DeepRLI: a multi-objective framework for universal protein–ligand interaction prediction

Abstract

Protein–ligand interaction prediction is a critical component of computer-aided drug design. Although recent deep learning scoring functions have demonstrated advantages over conventional scoring functions, accurate and efficient prediction of protein–ligand binding efficacy remains an intractable challenge. Most of those methods are tailored for specific tasks, such as binding affinity prediction, binding pose prediction, or virtual screening, and often fail to encompass all aspects. There are longstanding concerns that deep learning methods lack a comprehensive understanding of binding free energy and have limitations in generalization. Deep learning methods with a single optimization goal tend to struggle to achieve balanced performance in scoring, ranking, docking, and screening, thus failing to meet the needs of practical drug design research. To solve this challenge, we propose DeepRLI, a novel interaction prediction framework that is universally applicable across various tasks. The proposed model is trained with a multi-objective learning strategy that includes scoring, docking, and screening as optimization goals. It allows DeepRLI to have three relatively independent downstream readout networks, which can be optimized separately to enhance the task specificity of each output. Additionally, the model incorporates an improved graph transformer with a cosine envelope constraint, integrates a novel physics-informed module, and introduces a new contrastive learning strategy. With these designs, extensive evaluations across various benchmarks demonstrate that DeepRLI has superior comprehensive performance in broad applications, highlighting its potential as a fundamental tool for evaluating protein–ligand interactions in practical drug discovery and development.