Drug-Protein Interaction Prediction via Multi-View Variational Autoencoder and Cascade Deep Forests

\abstract{Accurate prediction of Drug-Protein Interactions (DPIs) is crucial for drug discovery and precision medicine. Despite recent progress, the high-dimensionality representation drug and protein characteristics and their interactions poses a significant challenge. In this study, we propose a novel deep learning framework, Multi-view Variational Auto-Encoder embedded Deep Forest (MVAE-DFDPnet) for DPI prediction. This framework learns low-dimensional drug/protein embeddings from heterogeneous drug/protein networks by multi-view variational auto-encoder, then feeds drug-protein embeddings into a cascade deep forest to predict yet unknown DPIs. Our experiments show that the MVAE-DFDPnet achieves 98.6\% out-of-sample accuracy in predicting DPIs on benchmark datasets, outperforming state-of-the-art methods with much lower embedding dimensions. We further demonstrate the robustness and generalizability of the method by testing on unseen drugs and proteins as well as on unseen drug classes, where the model still maintains high prediction accuracy. We have also successfully validated a number of novel DPIs by experimental evidence found in literature, demonstrating the potential of MVAE-DFDPnet in real-world applications.