MIT PottsMPNN Targets More Diverse, Stable Protein Designs

MIT researchers are questioning a common protein-design objective

MIT highlighted new machine-learning work on PottsMPNN on August 27, 2026. The underlying peer-reviewed paper by Foster Birnbaum and Amy Keating was published in the Proceedings of the National Academy of Sciences in July 2026, so this should be understood as newly highlighted research rather than a paper first released on August 27.

The central claim is methodological: many machine-learning systems for protein sequence design are trained and evaluated on native sequence recovery, or how well they reproduce the amino-acid sequence naturally paired with a known protein structure. The MIT researchers argue that optimizing too strongly for that target can be misaligned with the broader goal of finding many different sequences that could plausibly adopt a desired structure and remain energetically favorable.

PottsMPNN models residue interactions as an energy function

PottsMPNN represents a protein backbone as a graph. Nodes correspond to residues, while edges capture relationships between residues. The model learns both single-residue probabilities and a Potts-model energy function containing single-residue and pairwise interaction terms.

That energy representation can be used not only to generate candidate sequences but also to score sequence variants, including mutations. The architecture combines ideas from ProteinMPNN-style graph neural networks with an explicit pairwise energy model.

The researchers evaluated systems using measures intended to better reflect whether a generated sequence is compatible with the target fold and whether the model can predict the energetic effect of mutations. In their experiments, PottsMPNN achieved lower native-sequence recovery than some alternatives but improved sequence-structure self-consistency and energy prediction.

Adding noise helped reduce overfitting to natural sequences

One of the paper's counterintuitive findings is that adding coordinate noise to protein backbones during training could improve downstream design metrics even as native-sequence recovery fell. The authors interpret this as evidence that a model can overfit to reproducing the one known natural sequence instead of learning the wider sequence-energy landscape compatible with a structure.

They also trained with information from multiple sequence alignments, exposing the model to related sequences that share a fold. This gives the training process direct evidence that multiple amino-acid sequences can map to similar structural outcomes.

Across held-out tests, the researchers report that these changes improved sequence-structure consistency and mutation-energy prediction despite reducing the metric that conventional sequence-design training often tries to maximize.

Why the benchmark question matters

Protein design is a search problem over an enormous sequence space. Natural proteins are valuable examples, but evolution does not necessarily identify the single sequence that is optimal for every engineering objective. If a model is rewarded mainly for copying natural sequences, it may fail to explore other viable regions of sequence space.

PottsMPNN therefore contributes two things: a model architecture that explicitly estimates pairwise residue interactions and an argument for evaluating protein sequence models on biologically meaningful design properties rather than relying on native-sequence recovery alone.

The released research code includes training and inference tools plus examples for sequence design, sequence optimization and energy prediction, making the methodology inspectable by other researchers.

What the study does not establish

This work is computational protein-design research, not a new drug, therapeutic or experimentally validated general solution to protein engineering. The reported improvements come from the study's benchmark and held-out evaluation settings. Designing a sequence that appears structurally or energetically promising in silico does not by itself establish that a protein will express correctly, remain stable under all real conditions, perform a desired biological function or be safe.

The broader impact will depend on replication, experimental validation and how well the approach transfers to difficult real design tasks. The paper is nevertheless important because it challenges a widely used optimization target and provides a concrete alternative for learning the sequence-energy landscape.

The correct status is: PottsMPNN is a published research framework with public code and peer-reviewed results; it is not a commercial protein-design product or a claim of laboratory validation for every generated sequence.