Accurate and fast prediction of drug-target binding affinities (DTAs) is key for drug discovery; however, many methods, such as docking and empirical scoring, fail when generalizing to unseen cases. In this study, we introduced the MolXProt architecture, a novel transformer-based graph neural network that integrates graph ligand representations with protein language models using bidirectional multihead cross-attention. Our model is shown to be scalable to over 100000 protein-ligand pairs of mixed data sets, achieving 50% of predictions within ±1.0 kcal/mol and 80% within ±2.0 kcal/mol. We show that the architecture can explicitly learn residue-atom interactions while being computationally friendly via protein-token compression. By mapping token-residue interactions, we demonstrated that the model learns key binding pocket residues in benchmark complexes, such as CDK2-Staurosporine and DHFR-Methotrexate, but under-represents the hydrogen-bonding networks. Our calibration bias analysis revealed that the model overpredicted strong binders and underpredicted weak binders, which are linked to data imbalances and heteroscedastic noise. A simple posthoc isotonic correction partially mitigated the bias. Latent space analysis showed that the model learned continuous binding affinity manifolds without split leakage. Our work highlights a novel architecture that offers unique insights into binding mechanisms via transformer-based cross-attention and is computationally inexpensive.
Bruno Cucco (Mon,) studied this question.