acceptodds
Under review as a conference paper at ICLR 2027

Precision Inverse Folding: From Low-Order Mutant Measurements to Higher-Order Variant Prioritization

Abstract

Protein inverse folding models optimize sequence–structure compatibility but provide limited guidance on protein function. In practical protein engineering, only a small number of low-order mutants can typically be experimentally characterized, while promising designs may lie in a much larger space of higher-order combinations. We study this two-to-many setting, using only wild-type, single, and double-mutant wet-lab measurements to prioritize unseen variants with three or more mutations. Our approach constructs cross-order preferences from measured low-order variants and synthetic higher-order negatives derived from inactivity-associated mutations, focuses alignment on function-relevant mutations through critical-site preference optimization, and enriches structural representations with protein foundation-model embeddings. Across four protein fitness datasets and two inverse-folding architectures, our approach consistently improves higher-order variant prioritization, demonstrating that limited low-order experimental data can effectively guide exploration beyond the experimentally accessible mutational regime.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Precision Inverse folding, Protein fitness, Protein Engineering, Preference Alignment

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_mkbpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether low-order experimental measurements can be used to prioritize unseen higher-order protein variants. The authors formulate a two-to-many inverse-folding setting in which training uses only wild-type, single-mutant, and double-mutant fitness measurements, while evaluation is performed on variants with three or more mutations. PIF constructs preference pairs from measured low-order variants and synthetic higher-order negatives generated from inactivity-associated mutations. It further augments ProteinMPNN and ESM-IF representations with ESM2 residue embeddings through a FiLM module and applies preference optimization only at mutation sites that differ between paired variants. The method is evaluated on four FLIP2 protein fitness datasets using both inverse-folding backbones. Results are reported with top-K retrieval, enrichment, ranking metrics, and standard inverse-folding metrics, together with ablations for synthetic negatives and foundation-model conditioning.

Critical strengths
  1. The evaluation is reasonably broad, covering four FLIP2 protein fitness datasets with different assay types and two inverse-folding backbones, ProteinMPNN and ESM-IF. This reduces the chance that the reported gains are tied to a single protein or model architecture.

  2. The low-order-to-high-order setting is practically relevant for protein engineering, since experimentally characterizing single and double mutants is much cheaper than exhaustively assaying higher-order combinations. The paper evaluates this setting directly rather than relying only on standard inverse-folding recovery metrics.

Critical weaknesses
  1. The methodological novelty is limited and the positioning against prior preference optimization work is incomplete. The framework is closely related to FPA, and the paper explicitly adopts its FiLM-based functional conditioning and critical-site preference alignment. The main additional component is therefore the construction of cross-order preference pairs rather than a substantially new optimization framework. Moreover, the related-work discussion omits several closely related preference-based protein design methods, including InstructPLM-DPO [1], EnerBridge-DPO [2], CAPE-MPNN [3], and ProtAlign [4]. These works already apply preference optimization to inverse folding or protein sequence design using structural, energetic, or functional signals, which makes the incremental contribution of PIF less clear.

  2. The synthetic higher-order negative construction is empirically useful, but its underlying assumption is not sufficiently validated. The method assumes that recombining mutations associated with low-activity low-order variants produces useful higher-order negatives, even though the authors acknowledge that their actual fitness is unknown and may be affected by epistasis. The ablations show that these negatives improve downstream retrieval, so the strategy is clearly useful as a training heuristic; however, the paper does not measure how often the synthesized variants are actually lower-fitness than their paired positives, despite the FLIP2 benchmarks containing measured higher-order variants that could support such a check. This makes it difficult to tell whether the gain comes from biologically meaningful cross-order supervision or simply from a useful form of hard-negative regularization.

[1] Xu, Junde, et al. "Protein inverse folding from structure feedback." Advances in Neural Information Processing Systems 38 (2026): 24274-24299.

[2] Rong, Dingyi, et al. "EnerBridge-DPO: Energy-Aware Markov Bridge Inverse Folding for Protein Sequence Design." Journal of Chemical Information and Modeling 66.15 (2026): 9709-9720.

[3] Gasser, Hans-Christof, et al. "Tuning ProteinMPNN to reduce protein visibility via MHC Class I through direct preference optimization." Protein Engineering, Design and Selection 38 (2025): gzaf003.

[4] Liu, Junqi, et al. "Property-driven protein inverse folding with multi-objective preference alignment." International Conference on Learning Representations. Vol. 2026. 2026.

Questions for authors
  1. Could the authors clarify the main methodological difference between PIF and FPA, given that the proposed framework adopts FPA-style FiLM conditioning and critical-site preference optimization? It would also help to position PIF more explicitly against other recent preference-based protein design methods.

  2. The synthetic higher-order negatives improve downstream retrieval in the ablation, but their true fitness is unknown. Since the FLIP2 datasets contain measured higher-order variants, can the authors report how often the synthesized negatives that can be matched to measured variants are actually lower-fitness than their paired positives? This would help determine whether the gain comes from meaningful cross-order supervision or mainly from hard-negative regularization.

  3. Are synthetic higher-order sequences that exactly coincide with held-out test variants removed from the training preference set? This is particularly important for dense combinatorial datasets such as TrpB.

1 comment

Response to Reviewer mkbp

Thank you for the detailed assessment.

W1 / Q1. Difference from FPA and other preference-based methods

One clarification regarding FPA: FPA does not use FiLM-based foundation-model conditioning. Our contribution is not a new preference-optimization objective, but a framework designed for two-to-many protein inverse folding.

  • Problem and supervision. To our knowledge, PIF is the first preference-aligned inverse-folding framework studying the strict low-order-to-higher-order shift: only WT, single- and double-mutant measurements are available for training, while unseen 3+ variants are prioritized at test time. FPA instead builds preference pairs entirely in silico by perturbing residues presumed functionally important, without measured activity.
  • Cross-order preference construction. PIF anchors the preferred member in measured low-order activity and builds higher-order negatives by recombining substitutions enriched among inactive variants. The supervision is thus experimentally grounded on one side and deliberately extrapolative on the other.
  • Foundation-model conditioning. PIF adds residue-level ESM2 embeddings through a FiLM module with a learnable gate, enriching the structure-conditioned representation. This is not part of FPA.

We have also clarified "critical sites". In FPA they are chosen in advance from hypotheses about functional importance; in PIF the set is defined per preference pair as the positions where the two variants differ. Since all other residues are shared, these carry the mutation-specific signal behind the measured activity difference. This is now explicit in the revision.

Missing related work. We thank the reviewer for pointing out InstructPLM-DPO, EnerBridge-DPO, CAPE-MPNN and ProtAlign, which we add to Section 2. Their training signals differ from ours: InstructPLM-DPO uses folding-model structural feedback, EnerBridge-DPO energy-based preferences, CAPE-MPNN MHC-I visibility, and ProtAlign multi-objective property alignment. In each case the signal is evaluable at training time for an arbitrary sequence, so preferences can be generated on demand anywhere in sequence space. In our setting no such signal exists: activity is known only for a small low-order set, and the regime we must rank is unmeasured by construction. We revise the related-work section around this distinction rather than listing methods.

W2 / Q2. How often are the synthesized negatives actually lower-fitness than their paired positives?

We do not assume every recombinant remains inactive under epistasis; the construction is intended to make it more likely to have lower activity than its paired positive, providing a useful ordinal preference.

We test this over the full measured higher-order set, grouping all 265,686 measured 3+ mutants across the four datasets by how many inactivity-associated pool substitutions they carry. "Pool substitutions carried" is not the mutation count: it is how many of a variant's substitutions lie in the 25-substitution inactivity pool. IRED is for reference only: its negatives are measured inactive variants from the source publication, not recombinants.

DatasetPool substitutions carriednMean activity% > WT
Amylase02,0340.0413.59
Amylase11110.0100.00
Amylase2210.0100.00
IRED03,565−1.36811.05
IRED1584−1.5139.59
IRED229−1.5790.00
TrpB0200,9900.0400.51
TrpB111,028−0.0560.00
TrpB25,138−0.0780.00
TrpB3351−0.0930.00
NucB039,2410.65323.35
NucB12,5150.1202.31
NucB2740.0000.00

The relationship is monotonic on all four datasets. One pool substitution reduces the fraction exceeding WT from 23.35% to 2.31% on NucB, 3.59% to 0% on Amylase and 0.51% to 0% on TrpB; IRED shows a weaker but consistent trend. Inactivity-associated mutations identified independently from training and from held-out higher-order data also overlap substantially, indicating these patterns transfer to higher-order backgrounds rather than being artifacts of the training split.

Q3. Do synthetic higher-order sequences coincide with held-out test variants?

We measured this overlap directly.

Dataset# negatives# testOverlap% of negatives
Amylase1,8002,17000.00
IRED1,4524,17800.00
TrpB2,592217,50750.19
NucB9,04341,83100.00

Synthetic negatives are essentially disjoint from the evaluation pool: the overlap is zero for Amylase, IRED and NucB, and for TrpB only 5 of 2,592 (0.19%) coincide with measured test variants, or 0.0023% of the test pool. All five are genuinely low-fitness (mean activity −0.061 against a test-pool mean of +0.033) and therefore satisfy the intended negative preference.

Reviewer_cVwSOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a framework for protein inverse folding task which uses two-to-many setting to prioritize unseen higher-order mutants using only wild-type, single-, and double-mutant wet-lab measurements. The proposed method constructs cross-order preferences by combining measured low-order fitness differences with synthetic higher-order negative variants generated from inactivity-associated mutations. It then incorporates protein foundation model embeddings to enrich structural representations and uses critical-site preferences to focus alignment on mutated positions. Experiments on four FLIP2 datasets using ProteinMPNN and ESM-IF baselines demonstrate improvements in higher-order variant prioritization, although the performance gains are not consistent across all datasets and metrics.

Critical strengths

The paper addresses a practically challenging problem in protein engineering by leveraging low-order experimental wet-lab measurements to prioritize unseen higher-order mutants. The proposed framework combines experimental fitness supervision, synthetic negative preference construction, and foundation model representations within a unified pipeline.

Critical weaknesses
  1. the paper has a reasonable research motivation, but the novelty of preference construction and alignment mechanism remains questionable. The two-to-many setting, has already been explored in prior works such as FLIP2 and MULTI-evolve [1]. Simply renaming this established setting as “two-to-many protein inverse folding” does not constitute a novel research contribution.

  2. The proposed cross-order preference construction method appears to be straightforward heuristic rather than a substantial methodological innovation. The authors identify low-activity variants via experimental annotations/quantile thresholds, then recombine them to generate synthetic higher-order negatives. However, mutations associated with inactivity in low-order mutants are not necessarily deleterious in higher-order combinations because of epistatic interactions. More importantly, with such negatives, the model may simply learn to penalize mutation count or seen deleterious substitutions during training rather than meaningful functional preferences. In addition, converting experimental fitness measurements into pairwise preferences for inverse-folding alignment has already been explored in ProteinDPO [2].

[1] "Rapid directed evolution guided by protein language models and epistatic interactions", Science, 2026. [2] "Aligning protein-generative models to experimental fitness with ProteinDPO", Nature Methods, 2026.

Questions for authors

In Section 3.3, are the foundation model embeddings extracted from the wild-type or mutant sequences? If mutant-specific, how is target-sequence information leakage avoided?

Lines 260-262, why are positions where preferred and dispreferred sequences differ considered "critical sites"? How do the authors establish their functional importance rather than treating them as mutated positions?

In Table 1, the performance improvements of the proposed method are inconsistent. In several cases, the full model performs worse than its ablated variants. Could the authors explain these inconsistencies and provide further analysis?

Also can the authors do more experiments using other baselines? for example the MULTI-evolove or FLIP2 official baselines.

In Figure 1, including higher-order mutants (3M/4M) during training has substantially improves the performance, which goes beyond the claimed two-to-many setting. Could the authors clarify how these results support their central claim of generalization using only low-order measurements?

3 comments

Response to Reviewer cVwS (1/2)

Thank you for the detailed assessment. We address each point below.

W1. Novelty of the two-to-many setting

We do not claim that low-order-to-higher-order generalization is itself new. FLIP2 provides the number split and MULTI-evolve shows low-order measurements can guide higher-order engineering; we cite both as motivation, not as our contribution.

Our contribution is to study this mutation-order shift within protein inverse folding: only WT, single- and double-mutant measurements are available for alignment, while the structure-conditioned generative model is evaluated on unseen variants with three or more mutations. PIF combines cross-order preference construction with protein-foundation-model conditioning, so that limited low-order functional supervision can directly guide a structure-conditioned inverse-folding model toward higher-order prioritization.

W2. Novelty of the cross-order preference optimization

  • Cross-order preference construction. PIF pairs experimentally measured low-order positives with synthetic higher-order negatives built from inactivity-associated substitutions, extending functional supervision from the measured WT/single/double regime into the unseen 3+ regime without additional assays.
  • Difference from existing preference-based methods. ProteinDPO converts measured fitness into preference pairs, but its preferences remain within the measured mutational regime. FPA constructs preferences entirely in silico and does not use measured activity. PIF instead anchors the preferred member in measured low-order activity and deliberately constructs higher-order negatives, making the supervision explicitly cross-order. PIF further adds residue-level foundation-model embeddings through FiLM conditioning, which is not part of FPA.

Validity of the synthetic negatives. We do not assume every recombinant remains inactive under epistasis; the construction is intended to make it more likely to have lower activity than its paired positive. Grouping measured 3+ mutants by how many inactivity-associated substitutions they carry, measured activity decreases monotonically across all four datasets. "Pool substitutions carried" is not the variant's mutation count: it is how many of a variant's substitutions lie in the 25-substitution inactivity pool. IRED is for reference only: its training negatives are measured inactive variants from the source publication, not recombinants, so the pool below is reconstructed solely for this diagnostic.

DatasetPool substitutions carriednMean activity% > WT
Amylase02,0340.0413.59
Amylase11110.0100.00
Amylase2210.0100.00
IRED03,565−1.36811.05
IRED1584−1.5139.59
IRED229−1.5790.00
TrpB0200,9900.0400.51
TrpB111,028−0.0560.00
TrpB25,138−0.0780.00
TrpB3351−0.0930.00
NucB039,2410.65323.35
NucB12,5150.1202.31
NucB2740.0000.00

One pool substitution drops the fraction exceeding WT from 23.35% to 2.31% on NucB, 3.59% to 0% on Amylase and 0.51% to 0% on TrpB. Inactivity-associated mutations identified independently from training and from held-out higher-order data also overlap substantially.

Mutation-count-matched controls. Within a fixed mutation-count group every variant carries the same number of substitutions, so mutation count supplies no ranking signal. NucB is omitted because its measured labels take only four ordinal values: within each fixed-k group two random test variants share a label 41–46% of the time, so the ground truth is only partially ordered and the attainable Spearman is capped far below that of the continuous datasets. TrpB's test library contains only k = 3 and k = 4 variants.

DatasetModelk=3k=4k=5k=6
AmylaseProteinMPNN0.4230.3500.3830.279
+PIF0.6410.5280.5610.488
ESM-IF0.4330.3570.4380.221
+PIF0.6160.4960.5760.404
IREDProteinMPNN0.1130.126−0.0660.041
+PIF0.2640.183−0.0070.067
ESM-IF0.1540.115−0.0180.006
+PIF0.2390.165−0.0040.029
TrpBProteinMPNN0.2420.086——
+PIF0.6300.126——
ESM-IF0.4190.078——
+PIF0.6330.138——

PIF improves the pretrained model in every fixed-k setting, with the largest gains at k = 3. Because mutation count is constant within each comparison, these improvements cannot be explained by a shortcut that penalizes variants with more mutations. This is consistent with Figure 1, where PIF improves within fixed test mutation-order groups.

(continued in 2/2)

Response to Reviewer cVwS (2/2)

Q1. Are the foundation-model embeddings from wild-type or mutant sequences, and how is leakage avoided?

The ESM2 embeddings are extracted from the candidate mutant sequences. This introduces no additional sequence information, because the inverse-folding baselines (ProteinMPNN and ESM-IF) also receive the same candidate sequence S when computing the teacher-forced likelihood log p(S | backbone). The ESM2 embedding simply re-encodes an input the scorer already has. The encoder is frozen and the embeddings are precomputed offline; ESM2 is never fine-tuned and never sees any activity measurement.

No activity-label leakage occurs at training or inference: experimental measurements are used only to construct preference pairs among training variants, and no label from the held-out 3+ split enters the model.

Q2. Why are the differing positions called "critical sites"?

We use the term operationally at the preference-pair level: the critical-site set is the positions where the preferred and dispreferred variants differ, since these carry the mutation-specific signal associated with the measured activity difference. In FPA, critical sites instead refer to residues presumed functionally important for the protein and selected in advance from biological or model-based criteria.

In our setting the two members of a pair share the same wild-type background and differ at only a few positions, so the general notion reduces to the pair-differing set: all unchanged residues are shared, and the observed activity difference is attributable to the distinguishing mutations. Restricting the preference likelihood to these positions also avoids diluting the sparse signal with unchanged residues.

Q3. Inconsistency between full and ablated models in Table 1

We respectfully disagree. Our primary objective is higher-order top-K prioritization under a limited experimental budget, while Pearson and Spearman characterize global ranking over the full pool. Comparing PIF (Full) with either PIF w/o Neg or PIF w/o PFM, HC@10 and EF@10 improve or remain unchanged in every dataset and base-model setting on both base models. Small fluctuations in Pearson/Spearman or at larger K can therefore occur without conflicting with those gains.

The few exceptions occur only at K = 100. For the negative component, IRED/ESM-IF decreases slightly from HC@100 39 to 38. Notably, IRED is the only dataset where we do not introduce synthetic higher-order negatives, using measured inactive single and double mutants instead (Appendix A.7), so this component provides less new cross-order information there.

Q4. Additional baselines: FLIP2 and MULTI-evolve

Several of the suggested comparisons address a different task from ours.

  • Different task. PIF performs inverse folding with functional alignment: the model stays structure-conditioned and its likelihood is optimized for both structural compatibility and functional activity. Additive, pairwise, ESM2-based supervised predictors and standard FLIP2 baselines instead map sequence directly to fitness.
  • Different objective. We do not aim to regress the higher-order fitness landscape or train a separate activity predictor, but to incorporate limited functional supervision into the inverse-folding distribution.

These fitness predictors are therefore not directly comparable baselines for the problem studied here. Our comparisons instead include pretrained and preference-aligned inverse-folding methods: ProteinMPNN, ESM-IF, SPPO and FPA.

Different protocol from MULTI-evolve. It provides important evidence that low-order measurements can guide higher-order engineering, but relies on iterative model-guided evolution with experimental feedback each round. Our setting uses a fixed low-order measurement set and evaluates one-shot generalization to held-out higher-order variants, so MULTI-evolve is relevant motivation rather than a baseline executable under the same protocol.

Q5. Does Figure 1 undercut the two-to-many claim?

This concern rests on a misunderstanding of the experiment. The main PIF results in Tables 1–2 do not use 3- or 4-mutation measurements for training. They use only WT, single- and double-mutant measurements, exactly following the claimed two-to-many setting.

Figure 1 is a separate diagnostic that varies the maximum training mutation order from k_train = 1 to 4, to quantify how much additional benefit higher-order measurements would provide. It shows that much of the gain is already achieved at k_train = 2, while adding 3- and 4-mutation measurements yields smaller, less consistent improvements.

Figure 1 therefore supports rather than contradicts our central claim: accessible low-order measurements already capture much of the benefit available for higher-order prioritization. We have made this diagnostic purpose explicit in the caption and the main text.

I appreciate the authors' responses to my concerns and questions. Re the novelty, however, I remain of the view that the proposed approach represents a distinct implementation of existing techniques rather than a substantially novel alignment methodology. I agree with Reviewer j8Qq that the difference in model formulation does not make these baselines incomparable when they address the same downstream task. 

In addition, the authors' assumption is that the resulting synthetic negatives have lower activity than their paired positive sequences. The additional analysis reveals that inactivity-associated substitutions correlate with reduced activity in measured higher-order variants, However, it still remains unclear whether the observed trend can reliably be extrapolated to the specific recombinant negatives used for preference training, particularly under epistatic interactions.

Reviewer_2V2iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PIF, which aligns inverse-folding models using single- and double-mutant functional measurements to rank unseen higher-order variants. It combines synthetic higher-order negatives, preference optimization at differing residues, and ESM2-based representation conditioning. Experiments on four FLIP2 datasets and two architectures show improved retrieval over pretrained and preference-aligned baselines.

Critical strengths
  1. The low-order-to-higher-order setting addresses a practical protein-engineering problem, and evaluation emphasizes top-ranked candidates relevant to limited experimental budgets.

  2. Results across two architectures show useful retrieval improvements, with ablations providing some evidence for the proposed components.

Critical weaknesses
  1. The benefit of likelihood-based alignment over direct functional prediction remains unestablished. Training optimizes differing-site likelihoods, whereas inference uses full-sequence likelihood. Comparisons with ESM2-based supervised predictors and additive/pairwise baselines are missing. Given MULTI-evolve’s prior demonstration of low-order-to-higher-order engineering, these comparisons are important for establishing the additional contribution.

  2. The synthetic negatives rely on an unvalidated assumption: mutations associated with low activity remain detrimental when recombined. Compensatory epistasis could violate this assumption, and pairing low-order positives with higher-order negatives may introduce mutation-count shortcuts. The ablation gains are mixed, and the main results lack uncertainty estimates.

Questions for authors
  1. Under identical training splits, how does PIF compare with an ESM2-based functional predictor and additive/pairwise models?

  2. What evidence supports the synthetic preference labels? Do the gains persist with mutation-count-matched controls?

  3. Please clarify Appendix A.8.1: does “active mutations, drawn from all training sequences” include inactive sequences? If so, would filtering always trigger the fallback?

  4. Are the reported improvements consistent across training seeds?

2 comments

Response to Reviewer 2V2i (1/2)

Thank you for the detailed assessment. We address each point below.

W1 & Q1. Likelihood alignment versus direct functional prediction

The suggested methods address a different task from ours.

  • Different task. PIF performs inverse folding with functional alignment: the model stays structure-conditioned and its sequence likelihood is optimized for both structural compatibility and functional activity. Additive, pairwise and ESM2-based supervised predictors instead map sequence directly to fitness.
  • Different objective. We do not aim to regress the higher-order fitness landscape or train a separate activity predictor, but to incorporate limited functional supervision into the inverse-folding distribution.

These fitness predictors are therefore not directly comparable baselines for the problem studied here. Our comparisons are with pretrained and preference-aligned inverse-folding methods: ProteinMPNN, ESM-IF, SPPO and FPA.

Differing-site training versus full-sequence inference. Differing-site likelihood is used during preference optimization to avoid diluting the sparse functional signal with unchanged residues in low-order mutant pairs. At inference, candidates are scored independently, so the pair-dependent differing-site mask is unavailable; normalized full-sequence likelihood therefore provides a consistent inverse-folding score for higher-order prioritization.

Different protocol from MULTI-evolve. We agree it shows low-order measurements can guide higher-order engineering, and do not claim this premise as our contribution. However, MULTI-evolve is an iterative wet-lab-in-the-loop procedure requiring new measurements each round, whereas PIF uses a fixed low-order dataset and aligns once. Since the FLIP2 number split is a static offline benchmark with no further measurement rounds, MULTI-evolve cannot run under the same protocol. We treat it as prior motivation, and make this explicit in Section 2.

W2 / Q2. Evidence for the synthetic preference labels, and mutation-count shortcuts

The synthetic negatives are unmeasured, but FLIP2 contains measured 3+ mutants, so the assumption is directly testable. Variants carrying more inactivity-pool substitutions have systematically lower measured activity, confirming these substitutions remain deleterious when recombined into higher-order backgrounds. Compensatory epistasis is not absent — a minority remain active — but it is not the typical case, which is what the ordinal constraint requires.

"Pool substitutions carried" is not the mutation count: it is how many of a variant's substitutions lie in the 25-substitution inactivity pool. All variants come from the measured 3+ test split, so 0 means three or more substitutions, none in the pool. IRED is for reference only: its negatives are measured inactive variants from the source publication (Appendix A.7), not recombinants.

DatasetPool substitutions carriednMean activity% > WT
Amylase02,0340.0413.59
Amylase11110.0100.00
Amylase2210.0100.00
IRED03,565−1.36811.05
IRED1584−1.5139.59
IRED229−1.5790.00
TrpB0200,9900.0400.51
TrpB111,028−0.0560.00
TrpB25,138−0.0780.00
TrpB3351−0.0930.00
NucB039,2410.65323.35
NucB12,5150.1202.31
NucB2740.0000.00

Mutation-count-matched controls. Figure 1 already compares models within fixed test mutation loads, where mutation count carries no ranking information. We additionally report Spearman within fixed k = 3, 4, 5, 6 groups. NucB is omitted: its four ordinal classes leave 41–46% of variant pairs tied within each group, so rank correlations are not comparable with the continuous datasets. TrpB's library contains only k = 3 and k = 4.

DatasetModelk=3k=4k=5k=6
AmylaseProteinMPNN0.4230.3500.3830.279
+PIF0.6410.5280.5610.488
ESM-IF0.4330.3570.4380.221
+PIF0.6160.4960.5760.404
IREDProteinMPNN0.1130.126−0.0660.041
+PIF0.2640.183−0.0070.067
ESM-IF0.1540.115−0.0180.006
+PIF0.2390.165−0.0040.029
TrpBProteinMPNN0.2420.086——
+PIF0.6300.126——
ESM-IF0.4190.078——
+PIF0.6330.138——

PIF improves the pretrained model in every fixed-k setting, with the largest gains at k = 3. Because mutation count is constant within each comparison, these improvements cannot be explained by a shortcut that penalizes variants with more mutations.

Response to Reviewer 2V2i (2/2)

Q3. Appendix A.8.1 — does the filter always trigger the fallback?

The interpretation arises from an ambiguity in Appendix A.8.1. The positive- and negative-mutation pools are identified using different activity-based criteria. In the implementation, active mutations are drawn from the active sequence subset defined in Table 9, not from all training sequences, while negative mutations come from the inactive subset. The two pools are therefore not nested and have only limited overlap.

The filtering step removes mutations also supported by the positive-side criterion, making the remaining negative pool more specifically associated with inactivity. In practice a substantial number of inactive mutations survive filtering on every dataset:

DatasetActive seqs (Table 9)Inactive mutationsSurvive filterRetained (%)
Amylase231 (top 15%)23720285.2
IRED298 (≥ 0.5)1,4271,05073.6
TrpB540 (top 5%)29324583.6
NucB1,182 (class ≥ 2)2,0411,72884.7

The fallback is therefore not triggered on any of the four datasets; it is retained only as a safeguard. We have revised Appendix A.8.1 to state the active-mutation source explicitly and cross-reference Table 9.

Q4. Consistency across training seeds

We retrained PIF from scratch with three independent seeds (42, 1 and 2), seed 42 being the run reported in the paper, and evaluated each on the identical held-out 3+ mutant test pool. The table reports mean ± sample standard deviation over the three runs, alongside the pretrained base model and the published single run. The standard deviation applies to the +PIF (3 runs) rows only. All @K metrics report K = 10/100. NucB has discrete activity classes, so no correlations are reported for it.

DatasetModelHC@KEF@KNDCG@KPearsonSpearman
AmylaseProteinMPNN1 / 132.97 / 3.860.71 / 0.670.4490.388
+PIF (published)4 / 2411.89 / 7.130.73 / 0.780.6480.599
+PIF (3 runs)3.0±1.7 / 22±48.92±5.15 / 6.44±1.200.68±0.04 / 0.74±0.030.630±0.0160.582±0.015
ESM-IF0 / 160.00 / 4.760.56 / 0.620.4330.398
+PIF (published)3 / 188.92 / 5.350.79 / 0.750.5860.569
+PIF (3 runs)3.0±1.0 / 21±48.92±2.97 / 6.14±1.130.70±0.08 / 0.74±0.010.579±0.0090.563±0.008
IREDProteinMPNN4 / 213.71 / 1.950.52 / 0.600.1080.116
+PIF (published)6 / 425.57 / 3.890.68 / 0.710.1960.199
+PIF (3 runs)5.7±0.6 / 37±45.26±0.54 / 3.47±0.390.67±0.01 / 0.70±0.010.198±0.0070.202±0.005
ESM-IF4 / 223.71 / 2.040.60 / 0.630.1250.137
+PIF (published)7 / 386.50 / 3.530.67 / 0.700.1660.186
+PIF (3 runs)7.0±0.0 / 37±16.50±0.00 / 3.40±0.110.69±0.01 / 0.69±0.010.166±0.0020.185±0.002
TrpBProteinMPNN0 / 70.00 / 14.840.38 / 0.400.1810.059
+PIF (published)3 / 2863.60 / 59.350.59 / 0.590.3590.433
+PIF (3 runs)3.0±1.0 / 26±863.60±21.20 / 55.12±17.350.53±0.05 / 0.55±0.040.377±0.0280.446±0.014
ESM-IF4 / 2284.79 / 46.630.44 / 0.500.2430.119
+PIF (published)8 / 48169.60 / 101.760.70 / 0.660.4040.410
+PIF (3 runs)7.0±1.0 / 42±6148.40±21.20 / 88.33±11.680.66±0.06 / 0.62±0.040.388±0.0130.400±0.015
NucBProteinMPNN5 / 472.27 / 2.130.46 / 0.41——
+PIF (published)9 / 564.08 / 2.540.64 / 0.48——
+PIF (3 runs)7.7±1.2 / 55±43.48±0.52 / 2.48±0.190.60±0.04 / 0.47±0.02——
ESM-IF3 / 291.36 / 1.320.21 / 0.24——
+PIF (published)8 / 443.63 / 2.000.56 / 0.45——
+PIF (3 runs)4.7±2.9 / 40±42.12±1.31 / 1.81±0.180.44±0.11 / 0.40±0.05——

The improvement over the base model is consistent across seeds. Variability is concentrated in the small-K retrieval counts, as expected: HC@10 counts at most ten variants, so one reordering near the top moves it by a whole unit. The widest spread is TrpB EF@10 (63.60 ± 21.20), where only 0.47% of the 217,507 test variants are active, so one additional retrieved hit shifts the enrichment factor by 21.20. Even there the weakest run retrieves HC@100 = 17 against the base model's 7, with Spearman 0.431 against 0.059. The rank correlations are the most stable quantities reported, with Spearman standard deviations of 0.002–0.028.

These mean ± standard deviation values are reported in the revision.

Reviewer_j8QqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a method for fine-tuning pretrained protein inverse-folding models, evaluated on ProteinMPNN and ESM-IF, to improve the prioritization of higher-order mutants using fitness measurements from low-order mutants. The proposed method combines preference pairs constructed from experimentally measured low-order variants and synthetic higher-order negative samples, critical-site preference optimization, and protein foundation model embeddings. The authors demonstrate that their approach generally improves higher-order variant prioritization compared to the included baselines across four protein fitness datasets.

Critical strengths
  1. The paper addresses an important and challenging problem in protein engineering: predicting the fitness of higher-order mutants using only low-order experimental measurements.

  2. The idea of reformulating higher-order mutant prioritization as a direct preference optimization problem is creative and provides an interesting approach to leveraging pretrained inverse-folding models.

Critical weaknesses

There are several weaknesses that I consider important to address in order to strengthen this work.

1. Missing additive baseline for higher-order fitness prediction.

The paper lacks an important baseline for higher-order mutational fitness prediction. One of the major challenges in predicting higher-order mutational effects is capturing epistasis, where the combined effects of mutations cannot be explained by simply adding their individual effects. A successful method should demonstrate improvements over a simple additive model, particularly for variants exhibiting substantial epistatic interactions. Otherwise, reasonably accurate predictions may be obtained simply by summing the measured effects of individual mutations. Given that the proposed method uses both single- and double-mutant measurements, comparisons against additive and pairwise-interaction baselines would be particularly informative. The absence of such comparisons makes it difficult to assess the actual benefit of the proposed approach.

2. Limited effectiveness in retrieving the highest-fitness variants.

The results in Table 13 of the supplementary material raise concerns about the effectiveness of the proposed method in identifying the most promising variants. Despite the improvements reported in the main paper, the full model fails to retrieve any of the true top-10 variants in three out of four datasets, for both ProteinMPNN and ESM-IF. Although the method improves several ranking and enrichment metrics, these results suggest that its ability to identify the very best candidates remains limited. This is particularly important given that prioritizing high-fitness variants is the central motivation of the work.

3. Inconsistent effectiveness of synthetic negative samples.

The ablation results in Table 1 raise questions about the effectiveness of the proposed synthetic negative-sampling strategy. In particular, when applied to ESM-IF, the full model achieves Spearman correlations that are nearly identical to, or occasionally worse than, those of the model trained without synthetic negative samples. While the negative samples provide improvements on certain datasets and evaluation metrics, their contribution appears inconsistent across architectures and tasks. This raises concerns about the general effectiveness of this component and warrants further investigation into when and why synthetic negative samples are beneficial.

Questions for authors

Beyond the weaknesses mentioned above, I have a question regarding the differences in performance between ProteinMPNN and ESM-IF. Both are inverse-folding models trained with similar objectives, although they employ different architectures, training procedures, and pre-training data. Nevertheless, the two models exhibit different baseline performance, and these differences appear to become more pronounced in some cases after fine-tuning with the proposed method.

Do the authors have any insights into why the two models respond differently to the same fine-tuning framework? Could these differences arise from their underlying architectures, pretrained representations, pretraining data, or sensitivity to the fine-tuning procedure?

I find these differences particularly interesting, as understanding them might reveal additional insights beyond the main contributions of this paper.

2 comments

Response to Reviewer j8Qq

Thank you for the detailed assessment. We address each point below.

W1. Missing additive baseline for higher-order fitness prediction

The suggested comparisons address a different task from ours.

  • Different task formulation. PIF performs protein inverse folding with functional alignment: the model remains structure-conditioned, and its sequence likelihood is optimized to reflect both structural compatibility and functional activity. Additive models, pairwise-interaction models and ESM2-based supervised predictors instead map sequence directly to fitness.
  • Different objective. Our goal is not to regress the higher-order fitness landscape or train a separate activity predictor, but to incorporate limited functional supervision into the inverse-folding distribution, so that structural compatibility and functional activity are jointly reflected in one generative model.

These fitness predictors are therefore not directly comparable baselines for the problem studied here. Our comparisons are with pretrained and preference-aligned inverse-folding methods: ProteinMPNN, ESM-IF, SPPO and FPA.

W2. Limited effectiveness in retrieving the highest-fitness variants

Table 13 evaluates a much more stringent notion of success than Tables 1-2: exact identity overlap with the globally highest-ranked top-K variants. For TrpB the true top-10 is only 10/217,507 ≈ 0.0046% of the test pool, so exact recovery under this selection ratio is far harder than identifying highly active candidates, which is the practical objective of our setting.

The prioritization results already show substantial gains. On TrpB with ESM-IF, PIF improves HC@10 from 4 to 8, EF@10 from 84.79 to 169.60 and NDCG@10 from 0.44 to 0.70; with ProteinMPNN, HC@10 from 0 to 3 and EF@10 from 0 to 63.60. Similar gains hold on Amylase, IRED and NucB. Notably, although ESM-IF + PIF has zero exact top-10 overlap on TrpB, 8 of its 10 selected variants exceed wild-type activity. Zero exact overlap therefore does not imply failure to retrieve active candidates; it means only that the retrieved variants are not exactly the ten globally highest sequences in a very large pool.

We clarify this in the revision and additionally report the median rank percentile of the true top-K variants under each model, characterizing retrieval quality on a continuous scale.

W3. Inconsistent effectiveness of synthetic negative samples

We respectfully disagree. Synthetic negatives target higher-order candidate prioritization, and the ablation supports this directly: comparing PIF (Full) with PIF w/o Neg, HC@10 and EF@10 improve or remain unchanged in every dataset and base-model setting. The small Pearson/Spearman changes are not inconsistent with that objective, since those metrics characterize global ranking over the entire test pool, whereas our target setting is top-K prioritization under a limited experimental budget.

Effect of synthetic negatives on HC@K, K = 10/100:

DatasetMPNN w/o NegMPNN FullESM-IF w/o NegESM-IF FullHigher-order negatives?
NucB1 / 89 / 560 / 128 / 44yes
Amylase0 / 164 / 243 / 183 / 18yes
TrpB1 / 113 / 286 / 368 / 48yes
IRED5 / 376 / 425 / 397 / 38no

The varying contribution tracks how much new information the construction provides. Amylase, TrpB and NucB introduce recombinant higher-order negatives with r ≥ 3 substitutions, whereas IRED uses measured inactive variants carrying only one or two mutations (Appendix A.7). IRED is thus the only dataset where the component adds no synthetic higher-order examples, consistent with its smaller improvement and the slight decrease on ESM-IF at K=100 (HC@100: 39 to 38). The benefit therefore comes from exposing the model to informative higher-order negatives, not from simply adding more negatives.

Q. Why do ProteinMPNN and ESM-IF respond differently?

Across all three continuous-fitness datasets, ProteinMPNN receives consistently larger Pearson and Spearman gains after PIF, while the two base models show no consistent dominance in pretrained top-K performance.

We attribute this to their different architectures and pretrained representations. ESM-IF is a larger GVP-Transformer with large-scale structural pretraining and may already encode richer sequence-context dependencies, leaving ProteinMPNN more room to benefit from ESM2 conditioning and cross-order supervision. Both base models nevertheless improve on the higher-order prioritization objective, so the asymmetry concerns the size of the marginal gain rather than whether the framework is effective.

Thank you for your detailed response. I appreciate your efforts.

I would like to briefly restate my first weakness, as it remains my major concern regarding the effectiveness of this work.

Higher-order mutational effect prediction is a well-established and challenging problem, precisely because of the epistatic interactions I mentioned in my original review. Even with complete measurements of single mutants or other low-order mutants, we cannot perfectly predict higher-order mutational effects, as higher-order interactions cannot necessarily be inferred from lower-order measurements. This is exactly the challenge that PIF attempts to address, using low-order experimental measurements as supervision.

A simple additive model addresses the same problem by summing the experimentally measured effects of individual mutations, while ignoring epistasis entirely. When the relevant single-mutant measurements are available, this baseline requires no additional training and no structural information. It is a standard and fundamental baseline for supervised higher-order mutational effect prediction. Zero-shot prediction might be a different case, but PIF is clearly not operating in a zero-shot setting, as it explicitly uses experimental fitness measurements for training.

I understand that the proposed method is formulated as structure-conditioned inverse folding rather than direct fitness regression. However, the difference in model formulation does not make these baselines incomparable when they address the same downstream task. Both approaches can rank the same higher-order variants using the same low-order experimental measurements, and their effectiveness can be evaluated using exactly the same metrics.

In fact, the additive baseline does not even use structural information. If a more sophisticated model incorporating structural representations, pretrained protein models, and preference optimization cannot outperform this simple baseline, it is difficult to establish what additional value the proposed method provides for higher-order variant prioritization.

Therefore, I still consider the absence of this comparison a fundamental weakness of the paper, and I do not believe the response has addressed my original concern.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Manuscript Revisions

Summary of Rebuttal and Manuscript Revisions

We thank the reviewers for their comments. The table maps each main question to our response and the corresponding manuscript revision (additions marked in blue).

QuestionRaised byShort answerRevision
Are additive / pairwise / ESM2 / FLIP2 / MULTI-evolve appropriate baselines?j8Qq W1; 2V2i W1/Q1; cVwS Q4Different task. PIF stays structure-conditioned and aligns a sequence likelihood; these predictors regress fitness directly. MULTI-evolve needs iterative wet-lab rounds that a static benchmark cannot provide.Section 4.1 (Baselines); Section 2
What is novel in the two-to-many setting?cVwS W1/W2; mkbp W1/Q1We do not claim low-to-high-order generalization as new. Our contribution is to study this shift within inverse folding, combining cross-order preference construction with foundation-model conditioning.Section 2; Section 3.4
Are the synthetic negatives biologically meaningful?2V2i W2/Q2; cVwS W2; mkbp W2/Q2Yes. On measured 3+ variants, activity falls monotonically with more pool substitutions. One substitution drops the fraction exceeding WT from 23.35% to 2.31% (NucB), 3.59% to 0% (Amylase), 0.51% to 0% (TrpB).Section 3.2; Appendix E (new)
Could PIF be exploiting mutation count?2V2i W2/Q2; cVwS W2No. Within fixed-k groups (k = 3, 4, 5, 6), where mutation count carries no ranking signal, PIF improves the base model in every comparison, with the largest gains at k = 3.Appendix F (new); Figure 1 caption; Section 4.2
Are the full and ablated results inconsistent?j8Qq W3; cVwS Q3No, for the top-K prioritization. Against either ablation, HC@10 and EF@10 improve or hold in every setting. The negative component is weaker on IRED, the only dataset without synthetic higher-order negatives.Section 4.2 (Ablations)
Why do ProteinMPNN and ESM-IF respond differently?j8QqProteinMPNN gains more in Pearson/Spearman on all three continuous datasets, consistent with its pretrained representation having more room to benefit from ESM2 conditioning and cross-order supervision.Section 4.2
Does weak exact top-K recovery undermine the method?j8Qq W2No. Exact recovery is far stricter than retrieving actives: TrpB's true top-10 is 0.0046% of 217,507 candidates. ESM-IF+PIF still improves HC@10 from 4 to 8, EF@10 from 84.79 to 169.60 and NDCG@10 from 0.44 to 0.70.Table 13 caption; new rank-percentile table
Are key implementation details ambiguous?2V2i Q3; cVwS Q1/Q2The two mutation pools use different activity criteria; 73.6–85.2% of inactive mutations survive filtering and the fallback never fires. Conditioning uses the candidate sequence under a frozen, precomputed encoder; "critical sites" are the pair-differing positions.Section 3.3; Section 3.4; Appendix A.8.1
Does Figure 1 contradict the two-to-many claim?cVwS Q5No. Tables 1–2 use only WT/single/double measurements. Figure 1 is a diagnostic varying k_train = 1 to 4; most of the gain is already achieved at k_train = 2.Figure 1 caption; Section 4.2
Are the improvements robust across seeds?2V2i Q4Yes. Three independent runs improve over the base models. Variance concentrates in small-K retrieval counts, while rank correlations are stable (Spearman SD 0.002–0.028).Appendix D (new); Table 1–2 captions

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.