Accurate prediction of protein-ligand binding affinity is essential in drug discovery. However, the limited availability and high cost of experimentally resolved protein-ligand complex structures significantly hinder the generalizability and broad applicability of current structure-based deep learning approaches. To address this challenge, we present CompBind, a novel framework for binding affinity prediction that leverages latent interaction patterns learned from existing complex structures while eliminating the need for 3D structural inputs during inference. Specifically, CompBind integrates bidirectional cross-attention with a dual-objective pretraining strategy, where contrastive learning enforces feature-space consistency between monomer pairs and their corresponding complex structures, while generative learning reconstructs interaction features to model the bidirectional mapping between monomeric and complex representations. This enables the model to infer binding representations directly from protein and ligand sequences alone. Across challenging affinity prediction scenarios, including cold-start and sparse-label conditions, CompBind not only outperforms noncomplex-based methods but also competitively rivals complex-based prediction approaches. In a drug repurposing case study targeting glutathione peroxidase 4 (GPX4), a clinically relevant but traditionally undruggable protein, CompBind successfully ranked known inhibitors among the top candidates. Furthermore, the built-in attention mechanism enhances model interpretability by identifying key binding residues. By decoupling predictive accuracy from the availability of experimental complex structures, CompBind offers a scalable, generalizable, and practical solution for accelerating drug discovery pipelines.
Yi et al. (Wed,) studied this question.