Abstract
Machine learning has revolutionized protein structure and interaction prediction, yet its full potential for drug discovery is still emerging. In this study, we show that denoise diffusion-based co-folding methods-such as AlphaFold3 and Boltz-1/2-not only achieve highly accurate protein-ligand interaction predictions but can also separate active compounds from inactive ones. We introduce a simple and effective strategy, pairwise competitive docking, which ranks candidate molecules by directly comparing their relative binding to a protein's target site. Applied to 17 protein benchmark systems, the method generated rankings consistent with experimental trends, although the degree of agreement varied considerably by system, with concordance indices ranging from 0.52 (indicating no meaningful correlation) to 0.89 (indicating strong correlation). Notably, our rankings showed strong agreement with Boltz-2 affinity predictions, positioning our method as a practical alternative for inhibitor prioritization. Finally, we show how pairwise competitive docking can accelerate the identification of promising hits within a large chemical library and guide the de novo design of inhibitors with improved predicted potency. Collectively, these findings highlight how modern machine-learning models can make structure-based drug design faster, more reliable, and more cost-effective than relying solely on experimental workflows.