Effective drug-target affinity prediction via generative active learning

Abstract

Drug-target affinity (DTA) prediction is a critical early-stage task in drug discovery. Recently, deep learning has demonstrated remarkable efficacy in DTA prediction. However, acquiring experimentally verified data for target proteins proves to be a time-consuming, labor-intensive, and costly endeavor. In this study, we introduce an innovative generative active learning method for DTA prediction, referred to as GAL-DTA. GAL-DTA comprises two modules, data augmentation and generator fine-tuning. In the data augmentation module, the algorithm uses an optimized generator to produce informative and diverse molecules, thereby enhancing training of the predictor. The generator fine-tuning module introduces Fisher's informativeness and molecule diversity as objectives and employs the Pareto ranking algorithm to compute rewards. The final generator is fine-tuned using the policy-gradient method. GAL-DTA performs data augmentation by directly generating diverse and informative data, effectively reducing annotation costs while preserving model performance. Extensive experiments on independent test sets involving four target proteins consistently demonstrated that GAL-DTA achieves superior performance, resulting in an average reduction of 8.402% in mean squared error.