Synthetic lethal gene pair prediction method, device, terminal and medium based on graph convolutional neural network

Abstract

The present invention discloses a method, device, terminal, and medium for predicting synthetic lethal gene pairs based on a graph convolutional neural network. The synthetic lethal gene pair prediction method includes: obtaining protein structural features based on protein structure data; obtaining protein sequence features based on protein sequence data; obtaining protein functional features based on a protein-protein interaction network; merging and standardizing the protein structural features, sequence features, and functional features to obtain the gene features of the primary protein-producing genes; obtaining interactions between genes, and training a synthetic lethal gene pair prediction model based on a graph convolutional neural network using the gene interactions and gene features; obtaining a final feature representation for each gene based on the trained synthetic lethal gene pair prediction model, and predicting whether two genes form a synthetic lethal gene pair based on the final feature representation. This method improves the efficiency of feature extraction and the ability to predict gene interactions.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method, device, terminal and medium for predicting synthetic lethal gene pairs based on a graph convolutional neural network. Background Art

[0002] Among the many methods for cancer treatment, targeting the SL gene of the oncogene based on the principle of synthetic lethality (SL) can effectively kill cancer cells without damaging normal cell functions. Figure 1 As shown, synthetic lethality refers to the presence of gene pairs in human DNA that together guide the completion of some essential cellular life processes. When any one of these genes is inactivated, it does not have a fatal effect on the cell, and the remaining gene takes over the life process. However, when both are inactivated, the cell will no longer be able to complete its essential life processes, leading to cell death. For cancer patients, when the main pathogenic gene produced by the cancer cells in their bodies can be identified, the synthetic lethal gene of this gene can be targeted to attack the cancer cells. The first PARP inhibitor developed using the principle of synthetic lethality, which inhibits the repair of DNA damage in tumor cells and promotes apoptosis in tumor cells, was approved for clinical treatment in 2014.

[0003] Computational methods based on machine learning learn gene signatures from large amounts of data and construct predictive models to identify SL gene pairs. These methods can be divided into traditional machine learning methods and deep learning methods. Traditional machine learning methods mainly include support vector machines (SVM), random forests (RF), and logistic regression. These methods integrate multiple data sources, such as gene expression data, protein interaction networks (PPIs), and gene ontology (GO), to construct feature vectors and use machine learning algorithms for training and prediction. For example, the SL2MF method uses matrix decomposition technology to extract gene signatures from the gene differential expression similarity matrix to predict potential SL interactions. The advantage of these methods is that they can process large-scale datasets, but their disadvantage is that they require manual feature engineering and are easily affected by noisy data.

[0004] Deep learning methods automatically learn features from data by constructing complex network architectures, effectively capturing the complex relationships between genes. Graph convolutional neural networks (GCNs) learn low-dimensional representations of genes through message passing on gene interaction networks, thereby predicting SL gene pairs. The GCN model can automatically learn features from graphs, avoiding the tedious feature engineering used in traditional machine learning methods. The DDGCN model first applied graph neural networks (GNNs) to SL prediction, using a GCN strategy with a double dropout mechanism to effectively address data sparsity and prevent overfitting. The KG4SL model, based on the knowledge graph (KG), learns gene representations by sampling subgraphs around individual genes in the KG. The SLGNN model focuses on gene preferences in various relationships in the knowledge graph and designs a GNN model to effectively learn gene representations.

[0005] Although the above methods have achieved good results and interpretability in SL prediction, they are still limited to the feature representation of genetic data and ignore the possible role of protein-related data.

[0006] Therefore, the existing technology still needs to be improved and developed. Summary of the Invention

[0007] In view of the above-mentioned deficiencies in the prior art, the object of the present invention is to provide a method, device, terminal and medium for predicting synthetic lethal gene pairs based on a graph convolutional neural network, aiming to solve the problem that the existing synthetic lethal gene pair prediction methods do not take into account protein-related data, resulting in low prediction accuracy.

[0008] The technical solutions of the present invention are as follows:

[0009] In a first aspect, a method for predicting synthetic lethal gene pairs based on a graph convolutional neural network is provided, comprising:

[0010] Obtaining protein structural features based on protein structural data;

[0011] Obtaining protein sequence features based on protein sequence data;

[0012] Obtaining protein functional characteristics based on protein-protein interaction networks;

[0013] The structural, sequence, and functional characteristics of proteins are combined and standardized to obtain the gene characteristics of the main protein-producing genes;

[0014] Obtain gene interactions and use gene interactions and gene features to train a graph convolutional neural network-based synthetic lethal gene pair prediction model;

[0015] The final feature representation of each gene is obtained based on the trained synthetic lethal gene pair prediction model, and whether the two genes are a synthetic lethal gene pair is predicted based on the final feature representation.

[0016] A preferred technical solution, wherein obtaining the structural characteristics of a protein based on protein structure data specifically includes:

[0017] Alphafold2 is used to convert the protein sequence into the three-dimensional structure of the protein, and the three-dimensional structure is simplified into a contact map of the protein structure. The contact map is then processed using the node2vec method to obtain the structural characteristics of the protein.

[0018] A preferred technical solution, wherein the method of obtaining protein sequence features based on protein sequence data specifically includes:

[0019] The seq2vec method is used to extract the semantic information of protein sequences and generate feature embedding to obtain the sequence features of proteins.

[0020] A preferred technical solution, wherein the functional characteristics of proteins are obtained based on protein-protein interaction networks, specifically includes:

[0021] The node2vec method is used to extract the semantic information of nodes in the protein-protein interaction network and generate feature embeddings to obtain the functional characteristics of proteins.

[0022] The preferred technical solution is to obtain the interaction between genes, specifically including:

[0023] The interactions between genes were obtained from the SynlethDB database. Gene pairs with synthetic lethal relationships were used as positive samples, and gene pairs without synthetic lethal relationships were used as negative samples. Since there were fewer negative samples than positive samples, a generative adversarial network was used to generate negative samples so that the positive-to-negative sample ratio was 1:1.

[0024] The preferred technical solution, wherein the synthetic lethal gene pair prediction model based on graph convolutional neural network is trained based on gene interactions and gene features, specifically includes:

[0025] Genes are nodes, interactions between genes are edges between nodes, and gene features are node features;

[0026] set up is the initial eigenvector of node i, A is the adjacency matrix of the graph, D is the degree matrix, I is the identity matrix, and the graph convolution operation of the graph convolutional neural network model is expressed as:

[0027] in, It is obtained by adding self-connection to the original adjacency matrix A. yes The degree matrix, W (k) is the weight matrix of the kth layer, and σ is the nonlinear activation function.

[0028] A preferred technical solution, wherein the prediction of whether two genes are a synthetic lethal gene pair based on the final feature representation specifically includes:

[0029] The feature vectors of the two gene nodes u and v and Combined, a linear layer is used to predict whether they are a synthetic lethal gene pair. The predicted probability that gene nodes u and v are a synthetic lethal gene pair is expressed as:

[0030] In a second aspect, a synthetic lethal gene pair prediction device based on a graph convolutional neural network is provided, comprising:

[0031] A first acquisition module is configured to acquire structural features of the protein based on the protein structure data;

[0032] A second acquisition module is configured to acquire sequence features of proteins based on protein sequence data;

[0033] a third acquisition module, configured to acquire functional characteristics of proteins based on protein-protein interaction networks;

[0034] A fusion module is configured to merge and normalize the structural features, sequence features, and functional features of the protein to obtain the gene features of the main gene producing the protein;

[0035] A model building module is configured to obtain interactions between genes and train a synthetic lethal gene pair prediction model based on a graph convolutional neural network using the interactions between genes and gene features;

[0036] The prediction module is configured to obtain a final feature representation of each gene based on the trained synthetic lethal gene pair prediction model, and predict whether two genes are a synthetic lethal gene pair based on the final feature representation.

[0037] In a third aspect, a terminal is provided, comprising a memory, a processor, and a computer program stored in the memory and runnable on the processor, wherein when the processor executes the computer program, the method for predicting synthetic lethal gene pairs based on a graph convolutional neural network as described in the first aspect is implemented.

[0038] In a fourth aspect, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the synthetic lethal gene pair prediction method based on graph convolutional neural network described in the first aspect.

[0039] Beneficial effects: Compared with the existing technology, the present invention provides a method for predicting synthetic lethal gene pairs based on graph convolutional neural networks. This method comprehensively considers the interactions between proteins and the mutual influence or dependency of the genes that produce them, which not only improves the efficiency of feature extraction, but also enhances the prediction model's ability to predict gene interactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Schematic diagram of the synthetic lethality process.

[0041] Figure 2 Schematic diagram of the synthetic lethal gene pair prediction model based on graph convolutional neural network. DETAILED DESCRIPTION

[0042] The present invention provides a method, device, terminal, and medium for predicting synthetic lethal gene pairs based on a graph convolutional neural network. To make the objectives, technical solutions, and effects of the present invention clearer and more specific, the present invention is further described in detail below.

[0043] Proteins are the products of genes, and the effects of genes on biological traits are directly dependent on the proteins they produce. The inventors believe that the interactions between proteins can reflect the mutual influence or dependency of the genes they produce, that is, if there is a synthetic lethal effect between two genes, it is actually caused by the simultaneous inactivation of the proteins produced by the two genes. This means that protein is one of the direct actors in the occurrence of synthetic lethality, and the characteristic description of proteins should be able to well reflect the relevant mechanisms of synthetic lethality. Alphafold2 is a completely open source prediction model for protein spatial structure based on deep learning, which provides the most accurate prediction effect to date (Journal of chemical information and modeling, 2021, 61 (10): 4827-4831.). The spatial structure of proteins is closely related to their biological effects, and the characterization of protein spatial structure should be able to describe the effects of proteins. Based on this, the present invention proposes an improved synthetic lethal gene pair prediction method based on graph convolutional neural network, named Struct2SL, which comprehensively considers the characteristic characterization performance of protein sequence, PPI network and protein three-dimensional structure on gene interaction, so as to better predict potential SL pairs.

[0044] An embodiment of the present invention provides a method for predicting synthetic lethal gene pairs based on a graph convolutional neural network, comprising:

[0045] Obtaining protein structural features based on protein structural data;

[0046] Obtaining protein sequence features based on protein sequence data;

[0047] Obtaining protein functional characteristics based on protein-protein interaction networks;

[0048] The structural, sequence, and functional characteristics of proteins are combined and standardized to obtain the gene characteristics of the main protein-producing genes;

[0049] Obtain gene interactions and use gene interactions and gene features to train a graph convolutional neural network-based synthetic lethal gene pair prediction model;

[0050] The final feature representation of each gene is obtained based on the trained synthetic lethal gene pair prediction model, and whether the two genes are a synthetic lethal gene pair is predicted based on the final feature representation.

[0051] In one embodiment, obtaining the structural characteristics of the protein based on the protein structure data specifically includes:

[0052] Alphafold2 is used to convert the protein sequence into the three-dimensional structure of the protein, and the three-dimensional structure is simplified into a contact map of the protein structure. The contact map is then processed using the node2vec method to obtain the structural characteristics of the protein.

[0053] Specifically, for the structural characteristics of proteins, this embodiment uses Alphafold2 to predict the three-dimensional structure of proteins and simplifies the three-dimensional structure into a contact graph; the nodes in the contact graph are amino acids in the protein, and when the Euclidean distance between two amino acid molecules is less than 10 angstroms, it is recorded as contact between the two molecules, otherwise it is recorded as no contact; for convenience of calculation, this embodiment flattens the contact graph using the node2vec method, obtains the node vector representation in the contact graph for each amino acid node, and then flattens the feature matrix to obtain the feature vector; since the number of amino acids required for the composition of different proteins is not the same, the feature vector length is padded with trainable values, and finally the feature vector length is compressed to an acceptable range through a fully connected layer, thereby obtaining the structural characteristics of the protein.

[0054] First, alphafold2 is used to describe the three-dimensional structure of the protein. The pdb file data of the protein is output by alphafold2, and the Cα atomic coordinates in the three-dimensional structure of the protein are extracted. For each residue in the structure, there is: i =(x i ,y i ,z i ); where r i Represents the coordinate vector of the Cα atom of the i-th residue. Using the coordinates of the Cα atom, calculate the distance between all residue pairs and construct the distance matrix D:

[0055]

[0056] Among them, D ij represents the Euclidean distance between the i-th and j-th residues.

[0057] Next, the protein sequence is converted to a one-hot encoding representation to capture the composition information of the sequence. For a sequence containing V different amino acids, the one-hot encoding vector of the i-th amino acid is defined as:

[0058]

[0059] Then, according to the given threshold cmap_thresh, the distance matrix is ​​converted into the adjacency matrix A to represent the contact relationship in the protein structure:

[0060]

[0061] Finally, the adjacency matrix A and the one-hot encoding vector of the protein sequence are integrated together to form a graph representation of the protein structure. This graph representation can be used for subsequent graph convolutional network analysis to extract high-level features of the protein structure.

[0062] In one embodiment, obtaining protein sequence features based on protein sequence data specifically includes:

[0063] The seq2vec method is used to extract the semantic information of protein sequences and generate feature embedding to obtain the sequence features of proteins.

[0064] Specifically, protein sequences are composed of different arrangements of amino acids. Protein sequences can be considered as "sentences," allowing natural language processing methods to extract features from them. This example uses the ELMo model, whose parameters have been pre-trained, to extract features from protein sequences. The ELMo model uses a deep bidirectional recurrent neural network to learn subvector representations, thereby obtaining richer semantic information.

[0065] For protein sequence features, this example uses the deep learning-based seq2vec method to automatically extract semantic information from protein sequences and generate feature embeddings. This approach avoids the complexity of manually designing features and instead leverages a pre-trained deep learning model to capture complex patterns in sequences.

[0066] During the seq2vec conversion process, the ELMo model is first used to obtain an embedded representation of the sequence. The ELMo model is a pre-trained deep bidirectional language model that can capture the semantics of each element in the sequence (such as amino acids) in different contexts.

[0067] First, the protein sequence file is built into a sequence dictionary. Then, a function for processing embedding is defined, which processes the embedding representation output by the Elmo model and calculates the average value of all time steps to generate a single embedding vector for each protein:

[0068]

[0069] Among them, h t is the hidden state at time step t, and T is the length of the sequence.

[0070] Afterwards, the embedding representation is updated layer by layer in an iterative manner to obtain higher-level features:

[0071]

[0072] in, represents the embedding representation of entity i at layer l, N(i) represents the set of neighbor entities of entity i, and aggregate is an aggregation function used to integrate the information of the current entity and its neighbors.

[0073] Finally, the final protein embedding is obtained by accumulating the embedding representations of all layers:

[0074]

[0075] Where L is the number of layers in the network.

[0076] In one embodiment, the obtaining of functional characteristics of proteins based on a protein-protein interaction network specifically includes:

[0077] The node2vec method is used to extract the semantic information of nodes in the protein-protein interaction network and generate feature embeddings to obtain the functional characteristics of proteins.

[0078] Specifically, when proteins play their roles in life, their interactions with other proteins play a major role. Therefore, the interactions between proteins can also be used to describe the functional characteristics of proteins.

[0079] When processing protein-protein interaction (PPI) network data, this example uses a deep learning-based method, node2vec, to automatically extract semantic information about nodes (proteins) in the network and generate feature embeddings. This method effectively captures the topological structure and neighborhood information of the nodes in the network.

[0080] The node2vec method explores the network by simulating random walks to learn embedded representations of nodes. First, a starting node is selected from the network and a random walk is performed according to predefined transition probabilities. The transition probability is calculated by considering the neighborhood of the current node and parameters p and q. Specifically, for each neighbor of the current node, the probability of it being selected as the next node is calculated. This probability is proportional to the weight of the neighbor and is adjusted by parameters p and q. Mathematically, the transition probability can be expressed as:

[0081]

[0082] Among them, w uv is the edge weight between nodes u and v, and N(v) is the set of neighbors of node v.

[0083] Next, we repeat multiple random walks from each node to generate a series of walk paths. Each walk path can be regarded as a series of node sequences that contain the neighborhood information of the node.

[0084] These walk paths are then used as input to learn the embedding representations of the nodes using the Word2Vec model. The Word2Vec model is a pre-trained word embedding model that learns vector representations of words (or nodes) by optimizing the Skip-gram objective function. In this case, the node sequence in each walk path is treated as a sequence of words in a sentence, and the model learns the embedding vector for each node.

[0085] Finally, the final node embedding is obtained by averaging the embedding vectors of each node in all walk paths. These embedding vectors capture the structural and semantic information of the node in the network and extract data features, namely the functional characteristics of the protein, from the original PPI network data.

[0086] For each protein, this embodiment extracts its structural features, sequence features, and functional features. These three features are combined and multi-feature normalization is performed to obtain the protein features of each protein. Considering that each protein has a major gene for production, this embodiment uses a one-to-one approach to map gene features with protein features to describe the gene's feature vector. For cases where multiple proteins exist that correspond to a major gene for production, the gene features are described using the features of the major protein produced by the gene as indicated in the uniprot database to ensure the rationality of the model.

[0087] In one embodiment, obtaining the interaction between genes specifically includes:

[0088] The interactions between genes were obtained from the SynlethDB database. Gene pairs with synthetic lethal relationships were used as positive samples, and gene pairs without synthetic lethal relationships were used as negative samples. Since there were fewer negative samples than positive samples, a generative adversarial network was used to generate negative samples so that the positive-to-negative sample ratio was 1:1.

[0089] Specifically, SynLethDB is a comprehensive database of synthetic lethal gene pairs. Since the amount of real negative sample data in the SynLethDB database is relatively small, this embodiment uses a generative adversarial network (GAN)-based data enhancement, adopting a model architecture of a pair of generators (Generator) and discriminators (Discriminator) to generate realistic negative sample data through an adversarial process to enhance the original dataset.

[0090] The goal of the generator is to generate fake samples that are good enough to “fool” the discriminator. Given a random noise vector z, the generator G converts it into a fake data sample

[0091] Here, z is random noise sampled from a standard normal distribution, G consists of three hidden layers with 256, 128, and 64 neurons respectively. Each layer is followed by a ReLU activation function, and the last layer uses a Sigmoid activation function to ensure that the output is in the range [0, 1].

[0092] The goal of the discriminator is to distinguish between real samples and fake samples generated by the generator. The discriminator D accepts an input sample x and outputs a probability value indicating the probability that the sample is a real sample:

[0093] D(x)=σ(Wx+b);

[0094] Among them, x is the input sample (can be a real sample or a fake sample), W and b are the parameters of the discriminator, σ is the Sigmoid activation function, and the output value is in the range [0,1].

[0095] The process of training a GAN involves minimizing a combined loss function consisting of two parts: the discriminator loss and the generator loss.

[0096] The discriminator loss is the discriminator's attempt to maximize its ability to correctly classify real samples and fake samples. The discriminator's loss function L D Defined as:

[0097]

[0098] Among them, the first term is the loss of real samples, and the second term is the loss of fake samples generated by the generator.

[0099] The generator loss is the probability that the generator will try to minimize the probability that the discriminator will classify the fake samples it generates as fake. The generator loss function L G Defined as:

[0100]

[0101] During training, the parameters of the discriminator and generator are updated alternately. First, the generator parameters are fixed and the discriminator parameters are updated; then, the discriminator parameters are fixed and the generator parameters are updated. This process is repeated until a certain number of iterations is reached or the generated samples are realistic enough, that is, the error is small.

[0102] Through the above process, this embodiment can automatically learn the distribution characteristics of data from the original data and generate new data samples, thereby increasing the diversity and quantity of non-SL pair data sets, providing a richer and more accurate data foundation for subsequent link prediction.

[0103] In one embodiment, the training of a synthetic lethal gene pair prediction model based on a graph convolutional neural network using gene interactions and gene features specifically includes:

[0104] Genes are nodes, interactions between genes are edges between nodes, and gene features are node features;

[0105] set up is the initial eigenvector of node i, A is the adjacency matrix of the graph, D is the degree matrix, I is the identity matrix, and the graph convolution operation of the synthetic lethal gene pair prediction model based on the graph convolutional neural network is expressed as:

[0106]

[0107] in, It is obtained by adding self-connection to the original adjacency matrix A. yes The degree matrix, W (k) is the weight matrix of the kth layer, and σ is the nonlinear activation function.

[0108] Specifically, the graph convolutional neural network (GCN) is a deep learning model used to process graph structured data. It can retain the structural information between nodes when processing graph data, and can perform good link prediction on the graph by embedding node attributes in feature engineering.

[0109] This example proposes training a GCN-based synthetic lethal gene pair prediction model to predict the linkage of synthetic lethal gene pairs. The core of this model is to automatically extract complex relationship features between genes, thereby effectively capturing the interactions and functional associations between genes.

[0110] First, genes are represented as nodes in a graph, and interactions between genes are represented as edges between nodes. Based on this, the goal of GCN is to capture the structural information of the entire graph by iteratively updating the feature representation of the nodes, thereby predicting whether there is a link between gene pairs.

[0111] The core idea of ​​GCN is to update the feature representation of each node by aggregating information from neighboring nodes. This process can be described by the following mathematical formula:

[0112] set up is the initial eigenvector of node i, A is the adjacency matrix of the graph, D is the degree matrix, and I is the identity matrix. The graph convolution operation of GCN can be expressed as:

[0113]

[0114] in, It is obtained by adding self-connection to the original adjacency matrix A. yes The degree matrix, W (k) is the weight matrix of the kth layer, and σ is the nonlinear activation function.

[0115] This example defines a synthetic lethal gene pair prediction model based on graph convolutional neural network. The construction process is as follows: Figure 2 As shown in Figure 2, the model implements the above graph convolution operation through a series of linear layers and nonlinear activation functions. For each hidden layer, it first undergoes a linear transformation W. (k) Update the node's feature representation. Next, apply a nonlinear activation function (ReLU) to introduce nonlinearity and enhance the model's expressiveness. Finally, regularize with a Dropout layer to prevent overfitting. This allows for iterative updates of the node's feature representation.

[0116] After multiple layers of graph convolution, the final feature representation of each node can be obtained Where L is the number of layers of GCN. These feature representations will be used for the link prediction task, that is, predicting whether there is a link between two nodes (genes).

[0117] In one embodiment, L=256, that is, the synthetic lethal gene pair prediction model based on the graph convolutional neural network has 256 hidden layers.

[0118] In one embodiment, link prediction, i.e., predicting whether two genes are a synthetic lethal gene pair based on the final feature representation, specifically includes:

[0119] The feature vectors of the two gene nodes u and v and Combined, a linear layer or nonlinear layer is used to predict the existence of the link between them, that is, whether it is a synthetic lethal gene pair. The calculation formula is:

[0120]

[0121] in, is the predicted probability that there is a link between gene nodes u and v, that is, the predicted probability that gene nodes u and v are a synthetic lethal gene pair.

[0122] During model training, this embodiment uses a binary cross-entropy loss function to optimize model parameters and updates weights through a back-propagation algorithm. In addition, this embodiment also introduces regularization terms, such as L2 regularization, to further prevent overfitting.

[0123] Through the above method, this embodiment can automatically extract complex relationship features between genes from graph-structured data, complete the link prediction task, and identify synthetic lethal gene pairs. This method not only improves the efficiency of feature extraction, but also enhances the model's ability to predict gene interactions.

[0124] In a specific embodiment, as described above, the synthetic lethal gene pair prediction method (Struct2SL) based on graph convolutional neural network uses the protein sequence in the Uniprot database, the PPI network in the STRING database and the SynlethDB database as raw data, and converts the protein sequence into the three-dimensional structure of the protein through Alphafold2. The obtained three-dimensional structure is converted into a contact map of the protein structure, and processed to obtain structural features. At the same time, the PPI network is characterized by nodes to obtain the structural features, sequence features and functional features of the protein.

[0125] Proteins were mapped one-to-one to their primary gene-producing genes, thereby describing the genetic signatures of the genes. The SL network was reconstructed, incorporating node features (i.e., the gene features mentioned above) and edge features. Negative samples were generated using a GAN augmentation algorithm at a 1:1 ratio for the non-SL data. The training, validation, and test sets had a 7:1:2 ratio. The model learning rate was set to 0.004, and an early stopping strategy was employed. For all baseline methods, the hyperparameters and model parameters described in the original paper were used. The experimental results are as follows:

[0126]

[0127] The original papers of the baseline methods are as follows:

[0128] DDGCN:Cai R,Chen X,Fang Y,Wu M,Hao Y.Dual-dropout graph convolutionalnetwork for predicting synthetic lethality in humancancers.Bioinformatics.2020Aug 15;36(16):4458-4465.doi:10.1093 / bioinformatics / btaa211.PMID:32221609.

[0129] KG4SL:Li J,Lu X,Jiang K,Tang D,Ning B,Sun F.TARSL:Triple-attentioncross-network representation learning to predict synthetic lethality foranti-cancer drug discovery.IEEE J Biomed Health Inform.2023Aug 21;PP.doi:10.1109 / JBHI.2023.3306768.Epub ahead of print.PMID:37603479.

[0130] TAR4SL:Wang S,Xu F,Li Y,Wang J,Zhang K,Liu Y,Wu M,Zheng J.KG4SL:knowledge graph neural network for synthetic lethality prediction in human cancers.Bioinformatics.2021Jul12;37(Suppl_1):i418-i425.doi:10.1093 / bioinformatics / btab271.PMID:34252965;PMCID:PMC8336442.

[0131] SLGNN: Zhu Y, Zhou Y, Liu Y, Wang X, Li J. SLGNN: synthetic lethality prediction in human cancers based on factor-aware knowledge graph neuralnetwork.Bioinformatics.2023Feb3;39(2):btad015.doi:10.1093 / bioinformatics / btad015.PMID:36645245;PMCID:PMC9907046.

[0132] The performance of Struct2SL is compared with the four baseline models described previously based on three evaluation metrics: AUC, AUPR, and F1-score. As shown in the table, Struct2SL outperforms all baseline methods. Struct2SL achieves an AUC, AUPR, and F1-score of 0.9764, 0.9808, and 0.9425, respectively, which are 1.3%, 1.0%, and 3.7% higher than the next-best model, SLGNN.

[0133] The above comparison results show that GCN-based methods improve model performance by leveraging the topological structure of the graph to propagate information and exploiting the similarity of known SL interactions. However, baseline methods generally only use gene features to distinguish different relationships between genes when predicting SL. Among them, TAR4SL comprehensively considers the interaction relationships of the protein PPI network, but neither method considers the characteristic information of protein three-dimensional structure and protein sequence. From the comparison, it is clear that Struct2SL takes this into account and performs better in generating gene embeddings for SL interaction prediction.

[0134] Based on the same inventive concept, an embodiment of the present invention provides a synthetic lethal gene pair prediction device based on a graph convolutional neural network, comprising:

[0135] A first acquisition module is configured to acquire structural features of the protein based on the protein structure data;

[0136] A second acquisition module is configured to acquire sequence features of proteins based on protein sequence data;

[0137] a third acquisition module, configured to acquire functional characteristics of proteins based on protein-protein interaction networks;

[0138] A fusion module is configured to merge and normalize the structural features, sequence features, and functional features of the protein to obtain the gene features of the main gene producing the protein;

[0139] A model building module is configured to obtain interactions between genes and train a synthetic lethal gene pair prediction model based on a graph convolutional neural network using the interactions between genes and gene features;

[0140] The prediction module is configured to obtain a final feature representation of each gene based on the trained synthetic lethal gene pair prediction model, and predict whether two genes are a synthetic lethal gene pair based on the final feature representation.

[0141] Based on the same inventive concept, an embodiment of the present invention provides a terminal, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the synthetic lethal gene pair prediction method based on graph convolutional neural network as described above is implemented.

[0142] Based on the same inventive concept, an embodiment of the present invention provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable a computer to execute the synthetic lethal gene pair prediction method based on graph convolutional neural network as described above.

[0143] In summary, the key to the present invention is:

[0144] 1. Protein characteristics are used to describe gene characteristics. Proteins are the products of gene expression, and the effects of genes on biological traits are directly dependent on the proteins they produce. The present invention proposes that interactions between proteins can reflect the mutual influence or dependency of the genes that produce them. That is, if synthetic lethality exists between two genes, it is actually due to the simultaneous inactivation of the proteins produced by both genes.

[0145] 2. For protein feature extraction, the present invention obtains three aspects of features and fuses them, namely, sequence features extracted from protein sequence data, structural features extracted from protein structure data, and functional features extracted from PPI network. The three features are merged and standardized to obtain the genetic features of each protein.

[0146] 3. Considering the interactions between genes that rely on protein interactions. In addition to similarities with synthetic lethal genes themselves, potential synthetic lethal gene pairs may also be related to the biological process dependence between the proteins produced by the synthetic lethal genes.

[0147] 4. The protein's three-dimensional structure was described using Alphafold2. Using the PDB file data output by Alphafold2, the present invention extracted the Cα atomic coordinates within the protein structure. For each residue in the structure, the Euclidean distance between each residue was calculated and converted into an adjacency matrix using one-hot encoding, thereby obtaining a graphical representation of the protein's structural features.

[0148] 5. For the functional characteristics of proteins, the verified PPI network data provided by the public database STRING was used, and its vector feature representation was obtained through the node2vec method.

[0149] 6. For link prediction, the present invention defines a GCN-based synthetic lethal gene pair prediction model. For each hidden layer, the feature representation of the node is updated by linear transformation. Then, ReLU is used as the activation function to introduce nonlinear factors and enhance the expressive power of the model. Finally, regularization is performed through the Dropout layer to prevent overfitting and iteratively update the feature representation of the node. After multiple layers of graph convolution, the final feature representation of each node is obtained for the final prediction task, that is, predicting whether there is a link between two nodes (genes). The final feature representations of the two gene nodes can be combined and a linear layer can be used to predict the existence of a link between them (the output is the probability of synthetic lethality between the two gene nodes).

[0150] It should be understood that the application of the present invention is not limited to the above examples. For those skilled in the art, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A method for predicting synthetic lethal gene pairs based on graph convolutional neural networks, characterized in that: include: Obtaining protein structural features based on protein structural data; Obtaining protein sequence features based on protein sequence data; Obtaining protein functional characteristics based on protein-protein interaction networks; The structural, sequence, and functional characteristics of proteins are combined and standardized to obtain the gene characteristics of the main protein-producing genes; Obtain gene interactions and use gene interactions and gene features to train a graph convolutional neural network-based synthetic lethal gene pair prediction model; Based on the trained synthetic lethal gene pair prediction model, the final feature representation of each gene is obtained, and whether the two genes are a synthetic lethal gene pair is predicted based on the final feature representation; The obtaining of the interaction between genes specifically includes: Inter-gene interactions were obtained from the SynlethDB database. Gene pairs with synthetic lethality were used as positive samples, and gene pairs without synthetic lethality were used as negative samples. Since there were fewer negative samples than positive samples, a generative adversarial network was used to generate negative samples, so that the positive-to-negative sample ratio was 1:

1. The synthetic lethal gene pair prediction model based on graph convolutional neural network is trained based on gene interactions and gene features, specifically including: Genes are nodes, interactions between genes are edges between nodes, and gene features are node features; set up For nodes The initial eigenvector of is the adjacency matrix of the graph, is the degree matrix, is the identity matrix, and the graph convolution operation of the synthetic lethal gene pair prediction model based on graph convolutional neural network is expressed as: ； in, In the original adjacency matrix Add self-connection to the basis, yes The degree matrix of It is The weight matrix of the layer, is a non-linear activation function.

2. The method for predicting synthetic lethal gene pairs based on graph convolutional neural networks according to claim 1, characterized in that: The step of obtaining the structural characteristics of the protein based on the protein structure data specifically includes: Alphafold2 is used to convert the protein sequence into the three-dimensional structure of the protein, and the three-dimensional structure is simplified into a contact map of the protein structure. The contact map is then processed using the node2vec method to obtain the structural characteristics of the protein.

3. The method for predicting synthetic lethal gene pairs based on graph convolutional neural networks according to claim 1, characterized in that: The step of obtaining the sequence characteristics of the protein based on the protein sequence data specifically includes: The seq2vec method is used to extract the semantic information of protein sequences and generate feature embedding to obtain the sequence features of proteins.

4. The method for predicting synthetic lethal gene pairs based on graph convolutional neural networks according to claim 1, wherein: The method of obtaining the functional characteristics of proteins based on the protein-protein interaction network specifically includes: The node2vec method is used to extract the semantic information of nodes in the protein-protein interaction network and generate feature embeddings to obtain the functional characteristics of proteins.

5. The method for predicting synthetic lethal gene pairs based on graph convolutional neural networks according to claim 1, wherein: The predicting whether the two genes are a synthetic lethal gene pair based on the final feature representation specifically includes: The two gene nodes and The eigenvector of and Combined, a linear layer is used to predict whether they are synthetic lethal gene pairs, gene nodes and The predicted probability of a synthetic lethal gene pair is expressed as: 。 6. A synthetic lethal gene pair prediction device based on graph convolutional neural network, characterized in that: include: A first acquisition module is configured to acquire structural features of the protein based on the protein structure data; A second acquisition module is configured to acquire sequence features of proteins based on protein sequence data; a third acquisition module, configured to acquire functional characteristics of proteins based on protein-protein interaction networks; A fusion module is configured to merge and normalize the structural features, sequence features, and functional features of the protein to obtain the gene features of the main gene producing the protein; A model building module is configured to obtain interactions between genes and train a synthetic lethal gene pair prediction model based on a graph convolutional neural network using the interactions between genes and gene features; a prediction module configured to obtain a final feature representation of each gene based on the trained synthetic lethal gene pair prediction model, and predict whether two genes are a synthetic lethal gene pair based on the final feature representation; The obtaining of the interaction between genes specifically includes: Inter-gene interactions were obtained from the SynlethDB database. Gene pairs with synthetic lethality were used as positive samples, and gene pairs without synthetic lethality were used as negative samples. Since there were fewer negative samples than positive samples, a generative adversarial network was used to generate negative samples, so that the positive-to-negative sample ratio was 1:

1. The synthetic lethal gene pair prediction model based on graph convolutional neural network is trained based on gene interactions and gene features, specifically including: Genes are nodes, interactions between genes are edges between nodes, and gene features are node features; set up For nodes The initial eigenvector of is the adjacency matrix of the graph, is the degree matrix, is the identity matrix, and the graph convolution operation of the synthetic lethal gene pair prediction model based on graph convolutional neural network is expressed as: ； in, In the original adjacency matrix Add self-connection to the basis, yes The degree matrix of It is The weight matrix of the layer, is a non-linear activation function.

7. A terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, it implements the synthetic lethal gene pair prediction method based on graph convolutional neural network according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the synthetic lethal gene pair prediction method based on graph convolutional neural network according to any one of claims 1 to 5.

Images