Leveraging 3D Molecular Spatial Visual Information and Multi-Perspective Representations for Drug Discovery.

Abstract

Drug discovery remains a costly and time-intensive process, where accurate identification of drug associations is critical for therapeutic development. Existing computational approaches predominantly rely on sequence-derived or 2D molecular representations, often overlooking the intrinsic 3D complexity of small molecules. Here, a deep learning framework is presented that directly learns from 3D molecular spatial visual information, capturing geometric, topological, and stereochemical features from spatial renderings. By integrating this spatial information with traditional molecular descriptors, unified multi-perspective representations are constructed that better reflect molecular structure and function. Across benchmark tasks involving drug-microRNA, drug-drug, and drug-protein interaction prediction, this model consistently outperforms conventional fingerprint-based baselines. Interpretability analyses show that the model attends to biologically relevant substructures, highlighting the value of 3D molecular spatial visual information in molecular recognition. These findings demonstrate the potential of spatially informed learning to enhance predictive performance and provide mechanistic insights in computational drug discovery.