An Influenza Virus Antigen Prediction Method Based on ESMFold Protein Structure Prediction and Geometric Graph Learning

Abstract

This invention discloses an influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning, belonging to the field of bioinformatics. The method first uses ESMFold to predict the three-dimensional structure of the influenza virus hemagglutinin subunit 1 protein, constructs a protein graph based on the spatial adjacency relationships between amino acid residues, and extracts the protein's spatial geometric features as initial representations of nodes and edges. Subsequently, a Siamese network architecture is employed, incorporating a geometric graph learning algorithm and fusing key functional site information of the protein to perform characterization learning on node and edge features, obtaining a viral protein characterization that can simultaneously capture short-range and long-range dependencies between residues and comprehensively characterize residue interactions. Finally, the characterizations of two viruses output from the Siamese network are concatenated, input into a Transformer Encoder to obtain a fused characterization, and combined with a multilayer perceptron to achieve accurate prediction of antigen distance between virus strains.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, specifically to an influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning. Background Technology

[0002] Influenza A virus causes a huge morbidity and mortality burden globally. Effective control of seasonal influenza relies heavily on timely vaccination, and the protective efficacy of vaccines depends on accurate antigen matching. Therefore, there is an urgent need for accurate and rapid methods to assess antigenicity. Although traditional serological testing methods such as hemagglutinin inhibition assays are reliable, they are inherently labor-intensive, time-consuming, and have low throughput. In contrast, antigenic drift is mainly driven by mutations within the hemagglutinin subunit 1 (HA1) protein. Coupled with the accumulation of massive amounts of high-throughput sequencing data, these factors have provided support for the development of sequence-based antigen prediction models. Currently, several studies have attempted to construct antigenicity prediction models based on machine learning and deep learning algorithms. Early studies mostly used traditional machine learning algorithms to construct antigen prediction models based on the physicochemical properties of the influenza virus HA1 protein sequence. For example, those skilled in the art have constructed antigen prediction models based on Naive Bayes and have also used XGBoost regression models to estimate antigen distance. In recent years, the rapid development of deep learning technology has provided new ideas for influenza virus antigen prediction. Multiple studies have used entries from the AAindex database to encode viral sequences to characterize the physicochemical and biochemical properties of amino acids, and employed convolutional neural networks to enhance feature expression capabilities. Although current predictive models are practical, their antigen prediction performance is limited because they fail to incorporate protein structure data and thus cannot capture the spatial geometric features of proteins.

[0003] Given the current limitations of research, it is necessary to provide an efficient and accurate method for predicting influenza virus antigens based on protein structure. Summary of the Invention

[0004] To address the shortcomings of the aforementioned background technologies, this invention provides an influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning. This method first uses ESMFold to predict the three-dimensional structure of influenza virus proteins. Then, it constructs protein maps based on the three-dimensional viral protein structures. The constructed protein maps of virus pairs are input into a geometric graph learning model based on a Siamese network architecture to extract structure-aware viral protein representations for antigen prediction.

[0005] The first objective of this invention is to provide a method for predicting influenza virus antigens based on ESMFold protein structure prediction and geometric graph learning, comprising: The three-dimensional structure of the influenza virus hemagglutinin subunit 1 protein was obtained by using ESMFold prediction based on the influenza virus protein sequence. A protein graph was constructed based on the spatial adjacency relationships between residues in the three-dimensional structure of the influenza virus hemagglutinin subunit 1 protein. Spatial geometric features of the protein were extracted from the three-dimensional structure of the viral protein, and a local coordinate system was constructed to characterize the spatial constraint relationships between residues, thus obtaining the initial node features and edge features of the protein graph. The protein graphs of paired viruses are input into the geometric graph learning model in the Siamese network architecture. The node features and edge features are updated using a multi-scale attention mechanism to obtain residue representations that capture both long-range and short-range dependencies between residues. The residue representation vectors are then fused with embedding vectors based on key functional site information of proteins. Subsequently, the fused representations are input into a multilayer perceptron to obtain the representations of viral proteins. The protein characteristics of the two viruses are input into the Transformer Encoder to extract the nonlinear interaction features between them, and the antigen prediction results are output through a multilayer perceptron.

[0006] In one embodiment, the protein graph has nodes representing residues and edges representing spatial proximity, i.e., when the nodes are adjacent to each other... When the spatial distance between atoms is less than a certain threshold, an edge will be constructed between the two nodes.

[0007] In one embodiment, constructing a local coordinate system to characterize the spatial constraints between residues includes: A local coordinate system is constructed based on the main chain atoms of each residue; Based on this local coordinate system, the geometric features of the protein are extracted as the initial features of nodes and edges; wherein, the initial features of nodes include angle features, first distance features, and first direction features; the initial features of edges include relative orientation features, second distance features, and second direction features.

[0008] In one embodiment, the sine and cosine values ​​of the dihedral angles and bond angles of the protein backbone are extracted as angular features; radial basis functions are used to encode the spatial distances between atoms within residues to obtain the first distance feature; the distance between atoms in the residues relative to each other is calculated. The direction vector of the atom is obtained, and the direction vector is projected onto the local coordinate system to obtain the first directional feature.

[0009] In one embodiment, a relative rotation matrix between residue pairs is calculated using a quaternion function to characterize the relative orientation features; a second distance feature is obtained by encoding the atomic distance between residues using a radial basis function; and a second orientation feature is obtained by calculating the relative spatial position between residues relative to the local coordinate system.

[0010] In one embodiment, updating node features includes: The node update in the geometric graph learning model employs a multi-head attention mechanism, in the... Layers, nodes Features and nodes arrive The edge features are respectively represented as and Both were projected onto After creating a latent space, node features are updated as follows: In the formula, Represents a node Neighbors The attention weights are defined as follows: In the formula, , and Let represent the learnable weight matrices that map node features to the query space, key space, and value space, respectively. This represents a learnable weight matrix used to effectively incorporate geometric information encoded in the edges into the message aggregation process.

[0011] In one embodiment, the edge features are updated using information from neighboring nodes:

[0012] In the formula, This is a hyperparameter representing the importance of the source node, which is true if and only if the source node is located in an important region. ; This represents a multilayer perceptron used for edge updates; This indicates a join operation.

[0013] In one embodiment, the residue characterization process, which simultaneously captures both long-range and short-range dependencies between residues, also includes adaptively recalibrating nodes based on the global context using a gating mechanism. When recalibrating nodes, a global context vector is obtained by averaging the features of all nodes, and a gating mechanism is used to adaptively update the node representations based on the global context vector:

[0014] In the formula, This indicates the number of nodes in the protein graph; Represents element-wise multiplication; This represents the sigmoid activation function; This represents a gated multilayer perceptron used for adaptive control of the global context; Indicates the first The global context vector corresponding to the layer; Represents a node In the The intermediate representation before the layer gating update; Represents a node In the The final representation of the layer.

[0015] In one embodiment, the residue representation vector is fused with the embedding vector of protein key functional sites. The embedding vector based on the protein key functional sites refers to the learned embedding obtained by setting different initial seeds according to the importance category corresponding to the location of the residue and training. The importance category includes 10 categories, namely glycosylation site, 130-loop, 190-helix, 220-loop, antigenic epitope A, antigenic epitope B, antigenic epitope C, antigenic epitope D, antigenic epitope E, and background residues. The fusion method is to add the two together.

[0016] The second objective of this invention is to provide a system for predicting influenza virus antigens based on ESMFold protein structure prediction and geometric graph learning, comprising: The protein structure module is used to obtain the three-dimensional structure of the influenza virus hemagglutinin subunit 1 protein based on the influenza virus protein sequence using ESMFold prediction. The initial feature module is used to construct a protein graph based on the spatial adjacency relationship between residues in the three-dimensional structure of viral proteins, extract the spatial geometric structure features of proteins from the three-dimensional structure of viral proteins, construct a local coordinate system to characterize the spatial constraint relationship between residues, and obtain the initial node features and edge features of the protein graph. The protein characterization module is used to input the protein graphs of paired viruses into the geometric graph learning model in the Siamese network architecture. It uses a multi-scale attention mechanism to update the node features and edge features, thereby obtaining residue characterization that captures both long-range and short-range dependencies between residues. The residue characterization vector is then fused with the embedding vector based on the key functional site information of the protein. Subsequently, the fused characterization is input into a multilayer perceptron to obtain the characterization of the viral protein. The antigen prediction module is used to input the protein characteristics of the two viruses into the Transformer Encoder to extract the nonlinear interaction features between them, and output the antigen prediction results through a multilayer perceptron.

[0017] Compared with existing technologies, this invention provides an influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning, which has the following significant advantages: This invention addresses the problem of scarce influenza virus structural data and the limitation of antigen prediction performance due to the inability to capture key spatial geometric features such as residue conformation and inter-residue spacing using only sequence information. This method efficiently predicts the structure of the HA1 protein using a protein language model and explicitly captures key geometric features such as residue conformation and inter-residue spacing through geometric graph learning. By deeply mining the complex long- and short-range spatial dependencies between residues and organically integrating prior biological knowledge, it significantly enhances the biological semantic expression ability of protein characterization. While improving the accuracy and robustness of antigen prediction, it also possesses high interpretability and can accurately locate key sites driving antigen drift, providing scientific and reliable decision support for influenza variant monitoring and vaccine strain screening. Attached Figure Description

[0018] Figure 1 A framework diagram of the method of this invention; Figure 2 (a) Performance comparison of the method of the present invention with five baseline models (SVM, MFPAD, AdaBoost, CNN and GCN) in predicting antigen distance on H3 and H5 subtype data under cross-validation settings; Figure 2 (b) Performance comparison of the method of the present invention with eight baseline models (SVM, MFPAD, AdaBoost, PREDAC, CNN, CNN-PSO, PREDAC-CNN and GCN) in distinguishing antigen variants from non-variants on H3 and H5 subtype data under cross-validation settings; Figure 2 (c, d) show the performance comparison of the method of the present invention with three different feature encoding methods (AAindex, ESM-2, ProtTrans) on H3 and H5 subtype data under cross-validation settings; Figure 2 (eh) represents the performance of the method of the present invention when trained on a subset of data containing only 30%–90% randomly selected viruses or only 30%–90% randomly selected antigen-antiserum pairs; Figure 3 (a, b) show the performance comparison of the method of the present invention with five baseline models (SVM, MFPAD, AdaBoost, CNN and GCN) in predicting antigen distance on H3 and H5 subtype data under a retrospective test setting; Figure 3 (c, d) show the performance comparison of the method of the present invention with eight baseline models (SVM, MFPAD, AdaBoost, PREDAC, CNN, CNN-PSO, PREDAC-CNN and GCN) in distinguishing antigenic variants from non-variant strains on H3 and H5 subtype data in a retrospective setting. Figure 3 (e, f) represent the mean absolute error (MAE) for each year of the present invention, evaluated on the H3 and H5 subtypes over time spans of 13 and 7 years, respectively. Figure 3 (g, h) are antigen maps generated using FluNexus, showing the antigenic drift of the circulating virus in the current testing year relative to viruses collected in previous years for subtypes H3 and H5, respectively. For each subtype, the antigen map on the left shows large antigenic drifts, while the antigen map on the right shows small antigenic drifts. Figure 4 (a, b) are line graphs showing the importance scores of each amino acid site on the HA1 protein of H3 and H5 subtype influenza viruses inferred by the method of this invention. The 30 sites with the highest importance scores are marked with bold dots. Sites not located in any known functional regions are classified as "unknown"; Figure 4 (c, d) are bar charts showing the importance scores of the 30 most important amino acid sites on the HA1 protein of H3 and H5 subtype influenza viruses as inferred by the method of this invention. Figure 4 (e, f) represents the enrichment analysis of the specific functional regions containing the 30 most important amino acid sites on the HA1 protein of H3 and H5 subtype influenza viruses inferred by the method of this invention. Statistical significance is quantified by the negative logarithm of the P-value. Figure 4 (g, h) are schematic diagrams showing the mapping of the 30 most important amino acid sites on the HA1 protein of H3 and H5 subtype influenza viruses, deduced by the method of the present invention, onto the HA1 protein structures of H3 (PDB ID: 4KVN; A / Perth / 16 / 2009) and H5 (PDB ID: 4K62; A / Indonesia / 5 / 2005). Figure 5 The method of this invention, the empirical coverage of vaccines recommended by the WHO in the Northern Hemisphere, Southern Hemisphere seasons and "theoretical upper limit" (2012-2024). Detailed Implementation

[0019] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be further described below in conjunction with specific embodiments and accompanying drawings. However, the embodiments described are not intended to limit the present invention.

[0020] The purpose of this invention is to overcome the problem that existing antigen prediction models fail to incorporate protein structure data and thus cannot capture the spatial geometric features of proteins, thereby limiting antigen prediction performance. This invention provides an influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning. This method first uses ESMFold to predict the three-dimensional structure of the influenza virus hemagglutinin subunit 1 protein, constructs a protein graph based on the spatial adjacency relationships between amino acid residues, and extracts the protein's spatial geometric features as initial representations of nodes and edges. Subsequently, a Siamese network architecture is adopted, incorporating a geometric graph learning algorithm and fusing information from key functional sites of the protein to perform characterization learning on node and edge features, obtaining a viral protein characterization that can simultaneously capture short-range and long-range dependencies between residues and comprehensively characterize residue interactions. Finally, the two virus characterizations output from the Siamese network are concatenated and input into a Transformer Encoder to obtain a fused characterization, which is then combined with a multilayer perceptron to achieve accurate prediction of antigen distance between virus strains.

[0021] To achieve the above objectives, an influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning is proposed, comprising: S1. Based on the influenza virus protein sequence, the three-dimensional structure of the influenza virus hemagglutinin subunit 1 (HA1) protein was obtained by ESMFold prediction. The three-dimensional structure includes each atom in each amino acid residue of the viral HA1 protein (including...). , , and The three-dimensional coordinate information of the residues and the spatial distance relationship between them determined by it.

[0022] S2. Based on the spatial adjacency relationship between residues in the three-dimensional structure of influenza virus hemagglutinin subunit 1 protein, a protein graph is constructed, and the spatial geometric structure features of the protein are extracted from the three-dimensional structure of the viral protein. A local coordinate system is constructed to characterize the spatial constraint relationship between residues, and the initial node features and edge features of the protein graph are obtained. In one embodiment, based on the 3D protein structure predicted by ESMFold, each viral protein is modeled as a protein graph, where nodes represent residues and edges represent spatial proximity, i.e., when the distance between nodes is equal to the distance between them... When the spatial distance between atoms is less than a certain threshold, an edge will be formed between the two nodes. This threshold is the effective range for characterizing local spatial contact or interaction between residues. That is, when the distance between residues is within this range, their spatial conformation, non-covalent interaction, and local structural coupling effect are more likely to influence each other, and thus can be regarded as a potential structural association in the three-dimensional structure of the protein. When the distance exceeds this threshold, the direct spatial interaction between residues is weak, and edge connection may not be established.

[0023] When constructing a local coordinate system to characterize the spatial constraints between residues, the following steps are included: A local coordinate system is constructed based on the main chain atoms of each residue; Based on this local coordinate system, the geometric features of the protein are extracted as the initial features of nodes and edges; wherein, the initial features of nodes include angle features, first distance features, and first direction features; the initial features of edges include relative orientation features, second distance features, and second direction features.

[0024] The sine and cosine values ​​of the dihedral angles and bond angles of the protein backbone are extracted as angular features; the spatial distance between atoms within residues is encoded using radial basis functions to obtain the first distance feature; the distance between atoms in the residues relative to each other is calculated. The direction vector of the atom is obtained, and the direction vector is projected onto the local coordinate system to obtain the first directional feature.

[0025] The relative rotation matrix between residue pairs is calculated using quaternion functions to characterize the relative orientation features; the atomic distance between residues is encoded using radial basis functions to obtain the second distance feature; and the relative spatial position between residues with respect to the local coordinate system is calculated to obtain the second orientation feature.

[0026] S3. Input the protein graphs of paired viruses into the geometric graph learning model in the Siamese network architecture, use the multi-scale attention mechanism to update the node features and edge features, obtain the residue representation that captures both long-range and short-range dependencies between residues, and fuse the residue representation vector with the embedding vector based on the key functional site information of the protein. Then, input the fused representation into the multilayer perceptron to obtain the representation of the viral protein. The updating of node features includes: The node update in the geometric graph learning model employs a multi-head attention mechanism, in the... Layers, nodes Features and nodes arrive The edge features are respectively represented as and Both were projected onto After creating a latent space, node features are updated as follows: In the formula, Represents a node Neighbors The attention weights are defined as follows: In the formula, , and Let represent the learnable weight matrices that map node features to the query space, key space, and value space, respectively. This represents a learnable weight matrix used to effectively incorporate geometric information encoded in the edges into the message aggregation process.

[0027] The edge features are updated using information from neighboring nodes:

[0028] In the formula, This is a hyperparameter representing the importance of the source node, which is true if and only if the source node is located in an important region. ; This represents a multilayer perceptron used for edge updates; This indicates a join operation.

[0029] It should be noted that for the global context attention module in geometric graph learning, the global context vector is obtained by averaging the features of all nodes, and the node representation is adaptively updated based on the global context vector using a gating mechanism.

[0030] The process of characterizing residues that simultaneously capture both long-range and short-range dependencies between residues also includes adaptively recalibrating nodes based on the global context using a gating mechanism. When recalibrating nodes, a global context vector is obtained by averaging the features of all nodes, and a gating mechanism is used to adaptively update the node representations based on the global context vector:

[0031] In the formula, This indicates the number of nodes in the protein graph; Represents element-wise multiplication; This represents the sigmoid activation function; This represents a gated multilayer perceptron used for adaptive control of the global context; Indicates the first The global context vector corresponding to the layer; Represents a node In the The intermediate representation before the layer gating update; Represents a node In the The final representation of the layer.

[0032] The residue representation vector is fused with the embedding vector of protein key functional sites. The embedding vector based on the protein key functional sites refers to the learnable embedding obtained by setting different initial seeds according to the importance category corresponding to the location of the residue and training. The importance category includes 10 categories, namely glycosylation site, 130-loop, 190-helix, 220-loop, antigenic epitope A, antigenic epitope B, antigenic epitope C, antigenic epitope D, antigenic epitope E, and background residues. The fusion method is to add the two together.

[0033] S4. Input the protein characteristics of the two viruses into the Transformer Encoder to extract the nonlinear interaction features between them, and output the antigen prediction results through a multilayer perceptron.

[0034] In this invention, learnable functional position embeddings based on specific residue locations are introduced to enhance the geometric embeddings of residues learned through geometric graphs. These position embeddings capture the functional significance of residues, which are categorized into 10 different functional classes based on their location, including glycosylation sites, receptor binding sites (including 130-loop, 190-helix, and 220-loop), antigenic epitopes (AEs), and background residues. The updated residue node features are then aggregated using a multilayer perceptron to obtain a global protein representation vector. Finally, to predict the antigenic distance between strains, the protein representation vectors of paired viruses are concatenated and input into a Transformer Encoder to represent the nonlinear interaction between them, and the antigenic distance between strains is predicted using a multilayer perceptron.

[0035] To illustrate the influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning provided by this invention, a specific example is given.

[0036] See Figure 1 As shown, an influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning includes: Step 1: Collection and preprocessing of influenza virus hemagglutinin protein sequence and hemagglutination inhibition titer data: The influenza virus hemagglutinin protein sequences for influenza H3 and H5 subtypes were obtained from the GISAID database and the study by Smith et al. (https: / / github.com / acorg / Racmacs / blob / master / inst / extdata), and the HA sequence of the H5 subtype from within the BSL-3 laboratory was also included. Using the default settings of the FluNexus platform (https: / / flunexus.com / ), this embodiment performed sequence alignment and preprocessing for each subtype to extract the refined influenza virus hemagglutinin protein HA1 sequence. This subunit constitutes the globular head of the HA protein and contains key epitopes that determine the antigenic properties of H3 and H5 viruses.

[0037] For hemagglutination inhibition titers, this embodiment obtained hemagglutination inhibition detection data for the H3 subtype from studies conducted by the Global Influenza Centre at the Francis Crick Institute in London (https: / / www.crick.ac.uk / partnerships / worldwideinfluenza-centre / annual-and-interim-reports) and Smith et al. (https: / / github.com / acorg / Racmacs / blob / master / inst / extdata), while the H5 subtype hemagglutination inhibition titer data came from BSL-3 laboratories and were archived in the FluNexus repository (https: / / github.com / xingyili / FluNexus-methodbox / tree / main / H5_data). Invalid hemagglutination inhibition titers marked with '<' were excluded, while titers marked with '>512' were treated as double their value. To quantify the hemagglutination inhibition titer data, this embodiment used normalized antigen distance to calculate the antigen-antiserum pair antigen distance:

[0038] in Indicates antiserum Homologous titers, Indicates antigen Antiserum Heterologous titers. Viral pairs lacking homologous titers or corresponding HA1 sequences will be eliminated. Viral pairs with a normalized antigenic distance ≥2 are classified as antigenic variants; otherwise, they are classified as antigenic similarities.

[0039] Step 2: Predict the three-dimensional structure of influenza virus proteins using ESMFold: ESMFold is a protein language model designed to predict protein structures directly from amino acid sequences. Compared to other state-of-the-art methods, ESMFold achieves a good balance between lightweight architecture, prediction accuracy, and computational efficiency, enabling fast and scalable structure prediction on large-scale datasets. This embodiment utilizes ESMFold to efficiently generate the HA1 protein structure and feeds these structures as input to a geometric deep learning network.

[0040] Step 3: Construction of structure-aware protein maps: In this embodiment, each protein is modeled as a protein map. Here, nodes represent residues, and edges represent spatial proximity. Specifically, if a node... and of The Euclidean distance between atoms is less than a predefined threshold. For the H3 subtype, Set to 8 Å, H5 subtype If we set it to 6 Å, then there is an edge between the two nodes.

[0041] To extract geometric information contained in protein structure, based on residues The main chain atoms construct a local coordinate system. Given residues... coordinates Main chain orientation vector and and local coordinate system The definition is as follows:

[0042]

[0043] in, , and These represent the main chain direction, normal vector, and secondary tangent vector of the residue, respectively.

[0044] Based on this local coordinate system, the node and edge features are defined as follows: (i) Node Features: Node features integrate multiple geometric descriptors to comprehensively characterize the structural properties of each residue, including angle, distance, and orientation features. Dihedral angles were used to capture the geometric information of the main chain conformation. and key angle Furthermore, their sine and cosine values ​​are encoded as angular features. In addition, to capture spatial information within residues, the atomic spacing is encoded using radial basis functions. ,in ,and Represents residues atoms in, and Furthermore, in order to characterize the spatial orientation within the local coordinate system, each atom... Compared to The direction is normalized and projected onto local coordinates to serve as the direction feature. .

[0045] (ii) Edge features: Similar to node features, edge features include orientation, distance, and direction features. For each pair of residues... and Their relative orientation is determined by the rotation matrix. express:

[0046] in, Let denot a quaternion function, and let the distance feature be defined as follows: ,in ,and Representing nodes respectively and The atoms in the [structure / structure]. Furthermore, directional features capture directional information while maintaining rotational isotropy; they encode residues. In residues The relative spatial position in the local coordinate system. For each pair of residues The calculation method is as follows .

[0047] Step 4: Obtaining Virus Pair Representations Based on Siamese Network Architecture and Geometric Graph Learning: To capture the spatial geometric features of protein structures, the protein maps of the constructed virus pairs were input into a graph neural network layer based on a Siamese network architecture for geometric graph learning. Simultaneously, to model multi-scale inter-residue interactions, attention modules were employed at different levels (including node, edge, and global context levels).

[0048] (i) Node Update. Inspired by the Transformer, a multi-head attention mechanism is used for efficient message passing between nodes in the protein graph. In the first... Layers, nodes Features and slave nodes arrive The edge features are denoted as follows: and They are projected onto the graph neural network layer before entering the graph neural network layer. In the latent space of dimension . Subsequently, the node features will be updated in the following manner:

[0049] in, Represents a node Neighbors The attention weight is expressed as follows:

[0050] In the formula, , and These represent the learnable weight matrices that map node features to the query space, key space, and value space, respectively. This represents a learnable weight matrix designed to effectively incorporate the geometric information encoded in the edges into the message aggregation process; multiple attention heads are applied in parallel, and their outputs are concatenated to form an updated node representation.

[0051] (ii) Edge update: Edge features are updated using information from neighboring nodes.

[0052]

[0053] in, This is a hyperparameter used to represent the importance of the source node; to prioritize biologically important sites, this embodiment specifically assigns it an importance parameter if and only if the source node is located within an important region (glycosylation site, receptor binding site, and antigen epitope site). This amplifies the contribution of key residues during the update process; This represents a multilayer perceptron used for edge updates, while This indicates a join operation.

[0054] (iii) Global Context Attention: Although local interactions are crucial for learning node representations, global information has also been shown to enhance method performance. Therefore, this invention implements a global context attention module. This module computes a global context attention vector and adaptively recalibrates node representations based on the global context using a gating mechanism.

[0055] in, This indicates the number of nodes in the protein graph; Represents element-wise multiplication; It is the sigmoid activation function; This represents a gated multilayer perceptron for adaptive control of the global context.

[0056] To further enhance the geometric embeddings obtained from geometric graph learning, this embodiment introduces learnable and site-specific positional embeddings that capture the functional meaning of residues based on their location. Specifically, residues are categorized into 10 different functional classes based on their position within the HA1 protein, including glycosylation sites, receptor binding sites (including 130-loop, 190-helix, and 220-loop), antigenic epitopes (AEs), and background residues. Inspired by the positional encoding mechanisms used in language models such as BERT, this embodiment adds learnable positional embeddings defined by categories to the geometric embeddings, thereby obtaining node features that can both characterize the biological meaning of nodes and capture the antigenic features of viral proteins. Subsequently, these node features are processed using a multilayer perceptron to obtain a comprehensive, global protein characterization.

[0057] Step 5: Antigen prediction using Transformer Encoder and Multilayer Perceptron: To predict the antigen distance between antigens and antiserum, the embedding vectors of paired viruses are concatenated and fed into a Transformer Encoder layer to capture the nonlinear correlation between them. Subsequently, the fused embedding vectors are fed into a multilayer perceptron to predict the antigen distance between virus pairs.

[0058] Step Six: Experimental Verification To verify the feasibility of the method provided in this embodiment, antigen distance prediction (regression) and differentiation of antigenic variants and non-variants (classification) tests were conducted on human influenza H3 virus and avian influenza H5 virus datasets under a five-fold cross-validation and retrospective testing setting. Key structural residues driving antigenic variants were accurately identified based on the attention mechanism in the model. Finally, to verify the practicality of the model, a vaccine recommendation experiment was conducted and compared with vaccines recommended by the WHO. Specifically, to evaluate the accuracy, biointerpretability, and role of the method of this invention in influenza surveillance, public health management, and vaccine strain selection, the following four analyses were performed: (1) Accurately predict the antigenic characteristics of influenza A virus: To rigorously evaluate generalization ability and ensure the robustness and reliability of performance estimates, five-fold cross-validation (repeated five times) was used to evaluate the performance of the method in antigen prediction. Specifically, the root mean squared error (RMSE) and mean absolute error (MAE) were used to evaluate the accuracy of the method in predicting antigen distance, and the area under the receiver operating characteristic curve (AUC) and the area under the precision-recall curve (AUPRC) were used to evaluate its ability to distinguish antigen variants from non-variant strains. Notably, the method of this invention provides highly accurate antigen distance estimates ( Figure 2 a, H3 subtype MAE=0.715 and RMSE=0.941, H5 subtype MAE=0.907 and RMSE=1.276, and showed strong discriminatory ability in distinguishing antigenic variants from non-variant strains. Figure 2 b, H3 subtype AUC=0.887 and AUPRC=0.925, H5 subtype AUC=0.932 and AUPRC=0.987).

[0059] Furthermore, to evaluate the superiority of the method of this invention, it was compared with several antigen prediction models as benchmarks. These models included SVM, MFPAD, AdaBoost and PREDAC, CNN, CNN-PSO, PREDAC-CNN, and GCN. Under five-fold cross-validation, the method of this invention performed best in H3 and H5 subtypes. Figure 2 (a and b), which shows that the structure-aware features learned through geometric graph networks help to predict antigen properties more accurately.

[0060] To further evaluate the contribution of structural features to prediction performance, three encoding methods based solely on the HA1 sequence were introduced: AAindex, ESM-2, and ProtTrans. In experiments predicting antigen distance and distinguishing antigen variants from non-variants, the method of this invention demonstrated comprehensive advantages in both H3 and H5 subtypes. Figure 2 (c and d). These results demonstrate that utilizing structural features achieves higher accuracy compared to methods based solely on sequence information.

[0061] To further verify the robustness of the method under data-scarce conditions, two sampling strategies (virus-based and antigen-antiserum pair-based) were implemented, with data retention rates ranging from 30% to 90%. In virus-based sampling, a subset of viruses was randomly selected, and an antigen-antiserum pair was retained only if both viruses constituting that pair appeared in the subset. In antigen-antiserum pair-based sampling, a subset of antigen-antiserum pairs was randomly selected. The method of this invention exhibits excellent robustness, maintaining a high prediction accuracy even with a significant reduction in training data, highlighting its stability under data-constrained conditions. Figure 2 eh).

[0062] (2) Accurately predict the antigenic characteristics of influenza A virus for the next flu season: Timely identification of emerging antigenic variants and acquisition of antigenicity of circulating viruses are crucial for pandemic prevention and preparedness. To this end, the World Health Organization has established national influenza centers globally to track antigenic variants. Accordingly, this study conducted retrospective testing to simulate real-world scenarios and predict the antigenic characteristics of circulating viruses for the next influenza season. In retrospective testing, the isolation date of each antigen-antiserum virus pair was considered. Specifically, for a given target year, the method of this invention was trained using hemagglutination inhibition titers and HA1 sequence data prior to that year, and then tested using HA1 sequence data from the target year. Given the scarcity of historical data for earlier years, 13 years for the H3 (2012-2024) and 7 years for the H5 (2014-2020) subtypes were selected as test years for method evaluation. Furthermore, given the relatively slow antigenic evolution and sparse data of the H5 virus, each round of testing for the H5 subtype combined two consecutive years. In retrospective testing, the method of this invention demonstrated superior performance compared to competing models in predicting antigen distance and distinguishing between antigenic variants and non-variants for both H3 and H5 subtypes. Figure 3 ad, H3 subtype MAE=0.816, RMSE=1.074, AUC=0.820 and AUPRC=0.891, H5 subtype MAE=1.983, RMSE=2.380, AUC=0.739 and AUPRC=0.982).

[0063] It is worth noting that, disregarding the initial testing year, the prediction error of the method of this invention for the H3 and H5 subtypes increased significantly in 2018. Figure 3 e and f). To investigate this phenomenon, FluNexus (https: / / flunexus.com / atmap.html) was used to visualize antigen evolution trends and generate antigen maps to show the antigenic drift of H3 and H5 subtype circulating viruses relative to viruses collected in previous years. Figure 3 (g and h). The prevalent virus in 2018 was widely distributed compared to previous years, indicating significant antigenic drift. This significant antigenic change is likely the cause of the temporary decline in model performance. In contrast, the method of this invention exhibits robust predictive performance when no significant antigenic drift or only slight antigenic drift is observed in subsequent years. Figure 3 eh).

[0064] (3) Accurately identify sites in the HA1 protein associated with antigen drift: Increasing evidence suggests that site-specific substitutions within HA1 are a major driver of influenza antigen evolution. To quantify the impact of these substitutions on antigenicity, the method of this invention utilizes the attention mechanism in a geometric graph learning network to identify specific HA1 sites that have a dominant influence on antigenic drift. Figure 4 Specifically, the importance of each site is quantified by the attention weights on the edges connecting the nodes.

[0065] For the H3 subtype, 27 of the top 30 most important sites are located within key functional regions: antigenic epitopes (AEs) and receptor-binding regions (190-helix, 130-loop, and 220-loop). Figure 4 (a, c, and g). These regions have long been considered the main determinants of the antigenic properties of influenza A viruses, and amino acid substitutions within them have had a significant impact on the antigenic evolution of influenza A H3 and H5 viruses. Overall, these findings confirm that the method of this invention successfully integrates prior knowledge about key biological sites.

[0066] To further elucidate the potential functional roles of the most important sites identified in this invention in antigen evolution, a detailed analysis of representative sites was conducted. For the H3 subtype, residues 156, 159, and 186 within epitope B showed high importance scores (…). Figure 4 (a and c). Studies have shown that positions 156 and 159 have a significant impact on antigenic evolution, and the Y156F substitution is the main driving factor for antigenic drift from FU02 to the CA04 cluster. Furthermore, the G186D mutation is a key determinant of antigenic drift; this mutation stabilizes the N190 side chain conformation through hydrogen bonds, promoting direct interaction between viral proteins and human receptors, thereby driving the evolution of receptor binding patterns. For the H5 subtype, position 152 in epitope B and positions 141 and 127 in epitope A show high importance scores (…). Figure 4(b) and (d). Experimental evidence confirms that this site is a crucial antigenic determinant of the H5 subtype avian influenza virus clade 2.3.4.4. Meanwhile, antigenic evolution studies of the H5 subtype avian influenza virus have consistently identified immune escape mutants with substitution at site 152, highlighting the critical role of site 152 in mediating immune escape and driving antigenic drift. Furthermore, site 127 within epitope A is a key determinant of H5 subtype antigenic escape.

[0067] Furthermore, statistical analysis revealed that the key sites of the H3 subtype predicted by the method of this invention were significantly enriched in antigenic epitopes B, D, and E. Figure 4 e, g), while the key sites of the predicted H5 subtype are mainly enriched in epitopes A, D and E (e, g), , Figure 4 (f, h). Previous studies have confirmed that epitope B in the H3 subtype exhibits the highest immune dominance, while epitope A is a hotspot for H5 viral site mutations and a major target of the host immune response. This highlights the ability of the method of this invention to elucidate specific immune dominance epitopes that are crucial to antigen evolution.

[0068] (4) Screening and recommendation of influenza vaccine strains The effectiveness of a vaccine fundamentally depends on the antigenic similarity between the vaccine strain and the circulating virus. Therefore, to evaluate the capability of the method of this invention in influenza surveillance and vaccine recommendation, an experiment was designed to simulate the decision-making process of the World Health Organization Vaccine Components Conference—which selects vaccines for the next influenza season based on historical surveillance data. Given that the future circulating strain is uncertain, and that future circulating strains typically evolve from the dominant strains of the current season, targeting the antigenic centers of previously circulating viruses can maximize the probability of cross-protection against emerging variants. Therefore, coverage is introduced as an indicator of the antigenic centrality of candidate vaccine viruses, and this guides vaccine recommendations.

[0069] To recommend vaccines, a model is trained using only historical data prior to the target year, based on the predicted coverage calculated using the method of this invention, to predict the best candidate strain for recommendation. The virus with the highest predicted coverage is selected as the vaccine strain recommended by the method of this invention for the following year. To rigorously evaluate the performance of the method of this invention, the strain recommended at the "theoretical ceiling" is defined as the benchmark. Unlike the prediction model, the "theoretical ceiling" recommended strain is determined based on the empirical coverage actually observed in the target year (i.e., the coverage inferred by retrospective training using known data from the target year). Since the "theoretical ceiling" recommended strain uses a complete understanding of future epidemic data to identify the most effective candidate strain, it represents the theoretical upper limit of cross-protection by vaccines. The method of this invention, the "theoretical ceiling," and the specific vaccine strains recommended by the WHO each year are shown in Tables 1 and 2.

[0070] Table 1

[0071] Table 2

[0072] Due to the inherent sparsity of global hemagglutination inhibition titer data, it is impractical to directly calculate a reliable, data-driven empirical coverage rate from hemagglutination inhibition titer data to verify vaccine efficacy. Therefore, empirical coverage rate is used as a benchmark to quantify the cross-protective effect provided by each recommended strain against the circulating virus of the target year. To confirm whether the empirical coverage rate derived by the method of this invention accurately reflects the calculation results based on hemagglutination inhibition titer, a correlation analysis was performed between the empirical coverage rate estimated by the method of this invention and the empirical coverage rate calculated based on hemagglutination inhibition titer. The correlation test confirmed a significant correlation between the empirical coverage rate derived by the method of this invention and the empirical coverage rate derived based on hemagglutination inhibition titer (Pearson correlation coefficient). , Spearman's rank correlation coefficient , This provides a reliable basis for subsequent evaluation of vaccine efficacy.

[0073] The results of the vaccine efficacy assessment highlight the superior robustness and accuracy of the method of this invention. Figure 5 The vaccine recommended at the "theoretical upper limit" consistently achieves near-perfect empirical coverage (close to 1.0). Notably, the strain recommended by the method of this invention not only outperformed the WHO-recommended vaccine strain in empirical coverage for 10 out of 13 years, but its empirical coverage is also very close to that of the vaccine strain recommended at the "theoretical upper limit." In particular, in 2014, the coverage of the WHO-recommended strain was only 0.33, an extremely low level. This is consistent with widely reported findings during that influenza season that there was a severe antigenic mismatch between the WHO-recommended A / Texas / 50 / 2012 (H3N2) strain and the recommended circulating strain for the influenza season. Overall, these results demonstrate that, in practical surveillance scenarios, the method of this invention provides a highly reliable and superior alternative for vaccine strain selection.

[0074] This invention provides a system for an influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning, comprising: The protein structure module is used to obtain the three-dimensional structure of the influenza virus hemagglutinin subunit 1 protein based on the influenza virus protein sequence using ESMFold prediction. The initial feature module is used to construct a protein graph based on the spatial adjacency relationship between residues in the three-dimensional structure of viral proteins, extract the spatial geometric structure features of proteins from the three-dimensional structure of viral proteins, construct a local coordinate system to characterize the spatial constraint relationship between residues, and obtain the initial node features and edge features of the protein graph. The protein characterization module is used to input the protein graphs of paired viruses into the geometric graph learning model in the Siamese network architecture. It uses a multi-scale attention mechanism to update the node features and edge features, thereby obtaining residue characterization that captures both long-range and short-range dependencies between residues. The residue characterization vector is then fused with the embedding vector based on the key functional site information of the protein. Subsequently, the fused characterization is input into a multilayer perceptron to obtain the characterization of the viral protein. The antigen prediction module is used to input the protein characteristics of the two viruses into the Transformer Encoder to extract the nonlinear interaction features between them, and output the antigen prediction results through a multilayer perceptron.

[0075] Although embodiments of the present invention have been shown and described above, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning, characterized by, include: The three-dimensional structure of the influenza virus hemagglutinin subunit 1 protein was obtained by using ESMFold prediction based on the influenza virus protein sequence. A protein graph was constructed based on the spatial adjacency relationships between residues in the three-dimensional structure of the influenza virus hemagglutinin subunit 1 protein. Spatial geometric features of the protein were extracted from the three-dimensional structure of the viral protein, and a local coordinate system was constructed to characterize the spatial constraint relationships between residues, thus obtaining the initial node features and edge features of the protein graph. The protein graphs of paired viruses are input into the geometric graph learning model in the Siamese network architecture. The node features and edge features are updated using a multi-scale attention mechanism to obtain residue representations that capture both long-range and short-range dependencies between residues. The residue representation vectors are then fused with embedding vectors based on key functional site information of proteins. Subsequently, the fused representations are input into a multilayer perceptron to obtain the representations of viral proteins. The protein characteristics of the two viruses are input into the Transformer Encoder to extract the nonlinear interaction features between them, and the antigen prediction results are output through a multilayer perceptron.

2. The influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning according to claim 1, characterized in that, The protein graph, whose nodes represent residues, edges represent spatial proximity, i.e. when the spatial distance between atoms of two nodes is less than a certain threshold, an edge is built between the two nodes. The protein graph, whose nodes represent residues, edges represent spatial proximity, i.e. when the spatial distance between atoms of two nodes is less than a certain threshold, an edge is built between the two nodes. 3.The influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning according to claim 1, wherein, When constructing a local coordinate system to characterize the spatial constraints between residues, the following steps are included: A local coordinate system is constructed based on the main chain atoms of each residue; Based on this local coordinate system, the geometric features of the protein are extracted as the initial features of nodes and edges; wherein, the initial features of nodes include angle features, first distance features, and first direction features; the initial features of edges include relative orientation features, second distance features, and second direction features.

4. The influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning according to claim 3, characterized in that, The sine and cosine values of the extracted protein backbone dihedral angles and bond angles are taken as angle features; the spatial distances between atoms within a residue are encoded using a radial basis function to obtain first distance features; and direction vectors of the atoms in a residue relative to the local coordinate system are calculated and projected into the local coordinate system to obtain first direction features.

5. The influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning according to claim 3, characterized in that, The relative rotation matrix between residue pairs is calculated using quaternion functions to characterize the relative orientation features; the atomic distance between residues is encoded using radial basis functions to obtain the second distance feature. The relative spatial positions of residues with respect to the local coordinate system are calculated to obtain the second directional features.

6. The influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning according to claim 1, characterized in that, The updating of node features includes: The node update in the geometric graph learning model employs a multi-head attention mechanism, in the first... Layers, nodes Features and nodes arrive The edge features are respectively represented as and Both were projected onto After creating a latent space, node features are updated as follows: In the formula, Represents a node Neighbors The attention weights are defined as follows: In the formula, , and Let represent the learnable weight matrices that map node features to the query space, key space, and value space, respectively. This represents a learnable weight matrix used to effectively incorporate geometric information encoded in the edges into the message aggregation process.

7. The influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning according to claim 6, characterized in that, The edge features are updated using information from neighboring nodes: In the formula, This is a hyperparameter representing the importance of the source node, which is true if and only if the source node is located in an important region. ; This represents a multilayer perceptron used for edge updates; This indicates a join operation.

8. The influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning according to claim 1, characterized in that, The process of characterizing residues that capture both long-range and short-range dependencies between residues also includes adaptively recalibrating nodes based on the global context using a gating mechanism. When recalibrating nodes, a global context vector is obtained by averaging the features of all nodes, and a gating mechanism is used to adaptively update the node representations based on the global context vector: In the formula, This indicates the number of nodes in the protein graph; Represents element-wise multiplication; This represents the sigmoid activation function; This represents a gated multilayer perceptron used for adaptive control of the global context; Indicates the first The global context vector corresponding to the layer; Represents a node In the The intermediate representation before the layer gating update; Represents a node In the The final representation of the layer.

9. The influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning according to claim 1, characterized in that, The residue representation vector is fused with the embedding vector based on key functional sites of the protein. The embedding vector based on key functional sites of the protein refers to the learned embedding obtained by setting different initial seeds according to the importance category corresponding to the location of the residue and training. The importance category includes 10 categories, namely glycosylation site, 130-loop, 190-helix, 220-loop, antigenic epitope A, antigenic epitope B, antigenic epitope C, antigenic epitope D, antigenic epitope E, and background residues. The fusion method is to add the two together.

10. A system for the influenza virus antigen prediction method based on ESMFold protein structure prediction and geometric graph learning as described in claim 1, characterized in that, include: The protein structure module is used to obtain the three-dimensional structure of the influenza virus hemagglutinin subunit 1 protein based on the influenza virus protein sequence using ESMFold prediction. The initial feature module is used to construct a protein graph based on the spatial adjacency relationship between residues in the three-dimensional structure of viral proteins, extract the spatial geometric structure features of proteins from the three-dimensional structure of viral proteins, construct a local coordinate system to characterize the spatial constraint relationship between residues, and obtain the initial node features and edge features of the protein graph. The protein characterization module is used to input the protein graphs of paired viruses into the geometric graph learning model in the Siamese network architecture. It uses a multi-scale attention mechanism to update the node features and edge features, thereby obtaining residue characterization that captures both long-range and short-range dependencies between residues. The residue characterization vector is then fused with the embedding vector based on the key functional site information of the protein. Subsequently, the fused characterization is input into a multilayer perceptron to obtain the characterization of the viral protein. The antigen prediction module is used to input the protein characteristics of the two viruses into the Transformer Encoder to extract the nonlinear interaction features between them, and output the antigen prediction results through a multilayer perceptron.

Images