How do Boltz-1 and Chai-1 perform compared to older models for predicting kinase structures?

Abstract

Implicated in a variety of disease processes, protein kinases are key immunological targets in therapeutics. However, selective kinase targeting remains a significant challenge in inhibitor development, due to the highly conserved nature of the ATP-binding site to which most inhibitors bind. Furthermore, given the Protein Data Bank’s (PDB) bias towards active kinase structures, the development of inactive-state inhibitor development which often exhibits longer residence times is limited. AI models with the ability to predict kinase structures reaching experimental accuracy can be leveraged upon in accelerating drug discovery, especially for the data-scarce “dark kinome.” In this study, the publicly available Boltz-1 and Chai-1 models are benchmarked against older models - AlphaFold 2/3, RosettaFold, and ESMFold. Boltz-1 and Chai-1 demonstrated AlphaFold-level accuracy and improved predictions for conserved motifs compared to older models, highlighting their potential in reliable protein structure prediction, increasing confidence for direct application within experimental labs.