acceptodds
Under review as a conference paper at ICLR 2027

Steer, Select, and Stratify: Training-Free Sampling for Protein Motif Scaffolding

Abstract

Motif scaffolding embeds functional sites with specified geometry into protein scaffolds for therapeutics and biocatalysis. RFdiffusion is widely used for this task. However, two challenges remain: obtaining successful designs on difficult tasks and covering diverse successful structures within a finite candidate budget. We further diagnose the first challenge into two bottlenecks: generated-motif errors are associated with design failure, while accurate motif geometry alone does not ensure structural recovery after sequence design and refolding. For the second challenge, independent sampling can repeatedly produce similar successful structures, limiting finite-batch coverage. We therefore propose -Scaffold (Steer, Select, and Stratify), a training-free sampling framework with three corresponding components: gated gradient guidance, rollout reweighting, and Exploration via Marginal-Preserving Stratification (EMPS). Gated gradient guidance restores truncated geometric gradients without changing the denoiser's forward computation and uses clean predictions to guide scaffold updates gated by denoising time and motif error. Rollout reweighting combines scaffold consistency and refolded-motif errors from sequence design and refolding into a parent score for allocating independent continuations. EMPS jointly stratifies scaffold layouts and a geometry-dependent process coordinate of translational noise without additional model evaluations. Before rollout reweighting, we prove that EMPS preserves matched-reference marginals and expected success count while providing expected coverage of fixed successful categories no worse than independent or single-axis sampling. For the full framework, we establish finite-batch lower bounds on expected coverage under anchor-retention and successful-continuation conditions. Across all 30 MotifBench tasks at the standard budget of 100 backbones per task, -Scaffold raises RFdiffusion's published score from 21.95 to 30.27, solved tasks from 17 to 22, and successful clusters from 205 to 225. Component ablations and mechanism diagnostics support the complementary roles of geometric control, downstream feedback, and coordinated exploration.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein design, motif scaffolding, diffusion models, training-free sampling, inference-time guidance, stratified sampling, structural diversity

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_e3sPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates whether the performance of protein motif scaffolding can be improved by changing how candidates are sampled, rather than retraining the underlying generative model. Using pretrained RFdiffusion, the authors focus on two practical objectives: obtaining more designs that successfully recover the intended motif after sequence design and refolding, and exploring a broader range of successful scaffold structures within a limited candidate budget.

The proposed framework, S³-Scaffold, intervenes at different stages of the generation process. During denoising, it uses geometric gradients to steer scaffold residues toward better motif preservation, with an error-dependent gate to limit excessive corrections. At an intermediate sampling step, it evaluates candidate structures using ProteinMPNN and ESMFold, then assigns more independent continuations to promising candidates. In parallel, a stratified sampling scheme called EMPS distributes trajectories across scaffold layouts and diffusion-noise regions to reduce redundant exploration. The authors also analyze the conditions under which these interventions improve local geometry, sampling allocation, and coverage of distinct successful structures.

On 30 MotifBench tasks, S³-Scaffold increases the benchmark score from 21.95 to 30.27, the number of solved tasks from 17 to 22, and successful structural clusters from 205 to 225, using 100 final designs per task. The ablation results reveal a distinction between generating more successful candidates and discovering more distinct successful structures: rollout reweighting primarily increases the former, whereas EMPS improves the latter. Overall, the study presents an inference-time approach to improving the utilization of a pretrained protein generator, while requiring additional computation for geometric guidance and intermediate candidate evaluation.

Critical strengths

1. Effective integration of complementary sampling strategies

The paper presents a coherent sampling framework that addresses distinct failure modes in protein motif scaffolding, including motif geometry errors, downstream refolding failures, and limited structural diversity. The component-wise ablations provide useful evidence of their complementary effects. Notably, rollout reweighting increases the number of successful designs but reduces structural diversity, whereas EMPS improves structural coverage without increasing the number of successful designs. The combined framework improves the MotifBench score from 21.95 to 30.27 without retraining RFdiffusion.

2. Rigorous mechanism-level analysis and implementation transparency

The paper goes beyond reporting benchmark improvements by providing mathematical analyses and targeted diagnostics of the proposed sampling mechanisms. In particular, the restored geometric gradients are validated against finite differences, and local guidance interventions reduce motif errors in 31 of 32 tested cases. The EMPS component is accompanied by theoretical results on marginal preservation and expected coverage under explicitly stated assumptions. Detailed implementation settings and computational cost measurements further support the technical credibility and reproducibility of the work.

Critical weaknesses

1. Insufficient methodological novelty and significance beyond RFdiffusion

Although the proposed framework improves RFdiffusion's performance, its methodological novelty and broader significance appear insufficient for a main-track ICLR contribution. The approach primarily combines existing sampling strategies with implementation-level modifications tailored to a single pretrained model. Without demonstrating effectiveness across different diffusion architectures or establishing sufficiently novel model-independent insights, the work remains closer to an RFdiffusion-specific optimization than a broadly applicable methodological advance. While EMPS provides a theoretical contribution, its practical significance beyond the RFdiffusion setting has not been established.

2. Lack of comparisons under matched computational budgets

Although S³-Scaffold improves motif-scaffolding performance at a fixed number of final candidates, it requires substantially more computation than standard RFdiffusion. The authors provide detailed runtime measurements, but do not compare the proposed framework against generating additional independent RFdiffusion samples under an equivalent computational budget. In the reported pilot comparison, gradient guidance increases per-trajectory generation time by approximately 3.7–3.9 times relative to standard RFdiffusion, while rollout reweighting introduces additional sequence-design, refolding, and diffusion-continuation computations. Consequently, it remains unclear whether the reported improvements translate into better computational efficiency. A compute-matched comparison of successful designs and structural coverage is necessary to establish the practical advantages of the framework.

3. Compromised evaluation independence due to benchmark-informed hyperparameter selection

A major concern is the lack of independence between method development and final evaluation. As explicitly stated in Appendix A.3, all 30 MotifBench tasks were used during development, including the selection of the rollout boundary, and the same tasks were subsequently used for final performance evaluation. No held-out evaluation tasks were reserved. This creates a risk of benchmark-specific hyperparameter selection and potentially optimistic performance estimates, limiting the evidence for generalization to unseen motif-scaffolding tasks. This concern is further amplified by the reliance on a single benchmark and the absence of repeated independent runs with uncertainty estimates. Evaluation on genuinely held-out tasks or an independent benchmark would be necessary to establish the robustness and generalizability of the reported improvements.

Questions for authors

Q1. Methodological novelty and generalizability

The proposed framework is evaluated exclusively on the original RFdiffusion model. Could the authors clarify which methodological contributions are expected to generalize beyond RFdiffusion, particularly given the model-specific implementation of geometric gradient restoration? Have the authors evaluated any of the proposed sampling strategies on other diffusion-based protein generators, such as RFdiffusion2 or RFdiffusion3? If not, what evidence supports their broader applicability?

Q2. Computational efficiency under matched budgets

The proposed framework incurs substantial additional computational costs from gradient guidance, intermediate sequence design and refolding, and diffusion continuations. Could the authors compare S³-Scaffold against standard RFdiffusion using an equivalent GPU-time budget, allowing the baseline to generate additional independent samples? In particular, how do the methods compare in terms of successful designs and successful structural clusters as a function of computational cost?

Q3. Independence of hyperparameter selection and evaluation

Appendix A.3 explicitly states that all 30 MotifBench tasks participated in method development and that no held-out tuning split was used. Could the authors clarify the extent to which the final hyperparameters were selected based on performance on these same tasks? Given this overlap, how can the reported benchmark improvements be distinguished from potential benchmark-specific tuning effects? Could the authors provide an evaluation on previously unseen motif-scaffolding tasks or an independent public benchmark?

Q4. Statistical robustness of the reported improvements

The main experimental comparisons appear to rely on a single generated pool per configuration. Could the authors provide results from multiple independent random seeds, including mean performance and variability for MotifBench Score, solved tasks, successful designs, and successful structural clusters? This would help determine whether the reported improvements are stable across independent sampling runs.

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/2)

Thank you for the questions on transferability, computational value, evaluation independence, and sampling variability.

Q1. Methodological novelty and generalizability

S³-Scaffold provides a three-stage sampling strategy: Steer trajectories through geometry-aware, error-gated guidance; Select continuations using dual-criterion refolding feedback; and Stratify exploration through joint layout-noise sampling to improve structural diversity. Despite building on RFdiffusion, our method improves over all its versions and serves as a strong complementary module.

  • Steer: Gradient construction and error-gated guidance. Wu et al. (2023) reported unstable translation and rotation gradients when applied to RFdiffusion, often causing trajectory divergence. We address this by restoring coordinate derivatives through RFdiffusion's geometric operations while preserving its native forward predictions. Finite-difference tests validate the restored gradients at the tested states (Appendices B.2 and B.5.2). We further introduce an error gate that adapts guidance strength to predicted-motif error, reducing unnecessary scaffold perturbations once motif accuracy is sufficient. Full-benchmark ablations demonstrate improvements in Score, successful designs, and successful clusters, both with and without rollout (Appendices B.3, B.6, and I.1.2).

  • Select: Dual-criterion refolding feedback. Hartman et al. (2025) use sequence-design and relaxation rewards to resample RFdiffusion trajectories. Our approach introduces a joint refolding score that explicitly enforces both scaffold consistency and original-target motif recovery. Specifically, it takes the maximum of the two threshold-normalized errors from each refold and averages the resulting exponential scores across four refolds. This prevents a low error in one criterion from compensating for a high error in the other. The resulting scores guide the allocation of independent continuations from saved states (Section 4.2; Appendix C).

  • Stratify: Joint stratification for structural diversity. We introduce a joint stratification of scaffold layouts and geometry-dependent Gaussian-noise paths, requiring no additional denoiser evaluations. Our analysis establishes trajectory-marginal preservation before rollout reweighting and an expected-coverage ordering among joint, single-axis, and independent sampling under a fixed successful-category map and matched one-continuation reference, while preserving the expected success count (Theorems 1–2; Appendices D–E). Appendix F further establishes full-framework coverage lower bounds under retention and continuation conditions. Empirically, EMPS improves successful-structure coverage on both RFdiffusion (205 to 212 clusters, with 481 successful designs in both pools) and RFdiffusion3 (329 to 351 clusters, with Score increasing from 43.23 to 45.64), demonstrating its applicability beyond the original RFdiffusion model (Appendices I.2 and J).

RFdiffusion3 natively preserves fixed motif coordinates in its denoised predictions, so we transfer EMPS and rollout to RFdiffusion3. MotifBench evaluation is unchanged. Each configuration uses one 30-task run with 100 designs per task (main Table 1; Appendix J, Tables 33–34):

ConfigurationScoreSolved /30Successful designsSuccessful clustersRefolds
Native RFdiffusion343.23182889532924,000
RFdiffusion3 + EMPS45.63762991835124,000
RFdiffusion3 + EMPS + rollout46.1353281,09930736,000

EMPS improves all four outcomes at the same 24,000 refolding calls. Adding rollout further increases successful-design yield and Score, but reduces structural diversity at a higher cost of 36,000 calls. EMPS requires access to scaffold layouts and Gaussian-noise paths, whereas rollout requires resumable states and downstream evaluation. These results demonstrate that our training-free sampling framework generalizes beyond RFdiffusion and improves performance on the stronger RFdiffusion3 generator.


Q2. Computational efficiency under matched budgets

We compare S³-Scaffold against two expanded RFdiffusion baselines under complementary budget constraints: RF150 matches refolding calls, while RFtime matches GPU time. Across all 30 tasks, S³-Scaffold achieves a higher success rate (22.84 ± 0.57) than RF150 (16.11) and RFtime (15.92), and solves more tasks on average (21.67 ± 1.53 versus 17 and 19, respectively).

RF150 uses the same 1,200 refolding calls per task to evaluate 150 final backbones. RFtime uses the per-task GPU allocations measured from a complete S³-Scaffold run, totaling 237.179 GPU-hours on NVIDIA A800 80GB GPUs. S³-Scaffold results are reported as mean ± sample SD over three independent runs. All counts use complete evaluated pools, without 50-design subsampling.

AuthorsOpenReview ↗

Rebuttal by Authors (2/2)

Q2. Computational efficiency under matched budgets (continued)

  • Structural and task coverage on multi-segment motifs. S³-Scaffold's coverage advantage is most pronounced on the ten tasks with at least three motif segments (Table 20):

    ConfigurationFinal designs/runSuccessful designsSuccessful clustersSolved /10
    S³-Scaffold (3 runs)1,000148.00 ± 1.0017.33 ± 4.044.67 ± 0.58
    RF1501,50016462
    RFtime2,83730582

    All three S³-Scaffold runs outperform both baselines in successful clusters (22, 15, and 15) and solved tasks (5, 4, and 5). In particular, S³-Scaffold solves 24_1QY3 and 27_4XOJ in every run, whereas neither RFdiffusion baseline succeeds on either task (Table 22 and Appendix G.3).

    The additional yield of RFtime does not translate into broader structural coverage. From RF150 to RFtime, 185 of the 199 additional successful clusters (93.0%) come from single-segment tasks. On tasks with at least three segments, successful designs increase from 164 to 305, but solved tasks remain at two. Notably, 150 and 284 of these designs, respectively, come from task 22_1BCF and belong to a single cluster in each pool. Thus, RFdiffusion's increased designs fail to translate into increased structural diversity or task coverage.

  • Performance as a function of GPU cost. We further evaluate RFdiffusion at 25%, 50%, 75%, and 100% of the per-task GPU allocations, counting only designs whose final evaluation finishes within each budget (Appendix H, Table 21). At the full allocation, RFdiffusion produces more successful designs and clusters in total; RF150 also yields 725 successful designs and 265 clusters. However, S³-Scaffold distributes its successful designs across more tasks and achieves substantially greater coverage on complex, multi-segment motifs, highlighting the distinction between sampling yield and effective coverage.

    ConfigurationFinal designs/runGPU-hoursSuccessful designsSuccessful clustersSolved /30
    RFdiffusion / 25% allocation2,10858.82533514617
    RFdiffusion / 50% allocation4,236118.17468325317
    RFdiffusion / 75% allocation6,367177.4981,02735017
    RFdiffusion / 100% allocation8,501236.7691,35346419
    S³-Scaffold (3 runs)3,000237.179685.33 ± 17.21227.00 ± 5.2921.67 ± 1.53

Q3. Hyperparameter selection and external evaluation

1. Hyperparameter selection. We selected the rollout boundary τ=10\tau = 10 based on all 30 MotifBench tasks. Other sampling hyperparameters were not optimized (Appendix A.3).

2. External evaluation on GeomMotif. Without further hyperparameter tuning, we evaluated S³-Scaffold and RFdiffusion on all 57 GeomMotif tasks (Strashnov et al., 2026), generating 100 designs per task:

ConfigurationSUN (%)Successful designsSuccessful clusters
RFdiffusion37.97923,2692,517
S³-Scaffold40.08943,9522,620

S³-Scaffold achieves higher SUN, successful-design yield, and structural diversity than RFdiffusion, demonstrating its effectiveness on an external benchmark without benchmark-specific tuning. The full experimental protocol and per-task results are provided in Appendix K (Table 35).


Q4. Statistical robustness across independent runs

We conducted three independent S³-Scaffold runs, each generating 100 designs per task across all 30 tasks (Appendix G.1; per-task results in Tables 16–18):

RunScoreSolved /30Successful designsSuccessful clusters
Run 1 (original)30.273222678225
Run 228.375920673223
Run 331.761523705233
Mean ± sample SD30.14 ± 1.7021.67 ± 1.53685.33 ± 17.21227.00 ± 5.29

The three runs show consistent performance. Each independently outperforms the public RFdiffusion reference, demonstrating that the improvements are reproducible instead of driven by a single run.


References

Wu, L., Trippe, B. L., Naesseth, C. A., Blei, D. M., and Cunningham, J. P. (2023). Practical and asymptotically exact conditional sampling in diffusion models. NeurIPS, 36, 31372–31403.

Hartman, E., Wallin, J., Malmström, J., and Olsson, J. (2025). Controllable protein design with particle-based Feynman–Kac steering. arXiv:2511.09216v2.

Strashnov, P., Shevtsov, A., Meshchaninov, V., Kardymon, O., and Vetrov, D. P. (2026). GeomMotif: A benchmark for arbitrary geometric preservation in protein generation. ICLR.

I thank the authors for their detailed rebuttal and substantial additional experiments. The new results address several of my original concerns and improve the empirical support for the proposed approach. Nevertheless, I remain inclined to maintain my original Weak Reject recommendation for the following reasons.

Q1 — Generalizability

The RFdiffusion3 experiments substantially address my concern. The improvements from EMPS on a stronger generator provide meaningful evidence that its benefits are not confined to the original RFdiffusion implementation. I appreciate this additional validation and no longer consider the absence of experiments on a stronger generator a major concern.

Q2 — Compute-matched evaluation and practical contribution

I appreciate the addition of both refolding-matched and GPU-time-matched comparisons. These experiments directly address my original request.

However, the results materially affect my assessment of the paper's practical contribution. At approximately equal GPU cost, S³-Scaffold produces substantially fewer successful designs (685 vs. 1,353) and successful structural clusters (227 vs. 464) than expanded RFdiffusion sampling.

I acknowledge that S³-Scaffold achieves higher per-candidate success rates and solves more tasks, particularly among challenging multi-segment motifs. This is a meaningful advantage and is consistent with one of the objectives stated in the original submission.

Nevertheless, the original paper also emphasized improvements in successful-design yield and structural diversity under a fixed candidate budget. The compute-matched results reveal an important trade-off: these improvements do not translate into higher total design yield or structural coverage under a fixed computational budget.

While this does not invalidate the original fixed-candidate results, it substantially limits the demonstrated practical benefit of the method. I remain unconvinced that the improvement in difficult-task coverage sufficiently compensates for the substantial reduction in overall sampling throughput.

Q3 — Hyperparameter selection and independent evaluation

The authors acknowledge that the rollout boundary was selected using all 30 MotifBench tasks, which compromises the independence of the original benchmark evaluation.

The additional GeomMotif results provide valuable evidence of transfer beyond MotifBench. However, they do not remove the potential selection bias in the original evaluation. The extent of structural overlap between the benchmarks is also unclear.

Furthermore, the GeomMotif comparisons use matched final candidate counts rather than matched computational budgets. Thus, while the results support generalization to additional tasks, they do not establish superior computational efficiency.

I therefore consider this concern partially, but not fully, addressed.

Q4 — Reproducibility

The three independent runs provide useful evidence of stability, and I consider my original reproducibility concern substantially addressed. Although repeated baseline runs would allow a stronger statistical comparison, the additional experiments are a meaningful improvement.

##############################

I acknowledge the substantial effort invested in the rebuttal and the positive evidence for improved coverage of challenging motif-scaffolding tasks. My remaining concern is not primarily the absence of additional experiments, but the scope and significance of the contribution established by those experiments.

In particular, the compute-matched results reveal a substantial trade-off between difficult-task coverage and total successful-design yield and structural diversity. Combined with the benchmark-specific hyperparameter selection and the remaining uncertainty regarding the empirical advantages of joint stratification, I do not find the evidence sufficiently compelling to change my original recommendation.

I therefore maintain my Weak Reject rating, while recognizing that the rebuttal has substantially strengthened the paper.

Reviewer_rwW4OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes 3 components to improve motif scaffolding in RFDiffusion: (i) steering intermediate state to reduce motifRMSD, (ii) reweight the trajectories by intermediate states success, and (iii) stratify the scaffold layout and noises that were usually randomly sampled.

Critical strengths

Stratify (i.e. EMPS) portion of the method seems original.

Critical weaknesses
  1. The evaluation is matched final sample count not matched compute nor matched oracle calls. When producing 100 backbones, RFdiffusion calls 800 folding (oracle) calls, whereas this method calls 1200 refolding calls. This plus Tab6 in Appendix shows that S3 scaffold cost 3-4x wall clock time. Fair comparison is RFDiffusion at matched compute and matched oracle calls. With current setting, we cannot conclude whether the performance gain is coming from more structure evaluations or proposed method.

  2. Guidance and reward-based reweighting are established methods in diffusion models, thus method and theoretical claims are largely application of known theory rather than new method/ new theoretical contributions.

Questions for authors
  1. What is RFDiffusion performance at matched compute and matched refolding calls, respectively?
  2. The method is general to any protein design models. Why limit to RFDiffusion? Specifically why not apply to current SOTA RFDiffusion3?
  3. Does gains persist as combinatorial complexity of layout grows? How does the performance looks like by the number of motif segments (1, 2, 3+)
5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/2)

Thank you for recognizing EMPS's originality and requesting separate comparisons at matched refolding calls and GPU time. We have implemented all the requested tasks and shown our method's consistent improvements over existing approaches. Despite building on RFdiffusion, our method improves over all its versions and serves as a strong complementary module.

1. Matching refolding calls and GPU allocations (W1, Q1)

A complete 30-task S³-Scaffold run with 100 final designs per task uses 237.179 GPU-hours on NVIDIA A800 80GB GPUs. GPU allocations use this single-run timing profile; S³-Scaffold performance reports mean ± sample SD across three independent runs.

For equal refolding calls, S³-Scaffold uses 400 internal and 800 endpoint refolds per task; RF150 uses the same 1,200 calls to evaluate 150 final backbones. For equal GPU time, RFtime uses the measured per-task S³-Scaffold hours and admits only endpoints completed within each cutoff. Its admitted prefixes contain 281–287 backbones per task, totaling 8,501.

All counts and success rates use complete evaluated pools, without 50-design subsampling.

Results: main Table 2 and Appendix H, Table 22; S³-Scaffold run summaries: Appendix G.1.

ConfigurationFinal designs/runSuccessful designsSuccessful clustersSolved /30Success rate (%)
S³-Scaffold (3 runs)3,000685.33 ± 17.21227.00 ± 5.2921.67 ± 1.5322.84 ± 0.57
RF150: equal refolds4,5007252651716.11
RFtime: equal GPU hours8,5011,3534641915.92

Under both budgets, expanded RFdiffusion sampling gives more total successes and clusters. S³-Scaffold has higher per-candidate success and solves more tasks. Appendix H provides the budget definitions and RFdiffusion per-task results (Tables 22–23), while Appendix G.3 reports the three S³-Scaffold runs.

On tasks with at least three motif segments, S³-Scaffold also achieves broader structural and task coverage under matched cost (Response 4 below; Table 20), showing its stronger structural diversity and task coverage on more challenging tasks.


2. Methodological contributions relative to prior work (W2)

S³-Scaffold provides a three-stage sampling strategy: Steer trajectories through geometry-aware, error-gated guidance; Select continuations using dual-criterion refolding feedback; and Stratify exploration through joint layout-noise sampling to improve structural diversity.

  • Steer: Gradient construction and error-gated guidance. Wu et al. (2023) reported unstable translation and rotation gradients when applied to RFdiffusion, often causing trajectory divergence. We address this by restoring coordinate derivatives through RFdiffusion's geometric operations while preserving its native forward predictions. Finite-difference tests validate the restored gradients at the tested states (Appendices B.2 and B.5.2). We further introduce an error gate that adapts guidance strength to predicted-motif error, reducing unnecessary scaffold perturbations once motif accuracy is sufficient. Full-benchmark ablations demonstrate improvements in Score, successful designs, and successful clusters, both with and without rollout (Appendices B.3, B.6, and I.1.2).

  • Select: Dual-criterion refolding feedback. Hartman et al. (2025) use sequence-design and relaxation rewards to resample RFdiffusion trajectories. Our approach introduces a joint refolding score that explicitly enforces both scaffold consistency and original-target motif recovery. Specifically, it takes the maximum of the two threshold-normalized errors from each refold and averages the resulting exponential scores across four refolds. This prevents a low error in one criterion from compensating for a high error in the other. The resulting scores guide the allocation of independent continuations from saved states (Section 4.2; Appendix C).

  • Stratify: Joint stratification for structural diversity. We introduce a joint stratification of scaffold layouts and geometry-dependent Gaussian-noise paths, requiring no additional denoiser evaluations. Our analysis establishes trajectory-marginal preservation before rollout reweighting and an expected-coverage ordering among joint, single-axis, and independent sampling under a fixed successful-category map and matched one-continuation reference, while preserving the expected success count (Theorems 1–2; Appendices D–E). Appendix F further establishes full-framework coverage lower bounds under retention and continuation conditions. Empirically, EMPS improves successful-structure coverage on both RFdiffusion (205 to 212 clusters, with 481 successful designs in both pools) and RFdiffusion3 (329 to 351 clusters, with Score increasing from 43.23 to 45.64), demonstrating its applicability beyond the original RFdiffusion model (Appendices I.2 and J).

AuthorsOpenReview ↗

Rebuttal by Authors (2/2)

3. Application to RFdiffusion3 (Q2)

We began this project in early June, choosing RFdiffusion as an established general-purpose backbone generator with public MotifBench results. RFdiffusion3's three-run MotifBench results obtained with the corrected implementation were released in late August, after the project had started. We have now extended our training-free sampling components to RFdiffusion3 with its pretrained weights frozen.

RFdiffusion3 natively preserves fixed motif coordinates in its denoised predictions, so we transfer EMPS and rollout to RFdiffusion3. The results below are our evaluations, with one 30-task run per configuration, 100 designs per task, and eight fresh endpoint refolds per design, following the MotifBench protocol (Appendix J).

RFdiffusion3 transfer results (Appendix J, Table 33; summary in main Table 1).

ConfigurationScoreSolved /30Successful designsSuccessful clustersRefolds
Native RFdiffusion343.23182889532924,000
RFdiffusion3 + EMPS45.63762991835124,000
RFdiffusion3 + EMPS + rollout46.1353281,09930736,000

EMPS improves all four outcome metrics without additional model evaluations or refolds. Adding rollout increases successful-design yield and Score while reducing successful-cluster and solved-task counts. These results demonstrate that our training-free sampling components also improve performance on a stronger generator, beyond their original RFdiffusion implementation.


4. Performance by number of motif segments (Q3)

Table 20 groups tasks by motif-segment count, with ten tasks per group. Counts are group totals; S³-Scaffold reports mean ± sample SD across three runs.

Performance on MotifBench grouped by the number of motif segments (Appendix H, Table 20).

SegmentsConfigurationFinal designs/runSuccessful designsSuccessful clustersSolved /10
1S³-Scaffold (3 runs)1,000337.67 ± 10.69185.33 ± 4.738.33 ± 0.58
1RF1501,5003772388
1RFtime2,8327124238
2S³-Scaffold (3 runs)1,000199.67 ± 12.8624.33 ± 4.168.67 ± 0.58
2RF1501,500184217
2RFtime2,832336339
3+S³-Scaffold (3 runs)1,000148.00 ± 1.0017.33 ± 4.044.67 ± 0.58
3+RF1501,50016462
3+RFtime2,83730582

S³-Scaffold shows increasingly strong advantages as motif complexity grows. For three or more segments, S³-Scaffold covers more successful structures and solves more tasks than both expanded RFdiffusion baselines, despite producing fewer successful designs in total. In the two-segment group, it improves all three mean outcomes over RF150, while RFtime has higher totals. These results highlight the particular effectiveness of S³-Scaffold on more challenging motif scaffolding tasks.

Appendices G.2 and G.3 report the fixed-candidate comparison with the public RFdiffusion reference.


References

Hartman, E., Wallin, J., Malmström, J., and Olsson, J. (2025). Controllable protein design with particle-based Feynman–Kac steering. arXiv:2511.09216 (version 2, April 6, 2026).

Wu, L., Trippe, B. L., Naesseth, C. A., Blei, D. M., and Cunningham, J. P. (2023). Practical and asymptotically exact conditional sampling in diffusion models. Advances in Neural Information Processing Systems, 36, 31372–31403.

While I appreciate S3 scaffold solving more problems, especially at 3+ motif segments, the relevant metric here is number of successful clusters at matched compute not per candidate success rate. The new results table show at matched refolding calls or GPU hours base RFDiffusion sampling generates much much more # successful clusters than S3 scaffold. This really makes me skeptical of the advantage of the proposed method. Also the method doesn't transfers directly to newer/ different baselines. The motif coordinates part is specific to subset of protein design models, and RFdiffusion3 + EMPS + rollout underperforms base RFD3 in terms of # success clusters. Maybe the benefit of this method lies in more challenging motif scaffolding tasks with 3+ segments, and authors could explore this direction, but with current results I cannot change the rating.

Reviewer_DoxsOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes S3S^3-Scaffold, a training-free sampling method for protein motif scaffolding with frozen RFdiffusion. It combines three components: a) restoring geometric derivatives and applying gated gradient guidance, b) using sequence design and refolding to evaluate trajectory continuations, and c) spreading samples across scaffold layouts and diffusion noise ranges to improve structural diversity. On the MotifBench benchmark, both the reported score and the number of solved tasks increase.

Critical strengths
  1. A mathematically supported EMPS construction that preserves marginals and improves expected coverage of fixed successful categories under stated assumptions.

  2. The paper identifies specific coordinate dependencies truncated in RFdiffusion and restores their derivatives while preserving forward values.

  3. The method produces more successful designs and successful clusters in the reported evaluation, across the complete MotifBench task set.

Critical weaknesses
  1. The demonstrated contribution depends heavily on the original RFdiffusion implementation. The gradient restoration repairs dependencies missing from that implementation. Rollout and EMPS could have broader value, but neither is tested on a stronger generator. Practical usefulness also needs to be weighed against substantial additional computation and memory use.

  2. The evidence is confined to a benchmark used during development. All 30 MotifBench tasks participated in development, and each main configuration has one generated pool per task. Exact subset calculations remove scoring variation within those saved pools, but do not establish repeatability across newly generated batches. GeomMotif (ICLR'26) is a relevant missing evaluation because it tests complementary geometric preservation tasks and already evaluates RFdiffusion. Its omission leaves usefulness beyond MotifBench untested.

  3. The code and data is not provided. However, Appendix G.5 references to CSV and script files.

Questions for authors
  1. What establishes practical value beyond the RFdiffusion repair? Given the implementation-specific changes documented in B.2, what would rollout and EMPS require from another generator? Any existing evidence on a stronger generator would help assess how broadly useful these components are.

  2. How repeatable are the reported component gains? Are results from independently generated batches available, particularly for the smaller EMPS improvement? This is the most direct uncertainty affecting confidence in the measured gains.

  3. Does combining layout and process-noise stratification improve results over using either alone? The paper checks that sampling runs correctly on four tasks, but shows final results for layout-only, process-only, and joint EMPS on just one selected task. Are final results for these three versions available for the other three tasks?

2 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/2)

We appreciate your constructive feedback and recognition of our geometric-derivative restoration and EMPS analysis. We now include RFdiffusion3 and GeomMotif evaluations, matched-budget comparisons, three independent S³-Scaffold runs, and full-benchmark single-axis ablations. While our method builds on RFdiffusion, we demonstrate consistent improvements over all its versions, establishing our approach as a valuable complementary technique.

1. Transfer of rollout reweighting and EMPS to RFdiffusion3 (W1, Q1)

RFdiffusion3 natively preserves fixed motif coordinates in its denoised predictions, so we transfer EMPS and rollout to RFdiffusion3 without adding geometric guidance. EMPS requires access to layout sampling and Gaussian translational noise; rollout requires intermediate clean predictions, a downstream evaluator, and saved states supporting independent continuations. RFdiffusion3 exposes all required interfaces. Each configuration uses one 30-task run with frozen RFdiffusion3 and 100 designs per task. Final evaluation uses eight fresh sequence/refold pairs independent of parent scoring and retains the MotifBench protocol (Appendix J).

Revised main Table 1 and Appendix J, Tables 33–34:

ConfigurationScoreSolved /30Successful designsSuccessful clustersRefolds
Native RFdiffusion343.23182889532924,000
RFdiffusion3 + EMPS45.63762991835124,000
RFdiffusion3 + EMPS + rollout46.1353281,09930736,000

EMPS improves all four outcome metrics without additional model evaluations or refolds. Adding rollout increases successful-design yield and Score while reducing clusters and solved tasks.


2. Computational cost and memory (W1)

A complete 30-task S³-Scaffold run with 100 final designs per task uses 237.179 GPU-hours on NVIDIA A800 80GB GPUs. This timing profile sets RFdiffusion's GPU allocations; performance is reported as mean ± sample standard deviation across three independent S³-Scaffold runs.

We compare two expanded RFdiffusion baselines. RF150 matches refolding calls: both methods use 1,200 calls per task, with S³-Scaffold allocating 400 internally and 800 to final evaluation, while RF150 evaluates 150 final designs. RFtime matches the per-task A800 GPU hours measured for S³-Scaffold, admitting only designs whose final evaluation completes within each task's allocation.

For a fair comparison, all counts and success rates use complete evaluated pools, without 50-design subsampling.

Results: main Table 2 and Appendix H, Table 22; S³-Scaffold run summaries: Appendix G.1.

ConfigurationFinal designs/runSuccessful designsSuccessful clustersSolved /30Success rate (%)
S³-Scaffold (3 runs)3,000685.33 ± 17.21227.00 ± 5.2921.67 ± 1.5322.84 ± 0.57
RF150: equal refolds4,5007252651716.11
RFtime: equal GPU allocation8,5011,3534641915.92

S³-Scaffold has higher per-candidate success and solves more tasks, while expanded RFdiffusion sampling produces more successful designs and clusters in total. The key distinction is that S³-Scaffold distributes its successful designs more broadly across tasks and structural clusters, particularly on complex multi-segment motifs. For the ten tasks with three or more motif segments, S³-Scaffold gives 148.00 ± 1.00 successful designs, 17.33 ± 4.04 clusters, and 4.67 ± 0.58 solved tasks, versus 164/6/2 for RF150 and 305/8/2 for RFtime (Table 20).

We additionally report peak PyTorch-allocated GPU memory observed in selected generation runs on NVIDIA A800 GPUs, measured with torch.cuda.max_memory_allocated(). The recorded values were 17.90 GiB at protein length 125, 49.26–49.28 GiB at length 216, and 63.84 GiB at length 247.


3. Evaluation beyond MotifBench (W2)

We evaluated S³-Scaffold and RFdiffusion on all 57 GeomMotif tasks, using 100 designs per task and eight fresh sequence-design/refolding evaluations per design. Without further hyperparameter tuning, S³-Scaffold achieves improved design success and structural diversity, demonstrating its generalization beyond MotifBench.

SettingSUN (%)Successful designsSuccessful clustersSolved /57
RFdiffusion37.97923,2692,51756
S³-Scaffold40.08943,9522,62056

Appendix K provides the experimental protocol, SUN calculation, and complete results for all 57 tasks (Table 35).

AuthorsOpenReview ↗

Rebuttal by Authors (2/2)

4. Independent runs and evidence for the EMPS gain (W2, Q2)

For each MotifBench task, we conducted three independent S³-Scaffold runs, generating 3,000 designs per run:

MetricS³-Scaffold, mean ± sample SD
Score30.14 ± 1.70
Solved tasks21.67 ± 1.53
Successful designs685.33 ± 17.21
Successful clusters227.00 ± 5.29

Across the three independent S³-Scaffold runs, the total cluster counts are 225, 223, and 233, with expected 50-design subset coverages of 147.64, 145.77, and 154.53, respectively, demonstrating consistent performance across runs. Moreover, all three runs achieve higher expected coverage than the single w/o EMPS run (138.54, with 214 clusters), showing that EMPS improves structural diversity under a limited candidate budget (Appendix G.1 and Table 27).

The RFdiffusion3 comparison above also shows higher Score and structural coverage with EMPS at the same 24,000 refolding calls (Appendix J, Table 33).

Per-task results appear in Appendices G.3 and I.4 for MotifBench and Table 34 for RFdiffusion3.


5. Joint versus single-axis stratification (Q3)

We expanded the ablation to all 30 tasks, including the three additional diagnostic tasks requested. Guidance and rollout are disabled here; each configuration contributes one complete run with 100 designs per task.

Revised Tables 3 and 27; task details in Tables 29–32:

StratificationScore ↑Solved /30Successful designsSuccessful clusters
Layout-only22.754016484216
Process-only22.573017480214
Joint EMPS23.497917481212

Grouping the 30-task results by motif-segment count gives the following mean Scores; each group contains ten tasks.

Stratification1 segment2 segments≥3 segments
Layout-only42.4721.234.56
Process-only45.1817.015.53
Joint EMPS44.2620.715.53

The single-axis schemes lead in different task groups: Process-only performs the best on single-segment tasks, whereas Layout-only performs the best on two-segment tasks. Joint EMPS stays close to the best single-axis score in both groups and ties Process-only on tasks with three or more segments, achieving the highest overall Score.

Joint EMPS solves one additional task compared with Layout-only (17 versus 16). Joint and Process-only solve the same 17 tasks; Joint's overall Score advantage comes from the two-segment group, where Score increases from 17.01 to 20.71 and successful clusters from 14 to 19.


6. Code and data (W3)

We submitted the anonymized method code as supplementary material with this rebuttal (S3-Scaffold-anonymous.zip). The revised appendices document the evaluation protocol and provide the complete reported task-level results for the component, budget, and transfer experiments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.