Virus evolution trend prediction method and system

Abstract

The present invention provides a method and system for predicting the evolutionary trend of viruses, including: obtaining virus protein-target protein pairs with affinity annotations based on deep mutation scanning, dividing them into training sets, validation sets, and test sets; building a data encoder to encode data for a protein language model; performing layer-by-layer parameter training on the protein language model in the data encoder on downstream tasks; evaluating the two indicators of predicted affinity and experimental affinity with the correlation coefficient R2 on the training set and validation set, selecting the optimal model, and finally performing prediction verification on the test set; calculating the immune escape score based on structure and the immune escape score based on the language model; performing multiple rounds of iterations of virus adaptability calculation and genetic variation to screen out potential preset high-risk strains. The present invention can reveal the evolutionary pattern of viruses from a molecular level, effectively simulate the mutation law of viruses, and predict the effect of epidemic trends.

Description

Technical Field

[0001] The present invention relates to the field of trend prediction technology, and more particularly to a method and system for predicting virus evolution trends. In particular, the present invention relates to a method for predicting virus evolution trends based on protein language models, structures, and immune recognition mechanisms. Background Art

[0002] Existing techniques for predicting viral evolutionary trends include using pre-trained language models to predict protein adaptability. However, these methods suffer from numerous drawbacks. First, they require tedious multiple sequence alignments of the protein sequences to be predicted, providing protein homology information as input to infer evolutionary coupling information and local representations of homologous protein sequences. This information is then mixed with protein amino acid information as input, resulting in results that are dependent on the quality of the multiple sequence alignment. Furthermore, the model cannot handle proteins with poor homology, inevitably leading to reduced prediction accuracy. Second, the ESM-1b model developed by Facebook used in this technique is inferior to the subsequently upgraded ESM-2 model in terms of both model parameter count and accuracy, and has not been compared with other similar language models. Third, this model does not fully account for factors influencing viral evolution, such as immune escape and structural information. Fourth, when using the ESM-1b model, the original model parameters are not retrained and adjusted for specific data scenarios. Furthermore, ESM models are not trained for viral evolution prediction, so prediction accuracy cannot be guaranteed. Fifth, the model can only predict the fitness of known proteins, and cannot generate completely new protein sequences and predict protein evolution trends.

[0003] The protein fitness prediction method in the existing technology (Chinese patent application number CN202211290871.2, which discloses "a protein fitness prediction method based on deep learning") can predict the fitness of proteins, but in actual use, the considerations are not comprehensive enough, the prediction accuracy cannot be guaranteed, and it cannot be used for evolutionary prediction. Summary of the Invention

[0004] In view of the defects in the prior art, the purpose of the present invention is to provide a method and system for predicting virus evolution trends.

[0005] The method for predicting virus evolution trends provided by the present invention includes:

[0006] Step 1: Obtain viral protein-target protein pairs with affinity annotations based on deep mutation scanning in the virus database, divide them into training sets, validation sets, and test sets, and collect antibody-antigen complex structure data corresponding to different viral proteins;

[0007] Step 2: Build a data encoder to encode the data of the protein language model;

[0008] Step 3: Perform layer-by-layer parameter training on the protein language model in the data encoder for downstream tasks;

[0009] Step 4: Evaluate the predicted affinity and experimental affinity using the correlation coefficient R2 on the training set and validation set, select the optimal model, and finally perform prediction verification on the test set;

[0010] Step 5: Calculate the structure-based immune escape score and the language model-based immune escape score;

[0011] Step 6: Perform multiple rounds of viral fitness calculations and genetic variation iterations to screen out potential preset high-risk strains.

[0012] Preferably, the step 2 comprises:

[0013] Use a tokenizer to convert sequence information into model-readable vector information, and add special character markers cls and eos to the beginning and end of the input protein sequence;

[0014] The protein amino acid sequence information in the training, validation, and test sets is fed into a Transformer model with a multi-head attention mechanism. After passing through the hidden layer, a 1280-dimensional vector is output for each amino acid. While considering the overall information of the sequence, the vector sum for each amino acid site is averaged to extract the hidden protein information in the sequence.

[0015] The Transformer model first pre-trains sequences from the large database Uniref50 through masked language modeling. After training, it extracts the attention map, performs two operations, symmetrization and mean product correction, to convert the attention map into the required form, and then performs the regression task, applying logistic regression with L-1 regularization independently on each amino acid pair.

[0016] Preferably, the step 3 comprises:

[0017] The downstream task involves converting the 1280-character feature vectors representing the protein amino acid sequence information extracted by the data encoder into their corresponding affinity prediction values through a three-layer convolutional neural network. In the three-layer convolutional neural network, each convolutional layer uses a convolution kernel of size 3 to learn the relationship between adjacent amino acid features in the same sequence. The results are then output as affinity values required for experimental analysis through a fully connected layer of 512 neurons.

[0018] Model training uses mean squared error loss to calculate the error between the model-predicted affinity value and the actual affinity. To prevent overfitting, Adam is used as the optimizer for training the entire model, and a linear scheduler with preheating is used to adjust the learning rate.

[0019] To effectively retrain the protein language model, the following strategy is adopted: first, specific layers of the model are unfrozen while the rest of the layers are frozen, and trained together with the entire affinity prediction model; then, several layers after the Xth layer are unfrozen while the first X layers of the frozen protein language model are used for parameter training.

[0020] Preferably, the step 5 comprises:

[0021] The structure-based immune escape score is based on the antibody recognition site to establish a virus immune escape prediction model. Four levels are designed based on the frequency of antibody recognition, with different weights applied to each level. The immune escape score ΔE is calculated by calculating the input viral amino acid mutations. The expression is:

[0022]

[0023] Where: N mut represents the total number of mutation sites; i represents each mutation site; w i Indicates the weight corresponding to each mutation site;

[0024] The structure-based immune escape score uses a protein language model to calculate the semantic changes in the viral antigens relative to the original virus. The algorithm calculates the protein language changes of each new mutant strain relative to the original strain, and then calculates the L-1 norm of this change as the language model-based immune escape score ΔZ. In terms of the numerical scaling strategy, after parameter optimization and adjustment, this value is divided by 100 for use. The expression is:

[0025]

[0026] Where: Z represents the protein embedding vector of the antigen; Z ref Protein embedding vector representing wild-type antigen; S z Represents the scaling parameter.

[0027] Preferably, step 6 includes:

[0028] Virus fitness is a score based on target affinity and immune evasion score, calculated as follows:

[0029] fitness=bind_score+escape_score

[0030] escape_score = ΔE + ΔZ

[0031] Among them: fitness represents the calculated virus adaptability; bind_score represents the virus affinity; escape_score represents the virus immune escape ability.

[0032] The virus evolution trend prediction system provided by the present invention includes:

[0033] Module M1: Obtain viral protein-target protein pairs with affinity annotations based on deep mutation scanning in the virus database, divide them into training sets, validation sets, and test sets, and collect antibody-antigen complex structural data corresponding to different viral proteins;

[0034] Module M2: Build a data encoder to encode data for the protein language model;

[0035] Module M3: Perform layer-by-layer parameter training on the protein language model in the data encoder for downstream tasks;

[0036] Module M4: Evaluate the predicted affinity and experimental affinity using the correlation coefficient R2 on the training set and validation set, select the optimal model, and finally perform prediction verification on the test set;

[0037] Module M5: Calculate the structure-based immune escape score and the language model-based immune escape score;

[0038] Module M6: Perform multiple rounds of viral fitness calculations and genetic mutation iterations to screen out potential preset high-risk strains.

[0039] Preferably, the module M2 includes:

[0040] Use a tokenizer to convert sequence information into model-readable vector information, and add special character markers cls and eos to the beginning and end of the input protein sequence;

[0041] The protein amino acid sequence information in the training, validation, and test sets is fed into a Transformer model with a multi-head attention mechanism. After passing through the hidden layer, a 1280-dimensional vector is output for each amino acid. While considering the overall information of the sequence, the vector sum for each amino acid site is averaged to extract the hidden protein information in the sequence.

[0042] The Transformer model first pre-trains sequences from the large database Uniref50 through masked language modeling. After training, it extracts the attention map, performs two operations, symmetrization and mean product correction, to convert the attention map into the required form, and then performs the regression task, applying logistic regression with L-1 regularization independently on each amino acid pair.

[0043] Preferably, the module M3 includes:

[0044] The downstream task involves converting the 1280-character feature vectors representing the protein amino acid sequence information extracted by the data encoder into their corresponding affinity prediction values through a three-layer convolutional neural network. In the three-layer convolutional neural network, each convolutional layer uses a convolution kernel of size 3 to learn the relationship between adjacent amino acid features in the same sequence. The results are then output as affinity values required for experimental analysis through a fully connected layer of 512 neurons.

[0045] Model training uses mean squared error loss to calculate the error between the model-predicted affinity value and the actual affinity. To prevent overfitting, Adam is used as the optimizer for training the entire model, and a linear scheduler with preheating is used to adjust the learning rate.

[0046] To effectively retrain the protein language model, the following strategy is adopted: first, specific layers of the model are unfrozen while the rest of the layers are frozen, and trained together with the entire affinity prediction model; then, several layers after the Xth layer are unfrozen while the first X layers of the frozen protein language model are used for parameter training.

[0047] Preferably, the module M5 includes:

[0048] The structure-based immune escape score is based on the antibody recognition site to establish a virus immune escape prediction model. Four levels are designed based on the frequency of antibody recognition, with different weights applied to each level. The immune escape score ΔE is calculated by calculating the input viral amino acid mutations. The expression is:

[0049]

[0050] Where: N mut represents the total number of mutation sites; i represents each mutation site; w i Indicates the weight corresponding to each mutation site;

[0051] The structure-based immune escape score uses a protein language model to calculate the semantic changes in the viral antigens relative to the original virus. The algorithm calculates the protein language changes of each new mutant strain relative to the original strain, and then calculates the L-1 norm of this change as the language model-based immune escape score ΔZ. In terms of the numerical scaling strategy, after parameter optimization and adjustment, this value is divided by 100 for use. The expression is:

[0052]

[0053] Where: Z represents the protein embedding vector of the antigen; Z ref Protein embedding vector representing wild-type antigen; S z Represents the scaling parameter.

[0054] Preferably, the module M6 includes:

[0055] Virus fitness is a score based on target affinity and immune evasion score, calculated as follows:

[0056] fitness=bind_score+escape_score

[0057] escape_score = ΔE + ΔZ

[0058] Among them: fitness represents the calculated virus adaptability; bind_score represents the virus affinity; escape_score represents the virus immune escape ability.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] (1) The present invention improves the accuracy of the zero-shot model in affinity prediction by retraining the protein language model under viral evolution data, and solves the problem of long time consumption, high cost and low throughput of deep mutation scanning experiments to determine affinity changes. The accuracy on the validation set reaches R 2 = 0.85 accuracy, and can measure the virus-target protein affinity of thousands of sites within 2 minutes, which is much faster than experimental and conventional molecular dynamics simulation-based computational methods;

[0061] (2) The present invention solves the problem of predicting the evolutionary trend of viruses by adopting a structure that simulates the evolutionary process of viruses. It can reveal the evolutionary pattern of viruses at the molecular level, effectively simulate the mutation rules of viruses and predict the effects of epidemic trends, and provide ideas and guidance for the advanced design of vaccines and drugs.

[0062] (3) The present invention solves the problem of predicting the evolutionary trend of viruses by adopting a structure that simulates the evolutionary process of viruses. It can reveal the evolutionary pattern of viruses at the molecular level, effectively simulate the mutation rules of viruses and predict the effects of epidemic trends, and provide ideas and guidance for the advanced design of vaccines and drugs. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0064] Figure 1 Simulate system diagrams for genetic evolution;

[0065] Figure 2 This is a flow chart of the virus evolution trend prediction method of the present invention. DETAILED DESCRIPTION

[0066] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0067] Example 1:

[0068] like Figure 2 The present invention provides a method for predicting virus evolution trends based on protein language models, structures, and immune recognition mechanisms, including:

[0069] Step 1: Obtain viral protein-target protein pairs with affinity annotations based on deep mutation scanning from the virus database and divide them into training, validation, and test sets; viruses include SARS-CoV-2, HIV, respiratory syncytial virus, etc. Collect antibody-antigen complex structural data corresponding to different viral proteins.

[0070] Step 2: Build a data encoder. This encoder: ① Uses a tokenizer to convert sequence information into model-readable vectors and adds the special character markers cls and eos to the beginning and end of the input protein sequence. ② Feeds the protein amino acid sequence information from the training, validation, and test sets into a Transformer model with a multi-head attention mechanism. After passing through the hidden layers, a 1280-dimensional vector is output for each amino acid. While considering the overall sequence information, the vector sums for each amino acid position are averaged to extract hidden protein information within the sequence. The Transformer model is first pre-trained on sequences from a large database (Uniref50) using masked language modeling (MLM). After training, the attention map is extracted and converted to the required form through symmetry and mean product correction. Regression is then performed, with logistic regression with L-1 regularization applied independently to each amino acid pair. Different protein language model architectures have different model parameters and parameter counts. For example, esm2_t33_650M_UR50D has 33 layers and 650 million parameters.

[0071] Step 3: The protein language model in the data encoder is trained layer by layer on the downstream task. The downstream task refers to converting the 1280-length feature vector representing the protein amino acid sequence information extracted by the data encoder into its corresponding affinity prediction value through a three-layer convolutional neural network (CNN). In the three-layer convolutional neural network, each convolution layer uses a convolution kernel of size 3 to learn the relationship between adjacent amino acid features in the same sequence, and then outputs the result as the affinity value required for experimental analysis through a fully connected layer of 512 neurons. The model training uses the mean square error loss (MSELoss) to calculate the error loss between the affinity value predicted by the model and the actual affinity. To prevent overfitting, Adam is used as the optimizer to train the entire model, and a linear scheduler with preheating is used to adjust the learning rate.

[0072] To effectively retrain the protein language model, the following strategy was adopted: first, specific layers of the model were unfrozen while the remaining layers were frozen, and trained together with the entire affinity prediction model; then, several layers after the Xth layer were unfrozen while the first X layers of the frozen protein language model were used for parameter training.

[0073] The protein language models mentioned above include the current mainstream ProBERT, ProTrans-T5, and ESM-2 models. Retraining was performed according to the above steps, and the accuracy of the model's affinity prediction was significantly improved after retraining.

[0074] Step 4: Retrain the model and evaluate the two indicators of predicted affinity and experimental affinity using the correlation coefficient R2 on the training set and validation set to select the optimal model, and finally perform prediction verification on the test set.

[0075] Step 5: Calculate the structure-based immune escape score and the language model-based immune escape score.

[0076] Among them, the structure-based immune escape score established an immune escape prediction model for viruses based on the antibody recognition site, and designed four levels according to the frequency of antibody recognition, with different weights, as shown in Table 1.

[0077] The immune escape score ΔE was calculated by calculating the amino acid mutations of the input virus.

[0078]

[0079] Table 1: Parameters of structure-based immune escape model

[0080]

[0081] The structure-based immune escape score uses a protein language model to calculate the semantic changes in the overall viral antigen relative to the original virus. The algorithm calculates the embedding change of each new mutant relative to the original strain, then calculates the L-1 norm of this change as the language model-based immune escape score ΔZ. This value is then divided by 100 for numerical scaling after parameter optimization.

[0082]

[0083] Step 6: In the genetic evolution module, complete the calculation of virus fitness and multiple rounds of genetic mutation to screen out potential high-risk strains. Virus fitness is a score based on target affinity and immune escape score, calculated as follows:

[0084] fitness=bind_score+escape_score

[0085] escape_score = ΔE + ΔZ

[0086] The specific evolutionary process is Figure 1As shown in the figure, a population is constructed based on the input virus strain. By predicting the target affinity and immune evasion capabilities of different strains, fitness differences are calculated, and "natural selection" is carried out within the population. Then, based on the principles of viral evolution in nature, gene recombination and mutation are used, with a recombination probability of 0.6 and a mutation probability of 0.1, and these parameters are adjustable. New viral "offspring" are generated, and after fitness calculations, a new round of viral populations is selected. Through multiple rounds of evolutionary simulations, viral evolutionary trends can be captured and verified, and future mutation directions can be predicted. Unlike multiple rounds of scanning based on single-point saturation mutations, this method exponentially reduces the sequence search space, enabling more efficient exploration of the boundaries of viral evolution.

[0087] Example 2:

[0088] The present invention also provides a virus evolution trend prediction system, which can be implemented by executing the process steps of the virus evolution trend prediction method. That is, those skilled in the art can understand the virus evolution trend prediction method as a preferred implementation of the virus evolution trend prediction system.

[0089] The virus evolution trend prediction system provided by the present invention includes: module M1: obtaining virus protein-target protein pairs with affinity annotations based on deep mutation scanning in the virus database, and dividing them into training set, verification set and test set, and collecting antibody-antigen complex structure data corresponding to different viral proteins; module M2: building a data encoder to encode data for the protein language model; module M3: performing layer-by-layer parameter training on the protein language model in the data encoder on downstream tasks; module M4: evaluating the two indicators of predicted affinity and experimental affinity with the correlation coefficient R2 on the training set and verification set, selecting the optimal model, and finally performing prediction verification on the test set; module M5: calculating the structure-based immune escape score and the language model-based immune escape score; module M6: performing multiple rounds of virus adaptability calculations and genetic variation iterations to screen out potential preset high-risk strains.

[0090] The module M2 includes:

[0091] Use a tokenizer to convert sequence information into model-readable vector information, and add special character markers cls and eos to the beginning and end of the input protein sequence;

[0092] The protein amino acid sequence information in the training, validation, and test sets is fed into a Transformer model with a multi-head attention mechanism. After passing through the hidden layer, a 1280-dimensional vector is output for each amino acid. While considering the overall information of the sequence, the vector sum for each amino acid site is averaged to extract the hidden protein information in the sequence.

[0093] The Transformer model first pre-trains sequences from the large database Uniref50 through masked language modeling. After training, it extracts the attention map, performs two operations, symmetrization and mean product correction, to convert the attention map into the required form, and then performs the regression task, applying logistic regression with L-1 regularization independently on each amino acid pair.

[0094] The module M3 includes:

[0095] The downstream task involves converting the 1280-character feature vectors representing the protein amino acid sequence information extracted by the data encoder into their corresponding affinity prediction values through a three-layer convolutional neural network. In the three-layer convolutional neural network, each convolutional layer uses a convolution kernel of size 3 to learn the relationship between adjacent amino acid features in the same sequence. The results are then output as affinity values required for experimental analysis through a fully connected layer of 512 neurons.

[0096] Model training uses mean squared error loss to calculate the error between the model-predicted affinity value and the actual affinity. To prevent overfitting, Adam is used as the optimizer for training the entire model, and a linear scheduler with preheating is used to adjust the learning rate.

[0097] To effectively retrain the protein language model, the following strategy is adopted: first, specific layers of the model are unfrozen while the rest of the layers are frozen, and trained together with the entire affinity prediction model; then, several layers after the Xth layer are unfrozen while the first X layers of the frozen protein language model are used for parameter training.

[0098] The module M5 includes:

[0099] The structure-based immune escape score is based on the antibody recognition site to establish a virus immune escape prediction model. Four levels are designed based on the frequency of antibody recognition, with different weights applied to each level. The immune escape score ΔE is calculated by calculating the input viral amino acid mutations. The expression is:

[0100]

[0101] Where: N mut represents the total number of mutation sites; i represents each mutation site; w i Indicates the weight corresponding to each mutation site;

[0102] The structure-based immune escape score uses a protein language model to calculate the semantic changes in the viral antigens relative to the original virus. The algorithm calculates the protein language changes of each new mutant strain relative to the original strain, and then calculates the L-1 norm of this change as the language model-based immune escape score ΔZ. In terms of the numerical scaling strategy, after parameter optimization and adjustment, this value is divided by 100 for use. The expression is:

[0103]

[0104] Where: Z represents the protein embedding vector of the antigen; Z ref Protein embedding vector representing wild-type antigen; S z Represents the scaling parameter.

[0105] The module M6 includes:

[0106] Virus fitness is a score based on target affinity and immune evasion score, calculated as follows:

[0107] fitness=bind_score+escape_score

[0108] escape_score = ΔE + ΔZ

[0109] Among them: fitness represents the calculated virus adaptability; bind_score represents the virus affinity; escape_score represents the virus immune escape ability.

[0110] Those skilled in the art will appreciate that, in addition to implementing the system, device, and various modules provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, embedded microcontrollers, and the like by logically programming the method steps. Therefore, the system, device, and various modules provided by the present invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; the modules for implementing various functions can also be considered both software programs for implementing the method and structures within the hardware component.

[0111] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. A method for predicting virus evolution trends, characterized in that: include: Step 1: Obtain viral protein-target protein pairs with affinity annotations based on deep mutation scanning in the virus database, divide them into training sets, validation sets, and test sets, and collect antibody-antigen complex structure data corresponding to different viral proteins; Step 2: Build a data encoder to encode the data of the protein language model; Step 3: Perform layer-by-layer parameter training on the protein language model in the data encoder for downstream tasks; The downstream task involves converting the 1280-character feature vectors representing the protein amino acid sequence information extracted by the data encoder into their corresponding affinity prediction values through a three-layer convolutional neural network. In the three-layer convolutional neural network, each convolutional layer uses a convolution kernel of size 3 to learn the relationship between adjacent amino acid features in the same sequence. The results are then output as affinity values required for experimental analysis through a fully connected layer of 512 neurons. Step 4: Evaluate the predicted affinity and experimental affinity using the correlation coefficient R2 on the training set and validation set, select the optimal model, and finally perform prediction verification on the test set; Step 5: Calculate the structure-based immune escape score and the language model-based immune escape score; Step 6: Perform multiple rounds of viral fitness calculations and genetic mutation iterations to screen out potential high-risk strains; The step 5 comprises: The structure-based immune escape score is based on the antibody recognition site to establish a virus immune escape prediction model. Four levels are designed based on the frequency of antibody recognition, with different weights applied to each level. The immune escape score ΔE is calculated by calculating the input viral amino acid mutations. The expression is: Where: N mut represents the total number of mutation sites; i represents each mutation site; w i Indicates the weight corresponding to each mutation site; The structure-based immune escape score uses a protein language model to calculate the semantic changes in the viral antigens relative to the original virus. The algorithm calculates the protein language changes of each new mutant strain relative to the original strain, and then calculates the L-1 norm of this change as the language model-based immune escape score ΔZ. In terms of the numerical scaling strategy, after parameter optimization and adjustment, this value is divided by 100 for use. The expression is: Where: Z represents the protein embedding vector of the antigen; Z ref Protein embedding vector representing wild-type antigen; S z Represents the scaling parameter.

2. The method for predicting virus evolution trends according to claim 1, wherein: The step 2 includes: Use a tokenizer to convert sequence information into model-readable vector information, and add special character markers cls and eos to the beginning and end of the input protein sequence; The protein amino acid sequence information in the training, validation, and test sets is fed into a Transformer model with a multi-head attention mechanism. After passing through the hidden layer, a 1280-dimensional vector is output for each amino acid. While considering the overall information of the sequence, the vector sum for each amino acid site is averaged to extract the hidden protein information in the sequence. The Transformer model first pre-trains sequences from the large database Uniref50 through masked language modeling. After training, it extracts the attention map, performs two operations, symmetrization and mean product correction, to convert the attention map into the required form, and then performs the regression task, applying logistic regression with L-1 regularization independently on each amino acid pair.

3. The method for predicting virus evolution trends according to claim 1, wherein: The step 3 comprises: Model training uses mean squared error loss to calculate the error loss between the model's predicted affinity value and the actual affinity. Adam is used as the optimizer to train the entire model, and a linear scheduler with preheating is used to adjust the learning rate. To effectively retrain the protein language model, the following strategy is adopted: first, specific layers of the model are unfrozen while the rest of the layers are frozen, and trained together with the entire affinity prediction model; then, several layers after the Xth layer are unfrozen while the first X layers of the frozen protein language model are used for parameter training.

4. The method for predicting virus evolution trends according to claim 1, wherein: The step 6 comprises: Virus fitness is a score based on target affinity and immune evasion score, calculated as follows: fitness=bind_score+escape_score escape_score = ΔE + ΔZ Among them: fitness represents the calculated virus adaptability; bind_score represents the virus affinity; escape_score represents the virus immune escape ability.

5. A virus evolution trend prediction system, characterized in that: include: Module M1: Obtain viral protein-target protein pairs with affinity annotations based on deep mutation scanning in the virus database, divide them into training sets, validation sets, and test sets, and collect antibody-antigen complex structural data corresponding to different viral proteins; Module M2: Build a data encoder to encode data for the protein language model; Module M3: Perform layer-by-layer parameter training on the protein language model in the data encoder for downstream tasks; The downstream task involves converting the 1280-character feature vectors representing the protein amino acid sequence information extracted by the data encoder into their corresponding affinity prediction values through a three-layer convolutional neural network. In the three-layer convolutional neural network, each convolutional layer uses a convolution kernel of size 3 to learn the relationship between adjacent amino acid features in the same sequence. The results are then output as affinity values required for experimental analysis through a fully connected layer of 512 neurons. Module M4: Evaluate the predicted affinity and experimental affinity using the correlation coefficient R2 on the training set and validation set, select the optimal model, and finally perform prediction verification on the test set; Module M5: Calculate the structure-based immune escape score and the language model-based immune escape score; Module M6: Perform viral fitness calculations and multiple rounds of genetic variation iterations to screen out potential high-risk strains; The module M5 includes: The structure-based immune escape score is based on the antibody recognition site to establish a virus immune escape prediction model. Four levels are designed based on the frequency of antibody recognition, with different weights applied to each level. The immune escape score ΔE is calculated by calculating the input viral amino acid mutations. The expression is: Where: N mut represents the total number of mutation sites; i represents each mutation site; w i Indicates the weight corresponding to each mutation site; The structure-based immune escape score uses a protein language model to calculate the semantic changes in the viral antigens relative to the original virus. The algorithm calculates the protein language changes of each new mutant strain relative to the original strain, and then calculates the L-1 norm of this change as the language model-based immune escape score ΔZ. In terms of the numerical scaling strategy, after parameter optimization and adjustment, this value is divided by 100 for use. The expression is: Where: Z represents the protein embedding vector of the antigen; Z ref Protein embedding vector representing wild-type antigen; S z Represents the scaling parameter.

6. The virus evolution trend prediction system according to claim 5, characterized in that: The module M2 includes: Use a tokenizer to convert sequence information into model-readable vector information, and add special character markers cls and eos to the beginning and end of the input protein sequence; The protein amino acid sequence information in the training, validation, and test sets is fed into a Transformer model with a multi-head attention mechanism. After passing through the hidden layer, a 1280-dimensional vector is output for each amino acid. While considering the overall information of the sequence, the vector sum for each amino acid site is averaged to extract the hidden protein information in the sequence. The Transformer model first pre-trains sequences from the large database Uniref50 through masked language modeling. After training, it extracts the attention map, performs two operations, symmetrization and mean product correction, to convert the attention map into the required form, and then performs the regression task, applying logistic regression with L-1 regularization independently on each amino acid pair.

7. The virus evolution trend prediction system according to claim 5, characterized in that: The module M3 includes: Model training uses mean squared error loss to calculate the error loss between the model's predicted affinity value and the actual affinity. Adam is used as the optimizer to train the entire model, and a linear scheduler with preheating is used to adjust the learning rate. To effectively retrain the protein language model, the following strategy is adopted: first, specific layers of the model are unfrozen while the rest of the layers are frozen, and trained together with the entire affinity prediction model; then, several layers after the Xth layer are unfrozen while the first X layers of the frozen protein language model are used for parameter training.

8. The virus evolution trend prediction system according to claim 5, characterized in that: The module M6 includes: Virus fitness is a score based on target affinity and immune evasion score, calculated as follows: fitness=bind_score+escape_score escape_score = ΔE + ΔZ Among them: fitness represents the calculated virus adaptability; bind_score represents the virus affinity; escape_score represents the virus immune escape ability.

Images