PNAS study reports protein-design model that improves stability predictions

Researchers reported in PNAS that their PottsMPNN machine-learning framework improved protein-sequence generation and prediction of how mutations affect protein stability. The system models pairwise amino-acid interactions and uses evolutionarily related sequences during training to represent different sequences that can adopt the same folded structure. The team says the method can generate structurally feasible sequences that do not resemble native proteins and demonstrated improved structural compatibility and energy prediction in its study.