Multi-omics fusion based on attention mechanism for survival and drug response prediction in Digestive System Tumors

Abstract

In recent decades, digestive system tumors (DST) have become the primary cause of cancer-related deaths worldwide. Improving tumor prognosis and drug response prediction holds significant importance in personalized medicine. In order to construct an effective and interpretable model that integrates multiple omics data, the Multi-omics Fusion Graph Attention Network (MFGAN) is proposed for survival prediction and drug response prediction in DST. In this method, the similarity matrix of each kind of omics data is calculated and a new graph structure is learned by Graph Transformer (GT). Next, to learn the features of every kind of omics data from the TCGA and GDSC databases, the Graph Attention Network (GAN) is used. Lastly, the View Correlation Discovery Network (VCDN) combines different types of omics data features to predict survival risk and drug response. Among them, the survival prediction results showed an improvement of up to 9% relative to other methods in terms of the c-index metric, with a minimum improvement of 2.6%. In terms of drug response prediction performance, there was an improvement of 4%. The ablation experiment demonstrates MFGAN’s feature integration capability, and the functional enrichment analysis for significant genes explains the functional characteristics of the model. Furthermore, the prediction of unknown tumor drug response demonstrates the model’s prediction ability. The proposed high-prediction comprehensive model could have important potential value for the personalized medicine of DST.