Improving Drug-Target Interaction Prediction Using Enhanced Feature Selection

A deep understanding of drug-target interaction (DTI) mechanisms is essential for the design of effective and safe drugs. While traditional wet lab experiments play a critical role, they are often associated with significant time and financial investments. Therefore, computational methods based on interaction prediction can significantly reduce the search space for the mentioned experiments. This paper outlines an innovative multistage approach for the prediction of DTI. Feature vectors, comprising various descriptors and the drug-encoded FP2 fingerprint, are extracted from protein sequences and drug structures in the first step. One major obstacle in this field is the data imbalance caused by the limited number of known interactions. The FFS-RF-DT algorithm, a powerful combination of forward feature selection, random forest and decision tree classification, is subsequently used to optimize predictive performance. This effectively eliminates irrelevant features, resulting in a more focused and efficient model. At the final stage, the XGBoost classifier processes the balanced dataset containing the optimal features to predict DTIs, yielding new drug-target interactions. The proposed model demonstrated better performance, achieving an accuracy of 96.96%, representing a 5.5% improvement over other models. In addition to its superior classification performance, the proposed model exhibits a much lower time complexity than comparable models. This makes it a promising candidate for DTI applications.