Proceedings of the National Academy of Sciences · 2 July 2026 · From issue 01
Beyond native sequence recovery: Improved modeling of the sequence-energy landscape of protein structures
Being worse at guessing original protein sequences made a model better at real design tasks.
Foster Birnbaum, Amy E. Keating.
A common way to test protein design software may be measuring the wrong thing. Many computer models learn to design new proteins by first learning to guess the exact original sequence used to build a known protein structure. This paper shows that getting better at that guessing game does not mean a model is actually better at real design tasks. Real tasks include predicting how strongly a protein holds its shape, and how changes to its building blocks affect its stability. The authors built a new model called PottsMPNN. It learns to score how well different building blocks fit together at each spot in the protein, instead of just guessing the original sequence. This model was worse at guessing the exact natural sequence. But it made better designed sequences. It also made better predictions about how changes affect protein stability.
Integrity screen: passed (3 checks) Checked 6 July 2026. Read the source
The plain-language reading of this paper is in the full issue, for subscribers.
Read the full issue