A Cancer Driver Gene Prediction Method Based on Cross-Network Interaction Fusion

Abstract

This invention belongs to the field of bioinformatics and relates to a method for predicting cancer driver genes based on cross-network interactive fusion. First, multi-omics attributes, three biological networks, gene set functions, and structural inputs are constructed according to a unified gene order. Second, three parameter-independent multi-head graph attention branches are used to encode the three networks respectively, achieving network representation interaction for the same gene through cross-network attention, and node-level gating adaptive aggregation is utilized. Then, independent graph classification heads and auxiliary supervision are set to connect multi-omics, graph, functional, and structural representations to form the main classification branch input. Finally, the main classification result and graph branch result are fused in log-odds space using learnable parameters, and the probability of cancer driver genes is output via a sigmoid function. This invention realizes information interaction among three biological networks, node-level adaptive aggregation, and dual-branch joint prediction, providing a computational analysis tool for cancer driver gene identification and comprehensive analysis of multi-omics and biological networks.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, specifically relating to a method for predicting cancer driver genes based on cross-network interactive fusion. Background Technology

[0002] Identifying cancer driver genes is used to analyze functional genetic alterations during tumorigenesis and development. Computational methods typically combine multi-omics features such as gene expression, somatic mutations, and DNA methylation with biological networks for joint modeling. Protein-protein interaction networks reflect the physical or functional relationships between proteins, pathway co-occurrence networks reflect the joint participation of genes in biological pathways, and gene ontology semantic similarity networks reflect the similarity between functional annotations. These three types of networks have different adjacency relationships and information meanings. Existing methods, when processing these multiple networks, often use the same neighborhood averaging operator to encode each network separately and then directly weight and sum them, which has the following limitations: First, a fixed neighborhood normalization method makes it difficult to distinguish the contributions of neighbors for different nodes; second, compression occurs before information interaction occurs between the three network representations, and complementary features between networks are not explicitly modeled; third, when only a single classification output provides supervision, the network encoding result must undergo multiple subsequent feature processing before the classification error signal can be received, and there is a lack of direct constraints between the graph representation and the prediction target; fourth, graph information and other information are output through a single path, making it impossible to adjust the role of graph classification results in the final judgment based on training data. Therefore, there is a need for a cancer driver gene prediction method that can perform cross-network information exchange, node-by-node adaptive aggregation, and dual-classification branch joint training while retaining the three independent network topologies. Summary of the Invention

[0003] This invention provides a cancer driver gene prediction method based on cross-network interaction fusion, which improves the connectivity of various biological network representations in terms of interaction, aggregation, and supervision. The method includes the following four steps: Step 1: Construct input data: Generate multi-omics feature matrices, protein-protein interaction adjacency matrices, pathway co-occurrence adjacency matrices, gene ontology semantic similarity adjacency matrices, and gene set association matrices according to a unified gene order. Calculate random walk position encoding and PageRank centrality from the protein-protein interaction network. Output the attributes, functions, and structures of each gene.

[0004] Step 2: Generate cross-network graph representations: Set up two-layer multi-head graph attention encoders for the three biological networks respectively, and aggregate neighbor features within their respective adjacency relationships; use the three network representations of the same gene as three relation labels, exchange information through multi-head attention with cross-network masks, and then use node-by-node gating to obtain graph branch representations and graph branch classification log probabilities.

[0005] Step 3: Forming the dual-branch fusion result: The input of the main classification branch is composed of the basic attention aggregation result of three graph representations, multi-omics features, gene set functional embedding, random walk position encoding and PageRank centrality, to obtain the main classification log odds; the main classification log odds and the graph branch classification log odds are weighted by learnable parameters constrained by the Sigmoid function to obtain the final log odds.

[0006] Step 4: Joint training and output predictions: The joint objective function is constructed with the final classification loss as the main loss and the graph branch classification loss as the auxiliary loss. The model parameters are updated on the training set and the model state is selected according to the validation set. The final log odds are calculated for the test genes and converted into the probability of cancer driver genes by the Sigmoid function.

[0007] A cancer driver gene prediction method based on cross-network interaction fusion, the implementation process of step 1 is as follows: Let the set of genes to be analyzed be... ,in N This represents the number of genes. First, data from different sources is row-aligned based on gene identifiers to ensure that any row index... i All input matrices point to genes. vᵢ Multi-omics features include differences in gene expression, somatic mutation rate, and DNA methylation. For the pan-cancer task, these three features for each of the 16 cancer types are concatenated in a fixed order to obtain a 48-dimensional original attribute vector for each gene. After deleting rows with missing values, each feature dimension is standardized separately. Let... For genes i In the k The original values ​​of each feature μ k and σ k These are the mean and standard deviation of this dimension in terms of the genes preserved, respectively. The standardized value is: .in, ε To prevent positive numbers with a denominator of zero, all normalized vectors are arranged row-wise to form a node attribute matrix. In the pan-cancer task For cancer-specific tasks, a column is taken from each of the expression, mutation, and methylation positions of the corresponding cancer type to form... The node attribute matrix. The subsequent three network coding branches receive the same... This means that the branch differences originate from network adjacency relationships and branch parameters, rather than differences in the order of node inputs.

[0008] Construct protein-protein interaction networks respectively Pathway co-occurrence network Semantic similarity network with gene ontology For any network ,remember Its sparse adjacency matrix is .from Extract the row index, column index, and relation score from the non-zero elements to form the edge index. Sum of edges ,in For the network The number of edges. Edge indices are used to limit the neighborhood of graph attention; edge scores are retained for the edge bias encoding of the main classification branch and can be normalized and used as scalar edge attributes for graph attention when edge attributes are enabled. The three networks store their own edge indices, and the adjacency matrices are not added during computation. A gene set association matrix is ​​constructed. ,in The number of genes to be retained. Indicates gene Does it belong to the first A gene set. The sources of the gene set include pathway class sets and ontology class sets. In formation... Previously, the gene name table was re-indexed to fill missing associations with 0; at the same time, gene sets containing CANCER, TUMOR, NEOPLASM, CARCINOMA, LEUKEMIA, or SARCOMA in their names were removed to avoid directly replacing network learning with cancer naming information. (The selected gene set...) This is used only as a functional association input, without altering the edges of the three biological networks. The structural input is calculated from the protein-protein interaction adjacency matrix. Let... for The row-degree diagonal matrix, treating the degree of isolated nodes as 1, and the random walk transition matrix. and 3D position encoding Defined as: ,in, Indicates reservation Principal component transformation of each principal component Then calculate on the same protein interaction network. Centrality Damping coefficient Take 0.85: ,in, for A dimensional vector of all 1s The Each element is a gene. PageRank centrality Step 1 final output Edge indices and edge scores of the three networks, and gene set association matrix. Location coding and centrality vector Before training, binary labels consistent with the above gene sequence are read, and training index, validation index, and test index are saved for each fold. Each index is only used for its corresponding subset of data.

[0009] A cancer driver gene prediction method based on cross-network interaction fusion, the implementation process of step 2 is as follows: for Configure two separate multi-head graph attention encoders with independent parameters. Each encoder only receives data from the network. The edge indexes ensure that protein interactions, pathway co-occurrences, and gene ontology semantic similarity relationships remain independent before entering cross-network modules. Let the first... l Layer h The linear transformation matrix of each attention head is The attention vector is ,node i The input at this layer is For edges in the neighborhood First calculate the unnormalized fractions, then at the nodes i Intra-neighborhood normalization: , ,in, For the network Middle node The neighborhood of is represented by ⊕, where ⊕ denotes a vector connection. The normalization coefficient varies with the center node, neighboring nodes, network type, layer number, and attention head. The first layer will... The output connection of the head, and the application Activation function; the second layer continues to use Connect the heads and outputs to obtain the network. of Wei said: , Thus obtained , and The three encoders each optimize their own parameters, allowing the contribution of the same neighbor to the central gene to vary across different networks.

[0010] Stack the three representations of the same node as To preserve network identity, a learnable relation embedding is assigned to each of the three networks. Relationship marker .right After layer normalization, the three networks are treated as a sequence of labels of length 3. The queries, keys, and values ​​in cross-network multi-head attention are all obtained by linear projection of this sequence. For the first... A cross-network attention head: Set cross-network mask Its diagonal elements are negative infinity, and its off-diagonal elements are 0. This mask ensures that any network label reads information only from the other two networks, preventing its own representation from being counted repeatedly during the cross-network attention phase. The output of each head and the output after connecting all heads are as follows: , ,in, , This represents the number of attention heads across the network. Perform a residual join with the original relation label and perform layer normalization, then pass through the dimension. Feedforward networks; feedforward networks use Activate and discard operations, and then perform residual connections and layer normalization again to obtain the interactive tags. : During the calculation, Each node serves as the first dimension of the attention module, and each node's... , and The three network representations are formed as a sequence of network views of length 3. The input tensor places the three network labels of each node in a relation dimension of length 3, and the input shape across network modules is as follows: Cross-network attention is computed only along the relational dimension. The three network representations of the same node can interact, but they will not directly connect two different nodes within the same module. Local information between different nodes is transmitted through the aforementioned graph attention network, while global information is transmitted through the main classification described later. Processing. This allows us to distinguish the scope of cross-network interaction and cross-node interaction.

[0011] Node-level gating The three tags are connected as input, and then... Hidden layers After activating and discarding layers, three gated logarithmic values ​​are generated, which are then executed along the network dimensions. Different nodes calculate their own gating weights, and do not share a fixed three-network ratio. , The weights and biases of the gating network output layer are initialized to 0, so that all three gating values ​​are 1 / 3 at the beginning of training, and are subsequently updated by the classification error. Input layer normalization layer, Fully connected layer Activation layer, dropout layer, and single-output fully connected layer yield the log-odds ratio of independent graph branch classification. This output directly participates in auxiliary supervision, while retaining... Supply-demand analysis of three network pairs of nodes The gating ratio.

[0012] The independent graph classification head does not copy the parameters of the three sets of graph attention networks separately; instead, it shares the graph attention network outputs of the three networks with the main classification branch. During backpropagation training, the graph branch auxiliary loss is propagated back to the shared graph attention network along the graph classification head, node gating, and cross-network attention, while the final classification loss is propagated back to both the graph branch and the main classification branch along the learnable fusion path. Therefore, the shared graph attention network receives two types of classification constraints, while cross-network attention and node gating belong only to the graph branch.

[0013] A cancer driver gene prediction method based on cross-network interaction fusion, the implementation process of step 3 is as follows: The main classification branch and the graph branch share the three independent graph attention network representations obtained in step 2, but employ different subsequent processing. The main classification branch first processes each... Using a shared linear score vector Calculate the weights of the basic network and form a basic graph representation. This path does not use the labels resulting from cross-network interactions, thus allowing the main classification branch and the independent graph branch to provide two classification results with different structures: For the gene set association matrix in step 1 Set up a linear attention matrix and learnable gene set embedding table First, normalize the gene set weights within the row corresponding to each gene, then calculate the functional embedding of that gene. : The main classification branch also utilizes the edge scores of the three networks to form multi-head edge biases. For the network... m Middle (i, j) The row index, column index, and fraction join are Generated through a shared linear layer and the Sigmoid function Dimension the biases and sum the biases from different networks at the same location. Positions without edges are kept at 0. Standardize multi-omics vectors Basic diagram representation Function embedding Random walk position coding and centrality Join along the feature dimension. The resulting dimension is... Mapped to a linear projection d Dimensions constitute the initial representation of the main classification sequence. : , will all genes Stacked as pass through in sequence Layered Transformer. Each layer first performs edge biasing. Multi-head self-attention, followed by residual connections and layer normalization; then through... The feedforward network is then used again for residual connections and layer normalization. For the ... l Layer The attention matrix for this size is: The last layer output The Obtaining the primary class log odds by passing through a single-output linear layer Set the learnable scalar ,by Constrain it to Then, the main classification results and the graph branch classification results are fused in the log-odds space: ,in, For genes The predicted probability of cancer driver genes during training. Updated simultaneously with other parameters. Since fusion is performed before the sigmoid probability transformation, both branches participate in the weighting with uncompressed log-probability, avoiding redundant compression caused by first converting them separately to probabilities. The final output is... and For the sake of accuracy, Used only to assist in training and branch analysis.

[0014] A cancer driver gene prediction method based on cross-network interaction fusion, the implementation process of step 4 is as follows: For each training node i Given binary labels To ensure that hard-to-classify samples receive a larger weight in the loss, the main classification employs a focus loss based on binary cross-entropy. Let... logarithmic odds The corresponding node-by-node binary cross-entropy, Then the training index set The focal loss on is: , The main loss uses the final logarithmic odds. ℓ The auxiliary loss uses the graph branch log odds ℓᵍʳᵃᵖʰ, and both terms use the same training index. The joint objective function is defined as: ,in, λ The auxiliary term represents the non-negative auxiliary loss coefficient. The auxiliary term provides direct gradients for cross-network attention, node-level gating, and the graph classification head; the main term is fused with learnable parameters. β Simultaneously update the main category branch and the graph branch. When λ When =0, the gradient of the final result with respect to the graph branch is still retained, but the independent graph classification constraint is no longer applied. The model uses... The optimizer is updated, and training uses a ten-fold partitioning method. Each fold reads non-overlapping training, validation, and test indices; in each training epoch, joint loss and backpropagation are calculated only on the training indices, and computation is performed on the validation indices. And record verification The highest period. When the optimal model needs to be restored, all parameter states for that period are saved and restored before the current fold test; the test index does not participate in parameter updates and period selection. After completing a fold, the model and intermediate tensors are released, and then the model for the next fold is initialized. To make the execution process of each fold reproducible, the same record seed is set for the random number generator before training, and this seed is used for model parameter initialization, discard masking, and data indexing; when using a graphics processor, deterministic computation is enabled and automatic baseline selection is disabled. Different folds still create separate models and optimizers to avoid the parameters and optimizer momentum of the previous fold being passed to the next fold. During inference, a forward computation is performed on the full input formed in step 1, sequentially obtaining three network representations, cross-network interaction representations, node gating weights, graph branch log odds, main classification log odds, and final log odds, with the final probability being... When binary classification labels are required, a threshold of 0.5 should be used. If a gene is identified as a cancer driver gene candidate, it is considered a non-candidate. Threshold determination does not alter the continuous probability ranking. (Test set) Calculated in pairwise order of predicted probabilities for positive and negative samples: ,in, and These represent the number of positive and negative samples in the test set, respectively. This is an indicator function. After ten folds are completed, each fold... Calculate the arithmetic mean. The system outputs the final predicted probability for each gene and the aforementioned average. Node gating weights and fusion weights It can be stored as an intermediate value, but it does not replace the final probability.

[0015] The system preserves the independent topology of each network, realizes representation interaction through cross-network attention, and uses node-level gating to allocate network weights. The graph classification head and auxiliary supervision provide direct constraints on the graph representation. It can learn the role of bi-branch fusion in regulating the graph classification results on the final prediction, thereby improving the ability to identify cancer driver genes and providing strong support for the discovery of potential therapeutic targets. Attached Figure Description

[0016] Figure 1 This is a flowchart of the overall process for predicting cancer driver genes in this invention.

[0017] Figure 2 Independent multi-head graph attention encoding and cross-network attention structure graphs for three biological networks.

[0018] Figure 3It includes node-level gating, independent graph branch classification heads, main classification branches, and learnable fusion structure graphs.

[0019] Figure 4 A flowchart for joint training, model selection and testing output. Detailed Implementation

[0020] The present invention will be described below with reference to the accompanying drawings and specific embodiments. These embodiments are used to illustrate the transmission relationship between data items in the steps described, and are not intended to limit the data source, number of nodes, or network size.

[0021] Input the build implementation. For example... Figure 1 As shown, using a unified index of gene names for all inputs, multiple omics feature files are first read and row matching is performed according to gene names, ensuring that the same row always corresponds to the same gene node. For the pan-cancer task, each gene consists of three types of features—gene expression differences, somatic mutation rates, and DNA methylation differences across 16 cancer types—connected in a fixed order to form a 48-dimensional original attribute vector. After deleting gene rows containing missing values, the mean and standard deviation of each feature column of the retained data are calculated and standardized. The standardized node attribute matrix is ​​denoted as... ,in For the number of gene nodes, in the pan-cancer task =48. For cancer-specific tasks, a column is taken from the expression, mutation, and methylation positions of the corresponding cancer type to form... The node attribute matrix is ​​set to 3. The three graph encoding branches receive the same node attribute matrix, therefore the differences in subsequent representations are only generated by the network topology and branch parameters. Three sparse adjacency matrices—protein interaction, pathway co-occurrence, and gene ontology semantic similarity—are read. Sparse matrix normalization is performed on each matrix, merging non-zero items corresponding to the same row and column coordinates, and then extracting the row index, column index, and relation score of the non-zero items. For the network... Save the row index and column index as edge index according to the edge correspondence, and save the relation score as edge score. The edge index is the first... r The first column and edge fraction r Each element represents the same edge; the three networks are respectively represented by... , and The edge is stored as a key, without generating a joint adjacency matrix for the three networks. The edge index is used to limit the message-passing neighborhood of the corresponding graph attention network branch, and the edge score is used for the edge bias of the subsequent main classification branch, and is standardized before enabling edge attributes. After reading the gene set name and association matrix, the genes are re-indexed according to the gene order in the multi-omics data, and missing associations are filled with 0; at the same time, the corresponding columns of gene sets containing cancer-related words such as CANCER, TUMOR, NEOPLASM, CARCINOMA, LEUKEMIA, or SARCOMA in the gene set name are deleted. Here, only the columns in the gene set association matrix are deleted, not the gene nodes, gene sequences, or edges of the three networks. The structural input is calculated using the protein interaction adjacency matrix: first, the row degree of each node is normalized, and the degree of zero-degree nodes is set to 1 to avoid division by zero, forming a random walk transition matrix. Then, principal component analysis is used to reduce the transition vector of each node to 128 dimensions to obtain the random walk position encoding. Subsequently, PageRank centrality is calculated on the same protein interaction network with a damping coefficient of 0.85, and the centrality of each node is saved as a one-dimensional column vector. This results in node attributes, three independent edge relationships, gene set functional input, 128-dimensional random walk position encoding, and one-dimensional centrality input; before training, the binary classification labels corresponding to the unified gene order and the training, validation, and testing indexes for each fold are read.

[0022] Implemented across network graph branches. For example... Figure 2 As shown, the node attributes are sent simultaneously. , and Two-layer multi-head graph attention branches with three independent parameters. The branch only accepts the edge indices of the protein-protein interaction network. The branch only accepts the edge index of the path co-occurrence network. The branches only receive edge indices from semantically similar networks of the gene ontology, and the three branches do not share graph attention parameters before interacting across networks. Each branch consists of two graph attention convolutional layers, each with four attention heads. The hidden layer dimension and output layer dimension are both 128, and each attention head outputs a 32-dimensional value. The branch dropout rate is 0.1. For the... l Layer Each attention point, first based on the central node Neighbor nodes And calculate the unnormalized attention score for the corresponding network edge relationships, and then at the node i Normalization is performed within the network neighborhood, and the normalization weight is denoted as . The first layer concatenates the results of the four attention heads along the feature dimension and uses... The activation function is applied; the second layer continues to compute attention within the same network neighborhood and connects the outputs of each head, resulting in a 128-dimensional representation of the network. No additional activation function is applied to the output of the second layer; the three branches output separately. , and Network representation. For each gene node, the three 128-dimensional network representations are arranged according to... , , The nodes are stacked in a fixed order to form a network view label matrix with a shape of 3×128; after being arranged on all nodes, the input tensor across the network module has the following shape: ,in The number of gene nodes is denoted as . To preserve network identity, three 128-dimensional learnable relation embeddings are set, and the corresponding relation embeddings are added to the corresponding network labels. Cross-network attention first performs layer normalization on the three network labels of each node, then inputs the three labels as a network view sequence of length 3 into a 4-head multi-head attention module; here, attention is only calculated along the network view dimension and different gene nodes are not directly mixed in this module. A 3×3 cross-network attention mask is set, with three diagonal positions set to negative infinity and the remaining positions set to 0, so that... The tag can only be from and Mark the information to read. The tag can only be from and Mark the information to read. The tag can only be from and Labels are used to read information, avoiding repeated reading of the network's own representation during cross-network interaction. The attention output and input labels are added as residuals and layer normalization is performed before passing through a feedforward layer of 128 to 256 dimensions. Activation and discarding are performed, followed by backprojection from 256-dimensional to 128-dimensional and a second residual connection and layer normalization are executed. The three output labels of each node are concatenated into a 384-dimensional vector, which is then passed through a 128-dimensional gated hidden layer. After activating and discarding layers, three gated logarithmic values ​​are generated, which are then processed along the network dimension. To obtain the corresponding node , and The gating weights of the three networks are defined. The weights and biases of the gating output layer are initialized to 0, making all three gating weights 1 / 3 at the beginning of training, and then adaptively updated by the classification error. The three output labels are weighted according to the gating weights to obtain a 128-dimensional graph branch representation; this is then passed through layer normalization, a 128-dimensional full-text unified dimension to a 64-dimensional linear layer, and so on. By activating and discarding layers and linear layers from 64-dimensional to 1-dimensional, the log-odds of graph branches are obtained. The independent graph classification head shares the output with the three sets of graph attention network branches, but does not copy the graph attention network parameters; the graph branch auxiliary loss is backpropagated along the graph classification head, node-level gating, and cross-network attention to the shared graph attention network, and the final classification loss simultaneously constrains the graph branch and the main classification branch through the fusion path.

[0023] Dual-branch fusion and training implementation. For example... Figure 3 and Figure 4 As shown, the main classification branch receives the outputs of the three independent graph attention network branches from step 2, and calculates the base network weights for each of the three network representations using the same linear scoring layer. For each gene node, the three network weights are applied along the network dimensions. Make the sum of the three weights equal to 1, then... , and Three 128-dimensional network representations are weighted to form a 128-dimensional base graph representation. This base graph representation is used for the main classification branch, while the graph branch representations generated by cross-network attention and node-level gating are separately fed into the graph branch classification head. The two branches remain structurally independent in subsequent classification processes. For the gene set association matrix obtained in step 1, the association weight of each gene to different gene sets is first calculated through a gene set attention linear layer, and then the process is performed along the gene set dimension. Normalization is performed and multiplied with the learnable gene set embedding table to obtain a 128-dimensional functional representation for each gene. Subsequently, the 48-dimensional multi-omics attributes, 128-dimensional basic graph representation, 128-dimensional functional representation, 128-dimensional random walk position encoding, and 1-dimensional PageRank centrality are concatenated according to feature dimensions to form a 433-dimensional node input vector. This vector is then mapped to a 128-dimensional representation through a linear projection layer and arranged in gene node order as the main classification branch input sequence. The main classification sequence is input to a 3-layer Transformer encoder, with 4 self-attention heads and a dropout rate of 0.1 in each layer. For each edge in the three biological networks, its row index, column index, and relation score are concatenated to form an edge description vector, which is then mapped to the edge bias corresponding to the 4 attention heads through a shared linear layer, and then... The function limits the bias range and accumulates it to the corresponding network position; positions without corresponding edges remain at 0. Each Transformer layer first performs multi-head self-attention with edge bias, then passes through residual connections and layer normalization, followed by feedforward layers from 128 to 512 dimensions. After activation and a linear mapping from 512 dimensions to 128 dimensions, residual connections and layer normalization are performed again. The third layer after processing by a 3-layer Transformer... Each node represents the primary class log odds obtained through a single-output linear layer. The graph branch classification head receives the 128-dimensional graph branch representation generated by cross-network interaction and node-level gating in step 2, which is then normalized through a layer and reduced from 128 dimensions to 64 dimensions by a linear layer. Activation and discarding layers, and linear layers from 64 dimensions to 1 dimension, yield the graph branch log odds. The principal class log odds and graph branch log odds are... Weighted fusion is performed before function transformation, and a learnable scalar is set. And calculate the graph branch fusion weights. During the training process Updated simultaneously with other model parameters. The final log odds are a linear combination of the log odds of the two branches according to their current weights, then... The function is converted into the prediction probability of cancer driver genes. Training uses a focal loss function, with the focal loss of the classification branch as the main loss and the focal loss of the graph branch as the auxiliary loss. The joint loss is obtained by adding the main loss to the auxiliary loss of the graph branch by 0.3 times. The model uses the AdamW optimizer, with each fold training for 30 epochs. In each epoch, the joint loss is first calculated using the training index and backpropagation is used to update the model parameters. Then, the validation AUC is calculated using the validation index, and the epoch with the highest validation AUC is recorded. The test index does not participate in parameter updates or epoch selection. In this embodiment, all model parameters of the epoch with the highest validation AUC are restored before each fold test to ensure that the model state selected in the fold is used during the testing phase.

[0024] Predictive implementation. For the test fold, the gene order, three adjacency matrices, gene set association matrix, random walk position encoding, and PageRank centrality from the training phase are preserved. After restoring the model parameters determined in the validation phase, the model is switched to inference mode, random deactivation of dropout layers is disabled, and gradient recording is stopped. The model first computes three independent graph neural network representations for all gene nodes. The graph branch sequentially completes relation embedding addition, cross-network attention, residual and feedforward processing, node-level gating, and graph branch classification. The main classification branch sequentially completes the aggregation of basic weights from the three networks, gene set function embedding, structural input concatenation, edge bias construction, and Transformer classification. Then, the currently learnable fusion weights are used to merge the main classification results and graph branch results in log-odds space, and finally, the merged final log-odds is processed. The function calculates the continuous predicted probability of each gene belonging to a cancer driver gene. During prediction, standardized parameters, principal component transformations, or network edge indices are not refitted to avoid altering the input space during training. Each fold only calculates its evaluation value at its corresponding test index; the forward results from other nodes are not included in the fold's test statistics. When binary labels are required, a threshold of 0.5 is used. Genes with a predicted probability greater than 0.5 are considered cancer driver gene candidates, while those with a predicted probability less than or equal to 0.5 are considered non-candidates. This threshold is only used for label output and does not change the continuous probability ranking. After completing the ten-fold test, the final predicted probability of each gene at each test fold is saved, and the arithmetic mean of the AUC of the ten test folds is taken as the evaluation result for the pan-cancer task. The AUC is calculated based on the predicted probabilities of positive and negative samples in the test fold and is used to measure the model's ability to distinguish between positive and negative samples; at the node level... , , Gating weights and the fusion weights of main and graph branches ω It is stored as an intermediate analysis parameter and does not replace the final predicted probability.

[0025] To verify the effectiveness of this invention, on the pan-cancer data test set provided by the GRAFT project, the AUC of this invention was improved from 0.917 to 0.927 compared to the baseline method, demonstrating the effectiveness and superiority of this invention.

[0026] The above description, in conjunction with specific embodiments, illustrates the present invention. For those skilled in the art, any equivalent adjustments to feature dimensions, number of attention heads, number of network layers, auxiliary loss coefficients, or classification thresholds, as well as equivalent substitutions for biological network data sources, made without departing from the inventive concept, should fall within the scope of protection of this invention.

Claims

1. A cancer driver gene prediction method based on cross-network interaction fusion, characterized in that, Includes the following steps: Step 1: Construct input data: Form multi-omics feature matrix, protein interaction adjacency matrix, pathway co-occurrence adjacency matrix, gene ontology semantic similarity adjacency matrix and gene set association matrix according to a unified gene order, and calculate random walk position encoding and PageRank centrality from the protein interaction network, outputting the attributes, functions and structures of each gene as input; Step 2: Generate cross-network graph representations: Set up two-layer multi-head graph attention encoders for each of the three biological networks, and aggregate neighbor features within their respective adjacency relationships; The three network representations of the same gene are used as three relation labels. Information is exchanged through multi-head attention with cross-network mask, and then node-wise gating is used to obtain the graph branch representation and the log odds of graph branch classification. Step 3: Forming the dual-branch fusion result: The main classification branch input is composed of the basic attention aggregation result of three graph representations, multi-omics features, gene set functional embedding, random walk position encoding, and PageRank centrality, to obtain the main classification log odds; the main classification log odds and the graph branch classification log odds are weighted by learnable parameters constrained by the Sigmoid function to obtain the final log odds; Step 4: Joint training and output predictions: The joint objective function is constructed with the final classification loss as the main loss and the graph branch classification loss as the auxiliary loss. The model parameters are updated on the training set and the model state is selected according to the validation set. The final log odds are calculated for the test genes and converted into the probability of cancer driver genes by the Sigmoid function.

2. The method according to claim 1, characterized in that, The implementation process of step 1 is as follows: Input construction implementation: Using the same gene name table as a unified index for all inputs, multiple omics feature files are first read and row matching is performed according to gene names to ensure that the same row always corresponds to the same gene node; for pan-cancer tasks, each gene is formed by connecting three types of features—gene expression differences, somatic mutation rates, and DNA methylation differences—in a fixed order to form a 48-dimensional original attribute vector; after deleting gene rows containing missing values, the mean and standard deviation of each feature column of the retained data are calculated and standardized. The standardized node attribute matrix is ​​denoted as ,in For the number of gene nodes, in the pan-cancer task For cancer-specific tasks, a column is taken from each of the expression, mutation, and methylation positions corresponding to the cancer type, forming... The node attribute matrix; the three graph encoding branches receive the same node attribute matrix, so the differences in subsequent representations are only generated by the network topology and branch parameters; read the three sparse adjacency matrices of protein interaction, pathway co-occurrence, and gene ontology semantic similarity, perform sparse matrix normalization operations on each, merge the non-zero items corresponding to the same row and column coordinates, and then extract the row index, column index, and relation score of the non-zero items; for the network Save the row index and column index as edge index according to the edge correspondence, and save the relation score as edge score. The edge index is the first... The first column and edge fraction Each element represents the same edge; the three networks are respectively represented by... , and It is stored as a key and does not generate a joint adjacency matrix of the three networks; Edge indexes are used to limit the message passing neighborhood of the corresponding graph attention network branches, and edge scores are used for the edge bias of the subsequent main classification branches. They are standardized before enabling edge attributes. After reading the gene set names and association matrices, the genes are re-indexed according to the gene order in the multi-omics data, and missing associations are filled with 0. At the same time, the columns corresponding to gene sets containing cancer-related words such as CANCER, TUMOR, NEOPLASM, CARCINOMA, LEUKEMIA, or SARCOMA in the gene set names are deleted. Here, columns in the gene set association matrix are deleted, but gene nodes, gene sequences, or edges of the three networks are not deleted. The structural input is calculated using the protein-protein interaction adjacency matrix: first, the row degree of each node is normalized, and the degree of zero-degree nodes is set to 1 to avoid division by zero, forming a random walk transition matrix. Then, the transition vector of each node is reduced to 128 dimensions through principal component analysis to obtain the random walk position encoding. Subsequently, PageRank centrality is calculated on the same protein-protein interaction network with a damping coefficient of 0.85, and the centrality of each node is saved as a one-dimensional column vector. This forms node attributes, three independent edge relationships, gene set functional input, 128-dimensional random walk position encoding, and one-dimensional centrality input. Before training, the binary classification labels corresponding to the unified gene order and the training, validation, and testing indexes of each fold are read.

3. The method according to claim 1, characterized in that, The implementation process of step 2 is as follows: Implement across network graph branches; simultaneously input node attributes , and Two-layer multi-head graph attention branches with three independent parameters; The branch only accepts the edge indices of the protein-protein interaction network. The branch only accepts the edge index of the path co-occurrence network. The branches only receive edge indices from the semantically similar gene ontology network, and the three branches do not share graph attention parameters before interacting across networks; each branch includes two layers of graph attention convolutions, each layer has four attention heads, the hidden layer dimension and the output layer dimension are both 128, each attention head outputs 32 dimensions, and the branch dropout rate is 0.1; for the , Layer Each attention point, first based on the central node Neighbor nodes And calculate the unnormalized attention score for the corresponding network edge relationships, and then at the node Normalization is performed within the network neighborhood; The first layer concatenates the results of the four attention heads according to their feature dimensions and uses the ELU activation function; the second layer continues to compute attention within the same network neighborhood and concatenates the outputs of each head to obtain the 128-dimensional representation of the corresponding network. The second layer output does not apply an additional activation function; the three branches output separately. , and Network representation.

4. The method according to claim 3, characterized in that: For each gene node, the three 128-dimensional networks are represented according to... , and The nodes are stacked in a fixed order to form a network view label matrix with a shape of 3×128; after being arranged on all nodes, the input tensor across the network module has the following shape: ,in The number of gene nodes is denoted as 1; to preserve network identity, three 128-dimensional learnable relation embeddings are set, and the corresponding relation embeddings are added to the corresponding network labels; cross-network attention first performs layer normalization on the three network labels of each node, and then inputs the three labels as a network view sequence of length 3 into the 4-head multi-head attention module; The attention here is calculated only along the network view dimension, and different gene nodes are not directly mixed in this module; settings The cross-network attention mask has three diagonal positions set to negative infinity and the remaining positions set to 0, making... The tag can only be from and Mark the information to read. The tag can only be from and Mark the information to read. The tag can only be from and Mark the information to avoid repeatedly reading the network's own representation during cross-network interaction.

5. The method according to claim 4, characterized in that: The attention output and input label are residually summed and layer normalization is performed. Then, the result is passed through a 128-dimensional to 256-dimensional feedforward layer, GELU activation, and dropout processing. Subsequently, it is back-projected from 256-dimensional to 128-dimensional and undergoes a second residual connection and layer normalization. The three output labels of each node are concatenated into a 384-dimensional vector, which is then passed through a 128-dimensional gated hidden layer, GELU activation, and dropout layer to generate three gated logarithmic values. Finally, Softmax is performed along the network dimensions to obtain the corresponding value for that node. , and The gating weights of the three networks are initialized to 0, so that the weights and biases of the gating output layer are all 1 / 3 at the beginning of training, and then adaptively updated by the classification error. The three output labels are weighted according to the gating weights to obtain a 128-dimensional graph branch representation. Then, the graph branch log odds are obtained by passing the layer normalization, a linear layer from 128-dimensional to 64-dimensional, GELU activation, a dropout layer, and a linear layer from 64-dimensional to 1-dimensional. The independent graph classification head shares the graph attention network output with the three graph attention network branches, but does not copy the graph attention network parameters. The graph branch auxiliary loss is backpropagated to the shared graph attention network along the graph classification head, node-level gating, and cross-network attention. Finally, the classification loss constrains both the graph branch and the main classification branch through the fusion path.

6. The method according to claim 1, characterized in that, The implementation process of step 3 is as follows: Dual-branch fusion and training implementation: The main classification branch receives the outputs of the three independent graph attention network branches from step 2, and calculates the base network weights for each of the three network representations using the same linear scoring layer; for each gene node, the three network weights are subjected to Softmax along the network dimension to make the sum of the three weights equal to 1, and then... , and Three 128-dimensional network representations are weighted to form a 128-dimensional base graph representation. This base graph representation is used for the main classification branch, while the graph branch representations generated by cross-network attention and node-level gating are fed separately into the graph branch classification head. The two branches remain structurally independent in subsequent classification processes. For the gene set association matrix obtained in step 1, the association weight of each gene to different gene sets is first calculated through a gene set attention linear layer, then Softmax normalization is performed along the gene set dimension, and multiplied with the learnable gene set embedding table to obtain the 128-dimensional functional representation corresponding to each gene. Subsequently, the 48-dimensional multi-omics attributes, the 128-dimensional base graph representation, the 128-dimensional functional representation, the 128-dimensional random walk position encoding, and the 1-dimensional PageRank centrality are concatenated according to the feature dimensions to form a 433-dimensional node input vector, which is then mapped to a 128-dimensional table through a linear projection layer. The input sequence for the main classification branch is arranged in gene node order. The main classification sequence is input into a 3-layer Transformer encoder, with 4 self-attention heads and a dropout rate of 0.1 in each layer. For each edge in the three biological networks, its row index, column index, and relation score are concatenated into an edge description vector, which is then mapped to the edge bias corresponding to the 4 attention heads through a shared linear layer. The bias range is then limited by the Sigmoid function and accumulated to the corresponding network position. Positions without corresponding edges are kept at 0. Each Transformer layer first performs multi-head self-attention with edge bias, then performs residual connection and layer normalization, followed by a feedforward layer from 128 to 512 dimensions, ReLU activation, and a linear mapping from 512 to 128 dimensions, and then performs residual connection and layer normalization again. The 3-layer Transformer processed 3rd layer of the ... Each node represents the log-probability of the primary class obtained through a single-output linear layer.

7. The method according to claim 6, characterized in that: The graph branch classification head receives the 128-dimensional graph branch representation generated in step 2 through cross-network interaction and node-level gating. This representation undergoes layer normalization, a 128-dimensional to 64-dimensional linear layer, GELU activation, a dropout layer, and a 64-dimensional to 1-dimensional linear layer to obtain the graph branch log odds. The primary classification log odds and the graph branch log odds are weighted and fused before the Sigmoid function transformation, and a learnable scalar is set. And calculate the graph branch fusion weights. During training It is updated simultaneously with other model parameters; the final log odds are a linear combination of the log odds of the two branches according to the current weights, and then converted into the cancer driver gene prediction probability by the Sigmoid function.

8. The method according to claim 1, characterized in that, The implementation process of step 4 is as follows: Training employs a focal loss function, with the focal loss of the classification branch serving as the primary loss and the focal loss of the graph branch serving as the secondary loss. The joint loss is obtained by adding the primary loss to the graph branch secondary loss at a factor of 0.

3. The model uses the AdamW optimizer. In each epoch, the joint loss is first calculated using the training index, and the model parameters are updated via backpropagation. Then, the validation index is used to calculate the validation AUC, and the epoch with the highest validation AUC is recorded. The test index does not participate in parameter updates or epoch selection. In this embodiment, all model parameters of the epoch with the highest validation AUC are restored before each epoch of testing to ensure that the model state selected in the epoch is used during the testing phase.

9. The method according to claim 8, characterized in that: Predictive implementation: For the test fold, the gene order, three adjacency matrices, gene set association matrix, random walk position encoding, and PageRank centrality from the training phase are preserved. After restoring the model parameters determined in the validation phase, the model is switched to inference state, random deactivation of dropout layers is disabled, and gradient recording is stopped. The model first computes three independent graph attention network representations for all gene nodes. The graph branch sequentially completes relation embedding addition, cross-network attention, residual and feedforward processing, node-level gating, and graph branch classification. The main classification branch sequentially completes the aggregation of basic weights of the three networks, gene set function embedding, structural input splicing, edge bias construction, and Transformer classification. Then, the currently learnable fusion weights are used to merge the main classification results and graph branch results in log-odds space. Finally, the Sigmoid function is applied to the merged final log-odds to obtain... Each gene is assigned a continuous predicted probability as a cancer driver gene; during prediction, standardized parameters, principal component transformations, or network edge indices are not refitted to avoid test data altering the input space during training; each fold only calculates its evaluation value on its corresponding test index, and the forward results of other nodes are not included in the fold's test statistics; when binary labels need to be output, genes with a predicted probability greater than 0.5 are considered cancer driver gene candidates, and genes with a predicted probability less than or equal to 0.5 are considered non-candidates; threshold determination is only used for label output and does not change the continuous probability ranking; after completing the ten-fold test, the final predicted probability of each gene in the corresponding test fold is saved, and the arithmetic mean of the AUC of the ten test folds is taken as the evaluation result of the pan-cancer task; AUC is calculated based on the predicted probabilities of positive and negative samples in the test fold and is used to measure the model's ability to distinguish between positive and negative samples; node level. , , Gating weights and the fusion weights of main and graph branches It is stored as an intermediate analysis parameter and does not replace the final predicted probability.

Images