PottsMPNN

For any given backbone, the sequence nature happened to select is one of many that will fold into it, and it is usually not the most stable one. Sequence design models are nonetheless trained and ranked on native sequence recovery (NSR) — the fraction of positions at which a generated sequence matches the natural one. PottsMPNN, from Foster Birnbaum and Amy E. Keating in the Department of Biology at MIT, was built to argue that pushing NSR higher makes a model worse at the two things designers actually need: producing sequences that fold to the intended structure, and predicting how a mutation changes stability.

The model keeps ProteinMPNN's graph featurization and message passing but changes what comes out of the graph. A single linear layer converts each final edge embedding, self-edges included, into a Potts energy function: a table of per-residue self-energies and a 20×20 matrix of pair energies for every edge. Training uses a composite pseudo-likelihood loss that supervises amino acid identities at pairs of positions as well as singly, so the network learns an explicit energy over whole sequences rather than a set of independent per-position distributions. That energy can be sampled from, minimized directly in sequence space, or differenced between wild-type and mutant to give a ΔΔG.

Two further changes deliberately move the model away from native sequences: Gaussian noise on backbone coordinates during training, and a loss averaged over sequences drawn from a multiple sequence alignment of the target, giving the model explicit examples of different sequences adopting one fold. Both lower NSR and raise sequence-structure self-consistency and energy prediction — the paper's central demonstration that the standard metric and the useful ones pull in opposite directions.

#Key Features

• Pairwise energy output: The Potts head produces a full energy table over all amino acid pairs at every graph edge, capturing coupling between positions that a per-position probability distribution cannot express.
• Unified design and ΔΔG scoring: One frozen checkpoint both generates sequences and predicts mutational stability effects, with no energy data used in training and no per-target refitting.
• Homolog-averaged loss: Averaging the loss over many sequences sampled from a filtered MSA each iteration teaches sequence-structure degeneracy directly; sampling one homolog per iteration does not work.
• Direct energy optimization: Because the Potts function is explicit, sequences can be optimized against it rather than only sampled autoregressively, including joint optimization of fold stability and inter-chain binding energy at a chosen interface.
• Drop-in weight set: A ProteinMPNN-compatible checkpoint lets the model substitute for ProteinMPNN in an existing pipeline.

#Technical Details

PottsMPNN uses ProteinMPNN's architecture and hyperparameters: a 48-nearest-neighbour residue graph with edges featurized by radial basis functions over 25 backbone heavy-atom distance pairs including virtual Cβ, a three-layer message passing encoder, a three-layer autoregressive decoder, and a hidden dimension of 128. The released checkpoints hold 1,712,342 parameters. Models were trained on the CATH 4.2 set of 19,700 single-chain structures and on the larger PDB-clust set of 25,361 clusters at 30% sequence identity, with filtered MSAs of median depth 147 and 133 sequences respectively. Energy prediction was evaluated on three experimental datasets after removing proteins present in training — 202,804 point mutations across 232 proteins from Megascale, 2,301 across 64 proteins from FireProt, and roughly 4,000 single mutants from a SARS-CoV-2 receptor binding domain deep mutational scan. PottsMPNN's NSR is low by design, around 44% against roughly 58% for the highest-NSR baseline tested, while it achieves the best sequence-structure self-consistency by AlphaFold2 refolding and Rosetta rethreading, and the strongest ΔΔG correlations, of the models compared. Vanilla and soluble-only lines are released, each at three settings of training noise and MSA use, all at the epoch 100 checkpoint.

#Applications

PottsMPNN occupies the same slot as ProteinMPNN in a design pipeline: it fits a sequence to a backbone produced by a generative structure model, and its output is checked by refolding. The explicit energy function extends that role in two directions — zero-shot stability scoring for protein engineering, where a deep mutational scan over a structure needs no experimental training data, and energy-guided mutation search for binder optimization, for which the repository ships a loop coupling the search to AlphaFold3 prediction and interface scoring.

#Impact

The contribution is as much a critique of an evaluation habit as a model. Sequence design methods had been compared largely on NSR, and PottsMPNN shows that models topping that metric trail on refolding and on mutation-effect prediction, so an entry in that race can move away from usable design while appearing to improve. Its limits are worth stating: architecture and scale are inherited rather than enlarged, evaluation is entirely computational with no wet-lab validation of the designed sequences, and the authors note they did not explore the homolog reweighting that might extract more from the alignments.