Methods, devices, computer equipment, and media for epilepsy gene prediction based on transfer learning

Abstract

This application relates to a method, apparatus, computer device, and medium for predicting epilepsy genes based on transfer learning, comprising: S1, acquiring three-dimensional structural data and amino acid sequence data of the protein encoding a sample gene; S2, extracting structural feature vectors from the three-dimensional structural data and sequence feature vectors from the amino acid sequence data; S3, concatenating the structural feature vectors and sequence feature vectors to obtain a concatenated vector; S4, inputting the concatenated vector into a lightweight prediction module, outputting a probability score indicating that the sample gene belongs to an epilepsy gene; S5, adjusting the model parameters of the lightweight prediction module based on the model loss corresponding to the probability score, and repeating S2, S3, S4, and S5 based on the adjusted lightweight prediction module until the iteration stopping condition is met to obtain a trained lightweight prediction module, and predicting the probability score of the gene to be tested belonging to an epilepsy gene through the trained lightweight prediction module. This method can accurately capture epilepsy genes associated with epilepsy.

Description

Technical Field

[0001] This application relates to the fields of deep learning and bioinformatics technology, and in particular to a method, apparatus, computer device and medium for predicting epilepsy genes based on transfer learning. Background Technology

[0002] Epilepsy is one of the most common chronic neurological disorders, characterized by recurrent seizures. These seizures are caused by abnormal and sudden electrical activity in the brain, leading to sensory disturbances and behavioral changes such as loss of consciousness or muscle spasms. Epilepsy can significantly impact the quality of life of individuals with epilepsy, their families, and the community, posing a range of risks including physical injury and accidents, cognitive and psychological impairments, sudden epileptic death, and status epilepticus.

[0003] Epilepsy has diverse and complex causes, including genetic factors, tumors, infections, and brain injury. Among these, genetic factors play a significant role in many cases. Several genes have been identified as being associated with epilepsy, such as DEPDC5, SCN1A, KCNQ2, KCNQ3, and PCDH19. These genes encode proteins involved in various aspects of neuronal function, such as ion channels, signal transduction, cell adhesion, and synaptic transmission. Mutations in these genes disrupt the normal balance between excitation and inhibition in the brain, leading to increased susceptibility to epilepsy. Therefore, accurately identifying novel epilepsy-related genes is crucial for a deeper understanding of the molecular mechanisms of epilepsy and for developing novel diagnostic and therapeutic strategies.

[0004] Currently, although traditional technologies have made progress in identifying disease-associated genes, the traditional methods use the neural network model with the highest average score as the prediction model, which is prone to overfitting and thus cannot accurately capture some genes related to epilepsy. Summary of the Invention

[0005] Therefore, it is necessary to provide a transfer learning-based method, device, computer equipment, and medium for predicting epilepsy genes that can accurately capture epilepsy-related genes, addressing the aforementioned technical problems.

[0006] An epilepsy gene prediction method based on transfer learning, the method comprising:

[0007] S1. Obtain the three-dimensional structural data and amino acid sequence data of the protein encoded by the sample gene; the sample gene includes multiple normal genes and multiple epilepsy genes;

[0008] S2. Use a pre-trained structural feature extraction network to extract the structural feature vectors of each of the three-dimensional structural data, and use a pre-trained sequence feature extraction network to extract the sequence feature vectors of each of the amino acid sequence data;

[0009] S3. Concatenate the structural feature vector and the sequence feature vector of the same sample gene to obtain the concatenation vector of each sample gene;

[0010] S4. Input each of the spliced ​​vectors into the lightweight prediction module and output the probability score of each sample gene belonging to the epilepsy gene;

[0011] S5. Based on the model loss corresponding to each probability score, adjust the model parameters of the lightweight prediction module, and repeat S2, S3, S4 and S5 based on the adjusted lightweight prediction module until the iteration stopping condition is met to obtain the trained lightweight prediction module, so as to predict the probability score of the gene to be tested belonging to the epilepsy gene through the trained lightweight prediction module.

[0012] In this application, a pre-trained structural feature extraction network is used to extract structural feature vectors from each 3D structural data, and a pre-trained sequence feature extraction network is used to extract sequence feature vectors from each amino acid sequence data. The structural feature vectors and sequence feature vectors of the same sample gene are then concatenated to obtain the concatenated vectors for each sample gene, thus achieving a more comprehensive and in-depth characterization of the encoded protein. By inputting each concatenated vector into a lightweight prediction module, the probability score of each sample gene belonging to the epilepsy gene is output. Based on the model loss corresponding to each probability score, the model parameters of the lightweight prediction module are adjusted. The feature vectors and predicted probability scores are then repeatedly extracted based on the adjusted lightweight prediction module to obtain the trained lightweight prediction module. This effectively utilizes the pre-training knowledge of a large-scale dataset while significantly reducing the computational cost and time of training the lightweight prediction module. It also effectively avoids the overfitting problem that may occur when training the lightweight prediction module from scratch on a specific small dataset. Therefore, when predicting the probability score of a test gene belonging to the epilepsy gene using the trained lightweight prediction module, the accuracy of the predicted probability score is improved, thus accurately capturing epilepsy-related genes.

[0013] In one embodiment, the three-dimensional structural data includes multiple residues, each residue including multiple atoms, and the structural feature extraction network includes a multi-layer decoder layer and a multi-layer encoder layer, the decoder layer and the encoder layer being connected sequentially; the extraction process of the structural feature vector in step S2 includes:

[0014] For each residue in the three-dimensional structural data, calculate the Euclidean distance between each atom in the residue and each atom in the adjacent residues, and randomly generate the initial node features of each residue;

[0015] Based on each of the Euclidean distances of the residues and the initial node features, through The first feature vector output by the last encoder layer in the structural feature extraction network is obtained; It is the node feature of residue v in the l-th decoder layer. It is the node feature of residue u, which is the neighboring residue v, in the l-th decoder layer. Let M be the set of neighboring residues of residue v, M be the message function, and Aggregate be the aggregation function; e vu It is the edge relationship between residue v and adjacent residue u, and the edge relationship is determined based on the Euclidean distance;

[0016] The first feature vector is input to the decoder layer so that the first feature vector is decoded by the decoder layer to obtain the second feature vector in the intermediate sub-layer of each decoder layer;

[0017] By concatenating the second feature vectors, the structural feature vector is obtained.

[0018] In this application, for each residue in the three-dimensional structural data, the Euclidean distance between each atom in the residue and each atom in the adjacent residues is calculated, and initial node features of each residue are randomly generated. Based on each Euclidean distance and the initial node features of each residue, the method is further refined. The first feature vector output by the last encoder layer in the structural feature extraction network is obtained. The first feature vector is then input into the decoder layer to decode the first feature vector and obtain the second feature vector in the intermediate sub-layer of each decoder layer. The structural feature vector can then be obtained by concatenating the second feature vectors.

[0019] In one embodiment, the process of determining the edge relationship includes:

[0020] For each of the residues, determine the average Euclidean distance between each atom in the residue and each atom in the adjacent residues;

[0021] The average value is used as the edge relationship between the residue and the adjacent residue.

[0022] In this application, by determining the average value of the Euclidean distance between each atom in a residue and each atom in an adjacent residue, the average value is used as the edge relationship between the residue and its adjacent residue. This makes the determined edge relationship comprehensive and more accurate.

[0023] In one embodiment, the lightweight prediction module includes a lightweight attention submodule and a multilayer perceptron submodule, and step S4 includes:

[0024] The lightweight attention submodule is used to perform a linear transformation on the concatenated vector to obtain the key vector, value vector, and query vector of the concatenated vector;

[0025] Based on the key vector K, value vector V, and query vector Q of the concatenated vector, through Attention reweighting is performed to obtain the reweighted feature vector F. att ;

[0026] The multilayer perceptron submodule is used to perform nonlinear mapping on the reweighted feature vector to obtain the probability score that the sample gene belongs to the epilepsy gene.

[0027] In this application, a lightweight attention submodule is used to perform a linear transformation on the concatenated vector to obtain the key vector, value vector, and query vector of the concatenated vector. Based on the key vector K, value vector V, and query vector Q of the concatenated vector, the following steps are taken: Attention reweighting is performed to obtain the reweighted feature vector F. att By using a multilayer perceptron submodule to perform nonlinear mapping on the reweighted feature vector, a probability score is obtained that the sample gene belongs to the epilepsy gene. This allows for the optimization and purification of the spliced ​​vector, making the final output of the multilayer perceptron submodule more reliable in indicating that the sample gene belongs to the epilepsy gene. Based on this reliable probability score, an accurate model loss can be obtained, thus improving the optimization effect of the model parameters.

[0028] In one embodiment, the formula for the nonlinear mapping process includes: , and ;

[0029] Where h1 is the output vector of the first hidden network layer in the multilayer perceptron submodule, and W MLP1 b represents the weight parameters in the first hidden network layer. MLP1 h1 is the bias term in the first hidden network layer; h2 is the output vector of the second hidden network layer in the multilayer perceptron submodule; W MLP2 b represents the weight parameters in the second hidden network layer. MLP2 The bias term in the second hidden network layer; P is the probability score of the sample gene belonging to the epilepsy gene, W MLPo b represents the weight parameters of the output layer in the multilayer perceptron submodule. MLPo This is the bias term for the output layer.

[0030] In this application, by using , and By performing nonlinear mapping, the probability score of the final output can be made more reliable. This allows for the accurate determination of the model loss based on the reliable probability score, thereby improving the optimization effect of the model parameters.

[0031] In one embodiment, the process of determining the model loss includes:

[0032] Determine the true label of each sample gene;

[0033] Based on the true label and probability score of each sample gene, through Calculate model loss N is the total number of genes in the sample, Y i P is the true label for sample gene i. i This represents the probability score of gene i in the sample.

[0034] In this application, the true labels of each sample gene are determined, and based on the true labels and probability scores of each sample gene, the following steps are taken: Calculate model loss This can improve the optimization effect of model parameters when optimizing model parameters based on model loss.

[0035] In one embodiment, the method further includes:

[0036] Multiple sample genes are divided into multiple datasets, and multiple training and validation data sets are constructed using each dataset as a validation set and the remaining datasets as training sets. The number of datasets is consistent with the number of candidate lightweight prediction modules to be evaluated. Each training and validation data set includes one validation set and multiple training sets. There is a one-to-one mapping relationship between the training and validation data sets and the candidate lightweight prediction modules.

[0037] Based on the mapping relationship and the training set in the training and validation data set, the corresponding candidate lightweight prediction modules are trained to obtain multiple modules to be validated.

[0038] Based on the mapping relationship and the validation set in the training and validation data set, the prediction effect of the corresponding module to be validated is verified, and the performance indicators of each module to be validated are obtained.

[0039] The performance evaluation result of the candidate lightweight prediction module is determined based on the average performance index of each module to be verified.

[0040] In this application, multiple sample genes are divided into multiple datasets, and multiple training and validation datasets are constructed using each dataset as a validation set and the remaining datasets as training sets. Based on the mapping relationship and the training set in the training and validation datasets, corresponding candidate lightweight prediction modules are trained to obtain multiple modules to be validated. Based on the mapping relationship and the validation set in the training and validation datasets, the prediction effect of the corresponding modules to be validated is verified to obtain the performance index of each module to be validated. In this way, the performance evaluation result of the candidate lightweight prediction modules can be determined based on the average performance index of each module to be validated, and then the lightweight prediction modules can be selected based on the performance evaluation result.

[0041] An epilepsy gene prediction device based on transfer learning, the device being used to perform:

[0042] S1. Obtain the three-dimensional structural data and amino acid sequence data of the protein encoded by the sample gene; the sample gene includes multiple normal genes and multiple epilepsy genes;

[0043] S2. Use a pre-trained structural feature extraction network to extract the structural feature vectors of each of the three-dimensional structural data, and use a pre-trained sequence feature extraction network to extract the sequence feature vectors of each of the amino acid sequence data;

[0044] S3. Concatenate the structural feature vector and the sequence feature vector of the same sample gene to obtain the concatenation vector of each sample gene;

[0045] S4. Input each of the spliced ​​vectors into the lightweight prediction module and output the probability score of each sample gene belonging to the epilepsy gene;

[0046] S5. Based on the model loss corresponding to each probability score, adjust the model parameters of the lightweight prediction module, and repeat S2, S3, S4 and S5 based on the adjusted lightweight prediction module until the iteration stopping condition is met to obtain the trained lightweight prediction module, so as to predict the probability score of the gene to be tested belonging to the epilepsy gene through the trained lightweight prediction module.

[0047] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0048] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0049] In the aforementioned epilepsy gene prediction device, computer equipment, and medium based on transfer learning, structural feature vectors of each three-dimensional structural data are extracted using a pre-trained structural feature extraction network, and sequence feature vectors of each amino acid sequence data are extracted using a pre-trained sequence feature extraction network. The structural feature vectors and sequence feature vectors of the same sample gene are then spliced ​​together to obtain the spliced ​​vectors of each sample gene, thus achieving a more comprehensive and in-depth characterization of the encoded protein. By inputting each concatenated vector into a lightweight prediction module, the probability score of each sample gene belonging to the epilepsy gene is output. Based on the model loss corresponding to each probability score, the model parameters of the lightweight prediction module are adjusted. The feature vector and prediction probability score are repeatedly extracted based on the adjusted lightweight prediction module to obtain the trained lightweight prediction module. This effectively utilizes the pre-training knowledge of large-scale datasets, while greatly reducing the computational cost and time of training the lightweight prediction module. It also effectively avoids the overfitting problem that may occur when the lightweight prediction module is trained from scratch on a specific small dataset. Therefore, when predicting the probability score of the test gene belonging to the epilepsy gene through the trained lightweight prediction module, the accuracy of the predicted probability score is improved, thus accurately capturing epilepsy-related genes. Attached Figure Description

[0050] Figure 1 This is a diagram illustrating the application environment of a transfer learning-based epilepsy gene prediction method in one embodiment.

[0051] Figure 2 This is a flowchart illustrating an epilepsy gene prediction method based on transfer learning in one embodiment.

[0052] Figure 3 This is a schematic diagram showing the connection between the multi-layer decoder layer and the multi-layer encoder layer in a structural feature extraction network in one embodiment.

[0053] Figure 4 This is a schematic diagram of the overall process for epilepsy gene prediction in another embodiment;

[0054] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0056] The epilepsy gene prediction method based on transfer learning provided in this application can be applied to, for example... Figure 1In the application environment shown, terminal 102 interacts with server 104 via a wired / wireless channel. A data storage system can store the data that server 104 needs to process. S1. Obtain the three-dimensional structural data and amino acid sequence data of the protein encoding the sample gene; the sample gene includes multiple normal genes and multiple epilepsy genes; S2. Use a pre-trained structural feature extraction network to extract the structural feature vector of each of the three-dimensional structural data, and use a pre-trained sequence feature extraction network to extract the sequence feature vector of each of the amino acid sequence data; S3. Concatenate the structural feature vector and the sequence feature vector of the same sample gene to obtain the concatenated vector of each sample gene; S4. Input each concatenated vector into a lightweight prediction module and output the probability score of each sample gene belonging to the epilepsy gene; S5. Based on the model loss corresponding to each probability score, adjust the model parameters of the lightweight prediction module, and repeat S2, S3, S4 and S5 based on the adjusted lightweight prediction module until the iteration stopping condition is met to obtain the trained lightweight prediction module, so as to predict the probability score of the test gene belonging to the epilepsy gene. The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, etc. Server 104 can be a single server, a server cluster consisting of multiple servers, or a cloud computing center consisting of multiple servers.

[0057] In one embodiment, such as Figure 2 As shown, a method for predicting epilepsy genes based on transfer learning is provided, and this method is applied to... Figure 1 Taking server 104 as an example, the following steps are included:

[0058] S1. Obtain the three-dimensional structure data and amino acid sequence data of the proteins encoded by the sample genes; the sample genes include multiple normal genes and multiple epilepsy genes;

[0059] The number of normal genes and the number of epilepsy genes in the sample genes may be the same or different. Epilepsy genes are known positive sample genes associated with epilepsy, while normal genes are known negative sample genes not associated with epilepsy.

[0060] Normal gene sequences can be obtained through massive protein sequence recovery datasets and general-purpose protein databases. General-purpose protein databases include, but are not limited to, PDB (Protein Data Bank), Megascale Dataset (stability prediction benchmark database), and Fireprot Database (thermal stability database). Among these, massive protein sequence recovery datasets can be used to train sequence feature extraction networks.

[0061] Epilepsy genes can be obtained from epilepsy gene databases, including but not limited to GWAS (Genome-Wide Association Study), EpilepsyGene, Genes4Epilepsy (epilepsy gene), and DisGeNET (Disease Gene Network), among other publicly available epilepsy gene databases.

[0062] The protein encoded by the sample gene is a protein molecule that is "translated" according to the base sequence of the sample gene.

[0063] Three-dimensional structural data encoding proteins are coordinate information describing the specific location of each atom in the protein in three-dimensional space.

[0064] Amino acid sequence data encoding proteins is digital information describing the linear sequence of amino acid residues in a protein.

[0065] S2. Use a pre-trained structural feature extraction network to extract the structural feature vectors of each three-dimensional structural data, and use a pre-trained sequence feature extraction network to extract the sequence feature vectors of each amino acid sequence data.

[0066] Among them, the structural feature extraction network is a network model used to extract structural features from the three-dimensional structural data encoding proteins. For example, ProteinMPNN (Protein Message Passing Neural Network) is used as the structural feature extraction network. ProteinMPNN is a deep learning model based on graph neural networks, which is a model pre-trained based on the three-dimensional structural information of proteins.

[0067] Sequence feature extraction networks are network models used to extract sequence features from amino acid sequence data. For example, ESM2 (Evolutionary Scale Modeling 2) can be used as a sequence feature extraction network. ESM2 is a model pre-trained on large-scale protein amino acid sequences. A Transformer encoder pre-trained on the ESM2 model extracts a 1280-dimensional context embedding vector as a sequence feature vector from the last or a specified intermediate layer without fine-tuning. This embedding vector captures the evolutionary, structural, and functional information of amino acids in the sequence and is input together with the structural feature vector extracted by ProteinMPNN into a lightweight prediction module. The i-th dimension of the sequence feature vector represents the deep semantic meaning of the i-th amino acid in the amino acid sequence data.

[0068] Sequence feature vectors and structural feature vectors can be extracted using a dual-channel mode, i.e., one channel extracts sequence feature vectors and one channel extracts structural feature vectors, and the two channels can extract feature vectors simultaneously.

[0069] S3. Concatenate the structural feature vector and sequence feature vector of the same sample gene to obtain the concatenated vector of each sample gene;

[0070] Specifically, for each sample gene, a splicing operation is performed on its structural feature vector and sequence feature vector (splicing the structural feature vector and sequence feature vector), thereby obtaining the spliced ​​vector of each sample gene.

[0071] The concatenated expression is F comp =Concatenate(F struct ,F seq ), F comp For concatenating vectors, F struct F is the structural feature vector. seq For sequence feature vectors, Concatenate means concatenation.

[0072] S4. Input each spliced ​​vector into the lightweight prediction module and output the probability score of each sample gene belonging to the epilepsy gene.

[0073] The lightweight prediction module can process multiple spliced ​​vectors simultaneously and output the probability score of each sample gene belonging to the epilepsy gene at the same time. Alternatively, it can process multiple spliced ​​vectors sequentially and output the probability score of each sample gene belonging to the epilepsy gene sequentially.

[0074] In the process of outputting the probability score that a sample gene belongs to the epilepsy gene, the lightweight prediction module first uses a lightweight attention submodule to dynamically reweight the concatenated vector under context awareness using a self-attention mechanism. Then, the reweighted vector is input into a multilayer perceptron submodule containing two hidden network layers, ultimately outputting the probability score representing that the sample gene belongs to the epilepsy gene. The lightweight prediction module is designed to achieve a fast and scalable nonlinear mapping from high-dimensional features to the final prediction score using a lightweight architecture.

[0075] S5. Based on the model loss corresponding to each probability score, adjust the model parameters of the lightweight prediction module, and repeat S2, S3, S4 and S5 based on the adjusted lightweight prediction module until the iteration stopping condition is met to obtain the trained lightweight prediction module, so as to predict the probability score of the gene to be tested belonging to the epilepsy gene through the trained lightweight prediction module.

[0076] The iteration stopping conditions include, but are not limited to, reaching the required number of iterations and the model loss stabilizing.

[0077] The probability score is between 0 and 1, including 0 and 1.

[0078] The process of predicting the probability score of a gene belonging to the epilepsy gene using a trained lightweight prediction module involves: acquiring the three-dimensional structural data and amino acid sequence data of the protein encoding the gene; extracting the structural feature vector from the three-dimensional structural data of the gene using a pre-trained structural feature extraction network, and extracting the sequence feature vector from the amino acid sequence data of the gene using a pre-trained sequence feature extraction network; concatenating the structural feature vector and the sequence feature vector of the gene to obtain the concatenated vector of the gene; and inputting the concatenated vector of the gene into the pre-trained lightweight prediction module to output the probability score of the gene belonging to the epilepsy gene.

[0079] During the repetition of steps S2, S3, S4, and S5 using the adjusted lightweight prediction module, all network parameters of the structural feature extraction network and sequence feature extraction network, which serve as feature extractors, are frozen. Only the model parameters of the lightweight prediction module are updated iteratively through optimization algorithms. This allows the use of general knowledge about protein structural stability learned from the structural and sequence feature extraction networks, and the transfer of this knowledge to specific epilepsy gene prediction tasks, thereby achieving high-precision prediction performance without having to train a large model from scratch.

[0080] In the aforementioned transfer learning-based epilepsy gene prediction method, a pre-trained structural feature extraction network is used to extract structural feature vectors from each 3D structural data, and a pre-trained sequence feature extraction network is used to extract sequence feature vectors from each amino acid sequence data. The structural feature vectors and sequence feature vectors of the same sample gene are then concatenated to obtain the concatenated vectors for each sample gene, thus achieving a more comprehensive and in-depth characterization of the encoded protein. By inputting each concatenated vector into a lightweight prediction module, the probability score of each sample gene belonging to the epilepsy gene is output. Based on the model loss corresponding to each probability score, the model parameters of the lightweight prediction module are adjusted. The feature vectors and predicted probability scores are then repeatedly extracted based on the adjusted lightweight prediction module to obtain the trained lightweight prediction module. This effectively utilizes the pre-training knowledge of a large-scale dataset while significantly reducing the computational cost and time of training the lightweight prediction module. It also effectively avoids the overfitting problem that may occur when training the lightweight prediction module from scratch on a specific small dataset. Therefore, when predicting the probability score of a test gene belonging to the epilepsy gene using the trained lightweight prediction module, the accuracy of the predicted probability score is improved, thus accurately capturing epilepsy-related genes.

[0081] In one embodiment, the three-dimensional structural data includes multiple residues, each residue including multiple atoms, and the structural feature extraction network includes multiple decoder layers and multiple encoder layers, with the decoder layers and encoder layers connected sequentially; the extraction process of the structural feature vector in step S2 includes:

[0082] For each residue in the three-dimensional structural data, the Euclidean distance between each atom in the residue and each atom in the adjacent residues is calculated, and the initial node features of each residue are randomly generated.

[0083] Based on the Euclidean distance of each residue and the initial node features, through The first feature vector output by the last encoder layer in the structural feature extraction network is obtained; It is the node feature of residue v in the l-th decoder layer. It is the node feature of residue u, which is the neighboring residue v, in the l-th decoder layer. Let M be the set of neighboring residues of residue v, M be the message function, and Aggregate be the aggregation function; e vu It is the edge relationship between residue v and adjacent residue u, which is determined based on Euclidean distance;

[0084] The first feature vector is input to the decoder layer so that the first feature vector is decoded by the decoder layer to obtain the second feature vector in the intermediate sub-layer of each decoder layer;

[0085] By concatenating the second feature vectors, we obtain the structural feature vector.

[0086] A residue is a single repeating unit that makes up a biological macromolecule. The atoms in a residue include, but are not limited to, N (nitrogen), C (carbon), and O (oxygen). An adjacent residue refers to a residue that is next to the given residue.

[0087] The initial node features of each residue are randomly generated by: acquiring the topological information and number of residues from the three-dimensional structural data; for each residue, independently sampling from a preset random probability distribution to obtain the initial node features of each residue; the preset random probability distribution is one of a standard normal distribution, a uniform distribution, or a truncated normal distribution. Further, the three-dimensional spatial coordinates M of each residue are extracted from the three-dimensional structural data. i (x i ,y i ,z i ), and generate a random projection matrix W; where the elements of the random projection matrix W follow a mean of 0 and a variance of . Gaussian distribution; initial node features of each residue are calculated through linear mapping: , The initial node characteristics of residue i.

[0088] For each residue in the three-dimensional structural data, the Euclidean distance between atoms is calculated.

[0089] Adjacent residues refer to the residues that are physically closest to the original residue in three-dimensional space. Furthermore, the number of adjacent residues is 48, meaning that each residue is ordered from smallest to largest according to its physical distance to the original residue, and the top 48 are all adjacent residues.

[0090] In the formula In the middle, when l is 0, The initial node features of residue v, The initial node features of residue v's neighboring residue u. The edge relationship e between residue v and its neighboring residue u. vu It can be the Euclidean distance between any atom in residue v and any atom in adjacent residue u, or it can be the average of the Euclidean distances between all atoms in residue v and each atom in adjacent residue u. Each Euclidean distance of a residue refers to the Euclidean distance between each atom in the residue and each atom in the adjacent residues.

[0091] A schematic diagram of the connection between the multi-layer decoder layer and the multi-layer encoder layer in the structural feature extraction network is shown below. Figure 3 As shown.

[0092] When using a pre-trained structural feature extraction network to extract structural feature vectors from each 3D structural data, the information input to the network is the Euclidean distance-encoded Gaussian radial basis function of each residue in the 3D structural data and the initial node features. The Euclidean distance-encoded Gaussian radial basis function R(d... ij ) represents d ij Let be the Euclidean distance between atoms i and j. For the k-th preset center value, It is the width parameter that adjusts the ambiguity of distance coding. The value of k can be determined by preset, and the width parameter can also be obtained by preset or random initialization.

[0093] The encoder layer learns the structural context of each residue through message passing between residues.

[0094] Furthermore, the decoder layer has three layers. The first feature vector is input into the decoder layer to decode it, obtaining the second feature vector in the intermediate sub-layers of each decoder layer. This includes: inputting the first feature vector into the first decoder layer to obtain its output vector, and simultaneously extracting the second feature vector from the intermediate sub-layers of the first decoder layer; inputting the output vector of the first decoder layer into the second decoder layer to obtain its output vector, and simultaneously extracting the second feature vector from the intermediate sub-layers of the second decoder layer; and inputting the output vector of the second decoder layer into the third decoder layer to obtain its output vector, and simultaneously extracting the third feature vector from the intermediate sub-layers of the third decoder layer. For example, extracting a 128-dimensional second feature vector from the intermediate sub-layers of the first, second, and third decoder layers.

[0095] In this embodiment, for each residue in the three-dimensional structural data, the Euclidean distance between each atom in the residue and each atom in the adjacent residues is calculated, and initial node features of each residue are randomly generated. Based on each Euclidean distance and the initial node features of each residue, the method is then used to... The first feature vector output by the last encoder layer in the structural feature extraction network is obtained. The first feature vector is then input into the decoder layer to decode the first feature vector and obtain the second feature vector in the intermediate sub-layer of each decoder layer. The structural feature vector can then be obtained by concatenating the second feature vectors.

[0096] In one embodiment, the process of determining edge relationships includes:

[0097] For each residue, determine the average Euclidean distance between each atom in the residue and each atom in the adjacent residue;

[0098] The average value represents the edge relationship between residues and adjacent residues.

[0099] In a specific application, residue A includes atoms 1 and 2, and the adjacent residue B of residue A includes atoms 3 and 4. Then the edge relationship between residue A and residue B is (S... 13 +S 14 +S 23 +S 24 ) / 4, S 13 S is the Euclidean distance between atom 1 and atom 3. 14 S is the Euclidean distance between atom 1 and atom 4. 23 S is the Euclidean distance between atoms 2 and 3. 24 Let be the Euclidean distance between atom 2 and atom 4.

[0100] In this embodiment, by determining the average value of the Euclidean distance between each atom in a residue and each atom in an adjacent residue, the average value is used as the edge relationship between the residue and its adjacent residue. This makes the determined edge relationship comprehensive and more accurate.

[0101] In one embodiment, the lightweight prediction module includes a lightweight attention submodule and a multilayer perceptron submodule, and step S4 includes:

[0102] The lightweight attention submodule is used to perform a linear transformation on the concatenated vector to obtain the key vector, value vector, and query vector of the concatenated vector;

[0103] Based on the concatenated key vector K, value vector V, and query vector Q, through... Attention reweighting is performed to obtain the reweighted feature vector F. att ;

[0104] The multilayer perceptron submodule is used to perform nonlinear mapping on the reweighted feature vector to obtain the probability score of the sample gene belonging to the epilepsy gene.

[0105] The lightweight attention submodule and the multilayer perceptron submodule are connected in series.

[0106] The lightweight attention submodule mainly optimizes and refines the generated concatenation vector through a self-attention mechanism. Specifically, it performs a linear transformation on the concatenation vector to obtain the key vector, value vector, and query vector of the concatenation vector.

[0107] The specific formula for the linear transformation is expressed as follows: , It is a learnable weight matrix, d k Let D be the dimension of the key vector, and D be the dimension of the concatenated vector.

[0108] In this embodiment, a lightweight attention submodule is used to perform a linear transformation on the concatenated vector to obtain the key vector, value vector, and query vector of the concatenated vector. Based on the key vector K, value vector V, and query vector Q of the concatenated vector, through... Attention reweighting is performed to obtain the reweighted feature vector F. att By using a multilayer perceptron submodule to perform nonlinear mapping on the reweighted feature vector, a probability score is obtained that the sample gene belongs to the epilepsy gene. This allows for the optimization and purification of the spliced ​​vector, making the final output of the multilayer perceptron submodule more reliable in indicating that the sample gene belongs to the epilepsy gene. Based on this reliable probability score, an accurate model loss can be obtained, thus improving the optimization effect of the model parameters.

[0109] In one embodiment, the formula for nonlinear mapping processing includes: , and ;

[0110] Where h1 is the output vector of the first hidden network layer in the multilayer perceptron submodule, and W MLP1 b represents the weight parameters in the first hidden network layer. MLP1 h1 is the bias term in the first hidden network layer; h2 is the output vector of the second hidden network layer in the multilayer perceptron submodule; W MLP2 b represents the weight parameters in the second hidden network layer. MLP2 The bias term in the second hidden network layer; P is the probability score of the sample gene belonging to the epilepsy gene, W MLPo b represents the weight parameters of the output layer in the multilayer perceptron submodule. MLPo This is the bias term for the output layer.

[0111] The multilayer perceptron submodule encloses two hidden network layers. When using the multilayer perceptron submodule to perform nonlinear mapping on the reweighted feature vector, in the first hidden network layer, through... The output vector h1 of the first hidden network layer is obtained, and then the output vector h1 of the first hidden network layer is input into the second hidden network layer; in the second hidden network layer, through... The output vector h2 of the second hidden network layer is obtained, and then the output vector h2 of the second hidden network layer is input into the output layer; in the output layer, through Determine the probability score P of a sample gene belonging to the epilepsy gene. σ is the activation function, and the sigmoid function is used to compress the final output probability score to between 0 and 1.

[0112] In this embodiment, by using , and By performing nonlinear mapping, the probability score of the final output can be made more reliable. This allows for the accurate determination of the model loss based on the reliable probability score, thereby improving the optimization effect of the model parameters.

[0113] In one embodiment, the process of determining the model loss includes:

[0114] Determine the true gene labels for each sample;

[0115] Based on the true labels and probability scores of genes in each sample, through Calculate model loss N is the total number of genes in the sample, Y i P is the true label for sample gene i. iThis represents the probability score of gene i in the sample.

[0116] The true labels for each sample gene fall into two categories: those indicating a normal gene and those indicating an epilepsy gene. For example, a true label of "1" represents an epilepsy gene, while a true label of "0" represents a normal gene.

[0117] Furthermore, when optimizing parameters through model loss, an optimizer (such as AdamW) is used to optimize model parameters through backpropagation algorithm and model loss.

[0118] In this embodiment, by determining the true label of each sample gene, and based on the true label and probability score of each sample gene, through... Calculate model loss This can improve the optimization effect of model parameters when optimizing model parameters based on model loss.

[0119] In one embodiment, the method further includes:

[0120] Multiple sample genes are divided into multiple datasets, and multiple training and validation datasets are constructed using each dataset as a validation set and the remaining datasets as training sets. The number of datasets is consistent with the number of candidate lightweight prediction modules to be evaluated. Each training and validation dataset includes one validation set and multiple training sets. There is a one-to-one mapping relationship between the training and validation datasets and the candidate lightweight prediction modules.

[0121] Based on the mapping relationship and the training set in the training and validation dataset, the corresponding candidate lightweight prediction modules are trained to obtain multiple modules to be validated.

[0122] Based on the mapping relationship and the validation set in the training and validation dataset, the prediction effect of the corresponding module to be validated is verified, and the performance indicators of each module to be validated are obtained.

[0123] The performance evaluation results of the candidate lightweight prediction modules are determined based on the average performance metrics of each module to be verified.

[0124] This embodiment is essentially a specific method of five-fold cross-validation.

[0125] The number of training and validation datasets is the same as the number of datasets, and the validation sets in the training and validation datasets are not the same.

[0126] The model structure and hyperparameters of all candidate lightweight prediction modules are identical. However, the hyperparameters of the modules to be validated may not be identical.

[0127] The mapping relationship can be a manually defined relationship.

[0128] Performance metrics include, but are not limited to, accuracy, precision, recall, F1 score, and area under the curve.

[0129] Furthermore, if the performance evaluation results of the candidate lightweight prediction module indicate that the prediction effect of the candidate lightweight prediction module is good, then the candidate lightweight prediction module is taken as the lightweight prediction module in S2.

[0130] In this embodiment, multiple sample genes are divided into multiple datasets, and multiple training and validation datasets are constructed using each dataset as a validation set and the remaining datasets as training sets. Based on the mapping relationship and the training set in the training and validation datasets, corresponding candidate lightweight prediction modules are trained to obtain multiple modules to be validated. Based on the mapping relationship and the validation set in the training and validation datasets, the prediction effect of the corresponding modules to be validated is verified to obtain the performance index of each module to be validated. In this way, the performance evaluation result of the candidate lightweight prediction modules can be determined based on the average performance index of each module to be validated, and then the lightweight prediction modules can be selected based on the performance evaluation result.

[0131] In one specific embodiment, such as Figure 4 As shown, multiple epilepsy genes are obtained from an epilepsy gene database, and multiple normal genes are obtained from a massive protein sequence recovery dataset. Structural feature vectors of the three-dimensional structural data of normal and epilepsy genes are extracted using a structural feature extraction network, and sequence feature vectors of the amino acid sequence data of normal and epilepsy genes are extracted using a sequence feature extraction network. A lightweight prediction module consisting of a lightweight attention submodule and a multilayer perceptron submodule is used to predict probability scores and output the probability score of each sample gene belonging to the epilepsy gene.

[0132] To better illustrate the technical effects of this application, a cross-validation step was used to obtain an unbiased evaluation of its effectiveness. Table 1 lists the results of the ablation experiment conducted under five-fold cross-validation regarding epilepsy gene prediction:

[0133] Table 1. Epilepsy gene prediction results of this application under ablation experiments.

[0134]

[0135] Wherein, ACC represents accuracy, Sen represents sensitivity, Spec represents specificity, Pre represents precision, MCC represents the Matthews correlation coefficient, AUC represents the area under the curve, MLP represents a multilayer perceptron, and PSSM represents the position-specific scoring matrix. As can be seen from the table, this application achieved high evaluation scores in multiple experiments, indicating that it possesses excellent performance in epilepsy gene prediction.

[0136] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0137] Based on the same inventive concept, this application also provides a transfer learning-based epilepsy gene prediction device for implementing the above-described transfer learning-based epilepsy gene prediction method. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more transfer learning-based epilepsy gene prediction device embodiments provided below can be found in the limitations of the transfer learning-based epilepsy gene prediction method described above, and will not be repeated here.

[0138] In one embodiment, an epilepsy gene prediction device based on transfer learning is provided, the device being used to perform:

[0139] S1. Obtain the three-dimensional structural data and amino acid sequence data of the protein encoded by the sample gene; the sample gene includes multiple normal genes and multiple epilepsy genes;

[0140] S2. Use a pre-trained structural feature extraction network to extract the structural feature vectors of each of the three-dimensional structural data, and use a pre-trained sequence feature extraction network to extract the sequence feature vectors of each of the amino acid sequence data;

[0141] S3. Concatenate the structural feature vector and the sequence feature vector of the same sample gene to obtain the concatenation vector of each sample gene;

[0142] S4. Input each of the spliced ​​vectors into the lightweight prediction module and output the probability score of each sample gene belonging to the epilepsy gene;

[0143] S5. Based on the model loss corresponding to each probability score, adjust the model parameters of the lightweight prediction module, and repeat S2, S3, S4 and S5 based on the adjusted lightweight prediction module until the iteration stopping condition is met to obtain the trained lightweight prediction module, so as to predict the probability score of the gene to be tested belonging to the epilepsy gene through the trained lightweight prediction module.

[0144] The modules in the aforementioned epilepsy gene prediction device based on transfer learning can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0145] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores various types of data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a transfer learning-based method for predicting epilepsy genes.

[0146] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0147] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0148] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0149] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0150] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0151] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0152] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0153] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for predicting epilepsy genes based on transfer learning, characterized in that, The method includes: S1. Obtain the three-dimensional structural data and amino acid sequence data of the protein encoded by the sample gene; the sample gene includes multiple normal genes and multiple epilepsy genes; S2. Use a pre-trained structural feature extraction network to extract the structural feature vectors of each of the three-dimensional structural data, and use a pre-trained sequence feature extraction network to extract the sequence feature vectors of each of the amino acid sequence data; S3. Concatenate the structural feature vector and the sequence feature vector of the same sample gene to obtain the concatenation vector of each sample gene; S4. Input each of the spliced ​​vectors into the lightweight prediction module and output the probability score of each sample gene belonging to the epilepsy gene; S5. Based on the model loss corresponding to each probability score, adjust the model parameters of the lightweight prediction module, and repeat S2, S3, S4 and S5 based on the adjusted lightweight prediction module until the iteration stopping condition is met to obtain the trained lightweight prediction module, so as to predict the probability score of the gene to be tested belonging to the epilepsy gene through the trained lightweight prediction module.

2. The method according to claim 1, characterized in that, The three-dimensional structural data includes multiple residues, each residue including multiple atoms, and the structural feature extraction network includes a multi-layer decoder layer and a multi-layer encoder layer, the decoder layer and the encoder layer being connected in sequence. The process of extracting structural feature vectors in step S2 includes: For each residue in the three-dimensional structural data, calculate the Euclidean distance between each atom in the residue and each atom in the adjacent residues, and randomly generate the initial node features of each residue; Based on each of the Euclidean distances of the residues and the initial node features, through The first feature vector output by the last encoder layer in the structural feature extraction network is obtained; It is the node feature of residue v in the l-th decoder layer. It is the node feature of residue u, which is the neighboring residue v, in the l-th decoder layer. Let M be the set of neighboring residues of residue v, M be the message function, and Aggregate be the aggregation function; e vu It is the edge relationship between residue v and adjacent residue u, and the edge relationship is determined based on the Euclidean distance; The first feature vector is input to the decoder layer so that the first feature vector is decoded by the decoder layer to obtain the second feature vector in the intermediate sub-layer of each decoder layer; By concatenating the second feature vectors, the structural feature vector is obtained.

3. The method according to claim 2, characterized in that, The process of determining the edge relationship includes: For each of the residues, determine the average Euclidean distance between each atom in the residue and each atom in the adjacent residues; The average value is used as the edge relationship between the residue and the adjacent residue.

4. The method according to claim 1, characterized in that, The lightweight prediction module includes a lightweight attention submodule and a multilayer perceptron submodule. Step S4 includes: The lightweight attention submodule is used to perform a linear transformation on the concatenated vector to obtain the key vector, value vector, and query vector of the concatenated vector; Based on the key vector K, value vector V, and query vector Q of the concatenated vector, through Attention reweighting is performed to obtain the reweighted feature vector F. att ; The multilayer perceptron submodule is used to perform nonlinear mapping on the reweighted feature vector to obtain the probability score that the sample gene belongs to the epilepsy gene.

5. The method according to claim 4, characterized in that, The formula for the nonlinear mapping process includes: , and ; Where h1 is the output vector of the first hidden network layer in the multilayer perceptron submodule, and W MLP1 b represents the weight parameters in the first hidden network layer. MLP1 h1 is the bias term in the first hidden network layer; h2 is the output vector of the second hidden network layer in the multilayer perceptron submodule; W MLP2 b represents the weight parameters in the second hidden network layer. MLP2 The bias term in the second hidden network layer; P is the probability score of the sample gene belonging to the epilepsy gene, W MLPo b represents the weight parameters of the output layer in the multilayer perceptron submodule. MLPo This is the bias term for the output layer.

6. The method according to claim 1, characterized in that, The process of determining the model loss includes: Determine the true label of each sample gene; Based on the true label and probability score of each sample gene, through Calculate model loss N is the total number of genes in the sample, Y i P is the true label for sample gene i. i This represents the probability score of gene i in the sample.

7. The method according to claim 1, characterized in that, The method further includes: Multiple sample genes are divided into multiple datasets, and multiple training and validation data sets are constructed using each dataset as a validation set and the remaining datasets as training sets. The number of datasets is consistent with the number of candidate lightweight prediction modules to be evaluated. Each training and validation data set includes one validation set and multiple training sets. There is a one-to-one mapping relationship between the training and validation data sets and the candidate lightweight prediction modules. Based on the mapping relationship and the training set in the training and validation data set, the corresponding candidate lightweight prediction modules are trained to obtain multiple modules to be validated. Based on the mapping relationship and the validation set in the training and validation data set, the prediction effect of the corresponding module to be validated is verified, and the performance indicators of each module to be validated are obtained. The performance evaluation result of the candidate lightweight prediction module is determined based on the average performance index of each module to be verified.

8. An epilepsy gene prediction device based on transfer learning, used to perform the method according to any one of claims 1-7, characterized in that, The device is used to perform: S1. Obtain the three-dimensional structural data and amino acid sequence data of the protein encoded by the sample gene; the sample gene includes multiple normal genes and multiple epilepsy genes; S2. Use a pre-trained structural feature extraction network to extract the structural feature vectors of each of the three-dimensional structural data, and use a pre-trained sequence feature extraction network to extract the sequence feature vectors of each of the amino acid sequence data; S3. Concatenate the structural feature vector and the sequence feature vector of the same sample gene to obtain the concatenation vector of each sample gene; S4. Input each of the spliced ​​vectors into the lightweight prediction module and output the probability score of each sample gene belonging to the epilepsy gene; S5. Based on the model loss corresponding to each probability score, adjust the model parameters of the lightweight prediction module, and repeat S2, S3, S4 and S5 based on the adjusted lightweight prediction module until the iteration stopping condition is met to obtain the trained lightweight prediction module, so as to predict the probability score of the gene to be tested belonging to the epilepsy gene through the trained lightweight prediction module.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 7.

Images