acceptodds
Under review as a conference paper at ICLR 2027

Learning and Steering Protein Flexibility for Generative Protein Design

Abstract

Protein flexibility is a fundamental determinant of molecular recognition, allostery, and catalysis, yet predicting and engineering it at scale remains an open challenge. Existing approaches either rely on computationally prohibitive molecular dynamics simulations, resort to physics-based approximations that lack differentiability and miss sequence-dependent effects, or learn representations that are often tied to specific simulation protocols and can struggle to generalize across protein families and conditions. We introduce FlexFlow, a framework for designing proteins with prescribed per-residue root mean square fluctuation (RMSF) profiles via inference-time guidance on a frozen pretrained sequence–structure co-design model. At its core, FlexFlow is powered by FlexFormer, a differentiable predictor of per-residue backbone flexibility that achieves state-of-the-art accuracy and, unlike predictors that read only coordinates, responds to individual amino acid substitutions, tracking their measured effect on flexibility more faithfully than existing methods. Rather than retraining a conditional generative model, which may distort the structural prior learned from experimentally solved structures and sacrifices designability for flexibility, FlexFlow steers a pretrained joint sequence–structure co-design model at inference time using FlexFormer gradients, without any task-specific retraining. Each backbone is steered directly toward the target flexibility profile during generation, eliminating the post-hoc screening that existing methods require, yielding significant reduction in wall-clock time at substantially better structural quality.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein design, inference-time guidance, flow matching, protein flexibility, molecular dynamics, geometric deep learning

Reviews

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

This paper proposes a method for predicting protein structures by modeling their flexibilities. The main framework contains two components: FlexFormer and FlexFlow, where FlexFormer predicts the flexibility profile that also considers sequence-level mutations, and FlexFlow uses this profile to steer the starting structures via inference-time guidance. The method shows good performance in flexibility profile (RMSF) prediction and better structure quality.

Critical strengths
  1. The problem this paper targets is important: modeling the protein's dynamic flexibility and using this to guide protein generation. The two-stage framework is well explained and well motivated, and considering the inclusion of sequence mutation solves some real limitations of current methods
  2. Experiments are comprehensive and correspond well with the methods and initial motivation.
Critical weaknesses
  1. Using only the Person correlation coefficient as the objective in inference-time guidance may bring some problems because it's scale-insensitive. RMSF profiles with different scales can still have r near 1, which may make FlexFlow generate proteins that move too little or too much.
  2. It lacks analysis of what physical states of proteins this model generates. The large sc-RMSD in Table 3 is very large. It can be because either the structure itself is bad or it's just not designable, where the latter can be because the generated protein is not an energy-stable state. Though it's reasonable in the current flexible protein modeling setting, it would be good to clarify what proteins this model generates.
Questions for authors
  1. Following the first pointed weakness, have the authors tried other objectives in inference-time guidance, like direct MSE loss for RMSF profiles?
  2. Have you done any energy analysis of the generated proteins? I am curious if they are in the stable/metastable state or the transition state.
  3. It would be good to clarify the equivariance/invariance of the proposed methods. Based on my understanding, the FlexFormer should be invariant in predicting RMSF profiles, but the FlexFlow should be equivariant. Both are not fully discussed in the current paper.
3 comments

We sincerely thank the reviewer for the thoughtful and constructive evaluation of our work and for recognizing the importance of flexibility-guided protein design. These comments motivated additional experiments and helped clarify the scope of FlexFlow.

Comment 1: Pearson versus MSE guidance

We agree that Pearson correlation is insensitive to positive scaling and constant offsets:

Corr(ξ̂, aξ + b) = Corr(ξ̂, ξ), for a > 0.**

Our current guidance objective is designed to control the relative distribution of flexibility across residues, particularly the locations of flexible and rigid regions, rather than independently controlling absolute RMSF magnitude.

To directly investigate the reviewer's suggestion, we conducted a 24-target factorial ablation using the same frozen MultiFlow generator, independently varying predictor (BackFlip/HierEGT), guidance location (intermediate/predicted clean), and objective (MSE/Pearson). Strengths were calibrated on a separate validation set; generated flexibility was evaluated with GNM, independent of the learned predictors.

PredictorGuidance locationGNM Δr (Pearson − MSE)
BackFlipIntermediate+0.283
BackFlipPredicted clean+0.223
HierEGTIntermediate+0.194
HierEGTPredicted clean+0.390

All four comparisons have p ≤ 0.0025. Particularly noteworthy is the original BackFlip predictor at intermediate coordinates: changing only the objective yields +0.283. Over a further 32-fold strength sweep, MSE reached at most GNM r = 0.467, versus r = 0.663 for Pearson. These results support Pearson guidance for relative-profile agreement; absolute RMSF control would require an additional amplitude-sensitive term.

Comment 2: Structural self-consistency and physical states

FlexFlow generates candidate backbones; ProteinMPNN assigns sequences and ESMFold refolding assesses sequence–structure self-consistency. The large scRMSD values indicate challenges in recovering the generated backbone. scRMSD alone cannot distinguish geometric inconsistency, sequence incompatibility, and thermodynamic instability.

We therefore additionally examined threshold-based self-consistency and local geometry. At L = 300, 21/100 FlexFlow designs achieve scRMSD ≤ 5 Å, versus 0/100 for FliPS; under the stricter ≤ 2 Å criterion the counts are 8/100 versus 0/100. These results improve long-design self-consistency, although designability remains challenging. FliPS retains better local backbone geometry, so the methods exhibit different structural-quality trade-offs.

In a separate geometry-controlled analysis, reducing λ from 400 toward 200 maintained comparable independent GNM agreement while reducing Cα bond-length distortion and improving geometry-valid yield by approximately 1.5-fold.

We have not performed the free-energy landscape analysis required to distinguish stable, metastable, and transition states. Existing finite-time all-atom MD characterizes flexibility of selected candidates but does not establish thermodynamic-state identity. We will make this limitation explicit.

Comment 3: FlexFormer invariance and FlexFlow equivariance

FlexFormer predicts scalar per-residue RMSF and is designed to be invariant to global rigid-body transformations. Writing backbone coordinates as X, sequence as S, rotation as R ∈ SO(3), and translation as t:

f_FlexFormer(RX + t, S) = f_FlexFormer(X, S).

HierEGT uses equivariant geometric representations to produce invariant scalar predictions; Stage 2 adds ProtT5 sequence features independent of the global coordinate frame.

FlexFlow applies a scalar loss to predicted-clean coordinates:

ℒ(X, ξ*) = −Corr(f_HierEGT(X), ξ*)

Because this predictor output and loss are invariant, the coordinate gradient transforms equivariantly:

∇_{RX+t} ℒ = R ∇_X ℒ

Combined with the equivariant MultiFlow update and its coordinate-centering convention, this preserves the symmetry of the guidance formulation. Guidance changes backbone translations only, leaving the generator's rotational and sequence updates unchanged. We will clarify these mathematical properties in Methods.

We sincerely thank the reviewer again for the valuable suggestions. The new experiments strengthen the rationale for the Pearson objective and clarify the structural and geometric claims, while delineating the limitations of the present method.

We sincerely thank the reviewer again for the thoughtful and constructive evaluation. We have provided additional MSE-versus-Pearson guidance experiments and clarified the physical interpretation and symmetry properties of our framework. We would greatly appreciate any further feedback on whether our responses address your questions, and would be happy to provide additional clarification if helpful. Thank you so much again!

Thank you for the additional experiments and clarification of the limitations. Please incorporate these explanations and clarify the scope of the claims in the main text. Overall, I maintain my score of 3.

Reviewer_x6urOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies protein generation conditioned on prescribed per-residue flexibility profiles. It introduces FlexFormer, a two-stage RMSF predictor combining a backbone geometry encoder, HierEGT, with a pretrained protein language model, and FlexFlow, which uses HierEGT gradients to guide a frozen MultiFlow generator at inference time. Guidance modifies predicted clean translations using a negative Pearson correlation objective. Generated backbones are subsequently redesigned with ProteinMPNN.

The authors evaluate flexibility prediction on ATLAS and mdCATH, compare conditional generation against FliPS and unguided MultiFlow, and validate 40 generated designs using three independent 100 ns molecular dynamics simulations per design. They also introduce a fixed-backbone substitution-response metric intended to assess sensitivity to amino acid substitutions.

Critical strengths
  1. The problem is practically relevant, and the predictor evaluation goes beyond correlation. Controlling dynamic properties is a meaningful extension of protein design. Reporting calibration slopes, overall MAE, and errors in mobile regions helps identify limitations that correlation alone would conceal. FlexFormer improves several reported metrics across both benchmarks.

  2. The study includes substantial physics-based validation. The 40-design MD evaluation provides evidence beyond learned predictor scores. Showing the individual profiles, including the weaker results on hand-drawn targets, is helpful for assessing the capabilities and limitations of the complete design pipeline.

Critical weaknesses
  1. The methodological contribution and the validity of the main baseline comparison are not sufficiently established. The manuscript largely contrasts inference-time guidance with FliPS’s conditional training. However, the FliPS paper already introduces training-free BackFlip Guidance in Section 4.1 and evaluates guidance schedules and scaling strategies in Appendix A.13. Consequently, using a differentiable flexibility predictor to guide a frozen generator is not, by itself, an established novel contribution here. Applying guidance to predicted clean structures, using HierEGT, and adopting a different objective may provide meaningful improvements, but their individual contributions are not isolated. A direct comparison using the same generator, with BackFlip versus HierEGT and matched tuning budgets, is particularly important. There is also a potentially consequential mismatch in the flexibility variable. This manuscript uses globally aligned RMSF, whereas the original FliPS model is conditioned on locally aligned RMSF. These quantities describe different aspects of motion and are not interchangeable merely because they share physical units. The manuscript does not adequately explain whether FliPS was retrained for global RMSF or how the conditioning mismatch was handled. Retraining a BackFlip predictor for Table 1 would not, by itself, resolve the conditioning semantics of the FliPS generator. Until this issue is clarified, the main generation comparison cannot be confidently interpreted as evidence of methodological superiority.

  2. The evaluations do not adequately support the claims of mutation-effect prediction and prescribed flexibility control. Table 2 compares substitution-dependent predictions against an aggregate amino acid flexibility propensity estimated from training data, rather than against measured changes for the corresponding mutants. A context-independent predictor that simply returns this propensity for each substituted residue could achieve perfect correlation under the proposed metric while having no ability to predict position-specific mutation effects. The experiment therefore measures agreement with population-level residue propensities, but does not establish accurate prediction of actual mutational effects. The generation objective creates a related mismatch between the task and the evidence. Negative Pearson correlation is invariant to positive rescaling and constant offsets of the target profile. It therefore cannot distinguish targets with the same shape but different absolute flexibility magnitudes. This is especially consequential because the manuscript explicitly argues that amplitude calibration is essential for design. The results currently support control of relative flexibility patterns more directly than realization of prescribed absolute RMSF profiles. Finally, the MD evaluation contains no matched unguided or competing-method controls. Many natural-target profiles exhibit elevated terminal flexibility, which could contribute substantially to whole-chain correlations. Without terminal-excluded analyses, absolute-error measurements, and matched controls, the reported correlations do not establish how much physical improvement is attributable to guidance. The criteria used to select the 40 validated designs also need clarification.

Questions for authors
  1. Which FliPS checkpoint was used for generation? Was it retrained to condition on globally aligned RMSF? If not, how were the differences between local and global RMSF, including their numerical ranges, handled?
  2. What is the decisive technical improvement over the previously published BackFlip Guidance strategy? Can the authors provide a comparison on the same MultiFlow generator that separates predictor choice from guidance placement and objective choice?
  3. What score does a residue-identity-only predictor based on the training-set propensity vector achieve in Table 2? Is there paired wild-type/mutant MD or experimental evidence supporting the stronger mutation-effect claims?
  4. With the same random seed, how do generated outputs change when the target profile is multiplied by two or shifted by a constant? How should the claimed magnitude-control capability be interpreted under a Pearson-only objective?
  5. How were the 40 MD designs selected? Please report terminal-excluded correlations and absolute RMSF errors, and clarify whether matched unguided controls are available.
  6. Please reconcile the sequence-selection criteria: the main text and MD protocol refer to selection by pLDDT, whereas Appendix B describes selection by S2. Please also clarify in what sense the Stage-2 evaluator is independent when it incorporates the Stage-1 predictor used for guidance.
5 comments

We sincerely thank the reviewer for the detailed and technically insightful assessment. We particularly appreciate the distinction between inference-time guidance itself and the specific contributions of our formulation. Motivated by these comments, we conducted controlled experiments to isolate the components and revisited the released FliPS conditioning protocol.

Critical Weakness 1 / Q2: Distinction from BackFlip Guidance

We agree that training-free flexibility guidance was previously introduced by FliPS through BackFlip Guidance (BG). Our contribution is not the general idea of inference-time guidance, but a different steering formulation: FlexFlow evaluates flexibility on the predicted-clean backbone and optimizes negative Pearson correlation; BG uses MSE on intermediate coordinates.

We performed a 24-target factorial ablation on the same frozen MultiFlow generator, varying predictor (BackFlip or HierEGT), guidance location (intermediate or predicted clean), and objective (MSE or negative Pearson). Guidance strengths were calibrated on a separate validation cohort using a geometry-based criterion. We evaluated the outputs using GNM, which shares neither guidance predictor's learned architecture.

PredictorGuidance locationGNM Δr\Delta r (Pearson − MSE)
BackFlipIntermediate+0.283
BackFlipPredicted clean+0.223
HierEGTIntermediate+0.194
HierEGTPredicted clean+0.390

All comparisons have p≤0.0025p\leq0.0025. Most importantly, retaining BG's original predictor and intermediate-coordinate guidance while changing only the objective produces +0.283, isolating the contribution without changing the generator, predictor, or placement. An additional multi-seed analysis on 24 targets confirms +0.270 on full chains and +0.205 after terminal masking. Changing the predictor alone did not significantly improve GNM correlation (p=0.966p=0.966). Predicted-clean guidance is a distinguishing implementation choice, but its independent benefit was not statistically established.

Larger matched-generator validation

We compared the full FlexFlow and BG guidance formulations on 100 score-blind targets, again using the same frozen MultiFlow, matched initial conditions, and independently calibrated strengths:

MethodFull-chain GNM rrInterior GNM rr
BG0.3830.064
FlexFlow guidance0.6110.176
Paired improvement+0.228+0.112
95% CI[0.189, 0.266][0.069, 0.153]
Targets favoring FlexFlow87/10078/100

Interior evaluation removes 10% of residues at each terminus. A separate 32-fold guidance-strength study found a maximum tested GNM correlation of 0.467 for MSE versus 0.663 for Pearson. These findings identify the correlation objective as the principal improvement for relative flexibility-pattern steering, without claiming that clean-coordinate placement alone is decisive.

Q1: FliPS checkpoint and conditioning

We verified that the released FliPS generator checkpoint flips_final.ckpt was trained using locally aligned local_flex and was not retrained for globally aligned RMSF. Its conditioning interface clamps scalar values at 3 Å, representing them with a continuous feature and six bins. On 100 reference targets, 12.16% of globally aligned RMSF values exceed this threshold, versus 2.71% of locally aligned values. Local and global RMSF are therefore not interchangeable.

Accordingly, we interpret the original full-pipeline comparison as a comparison against the available released FliPS configuration, not an isolated test of guidance superiority under identical conditioning semantics. We will clarify the checkpoint provenance and exact inference-time preprocessing. Crucially, the matched-generator ablations do not depend on comparing different conditional generators; they directly assess the guidance formulations on frozen MultiFlow.

We thank the reviewer for the careful questions regarding the interpretation of fidelity, the Pearson objective, and whether FlexFlow can respond to changes in prescribed interior flexibility. These distinctions are important for precisely characterizing what FlexFormer predicts and what FlexFlow controls.

Critical Weakness 2 / Q3: Fidelity and mutation effects

For each position i, we substitute each of the 20 amino acids while holding backbone coordinates fixed. We obtain the predicted RMSF at that position using the same FlexFormer prediction head; no separate mutation-effect head is used. Fidelity measures Pearson agreement between these substitution-dependent predictions and amino-acid flexibility propensities estimated from the training data.

As the reviewer correctly points out, a residue-identity-only predictor returning the propensity vector would achieve fidelity = 1.0 by construction. We therefore clarify that manuscript Table 2 measures agreement with residue-level flexibility propensities, rather than accuracy on experimentally measured context-specific mutation effects. We do not currently have paired wild-type/mutant MD or experimental measurements establishing the stronger claim. We will revise the terminology and interpretation accordingly.

Q4: Target scaling and absolute RMSF control

We appreciate the reviewer's precise mathematical observation. Pearson correlation is invariant to positive affine transformations of the target:

Corr(predicted RMSF, a × target RMSF + b) = Corr(predicted RMSF, target RMSF), for a > 0.

We directly tested this property across four protein lengths and three random seeds with frozen checkpoints and matched initial conditions. For the transformation target → 2 × target, all 12 generated backbones were bit-identical, confirming that the objective does not encode absolute target scale.

For target → target + 1, final coordinates could differ substantially (median maximum Cα displacement: 9.70 Å) despite the mathematical invariance. At fixed coordinates, however, the objective differed by only 1.1 × 10⁻⁸, and the relative guidance-gradient difference was 4.4 × 10⁻⁷. This is consistent with amplification of finite-precision differences during iterative sampling, not meaningful control of absolute RMSF magnitude.

Thus, FlexFlow steers relative flexibility patterns rather than independently prescribing absolute RMSF amplitudes. The manuscript's amplitude-calibration results characterize FlexFormer as a predictor, whereas FlexFlow's Pearson-only generator objective is designed for relative-profile control. We will explicitly distinguish these claims.

Counterfactual interior steering

To investigate relative flexibility control independently of mobile termini, we conducted a further experiment with 100 pairs of prescribed profiles and three generation seeds. Profiles within each pair had identical value distributions and flat termini, but their main peaks occurred at different interior positions. The frozen MultiFlow checkpoint, initial noise, and sampling configuration were held fixed, changing only the target profile.

Using independent GNM evaluation, we measured how much better each generated structure matches its own target than the alternative. FlexFlow achieved a target-discrimination score D = 0.800 (95% CI [0.774, 0.826]), versus D = 0.015 for calibrated BG on the same generator and D = 0 for unguided MultiFlow. All 100 profile pairs showed positive discrimination.

This provides direct evidence of target-dependent steering of interior flexibility locations, rather than recovering a fixed pattern of naturally flexible termini. We interpret this as control under an independent structural flexibility proxy, not as a demonstration of absolute RMSF magnitude control or a guarantee of physically valid structures.

We appreciate the reviewer's request for more detailed physical evaluation and clarification of the original MD-selection protocol. We revisited the simulations and retained all 40 original designs in the new analysis.

Q5: MD selection, terminal masking, and absolute RMSF error

The original 40 MD designs were computationally prioritized using S2 agreement, then independently evaluated with all-atom MD. To quantify terminal contributions, we recomputed Pearson correlation after removing 10% of residues from each terminus, together with full-chain RMSF MAE:

DatasetnFull rInterior rFull MAE (Å)
ATLAS100.8190.0341.241
mdCATH100.8990.2520.943
Hand-drawn200.5090.4430.714

The terminal-masking analysis reveals different behavior across target classes. Whole-chain agreement for natural ATLAS and mdCATH profiles is strongly influenced by their flexible terminal regions. In contrast, the hand-drawn cohort, which includes prescribed interior flexibility patterns, retains more agreement after masking (r = 0.509 full-chain versus r = 0.443 interior). These results highlight the importance of distinguishing whole-chain agreement from control of interior flexibility.

The original simulations provide physical RMSF measurements for S2-prioritized candidates, not an unbiased estimate over all generated designs. They did not include matched unguided or FliPS MD controls. We have initiated a separate score-blind, baseline-controlled MD evaluation to investigate comparative physical performance; we do not present incomplete controls as completed validation. Our new GNM counterfactual experiment (Part 2) separately tests target-dependent interior steering under a structural proxy.

Q6: Sequence selection and S2 independence

We thank the reviewer for identifying the inconsistent sequence-selection descriptions. We verified the original selection against all 40 simulated structures: designs were prioritized by S2 agreement, and the highest-S2 ProteinMPNN sequence was selected within each retained backbone. Highest pLDDT was not the MD sequence-selection rule. Appendix B correctly describes S2 selection; the MD protocol's reference to pLDDT was a documentation error, which we will correct.

The selection procedure prioritized promising candidates for computationally expensive physical evaluation, whereas the subsequent flexibility measurements were obtained independently through all-atom MD. This distinction is important: the simulation measurement itself is independent of S2, but the set of simulated designs is S2-selected.

We also agree that Stage 2 is not fully independent of Stage 1, since Stage 2 incorporates Stage-1 predictions and additional ProtT5 sequence representations. Our new guidance experiments use GNM, and the expanded 400-target de novo evaluation uses ANM. Neither structural evaluator shares the learned FlexFormer architecture; they provide complementary assessments beyond S2, while all-atom MD measures finite-time physical flexibility.

Finally, the original full-system comparison involves differences in candidate screening and sequence-selection conventions between FlexFlow and FliPS. We therefore do not attribute the full-pipeline differences solely to guidance. The matched-generator factorial and 100-target comparisons in Part 1 directly isolate the contribution of the guidance formulation under common generation and calibration conditions.

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

The paper proposes FlexFlow to design protein sequences according to prescribed per-residue flexibility profile. A flexibility predictor, FlexFormer, which integrates an equivariant HierEGT and Prot5 based protein language model. The two-staged model combines both coordinate and evolutionary information, showing promising benchmark results in flexibility prediction. The flexibility guided gradients of the translations of residues are used to guide the generation process of MultiFlow, enabling a training-free guidance without disturbing the learned structure prior by the flow matching model. Flexibility prediction accuracy, correlation of flexibility of designed sequence and the target, quality of structure, and runtime are compared across existing baselines. Molecular dynamics over 40 generated backbones verify the effectiveness of the guided design.

Critical strengths
  • The proposed method is based on a training-free guidance. Thus, can be easily combined with many existing de novo design methods based on diffusion or flow-matching.

  • The paper provides thorough benchmark studies over metrics covering flexibility prediction accuracy and flexibility consistency of designed sequence with the target profile over existing baselines. Experiments show a good trade-off between the steering effect and structure quality preservation.

  • The paper is well written overall with clear idea.

Critical weaknesses
  • The advantage of FlexFlow is not clear enough. For example, FlexFlow shows better S2 correlation than FliPS in mdCATH, but is less favorable for ATLAS and OOD random profiles. From Table 3, it seems that FlexFlow achieves a trade-off between structure quality and flexibility correlation. However, the trade-off should heavily rely on the strength of guidance for a training-free guided generation method, which is not discussed in the paper. For the 40 cases validated by MD, FlexFlow performs good on those profiles with flexible loops in its ends and with higher rigidity in the middle of the sequence, which is a highly consistent pattern. But for profiles with high flexibility at random positions, the performance downgraded strongly, as can be seen in Figure 7 from the appendix. The same performance gap can be seen from the S2 Pearson in Figure 2 in the main text.

  • Training-free guidance is already proposed and discussed by the FliPS paper, which explicitly declaring that strong guidance may hurt the structure quality. The difference of FlexFlow with the training-free guidance in FliPS should be carefully discussed.

  • The effectiveness of the design with prescribed flexibility is mainly demonstrated by S2 Pearson, which is a T5-based RMSF evaluator, instead of flexibility from MD. The ProT5 RMSF evaluator is from Flexpert whose architecture shared with the FlexFormer. It is unclear whether this sharing of model backbone can contribute to the S2 Pearson due to similar flexibility prediction pattern from the model architecture. The demonstrated results of 40 MD cases lack comparison with baselines.

  • The value of the proposed fidelity metric is unclear. The definition of fidelity is not clear enough. For example, it is unclear that how prediction at a position edited to a vav_a is calculated or it is predicted with a separate head. It is also unclear from the argument about how higher fidelity contributes to the quality of design.

Questions for authors
  • Why the guidance is only applied from RMSF prediction from stage 1 and on translation only? What’s the gap of accuracy between stage 1 and stage 2 RMSF prediction?

  • What’s the effect of strength of steering on the results?

  • Other questions mentioned in the weakness part.

5 comments

We sincerely thank the reviewer for carefully reading our paper and providing thoughtful, constructive feedback. We particularly appreciate the questions regarding the flexibility–structure trade-off and the feasibility of prescribed flexibility profiles. We conducted additional experiments to investigate both aspects.

Question 1: Guidance strength

We evaluated 12 guidance strengths on 139 ATLAS proteins. Stronger guidance improves flexibility agreement but reduces structural quality and designability (selected strengths shown):

λ\lambdaS2 Pearson ↑pLDDT ↑Designable (%) ↑
200.46680.616.6
2000.64477.312.2
400 (submitted)0.70373.47.9
6000.71969.65.0
10000.74864.71.4

For example, increasing λ\lambda from 400 to 1000 improves S2 correlation from 0.703 to 0.748, but decreases designability (scRMSD ≤2\leq 2 Å) from 7.9% to 1.4%. A separate geometry-controlled analysis shows that λ≈200\lambda\approx200 can achieve comparable GNM agreement with substantially better backbone geometry. We will clarify this trade-off rather than presenting λ=400\lambda=400 as universally optimal.

OOD profiles and interior flexibility

We agree that designing proteins with arbitrary flexibility profiles is challenging: randomly generated targets are not necessarily compatible with a stable, foldable structure. We included these profiles to evaluate FlexFlow beyond naturally occurring patterns, while recognizing that not every prescribed profile is physically realizable.

In Figure 2, FlexFlow achieves higher S2 correlation than FliPS for shorter proteins (L=70,100L=70,100), while FliPS performs better for L=200,300L=200,300. However, at L=300L=300, FlexFlow achieves better sequence-refolding self-consistency (scRMSD 11.49 vs. 19.12 Å) and folding confidence (pLDDT 67.6 vs. 58.8), despite lower flexibility correlation (0.47 vs. 0.59). This illustrates why flexibility and structural quality must be evaluated together.

To investigate control at unusual interior positions, we constructed 100 pairs of synthetic RMSF profiles with three generation seeds. Each pair has identical flexibility distributions and flat termini, but peaks at different interior locations (e.g., 30% vs. 70%). We held the frozen MultiFlow checkpoint, initial noise, and seed fixed, changing only the target. We compared against calibrated BackFlip Guidance (BG) applied to the same frozen MultiFlow, not the separately trained FliPS conditional generator.

We define discrimination DD as the average advantage of each generated structure's agreement with its own target over the paired alternative. Flexibility was assessed using GNM, a structural-contact-based model independent of the learned predictors.

MethodGNM discrimination DD ↑
Unguided MultiFlow0.000
Calibrated BG on MultiFlow0.015
FlexFlow0.800

FlexFlow achieves D=0.800D=0.800 (95% CI [0.774, 0.826]), positive for all 100 pairs. This demonstrates responsiveness to prescribed interior flexibility locations, rather than simply reproducing flexible termini. Nevertheless, relocation is partial, and only 37% of designs at λ=400\lambda=400 satisfy the geometry-validity criterion.

We see region-specific flexibility design, for example at enzyme loops or antibody CDRs with a preserved scaffold, as a promising future extension rather than an application validated here. We acknowledge that the complete FliPS pipeline achieves higher interior agreement on several existing benchmarks; our counterfactual study specifically establishes target-responsive steering under an independent structural proxy.

We thank the reviewer for highlighting the importance of distinguishing our contribution from prior guidance and of independently evaluating flexibility and structural quality.

Question 2: Relation to BackFlip Guidance

We agree that FliPS previously introduced training-free BackFlip Guidance (BG). FlexFlow differs in evaluating flexibility on MultiFlow's predicted-clean backbone rather than intermediate noisy coordinates and optimizing negative Pearson correlation rather than RMSF MSE.

To compare the formulations directly, we applied both to the same frozen MultiFlow generator, with matched targets and sampling budgets and strengths calibrated on a separate validation set. On 100 score-blind targets, independent GNM evaluation shows:

Guidance methodFull-chain GNM rrInterior GNM rr
BG0.3830.064
FlexFlow0.6110.176
Paired Δr\Delta r+0.228+0.112
95% CI[0.189, 0.266][0.069, 0.153]
Targets favoring FlexFlow87/10078/100

Interior evaluation excludes 10% of residues from each terminus. Strengths were separately calibrated (λ≈200\lambda\approx200 for FlexFlow and λ≈86.8\lambda\approx86.8 for BG), rather than using the submitted λ=400\lambda=400 setting.

A further factorial ablation varied predictor, guidance location, and objective. Retaining BG's predictor and noisy-coordinate guidance while replacing MSE with Pearson improved GNM correlation by +0.270 on full chains and +0.205 on interior residues in the multi-seed analysis. Across a 32-fold strength sweep, tested MSE configurations reached at most r=0.467r=0.467, versus r=0.663r=0.663 for Pearson. These results identify the correlation objective as the main driver of improved relative steering; the independent benefit of predicted-clean placement was not statistically established. This is a guidance-level comparison on MultiFlow, not a comparison against FliPS's conditional generator.

Comment 3: Independent flexibility evaluation and MD validation

We re-evaluated all 400 Figure 2 designs with the Anisotropic Network Model (ANM), independent of FlexFormer, ProtT5, and S2. The resulting flexibility–structure trade-off is:

LengthFlexFlow − FliPS ANM Δr\Delta rscRMSD (Å), FF / FliPS ↓pLDDT, FF / FliPS ↑scRMSD ≤5\leq5 Å (%), FF / FliPS
70−0.0639.35 / 8.8569.0 / 76.123 / 35
100−0.1048.98 / 10.3668.9 / 77.525 / 24
200−0.1719.84 / 13.4375.3 / 68.829 / 11
300−0.16611.49 / 19.1267.6 / 58.821 / 0

FliPS achieves higher ANM flexibility correlation; the short-length S2 advantage is not reproduced under ANM. Conversely, FlexFlow has better sequence-refolding consistency at longer lengths. At L=300L=300, 21/100 vs. 0/100 achieve scRMSD ≤5\leq5 Å (at ≤2\leq2 Å: 8/100 vs. 0/100). FliPS retains better local backbone geometry. The original full pipelines also use different screening budgets: FliPS selects from a screened candidate pool while FlexFlow generates one backbone per target. We interpret these as trade-offs, not an isolated guidance comparison.

Regarding the 40 MD designs, candidates were prioritized by S2 agreement, and the highest-S2 ProteinMPNN sequence was selected for each backbone. We verified this against all 40 structures. Appendix B correctly specifies S2 selection; the MD protocol's highest-pLDDT description is a documentation error. All-atom MD measured RMSF independently of S2, but this prioritized cohort is not an unbiased sample or a matched-baseline MD comparison.

Comment 4: Fidelity

For each position ii, we substitute all 20 amino acids while keeping backbone XX fixed; the same FlexFormer predictor returns vi,a=fFlexFormer(Si→a,X)iv_{i,a}=f_{\mathrm{FlexFormer}}(S_{i\to a},X)_i. Fidelity correlates these predictions with training-set amino-acid flexibility propensities; no separate mutation head is used. An identity-only predictor reproducing that propensity vector would attain fidelity = 1.0. We therefore clarify that fidelity is an auxiliary residue-propensity diagnostic, not direct evidence of context-specific mutation-effect accuracy or downstream design quality. We will revise this interpretation.

We sincerely thank the reviewer again for the constructive feedback. We have provided additional experiments addressing the guidance-strength trade-off, interior flexibility control, and independent evaluation. As the discussion period approaches its end, we would greatly appreciate any further feedback on whether our responses address your main concerns. We would be happy to provide additional clarification if needed. Thank you so much again!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.