SVCPI: A Soft Voting Ensemble-Based Model for Compound-Protein Interaction Prediction

Understanding the interaction between compounds and proteins is a crucial step in the process of discovering and developing drugs. To assist biologists and medicinal chemists, predicting compound-protein interaction (CPI) via computational methods has proven highly valuable. Deep learning has demonstrated its effectiveness in this domain. Modern studies have striven to improve performance with complex schemes for extracting features or fusing different features to obtain rich information. Notably, integrating multiple basic classifier models using voting approach can improve model performance. In this research, we introduced the SVCPI, a soft voting ensemble model. SVCPI employs a soft voting strategy to integrate basic classifiers that rely on GCN features and molecular fingerprint features. Through comprehensive testing on diverse datasets, we have determined that the SVCPI algorithm outperforms the basic classifiers. Furthermore, SVCPI delivers competitive results when compared to classical machine learning techniques and recently reported state-of-the-art approaches across five benchmark datasets. More importantly, on the Kinases dataset, SVCPI showcases superior performance. Compared with existing leading-edge methodologies, the AUC-ROC and AUC-PR of SVCPI are increased by more than 10% and 20%. These experimental findings strongly indicate that SVCPI is an effective and feasible method for CPI prediction task.