Synthetic Lethality Gene Interaction Prediction Algorithm Based On Multi-view Graph Convolutional Networks

Synthetic lethality is a specific interaction between genes that can result in cell death when two genes that form a synthetic lethality interaction become defective at the same time.The development of targeted anti-cancer drugs using synthetic lethal interactions is an important part of modern cancer treatment theory.In traditional wet-lab experiments,unknown synthetic lethality gene pairs need to be detected through RNA screening and other means,which are costly and inefficient.Using in-silico methods to predict synthetic lethal interactions can provide a target guide for wet-lab experiments,thus improving the efficiency of detection experiments and reducing the cost.As research progressed,more and more genetically relevant data was generated and researchers began to use this data to develop algorithms to predict synthetic lethal interactions,attempting to mine hidden patterns from known synthetic lethal interactions to achieve more accurate predictions.Recently,researchers attempted to combine multi-omics data to further improve the accuracy of synthetic lethality prediction,making synthetic lethality prediction became a multi-view learning problem.However,the introduction of multi-omics data inevitably introduces a large amount of noise at the same time,making it difficult to extract information related to synthetic lethality from the multi-omics data.To address the above difficulties,this paper proposed a supervised multi-view graph auto-encoder framework and implemented the basic algorithm under this framework as well as an improved algorithm incorporating the attention mechanism.Graph auto-encoder considers genes as nodes and gene-gene interactions as edges and models the inter-gene relationships as a graph.We introduced a known synthetic lethality interaction as a supervised signal,and also performed supervised training on local single-view and global multi-view reconstruction to obtain information related to synthetic lethality in each view at a fine-grained level.Finally,the reconstructed synthetic lethality interactions from multi-views are fused to obtain the final prediction results.However,in the current multi-omics genetic data,the correlation between the data from each view and the synthetic lethality interactions is different.Therefore,in the enhanced algorithm,we further distinguish the multi-view data between the main view and support views and enables the algorithm to better distinguish between target and noisy data through parameter control.At the same time,the attention mechanism further improves the effectiveness of the integration of multi-view data,so that data with high correlation with synergistic lethality in multi-view data are given more weight and contribute more to the prediction results to improve the prediction accuracy of the algorithm.The effectiveness of the algorithm framework is verified through comparative experiments on Syn Leth DB and Gene Interactions datasets.Besides,the validity and rationality of the model structure are verified through model elimination experiments,and the role of each hyperparameter of the model is tested through parameter sensitivity analysis experiments.Finally,the case studies on novel predicted synthetic lethality interactions also illustrate the effectiveness of our proposed method.