acceptodds
Under review as a conference paper at ICLR 2027

Ensemble-IF: Towards Protein Design for Target Conformational Populations

Abstract

Designing amino acid sequences that adopt one or more target conformations is an important problem in protein engineering, with applications in antibody and GPCR design. However, many use cases require controlling not only which conformational states are accessible, but also their equilibrium populations, a property that existing inverse folding-based design methods leave unconstrained. In this work, we introduce Ensemble-IF, an inverse folding-based pipeline for designing protein sequences that exhibit target equilibrium populations. Ensemble-IF combines pretrained inverse folding models with machine-learned conformational ensemble generation, scoring sequence-structure compatibility across sampled conformations to identify mutations that directionally reshape the conformational landscape. We validate designs with molecular dynamics simulations on events defined from reference trajectories, achieving an over 90% reduction in mean target-band distance error relative to baselines while maintaining structural integrity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein inverse folding, protein dynamics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_YfU7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ENSEMBLE-IF, a protein sequence design framework that aims to control the equilibrium populations of target conformational states rather than simply designing sequences compatible with one or multiple static structures. The method combines ProteinMPNN with BioEmu in an iterative design loop. At each iteration, it first uses KL projection to reweight conformations according to a specified target population interval. The resulting weights are then used to combine ProteinMPNN logits across the conformational ensemble and generate candidate sequences. BioEmu subsequently predicts the conformational populations of these candidates, and the best-performing sequence is selected for the next iteration. The authors evaluate the method on six target population intervals across three protein domains selected from mdCATH, together with a ubiquitin conformational switching case study. All-atom molecular dynamics simulations are used to assess population shifts and structural integrity.

Critical strengths
  1. Extending protein inverse folding from static structural compatibility to controlling equilibrium conformational populations is an interesting and biologically meaningful direction. Existing inverse folding methods largely focus on whether a sequence can adopt a target structure, while many protein functions depend on the relative populations of different conformational states. The problem addressed here is relevant to allosteric regulation and dynamic protein engineering, and I find the proposed design setting worth exploring.

  2. The experimental validation goes beyond the usual structure-prediction-based evaluation. The authors systematically identify conformational events from mdCATH trajectories using criteria such as replica consistency, state persistence, and threshold stability. They also perform explicit-solvent all-atom MD simulations to evaluate the resulting designs. This is a meaningful effort, as equilibrium population control cannot be adequately assessed through static structure prediction alone.

Critical weaknesses
  1. The connection between the proposed KL projection and actual equilibrium population control is not established. For the binary events considered in this paper, the projection essentially redistributes probability mass between two groups of conformations. After substituting this into the sequence decoding equations, the method reduces to a contrast between the average ProteinMPNN logits of the two states, scaled by the target population difference and a manually chosen guidance strength η. However, ProteinMPNN probabilities are not calibrated against conformational free energies, so reweighting these logits does not establish a quantitative relationship with the equilibrium populations of the generated sequences. The method ultimately relies on repeated BioEmu evaluation and candidate selection to approach the target interval. More importantly, the paper provides no component ablations to isolate the contribution of KL projection, and no analysis of how accurately BioEmu predicts the populations subsequently measured by MD. The current results demonstrate the performance of the overall search pipeline, but do not establish that the proposed projection is responsible for the improvement or that it provides reliable quantitative control over conformational populations.

  2. The experimental comparisons are not sufficiently controlled, and the MD results do not convincingly establish reliable equilibrium population control. ENSEMBLE-IF allows unrestricted sequence redesign using full conformational ensembles and multiple rounds of BioEmu feedback, whereas the main Conformational Biasing baseline uses two representative conformations and searches only single-point mutations. Even the additional CB-5 experiment does not match the sequence search space, structural information, or iterative feedback available to ENSEMBLE-IF. The reported performance advantage therefore cannot be attributed to the proposed algorithm itself. The evaluation is also limited to three top-ranked domains selected from 199 filtered domains, with only three of the six target intervals actually reached. For the 2v0cA03 low-population target of 14.40%–20.40%, the final MD occupancy is only 2.33%. Some high-population trajectories remain almost entirely within one state, while other cases show substantial disagreement across MD replicas. These results raise a more fundamental concern about whether the reported occupancies represent converged equilibrium distributions or finite-time trapping in particular conformational states. Given that controlling equilibrium populations is the central objective of the paper, the current validation is insufficient.

Questions for authors
  1. How many mutations does each final ENSEMBLE-IF design introduce relative to its native sequence? ENSEMBLE-IF allows unrestricted sequence redesign, while the main CB baseline is limited to a single mutation. How would the results change if both methods were evaluated under matched mutation budgets, BioEmu evaluation counts, and overall search budgets? The CB-5 experiment in the appendix still does not provide a fully matched comparison.

  2. For binary conformational events, the KL projection ultimately becomes a state-contrastive logit bias scaled by the target population difference and the guidance parameter η. Given that η must also be adjusted across systems, what is the actual benefit of the KL projection over simpler weighting schemes? How would the method perform with fixed positive/negative state weights, or with BioEmu-based iterative selection alone?

  3. How well do the conformational populations predicted by BioEmu agree with those measured by the final all-atom MD simulations? I would like to see a direct comparison of predicted and MD-measured occupancies for all evaluated designs and baseline sequences, including an analysis of systematic prediction errors across proteins and target intervals. Since BioEmu determines which candidates are accepted during optimization, its predictive reliability is essential to the proposed framework.

  4. How do the authors establish that the MD trajectories have adequately sampled the equilibrium conformational distributions? Several trajectories show very few apparent state transitions, while others exhibit large differences between replicas. All replicas for a given sequence also originate from the same equilibrated starting structure. I would like to see state transition counts, population convergence across trajectory segments, and simulations initialized from different conformational states. The use of three replicas for ENSEMBLE-IF but four for CB in the ubiquitin experiment also needs an explanation.

  5. For the 2v0cA03 low-population target of 14.40%–20.40%, ENSEMBLE-IF achieves only 2.33% occupancy in MD, while its predicted pLDDT drops to 57.17. Why did the BioEmu-based selection procedure fail to identify this design as problematic? Could the authors analyze the discrepancy between its BioEmu prediction and MD outcome, together with the structural quality of the resulting sequence? It would also be useful to examine whether incorporating folding-confidence or sequence constraints could improve the reliability of the design process.

2 comments

We thank the reviewer for their feedback and for recognizing the value of population-directed sequence design and MD evaluation. Please see our point-by-point responses below.

Weakness 2/Question 1: η=0\mathbf{\eta=0} control, mutation counts, and comparisons under matched budgets.

Regarding mutation counts and matched budgets, the manuscript Ensemble-IF designs introduce 25-56 substitutions, compared with one for each CB design; our response to Reviewer kgi4, Weakness/Question 1, lists the counts for every design.

The three-substitution comparison in our response to Reviewer N7W7, Weakness/Question 2, uses the same sampling budgets and gives Ensemble-IF the lowest mean error in 3/3 settings. On 1kwgA03, Ensemble-IF reaches the band in all three searches, compared with none for η=0\eta=0 and two for CB. The shared mutation ceiling addresses the unrestricted-redesign concern, while η=0\eta=0 retains the same ensemble-conditioned proposals and screening to isolate the guidance correction. Together, these controls show that guidance contributes to targeting under a common mutation allowance.

Weakness 1/Question 2: KL projection and alternative weighting schemes.

Regarding the role of KL projection, it gives the minimum change in ensemble weights, measured by KL divergence, needed to reach the requested population band. For binary events, it preserves relative weights within each state and changes the two state masses. This produces the state contrast identified by the reviewer,

zˉj(η)=zˉjw+η(uB−u)(zˉj+−zˉj−),\bar z_j^{(\eta)} = \bar z_j^w + \eta(u_B-u)(\bar z_j^+ - \bar z_j^-),

where uBu_B is the closest point in the band to the current population uu, and zˉj+\bar z_j^+ and zˉj−\bar z_j^- are conditional mean logits. The correction vanishes in band and reverses after crossing it; η\eta sets its strength. This rule directs sequence proposals. Each evaluated candidate receives its own BioEmu ensemble, followed by MD for selected designs.

Regarding fixed weighting, we implemented a fixed-contrast control that holds the signed state-contrast coefficient at its initial adaptive value throughout search. Specifically, at search round tt,

=\bar z_{j,t}^{w} +\beta_0\left(\bar z_{j,t}^{+}-\bar z_{j,t}^{-}\right), \qquad \beta_0=\eta\left(u_{B,0}-u_0\right),$$ where $u_0$ is the event population in the initial reference ensemble and $u_{B,0}$ is its closest point in the target band. The ensemble and conditional mean logits are updated as search proceeds, but $\beta_0$ remains fixed. For experiments, we reevaluated all six domain/target settings under unrestricted sequence redesign between $\eta=0$ and fixed contrast. The contrast coefficients were fixed before search using the formula above. We designated the first search seed in each setting and arm for MD before observing outcomes, with three 200 ns replicas per selected sequence and one fresh ESMFold2 structure per sequence. We use 200 ns due to the time intensiveness of running MD given that, for occupancy, most methods' trends stabilized within this time frame. The table below reports the distance of the mean replica occupancy to the target band, in percentage points. Every column uses the first 200 ns of production for fair comparison with the paper-reported numbers (normal Ensemble-IF and CB). |Domain|Target band (%)|$\eta=0$ error (pp)|Fixed-contrast error (pp)|Ensemble-IF error (pp)|CB error (pp)| |---|---|---:|---:|---:|---:| |1divA02|31.89–37.89|60.11|45.78|0.00|26.94| |1divA02|91.89–97.89|1.94|0.11|0.61|0.00| |1kwgA03|32.56–38.56|52.28|5.89|2.89|20.94| |1kwgA03|77.56–83.56|9.78|7.61|0.00|33.89| |2v0cA03|96.00–100.00|0.83|0.00|0.00|93.17| |2v0cA03|14.40–20.40|0.00|26.93|11.73|24.10| Fixed contrast lowers MD target-band error relative to $\eta$=0 in four settings, reducing mean error from 24.60 to 17.24 pp. Because these two controls share the same BioEmu screening budget and unrestricted mutation allowance, this improvement supports the contribution of state-aware proposals beyond screening alone. The full Ensemble-IF pipeline achieves the strongest overall targeting in this comparison. Across those same five settings, its mean band error is 2.92 pp, compared with 17.24 pp for fixed contrast, 24.60 pp for $\eta$=0, and 39.81 pp for CB. Ensemble-IF improves upon fixed contrast in four settings and ties it on high-target 2v0cA03. Thus, fixed contrast provides a useful population-directed proposal mechanism, while the complete Ensemble-IF pipeline delivers substantially better overall targeting in the reported results.

Question 3: BioEmu-MD occupancy correlation and prediction reliability.

Regarding prediction reliability, our response to Reviewer kgi4, Weakness 3/Questions 3-4, compares BioEmu and MD for every manuscript design and CB baseline. Across the 12 mdCATH pairs, Pearson correlation is 0.761, and the six Ensemble-IF designs have 4.22 pp mean absolute disagreement, exhibiting high agreement.

Weakness 2/Question 4: MD convergence, state transitions, and starting-structure sensitivity.

Regarding temporal stability and state transitions, the table below compares the two halves of the original 400 ns Ensemble-IF trajectories. Transitions count changes in the binary event between adjacent saved frames.

DomainTarget band (%)First 200 ns occupancy (%)Second 200 ns occupancy (%)Event transitions per replica, full 400 ns
1divA0231.89-37.8933.6743.3311 / 97 / 56
1divA0291.89-97.8998.5099.334 / 6 / 16
1kwgA0332.56-38.5629.6739.3364 / 53 / 31
1kwgA0377.56-83.5678.8879.5081 / 79 / 76 / 66
2v0cA0396.00-100.0097.5097.0020 / 16 / 22
2v0cA0314.40-20.402.672.0022 / 18 / 7

High-target 1kwgA03 and 2v0cA03 have similar occupancies in both halves, with repeated event transitions in every replica. The lower targets for 1divA02 and 1kwgA03 show larger shifts. The complementary uncertainty analysis in our response to Reviewer N7W7, Weakness 1/Question 1, gives 107-241 effective samples per replica for high-target 2v0cA03, with nearly identical occupancy intervals using 20 and 50 ns blocks. This design combines persistence near the requested population with repeated transitions and a precise occupancy estimate within the sampled trajectories.

Regarding robustness to MD starts, we provide an ESMFold2 comparison in our response to Reviewer N7W7, Weakness 1/Question 1, using differing predicted structures to initialize MD runs. Ensemble-IF remains closer to the band than CB in 5/6 settings. High-target 2v0cA03 reaches its 96-100% band in all three new replicas; low-target 2v0cA03 favors CB, with both methods outside the band. Keeping the sequences fixed makes this a direct test of initialization sensitivity: the main targeting advantage persists when the starting structures are regenerated.

We also apologize for using 4 replicas for ubiquitin instead of 3; this was an experimental error unchecked by our visualization pipeline. We have updated the manuscript accordingly. Their mean occupancies range from 26.28% to 40.32%, compared with Ensemble-IF's 61.58% and the 64.29-74.29% target. Ensemble-IF remains closer to the target in every comparison with equal replica counts.

Weakness 1/Weakness 2/Question 5: 2v0cA03 low-population target, BioEmu-MD discrepancy, and structural quality.

Regarding the low-target 2v0cA03 design, BioEmu predicted 14.81% occupancy (8/54 conformations), within the 14.40-20.40% band. Selection used population error without a folding-confidence filter, so pLDDT 57.17 did not reject it. BioEmu itself does not have pLDDT or confidence rejection filters. In MD, occupancy is 2.33% despite retaining 83.67% of native contacts. The BioEmu ensemble also shows structural distortion: event-positive conformations retain 37.71% of native contacts, versus 59.68% for event-negative conformations. We see this as evidence that structural integrity is indeed important for ensuring good occupancy performance, but not sufficient per our response to reviewer kgi4 under Weakness 4.

Regarding structural filtering, we applied identical ESMFold2 criteria to the separate high-target 2v0cA03 candidate pools (96-100%): pLDDT ≥ 74.644, native-contact retention ≥ 94.145%, and C$\alpha$ RMSD ≤ 1.522 Å. These thresholds correspond to the native-sequence reference values minus 5 pLDDT points, minus 5 percentage points of contact retention, and plus 1 Å RMSD, respectively. To evaluate the effect in MD, we simulated the three unfiltered sequences and the shared filtered native sequence from one search, using three 200 ns replicas per sequence with a maximum of 3 mutations. The table below gives occupancy with replica/20 ns-block intervals and fold retention.

Selected endpointMD occupancy % [95% CI]Target-band error (pp)Native contacts (%)RMSD (Å)
Unfiltered η=0\eta=033.17 [2.00, 58.34]62.8336.6610.61
Unfiltered CB20.83 [6.67, 34.67]75.1735.3210.60
Unfiltered Ensemble-IF95.67 [91.50, 98.33]0.3374.442.99
Quality-filtered native sequence, shared by all methods78.83 [72.00, 85.00]17.1792.971.40

The unfiltered guided design is closest to the target and retains more native contacts than either unfiltered control, although it lowers native contact % and RMSD. We hypothesize that, after a certain point, structural quality is only a necessary but insufficient condition for adequate event population control.

Reviewer_kgi4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Ensemble-IF, a protein sequence design framework that aims to control the equilibrium populations of user-defined conformational states. Unlike conventional single-state or multi-state inverse folding, which primarily seeks compatibility with specified structures, Ensemble-IF explicitly targets the relative occupancy of conformational states. The method consists of three components: (1) a KL-divergence projection that reweights a reference conformational ensemble toward a desired population interval, (2) population-guided sequence generation using weighted ProteinMPNN logits, and (3) iterative candidate selection using BioEmu-predicted conformational ensembles. The authors evaluate the method on six target population intervals across three mdCATH protein domains, with independent molecular dynamics (MD) simulations used for validation. They additionally provide a ubiquitin case study involving a biologically relevant pincer motion. The reported results show substantial improvements over Conformational Biasing (CB), including a reduction of more than 90% in mean target-band distance error across the six mdCATH settings.

Critical strengths
  1. An important and relatively underexplored protein design objective. The paper addresses a meaningful limitation of existing inverse folding methods: compatibility with multiple conformational states does not imply control over their equilibrium populations. The distinction is well motivated and potentially important for applications involving conformational selection and allosteric regulation. The target-population formulation provides a useful perspective beyond conventional multi-state protein design.

  2. A simple and interpretable population-guided design mechanism. The KL projection and its closed-form solution for binary conformational events are mathematically elegant. The resulting state-contrast term in Eq. (6), together with the guidance strength in Eq. (7), provides an interpretable connection between the desired population shift and sequence-level preferences. The method also leverages pretrained inverse folding models without requiring additional model training.

  3. Validation through molecular dynamics simulations. The use of independent all-atom MD simulations, rather than relying exclusively on BioEmu predictions, strengthens the experimental evaluation. The reported reductions in target-band distance are substantial, and the ubiquitin case study adds biological relevance to the otherwise geometry-based benchmark. The additional fold-quality measurements are also valuable for assessing whether the designed sequences preserve structural integrity.

Critical weaknesses
  1. The comparison with Conformational Biasing is not fully controlled for mutation budget. The primary CB baseline considers single-residue mutations, whereas Appendix A explicitly states that Ensemble-IF does not impose a limit on the number of mutations per proposal. Consequently, the two approaches may operate over substantially different sequence search spaces. Although Appendix E investigates a five-mutation CB variant, this comparison is limited to selected target settings and does not establish matched mutation or candidate-evaluation budgets. This is important because some of the reported improvement may result from the greater flexibility of multiple simultaneous mutations rather than from the proposed population-guided projection. The authors should report the number of mutations in each final design and provide comparisons under matched mutation counts and BioEmu evaluation budgets.

  2. Missing ablations make it difficult to isolate the contribution of population-aware guidance. The proposed framework combines weighted inverse folding with BioEmu-based candidate screening. However, the experiments do not sufficiently distinguish the contribution of these components. In particular, the candidate selection procedure directly minimizes the BioEmu-predicted distance to the target population interval. It is therefore unclear whether comparable improvements could be achieved using standard ensemble-conditioned ProteinMPNN proposals followed by the same BioEmu screening process. An especially important ablation would set the guidance strength to eta = 0 while preserving the proposal and evaluation budgets. This would directly test whether target-aware logit reweighting contributes improvements beyond ensemble-based proposal generation and oracle-guided selection. Additional comparisons with random or standard ProteinMPNN proposals followed by identical BioEmu screening would further strengthen the central methodological claim.

  3. Limited experimental coverage and uncertainty in equilibrium population estimates. The principal benchmark contains only three protein domains and six target intervals. Although Ensemble-IF outperforms CB on target-band distance across these settings, its MD-derived occupancy falls within the specified target interval in only three of the six experiments. Furthermore, the reported MD trajectories exhibit substantial variability between replicas in several cases. Given that the principal claim concerns equilibrium populations, the paper would benefit from a more detailed statistical analysis of occupancy estimation. In particular, Table 1 reports uncertainty for native contact recovery but not for event occupancy. Confidence intervals, effective sample sizes, time-correlation analyses, and convergence diagnostics would make the equilibrium-population claims more convincing. The authors should also examine the agreement between BioEmu-predicted populations and independently estimated MD populations to assess the reliability of the optimization oracle.

  4. Some baseline failures may be confounded by structural instability. For the 2v0cA03 domain, the CB designs exhibit substantially lower native contact recovery than Ensemble-IF. Appendix D also reports poor folding-quality metrics for CB on this domain. This raises the possibility that part of the occupancy-performance difference reflects structural disruption in the baseline rather than superior population control alone. A comparison that controls for structural quality, or applies similar fold-preservation criteria during candidate selection, would help clarify this issue.

Overall, these concerns do not undermine the relevance of the research direction, but they limit the extent to which the experiments establish the proposed guidance mechanism as the main source of the reported improvements.

Questions for authors

Q1. What are the Hamming distances between the final designed sequences and their corresponding native sequences? Can the authors provide a comparison with CB under matched mutation and BioEmu evaluation budgets? Q2. How does Ensemble-IF perform when the guidance strength is set to eta = 0? To what extent do the reported improvements arise from population-aware sequence guidance rather than BioEmu-based candidate screening? Q3. How well do BioEmu-predicted event populations correlate with occupancies measured from independent MD simulations? Are prediction errors systematically larger for extreme target populations? Q4. Can the authors provide statistical uncertainty estimates and convergence diagnostics for the MD-derived event populations, particularly in settings with substantial replica-to-replica variability?

4 comments

We thank the reviewer for their feedback and for recognizing the population-targeting objective and interpretable guidance rule. Please see our point-by-point responses below.

Weakness/Question 1: Mutation comparison and performance on shared budgets.

Regarding mutation counts, the table below gives the Hamming distance of each manuscript design from its native sequence. Ensemble-IF used unrestricted redesign, while the original CB baseline selected single mutants.

DomainTarget band (%)Ensemble-IF substitutions / lengthCB substitutions / length
1divA0231.89-37.8956 / 901 / 90
1divA0291.89-97.8944 / 901 / 90
1kwgA0332.56-38.5625 / 501 / 50
1kwgA0377.56-83.5631 / 501 / 50
2v0cA0396.00-100.0033 / 571 / 57
2v0cA0314.40-20.4033 / 571 / 57
Ubiquitin64.29-74.2942 / 761 / 76

Regarding matched-budget comparisons, our response to Reviewer N7W7, Weakness/Question 2, reports a three-substitution comparison with the same BioEmu evaluation budget across methods. Ensemble-IF has the lowest mean target-band error in 3/3 settings. On 1kwgA03, it reaches the band in 3/3 searches with 1.67 substitutions on average, compared with 2/3 searches and 2.33 substitutions for CB. Ensemble-IF's advantage therefore persists when unrestricted redesign is removed from the comparison.

Weakness/Question 2: η=0\mathbf{\eta=0} control and results with standard proposals.

Regarding standard proposals with identical BioEmu screening, the η=0\eta=0 control uses the same ensemble-conditioned ProteinMPNN, reference structures, candidate budget, and selection objective as Ensemble-IF, with the target-dependent logit correction disabled. This directly tests the standard-proposal alternative.

In our capped three-substitution study (Reviewer N7W7, Weakness/Question 2), guidance lowers mean band error from 33.87 to 8.84 pp on 1divA02 and from 38.83 to 0.00 pp on 1kwgA03, across three searches per method. Both methods remain outside the high-target 2v0cA03 band. Because the guided and unguided candidates receive identical screening, the gains on the first two targets show that changing the proposals improves targeting beyond what screening achieves with standard proposals.

We also tested fixed state-contrast proposals against η=0\eta=0 without a mutation cap and evaluated the selected sequences in MD. Our response to Reviewer YfU7, Weakness 1/Question 2, reports lower target-band error with fixed contrast in 5/6 settings, including a reduction from 52.28 to 5.89 pp on 1kwgA03. This provides further evidence that state-directed proposals can improve the resulting MD populations.

Weakness 3/Question 3/Question 4: BioEmu-MD occupancy correlation, statistical analysis.

Regarding BioEmu-MD agreement, the table below compares predictions and MD occupancies for every manuscript design and CB baseline. Across the 12 mdCATH sequence-target pairs, Pearson and Spearman correlations are 0.761 and 0.734, respectively. Ubiquitin is shown separately.

DomainTarget band (%)Ensemble-IF BioEmu (%)Ensemble-IF MD (%)CB BioEmu (%)CB MD (%)
1divA0231.89-37.8938.8938.5072.4263.50
1divA0291.89-97.8992.1998.9281.7880.83
1kwgA0332.56-38.5634.2934.5079.4441.58
1kwgA0377.56-83.5682.7679.1985.5135.00
2v0cA0396.00-100.0095.3197.2536.966.08
2v0cA0314.40-20.4014.812.3330.8058.25
Ubiquitin64.29-74.2969.8461.5848.7440.31

For the six Ensemble-IF mdCATH designs, mean absolute BioEmu-MD disagreement is 4.22 pp, exhibiting relatively high correlation. Examining other events from other proteins and domains is left for future work, as BioEmu's training data is limited in scope despite its large size. We encourage testing with superior oracles, e.g. BioEmu2.

Regarding extreme target populations, the pattern differs between cohorts. For the six manuscript designs, targets with midpoints ≤25% or ≥75% have mean disagreement of 6.18 pp, versus 0.30 pp for middle targets.

Regarding MD uncertainty, our response to Reviewer N7W7, Weakness 1/Question 1, reports occupancy intervals using 20 and 50 ns blocks and per-replica ESS. All three high-target Ensemble-IF occupancy intervals lie entirely above CB's at both block lengths, so the higher populations remain clear after accounting for temporal correlation. Low-target 1divA02 and 1kwgA03 have broader intervals and less precisely resolved populations, although occupancy superiority of Ensemble-IF is demonstrated. We also provide commentary regarding effective sample sizes per-trajectory.

Weakness 4: Impact of structural quality on designs.

Regarding structural quality and population targeting, CB retains about 88% of native contacts on both 1kwgA03 targets, compared with 83-86% for Ensemble-IF. However, the occupancy comparison under Weakness 3/Questions 3-4 shows that both Ensemble-IF means fall in their target bands, while both CB means fall outside. Better native-contact retention alone therefore does not account for Ensemble-IF's population-targeting advantage on these designs. To compare population targeting at similar fold quality, we examined the selected three-substitution 1divA02 designs targeting [31.89-37.89] with three 200 ns MD replicas per method. The table reports occupancy, distance of the mean occupancy to the target band, and fold retention.

MethodNative contacts (%)C$\alpha$ RMSD (Å)
η=0\eta=088.101.22
CB86.441.32
Ensemble-IF85.241.86

All three methods retain 85-88% of native contacts, yet Ensemble-IF has the lowest MD band error per our reply to reviewer N7W7 under Weakness/Question 2. This comparison addresses the structural-instability concern: the guided population is closer to the target even when both controls retain the fold. We clarify that all methods and baselines use the same fold-preservation criteria during design. Extending the same replicas to 400 ns preserves this ordering. Each method uses one fixed sequence and three MD replicas; the guided mean remains above the band, with substantial replica variability.

Regarding explicit fold-preservation criteria, we tested sequences using a structural filter applied to all methods' high-target 2v0cA03 candidate pools. Our response to Reviewer YfU7, Weakness 1/Weakness 2/Question 5, reports that filtering selected the native sequence in seven of nine pools. The MD comparison shows improved fold retention but a population farther from the target than the unfiltered guided design. This reinforces the need to assess the requested population alongside structural quality when selecting sequences.

Thanks for the detailed response and additional experiments. The matched mutation-budget comparisons and η = 0 ablations address my main concerns about the baseline comparison and the contribution of population-aware guidance. I still have some reservations about the limited number of systems and the reliability of BioEmu predictions, but the new results provide stronger support for the method. I will keep my rating at 3.

Thank you for following up, and we are glad the matched-budget comparisons and η=0\eta=0 ablations helped address your main concerns. We agree that broader system coverage and BioEmu prediction reliability are important limitations, and we believe scaling multi-replica, multi-target MD validation across a wider set of systems and reducing oracle error are natural priorities for future work. We have incorporated clearer wording of these limitations into the manuscript. Thank you again for your thoughtful review and for the constructive engagement.

Reviewer_N7W7OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Ensemble-IF designs sequence mutations toward target conformational populations by reweighting ensembles, combining ProteinMPNN logits, and screening candidates with BioEmu. The paper validates selected designs with molecular dynamics on three mdCATH domains and a ubiquitin case study.

Critical strengths
  1. I like the distinction between supporting multiple conformations and controlling their relative populations. The target-band formulation and weighted-logit proposal rule provide a clear way to address this more specific design objective.

  2. The evaluation uses molecular dynamics beyond the BioEmu screening oracle and checks native-contact recovery. Table 1 shows lower target-band error than Conformational Biasing in all six mdCATH settings, with better contact recovery in four.

Critical weaknesses
  1. The precision of the population-control claim is not yet clear: only three of six mdCATH means fall inside their target bands, and Figure 3 shows substantial replica variability. Confidence intervals and convergence diagnostics for the designed sequences are needed to assess whether the narrow bands are reliably achieved.

  2. The benefit of population reweighting is not isolated from candidate search and BioEmu screening. Appendix E adds multi-round CB, but there is no matched-budget η = 0 control, and Ensemble-IF proposals have no mutation cap.

Questions for authors
  1. Are the reported target-band successes stable across independent MD starts and after accounting for trajectory autocorrelation?

  2. How does η = 0 perform with the same proposal, mutation, and BioEmu evaluation budgets?

2 comments

We thank the reviewer for their feedback and for highlighting our distinction between structural compatibility and population targeting. Please see a point-by-point response to each weakness/question below.

Weakness 1/Question 1: Stability across MD starts and reliability.

Regarding robustness to MD starts, we kept each design fixed and generated new ESMFold2 structures per original replica. We reused the production seeds and compared matched 200 ns windows after fresh preparation and equilibration. We use 200 ns due to the time intensiveness of running MD given that, for occupancy, most methods' trends stabilized within this time frame. The table below includes the mean results for these new structures and their original counterparts across both methods and six targets.

DomainTarget band (%)Ensemble-IF, original start (%)Ensemble-IF, new starts (%)CB, original start (%)CB, new starts (%)
1divA0231.89-37.8933.6744.3364.8391.50
1divA0291.89-97.8998.5099.5094.5077.67
1kwgA0332.56-38.5629.6722.6759.5068.50
1kwgA0377.56-83.5678.8873.1343.6742.67
2v0cA0396.00-100.0097.5099.172.8323.33
2v0cA0314.40-20.402.674.1744.5025.83

Across the six settings, Ensemble-IF's new-start mean is closer to the target band than CB's in 5/6 comparisons. The exception is low-target 2v0cA03, where CB has 5.43 pp error versus 10.23 pp for Ensemble-IF and both means remain outside the band. Regarding consistent attainment of a narrow target band, the clearest result is 2v0cA03 at 96-100%. All three new ESMFold2-start occupancies remain in band (100%, 98.5%, and 99%). Across these main benchmark systems, Ensemble-IF appears stable across structure initializations.

Regarding MD sampling uncertainty, we also quantified the uncertainty within our reported 400 ns trajectories. Using N=400N=400 frames saved at 1 ns intervals, we computed effective sample sizes (ESS) separately for each binary event via

\widehat\tau = \max\left(1,\;1+2\sum_{k=1}^{K}\widehat\rho(k)\right).$$ Here $\widehat\rho(k)$ is the mean-centered sample autocorrelation at lag $k$. We used Geyer's initial-positive-sequence rule, retaining adjacent lag pairs starting at lag zero until the first nonpositive pair. $K$ is the last retained lag. The lower bound on $\widehat\tau$ caps ESS at $N$. To estimate occupancy uncertainty, we resampled replicas and contiguous 20 ns time blocks. The table below reports the original 400 ns occupancies with 20 ns-block 95% intervals and **per-replica ESS for all six settings and both methods**. | Domain | Target band (%) | Ensemble-IF occ. (%) | Ensemble-IF 20 ns CI | Ensemble-IF 50 ns CI | CB occ. (%) | CB 20 ns CI | CB 50 ns CI | Ensemble-IF ESS per replica | CB ESS per replica | |---|---|---|---|---|---|---|---|---|---| |1divA02|31.89-37.89| 38.50| [10.66, 72.17]|[8.33, 70.92]|63.50|[15.41, 91.25] | [14.83, 91.58] | 14.2, 32.9, 18.5 | 68.5, 79.0, 11.1 | |1divA02|91.89-97.89| 98.92| [97.75, 99.67]| [97.92, 99.67] | 80.83 | [67.25, 94.84] | [66.00, 95.25] | -, -, 256.5 | 14.6, 7.1, 219.7 | |1kwgA03|32.56-38.56| 34.50| [19.50, 51.92] | [18.25, 53.42] | 41.58 | [19.67, 62.59] | [19.00, 63.75] | 22.7, 7.1, 12.1 | 5.4, 57.9, 6.9 | |1kwgA03 |77.56-83.56| 79.19| [72.19, 84.69] | [71.75, 85.13] | 35.00 | [18.91, 55.42] | [17.50, 56.25] | 152.0, 143.7, 31.4, 34.6 | 19.6, 17.1, 98.0 | |2v0cA03|96.00-100.00| 97.25 | [95.67, 98.50] | [95.75, 98.50] | 6.08 | [0.00, 14.17] | [0.00, 13.67] | 240.5, 202.8, 107.4 | 85.8, 102.8, - | |2v0cA03|14.40-20.40| 2.33 | [0.83, 4.42] | [0.83, 4.50] | 58.25 | [28.16, 85.25] | [27.08, 85.75] | 102.0, 400.0, -| 130.9, 17.5, 4.4 | ESS entries follow replica order. A dash indicates a constant event trace, whose normalized autocorrelation is undefined, or fewer than ten event changes, for which we withhold ESS because sparse switching makes the estimate unreliable. For targets near saturation, persistent occupancy can be the desired outcome. The two high-target 1divA02 replicas with unreported ESS maintain 99.5% and 99.25% occupancy, with four and six event changes, favoring the requested state while lying slightly above the 91.89–97.89% band. The high-target population shifts remain clear after accounting for temporal correlation: all three Ensemble-IF occupancy intervals lie entirely above CB’s. High-target 2v0cA03 combines 97.25% occupancy with 107-241 effective samples per replica and integrated event correlation times of 1.7-3.7 ns, supporting precise occupancy estimates within these trajectories. Its interval barely changes with 50 ns blocks, from 95.67-98.50% to 95.75-98.50%. Low-target 1divA02 and 1kwgA03 have broader intervals and less precisely resolved populations, with the latter’s 50 ns-block interval of 18.25-53.42%.

Weakness/Question 2: η=0\mathbf{\eta=0} control and mutational budget results.

Regarding the effect of guidance under matched budgets, we compared Ensemble-IF, η=0\eta=0, and CB with a maximum of 3 substitutions from the native sequence, keeping all other sampling hyperparameters constant. The η=0\eta=0 control uses the same ensemble-conditioned ProteinMPNN proposals and BioEmu screening as Ensemble-IF, with the target-dependent logit correction disabled. The table below reports mean target-band error ± SD across three searches, measured in percentage points (pp). We report results across 3 bands for our main mdCATH benchmark due to compute constraints.

DomainTarget band (%)Ensemble-IF error (pp)η=0\eta=0 error (pp)CB error (pp)In-band searches, Ensemble-IF / η=0\eta=0 / CBMean substitutions, same order
1divA0231.89-37.898.84 ± 11.4633.87 ± 4.1633.56 ± 3.081/3 / 0/3 / 0/33.00 / 0.67 / 1.33
1kwgA0332.56-38.560.00 ± 0.0038.83 ± 6.471.02 ± 1.763/3 / 0/3 / 2/31.67 / 2.67 / 2.33
2v0cA0396.00-100.0046.75 ± 5.6647.91 ± 4.9851.79 ± 2.210/3 / 0/3 / 0/32.33 / 1.33 / 1.33

Ensemble-IF has the lowest mean error in 3/3 settings, with the largest gains on 1divA02 and 1kwgA03. On 1kwgA03, all three guided searches reach the band, compared with none for η=0\eta=0 and two for CB. Ensemble-IF also uses fewer substitutions than CB on this target, averaging 1.67 versus 2.33. These results show improved targeting under the same mutation ceiling and BioEmu evaluation budget, validating our approach in Ensemble-IF.

Comments on the paper

2 comments

Thank you to the reviewers for your careful reading and constructive feedback. We have revised the manuscript to correct 2 central presentation concerns with edits highlighted in blue.

  1. We have corrected an issue where Figure 4 had no caption.
  2. We have corrected an issue pointed out by reviewer YfU7 where the CB baseline for Ubiquitin in Figure 4 had 4 replicas instead of 3. We have re-rendered Figure 4 with the appropriate replicas, excluding the trajectory result from the non-identical seed. The table in our reply to kgi4 under Weakness 3/Question 3/Question 4 also reflects this now.

We appreciate the reviewers’ feedback helping us clarify the paper and examine the method’s strengths and limitations. We hope these revisions and our responses address their concerns, and we welcome any further questions and score revisions in light of our rebuttal.

AuthorsOpenReview ↗

Follow-up on Rebuttal Responses

Dear Reviewers,

Thank you again for your time and valuable feedback on our work. We would love to hear your thoughts on our rebuttal responses, which we believe address the key concerns brought up during the review process.

We greatly appreciate any further discussion and would be happy to clarify any unresolved questions. If our responses have addressed your concerns, we would also be grateful if you would consider increasing your scores accordingly.

Thank you again for your valuable time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.