PottsMPNN Improves Protein Design for 20 Amino Acids, Moving Beyond Native Sequences
Updated
Updated · MIT News · Aug 27
PottsMPNN Improves Protein Design for 20 Amino Acids, Moving Beyond Native Sequences
1 articles · Updated · MIT News · Aug 27
Summary
PottsMPNN, a new machine-learning framework reported in PNAS, improved protein sequence generation and mutation-stability prediction by modeling how sequences map onto a target fold rather than trying to reproduce evolution’s chosen sequence.
A key change is its pairwise distribution over all 20 amino-acid options at two positions, which better captures physical interactions and the sequence-energy landscape that governs protein stability.
The model also adds training noise to reduce overfitting to native proteins and uses evolutionarily related sequence sets to teach that many different sequences can adopt the same structure.
Tests showed that as reliance on native-sequence similarity fell, structural compatibility and energy prediction improved, including for proteins with no natural sequence counterpart.
That could widen protein design beyond nature-derived templates, helping researchers build new-to-nature proteins for tasks such as binding disease-related molecules and other biological engineering applications.