Robust and interpretable machine learning for affinity prioritization of designed protein interactors

Recent advances in computational protein design have resulted in a large pool of de novo interactors that necessitate efficient and robust tools for candidate prioritization. Here, we introduce Proffinity, a one-stop computational workflow employing explainable machine learning (ML) for prioritizing high-affinity binders for protein design. Proffinity integrates knowledge-based contact potential and ML to extrapolate detailed residue-level interactions from protein structures. Trained on the structural mutants and de novo protein interactors, our model achieves robust performance for binding affinity estimation compared to the experimental ground truth. Application of the Proffinity workflow in the design of the ubiquitin variant raised for Rsp5 E3 ligase successfully identified tight binders with potent binding affinity. Our tool offers a timely, cost-effective approach to rapidly identifying high-affinity binders from large candidate pools, thus facilitating state-of-the-art protein design (https://github.com/yuchen-lo/Proffinity).