Liu, J.; Pan, T.; Guo, X.; Ran, Z.; Hao, Y.; Yang, Y.; Ng, A. P.; Pan, S.; Song, J.; Li, F.

Understanding how small molecules interact with protein binding pockets is central to structure-based drug discovery. Accurately modelling these interactions requires capturing the 3D geometry and physicochemical complementarity of binding interfaces, yet existing computational approaches encode proteins and ligands separately or rely on simplified structural representations that do not explicitly model cross-entity spatial relationships. Such decoupled representations restrict their capacity to capture interface-level geometric constraints that arise from protein-ligand co-organisational. Here we present MolX, a Graph Transformer foundation model that jointly learns geometric and chemical representations of protein pockets and ligands from large-scale 3D structural data. Integrating over 3 million protein pockets and 5 million molecules, MolX represents both entities as E(3)-equivariant graphs to preserve spatial geometry and chemical context. The architecture employs dual E(3)-equivariant graph Transformer encoders to model pocket and ligand embeddings, ensuring representations remain invariant to rotation, translation, and reflection. MolX is pretrained using a hybrid learning paradigm that combines supervised biochemical objectives, logP and energy-gap regression, with self-supervised geometric objectives, coordinate reconstruction, and atom-type prediction, fostering generalisable molecular understanding. Across eight downstream benchmarks, including antibody-drug conjugates (ADC), proteolysis-targeting chimera (PROTAC), molecular glue, and PCBA activity prediction, as well as binding affinity and physicochemical property regression, MolX achieves consistent state-of-the-art performance and strong cross-domain generalisation. Furthermore, MolX incorporates a sparse autoencoder module to decompose latent representations into interpretable biological components, thereby revealing the pocket-ligand interactions that drive prediction outcomes. Together, MolX establishes a scalable and interpretable foundation model for molecular representation learning, providing a unified framework for predicting and interpreting complex small-molecule-protein interactions.