Protein mutation stability prediction method based on generative adversarial network data generation

Abstract

The present invention discloses a protein mutation stability prediction method based on generative adversarial network data generation, which belongs to the field of bioinformatics technology and includes the following steps: constructing a protein mutation data set; extracting protein features from a Megascale data set and a protein mutation data set; performing a five-fold cross validation on the generative adversarial network and saving the generator G; generating mutation samples through the generator G; constructing a downstream task network model; optimizing the downstream task network model parameters through an adaptive gradient method; verifying the model performance through a validation set calculation evaluation index; and realizing prediction through a downstream task network model. The present invention applies a generative adversarial network to generate protein mutation samples, which effectively solves the problem of data scarcity in protein stability research; the generative adversarial mechanism combines the cycle consistency loss to ensure the high quality of the generated data, successfully expand the data set, and improve the model's adaptability to different mutation types.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a protein mutation stability prediction method based on generative adversarial network data generation. Background Art

[0002] Missense mutations often affect protein function, affecting protein expression, polypeptide chain folding, interactions with other proteins, and changes in catalytic processes. In addition, it may lead to disease occurrence and drug resistance. Due to the association between mutations and diseases, predicting the effects of mutations on proteins has become a focus of attention, which usually involves evaluating changes in thermodynamic stability. Therefore, accurately evaluating the changes in thermodynamic stability caused by mutations can help to gain a deeper understanding of the effects of mutations on protein function.

[0003] Researchers have invested a lot of effort in developing computational methods for predicting changes in thermodynamic stability of mutations. These methods can be roughly divided into biophysical methods and deep learning methods. Biophysical methods are usually based on biophysical free energy functions derived from molecular mechanics, rely on accurate biophysical modeling of amino acid interactions, and combine with conformational sampling of protein structures, but such methods are usually computationally intensive and have low throughput. Deep learning methods reveal potential laws that may be overlooked by biophysical methods by extracting multiple patterns of amino acid sequences, evolutionary information, and protein structural features, significantly reducing time costs and greatly improving the flexibility of predictions.

[0004] However, the effectiveness and accuracy of deep learning methods are highly dependent on large-scale, high-quality data. During the training process of deep learning models, a large amount of data is used to capture complex feature patterns and distribution information to ensure the generalization ability of the model. Therefore, for the prediction of changes in thermodynamic stability of protein mutations, it is crucial to have sufficient protein stability data, which not only helps the model accurately identify the effects of mutations on protein stability and function, but also ensures that the model has good performance under a wider range of mutation types and environments. Due to the high demand for experimental resources, diverse mutation types and expensive measurement costs, the accumulation of protein stability data is slow, resulting in relatively scarce data resources. Therefore, how to generate high-quality protein stability data to give deep learning models more powerful generalization capabilities is still a difficult problem that needs to be solved urgently. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a protein mutation stability prediction method based on generative adversarial network data with high prediction accuracy and good robustness.

[0006] The technical solution of the present invention to solve the above technical problem is: a protein mutation stability prediction method based on generative adversarial network data, comprising the following steps:

[0007] S1: Manual construction of protein mutation dataset;

[0008] S2: extracting protein features from the Megascale dataset and the protein mutation dataset constructed in step S1 respectively to obtain a Megascale feature dataset and a protein mutation feature dataset, and dividing the protein mutation feature dataset into a training set, a validation set, and a test set;

[0009] S3: Use the Megascale feature dataset obtained in step S2 to perform a five-fold cross validation on the generative adversarial network, and save the generator G after training;

[0010] S4: Generate mutation samples through the generator G, and combine the generated mutation samples with the training set and validation set in step S2 to obtain new training set and validation set;

[0011] S5: Build a downstream task network model, and use the new training set obtained in step S4 to train the downstream task network model;

[0012] S6: Use the mean square error loss function to obtain the overall loss value of the downstream task network model, and use the adaptive gradient method to optimize the parameters of the downstream task network model;

[0013] S7: Input the new validation set obtained in step S4 into the trained downstream task network model, and use the Pearson correlation coefficient PCC and root mean square error RMSE to calculate the prediction performance of the validation set;

[0014] S8: Input the test set in step S2 into the downstream task network model to predict the stability of protein mutations.

[0015] In the above-mentioned protein mutation stability prediction method based on generative adversarial network data generation, the specific process of step S1 is as follows:

[0016] S11: Several publicly available protein mutation stability datasets were selected as experimental datasets;

[0017] S12: Remove duplicate samples in the experimental dataset.

[0018] In the above-mentioned protein mutation stability prediction method based on generative adversarial network data generation, in step S2, for the samples in the Megascale data set, the mutated amino acid and its surrounding 16 amino acids are selected, and after extracting the structure and sequence information of all amino acids, the mutated amino acid is used as a node, and the interactions between the surrounding 16 amino acids are used as edges, and complete graphs of the samples before and after the mutation are constructed respectively, thereby obtaining the Megascale feature data set.

[0019] In the above-mentioned protein mutation stability prediction method based on generative adversarial network data generation, in step S2, for samples in the protein mutation data set, protein features are extracted from three levels: atoms, amino acids, and surfaces, thereby obtaining a protein mutation feature data set;

[0020] At the amino acid level, the mutated amino acid and its surrounding 16 amino acids were selected. After extracting the structural and sequence information of all amino acids, the mutated amino acid was used as a node, and the interactions between the surrounding 16 amino acids were used as edges to construct the complete graphs of the samples before and after the mutation.

[0021] At the atomic level, the protein structure is viewed as a collection of voxels in three-dimensional space, collecting Voxel grid, each voxel is divided according to the formula The seven chemical properties are parameterized using attribute channels of seven different chemical properties, namely hydrophobicity, aromaticity, hydrogen bond acceptor, hydrogen bond donor, positive ionization, negative ionization, and occupancy, and finally a A tensor of represents the van der Waals radius, represents the distance from the atom to the voxel center, Indicates the saturation of each voxel;

[0022] At the surface level, selection is based on the mutation site The coordinates of atoms within 5 angstroms centered on the carbon atom were used to identify the mutation site using the molecular surface interaction fingerprint MaSIF. The carbon atom and all the atoms within 5 angstroms around it are triangulated, and the shape index, curvature, hydrogen bonding potential, charge and hydrophobicity characteristics of each surface point are calculated. Then, the surface points are projected onto a two-dimensional plane through the multidimensional scaling MDS algorithm to obtain a set of two-dimensional images, in which each image corresponds to a geometric or biochemical property, and the pixel intensity represents the characteristic value;

[0023] The protein mutation feature dataset was divided into training set, validation set and test set in a ratio of 0.8:0.1:0.1, ensuring that there was no protein from the same protein in different sets.

[0024] In the above-mentioned protein mutation stability prediction method based on generative adversarial network data generation, the generative adversarial network in step S3 includes two generators G and F and two discriminators D X , D Y , both extract information through adaptive feature propagation mechanism; firstly, the discriminator D is trained using the Megascale dataset X and D Y , so that D X and D YIt can judge whether it is a real sample, and then train the generator G and F, so that the generator G converts the feature sample before the mutation into the feature sample after the mutation according to the label information, and generates the stability change value caused by the mutation, thereby deceiving the discriminator D Y , the generator F converts the mutated feature samples into the feature samples before mutation according to the label information, thereby deceiving the discriminator D X .

[0025] In the above-mentioned protein mutation stability prediction method based on generative adversarial network data generation, in step S3, a loss function is used to constrain the generative adversarial network training process, specifically:

[0026] Using the formula Optimizing the Discriminator and performance; among them, represents the loss of the discriminator, Represents the distribution of real sample data Calculate the expected value of the sample y in Represents the distribution of real sample data Calculate the expected value of the sample x in Represents sample y as the discriminator The probability of being judged as real data, Represents sample x as the discriminator The probability of being judged as real data;

[0027] Using the formula Ensure that the samples generated by the generator G are real and meet the conditional labels; represents the loss of the generator G, Indicates that the generator G receives input features and conditional tags Generate mutated feature samples and conditional labels Includes the mutation position and the amino acid type after mutation;

[0028] Using the formula Ensure that the samples generated by generator F are real and meet the conditional labels; represents the loss of generator F, Indicates that the generator F receives input features and conditional tags Generate feature samples before mutation;

[0029] Using the formula Ensure that the samples after the generator are unchanged from the input feature samples; represents the cycle consistency loss, and Represents the cyclic conversion process, represents the L1 norm, that is, the sum of the absolute values ​​of the elements. The final loss combines the conditional adversarial loss and the cycle consistency loss. The conditional adversarial loss is the sum of the loss of the generator G and the loss of the generator F. The final loss formula is:

[0030] ;

[0031] in, is a weight parameter used to balance the impact of conditional adversarial loss and cycle consistency loss.

[0032] In the above-mentioned protein mutation stability prediction method based on generative adversarial network data generation, in step S3, the adaptive feature propagation mechanism of the generative adversarial network is specifically:

[0033] Using the formula , the node's feature representation is updated by its neighbor nodes and adjacent edges, the context association of the node features is strengthened, and the interaction relationship between amino acids is captured; among them, is the adjacent edge at level l-1 The characteristics of, N(i) represents the set of neighbor nodes of node i, Represents the l-1th layer node Features, represents the weight mapping function, Represents the features of node i in layer l;

[0034] Using the formula , the feature representation of the edge is updated by the features of the neighboring edges to strengthen the consistency of the local and global structures of the protein; where N(j) represents the set of neighboring nodes of node j, is the l-1th layer node To Node The characteristics of the edge ik, is the l-1th layer node The characteristics of edge kj to node j, Represents element-wise product.

[0035] In the above protein mutation stability prediction method based on generative adversarial network data generation, in step S3, the discriminator further performs the following operations in order to discriminate the authenticity of the sample and obtain the final confidence score:

[0036] Using the formula Capture long-range dependencies between amino acids; represents the amino acid feature representation after the adaptive feature propagation mechanism, represents the attention mechanism, represents the feature representation of the output of the t-th attention head, , , are the mapping matrices of query, key, and value respectively;

[0037] Using the formula Capture short-range dependencies between amino acids; Indicates the use of one-dimensional convolution processing, represents the learnable matrix, represents the final learned feature embedding, and T represents the total number of attention heads;

[0038] Using the formula , after the first fully connected layer and the second fully connected layer, we get a The output vector of is used to measure the possibility that the sample is a true sample; among them, and denote the learnable matrices of the first fully connected layer and the second fully connected layer, respectively. and Represent the bias values ​​of the first fully connected layer and the second fully connected layer respectively, represents the activation function, Indicates the possibility that the sample is a true sample.

[0039] In the above-mentioned protein mutation stability prediction method based on generative adversarial network data generation, the specific process of step S4 is as follows:

[0040] S41: Count all proteins included in the training set and validation set divided in step S1, obtain the situation where all amino acids on all proteins mutate into another 19 amino acids, exclude the mutation samples in the training set and validation set, obtain the training set and validation set without mutation, input the training set, validation set and mutation label without mutation into the generator G, generate the corresponding post-mutation structure diagram and the stability change value caused by the mutation;

[0041] S42: The training set without mutation, the validation set and the generated structure diagram after mutation are merged. After the merger, each sample in the training set and the validation set has two labels, which are the stability change value caused by the mutation and whether it is a mutation sample generated by the generator G.

[0042] In the above-mentioned protein mutation stability prediction method based on generative adversarial network data generation, in step S5, the specific process of constructing the downstream task network model is:

[0043] S51: Using formula Capture the spatial information of the mutated atom and its surrounding atoms; represents the characteristic matrix of an atom, represents a three-dimensional convolution, represents the atoms after feature extraction;

[0044] S52: Using adaptive feature propagation mechanism to capture the interaction relationship between amino acids;

[0045] S53: At the surface level, the obtained multi-channel image is positionally encoded and then used using the formula Extract surface features; where , , , Q, K, and V represent the query, key, and value matrices in the process of surface information feature extraction, respectively. represents the characteristic matrix of the surface, Indicates the dimension of the key, used for scaling factors, For attention weight normalization, Represents the surface information after feature extraction, , , denote the query, key, and value learnable matrices respectively;

[0046] S54: Generate embeddings of mutation sites using large language models to capture deep features in protein sequence and structure;

[0047] S55: Utilizing formulas Captures the interrelationships between amino acid, atomic, and surface-level features; , , , , , represent the query, key, and value matrices in the process of fusing amino acid, atom, and surface level information, respectively. , , Respectively represent the amino acid, atom, and surface query feature matrices after feature extraction, , , Respectively represent the amino acid, atom, and surface bond feature matrices after feature extraction, , , Respectively represent the amino acid, atom, and surface value feature matrices after feature extraction, represents the matrix concatenation operation, It represents the feature matrix after feature fusion of amino acid, atom and surface information;

[0048] S56: Using formula The information from the three levels is effectively integrated with the mutation embedding generated by the large language model; represents the embedding of mutation sites generated by large language models, represents the activation function, represents the fused embedded representation;

[0049] S57: Using formula After the third and fourth fully connected layers, we get The output vector , which is the stability change value caused by the mutation and the confidence that the sample is a generated sample; where, and Represent the learnable matrices of the third and fourth fully connected layers, respectively. and Represents the bias values ​​of the third and fourth fully connected layers.

[0050] The beneficial effects of the present invention are:

[0051] 1. The present invention innovatively applies generative adversarial networks to generate protein mutation samples, effectively solving the problem of data scarcity in protein stability research. The generator of the generative adversarial network can generate characteristic samples of mutated proteins based on mutation positions and amino acid type labels, while the discriminator ensures that the generated samples are realistic and meet the conditional labels. This generative adversarial mechanism combined with cycle consistency loss ensures the high quality of generated data, successfully expands the data set, and improves the model's adaptability to different mutation types.

[0052] 2. The present invention uses the generated mutation data for downstream task training, extracts potential features from three levels: atoms, amino acids, and surfaces, and generates embeddings corresponding to mutation sites through a large language model during fusion to enrich protein information, effectively improving the accuracy of stability change predictions. This strategy of combining data generated by generative adversarial networks with downstream tasks significantly improves the accuracy and robustness of protein mutation stability predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of the present invention.

[0054] Figure 2 Schematic diagram of the structure of the generative adversarial network in the present invention.

[0055] Figure 3 Schematic diagram of node update of the adaptive feature propagation mechanism of the generative adversarial network in the present invention.

[0056] Figure 4 Schematic diagram of edge update of the adaptive feature propagation mechanism of the generative adversarial network in the present invention.

[0057] Figure 5This is a schematic diagram of data expansion in the present invention.

[0058] Figure 6 Schematic diagram of the structure of the downstream task network model in the present invention. DETAILED DESCRIPTION

[0059] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0060] like Figure 1 As shown, a protein mutation stability prediction method based on generative adversarial network data generation includes the following steps:

[0061] S1: Manual construction of protein mutation dataset.

[0062] The specific process of step S1 is:

[0063] S11: 14 publicly available and widely used protein mutation stability datasets (S1784, S2156, S339, S625, S1615, S388, S2648, S1626, S3421, S306, S605, S2204, S1948, S1791) were selected as experimental datasets.

[0064] S12: Remove duplicate samples from the experimental dataset.

[0065] S2: Extract protein features from the Megascale dataset and the protein mutation dataset constructed in step S1 respectively to obtain a Megascale feature dataset and a protein mutation feature dataset, and divide the protein mutation feature dataset into a training set, a validation set, and a test set.

[0066] In step S2, for the samples in the Megascale data set, the mutated amino acid and the 16 amino acids around it are selected, and after the structural and sequence information of all amino acids is extracted, the mutated amino acid is used as a node, and the interaction between the surrounding 16 amino acids is used as an edge to construct a complete graph of the samples before and after the mutation, thereby obtaining a Megascale feature data set. It should be noted that for the Megascale data set, only the information at the amino acid level is extracted, and the protein structure after the mutation is generated using foldx, thereby constructing the 16 amino acid interaction graph around the mutation, and an additional label is included in the node embedding as a feature, the label before the mutation is set to 0 by default, and the label after the mutation is set to the stability change value.

[0067] For samples in the protein mutation dataset, protein features are extracted from three levels: atoms, amino acids, and surfaces, thus obtaining a protein mutation feature dataset.

[0068] At the amino acid level, the mutated amino acid and its surrounding 16 amino acids were selected. After extracting the structure and sequence information of all amino acids, the mutated amino acid was used as a node and the interactions between the surrounding 16 amino acids were used as edges to construct the complete graphs of the samples before and after mutation.

[0069] At the atomic level, the protein structure is viewed as a collection of voxels in three-dimensional space, collecting Voxel grid, each voxel is divided according to the formula The seven chemical properties are parameterized using attribute channels of seven different chemical properties, namely hydrophobicity, aromaticity, hydrogen bond acceptor, hydrogen bond donor, positive ionization, negative ionization, and occupancy, and finally a A tensor of represents the van der Waals radius, represents the distance from the atom to the voxel center, Indicates the saturation of each voxel.

[0070] At the surface level, selection is based on the mutation site The coordinates of atoms within 5 angstroms centered on the carbon atom were used to identify the mutation site using the molecular surface interaction fingerprint MaSIF. The carbon atom and all the atoms within 5 angstroms around it are triangulated, and the shape index, curvature, hydrogen bonding potential, charge and hydrophobicity characteristics of each surface point are calculated. The surface points are then projected onto a two-dimensional plane using the multidimensional scaling (MDS) algorithm to obtain a set of two-dimensional images, where each image corresponds to a geometric or biochemical property and the pixel intensity represents the characteristic value.

[0071] The protein mutation feature dataset was divided into training set, validation set and test set in a ratio of 0.8:0.1:0.1, ensuring that there was no protein from the same protein in different sets.

[0072] S3: Use the Megascale feature dataset obtained in step S2 to perform a five-fold cross-validation on the generative adversarial network. When the number of iterations reaches the set value of 150, stop the training process and save the generator G after training.

[0073] like Figure 2 As shown, the generative adversarial network consists of two generators G, F and two discriminators D X , D Y , both extract information through adaptive feature propagation mechanism; firstly, the discriminator D is trained using the Megascale dataset X and D Y , so that D X and D YIt can judge whether it is a real sample, and then train the generator G and F, so that the generator G converts the feature sample before the mutation into the feature sample after the mutation according to the label information, and generates the stability change value caused by the mutation, thereby deceiving the discriminator D Y , the generator F converts the mutated feature samples into the feature samples before mutation according to the label information, thereby deceiving the discriminator D X .

[0074] The loss function is used to constrain the training process of the generative adversarial network, specifically:

[0075] Using the formula Optimizing the Discriminator and performance; among them, represents the loss of the discriminator, Represents the distribution of real sample data Calculate the expected value of the sample y in Represents the distribution of real sample data Calculate the expected value of the sample x in Represents sample y as the discriminator The probability of being judged as real data, Represents sample x as the discriminator The probability of being judged as real data;

[0076] Using the formula Ensure that the samples generated by the generator G are real and meet the conditional labels; represents the loss of the generator G, Indicates that the generator G receives input features and conditional tags Generate mutated feature samples and conditional labels Includes the mutation position and the amino acid type after mutation;

[0077] Using the formula Ensure that the samples generated by generator F are real and meet the conditional labels; represents the loss of generator F, Indicates that the generator F receives input features and conditional tags Generate feature samples before mutation;

[0078] Using the formula Ensure that the samples after the generator are unchanged from the input feature samples; represents the cycle consistency loss, and Represents the cyclic conversion process, represents the L1 norm, that is, the sum of the absolute values ​​of the elements. The final loss combines the conditional adversarial loss and the cycle consistency loss. The conditional adversarial loss is the sum of the loss of the generator G and the loss of the generator F. The final loss formula is:

[0079] ;

[0080] in, is a weight parameter used to balance the impact of conditional adversarial loss and cycle consistency loss.

[0081] like Figure 3 , Figure 4 As shown in Figure 2, the adaptive feature propagation mechanism of the generative adversarial network is specifically:

[0082] Using the formula , the node's feature representation is updated by its neighbor nodes and adjacent edges, the context association of the node features is strengthened, and the interaction relationship between amino acids is captured; among them, is the adjacent edge at level l-1 The characteristics of, N(i) represents the set of neighbor nodes of node i, Represents the l-1th layer node Features, represents the weight mapping function, Represents the features of node i in layer l;

[0083] Using the formula , the feature representation of the edge is updated by the features of the neighboring edges to strengthen the consistency of the local and global structures of the protein; where N(j) represents the set of neighboring nodes of node j, is the l-1th layer node To Node The characteristics of the edge ik, is the l-1th layer node The characteristics of edge kj to node j, Represents element-wise product.

[0084] In order to determine the authenticity of the sample and obtain the final confidence score, the discriminator also performs the following operations:

[0085] Using the formula Capture long-range dependencies between amino acids; represents the amino acid feature representation after the adaptive feature propagation mechanism, represents the attention mechanism, represents the feature representation of the output of the t-th attention head, , , are the mapping matrices of query, key, and value respectively;

[0086] Using the formula Capture short-range dependencies between amino acids; Indicates the use of one-dimensional convolution processing, represents the learnable matrix, represents the final learned feature embedding, T represents the total number of attention heads; using the formula , after the first fully connected layer and the second fully connected layer, we get a The output vector of is used to measure the possibility that the sample is a true sample; among them, and denote the learnable matrices of the first fully connected layer and the second fully connected layer, respectively. and Represent the bias values ​​of the first fully connected layer and the second fully connected layer respectively, represents the activation function, Indicates the possibility that the sample is a true sample.

[0087] S4: Generate mutation samples through generator G, and combine the generated mutation samples with the training set and validation set in step S2 to obtain new training set and validation set.

[0088] like Figure 5 As shown, the specific process of step S4 is:

[0089] S41: Count all proteins included in the training set and validation set divided in step S1, and obtain the situation where all amino acids on all proteins mutate into another 19 amino acids. After excluding the mutation samples in the training set and validation set, the training set and validation set without mutation are obtained. The training set without mutation, validation set and mutation label are input into the generator G to generate the corresponding post-mutation structure diagram and the stability change value caused by the mutation.

[0090] S42: The training set without mutation, the validation set and the generated structure diagram after mutation are merged. After the merger, each sample in the training set and the validation set has two labels, which are the stability change value caused by the mutation and whether it is a mutation sample generated by the generator G.

[0091] S5: Build a downstream task network model and use the new training set obtained in step S4 to train the downstream task network model. Since voxels are used to represent the information around the mutation at the atomic level, 3D-CNN is used to extract the atomic information; and the amino acid level uses a graph to represent the interaction between amino acids, so the adaptive feature propagation mechanism is still used; ViT is used to extract the protein surface information after mapping into a two-dimensional image.

[0092] like Figure 6As shown in Figure 2, the specific process of building the downstream task network model is as follows:

[0093] S51: Using formula Capture the spatial information of the mutated atom and its surrounding atoms; represents the characteristic matrix of an atom, represents a three-dimensional convolution, represents the atoms after feature extraction;

[0094] S52: Using adaptive feature propagation mechanism to capture the interaction relationship between amino acids;

[0095] S53: At the surface level, the obtained multi-channel image is positionally encoded and then used using the formula Extract surface features; where , , , Q, K, and V represent the query, key, and value matrices in the process of surface information feature extraction, respectively. represents the characteristic matrix of the surface, represents the dimension of the key, used for scaling factors, For attention weight normalization, Represents the surface information after feature extraction, , , denote the query, key, and value learnable matrices respectively;

[0096] S54: Generate embeddings of mutation sites using large language models to capture deep features in protein sequence and structure;

[0097] S55: Utilizing formulas Captures the interrelationships between amino acid, atomic, and surface-level features; , , , , , represent the query, key, and value matrices in the process of fusing amino acid, atom, and surface level information, respectively. , , Respectively represent the amino acid, atom, and surface query feature matrices after feature extraction, , , Respectively represent the amino acid, atom, and surface bond feature matrices after feature extraction, , , Respectively represent the amino acid, atom, and surface value feature matrices after feature extraction, represents the matrix concatenation operation, It represents the feature matrix after feature fusion of amino acid, atom and surface information;

[0098] S56: Using formula The information from the three levels is effectively integrated with the mutation embedding generated by the large language model; represents the embedding of mutation sites generated by large language models, represents the activation function, represents the fused embedded representation;

[0099] S57: Using formula After the third and fourth fully connected layers, we get The output vector , which is the stability change value caused by the mutation and the confidence that the sample is a generated sample; where, and Represent the learnable matrices of the third and fourth fully connected layers, respectively. and Represents the bias values ​​of the third and fourth fully connected layers.

[0100] S6: Use the mean square error loss function to obtain the overall loss value of the downstream task network model, and use the adaptive gradient method to optimize the parameters of the downstream task network model.

[0101] S7: Input the new validation set obtained in step S4 into the trained downstream task network model, and use the Pearson correlation coefficient PCC and root mean square error RMSE to calculate the prediction performance of the validation set.

[0102] S8: Input the test set in step S2 into the downstream task network model to predict the stability of protein mutations.

Claims

1. A protein mutation stability prediction method based on generative adversarial network data generation, characterized in that: The following steps are involved: S1: Manual construction of protein mutation dataset; S2: extracting protein features from the Megascale dataset and the protein mutation dataset constructed in step S1 respectively to obtain a Megascale feature dataset and a protein mutation feature dataset, and dividing the protein mutation feature dataset into a training set, a validation set, and a test set; In step S2, for the samples in the Megascale data set, the mutated amino acid and the 16 amino acids around it are selected, and after the structure and sequence information of all the amino acids are extracted, the mutated amino acid is used as a node, and the interactions between the surrounding 16 amino acids are used as edges, and the complete graphs of the samples before and after the mutation are respectively constructed, thereby obtaining the Megascale feature data set; S3: Use the Megascale feature dataset obtained in step S2 to perform a five-fold cross validation on the generative adversarial network, and save the generator G after training; The generative adversarial network in step S3 includes two generators G and F and two discriminators D X , D Y , both extract information through adaptive feature propagation mechanism; firstly, the discriminator D is trained using the Megascale dataset X and D Y , so that D X and D Y It can judge whether it is a real sample, and then train the generator G and F, so that the generator G converts the feature sample before the mutation into the feature sample after the mutation according to the label information, and generates the stability change value caused by the mutation, thereby deceiving the discriminator D Y , the generator F converts the mutated feature samples into the feature samples before mutation according to the label information, thereby deceiving the discriminator D X ; S4: Generate mutation samples through the generator G, and combine the generated mutation samples with the training set and validation set in step S2 to obtain new training set and validation set; S5: Build a downstream task network model, and use the new training set obtained in step S4 to train the downstream task network model; S6: Use the mean square error loss function to obtain the overall loss value of the downstream task network model, and use the adaptive gradient method to optimize the parameters of the downstream task network model; S7: Input the new validation set obtained in step S4 into the trained downstream task network model, and use the Pearson correlation coefficient PCC and root mean square error RMSE to calculate the prediction performance of the validation set; S8: Input the test set in step S2 into the downstream task network model to predict the stability of protein mutations.

2. The protein mutation stability prediction method based on generative adversarial network data generation according to claim 1, characterized in that: The specific process of step S1 is: S11: Several publicly available protein mutation stability datasets were selected as experimental datasets; S12: Remove duplicate samples from the experimental dataset.

3. The protein mutation stability prediction method based on generative adversarial network data generation according to claim 1, characterized in that: In step S2, for samples in the protein mutation dataset, protein features are extracted from three levels: atoms, amino acids, and surfaces, thereby obtaining a protein mutation feature dataset; At the amino acid level, the mutated amino acid and its surrounding 16 amino acids were selected. After extracting the structural and sequence information of all amino acids, the mutated amino acid was used as a node, and the interactions between the surrounding 16 amino acids were used as edges to construct the complete graphs of the samples before and after the mutation. At the atomic level, the protein structure is viewed as a collection of voxels in three-dimensional space. A 16×16×16 voxel grid around the mutation site is collected and each voxel is transformed according to the formula The seven chemical properties are parameterized using attribute channels of different chemical properties, namely hydrophobicity, aromaticity, hydrogen bond acceptor, hydrogen bond donor, positive ionization, negative ionization, and occupancy, and finally a 16×16×16×7 tensor is obtained; where r vdw represents the van der Waals radius, d represents the distance from the atom to the voxel center, and f(d) represents the saturation of each voxel; At the surface level, the coordinates of atoms within 5 angstroms centered on the α carbon atom of the mutation site are selected, and the α carbon atom of the mutation site and all atoms within 5 angstroms around it are triangulated using the molecular surface interaction fingerprint MaSIF. The shape index, curvature, hydrogen bonding potential, charge and hydrophobicity characteristics of each surface point are calculated, and then the surface points are projected onto a two-dimensional plane through the multidimensional scaling MDS algorithm to obtain a set of two-dimensional images, in which each image corresponds to a geometric or biochemical property, and the pixel intensity represents the characteristic value; The protein mutation feature dataset was divided into training set, validation set and test set in a ratio of 0.8:0.1:0.1, ensuring that there was no protein from the same protein in different sets.

4. The protein mutation stability prediction method based on generative adversarial network data generation according to claim 1, characterized in that: In step S3, a loss function is used to constrain the training process of the generative adversarial network, specifically: Using the formula Optimize the discriminator D Y and D X performance; Among them, Loss D represents the loss of the discriminator, Represents the distribution of real sample data p data(y) Calculate the expected value of the sample y in Represents the distribution of real sample data p data(x) The expected value of the sample x in D is calculated. Y (y) represents the sample y discriminator D Y The probability of being judged as real data, D X (x) represents the sample x discriminator D X The probability of being judged as real data; Using the formula Ensure that the samples generated by the generator G are real and meet the conditional labels; among them, Loss GAN (G,D Y ,X,Y) represents the loss of the generator G, G(x|c) represents that the generator G receives the input feature x and the conditional label c to generate the mutated feature sample, and the conditional label c contains the mutation position and the mutated amino acid type; Using the formula Ensure that the samples generated by generator F are real and meet the conditional labels; among them, Loss GAN (G,D X ,Y,X) represents the loss of generator F, G(y|c′) represents the feature sample before the generator F receives the input feature y and the conditional label c′ to generate the mutation; Using the formula Ensure that the samples after the generator are unchanged from the input feature samples; among them, Loss Cycle (G, F) represents the cycle consistency loss, F(G(x|c)) and G(F(y|c)) represent the cycle conversion process, ||·||1 represents the L1 norm, that is, the sum of the absolute values ​​of the elements. The final loss combines the conditional adversarial loss and the cycle consistency loss. The conditional adversarial loss is the sum of the loss of generator G and the loss of generator F. The final loss formula is: Loss(G,F,D X ,D Y )=Loss GAN (G,D Y ,X,Y)+Loss GAN (G,D X ,Y,X)+λLoss Cycle (G,F) Among them, λ is a weight parameter used to balance the impact of conditional adversarial loss and cycle consistency loss.

5. The protein mutation stability prediction method based on generative adversarial network data generation according to claim 4, characterized in that: In step S3, the adaptive feature propagation mechanism of the generative adversarial network is specifically: Using the formula The node feature representation is updated by the node’s neighbor nodes and adjacent edges, strengthening the contextual association of node features and capturing the interaction relationship between amino acids; ij (l-1) is the feature of the neighbor edge ij at the l-1th layer, N(i) represents the set of neighbor nodes of node i, h j (l-1) represents the feature of node j in the l-1th layer, k(·) represents the weight mapping function, and h i (l) Represents the features of node i in layer l; Using the formula The feature representation of the edge is updated by the features of the neighboring edges to strengthen the consistency of the local and global structures of the protein; where N(j) represents the set of neighboring nodes of node j, e ik (l-1) is the feature of the edge ik from node i to node k in the l-1th layer, e kj (l-1) is the feature of edge kj from node k to node j in the l-1th layer, Represents element-wise product.

6. The protein mutation stability prediction method based on generative adversarial network data generation according to claim 4, characterized in that: In step S3, the discriminator further performs the following operations in order to discriminate the authenticity of the sample and obtain the final confidence score: Use the formula head t =attention(AW Q ,AW K ,AW V ) captures the long-range dependencies between amino acids; where A represents the amino acid feature representation after the adaptive feature propagation mechanism, attention(·) represents the attention mechanism, and head t represents the feature representation of the output of the t-th attention head, W Q , W K , W V are the mapping matrices of query, key, and value respectively; Using the formula Capture the short-distance dependencies between amino acids; Conv1D(·) represents the use of one-dimensional convolution processing, W O represents the learnable matrix, f represents the final learned feature embedding, and T represents the total number of attention heads; Using the formula z=W2·σ(W1·f+b1)+b2, a 1×1 output vector is obtained after transformation through the first fully connected layer and the second fully connected layer, which is used to measure the possibility that the sample is a true sample; where W1 and W2 represent the learnable matrices of the first fully connected layer and the second fully connected layer, respectively, b1 and b2 represent the bias values ​​of the first fully connected layer and the second fully connected layer, respectively, σ(·) represents the activation function, and z represents the possibility that the sample is a true sample.

7. The protein mutation stability prediction method based on generative adversarial network data generation according to claim 6, characterized in that: The specific process of step S4 is as follows: S41: Count all proteins included in the training set and validation set divided in step S1, obtain the situation where all amino acids on all proteins mutate into another 19 amino acids, exclude the mutation samples in the training set and validation set, obtain the training set and validation set without mutation, input the training set, validation set and mutation label without mutation into the generator G, generate the corresponding post-mutation structure diagram and the stability change value caused by the mutation; S42: The training set without mutation, the validation set and the generated structure diagram after mutation are merged. After the merger, each sample in the training set and the validation set has two labels, which are the stability change value caused by the mutation and whether it is a mutation sample generated by the generator G.

8. The protein mutation stability prediction method based on generative adversarial network data generation according to claim 7, characterized in that: In step S5, the specific process of constructing the downstream task network model is as follows: S51: Using formula H atom =Conv3D(X atom ) captures the spatial information of the mutant atom and its surrounding atoms; where X atom represents the feature matrix of the atom, Conv3D(·) represents the three-dimensional convolution, H atom represents the atoms after feature extraction; S52: Using adaptive feature propagation mechanism to capture the interaction relationship between amino acids; S53: At the surface level, the obtained multi-channel image is positionally encoded and then used using the formula Extract surface features; where Q = X surf W Q , K = X surf W K , V = X surf W V , Q, K, and V represent the query, key, and value matrices in the process of surface information feature extraction, respectively, and X surf represents the characteristic matrix of the surface, d k represents the dimension of the key, used for scaling factor, softmax(·) for attention weight normalization, H surf represents the surface information after feature extraction, W Q , W K , W V denote the query, key, and value learnable matrices respectively; S54: Generate embeddings of mutation sites using large language models to capture deep features in protein sequence and structure; S55: Utilizing formulas Captures the interrelationships between amino acid, atomic, and surface level features; where Q inter =[Q res ;Q atom ;Q surf ]·W Q , K inter =[K res ; K atom ; K surf ]·W K , V inter =[V res ; V atom ; V surf ]·W V , Q inter , K inter 、V inter represents the query, key, and value matrices in the process of fusing amino acid, atom, and surface level information, respectively. res , Q atom , Q surf They represent the amino acid, atom, and surface query feature matrices after feature extraction, respectively. res , K atom , K surf Respectively represent the amino acid, atom, and surface bond feature matrices after feature extraction, V res 、V atom 、V surf Respectively represent the amino acid, atom, and surface value feature matrices after feature extraction, [;] represents the matrix concatenation operation, H inter It represents the feature matrix after feature fusion of amino acid, atom and surface information; S56: Using the formula y′=σ(H inter p) effectively fuses the information fused from the three levels with the mutation embedding generated by the large language model; where p represents the embedding of the mutation site generated by the large language model, σ(·) represents the activation function, and y′ represents the fused embedding representation; S57: Using the formula x′=W4·σ(W3·y+b3)+b4, after transformation through the third fully connected layer and the fourth fully connected layer, a 1×2 output vector x′ is obtained, which is the stability change value caused by the mutation and the confidence that the sample is a generated sample; wherein W3 and W4 represent the learnable matrices of the third fully connected layer and the fourth fully connected layer respectively, and b3 and b4 represent the bias values ​​of the third fully connected layer and the fourth fully connected layer.

Images