acceptodds
Under review as a conference paper at ICLR 2027

POTTSENERGY: LEARNING INTERACTIONS FOR COMBINATORIAL PROTEIN STABILITY PREDICTION

Abstract

Predicting the stability effects of combined mutations requires interaction models that generalize beyond individual substitutions. To meet this need, we introduce PottsEnergy, a compact framework that jointly designs structural encoding, explicit Potts parameterization, and sitewise self-supervision. Specifically, a backbone graph network predicts fields and pair couplings, while native-context pseudolikelihood and masked-context residue reconstruction train the same potentials used for mutation scoring. As a result, this design turns residue-level learning into a reusable sequence energy that combines single-site effects with explicit pair corrections. To evaluate this representation, we curate a higher-order stability benchmark from ten data sources. Even without stability-label training, PottsEnergy achieves state-of-the-art aggregate zero-shot ranking on this benchmark among the evaluated model families. Matched additive controls further show that the predictive contributions of the pair corrections are dataset-dependent. Separate diagnostics then distinguish these gains from interaction-sign accuracy. In addition, single-mutant adaptation tests the compact representation further. Together, these results connect residue reconstruction, combinatorial stability prediction, and candidate design through an explicit interaction-energy representation. More broadly, coupling learned interaction energies with quantum-assisted sampling could broaden the exploration of combinatorial sequence space for protein design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Combinatorial protein stability prediction, Protein design, Potts models, Structure-conditioned learning, Self-supervised learningCombinatorial protein stability prediction, Self-supervised learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_TXTnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes PottsEnergy, a compact network that converts a protein backbone into explicit Potts fields and pairwise couplings for stability prediction. Residue reconstruction trains these potentials, and mutation effects are subsequently calculated from the same energy. A combined mutation score decomposes exactly into single-mutant effects plus pair corrections, allowing inexpensive scoring after the energy tables are generated. The paper assembles a benchmark of 306k variants with at least three substitutions and demonstrates candidate retrieval using position proposals from a coherent Ising machine (CIM).

Critical strengths
  • The proposed a reusable representation based on a learned Potts energy. The same energy support reconstruction, stability scoring, and optimization.
  • Although the proposed model is a small one with only ~1M parameters, it achieves comparable performance with larger domain-specific models like SPURS.
  • The paper provides comprehensive baseline studies and ablations to further diagnose the performance of the proposed model. Code is also provided for good reproducibility.
Critical weaknesses
  • Most critically, the core idea of using the Potts model for modeling energy and related stability tasks has been explored in PottsMPNN, which is cited in this paper. The Potts pseudolikelihood training objective follows the spirit of PottsMPNN, although it adds auxiliary terms/regularizers. PottsMPNN also explores the energy optimization idea. I do not see a significant methodological contribution in this work.
  • The motivation of fitting an energy as stability predictor is less justifiable than standand inverse folding tasks, as there is no guarantee that the learned energy actually aligns with physical ddG. Currently, the model also only offer binary labels. The regression tasks in the appendix requires additional affine calibration and no raw RMSE is reported.
  • For the same reason, the paper does not provide any comparison with ddG regressors, which offer substantial more information at a finer granularity. Existing works on the same task include [1] (ICLR 2023) and more recent [2], both demonstrating the feasibility of learning the ddG with strong performance.
  • Dataset construction may leak stability/mutation information. Although the paper mentions that the test proteins are held out, no structure and/or sequence clustering were run for the dataset construction. This may leak homologs to the proposed model during pretraining or finetuning, making the final results less credible.
  • Some baselines including PottsMPNN, ProteinMPNN, and SPURS are directly evaluated using the official checkpoints, whereas the proposed model was trained on a larger vocab, leading to potentially unfair comparison.

[1] Luo, S., Su, Y., Wu, Z., Su, C., Peng, J., & Ma, J. (2023). Rotamer density estimator is an unsupervised learner of the effect of mutations on protein-protein interaction. ICLR, 2023.

[2] Cho, Y., Tsuboyama, K., Litberg, T. J., Jung, M. D., Obisesan, A., Wang, Q., ... & Rocklin, G. J. (2026). Accurate protein stability prediction for small domains using mega-scale experiments. bioRxiv, 2026-05.

Questions for authors

Please refer to the weaknesses.

Reviewer_giYVOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes PottsEnergy, a parametric Potts model where the Potts model's potentials are parameterized by GNN embeddings. The model is trained with pseudolikelihood and MLM; the learnable parameters are both the GNN's parameters and the those that map the GNN embeddings to the Potts model potentials. Once the model is fit, it can be used for scoring mutations and for design. Importantly, since the model is a (conditional) PottsModel, the design problem has a quadratic structure that can be exploited. Specifically, the paper proposes to factor the design problem into two steps: (i) determining whether a binary mask is promising (by using a theoretical lower bound), and (ii) given a promising binary mask, evaluating its expansions. This leads to faster exploration of the design space and ultimately superior designs as shown in benchmarks.

Critical strengths
  • The approach is compelling: fit a complex model whose "last layer" is a Potts Model, thereby facilitating the combinatorial design problem.
  • The method performs well on benchmarks.
Critical weaknesses
  • The paper is very hard to follow, plagued with dense, cryptic phrases and unconventional language. It was extremely frustrating to read. The paper cannot be published in its current form.
Questions for authors

Please rewrite the work using language and phrasing that is human-readable. The paper cannot be published in its current form. Here are some concrete examples of abstruse phrasing:

  • "Matched additive controls" => What is a "matched" additive control? Do you mean "additive baselines"?
  • "Separate diagnostics then distinguish these gains from interaction-sign accuracy" => what does this mean?
  • "More broadly, coupling learned interaction energies with quantum-assisted sampling [...]" => what is "quantum-assisted sampling"?
  • "Extending it to combination prediction requires a consistent score across mutation contexts" => what that does this mean? Also what is a "consistent score"? "consistent" has a precise statistical meaning (as in a consistent estimator).
  • "Explicit Potts fields and couplings provide a common sequence energy. " => What is an "Explicit" Potts field? Remove the word if it does not actually say anything.
  • "Sitewise reconstruction trains structure-conditioned fields and couplings" => It would be much clearer to say "Sitewise reconstruction trains structure-conditioned fields f_i and couplings J_{ij}", referencing the mathematical quantities in the text. This is a general comment: many parts of the text use natual language to refer to mathematical quantities in indirect (and sometimes cryptic) ways, making reading the text hard.
  • "The position QUBO supplies a sampling interface" => what is a "sampling interface"?
  • "Candidate search through reusable interaction energies" => What is a "reusable interaction energies"? Isn't it clearer to say "Candidate search by precomputing Potts potentials"?
  • "Single-mutant adaptation transfers representations to stability labels " => What does this mean?
  • "Binary quadratic optimization (Kochenberger et al., 2014) supplies the search interface" => Again, what is a "search interface"? Isn't "We use binary quadratic optimization (Kochenberger et al., 2014) to solve the design problem" clearer?
  • "Reciprocal-kNN graph" => What does "reciprocal" mean? Do you mean "undirected-kNN graph". kNN graphs are also undirected by default, so why not just say "kNN graph"?
  • "Scalar sequence models use single and four-state differences" => what are "four-state differences".
  • "Four width-110 node–edge co-updates propagate structural context while retaining explicit edge states." => what does this mean?
  • "Site-field and factorized coupling heads produce reduced coordinates" => what are "reduced coordinates"?
  • "reciprocal symmetrization" => what is "reciprocal symmetrization"?
  • "The four-state cycle cancels fields and unchanged-neighbor interactions." => What does this mean? This is cryptic.
  • "The matched additive control removes only target–target corrections, which are zero outside G." => what does this mean?
  • "Reverse-edge transpose averaging enforces reciprocity." => Another cryptic phrease. What does this mean?
  • " The 21 states include 20 amino acids and X" => X is overloaded here, it referred to structure earlier. Say that "X" represents an unknown amino acid.
  • "O(|G|q^2)" => define |G| for clarity (e.g. #nodes + #edges).
  • "an exposure/packing prior enters the fields" => what does this mean?
  • "visible identities enter the conditional readout" => what does this mean?
  • Section 3.4: This is a key part of the work. (i) Please make this section clearer. Provide defining equations for u_i, v_{ij}. (ii) Say that P_k is called the "position bound". (iii) " On exactly-k masks, the unpenalized position objective equals the minimum energy over amino-acid assignments for k = 2 and lower-bounds that minimum for higher orders. " => define "exactly-k masks" (I get what you mean, but please define these custom phrases); ptove the lower bound or cite a proof. (iv) "original-energy lower bounds" => define this too. (v) "This architecture makes binary sampling reusable across mutation orders while retaining the learned energy for sequence ranking." => Another dense phrase -- what are you trying to say? (vi) "Expansion enforces a consistent amino-acid assignment across pairs; final ranking uses the original float64 energy" => how does "float64" belong here?
  • "Thermodynamic endpoints and abundance proxies retain separate metric blocks" => What's a "Thermodynamic endpoint", what's an "abundance proxy", what's a "metric block"?
  • "Complete-score sensitivity additionally evaluates full-sequence differences where defined" => What does this mean?
  • "VenusMutHub remains exploratory after 28 local configurations." => what does this mean?
  • Table 3: Consider adding SEs (e.g. via bootstrapping).
  • "The interventions share inputs and recorded batches" => what does this mean?
  • "their removal raises the higher-order point estimates" => what does this mean? Do you mean that "ablating this component deteriorates performance"?
  • "The tested four-sweep Gibbs negative-phase run reaches −0.0073 under the common score" => what does this mean?
  • Table 5: Definitely add SEs. Your method beats SPURS on SPCC at the fourth signifiacnt digit only. This is likely a statistical tie.
  • "Its averaged float64 tables" => what is "float64" doing here?
  • "with 2,000 successful logical slots per protein" => what is a "logical slot"? You mean simply positions?
  • "The ICM reference anchors site and pair energy changes" => what does this mean? Is this a contribed way of saying that energy changes are measured wrt the reference sequence (i.e. a trivial statement)?
  • "Physical CIM receives the Ising matrix at 14-bit coefficient precision" => How does "14-bit coefficient precision" belong here?
  • "Candidates are ranked by the original float64 Potts energy." => float64 again. How does this computer architecture detail belong in the Figure 3 caption.
  • "target-pair cycles alter scores when both sites change" => what is a "target-pair cycle".
  • "The frozen energy provides a common target for evaluating future quantum-assisted sampling interfaces." => what is a "quantum-assisted sampling interface".
Reviewer_GsjTOpenReview ↗
Rating
1(−2)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PottsEnergy, a compact framework for predicting the stability effects of combinatorial protein mutations. The model uses a backbone graph network to predict explicit Potts fields and pairwise couplings. Native-context pseudolikelihood and masked-context residue reconstruction jointly train the same energy function later used for mutation scoring. This allows the model to capture interactions between mutated positions without requiring stability-label training in its primary zero-shot setting. The authors also construct a higher-order stability benchmark from ten data sources and evaluate the model against 13 model families. PottsEnergy achieves the highest aggregate normalized ranking in the reported multi-site comparison. Finally, the paper introduces a candidate search pipeline combining position-level QUBO proposals, a coherent Ising machine, full-alphabet expansion, and energy-based ranking.

Critical strengths
  1. A unified representation for combinatorial mutation scoring The main strength is the use of a shared Potts energy for both residue reconstruction and mutation scoring. Explicit pairwise couplings provide a consistent way to account for interactions between mutated positions, rather than relying entirely on the addition of independent single-mutation effects. The model is also relatively compact, with approximately 1.05 million parameters, making the representation interesting from a computational perspective.
  2. A substantial benchmark and a clear search pipeline The benchmark contains 306,014 higher-order variants from ten data sources, with 679 assay-condition groups. The authors also distinguish thermodynamic measurements from abundance proxies, which is important when evaluating protein stability. The candidate search experiment is technically interesting. On ten proteins, the proposed pipeline recovers all of the Top-1 and Top-5 sequences under the frozen energy ranking while substantially reducing the number of complete sequences that need to be evaluated.
Critical weaknesses
  1. The aggregate ranking does not fully establish statistical superiority The main comparison uses a normalized rank score across 13 model families and multiple metric blocks. Although PottsEnergy obtains the highest aggregate score, this ranking does not directly quantify the magnitude or statistical significance of its advantage. The component ablations also use a single training seed. Moreover, the authors acknowledge that the contribution of pairwise corrections depends on the dataset and endpoint, and that improved total-effect ranking does not necessarily translate into better interaction-sign prediction. Another limitation is that the model uses pairwise Potts couplings. It can score variants containing many substitutions, but it does not explicitly represent genuine three-body or higher-order interaction terms.
  2. The candidate search experiment has limited external validation The search experiment evaluates whether the proposed pipeline can recover the best candidates according to its own frozen Potts energy. It does not establish that these candidates are experimentally more stable. Furthermore, the search uses a separately supervised ensemble rather than exactly the same zero-shot configuration used in the main benchmark. The experiments involve only ten proteins, and the reported 475-fold and 2,747-fold reductions refer to the number of complete sequence evaluations, not equivalent wall-clock speedups over classical optimization methods. These distinctions should be made clearer when discussing the practical advantages of the search framework.
Questions for authors
  1. Can the authors provide multi-seed results and statistical uncertainty for the aggregate benchmark ranking and component ablations? Reporting per-source absolute metrics would also help clarify whether the improvement is consistent across datasets.
  2. Can the authors compare the CIM-based search with competitive classical QUBO solvers and report actual end-to-end runtime? It would also be valuable to validate the selected candidates against experimental stability measurements rather than only the model's own energy ranking.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.