Researchers at MIT's Department of Biology reported a machine-learning framework, PottsMPNN, that incorporates physical principles governing protein structure and stability. The work, described in a paper published in PNAS and reported on Aug. 27, 2026, aims to improve sequence generation and prediction of how mutations affect stability.

Senior author Amy E. Keating said the field has long measured success by whether a model reproduces the sequence evolution selected, and argued this is not the best metric for protein design. Lead author Foster Birnbaum said that for a completely novel designed structure there is no native sequence for comparison, and that what matters is how likely generated sequences are to fold into desired structures.

The framework uses noise during training, a pairwise distribution to capture interactions between amino acids, and sets of evolutionarily related sequences. The authors reported that as the model depends less on native sequences, structural compatibility and energy prediction improved, including for novel proteins. Keating said the methods move the field toward designing new-to-nature proteins while providing a foundation for future advances. The paper appears in PNAS; no later outcome is asserted here.