PLMCA: A General Multimodal Protein–Ligand Cross-Attention Framework for Pocket Identification and Binding Affinity Prediction

Accurate prediction of drug-target binding affinity (DTA) remains a central challenge in drug discovery due to the need to integrate heterogeneous sequence, structural, and physicochemical information. Here, we propose PLMCA, a multimodal protein-ligand cross-attention framework that unifies protein sequence embeddings from two protein language models, three-dimensional geometric features, physicochemical descriptors, and ligand molecular graph representations within a single architecture. PLMCA further incorporates experimental assay conditions from the ChEMBL database as auxiliary inputs to mitigate batch effects and reduce measurement noise. On the PDBbind21 data set, PLMCA performs competitively or outperforms state-of-the-art methods under random, unseen-ligand, and unseen-protein splits for Kd and Ki prediction. On the ChEMBL_mini data set, PLMCA achieves R² values of 0.531, 0.635, and 0.519 for IC₅₀, Kd, and Ki prediction, respectively. In addition, PLMCA demonstrates robust protein binding pocket prediction performance, achieving an AUPR of up to 0.655 under the unseen-protein setting.