Unifying Structure- and Ligand-based Drug Design via Contrastive Geometric Learning

ICLR 2026 Conference Submission19074 Authors

Keywords: contrastive learning, geometric deep learning, protein-ligand interaction, structure-based drug design, ligand-based drug design

TL;DR: We introduce ConGLUDe, an approach that combines structure- and ligand-based training data with geometric and contrastive learning, enabling multiple drug discovery tasks to be solved with a single model.

Abstract:

Structure-based computational drug design, which employs methods trained on large datasets of protein-ligand complex structures,
has been revolutionized by breakthroughs such as AlphaFold. In parallel, ligand-based computational drug design, driven by models
trained on extensive bioactivity resources, has impacted drug discovery by enabling the simultaneous prediction of numerous biological effects of small-molecule ligands. Yet, despite recent advances in both structure- and ligand-based approaches, no existing method integrates them effectively at scale. We introduce **Con**trastive **G**eometric **L**earning for **U**nified Computational **D**rug D**e**sign (ConGLUDe), an approach that leverages both structure- and ligand-based training data through geometric and contrastive learning. The ConGLUDe architecture combines a geometric protein encoder, producing both spatial binding pocket and global protein representations, with a ligand encoder. The encoders are trained jointly via contrastive learning on 20K protein-ligand complexes from PDBbind and 77M ligand-based datapoints from ChEMBL, PubChem, and BindingDB. With ConGLUDe, multiple key drug discovery tasks, including virtual screening, binding pocket prediction, ligand-conditioned pocket selection and target fishing, can be addressed within a single model. ConGLUDe achieves state-of-the-art performance on virtual screening benchmarks and strong results across other tasks, demonstrating the benefit of joint structure-ligand training. By replacing a set of specialized models with a single system and by unifying structure- and ligand-based paradigms, ConGLUDe represents a major step toward foundation models for drug discovery.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 19074