Geometric Protein Optimization

The vastness of the space of possible protein variations is often regarded an insurmountable challenge for effective optimization. Traditional approaches focus primarily on substitutions near the binding pocket but are frequently hindered by epistatic effects and non-convex optimization landscapes. Current AI-aided methods can accelerate laboratory procedures but largely target the same substitutions. Here, we propose a complementary approach, Geometric Protein Optimization (GPO), an AI-native framework that fine-tunes the global geometry of the protein by combining a large number of substitutions from diverse locations along the sequence. GPO leverages the strong inverse power-law dependence of electrostatic forces where small adjustments can have large effects on binding affinity. We make the surprising discovery that the inclusion of distal substitutions leads to a smoother and approximately separable optimization landscape. Our empirical investigations reveal three stylized facts about this landscape that we use as a guide to develop BuildUp, a baseline algorithm for GPO. Results show that it is able to navigate this landscape much more effectively and achieve significant improvements in in silico binding affinity (Kd) across diverse protein-ligand complexes. Evaluations of the derived variants through protein-ligand interaction profiling, docking simulations, and molecular dynamics simulations confirm that GPO can achieve beneficial effects even with a sequence-based scoring function.