MIT's PottsMPNN improved protein design beyond native sequences

Protein design pipelines often fix a target structure, then ask a model to generate amino-acid sequences that fold into it. The field had often scored the result by native-sequence recovery—whether the model could reproduce the sequence evolution chose. That metric rewards resemblance to nature, but it breaks down for genuinely novel proteins whose structures have no native sequence.

MIT Department of Biology developed PottsMPNN to model the sequence-energy landscape instead: the relationship between amino-acid identity and protein stability. The training recipe adds noise to protein structures to reduce native-sequence mimicry, models pairwise interactions across the 20 amino acids, and feeds in evolutionarily related sequence sets so the model learns that different sequences can adopt the same fold.

The reported result is a better design target and better behavior: PottsMPNN reduces native-sequence mimicry while improving structural-compatibility and energy prediction, including for novel proteins, and improves prediction of mutation effects on stability. The research moves protein design from asking ‘did it copy evolution?’ to asking ‘will it fold and remain stable?’

Optimize foldability, not native-sequence recovery

Novel structures have no native sequence to copy, so recovery cannot be the success test.

Noise discourages the model from overfitting to evolution's selected sequences.

Pairwise amino-acid interactions model the physical dependencies that determine stability.

Related-sequence families teach that multiple sequences can realize one fold.

If a benchmark rewards resemblance to history, it can punish useful novelty. Replace it with a causal constraint—here, foldability and sequence-energy fit—and train the model to explore it.

The authors say the framework could be fine-tuned for specific tasks, such as predicting the consequence of a mutation. Its larger promise is a design pipeline that can propose useful new-to-nature proteins without needing a natural sequence as the template.