PottsMPNN scores the many sequences that fit a fold rather than the one evolution kept
Protein design has graded its models on whether they can guess the sequence evolution happened to pick. An MIT group says that measures the wrong thing, and reports that leaning less on native sequences improves stability prediction.
The Scientist · Science desk
What happened
- Her group's framework, PottsMPNN, builds the physical principles governing structure and stability into both sequence generation and prediction of how mutations change stability.
- Training added noise to the input structures, which the team says lowers the model's tendency to copy native sequences and widens the set of structures it can write sequences for.
- The reported result is directional: as reliance on native sequences dropped, structural compatibility and energy prediction improved, including for proteins with no natural counterpart.
Why it matters
- What's now possible capability A team designing a binder for a target with no natural template gains a way to rank candidate sequences that a homology-based score cannot rank at all, because there is nothing to be homologous to.
- Who has to decide decision Groups that evaluate design models must now decide whether to keep recovery, which is cheap and comparable across papers, or fund the held-out mutational stability data that the alternative metrics require.
- What it sets up precedent If the metric critique takes hold, reviewers will start expecting fold-compatibility and mutation-effect numbers in every new design paper, which raises the evidentiary cost of publishing a model.
Sequence recovery has a denominator problem. Many different amino acid sequences fold into the same structure [3], so a model that proposes something unlike the native protein may have found another workable answer, and a recovery score marks it wrong either way. Amy Keating, who heads MIT's Department of Biology and is senior author on the PNAS paper, frames this as a measurement error rather than a modelling one: the field has asked whether a model can reproduce the sequence "that evolution happened to select", and she says "this isn't the best metric for protein design" [1] [2].
What lead author Foster Birnbaum offers instead are quantities you can score without a native reference: how likely generated sequences are to adopt the target fold, how well the model grasps the sequence-energy landscape, and how well it predicts the effect of a mutation on stability [5]. PottsMPNN is built for the last two. Rather than treating positions one at a time, it carries a pairwise distribution over the physical interactions between all 20 amino acid options at each of two positions, which the team gives as a key reason it models the sequence-energy landscape more accurately than other methods [6]. Multiply it out and that is 400 joint combinations for every pair the model tracks, against 20 numbers per position for a model that scores positions separately [14].
Two training choices push the model off the native answer. Noise, meaning variations added to the structure during training, reduces the tendency to mimic native sequences and widens the range of structures the model can write sequences for [7]. Sets of evolutionarily related sequences teach it that many sequences occupy one fold [8]. Birnbaum concedes the second is, in some ways, still a reliance on natural sequences [9]; the reported direction is that as that reliance falls, structural compatibility and energy prediction improve, including for novel proteins [16].
The thing these accounts do not tell you is how much. Neither the MIT release nor the phys.org version, which reproduces the same text [12], gives an effect size, a comparison figure, or a bench test showing that designed sequences fold and stay folded [13]. The phys.org headline keeps the hedge where it belongs, on "potentially stable" [10]. And the most widely used model in the field was released in 2022 and, by Birnbaum's account, has not been surpassed [15], which is a reminder that a metric argument is how a challenger earns a look rather than how it takes over someone else's pipeline. Birnbaum also says that being able to design any protein at will "enables us to do a potentially scary amount of biological engineering" [11], a sentence worth noting because it comes from the person building the tool.
What to watch
- The PNAS tables: whether PottsMPNN's mutation-effect and fold-compatibility results are reported against the 2022 default on held-out data.
- Wet-lab expression and thermal stability tests on designs that have no native homolog to compare against.
- Whether other design groups start reporting energy-landscape scores alongside, or instead of, sequence recovery.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence30
- Adoption
- Not scored
- Hype gap+32
- Incentives68
- Confidence58
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Amy E. Keating: "For years, the field has measured success by asking whether a model can reproduce the protein sequence that evolution happened to select" and "our work shows that this isn't the best metric for protein design."
- [2]
Amy E. Keating is head of the Department of Biology, Jay A. Stein (1968) Professor of Biology, professor of biological engineering, and senior author of a paper recently published in PNAS.
- [3]
In nature, many different amino acid sequences can fold into the same structure, and one sequence can potentially adopt different structures depending on the protein's flexibility or a functional trigger.
- [4]
Many methods for designing novel proteins are two-step: the structure comes first, then a machine-learning framework generates a repertoire of sequences that could potentially adopt that structure.
- [5]
Birnbaum, graduate student and lead author: for a completely novel designed structure there would be no native sequence to compare to, and what matters is how likely generated sequences are to fold into the desired structures, how well the model understands the sequence-energy landscape, and how well it predicts the effect of mutations on stability.
- [6]
PottsMPNN uses a pairwise distribution to capture interactions between amino acids; the ability to account for physical interactions between all 20 possible sequence options at a pair of positions is a key reason it models the sequence-energy landscape more accurately than other methods.
- [7]
"Noise," meaning variations added to a protein structure during training, decreases the model's tendency to overly mimic native sequences and increases the diversity of structures for which it can generate sequences.
- [8]
The team introduced sets of evolutionarily related sequences into PottsMPNN's training to teach the model how different sequences can adopt the same folded structure.
- [9]
Birnbaum acknowledges that incorporating evolutionary information is, in some ways, still a reliance on native sequences.
- [10]
The phys.org headline reads "AI model designs potentially stable protein sequences beyond those found in nature."
- [11]
Birnbaum: "Once we can design any protein we want, that enables us to do a potentially scary amount of biological engineering," adding that it is a difficult task but he is optimistic about this century's progress in biology.
- [12]
The phys.org item and the MIT News item carry the same text, so the two published accounts are one press release rather than two independent reports.
- [13]
Neither published account reports a numeric benchmark result, an effect size, or an experimental test that PottsMPNN-designed sequences fold and remain stable.
- [14]
Accounting for all 20 options at each of two positions means 400 joint amino acid combinations per position pair, against 20 values per position for a model that scores positions independently.
- [15]
The most widely used protein model today was released in 2022, and Birnbaum says that for a field moving as fast as machine learning in biology, that model has not been surpassed.
- [16]
PottsMPNN demonstrated that as the model depends less and less on native sequences, structural compatibility and energy prediction, including for novel proteins, improve.
- [17]
PottsMPNN, a new machine-learning framework developed in MIT's Department of Biology, incorporates the physical principles that govern protein structure and stability, improving sequence generation and the ability to predict how mutations will affect a protein's stability.
- [18]
Adding the framework to a protein design pipeline will allow researchers to design structurally feasible proteins with sequences that don't resemble those of any native protein.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- news.mit.eduLooking beyond natural sequences
1 article · August 27, 2026
- phys.orgAI model designs potentially stable protein sequences beyond those found in nature
1 article · August 27, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.