is performed to obtain the intermediate feature vector corresponding to the peptide sequence with a dimension of B×256 , and the intermediate eigenvector Perform batch normalization to obtain the convolution feature vector corresponding to the peptide sequence with a dimension of B×256 ; (3-14) For each peptide sequence in the training set obtained in step (3-1), the graph representation vector corresponding to the peptide sequence obtained in step (3-8) with a dimension of B×128 after the average pooling layer is processed The convolution feature vector of dimension B×256 corresponding to the peptide sequence obtained in step (3-13) Input the multimodal feature concatenation layer in the feature fusion module of the universal HLA antigen presentation prediction model to convolute the feature vector and graph representation vector Perform splicing to obtain a fusion representation of the peptide sequence with a dimension of B×384 ; (3-15) For each peptide sequence in the training set obtained in step (3-1), the fusion representation of the dimension B×384 corresponding to the peptide sequence obtained in step (3-14) is Input the multimodal feature concatenation layer in the feature fusion module in the universal HLA antigen presentation prediction model to represent the fusion Perform full connection layer processing to obtain a vector of dimension B×64 corresponding to the peptide sequence , and the vector of dimension B×64 corresponding to the peptide sequence Perform linear classification processing to obtain the presentation probability prediction value corresponding to the peptide sequence with a dimension of B×1 ; (3-16) For each peptide sequence in the training set obtained in step (3-1), the predicted value of the presentation probability of dimension B×1 corresponding to the true label corresponding to the peptide sequence obtained in step (3-1) and the peptide sequence obtained in step (3-15) is A binary cross entropy loss function is constructed, and the universal HLA antigen presentation prediction model is iteratively optimized and trained using the back propagation algorithm until the binary cross entropy loss function converges, thereby obtaining a preliminarily trained universal HLA antigen presentation prediction model; the loss function used in this step is: ； Where S represents the total number of peptide sequences in the training set; ∈{0,1} represents the true label of the i-th peptide sequence in the training set, where 1 indicates that the peptide sequence has presentation ability, and 0 indicates that the peptide sequence has no presentation ability; ∈{0,1} represents the probability that the universal HLA antigen presentation prediction model predicts that the i-th sample is presented; (3-17) Using the test set obtained in step (3-1), the preliminarily trained universal HLA antigen presentation prediction model obtained in step (3-16) is verified, and the universal HLA antigen presentation prediction model with the best performance in the verification results is selected as the final trained universal HLA antigen presentation prediction model.

9. The universal HLA antigen presentation prediction method based on protein language model and multimodal neural network according to claim 8, characterized in that: The specific process of generating a negative sample peptide sequence set equal in number to the positive sample peptide sequence set in step (3-1) is as follows: first, each positive sample peptide sequence in the positive sample peptide sequence set is aligned to a non-redundant protein database using the dual-index next-generation sequencing data alignment tool DIAMOND to locate the source protein corresponding to the positive sample peptide sequence; and then a plurality of peptide sequences are randomly intercepted from other positions in the source protein except for the known epitope region; CD-HIT, a tool for removing sequence redundancy and clustering, was used to remove highly similar peptide sequences from the obtained multiple peptide sequences, ultimately generating a negative sample peptide sequence set. The fixed-length processing in step (3-2) specifically involves padding the ends of peptide sequences shorter than the fixed length with "1". That is, for HLA-I peptide sequences, the fixed amino acid sequence length is 14; for HLA-II peptide sequences, the fixed amino acid sequence length is 21, to meet the input requirements of the universal HLA antigen presentation prediction model. Steps (3-9) are as follows: First, embed the sequence matrix Perform transposition to obtain a sequence embedding feature with a dimension of B×1280×N , then, embed the sequence into features Perform regional convolution and edge filling to obtain the sequence embedding feature tensor corresponding to the peptide sequence with a dimension of B×256×N ; Steps (3-12) are as follows: first, the local area features of the peptide sequence with a dimension of B×256×N after the regional convolution initialization layer processing are processed. The feature vector corresponding to the peptide sequence with a dimension of B×256×N after processing by the second convolutional layer After residual connection processing, the feature vector corresponding to the peptide sequence with a dimension of B×256×N after residual processing is obtained. Then, the feature vector corresponding to the peptide sequence with a dimension of B×256×N after residual processing is Perform one-dimensional global maximum pooling to obtain the CNN feature vector corresponding to the peptide sequence with a dimension of B×256×1 after the maximum pooling layer processing .

10. A universal HLA antigen presentation prediction method based on protein language model and multimodal neural network, characterized in that: The following steps are involved: The first module is used to obtain a peptide sequence to be predicted and perform fixed-length processing on the peptide sequence to obtain a peptide sequence after fixed-length processing; The second module is used to input the peptide sequence obtained by the first module after fixed-length processing into the protein language model to obtain the sequence embedding matrix and residue contact probability matrix corresponding to the peptide sequence; The third module is used to input the pre-trained universal HLA antigen presentation prediction model into the sequence embedding matrix and residue contact probability matrix corresponding to the peptide sequence obtained in the second module to obtain the presentation probability prediction value of the peptide sequence.

Images