A system and method for predicting the evolutionary trajectory of influenza viruses

Abstract

The application discloses a system and method for predicting an influenza virus evolution track, and the system comprises a virus evolution track search model and a virus track prediction module; the virus track diffusion model comprises a diffusion training module, a fusion module, a time attention module, a space attention module and a track optimization module; the application proposes a sampling strategy for the influenza virus evolution track based on the low genetic diversity and antigenicity of the influenza virus, constructs an influenza virus evolution track dataset through the proposed sampling strategy, generates the influenza virus evolution track by using a diffusion model with the fusion module and the space-time attention module, and predicts the evolution track of the influenza virus.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and more particularly to a system and method for predicting the evolutionary trajectory of viruses. Background Technology

[0002] Phylogenetic trees are powerful tools for elucidating evolutionary relationships between species, pathogens, and individual human cells. However, constructing phylogenetic trees from large-scale sequence data remains computationally intensive. Because phylogenetic tree construction is largely based on sequence variation, it often ignores biological constraints that may influence the evolutionary process. One clustering-based approach samples evolutionary trajectories by first grouping influenza strains by year and then applying K-means clustering to the sequences for each year to generate clusters. During sampling, a strain is randomly selected from the current year's cluster that is closest in Euclidean distance to the previous year's cluster. However, this method still does not consider the biological characteristics of the virus.

[0003] Due to the frequent mutations of viruses, predicting their evolution has attracted widespread attention. In site-specific approaches, Temple utilizes LSTM to capture historical residue information, enabling the prediction of mutations at specific sites and thus predicting influenza evolution. Some works model temporally resolved frequency patterns of mutations within viral genome fragments to predict viral evolution and select representative strains for influenza vaccines. In sequence-specific approaches, two-dimensional CNN architectures are used to extract vector spaces with distributed amino acid representations and predict influenza antigen evolution. A generative adversarial network, combining sequence-to-sequence recurrent neural network generators, has been used to accurately predict the genetic mutations and evolution of future biological populations. However, these methods cannot simultaneously capture global sequence information and temporal evolutionary information, making it difficult to comprehensively reflect viral evolution.

[0004] Diffusion models, as a highly regarded generative model, have achieved significant success in image synthesis. Recent research advancements have further demonstrated their effectiveness in generating high-quality text data. Notably, emerging explorations in video generation highlight the model's ability to handle time-sensitive data. This technical characteristic bears an interesting resemblance to viral evolution: each individual video frame can be likened to a virus, while the video corresponds to the virus's evolutionary process. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a diffusion model for predicting the evolutionary trajectory of influenza viruses, which can predict the future sequence of the virus based on historical sequence information of influenza viruses.

[0006] The objective of this invention is achieved through the following technical solution: A system for predicting the evolutionary trajectory of an influenza virus, the system comprising an influenza virus evolutionary trajectory search model and an influenza virus evolutionary trajectory prediction model; the influenza virus evolutionary trajectory prediction model includes a diffusion training module, a fusion module, a temporal attention module, a spatial attention module, and a trajectory optimization module; comprising the following steps: The virus evolution trajectory search module collects influenza virus sequences from different time points to construct the optimal evolution trajectory of the influenza virus; wherein: The influenza virus data were divided according to different years to obtain the sequence data of influenza virus strains and the antigenic phenotype data of influenza viruses. A directed graph of viral evolution is constructed based on the segmented influenza virus strains as nodes; The trajectory of viral genetics and antigenic evolution is obtained by calculating the connection relationships between viral nodes in the directed viral evolution graph using the sequence data and antigenic phenotype data of influenza virus strains according to the following formula.

[0007] in: This indicates the virus strain sequence from that year. This indicates the viral strain sequence for the following year. Indicates the genetic difference distance between two sequences; Indicates the antigen distance between two sequences; and The weights of genetic distance and antigenic distance are respectively represented. Set to 1, Set to 10; The shortest distance between the genetic and antigenic evolutionary trajectories of the virus is used as the optimal evolutionary trajectory of the influenza virus stream, based on a graph search algorithm. The influenza virus evolution trajectory prediction model uses deep representation learning and pattern mining to predict the future evolution trajectory of the influenza virus by predicting the optimal evolution trajectory of the influenza virus.

[0008] Furthermore, the influenza virus evolution trajectory prediction model uses deep representation learning and pattern mining to predict the future evolution trajectory of the influenza virus by performing deep representation learning and pattern mining on the optimal evolution trajectory of the influenza virus, including: The diffusion training module processes parameters through a forward process. Influenza virus samples are generated by embedding discrete amino acid sequences of influenza viruses. ; The fusion module fuses the sequence information and evolutionary gradient information of the influenza virus in its evolutionary trajectory to generate an influenza virus conditional feature vector.

[0009] in: G represents historical information, and G represents the evolutionary gradient. It is a neural network used as a feature extractor; The diffusion training module processes samples through a reverse process. With parameters After learning, embedding the conditional feature vector of influenza virus Generate the evolutionary trajectory of influenza viruses ; The spatial module will assign attention weights to the correlations between amino acids. Embedding the evolutionary trajectory of influenza virus to obtain the spatially related evolutionary trajectory of influenza virus ; The time attention module assigns weights to amino acids at different time points. Embedded in the spatial evolutionary trajectory of popular viruses Obtaining the spatiotemporal evolution trajectory of the influenza virus ; The trajectory optimization module uses a loss function to minimize the difference between the spatiotemporal evolution trajectory of the influenza virus and the actual trajectory, generating the future evolution trajectory of the influenza virus.

[0010] Attention weights for the correlations between the amino acids The statement is as follows:

[0011] in: , , These represent the query, key, and value in the attention mechanism, respectively. This represents the attention weights that reflect the correlation between amino acids. Indicates the updated number Year virus embedded.

[0012] Furthermore, the weights of the amino acids at different time points The statement is as follows:

[0013] in: , , ⊙ represents the multiplication of corresponding elements in the matrix. The MASK is responsible for maintaining the temporal order of amino acid evolution. Indicates the first Update the embedding at each site.

[0014] The present invention can also adopt the following technical solutions: A method for predicting the evolutionary trajectory of influenza viruses includes the following steps: S1 collected influenza virus sequences from different time points to construct the optimal evolutionary trajectory of the influenza virus; among which: 101. By dividing the influenza virus data according to different years, sequence data of influenza virus strains and antigenic phenotype data of influenza viruses were obtained. 102. Construct a directed graph of viral evolution based on the divided influenza virus strains as nodes; 103. Using the sequence data and antigenic phenotype data of influenza virus strains, calculate the connection relationships between virus nodes in the directed viral evolution graph according to the following formula to obtain the trajectory of viral genetics and antigenic evolution.

[0015] in: This indicates the virus strain sequence from that year. This indicates the viral strain sequence for the following year. Indicates the genetic difference distance between two sequences; Indicates the antigen distance between two sequences; and The weights of genetic distance and antigenic distance are respectively represented. Set to 1, Set to 10; 104. The shortest distance between the viral genetic and antigenic evolutionary trajectories is used as the optimal evolutionary trajectory for the influenza virus, based on a graph search algorithm. S2 uses deep representation learning and pattern mining to predict the future evolutionary trajectory of influenza viruses by applying the optimal evolutionary trajectory of the virus; among which: 201. Parameters are passed through a forward process. Influenza virus samples are generated by embedding discrete amino acid sequences of influenza viruses. ; 202. Fuse the sequence information and evolutionary gradient information of the influenza virus in its evolutionary trajectory to generate a conditional feature vector for the influenza virus:

[0016] in: G represents historical information, and G represents the evolutionary gradient. It is a neural network used as a feature extractor; 203. Samples are processed through a reverse process. With parameters After learning, embedding the conditional feature vector of influenza virus Generate the evolutionary trajectory of influenza viruses ; 204. Attention weighting of the correlation between amino acids Obtaining the spatial evolutionary trajectory of influenza viruses by embedding their evolutionary trajectory ; 205. Weighting of amino acids at different time points Embedded in the spatial evolutionary trajectory of influenza viruses Obtaining the spatiotemporal evolution trajectory of the influenza virus ; 206. The future evolutionary trajectory of the influenza virus is generated by minimizing the difference between the spatiotemporal evolutionary trajectory of the epidemic virus and the actual trajectory by using a loss function.

[0017] Beneficial effects Compared with the prior art, the beneficial effects of the technical solution of the present invention are: 1. The sampling method for viral evolutionary trajectory in this invention differs from previous methods. Previous methods only focused on sequence changes in viral strains, while the sampling method in this invention models both genetic diversity and antigenicity, assigning weights to each to flexibly adjust the biological properties of the influenza virus evolutionary trajectory. First, the influenza virus data is divided by year, and edges are used to connect strains from adjacent years, constructing a directed graph. This ensures that the viral evolutionary trajectory strictly follows chronological order, preventing unrealistic "time reversal" or long-distance leaps, and conforming to the observation of the gradual, year-by-year evolution of the influenza virus.

[0018] 2. This invention uses a diffusion model with a fusion module and a spatiotemporal attention module as the main model. It utilizes neural networks to extract interaction information between amino acids in the spatial dimension and the evolutionary trend of influenza viruses in the temporal dimension. This model can predict influenza virus strains in future years based on historical influenza virus information. This invention maps discrete amino acid sequences to a continuous vector space and automatically learns the biochemical semantics of amino acids (such as the hydrophilicity of antigenic sites and the structural stability of conserved regions), overcoming the limitations of traditional methods that rely on manually designed features. This invention integrates evolutionary gradients into the evolutionary trajectory, quantifying the direction and magnitude of each step of change, enabling explicit guidance and constraint of the evolutionary prediction process. This invention extracts long-range dependencies between sites within a single sequence in the spatial dimension, resolving protein spatial structural constraints; and models the directional evolutionary trend of the same site across historical sequences in the temporal dimension, identifying evolutionary signals driven by immune selection pressure. The combination of these two approaches allows the model to focus on regions of high biological value while outputting interpretable attention weights, providing a scientific basis for mechanism analysis. Attached Figure Description

[0019] Figure 1 This is an example diagram of the sampling strategy in the method of the present invention.

[0020] Figure 2 This is an example diagram of the model used in the method of the present invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0022] like Figure 1 , Figure 2 As shown, the diffusion system for predicting the evolutionary trajectory of influenza viruses proposed in this invention comprises two parts: a sampling strategy for influenza virus evolutionary trajectories based on the low genetic diversity and antigenicity of influenza viruses, and the construction of an influenza virus evolutionary trajectory dataset using the proposed sampling strategy; and the generation of influenza virus evolutionary trajectories using a diffusion model with a fusion module and a spatiotemporal attention module for predicting influenza viruses. The system includes a virus evolutionary trajectory search model and a virus trajectory prediction module; the virus trajectory diffusion model includes a diffusion training module, a fusion module, a temporal attention module, a spatial attention module, and a trajectory optimization module. Specific process: 1. Sampling Strategy The evolutionary trajectory of influenza viruses is sampled based on two properties they satisfy during their evolution: low genetic diversity and antigenicity. Low genetic diversity means that although influenza viruses continuously mutate, the diversity of actual circulating strains at a specific point in time is limited relative to their vast potential sequence space. This implies that viral evolution does not wander randomly in all directions, but rather proceeds along certain "trunks" or "trajectories." Antigenicity refers to the fact that the main driving force behind influenza virus evolution is immune pressure. Viruses evade immune recognition by altering antigenic sites on their surface proteins. Therefore, changes in antigenicity are a core indicator for measuring the direction and speed of evolution.

[0023] 1.1 The influenza virus data is divided by year.

[0024] The data used in this invention comprises two parts: sequence data of influenza virus strains, containing the genetic information of the influenza virus; and antigenic phenotype data of the influenza virus, which are measured by hemagglutinin (HI) assays, using normalized HI titers (NHT) to measure the antigenic distance between two virus strains. A higher NHT value indicates a greater antigenic distance. The calculation formula is as follows:

[0025] Among them, homology titer and heterologous titer These represent the inhibitory reference virus isolates. and test virus isolates Antiserum required for cell agglutination The reciprocal of the maximum dilution.

[0026] The two sets of data are used to represent the heritability and antigenicity of the influenza virus, respectively.

[0027] 1.2 Constructing a Directed Graph The earliest virus strain sequence was used as the ancestor node. And connect each strain of the current year with all strains of the following year. For nodes and nodes The edge weights between them are calculated as follows:

[0028] in, This indicates the virus strain sequence from that year. Indicates the virus strain sequence for the following year. and These represent the edit distance and antigen distance between the two sequences, respectively. and These represent the weights of the edit distance and the antigen distance, respectively. Set to 1, Set it to 10.

[0029] The constructed directed graph represents all the evolutionary trajectories of the virus, and the edge weights are represented by the combination of edit distance and antigenic distance between viruses, which includes two biological properties of influenza virus: low genetic diversity and antigenic variation.

[0030] 1.3 Selecting the shortest trajectory Apply Dijkstra's algorithm to the graph to find the shortest trajectory from the earliest year to the latest year, and select it as the evolutionary trajectory.

[0031] 2. Diffusion model with fusion module and spatiotemporal attention module A diffusion model with a fusion module and a spatiotemporal attention module predicts viral strain sequences based on historical sequence information. It consists of two key components: a fusion module that embeds evolutionary gradients into the trajectory; and a spatiotemporal attention module that uses self-attention mechanisms in both temporal and spatial dimensions to predict and remove noise at each step, ultimately generating realistic viral evolutionary trajectory samples. The fusion module uses a feature extractor composed of neural networks to extract evolutionary information from the evolutionary gradients and integrates it into the influenza virus evolutionary trajectory, where the evolutionary gradient is obtained by calculating the differences between viral sequence embeddings in the evolutionary trajectory. The spatiotemporal attention module extracts global sequence information and evolutionary trends from the influenza virus strain sequence, respectively. This module simultaneously considers the interactions between amino acids within a single sequence and the changes at specific sites across sequences over time.

[0032] 2.1 Diffusion Process Diffusion processes include forward and reverse processes.

[0033] In the forward process, the sequence used for prediction It gradually transforms into random noise. Compared to the standard forward process, it initially introduces a noise source... Parameterized Markov transformations convert discrete amino acid sequences into continuous embeddings. Through... The extension of the forward process allows for iterative diffusion. For each time step This invention applies diffusion distribution Generate a noisier sample. Finally, the sequence... It becomes almost pure random noise. It follows a standard Gaussian distribution.

[0034] In the reverse process, considering the relationship between the predicted sequence and the historical sequence, for each time step... A by Parameterized learning denoising distribution Generate samples The condition is that the previous sample had more noise. and conditional features after integrating historical information and evolutionary gradient ,Right now When the reverse denoising process reaches... At that time, the present invention distributes by rounding. Will The sequence is rounded to the nearest one in the embedding space to generate the final sequence. In practical implementation, this invention will... The input is fed into the spatiotemporal attention module for denoising, thereby achieving... Sampling.

[0035] 2.2 Fusion Module Evolutionary gradients characterize the directional trend of viral evolution within a genetic-antigen landscape, providing a quantitative measure of adaptation under selective pressure. The evolutionary gradient is obtained by calculating the differences between viral sequence embeddings in the evolutionary trajectory, extracting evolutionary information through a feature extractor, and integrating this information into the evolutionary trajectory. This process can be represented as:

[0036] in, G represents historical information, and G represents the evolutionary gradient. It is a neural network used as a feature extractor.

[0037] 2.3 Spatiotemporal Attention Module When learning from viral evolutionary trajectory data, it is essential to consider both the interactions between amino acids within a single sequence and the changes in specific amino acid sites across sequences over time. The spatiotemporal attention module addresses this dual consideration.

[0038] The spatial attention module is responsible for capturing global sequence information. The spatial attention module acts on... ,in , indicating the first Embedding of the virus sequence from that year. This module uses a self-attention mechanism to calculate the attention weight matrix. ,in Quantified the first The and the first The correlation between amino acids. The attention weight matrix is ​​used to encode the interaction information between amino acids into the embedding. The detailed formula is as follows:

[0039] in: , , These represent the query, key, and value in the attention mechanism, respectively. This represents the attention weight that reflects the correlation between amino acids. Indicates the updated number Year virus embedding. The updated evolutionary trajectory is represented as .

[0040] The temporal attention module is used to extract evolutionary trends. After processing by the spatial attention module, the temporal attention module embeds the evolutionary trajectories at specific sites. As input, where , representing the first in the evolutionary trajectory The module describes the amino acid evolution process at each site. Since amino acid mutations at a specific time point depend only on historical information, it employs a masking mechanism to prevent the influence of amino acids at the same site in the future, thus clearly defining the time dependence of the evolutionary trajectory. The detailed formula is as follows:

[0041] in , , , The contribution of amino acids at different time points to the evolutionary trajectory is measured. ⊙ represents the multiplication of corresponding elements in the matrix. MASK is responsible for maintaining the temporal order of amino acid evolution. Indicates the first The updated embeddings at each site. The final evolutionary trajectory is represented as follows: .

[0042] 2.4 Loss Function The formula for calculating the training objective is as follows:

[0043] in It is the loss function of the diffusion model. It is the loss function for converting continuous embeddings of evolutionary trajectories into discrete amino acid sequences. It is posterior The mean, yes The predicted mean.

[0044] In this invention, four evaluation metrics are used to assess the predictive performance of the model: sequence accuracy, epitope accuracy, mean squared error of antigen distance (MSE), and mean absolute error of antigen distance (MAE). Sequence accuracy refers to the sequence similarity between the predicted sequence and the true sequence. Epitope accuracy refers to the epitope similarity between the predicted sequence and the true sequence. MSE and MAE quantify the difference in antigenic distance between the predicted strain and the true strain and ancestral strain. The formulas are as follows:

[0045] in It represents the total number of influenza virus evolutionary trajectories, while SAP is a model for predicting antigen distance. They represent the first Predicted strains, actual strains, and ancestral strain sequences of an evolutionary trajectory.

[0046] Table 1: Comparison of evaluation results across four indicators

[0047] Table 1 shows the performance of this model on the two datasets.

[0048] This invention is not limited to the embodiments described above. The above description of specific embodiments is intended to illustrate and explain the technical solutions of this invention. The specific embodiments described above are merely illustrative and not restrictive. Without departing from the spirit and scope of the claims, those skilled in the art can make many specific modifications based on the teachings of this invention, and these modifications all fall within the scope of protection of this invention.

Claims

1. A system for predicting the evolutionary trajectory of influenza viruses, characterized in that: The system includes an influenza virus evolution trajectory search model and an influenza virus evolution trajectory prediction model; the influenza virus evolution trajectory prediction model includes a diffusion training module, a fusion module, a temporal attention module, a spatial attention module, and a trajectory optimization module; and includes the following steps: The virus evolution trajectory search module collects influenza virus sequences from different time points to construct the optimal evolution trajectory of the influenza virus; wherein: The influenza virus data were divided according to different years to obtain the sequence data of influenza virus strains and the antigenic phenotype data of influenza viruses. A directed graph of viral evolution is constructed based on the segmented influenza virus strains as nodes; The trajectory of viral genetics and antigenic evolution is obtained by calculating the connection relationships between viral nodes in the directed viral evolution graph using the sequence data and antigenic phenotype data of influenza virus strains according to the following formula. in: This indicates the virus strain sequence from that year. This indicates the viral strain sequence for the following year. Indicates the genetic difference distance between two sequences; Indicates the antigen distance between two sequences; and The weights representing genetic distance and antigenic distance are respectively... Set to 1, Set to 10; The shortest distance between the genetic and antigenic evolutionary trajectories of the virus is used as the optimal evolutionary trajectory of the influenza virus, based on a graph search algorithm. The influenza virus evolution trajectory prediction model uses deep representation learning and pattern mining to predict the future evolution trajectory of the influenza virus by predicting the optimal evolution trajectory of the influenza virus.

2. The system for predicting the evolutionary trajectory of influenza viruses according to claim 1, characterized in that: The influenza virus evolution trajectory prediction model uses deep representation learning and pattern mining to predict the future evolution trajectory of the influenza virus by applying the optimal evolutionary trajectory of the influenza virus. This includes: The diffusion training module processes parameters through a forward process. Influenza virus samples are generated by embedding discrete amino acid sequences of influenza viruses. ; The fusion module fuses the sequence information and evolutionary gradient information of the influenza virus in its evolutionary trajectory to generate an influenza virus conditional feature vector. in: G represents historical information, and G represents the evolutionary gradient. It is a neural network used as a feature extractor; The diffusion training module processes samples through a reverse process. With parameters After learning, embedding the conditional feature vector of influenza virus Generate the evolutionary trajectory of influenza viruses ; The spatial attention module assigns attention weights to the correlations between amino acids. Embedding the evolutionary trajectory of influenza virus to obtain the spatially related evolutionary trajectory of influenza virus ; The time attention module assigns weights to amino acids at different time points. Embedded in the spatial evolutionary trajectory of popular viruses Obtaining the spatiotemporal evolution trajectory of the influenza virus ; The trajectory optimization module uses a loss function to minimize the difference between the spatiotemporal evolution trajectory of the influenza virus and the actual trajectory, generating the future evolution trajectory of the influenza virus.

3. A system for predicting the evolutionary trajectory of influenza viruses according to claim 2, characterized in that: Attention weights for the correlations between the amino acids The statement is as follows: in: , , These represent the query, key, and value in the attention mechanism, respectively. This represents the attention weights that reflect the correlation between amino acids. Indicates the updated number Year virus embedded.

4. The system for predicting the evolutionary trajectory of influenza viruses according to claim 2, characterized in that: The weights of amino acids at different time points The statement is as follows: in: , , ⊙ represents the multiplication of corresponding elements in the matrix. The MASK is responsible for maintaining the temporal order of amino acid evolution. Indicates the first Update the embedding at each site.

5. A method for predicting the evolutionary trajectory of an influenza virus, said method being based on a system implementation of any one of claims 1-4, characterized in that, Includes the following steps: S1 collected influenza virus sequences from different time points to construct the optimal evolutionary trajectory of the influenza virus; among which:

101. The influenza virus data were divided according to different years to obtain the sequence data of influenza virus strains and the antigenic phenotype data of influenza virus. 102 Construct a directed graph of viral evolution based on the divided influenza virus strains as nodes; 103. Using the sequence data and antigenic phenotype data of influenza virus strains, the connection relationships between virus nodes in the directed viral evolution graph are calculated according to the following formula to obtain the trajectory of viral genetics and antigenic evolution. in: This indicates the virus strain sequence from that year. This indicates the viral strain sequence for the following year. Indicates the genetic difference distance between two sequences; Indicates the antigen distance between two sequences; and The weights representing genetic distance and antigenic distance are respectively... Set to 1, Set to 10; 104. The shortest distance between the viral genetic and antigenic evolutionary trajectories is used as the optimal evolutionary trajectory of the influenza virus based on the graph search algorithm. S2 uses deep representation learning and pattern mining to predict the future evolutionary trajectory of influenza viruses by applying the optimal evolutionary trajectory of the virus; among which: 201 uses a forward process to transfer parameters Influenza virus samples are generated by embedding discrete amino acid sequences of influenza viruses. ; 202. The sequence information and evolutionary gradient information of the influenza virus in its evolutionary trajectory are fused to generate a conditional feature vector of the influenza virus: in: G represents historical information, and G represents the evolutionary gradient. It is a neural network used as a feature extractor; 203 The sample is processed through a reverse process. With parameters After learning, embedding the conditional feature vector of influenza virus Generate the evolutionary trajectory of influenza viruses ; 204 Attention weights for the correlation between amino acids Obtaining the spatial evolutionary trajectory of influenza viruses by embedding their evolutionary trajectory ; 205. Weights of amino acids at different time points Embedded in the spatial evolutionary trajectory of influenza viruses Obtaining the spatiotemporal evolution trajectory of the influenza virus ; 206 uses a loss function to minimize the difference between the spatiotemporal evolution trajectory of the prevalent virus and the actual trajectory to generate the future evolution trajectory of the influenza virus.

Images