acceptodds
Under review as a conference paper at ICLR 2027

Generalizing Protein Fitness Prediction Across Distribution Shifts via Structural and Co-evolutionary Priors

Abstract

Protein fitness prediction is essential for protein engineering. Recent fitness-landscape benchmarks such as FLIP2 have shown that methods performing well under conventional train–test splits can fail when the test set is shifted out of distribution (OOD), including mutations at unseen positions and variants from unseen protein backbones. In particular, current protein language models (pLMs) often suffer substantial performance degradation under such distribution shifts, limiting their practical use in protein engineering. We therefore develop a framework that augments sequence representations with structural and co-evolutionary information for fitness prediction. Specifically, we pair structure-aware pLMs with features pooled from an AlphaFold-style trunk applied to each sequence and its multiple sequence alignment (MSA), and use a lightweight prediction probe to map these frozen representations to fitness scores. Our framework consistently outperforms the best previously reported FLIP2 scores across several types of distribution shift. We then systematically study the contribution of each component in our framework to generalization. Structural priors yield the largest performance gains on held-out protein backbones, whereas trunk features improve generalization to held-out mutation positions. Rather than proposing a single best method, we characterize which modalities are effective for different protein families and distribution shifts, providing practical guidance for choosing representations according to the generalization regime. These findings support the development of more generalizable fitness predictors that can serve as effective reward functions or oracles for protein design toward desired functions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
machine learning, protein engineering, protein language models, protein structure

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_PdTUOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies out-of-distribution protein fitness prediction on FLIP2 and two additional settings. It augments frozen protein language model representations with two sources of prior information: structure-aware protein representations and pooled features from an AlphaFold-style trunk that processes each sequence and its multiple-sequence alignment. A lightweight probe maps the frozen features to fitness scores.

The experiments cover several protein encoders and multiple distribution shifts, including unseen mutation positions, mutation counts, mutation types, fitness ranges, and protein backbones. The strongest gains occur on the AMYLASE position splits. The trunk features also improve some held-out-backbone splits and two evaluations outside FLIP2. The paper further separates structural and co-evolutionary contributions with feature-block and MSA-Pairformer controls. Its main conclusion is conditional: structural priors help most across protein backbones, while trunk features help most across mutation positions. No representation is best in every regime.

Critical strengths
  1. The evaluation targets practical distribution shifts and covers several protein families. This matters because standard random splits can overstate fitness-prediction utility. The paper evaluates the settings that are important in iterative protein engineering, including unseen positions and unseen backbones.

  2. The component analysis produces a useful scientific conclusion. The paper does not only report a best model. It shows that structure-aware encoders and trunk features help in different regimes. The MSA-Pairformer comparison further suggests that the position-shift gains are not explained only by co-evolutionary information.

Critical weaknesses
  1. The main comparison is not fully matched. The pLM-only and pLM-plus-trunk rows use different feature-normalization settings. The appendix suggests that normalization does not explain the main gains, but a matched comparison would make the contribution of the trunk features easier to interpret.

  2. Uncertainty is not reported consistently for the main FLIP2 results. The paper provides uncertainty for some analyses, but the main comparison mostly reports point estimates selected from several encoders and feature combinations. This does not undermine the large position-split gains, but it makes the smaller improvements less clear.

Questions for authors
  1. Could the authors provide a matched pLM-only versus pLM-plus-trunk comparison using the same normalization and probe setup?

  2. Could the authors report uncertainty for the main FLIP2 results, especially for the smaller gains?

  3. Could the authors comment on whether FLIP2 test proteins or close homologs may have appeared in the pretraining data of the structure-aware models or trunk? A complete audit may not be possible, but the likely overlap would help interpret the held-out-backbone results.

  4. If feasible, could one strong published FLIP2 baseline be evaluated in the same probe and preprocessing pipeline? This would provide a useful matched reference for the reported-number comparison.

2 comments

We would like to thank the reviewer for considering our paper provides useful scientific insights to the field. We also appreciate thoughtful feedback, and hope to address the posted questions below. We have revised the manuscript accordingly and the changes are marked in blue.

Matched comparison of feature standardization

Could the authors provide a matched pLM-only versus pLM-plus-trunk comparison using the same normalization and probe setup?

We thank the reviewer for the suggestion for comparing different methods under the same normalization and probe schemes. We have updated Table 1 to use the same normalized MLP probes for both pLM-only and pLM with Protenix’s trunk features. Our main findings remain unchanged, particularly the substantial trunk gains on unseen mutation positions. Originally, we used unnormalized probes for pLM-only models because normalization did not consistently improve performance. In contrast, standardization was more beneficial for pLM + trunk models due to differences in feature scales from different pre-trained models. We provide further evidence in Appendix C.3 (Tables C3–C4), showing that standardization has mixed effects on pLM-only models but more consistently improves pLM-plus-trunk models. These analyses explain our original choices, while the revised Table 1 isolates the contribution of trunk features.

Uncertainty Quantification in Main Table

Could the authors report uncertainty for the main FLIP2 results, especially for the smaller gains?

We have updated Table 1 to report mean performance and standard deviations across 5 seeds, with 95% confidence intervals estimated from 2,000 bootstrap resamples. Our key findings remain unchanged. On AMYLASE unseen-position splits, trunk features yield significant gains across all eight encoders, ranging from +0.071 ± 0.019 to +0.263 ± 0.033 on close-to-far and +0.075 ± 0.021 to +0.221 ± 0.035 on far-to-close. All gains have 95% bootstrap confidence intervals surpassing zero. On RHOMAX, trunk features also moderately improve mean performance across all of the base pLMs, but with greater variability (e.g., ESM-C: +0.185 ± 0.175).

Homologs in pre-training data of structure-aware models

Could the authors comment on whether FLIP2 test proteins or close homologs may have appeared in the pretraining data of the structure-aware models or trunk? A complete audit may not be possible, but the likely overlap would help interpret the held-out-backbone results.

We appreciate the reviewer’s question regarding the potential leakage of structural information from FLIP2’s held-out set into structure-aware pLMs and even Protenix. To investigate this, we further evaluated our method on the held-out wild-type rhodopsin proteins, stratifying the test proteins according to their similarity to sequences in each model’s pre-training set. Because we have access to the pre-training datasets of Protenix-v2, ESM3, DPLM2, and SaProt, we considered a test protein to be represented in a test protein to have close homologues in a model’s pre-training set if at least one training chain shared more than 95% sequence identity over 70% of its length. By this criteria, 22 of the 36 test wild types (63% of test set) overlap with at least one model’s pre-training set. These 14 “unseen” wild types with no match, we observed significantly lower correlation (See table below). This suggests that the trunk’s contribution may be limited when test sequences are entirely novel or far distant from any sequences represented in the models’ pre-training data. We’ve added this analysis in ablation study section in the revised manuscript.

ModelCorrelation of base features on 14 unseen sequencesCorrelation of base + trunk features on 14 unseen sequencesCorrelation of base features on 22 seen sequencesCorrelation of base + trunk features on 22 seen sequences
SaProt−0.02 [−0.21, 0.25]−0.09 [−0.30, 0.28]0.12 [−0.15, 0.33]0.30 [−0.04, 0.50]
DPLM-20.13 [−0.10, 0.36]−0.00 [−0.16, 0.17]0.31 [0.00, 0.51]0.37 [0.01, 0.58]
ESM-30.02 [−0.15, 0.32]0.10 [−0.17, 0.45]0.20 [−0.11, 0.52]0.47 [0.09, 0.65]

Matched reference of FLIP2 baselines

If feasible, could one strong published FLIP2 baseline be evaluated in the same probe and preprocessing pipeline? This would provide a useful matched reference for the reported-number comparison.

We thank the reviewer for the suggestion to report FLIP2 baselines using the same probes. We note that the original FLIP2 evaluation considers a broader range of baseline methods, including not just MLP probes but ridge regression on one-hot features and zero-shot sequence-likelihood scoring that is not applicable for trunk features. Nevertheless, to ensure a fair comparison, we therefore report the performance of the best model re-evaluated using the same probes additionally in Table C8 in Appendix.

We hope these responses and additional experiments clarify our paper’s key contributions and address your concerns. We would appreciate your considering raising your score if these points are addressed sufficiently, or following up with any remaining questions.

Reviewer_19bQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on protein fitness prediction under distribution shifts, where protein language models (pLMs) often suffer substantial performance degradation. The authors develop a framework that augments sequence representations with structural and co-evolutionary information for fitness prediction. Specifically, this framework pairs structure-aware pLMs with features pooled from an AlphaFold-style trunk applied to each sequence and its multiple sequence alignment, and use a prediction probe to map these frozen representations to fitness scores. The authors evaluate this framework on FLIP2, FireProt, and new-to-nature functions. The experiment results show that the framework outperforms previously reported FLIP2 scores across several types of distribution shift. Further experiments study the contribution of each component in the framework, the results shows that structural priors yield the largest performance gains on held-out protein backbones, whereas trunk features improve generalization to held-out mutation positions.

Critical strengths
  1. This paper propose a framework that augments sequence representations with structural and co-evolutionary information for fitness prediction. Specifically, this framework pairs structure-aware pLMs with features pooled from an AlphaFold-style trunk applied to each sequence and its multiple sequence alignment, and use a prediction probe to map these frozen representations to fitness scores.

  2. The authors evaluate this framework on FLIP2, FireProt, and new-to-nature functions. The experiment results show that the framework outperforms the best previously reported FLIP2 scores across several types of distribution shift. The ablation studies shows that structural information improve generalization to held-out protein backbones, while co-evolutionary and geometric information improve generalization to held-out mutation positions.

Critical weaknesses
  1. This paper's innovation is not strong as it essentially combines features from existing models using pooling and MLP rather than designing a new method for representation learning, nor propose a new representation object.

  2. The paper states that "for pLM-only models, we use unnormalized MLP probes, whereas pLM models with trunk features use feature-standardized MLP probes", but the authors do not fully explain why these two kinds models uses different scale method. Table 1 compares the results between unnormalized MLP probes trained on only pLM and normalized MLP probes with pLM and trunk features. That means the difference is not just trunk features, but also includes the normalization method. Looking at Tables 1 and Table C6 together, it appears that, in some cases, performance improvements stem primarily from the addition of a normalization method. The gains derived from adding "trunk features" are not as substantial as those suggested in Table 1.

Questions for authors
  1. Could the authors clarify the key method contributions beyond combining existing model features through pooling and MLPs? How does the proposed approach advance representation learning compared with existing methods?

  2. Could the authors explain why different normalization strategies are used for pLM-only and pLM-plus-trunk models? Can they provide a controlled comparison using the same normalization strategy to isolate the actual contribution of trunk features?

3 comments

We thank the reviewer for constructive feedback and thoughtful questions that we will address below. We have revised the manuscript accordingly and the changes are marked in blue.

Methodological Contributions

Could the authors clarify the key method contributions beyond combining existing model features through pooling and MLPs? How does the proposed approach advance representation learning compared with existing methods?

We agree that our approach uses existing pretrained representations with a lightweight pooling and MLP probe. This design allows us to investigate the contribution of structural and co-evolutionary information without confounding the results with encoder fine-tuning or additional architectural complexity. Our novel contribution is a systematic empirical investigation of representation generalization under FLIP2’s challenging OOD shifts that reflected in actual protein engineering practice. Specifically, we examine three questions:

  • OOD generalization (Tables 1–2): How do frozen structural representations perform across mutation-number, position, fitness, and wild-type shifts? Do their improvements reflect genuine OOD generalization or merely better in-distribution fitting?
  • Co-evolutionary contribution (Table 4): How much of the trunk gains can be recovered using MSA-only representations, and what additional benefits do geometry-aware representations provide?
  • Mutation-indexed feature ablation (Table 5): Does mutation-position-specific pooling improve unseen-position generalization beyond global pooling?

We have revised the Introduction to emphasize these empirical contributions rather than the prediction framework (line 092).

Feature Standardization

Could the authors explain why different normalization strategies are used for pLM-only and pLM-plus-trunk models? Can they provide a controlled comparison using the same normalization strategy to isolate the actual contribution of trunk features?

We agree that the original Table 1 confounded the effects of trunk features and standardization. We have therefore updated Table 1 to use feature-standardized MLP probes for both pLM-only and pLM with Protenix’s trunk features, ensuring a fair comparison. Our main findings remain unchanged, particularly the substantial trunk gains on unseen mutation positions. Originally, we used unstandardized probes for pLM-only models because standardization did not consistently improve performance. In contrast, standardization was more beneficial for pLM + trunk models, likely due to differences in feature scales. We provide further evidence in Appendix C.3 (Tables C3–C4), showing that standardization has mixed effects on pLM-only models but more consistently improves pLM + trunk models. These analyses explain our original choices, while the revised Table 1 isolates the contribution of trunk features.

We hope these responses and additional experiments clarify our paper’s key contributions and address your concerns. We would appreciate your considering raising your score if these points are addressed sufficiently, or following up with any remaining questions.

I appreciate authors' rebuttal in addressing my concerns.

Regarding the first weakness, the authors mentioned that the novel contribution is a systematic empirical investigation. While these experimental analyses are indeed valuable, I think they are not significantly enough for ICLR given that this paper still lacks a method innovation and a core technical idea.

Regarding to the second weakness, the revisions have addressed my concern.

Given my remaining concern on the lack of method contribution, I would like to keep the original score.

Dear Reviewer 19bQ,

We thank the reviewer for the reply. Our main contribution is an empirical study about how protein representations generalise. The ICLR 2027 Reviewer Guide recognises this type of contribution in its significance criterion.

Submissions bring value to the ICLR community when they convincingly demonstrate new, relevant, impactful knowledge (incl., empirical, theoretical, for practitioners, etc).

Reviewer_jyBtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

Protein fitness prediction on novel variants not seen during training time, such as mutations at unobserved positions, mutations of higher multiplicity, or whole new wild-type proteins. State of the art pLM-based approaches cannot handle these FLIP2 benchmarks settings well. The paper uses pre-trained encoders, which are left unchanged, and concatenate the pLM embeddings with Foldseek structural tokens and single and paired protein representations from the trunk of Protenix-v2. The concatenated features are condensed into nine fixed-size blocks and passed to a two-layer MLP, which is the only thing to be learned. Contributions are: (1) a novel predictor which apparently improves upon the FLIP2 benchmarks on most splits; (2) an investigation of structural and coevolutionary signal contribution to generalization.Experiments are run on 11 FLIP2 dataset/split combinations, 8 encoders, FireProt, and a new-to-nature protein function set.

Critical strengths
  1. Fitness predictors are used as oracles in iterative design methods, where later rounds leave the training distribution. There is a genuine problem that is identified by the authors regarding in distribution and out of distribution performance of existing models, and why out of distribution performance matters more than benchmark score.
  2. The analysis separating which parts of the gain come from MSA signal versus geometric priors is well-designed.
  3. The scope of the experiments is large with eight encoders spanning sequence-only, joint, and structure-input families, each paired with and without trunk features across 11 splits.
Critical weaknesses
  1. The central claim of the paper isn't satisfied in that the abstract and discussion do claim to provide “practical guidance for choosing representations according to the generalisation regime". In Table 1, the best configuration varies across ESM-3 + trunk, ESM2 + trunk, ESM-C + trunk, SaProt and CHEAP with no configuration that can be pre-selected by the user for a new protein and a known shift type.
  2. The main result which is shown in Table 1 provides no uncertainty quantification. The values are point estimates with no error bars, standard deviations, or confidence intervals.
  3. The actual novelty is very limited and is under-argued in the paper given that AlphaMissense has already used the AlphaFold-2 Evoformer single/pair representations for variant effect prediction. The argument that they are targeting multi-mutant fitness rather than single substitutions is reasonable, but the paper asserts the gap without arguing why bridging it is non-trivial. Beyond the 4 pair representation reductions, the components are taken off the self and the probe is a standard 2 layer MLP on the concatenated features.
  4. The new-to-nature experiment has insufficient samples to allow for any directional conclusions. Figure 3 shows the Spearman correlations on 12, 12 and 5 variants respectively. The correlation on 5 points has a confidence interval that spans almost the entire range and the one on 12 points could easily be from noise. The text goes on to draw directional conclusions from both of these which the sample sizes do not allow.
Questions for authors
  1. Can you report seed variance for Table 1 to see if the improvement survive once seed-based variance is considered, particularly on the splits where the margin is small?
  2. How should a practitioner use your results to select a configuration for a new protein and a known shift?
  3. Could you report confidence intervals for Figure 3, and reframe these results as indicating a direction rather than supporting strong claims of directionality?
  4. Can you argue why bridging the gap to multi-mutant fitness is non-trivial relative to AlphaMissense’s use of Evoformer single/pair representations?
2 comments

We thank the reviewer for acknowledging our comprehensive experiments and the value of generalization test for fitness prediction. We hope to address your concerns and questions below. We have revised the manuscript accordingly and the changes are marked in blue.

Overstating Practical Guidance

We thank the reviewer for making a suggestion. To address this concern, we removed this phrase from the abstract: "practical guidance for choosing representations" and revised the discussion. Our intended contribution is instead a systematic empirical investigation of how different protein representations, including structural and co-evolutionary features, affect out-of-distribution fitness prediction across different generalization regimes. Rather than prescribing a particular representation for a new protein, our results demonstrate where combining these representations can be beneficial and where their contributions remain inconsistent.

How should a practitioner use your results to select a configuration for a new protein and a known shift?

In a real-world protein engineering campaign, these insights can guide the further development of a specialized model for specific fitness prediction tasks. For example, a recent study highlights this same finding with AlphaEnzyme (released in Nov 2026), which uses trunk features and trains the additional probes to predict catalytic specificity and kinetics [1]. Furthermore, when the mutants in the test set are expected to have mutation at previously unseen positions, incorporating trunk features would be particularly beneficial, as we have shown its contributions showed consistent gains in FLIP2’s mutation position split.

Uncertainty Quantification in Table 1

We have updated Table 1 to report mean performance and standard deviations across 5 seeds, with 95% confidence intervals estimated from 2,000 bootstrap resamples. Our key findings remain unchanged. On AMYLASE unseen-position splits, trunk features yield significant gains across all eight encoders, ranging from +0.071 ± 0.019 to +0.263 ± 0.033 on close-to-far and +0.075 ± 0.021 to +0.221 ± 0.035 on far-to-close. All gains have 95% bootstrap confidence intervals excluding zero. On RHOMAX, trunk features also improve mean performance across all of the base pLMs, but with greater variability (e.g., ESM-C: +0.185 ± 0.175).

Novelty and Key Contribution

Can you argue why bridging the gap to multi-mutant fitness is non-trivial relative to AlphaMissense’s use of Evoformer single/pair representations?

We agree that AlphaMissense has shown the usefulness of AlphaFold-style representations for single-amino-acid substitution effects. However, this does not establish their ability to generalize to multi-mutant fitness prediction under OOD shifts. Multiple mutations can interact through epistasis, while unseen mutation positions and protein backbones introduce additional generalization challenges which are relevant to real-world protein engineering. For example, FLIP2's RHOMAX dataset involves predicting peak absorption wavelengths across 75 different starting sequences with up to six mutated positions. AlphaMissense's single-substitution predictions do not directly address the combined effects of these mutations or generalization to unseen starting sequences. Beyond extending the evaluation to multi-mutant OOD settings, our study addresses three empirical questions not investigated in AlphaMissense:

  • OOD generalization (Tables 1–2): How do frozen structural representations generalize across mutation-number, position, fitness, and wild-type shifts? Do the improvements reflect genuine OOD generalization or merely better in-distribution fitting?
  • Co-evolutionary contribution (Table 4): How much of the OOD improvement from trunk features can be recovered using co-evolutionary information alone?
  • Mutation-indexed feature ablation (Table 5): Do mutation-position-specific representations improve unseen-position generalization beyond global pooling?

Our contribution is therefore a systematic empirical investigation of representation generalization, rather than a new prediction architecture. We have revised the Introduction to acknowledge AlphaMissense (line 064), explain why extending to multi-mutant OOD prediction is non-trivial, and emphasize these empirical contributions (line 092).

New-to-nature experiment

Regarding these comments and questions

The new-to-nature experiment has insufficient samples to allow for any directional conclusions. Figure 3 shows the Spearman correlations on 12, 12 and 5 variants respectively. The correlation on 5 points has a confidence interval that spans almost the entire range and the one on 12 points could easily be from noise. The text goes on to draw directional conclusions from both of these which the sample sizes do not allow.

Could you report confidence intervals for Figure 3, and reframe these results as indicating a direction rather than supporting strong claims of directionality?

It is indeed true that the new-to-nature experiment lacks statistical power. However, limited experimental data for proteins with novel functions is an inherent challenge, representing an extreme out-of-distribution (OOD) regime. We deliberately extend our generalization evaluation to this particularly challenging setting, which is also highly relevant to practical de novo enzyme design campaigns. Despite the limited sample size, the phosphotriesterase campaign shows a notable shift in correlation, from negative (-0.15)) to positive (+0.35). While this result should be interpreted cautiously given the limited data, it provides encouraging evidence that combining pLM representations with structural and trunk information can yield meaningful predictive signals even in extreme OOD setting. Nevertheless, we agree with the reviewer to also report the confidence intervals, so we updated them on Figure 3 and revised the discussion to frame these findings as indicative of directional trends rather than as strong evidence about directionality.

We hope these responses and additional experiments clarify our paper’s key contributions and address your concerns. We would appreciate your considering raising your score if these points are addressed sufficiently, or following up with any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.