acceptodds
Under review as a conference paper at ICLR 2027

EvoTilt-MPNN: Distributional Tilting for Variation-Aware Binder Design

Abstract

Designed protein binders can fail as their targets evolve: a single interface point mutation can abolish binding, making escape a major failure mode in therapeutic and diagnostic applications. However, current binder-design methods condition on a single target, making their designs vulnerable to mutations. To address this, we introduce a structure-based binder-design approach in which each target-interface residue is specified by a distribution over likely mutations. During autoregressive design, we sample each binder residue from an inverse-folding model's conditional distribution, tilted toward residues that reduce the KL divergence between and the model's predicted distribution at the target interface. Memoization computes this tilt at a small additional cost. Tilting produces a consistent distributional response that transfers to held-out inverse-folding models across twenty-one targets. In silico co-folding on fifteen of these shows a small, heterogeneous, tail-concentrated improvement in worst-tail confidence across sampled target mutations, relative to an untilted ProteinMPNN baseline. Our method provides an efficient way to incorporate plausible target variation into sequence design. Because it accepts any per-residue distribution, can be drawn from deep mutational scanning, phylogenetic frequencies, or escape predictors, making the approach a general route to variation-aware binder design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein design, robustness, inverse folding, sequence design, generative models, boltzgen, proteinMPNN

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jBSDOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents EvoTilt-MPNN, a protein binder design method aimed at target sequence variation. Unlike the conventional practice of designing sequences against a single fixed wild-type target, the method describes the potential mutations at each target-interface residue as a per-residue amino-acid probability distribution π, and introduces a KL-divergence-based distributional tilt during ProteinMPNN's autoregressive decoding, so that the tilted inverse-folding posterior on the target side moves closer to π. This biases the generated binder toward tolerating these potential substitutions without explicitly enumerating mutant structures. To contain the extra cost, the authors design a cache-based leave-one-out decoding scheme that keeps the runtime of the tilted procedure on the same order as vanilla ProteinMPNN.

Experiments span multiple target families. The results show that the method consistently reduces the distributional divergence at the target interface, and that this effect partially transfers to inverse-folding models not used for optimization. On the mutant co-folding evaluation for a subset of these families, mild tilting yields a small, target-heterogeneous, worst-tail-concentrated improvement in structural confidence, whereas stronger tilting does not sustain the gain and causes some metrics to fall back below baseline. The paper demonstrates the feasibility of incorporating a mutation prior directly into sequence generation, but the improvement in actual binding robustness remains insufficiently validated.

Critical strengths

The problem studied is of clear biological significance. Compared with designing sequences against a single target structure, modeling potential target mutations as a per-residue probability distribution offers a general route to variation-aware binder design that does not require explicit enumeration of mutant structures and can be fed priors from DMS, phylogenetic frequencies, or escape predictors. Using the target-side inverse-folding posterior for distributional control at the generation stage carries some methodological novelty.

The computational optimization is sound and well validated. By caching the leave-one-out decoding states, the authors substantially reduce the overhead of distributional tilting, keeping its runtime on the same order as vanilla ProteinMPNN; more importantly, multiple exact implementations produce identical sequences, which gives the "exactness" claim hard support rather than resting on a speedup figure alone.

The authors test distribution matching across multiple families and verify cross-model transfer using different ProteinMPNN checkpoints and ESM-IF1. By decomposing the changes in WT probability versus non-WT preferences, they provide evidence that the KL improvement under mild tilting stems mainly from preference reallocation among non-wild-type residues rather than simply inflating the predicted probability of the wild-type residue, effectively ruling out the trivial explanation that tilting merely makes the model more confident in the WT.

Critical weaknesses

The main weakness is that the core improvement in binder robustness lacks sufficiently strong evidence. The distribution-matching metrics are stable, but in the primary structural evaluation the gain in worst-decile interface iPTM is small and positive in only about half of the families; as the paper itself notes, most worst-decile readouts improve under mild tilting yet largely drift back toward baseline under stronger tilting. With such a narrow effective window and no reported cross-family statistical significance (e.g., a paired test or bootstrap CIs), it is hard to judge whether this small, heterogeneous improvement is reliable—especially since the number of interface positions hit by the mutation panel varies across designs, further affecting worst-tail stability.

The method is also not validated against adequate competitive baselines. The experiments mainly compare tilt strengths against vanilla ProteinMPNN, without a matched-budget comparison to post-hoc reranking under the same KL objective. Since decoding-time tilting can reach sequences outside the vanilla sampling pool while reranking is confined to existing candidates, the two explore different spaces; absent such a comparison, it is unclear whether tilting at decoding time is actually superior to simple post-generation filtering.

A clear gap remains between distribution matching and genuine binding robustness. All experiments use a single ESM-2-inferred prior, without relating it to real mutation frequencies, measured mutational tolerance, or binder escape; the paper also concedes that the interface divergence weights all positions equally, whereas binding is decided by a subset π knows nothing about—so the weak effect may stem from a prior misaligned with binding-sensitive positions rather than from the method. The structural evaluation further relies mainly on Boltz-2 with BoltzGen-generated backbones, which the authors admit are not fully independent, and the held-out transfer verifies only the distributional proxy, not structure or function. The results thus more directly support a small gain in structure-prediction confidence than in actual binding robustness.

Finally, some controls and approximations are under-validated: the anti-π arm is close to uniform, limiting its power as an adversarial control; the prior-specificity result is exploratory with CIs including zero; and the greedy per-position approximation to the joint objective of Equation 2 is never quantified, leaving open whether it partly explains the "strong distribution matching but weak structural gains" pattern.

Questions for authors
  1. The worst-decile interface iPTM improves in only about half of the families. Could you report a paired statistical test across families, bootstrap confidence intervals, and leave-one-family-out sensitivity analysis? Also, how many target-side interface positions are covered by the mutation panel per design, and is this coverage sufficient for a reliable worst-decile estimate?
  2. Under matched candidate counts and compute budgets, how does EvoTilt-MPNN compare with generating sequences using the original model and reranking them using the same KL-based objective? Does decoding-time tilting provide worst-tail robustness gains beyond those achievable through post-hoc reranking, particularly as the candidate pool increases?
  3. Since the backbones are generated by BoltzGen and structural robustness is mainly evaluated using Boltz-2, could you validate the results using an independent structure predictor or physics-based interface assessment? Does the worst-tail improvement persist under this independent evaluation?
  4. The method relies on ESM-2 marginals as mutation priors. For targets with available deep mutational scanning (DMS) data, could you construct an alternative prior from experimentally measured mutation preferences and compare the resulting structural robustness? This would help determine whether the limited structural improvement is primarily due to the quality of the mutation prior or the tilting method itself.
2 comments

We thank Reviewer jBSD for the careful and technically engaged review. We appreciate their recognition of the significance of variation-aware binder design, the methodological novelty of using the target-side posterior for distributional control, and the soundness and exact implementation agreement of the cached decoding optimization. Their questions motivated a substantially stronger evaluation and clearer statement of the method's scope.

1. Is the structural improvement statistically reliable, and is panel coverage sufficient? We agree that the original paper did not provide sufficient uncertainty analysis. The primary worst-decile interface-iPTM effect is +0.0176 (family-bootstrap 95% CI [+0.0001,+0.0374]; paired t-test p=0.098; 9/15 families favorable). It is therefore not resolved by the paired test at the available family-level power. However, the mean remains favorable under every leave-one-family-out omission. The response also persists as the tail broadens: +0.0106 at eta=0.25, +0.0051 at eta=0.50, and +0.0044 for the full on-interface mean. Complex PDE moves concordantly (-0.0252; 95% CI [-0.0495,-0.0006]; paired p=0.072; 10/15 favorable), and 12/13 worst-tail confidence readouts are favorable at beta=10 (Section 5.4; Appendices B and F).

We also quantified the requested panel coverage. The pi panel intersects a median of 10 distinct target-interface positions per design (IQR 7--12; range 2--30), or 25.5% of the modeled target interface. Because CVaR_0.10 uses max(1,floor(0.10 n_on)) mutations, its tail contains one mutation for 94.9% of primary design-condition rows. We now describe this endpoint as a conservative worst-observed on-interface stress test rather than a precisely resolved within-design decile; the broader-tail estimates above provide the corresponding higher-coverage sensitivities (Appendix A.6).

2. Does decoding-time tilting improve over post-hoc reranking as the candidate pool grows? We computed terminal KL for all 7,249 eligible vanilla candidates across the 15 families and selected the same 40-design output count. Pure-KL selection reduced terminal KL, but WT interface iPTM fell from 0.8283 to 0.6564 and mutant CVaR_0.10 from 0.6300 to 0.4920 (Delta=-0.1381; 95% CI [-0.1668,-0.1100]). Confidence-selected EvoTilt retained WT iPTM at 0.8302 and achieved mutant CVaR_0.10 of 0.6476, exceeding vanilla pure-KL selection by +0.1557 ([+0.1256,+0.1843]; favorable in 15/15 families).

We also evaluated the full model-free confidence--KL curve and structurally tested an intermediate M=60 policy, which selects the lowest-KL 40 designs among the 60 highest-confidence candidates. This compromise did not improve absolute CVaR for either vanilla or EvoTilt candidates. Thus the tested post-hoc selectors do not reproduce EvoTilt's confidence--robustness trade-off as the candidate pool increases (Section 5.6, Table 3; Appendix G). Because the selected cohorts differ, this is a comparison of complete generation-and-selection policies rather than a same-sequence causal decoder ablation.

3. Does the response persist under an alternative mutation-effect evaluation? We evaluated the same frozen 15-family cohort with FoldX and StaB-ddG. For the pi-arm, design-interface worst tail, FoldX changed favorably by -0.330 (95% CI [-0.552,-0.141]; 12/15 families; multiplicity-adjusted p=0.0061). StaB-ddG changed by -0.071 ([-0.133,-0.011]; 10/15; adjusted p=0.0891), providing directional but not multiplicity-adjusted evidence. The FoldX direction persisted after adding each design's WT interaction energy (-1.314; [-2.168,-0.580]; 12/15 families), arguing against an apparent mutant gain caused only by worse WT complexes (Section 5.7, Table 4; Appendix H). These alternative mutation-effect scores broaden the endpoint beyond Boltz-2, but they remain computational and start from Boltz-derived structures; we therefore do not present them as a fully independent structure-generation test or as experimental binding validation.

4. Does the method generalize beyond the ESM prior, including to DMS-derived priors? We agree that a functional DMS-derived prior would be the most direct test of whether limited structural improvement is caused by prior quality. Constructing and biologically validating such a prior is an important but distinct task from the problem addressed here—(designing against a supplied prior) We did not complete a DMS-prior structural campaign but have added it as an obvious next step and future work. We therefore cannot separate those two explanations experimentally. We did test the narrower question of whether the distributional mechanism is specific to ESM by repeating inverse design for all 15 primary families with an independently constructed, untruncated MSA-frequency prior. Target-side KL decreases in 15/15 families at beta=10 and beta=20. Reciprocal cross-scoring favors the matched prior in both directions: the matched-minus-crossed gain is +0.00946 at beta=10 (95% CI [+0.00683,+0.01268]) and +0.01970 at beta=20 ([+0.01412,+0.02640]; Section 5.2; Appendix E). This establishes posterior-level transfer across two prior sources, not structural generality of the MSA campaign or validation of homolog frequency as functional mutational tolerance.

Additional concerns: anti-prior control and the stepwise approximation We agree that the anti-pi panel is close to uniform and therefore only mildly adversarial; the revision no longer treats it as strong falsification evidence. Structural prior-specificity is also reported as exploratory: at beta=10, the on-interface pi-versus-uniform contrast is +0.0104 (95% CI [-0.0037,+0.0245]), while the off-interface contrast is near zero. Stronger evidence that the posterior response is prior-directed comes from the reciprocal MSA/ESM cross-scoring above and from the conditional-mutant objective, for which the pi-weighted reduction exceeds the uniform-weighted reduction in all 21 families.

Finally, caching is exact for the defined stepwise tilted proposal: independent exact implementations produce identical sequences. The greedy per-position proposal remains an approximation to the global joint tilt in Equation 2. Evaluating the terminal joint cost for a completed sequence is tractable and shows consistent KL reduction, but exact sampling from the global tilt would require marginalizing over exponentially many sequence completions. We did not quantify the distance between these two distributions and therefore cannot exclude this approximation as one contributor to the distributional--structural gap. The revised Methods and Limitations make this distinction explicit. However, as a small toy test case, we separated cache exactness from the greedy approximation for a single A1M β=10\beta=10 case: exact enumeration over 484^8 combinations at the eight most affected binder positions gave KL(Pβ∥Q)=0.0035KL(P_\beta\|Q)=0.0035 and TV =0.033=0.033, while full-sequence importance sampling over all 93 binder positions gave KL=0.019KL=0.019, TV =0.077=0.077, and showed that the stepwise proposal retained 93% of the ideal objective reduction. This single case study suggests that the approximation is modest in the tested setting, but does not establish equivalence across decoding orders, backbones, or families..

We thank the reviewer again: these analyses sharpen both the evidence for mild-tilt distributional and structural effects. We hope the new analysis help address the most pressing concerns.

Reviewer_VA69OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the challenge of designing protein binders that remain effective when their target proteins undergo mutations. While conventional binder-design approaches typically optimize sequences against a single wild-type target, existing multistate design methods can account for target variation by explicitly considering multiple mutant structures, often at considerable computational cost. The authors propose EvoTilt-MPNN, a method that incorporates plausible target mutations into binder sequence design without requiring individual mutant structures to be generated and evaluated during optimization.

EvoTilt-MPNN builds on the pretrained ProteinMPNN model and modifies its amino acid sampling probabilities rather than changing the neural network architecture or retraining its parameters. The method first uses ESM-2 to estimate the distribution of plausible amino acid substitutions at each target position. During binder sequence generation, it then favors amino acids that make the model's predicted target-interface residue distributions more consistent with these mutation priors, using a KL divergence-based objective. To make this procedure computationally practical, the authors introduce a cached decoding strategy that limits the additional runtime to approximately 1.4 times that of standard ProteinMPNN.

The authors evaluate their approach across 21 target families and demonstrate consistent improvements in agreement with the specified mutation priors, including partial transfer to independent inverse-folding models. In structural evaluations involving 15 target families, moderate tilting largely maintains wild-type structural confidence while modestly improving the worst-decile interface iPTM for sampled single-residue target mutations from 0.6300 to 0.6476. However, these gains are not consistent across all targets and diminish at higher tilting strengths.

Critical strengths
  1. Efficient Integration of Target Variation into Binder Design

EvoTilt-MPNN incorporates predicted target mutations into ProteinMPNN's sequence generation without retraining or explicitly modeling individual mutant structures. Its cached decoding algorithm limits the computational overhead to approximately 1.4× that of standard ProteinMPNN, making the approach practical for large-scale binder design.

  1. Systematic Validation of Distributional Effects

Experiments across 21 target families demonstrate consistent improvements in agreement with mutation priors, with partial transfer to independent inverse-folding models. The authors also examine the trade-off between tilting strength and structural confidence, providing a clear assessment of the method's effectiveness and limitations.

Critical weaknesses
  1. Limited Methodological Novelty and Lack of Comparisons with Existing Variation-Aware Design Approaches

EvoTilt-MPNN primarily extends ProteinMPNN through inference-time probability reweighting, with its main technical contributions being the mutation-aware objective and efficient cached decoding. However, the paper does not provide direct comparisons with existing approaches capable of addressing target sequence variation, such as ProteinMPNN combined with variant-based post-design screening or multistate design methods such as Rosetta. Although these approaches have different computational requirements, comparisons under matched computational budgets would help establish the relative advantages of EvoTilt-MPNN. Without such baselines, its methodological novelty and practical advantages over existing approaches remain insufficiently demonstrated.

  1. Limited and Inconsistent Improvements in Binder Design Performance

Although EvoTilt-MPNN consistently improves agreement with the specified mutation priors, the improvements in predicted mutation robustness are modest and vary considerably across targets. The primary worst-decile interface iPTM increases from 0.6300 to 0.6476, with improvements observed in only 9 of 15 target families. Furthermore, the evaluation relies on predicted structural confidence rather than direct measurements of binding affinity or mutation tolerance. Consequently, the results do not yet provide convincing evidence that EvoTilt-MPNN delivers a substantial improvement in mutation-robust binder design.

  1. Insufficient Evaluation of Generalizability Across Mutation Priors

The proposed method is evaluated using only a single mutation prior derived from ESM-2. While the authors assess designs against mutation panels sampled from the prior, uniform distributions, and anti-prior distributions, they do not investigate the effects of using alternative priors during binder generation. Since ESM-2 residue probabilities do not necessarily reflect observed mutation frequencies or binding escape, it remains unclear whether the reported improvements would generalize to priors derived from other protein language models, deep mutational scanning, or experimentally observed mutation distributions. Given that the mutation prior is central to the proposed method, this lack of evaluation limits the evidence for its generalizability.

Questions for authors
  1. Comparison with Existing Variation-Aware Binder Design Approaches

Could the authors compare EvoTilt-MPNN with existing approaches for designing mutation-tolerant binders, particularly ProteinMPNN followed by variant-based post-design screening and multistate design methods such as Rosetta? A comparison under matched computational budgets would help clarify whether the proposed approach offers advantages in mutation robustness or design efficiency beyond standard ProteinMPNN.

  1. Significance and Consistency of Binder Robustness Improvements

The improvement in worst-decile interface iPTM is relatively modest (0.6300 to 0.6476), with favorable results in 9 of 15 target families. Could the authors provide statistical significance analyses and further investigate why some targets benefit from tilting while others do not? In particular, can the authors demonstrate that the observed improvements are robust to variations in candidate selection and evaluation settings?

2 comments

No rebuttal, no discussion. I keep my rating.

We thank Reviewer VA69 for the careful and constructive review. We appreciate their recognition of the practical value of incorporating target variation without retraining or enumerating mutant structures, the efficiency of cached decoding, and the systematic distributional validation. Their questions motivated clearer comparisons with post-hoc selection, stronger analysis of heterogeneous structural effects, and an alternative-prior experiment.

Weakness 1 and Question 1: methodological novelty and existing variation-aware approaches We agree that direct, compute-matched comparisons with explicit post-design variant screening and Rosetta multistate design would be valuable. We did not complete those head-to-head experiments and therefore do not claim superiority to multistate methods. An expensive Rosetta design campaign with our selected 120 mutants, coupled with our 216,000-mutant co-folding evaluation was difficult to complete during the rebuttal period. We have revised the Limitation section and name this missing comparison explicitly.

The contribution is narrower than a new network architecture but more than generic probability reweighting: EvoTilt formulates variation-aware design as distributional control of the target-side inverse-folding posterior, then uses an exact cache for the defined stepwise proposal to make that objective practical during autoregressive decoding. Unlike explicit screening and multistate design, it consumes a per-position prior without enumerating or materializing mutant target structures during design, and runs at 1.40x vanilla ProteinMPNN.

We added the closest matched post-generation baseline available under the same 40-design output and downstream evaluation: vanilla ProteinMPNN followed by the same terminal-KL objective. Pure KL reranking improves its surrogate but substantially reduces WT and mutant structural quality; a confidence-constrained M=60 policy also fails to improve the structural endpoint (Section 5.6, Table 3; Appendix G). Thus the tested post-hoc selectors do not reproduce EvoTilt's trade-off, although this does not substitute for a future direct multistate comparison.

Weakness 2 and Question 2: modest, heterogeneous structural effects

We agree that the primary structural response is modest and not family-universal. Its effect is +0.0176 (95% CI [+0.0001,+0.0374]; paired p=0.098; 9/15 favorable families), so we do not present it as statistically resolved at the available power. The response is nevertheless favorable under every leave-one-family-out omission and across broader tail fractions. Family heterogeneity is partly explained by objective engagement: at mild tilt, larger family-level reductions in the design-time KL associate with larger absolute structural responses (Spearman rho=0.782, permutation p<0.001, all leave-one-family-out correlations remained positive; Section 5.5, Figure 4). Defining the interface from each WT co-folded structure gives, as opposed to the original design gives +0.0166 (12/15 favorable), and restricting to refold--cofold challenge pose-consistent designs gives +0.0233 (12/15), the latter reported as post-hoc sensitivity (Section 5.4; Appendix B). Evaluation-setting sensitivity is partially addressed on the same frozen 15-family cohort with FoldX and StaB-ddG: FoldX is favorable after multiplicity correction (95% CI [−0.552,−0.141][-0.552,-0.141]; 12/15 families; multiplicity-adjusted p=0.0061p=0.0061), while StaB-ddG is directionally favorable but not significant after correction ([−0.133,−0.011][-0.133,-0.011]; 10/15; adjusted p=0.0891p=0.0891) (Section 5.7, Table 4; Appendix H). These results support converging mild-tilt evidence, not a universal effect across targets, selectors, or evaluators in a low signal to noise ratio regime.

Weakness 3: generalization across mutation priors We agree that an ESM-derived prior alone cannot establish generality to functional DMS or escape priors. The revised paper separates constructing and biologically validating a prior from the problem addressed here: designing against a supplied prior. To test whether the mechanism itself is ESM-specific, we repeated inverse design for all 15 primary families with an independently constructed, untruncated MSA-frequency prior. Target-side KL decreases in 15/15 families at both beta=10 and beta=20. Reciprocal cross-scoring favors the matched prior in both directions, with a matched-minus-crossed gain of +0.00946 at beta=10 (95% CI [+0.00683,+0.01268]) and +0.01970 at beta=20 ([+0.01412,+0.02640]; Section 5.2; Appendix E). This establishes posterior-level transfer across two materially different priors; it does not establish structural generality of the MSA campaign or generality to functional DMS priors, which remain future work.

We hope the revisions help to address your major concerns. Thank you for the careful review.

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

The paper introduces EvoTilt-MPNN, a variation-aware protein binder design method that aims to generate binders robust to mutations in their target proteins. Instead of designing against a single target sequence or explicitly enumerating target variants, the authors represent plausible substitutions at each target-interface position using a probability distribution. During autoregressive sequence generation with ProteinMPNN, the sampling distribution is adjusted to reduce the KL divergence between the target mutation prior and the model's predicted target-side residue distribution. The authors also introduce a caching strategy to efficiently compute the adjusted sampling probabilities. Experiments across 21 target families show consistent improvements in distributional alignment, which partially transfer to held-out inverse-folding models, including ESM-IF1. Additional co-folding experiments using Boltz-2 on 15 target families show modest improvements in worst-decile interface confidence under mild tilting, with interface iPTM increasing from 0.6300 to 0.6476 at β = 10. The proposed caching method reduces the computational overhead to approximately 1.40 times that of vanilla ProteinMPNN.

Critical strengths
  • The paper addresses an important problem in protein binder design by considering potential target mutations during sequence generation rather than relying solely on a fixed target sequence.
  • The distributional tilting formulation is simple and well motivated. It allows the incorporation of target mutation priors without explicitly enumerating all possible target variants and their structures.
  • The proposed caching strategy substantially reduces computational overhead, making the method practical for large-scale binder sequence generation.
  • The authors provide extensive analyses across different target families, tilt strengths, and evaluation settings, including held-out inverse-folding models, variant challenges, and controls using different mutation distributions.
Critical weaknesses
  • Although distributional tilting consistently improves alignment with the target mutation prior, the improvements in downstream structural robustness are relatively small and inconsistent. For example, worst-decile interface iPTM increases from 0.6300 to 0.6476 at β = 10, but only 9 of 15 target families show improvements. The gains also diminish at stronger tilt strengths. More rigorous statistical analysis of the primary robustness improvements would strengthen the paper.
  • The downstream design experiments primarily compare EvoTilt-MPNN against vanilla ProteinMPNN. However, a simpler alternative would be to generate candidates using vanilla ProteinMPNN and rerank them using the same KL-based distributional objective. Comparing against such a baseline under matched computational budgets would help establish whether modifying the autoregressive sampling distribution provides additional benefits over post-hoc candidate selection.
  • The mutation priors are generated exclusively using ESM-2, with each distribution truncated to the top 60% of its probability mass. Although the authors motivate the approach using evolutionary frequencies, deep mutational scanning, and experimental escape information, these alternative sources are not evaluated. The authors include uniform and anti-prior mutation challenges, but it remains unclear whether the improvements would generalize to different mutation priors.
  • The downstream robustness evaluation relies on predicted structural confidence from Boltz-2 rather than direct binding measurements. Although the authors evaluate distributional alignment using held-out inverse-folding models, these experiments do not directly establish binding robustness. Moreover, the co-folding evaluation is not fully independent of the BoltzGen backbone-generation pipeline. Additional evaluation using an independent computational approach would strengthen the results.
Questions for authors
  • The primary robustness improvement is relatively small and observed in only 9 of 15 target families. Could the authors report the statistical significance or confidence interval of the improvement in worst-decile interface iPTM? It would also be helpful to understand why some target families benefit from tilting while others perform worse.
  • Have the authors considered generating candidates using vanilla ProteinMPNN and reranking them according to the same KL-based distributional objective? How much additional benefit does modifying the autoregressive sampling distribution provide over post-hoc selection under comparable computational budgets?
  • The mutation priors are constructed using ESM-2, while the framework is motivated as being applicable to evolutionary frequencies and experimentally measured mutations. How sensitive is the proposed method to the choice of mutation prior and the truncation threshold? Would the results remain consistent with alternative prior distributions?
  • The proposed distributional objective minimizes the KL divergence between the mutation prior and the target-side predictions of an inverse-folding model. Does a larger reduction in KL divergence correlate with stronger improvements in the downstream Boltz-2 robustness evaluation? In particular, is this relationship consistent across different target families?
1 comment
AuthorsOpenReview ↗

Response to Reviewer QcWW

We thank Reviewer QcWW for the careful and constructive review. We appreciate their recognition of the importance of variation-aware binder design, the motivation of the distributional formulation, and the practical value of cached decoding and broad evaluation. Their questions motivated stronger statistical analysis, direct post-hoc reranking controls, an alternative-prior experiment, and a quantitative link between objective reduction and downstream structural response.

Question 1: We now report family-bootstrap intervals, paired tests, favorable-family counts, and leave-one-family-out sensitivity (Section 5.4; Appendix F). The primary effect is +0.0176, with 95% CI [+0.0001,+0.0374], paired p=0.098, and 9/15 favorable families; every leave-one-family-out mean remains positive. We do not claim a family-universal or statistically resolved effect. The response is not confined to eta=0.10: the interface-iPTM effect remains directionally favorable at eta=0.25, eta=0.50, and the full on-interface mean, and complex PDE is also favorable (-0.0252; 95% CI [-0.0495,-0.0006]; paired p=0.072; 10/15 families). While the structural effects are small, we have additionally introduced an independent scoring with FoldX and StaB-ddg ML based method. On the exact 15-family cohort, the beta=10 FoldX worst-tail complex total-energy response is -0.330 (95% CI [-0.552,-0.141]; 12/15 families; multiplicity-adjusted p=0.0061). StaB-ddG's predicted binding ΔΔG response is -0.071 ([-0.133,-0.011]; 10/15; raw p=0.0446, adjusted p=0.0891). Adding FoldX WT interaction energy retains a favorable WT-anchored response (-1.314, [-2.168,-0.580]) (Section 5.7, Table 4; Appendix H). We

Question 2: We scored terminal KL for all vanilla candidates and selected the same cohort count used by EvoTilt. Pure KL selection lowers terminal KL by 0.1991, but relative to vanilla confidence selection it reduces WT interface iPTM by 0.1719 and absolute mutant CVaR_0.10 by 0.1381. Confidence-selected EvoTilt exceeds vanilla pure-KL selection by +0.1557 absolute CVaR_0.10 (95% CI [+0.1256,+0.1843]; favorable in 15/15 families). Refolding confidence selection is essential. A less extreme policy selecting the lowest-KL 40 among the top-confidence 60 also fails to improve the structural endpoint at beta=0 or beta=10 (Section 5.6, Table 3; Appendix G).

Question 3: We acknowledge that the current paper did not evaluate multiple prior sources, due to the overall cost on the evaluation side (generation and design is cheap). We however repeated inverse design on all 15 primary families using an MSA-frequency prior. Target-side KL decreases in 15/15 families at both beta=10 and beta=20. Reciprocal cross-scoring shows that ESM-prior and MSA-prior designs preferentially move toward their matched prior: the matched-minus-crossed gain is +0.00946 at beta=10 (95% CI [+0.00683,+0.01268]) and +0.01970 at beta=20 ([+0.01412,+0.02640]) (Section 5.2; Appendix E).

Question 4: The new family-response analysis provides a partial explanation for heterogeneity: families with larger beta=10 objective reductions have larger absolute structural responses (rho=0.782, permutation p<0.001, bootstrap CI [0.489,0.920]), with a positive association under every leave-one-family-out omission. The association remains when KL is measured within the selected co-folding cohorts (rho=0.596, p=0.021, BH q=0.026). The relationship weakens at beta=20, matching the observed non-monotonic dose response (Section 5.5, Figure 4).

We hope the revision answers your concerns. Thank you for the careful review.

Comments on the paper

2 comments
Area_Chair_yMVSOpenReview ↗

Authors: please respond to your reviewers

To the paper authors: Please acknowledge and respond to the thoughtful replies by reviewers. They have raised some going points about strengths and weaknesses of your paper, as well as questions that merit replies.

We thank the reviewers for their constructive feedback, and apologize for submitting this rebuttal late in the response period. Addressing the requested experiments required several large-scale computational campaigns with long runtimes, which took longer than anticipated. The reviewers' concerns converge on four questions, which we addressed with new analyses and corresponding revisions to the paper.

  1. How reliable is the observed small structural response? We added family-bootstrap intervals, paired tests, leave-one-family-out analyses, broader-tail summaries, and effective-tail counts. The primary effect is +0.0176 ([+0.0001,+0.0374]; p=0.098; 9/15 families), so it is not statistically resolved at the available power, although its direction remains favorable under every family omission and across broader tail fractions (Section 5.4; Appendices A, B, and F).

  2. Does tilting outperform simpler post-hoc selection? We evaluated pure KL reranking over all eligible vanilla candidates and a confidence-constrained M=60 policy. Both fail to reproduce the EvoTilt trade-off (Section 5.6, Table 3; Appendix G).

  3. Does the mechanism generalize beyond the ESM-derived prior? Biologically grounded priors from DMS, surveillance, or measured escape are important, but their construction and validation are a separate problem beyond this paper's scope. We factorize the task into (i) constructing a biologically relevant prior (not this work) and (ii) designing against a supplied prior (EvoTilt addresses this one). To test whether that mechanism is specific to ESM, we used an independently constructed, untruncated MSA-frequency prior: KL decreases in 15/15 families at beta=10 and beta=20, and reciprocal cross-scoring favors the matched prior in both directions (Section 5.2; Appendix E). This establishes cross-prior transfer, not biological validation of the MSA prior or generality to functional DMS priors.

  4. Does objective reduction connect to downstream effects under other evaluations? Objective reduction associates with structural response across families (rho=0.782, p<0.001) and within the selected cohorts (rho=0.596, p=0.021, BH q=0.026; Section 5.5, Figure 4). Exact 15-family FoldX and StaB-ddG analyses further broaden evaluation: FoldX is favorable after multiplicity correction, while StaB-ddG is directionally favorable but not significant after correction (Section 5.7, Table 4; Appendix H).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.