Synthetic Lethal Prediction Method Based On Heterogeneous Graph Neural Network

Cancer is a complex class of diseases affected by multiple genetic mutations.Searching for synthetic lethality associated with mutated genes in cancer and targeting them as drug targets has important implications for the prevention and treatment of cancer.The method of high-throughput wet experiment has limitations such as high cost,inconsistent cell lines,off-target effects,and time-consuming.With the development of related technologies such as machine learning and neural networks,models of various computational methods have been successfully developed and used to synthesize predictions of lethality.These computational methods are also a good complement to high-throughput experimental screening of synthetic lethality in human cancers.Current computational methods can effectively predict synthetic lethal interactions to a certain extent,but they ignore the potential relationship between the diverse biological data related to genes and synthetic lethal associations,and cannot make good use of the diverse and abundant biological Learning data to predict synthetic lethal associations.In response to these problems,this paper proposes two synthetic lethal potential association prediction methods based on biological data:(1)We proposed a Graph Convolution Networks on Multiple Graph for Predicting Synthetic Lethality named MGGCN.MGGCN first calculates gene GO functional similarity feature and PPI topological structure feature on GO map and PPI map,respectively.Then,based on bi-graph features and known synthetic lethal correlations,a bi-graph convolutional neural network is used to learn the information of neighbor nodes,and the co-dimensional feature representation matrix of GO features and PPI features is obtained.Secondly,this method extends the graph attention to two graphs,learns the contribution of the two graphs to the genes,and obtains the feature fusion matrix.A graph convolutional neural network is then used on this feature matrix to obtain a lower-dimensional feature representation.Finally,the predicted synthetic lethal matrix is obtained through the inner product decoder,and the gene pair most likely to become a synthetic lethal pair is selected according to the threshold as the final prediction result.Experimental results show that this method can accurately identify potential synthetic lethal associations.(2)A synthetic lethal interaction prediction method BSGAT based on feature similarity and bidirectional attention is proposed.BSGAT first uses convolutional neural network and graph attention neural network to obtain randomly initialized twodimensional gene feature information for protein amino acid sequence and protein correlation network,respectively.Then,this model introduces the concept of feature similarity and models the feature similarity,obtains the information transfer matrix between the two features,and calculates the two-way attention degree between the two features on this basis,and innovatively introduces The multi-head attention mechanism updates the information of nodes.Finally,using the obtained node embeddings,multilayer perception is used to calculate the degree of correlation between genes,the greater the degree of correlation,the more likely it is to be a synthetic lethal pair.Experimental results show that BSGAT can accurately identify potential synthetic lethal pairs and is more stable in small samples.