acceptodds
Desk rejected at ICLR 2027

Prefix-Reuse Evolutionary Search with Normalized Proxy Guidance for Protein Binder Design

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Test-time steering guides a pretrained diffusion model toward high-reward outputs without retraining. Effective guidance should be computed at the clean-sample estimate rather than at a noisy intermediate state. the key requirement is a useful local direction, the gradient of an objective the sampler explicitly optimizes. Moreover, inexpensive scores computed along unfinished trajectories can be unreliable, making the observed score of a completed candidate the more trustworthy basis for selection. In this work, we propose PRES-NPG, a framework that combines Prefix-Reuse Evolutionary Search (PRES) with Normalized Proxy Guidance (NPG). NPG differentiates a proxy at the clean-sample estimate to obtain normalized local directions for inference-time steering. PRES evaluates completed candidates, selects their saved intermediate states based on these evaluations, and generates new continuations from the selected states across generations. We apply PRES-NPG to protein binder design using two models, PXDesign-d and Complexa. PRES-NPG improves binder quality through local directional guidance and produces a higher-quality retained set than both its registered random-parent control and independent best-of-\(N\) sampling, at a matched number of scored candidates. The code can be found at this anonymous link: https://anonymous.4open.science/r/PRES-NPG-8E22

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Test-time diffusion steering, Prefix-reuse evolutionary search, Protein binder design

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

The central premise of this paper is that computational effort in protein binder design can be used more effectively by leveraging completed sampling trajectories rather than treating every generation attempt independently. The authors propose PRES-NPG, an inference-time framework that combines gradient-based structural guidance with evolutionary search over previously generated candidates. The method operates on pretrained protein generators without updating their parameters.

PRES-NPG introduces two complementary mechanisms. Normalized Proxy Guidance (NPG) modifies denoising updates using normalized gradients of differentiable structural objectives, including steric clash avoidance, compactness, interface contacts, and hotspot proximity. These gradients are computed from predicted clean structures without backpropagating through the denoiser. Prefix-Reuse Evolutionary Search (PRES) evaluates fully generated binders using a terminal fitness score, identifies the most promising candidates, and resumes their saved intermediate denoising states with fresh randomness to produce new generations. This allows the method to concentrate sampling around promising structural configurations while avoiding redundant computation.

The authors evaluate the framework using PXDesign-d and Proteina-Complexa, with experiments covering ProtDBench and an additional BindCraft-derived target panel. On ProtDBench, PRES-NPG improves AF2 ipTM from 0.7508 to 0.8061 and Boltz-2 ipTM from 0.8035 to 0.8321 relative to independent guided best-of-64 sampling for the eight retained designs. These improvements are accompanied by a substantial reduction in structural diversity, reflecting a trade-off between exploiting high-scoring regions and exploring distinct binder configurations. The authors also report that prefix reuse reduces denoising computation by approximately 60%, although the measured end-to-end execution is only 1.2 times faster. Overall, the study demonstrates how completed-candidate evaluations and local structural guidance can be integrated to improve the predicted quality of protein binders within a fixed candidate-evaluation budget.

Critical strengths

1. Well-controlled empirical evaluation of fitness-guided search.

The strongest aspect of this paper is its empirical evaluation of fitness-guided evolutionary search for protein binder design. The authors compare PRES-NPG against both random-parent branching and independent guided sampling under a matched candidate-evaluation budget, providing evidence that fitness-based parent selection improves predicted binder quality. The improvements are also observed using Boltz-2, partially mitigating concerns about overoptimization toward the AF2-based selection metric. Although the underlying search principles are not fundamentally new, the experiments provide useful evidence of their effectiveness in this application.

2. Practical integration with distinct pretrained protein generators.

The framework is demonstrated on both PXDesign-d and Proteina-Complexa, representing diffusion-based and flow-matching-based generation, respectively. Its ability to operate without retraining the underlying models is a practical advantage. The use of saved intermediate states also reduces redundant denoising computation, although the end-to-end runtime improvement is relatively modest. Overall, the paper's main strength lies in its empirical validation and practical implementation rather than methodological novelty. The experiments support the usefulness of fitness-guided trajectory reuse for improving computational binder-design outcomes, even though closely related search strategies have already been explored in prior work.

Critical weaknesses

1. Limited Methodological Novelty and Insufficient Comparison with Prior Work

The central ideas of fitness-guided evolutionary search and trajectory reuse have already been explored in prior studies, including Diffusion-ES and Diffusion Tree Sampling. Although PRES-NPG introduces a specific implementation based on resuming saved intermediate states, its methodological advancement over existing approaches remains unclear. The absence of direct comparisons with closely related search algorithms makes it difficult to establish the significance of the proposed contribution.

2. Inconclusive Evidence for the Benefit of Fitness-Guided Parent Selection

One of the paper's central claims is that selecting promising parents based on completed-candidate fitness improves subsequent generation. However, on the additional BindCraft-derived benchmark, the improvement over random-parent branching is not statistically conclusive. The reported 95% confidence interval for the retained-set AF2 ipTM difference includes zero (−0.0145 to +0.0762). Although the main ProtDBench results are favorable, this uncertainty raises questions about the robustness and generalizability of the proposed fitness-guided selection strategy.

3. Incomplete Computational Efficiency Comparison

The primary comparisons use the same number of evaluated candidates rather than equivalent computational budgets. While prefix reuse reduces denoising computation by approximately 60%, the actual end-to-end runtime improvement is only about 1.2 times. It remains unclear whether PRES-NPG would maintain its performance advantage over alternative sampling strategies under matched wall-clock time or GPU budgets.

Questions for authors

1. Methodological Advancement over Prior Work

How does the proposed prefix-reuse strategy differ fundamentally from existing evolutionary diffusion search and trajectory-reuse methods, such as Diffusion-ES and Diffusion Tree Sampling? Beyond avoiding re-noising of completed samples, what specific algorithmic advantage does PRES provide? Could the authors provide a direct comparison with closely related approaches under matched computational budgets?

2. Statistical Robustness of Fitness-Guided Parent Selection

On the additional BindCraft-derived benchmark, the improvement of PRES-NPG over random-parent branching in retained-set AF2 ipTM has a 95% confidence interval of [−0.0145, +0.0762], which includes zero. Given that fitness-guided parent selection is a central component of the proposed framework, how do the authors interpret this statistically inconclusive result? Would additional targets or independent runs help establish whether fitness-based selection consistently outperforms random selection?

3. Computational Efficiency under Matched Budgets

The experiments match the number of evaluated candidates rather than total computational cost. Although prefix reuse reduces denoising computation by approximately 60%, the reported end-to-end speedup is only about 1.2 times. Would PRES-NPG still outperform independent guided sampling if both approaches were allocated the same wall-clock time or GPU budget?

6 comments
AuthorsOpenReview ↗

Title: Author Response To Reviewer (1/3)

Dear Reviewer czSb,

Thank you for your thoughtful comments. Key changes are highlighted in blue\color{blue}{\text{blue}} in the revised manuscript. We respond to your concerns below.


W1: Limited Methodological Novelty and Insufficient Comparison with Prior Work. The central ideas of fitness-guided evolutionary search and trajectory reuse have already been explored in prior studies, including Diffusion-ES and Diffusion Tree Sampling. Although PRES-NPG introduces a specific implementation based on resuming saved intermediate states, its methodological advancement over existing approaches remains unclear. The absence of direct comparisons with closely related search algorithms makes it difficult to establish the significance of the proposed contribution.

We thank the reviewer for this insightful suggestion. We agree that fitness-guided evolutionary search and reuse of intermediate states are established ideas, and we do not claim them as new. Our contribution is the integration and characterization of normalized structural-proxy guidance with fitness-guided allocation of fresh continuations for protein binder design.

PRES-NPG combines two complementary operations. Within each trajectory, NPG uses a differentiable structural proxy to determine the guidance direction and scales the guidance magnitude relative to the base denoising step. Across generations, PRES uses completed-candidate fitness to select which saved checkpoints receive additional continuations. This allows an external, potentially nondifferentiable evaluator to direct sampling while local guidance remains inexpensive and explicitly controlled.

Compared with Diffusion-ES: Diffusion-ES targets autonomous and instruction-guided driving, while PRES-NPG addresses protein binder design. PRES changes how subsequent candidates are initialized and allocated. Diffusion-ES re-noises selected completed outputs and uses stochastic reward-weighted resampling. PRES instead resumes recorded intermediate sampler states and allocates a fixed number of fresh continuations to each selected parent. These differences define the search policy. Their empirical effects are evaluated under the specified adaptations.

Compared with Diffusion Tree Sampling: Diffusion Tree Sampling studies reward-aligned image and text generation, while PRES-NPG targets protein binder design. PRES directly ranks completed trajectories from the current generation rather than maintaining recursively backed-up subtree values. It allocates equal continuation quotas to the selected saved states. Our comparison concerns the specified protein-domain adaptation, DTS-L, rather than all configurations of the original method.

Matched scored-candidate comparison. We compared PRES-NPG with re-noising under PRES allocation, a Diffusion-ES adaptation, and a DTS-style adaptation denoted DTS-L. All methods used the same generator, NPG guidance, fitness evaluator, branch schedule, and final-retention rule, with 64 scored candidates per target–seed run.

Table 1. Whole-pool search comparisons at a matched budget of 64 scored candidates per run on ProtDBench (ten targets. 30 paired target–seed cells per cohort). Scores average all candidates before final retention. Δ: PRES-NPG − comparator.

Comparator (all + NPG)CohortΔ AF2 ipTMΔ held-out Protenix-mini binder ipTM
Re-noising with PRES allocationconfirmatory+0.0153 [+0.0079, +0.0256]+0.0418 [+0.0205, +0.0668]
extension+0.0135 [+0.0062, +0.0364]+0.0173 [+0.0092, +0.0280]
Diffusion-ES adaptationconfirmatory+0.0151 [+0.0088, +0.0192]+0.0291 [+0.0010, +0.0580]
extension+0.0092 [+0.0031, +0.0176]+0.0150 [+0.0083, +0.0232]
DTS-L adaptationconfirmatory+0.0139 [+0.0092, +0.0216]+0.0226 [+0.0130, +0.0478]
extension+0.0127 [+0.0065, +0.0183]+0.0112 [+0.0075, +0.0189]

Table 1 shows positive whole-pool AF2 and held-out Protenix-mini binder ipTM differences against all three tested adaptations in both cohorts. The re-noising comparison supports checkpoint reuse under the tested allocation and mutation settings. These results concern whole-pool predicted quality at equal scored-candidate count.

AuthorsOpenReview ↗

Title: Author Response To Reviewer (2/3)

W2: Inconclusive Evidence for the Benefit of Fitness-Guided Parent Selection. One of the paper's central claims is that selecting promising parents based on completed-candidate fitness improves subsequent generation. However, on the additional BindCraft-derived benchmark, the improvement over random-parent branching is not statistically conclusive. The reported 95% confidence interval for the retained-set AF2 ipTM difference includes zero (−0.0145 to +0.0762). Although the main ProtDBench results are favorable, this uncertainty raises questions about the robustness and generalizability of the proposed fitness-guided selection strategy.

We agree that the original BindCraft comparison did not establish a clear average advantage over random-parent branching. That comparison included eight targets and two seeds per target, as stated in the Table 7 in Appendix E. We therefore evaluated two additional independent seeds on each of the same eight targets. When analyzed separately, the additional runs showed a positive mean difference with a 95% bootstrap confidence interval excluding zero. The combined four-seed analysis also favored PRES-NPG, with positive target-level mean differences on all eight targets.

Table 2a. Retained-eight AF2 ipTM, PRES-NPG − random-parent branching, BindCraft panel. Positive values favor PRES-NPG. bold = the interval excludes zero.

AnalysisTargets × seedsPaired target–seed cellsPRES-NPG − random-parent: AF2 ipTM [95% CI]Positive target means
Original submission8 × 216+0.0293 [−0.0145, +0.0762]7/8
New independent seeds only8 × 216+0.0815 [+0.0204, +0.1730]8/8
Combined analysis8 × 432+0.0554 [+0.0155, +0.1193]8/8

The benefit nevertheless varies across runs. CLDN1 illustrates this variability: two seeds favor each method, and its four-seed average advantage is small.

Table 2b. Mean AF2 ipTM of the retained eight candidates for bcCLDN1, by seed.

SeedPRES-NPGRandom-parent branchingPRES-NPG − random-parent
300110.838750.62000+0.21875
300120.305000.70625−0.40125
30013 (additional)0.528750.58625−0.05750
30014 (additional)0.656250.34750+0.30875
Four-seed average0.582190.56500+0.01719

Exploratory lineage analysis of the four CLDN1 seeds found that high AF2 scores clustered within families, while scores among parents with AF2 ipTM below 0.5 showed little association with offspring AF2 scores. This suggests that variability partly reflects which productive lineages emerge and survive selection. Overall, the results strengthen evidence for an average retained-set AF2 gain on this panel.


W3: Incomplete Computational Efficiency Comparison. The primary comparisons use the same number of evaluated candidates rather than equivalent computational budgets. While prefix reuse reduces denoising computation by approximately 60%, the actual end-to-end runtime improvement is only about 1.2 times. It remains unclear whether PRES-NPG would maintain its performance advantage over alternative sampling strategies under matched wall-clock time or GPU budgets.``

We agree that matching the number of scored candidates does not establish equivalent computational cost. In the original 64-candidate comparison, prefix reuse reduced the denoiser-step count by 60%, while the measured end-to-end speedup was 1.20× after including scoring and initialization (Appendix G, p. 22; runtime breakdown in Table 12).

To address this concern, we added a timing-model-adjusted comparison across ten targets and three seeds per target. Using measured pipeline runtimes under paired concurrency settings on shared GPUs, we estimated how many independent candidates could be generated and evaluated within each paired PRES-NPG run’s measured elapsed time. The comparison includes a single-invocation guided baseline, with initialization charged only once.

For each budget-adjusted candidate set, we evaluated the mean AF2 ipTM of the eight highest-scoring candidates, without a structural-diversity constraint. We calculated 95% confidence intervals by bootstrapping target-level paired mean differences after averaging seeds within each target.

AuthorsOpenReview ↗

Title: Author Response To Reviewer (3/3)

Table 3a. Timing-model-adjusted comparison at PRES-NPG’s measured wall-clock budget across ten targets and three seeds.

MethodMedian designs countedTop-eight AF2 ipTMPRES-NPG − baseline [95% CI]
PRES-NPG640.790—
Independent guided, repeated batches480.728+0.0621 [0.0333, 0.0952]
Independent unguided, repeated batches480.683+0.1069 [0.0580, 0.1613]
Independent guided, single invocation620.740+0.0497 [0.0264, 0.0745]

PRES-NPG achieved higher mean retained-eight AF2 ipTM than all three baselines, with all three 95% confidence intervals excluding zero. The advantage persisted against the single-invocation guided baseline (+0.0497; 95% CI [0.0264, 0.0745]). These results support an advantage under the estimated end-to-end time budgets and shared-GPU conditions tested here.


Q1: Methodological Advancement over Prior Work. How does the proposed prefix-reuse strategy differ fundamentally from existing evolutionary diffusion search and trajectory-reuse methods, such as Diffusion-ES and Diffusion Tree Sampling? Beyond avoiding re-noising of completed samples, what specific algorithmic advantage does PRES provide? Could the authors provide a direct comparison with closely related approaches under matched computational budgets?

Please see W1 and Table 1 for the methodological distinctions and comparisons with the Diffusion-ES and DTS-L adaptations. Beyond checkpoint reuse, PRES assigns fixed continuation quotas to selected parents, ensuring that each selected parent receives fresh continuations and making allocation deterministic conditional on parent selection. These search-method comparisons use a matched scored-candidate budget. W3 and Table 3a separately assess independent sampling under timing-model-adjusted wall-clock budgets.


Q2: Statistical Robustness of Fitness-Guided Parent Selection. On the additional BindCraft-derived benchmark, the improvement of PRES-NPG over random-parent branching in retained-set AF2 ipTM has a 95% confidence interval of [−0.0145, +0.0762], which includes zero. Given that fitness-guided parent selection is a central component of the proposed framework, how do the authors interpret this statistically inconclusive result? Would additional targets or independent runs help establish whether fitness-based selection consistently outperforms random selection?

As reported in W2 and Table 2a, we evaluated two additional independent seeds on each of the same eight BindCraft targets. Both the additional-seed analysis and the combined four-seed analysis yielded positive average differences with 95% bootstrap confidence intervals excluding zero. These results strengthen evidence for an average retained-set AF2 ipTM advantage and repeatability across independent runs on the same eight-target panel, although effects vary across seeds. They do not broaden the evaluated target distribution; generalization to additional targets remains to be tested.


Q3: Computational Efficiency under Matched Budgets. The experiments match the number of evaluated candidates rather than total computational cost. Although prefix reuse reduces denoising computation by approximately 60%, the reported end-to-end speedup is only about 1.2 times. Would PRES-NPG still outperform independent guided sampling if both approaches were allocated the same wall-clock time or GPU budget?

Please see W3 and Table 3a. Across ten targets and three seeds, PRES-NPG achieved higher mean retained-eight AF2 ipTM than independent guided sampling under the timing-model-adjusted wall-clock budgets. This advantage persisted against the single-invocation baseline, which incurs initialization only once (+0.0497; 95% CI [0.0264, 0.0745]). The timing accounting includes scoring and initialization.


We sincerely appreciate your deep engagement in reviewing our paper and your valuable feedback. Your suggestions have helped improve our paper in many aspects, and we hope that our detailed responses have addressed each of your concerns.

Reviewer_dSk8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRES-NPG, a test-time framework for improving protein binder generation without retraining pretrained models. NPG uses normalized gradients from a differentiable structural proxy evaluated at the predicted clean structure to guide denoising, while PRES selects parents based on completed-candidate fitness and reuses saved intermediate states to generate new continuations. The method is evaluated using PXDesign-d and Proteina-Complexa, with additional experiments on ProtDBench and BindCraft targets. The results show improvements in predicted interface confidence and several other computational metrics over random-parent selection and independent guided sampling at matched scored-candidate budgets, although these improvements come with reduced structural diversity.

Critical strengths

The separation between local differentiable guidance and terminal-score-based search is well motivated. The paper includes meaningful controls for parent selection and generational search, and evaluates both the retained candidates and the full generated pool. The experiments across different samplers, together with cross-verifier evaluations, provide useful evidence for the method's effectiveness. The authors also provide implementation details and code, discuss the limitations of their guidance approximation, and report actual runtime costs rather than relying only on theoretical generation-step counts.

Critical weaknesses

The methodological novelty appears relatively incremental. NPG shares its core idea with existing clean-estimate and Jacobian-free guidance methods, including the recent QGF work, which the paper itself cites. The main addition is the normalization strategy and relative step size, but the paper does not directly compare normalized guidance with an unnormalized proxy-gradient alternative, making it difficult to isolate the contribution of normalization itself. Similarly, PRES builds on existing evolutionary diffusion search, and the effects of its specific parent-selection and branching choices are not fully explored.

My primary concern is that the improvements do not transfer consistently across evaluation panels. Although the method improves on the AF2-based selection metric, the gains largely disappear on others—on the BindCraft panel, for instance, the retained-set AF2 ipTM improves clearly over independent guided sampling, while the corresponding Protenix-mini ipTM remains almost unchanged. Because AF2 ipTM simultaneously drives parent selection, final candidate filtering, and primary reporting, it is hard to tell how much of the gain reflects a genuine cross-model improvement in design quality rather than optimization toward the selection score. This concern is reinforced by the fact that some reported confidence intervals include zero, particularly for the retained-set comparison against random-parent selection, even though other comparisons do show clearer improvements.

A further concern is the reduction in structural diversity: the method produces higher-scoring but more concentrated candidate sets, and it remains unclear whether the advantage would persist under matched diversity constraints—a point that matters for practical binder discovery, where diverse backbones are often more valuable than higher average confidence.

Questions for authors
  1. Could the authors compare NPG with unnormalized clean-estimate proxy-gradient guidance using the same structural proxy, generator, and comparable tuning and evaluation budgets? This would help isolate the contribution of gradient normalization.
  2. Since AF2 drives selection, filtering, and reporting, how much of the gain reflects genuine cross-model quality rather than optimization toward the selector? Please reconcile the inconsistent results across evaluation panels.
  3. The retained set shows a clear drop in structural clusters and unique designs. Could the authors compare quality under a matched diversity constraint and show that the method yields more high-quality and structurally distinct candidates, which is what matters for binder discovery?
  4. The search uses a fixed number of parents/children and fixed branch positions, and alternative scorers are only tested over a fixed archive rather than by re-running the full search. How sensitive is performance to these choices, and do the gains persist when a non-AF2 evaluator drives full cross-generation parent selection?
  5. The budget is matched by scored-candidate count, not compute; the end-to-end wall-clock advantage is modest once scorer overhead is included. Could the authors compare under an equal wall-clock / GPU-hour budget, where independent sampling can generate and screen more candidates in the same time?
10 comments
AuthorsOpenReview ↗

Title: Author Response To Reviewer (1/6)

Dear Reviewer dSk8,

We thank the reviewer for the constructive feedback, which has helped us improve the clarity and quality of our manuscript. Key changes are highlighted in blue\color{blue}{\text{blue}} in the revised manuscript. Below, we provide point-by-point responses to the reviewer’s comments.


W1: The methodological novelty appears relatively incremental. NPG shares its core idea with existing clean-estimate and Jacobian-free guidance methods, including the recent QGF work, which the paper itself cites. The main addition is the normalization strategy and relative step size, but the paper does not directly compare normalized guidance with an unnormalized proxy-gradient alternative, making it difficult to isolate the contribution of normalization itself. Similarly, PRES builds on existing evolutionary diffusion search, and the effects of its specific parent-selection and branching choices are not fully explored.

Thank you for highlighting these concerns. Our contribution centers on integrating differentiable proxy guidance with non-differentiable search for protein-binder design: NPG uses gradients from a differentiable structural proxy, normalizing each candidate’s gradient and scaling its displacement relative to the base denoising step, while PRES performs gradient-free search by selecting saved intermediate states according to completed-rollout fitness for further guided continuations. We additionally assess these design choices through the following comparisons.

Guidance-magnitude control. To isolate the contribution of gradient normalization, we compare normalized and unnormalized proxy-gradient guidance on ten ProtDBench targets using three paired seeds per target, retaining the same dependence on the base denoising displacement. For steps within the shared guidance window, t∈Wt\in\mathcal W, with ∥gt∥>δg\|g_t\|>\delta_g, the updates are

utNPG=ρ∥Δt∥gt∥gt∥,utunnorm=ρ∥Δt∥gtc.u_t^{\mathrm{NPG}} = \rho\lVert\Delta_t\rVert\frac{g_t}{\lVert g_t\rVert}, \qquad u_t^{\mathrm{unnorm}} = \rho\lVert\Delta_t\rVert\frac{g_t}{c}.

Here, gtg_t is the proxy gradient after coordinate restriction, Δt\Delta_t is the base denoising displacement, and c>0c>0 is a fixed reference gradient magnitude. Both variants use the same liveness threshold δg\delta_g and activation rule as Equation (8): both updates are zero when t∉Wt\notin\mathcal W or ∥gt∥≤δg\|g_t\|\leq\delta_g. Both variants use ρ=0.04\rho=0.04. We set cc to the mean active-gradient norm from separate preliminary calibration runs to approximately match the average guidance-to-denoising magnitude ratio, then hold it fixed across evaluation targets, seeds, and steps.

We hold the generator, structural proxy, guidance window, search configuration, candidate budget, and final selection rule fixed. Both variants retain the same relative-step factor ρ∥Δt∥\rho\|\Delta_t\|. On active steps, only the normalized variant makes the guidance magnitude independent of the instantaneous gradient norm. This comparison therefore isolates the contribution of gradient normalization within the relative-step guidance scheme.

Table 1a. Gradient-normalization ablation with relative-step scaling retained in both variants.

MetricSetNormalizedUnnormalizedNormalized − unnormalized [95% CI]
AF2 ipTMRetained eight0.80610.7962+0.0100 [+0.0018, +0.0257]
AF2 ipTMWhole pool0.49320.4747+0.0185 [+0.0044, +0.0320]
Protenix-mini ipTM — held outRetained eight0.63970.6281+0.0116 [+0.0047, +0.0362]
Protenix-mini ipTM — held outWhole pool0.56600.5570+0.0090 [+0.0034, +0.0266]

Normalized guidance achieved higher mean ipTM scores than the magnitude-calibrated unnormalized baseline for both the retained eight candidates and the full candidate pool, as evaluated by AF2 and held-out Protenix-mini. All four 95% confidence intervals for the differences were strictly positive. With relative-step scaling retained in both variants, these results support the benefit of gradient normalization within the tested guidance scheme.

Parent selection. On the same targets and paired seeds, we compared fitness-guided parent selection with uniform-random parent selection. Both conditions used the same guidance, branch schedule, four continuations per parent, and 64-candidate budget, followed by retention of the eight candidates with the highest AF2 ipTM scores. Only the parent-selection rule differed: PRES selected the four highest-scoring completed trajectories from the preceding generation, whereas the control sampled four distinct parents uniformly without replacement.

AuthorsOpenReview ↗

Title: Author Response To Reviewer (2/6)

Table 1b. Effect of fitness-guided parent selection.

MetricSetFitness-guided parents (PRES-NPG)Random-parent branchingPRES − random [95% CI]
AF2 ipTMRetained eight0.80610.7255+0.0806 [+0.0302, +0.1523]
AF2 ipTMWhole pool0.49320.3420+0.1513 [+0.1085, +0.1943]
Protenix-mini ipTM — held outRetained eight0.63970.6291+0.0106 [+0.0069, +0.0291]
Protenix-mini ipTM — held outWhole pool0.56600.5054+0.0606 [+0.0227, +0.1128]

Fitness-guided parent selection achieved higher mean AF2 and held-out Protenix-mini ipTM scores for both the retained eight candidates and the whole candidate pool. All four 95% confidence intervals for the differences excluded zero. These results support the benefit of fitness ranking over uniform-random parent selection under the fixed configuration and show that the gains extend beyond the AF2 selection metric.

Branching choices. We additionally examined sensitivity to parent–offspring allocation and branch positions on BHRF1, H1, and IL17A using two paired seeds. All configurations used four generations and a fixed budget of 64 completed candidates.

Table 1c. Exploratory sensitivity to parent/offspring allocation and branch positions.

Completed variantRetained-eight AF2 difference versus default
Two parents × eight children−0.0231
Eight parents × two children−0.0383
Earlier branches: 0.70/0.75/0.80−0.0492
Later branches: 0.80/0.85/0.90+0.0052

The default configuration used four parents with four children each and branch positions of 0.70/0.80/0.90. Across the three targets and two seeds, it achieved higher mean AF2 ipTM for the retained eight candidates than the two-parent, eight-parent, and earlier-branch variants. Later branching yielded a small numerical increase of 0.0052 over the default. This exploratory analysis indicates sensitivity to the tested branching configurations but does not establish an optimal choice.


W2: My primary concern is that the improvements do not transfer consistently across evaluation panels. Although the method improves on the AF2-based selection metric, the gains largely disappear on others—on the BindCraft panel, for instance, the retained-set AF2 ipTM improves clearly over independent guided sampling, while the corresponding Protenix-mini ipTM remains almost unchanged. Because AF2 ipTM simultaneously drives parent selection, final candidate filtering, and primary reporting, it is hard to tell how much of the gain reflects a genuine cross-model improvement in design quality rather than optimization toward the selection score. This concern is reinforced by the fact that some reported confidence intervals include zero, particularly for the retained-set comparison against random-parent selection, even though other comparisons do show clearer improvements.

We agree that the original retained-set BindCraft results did not provide clear evidence of cross-verifier improvement. AF2 improvement alone cannot resolve this concern because AF2 ipTM drives parent selection, final retention, and primary reporting. We therefore distinguish two questions: whether AF2-guided search improves scores under held-out predictors, and whether a search advantage persists when a different predictor selects parents. Tables 2a–c address the first question while the full-search experiment in Table 2d addresses the second. The compared arms use a matched budget of 64 scored candidates per run, with eight candidates retained by AF2.

Table 2a. BindCraft seed expansion: retained-eight AF2 ipTM, PRES-NPG minus random-parent branching.

AnalysisTargets × seedsPaired target–seed cellsAF2 ipTM Δ [95% CI]Targets with positive mean differences
Original submission8 × 216+0.0293 [−0.0145, +0.0762]7/8
New independent seeds only8 × 216+0.0815 [+0.0204, +0.1730]8/8
Combined analysis8 × 432+0.0554 [+0.0155, +0.1193]8/8

The original confidence interval included zero, whereas the new-seed-only and combined intervals exclude zero. These results strengthen the evidence for the retained-set AF2 advantage. Transfer to other predictors is assessed separately below. The expanded BindCraft results in Table 2b use the combined four-seed analysis.

AuthorsOpenReview ↗

Title: Author Response To Reviewer (3/6)

Table 2b. AF2 and held-out Protenix-mini evaluation across panels. ProtDBench uses ten targets and three paired seeds. Expanded BindCraft uses eight targets and four paired seeds. Retained sets contain the eight candidates selected by AF2 ipTM from 64 scored candidates. Δ denotes PRES-NPG minus the comparator.

PanelComparatorEvaluated setAF2 ipTM Δ [95% CI]Held-out Protenix-mini ipTM Δ [95% CI]
ProtDBenchIndependent guided samplingRetained eight+0.0553 [+0.0232, +0.1033]+0.0127 [+0.0062, +0.0309]
ProtDBenchIndependent guided samplingWhole pool+0.1349 [+0.0939, +0.1721]+0.0579 [+0.0173, +0.1100]
ProtDBenchRandom-parent branchingRetained eight+0.0806 [+0.0302, +0.1523]+0.0106 [+0.0069, +0.0291]
ProtDBenchRandom-parent branchingWhole pool+0.1513 [+0.1085, +0.1943]+0.0606 [+0.0227, +0.1128]
BindCraft, expandedIndependent guided samplingRetained eight+0.0624 [+0.0139, +0.1281]+0.0179 [+0.0078, +0.0280]
BindCraft, expandedIndependent guided samplingWhole pool+0.0739 [+0.0492, +0.1016]+0.0189 [+0.0034, +0.0414]
BindCraft, expandedRandom-parent branchingRetained eight+0.0554 [+0.0155, +0.1193]+0.0193 [+0.0073, +0.0458]
BindCraft, expandedRandom-parent branchingWhole pool+0.0736 [+0.0446, +0.1030]+0.0267 [+0.0021, +0.0549]

All eight held-out Protenix-mini contrasts are positive, with 95% confidence intervals excluding zero. The expanded BindCraft results strengthen both the retained-set AF2 evidence and held-out transfer: the retained-set Protenix-mini gains are +0.0179 over independent guided sampling and +0.0193 over random-parent branching. These gains were numerically smaller than the corresponding AF2 gains but provide evidence of improvement under a predictor not used for parent selection or final retention in these comparisons. Whole-pool gains further show that improvement extends beyond final top-eight filtering, while still reflecting AF2-guided parent selection during search.

Using corrected target-complex inputs, we additionally evaluated the AF2-retained eight candidates and the complete candidate pool with Boltz-2, which was not used for parent selection or final retention. All scores are binder-specific ipTM.

Table 2c: Boltz-2 binder-specific ipTM on ProtDBench. Δ denotes PRES-NPG minus the comparator.

Evaluated setComparatorPRES-NPGComparator scoreΔ [95% CI]
Retained eightIndependent guided sampling0.83210.8035+0.0286 [+0.0118, +0.0493]
Retained eightRandom-parent branching0.83210.7965+0.0356 [+0.0158, +0.0549]
Whole poolIndependent guided sampling0.75230.7003+0.0520 [+0.0288, +0.0835]
Whole poolRandom-parent branching0.75230.6928+0.0595 [+0.0299, +0.1011]

All four Boltz-2 contrasts were positive, with 95% confidence intervals excluding zero. Together, Tables 2b–c support transfer of the observed score improvements to Protenix-mini across both evaluated panels and to Boltz-2 on ProtDBench.

To test dependence on the parent-selection evaluator, we reran the complete search on ten ProtDBench targets with three seeds per target, using Protenix-mini to select four parents at each branching generation. Guidance, continuation schedule, the 64-candidate budget, and final AF2 retention remained fixed.

Table 2d. Full-search parent-scorer swap evaluated by AF2 ipTM.

Evaluated setComparatorProtenix-mini-parent PRES − comparator [95% CI]
Retained eightAF2-parent PRES-NPG−0.0413 [−0.0845, −0.0091]
Whole poolAF2-parent PRES-NPG−0.0585 [−0.0885, −0.0257]
Retained eightIndependent guided sampling+0.0140 [+0.0039, +0.0240]
Whole poolIndependent guided sampling+0.0763 [+0.0356, +0.1262]
Reviewer_weYkOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PRES-NPG, a framework that combines strengths of Prefix-Reuse Evolutionary Search and Normalized Proxy Guidance for binder design.

Critical strengths
  1. The work is well-movatived.
  2. The NPG is designed simply and has theoretical basis.
  3. The experiment is quite comprehensive.
Critical weaknesses
  1. The argument 'intermediate states are unreliable' lacks experimental support, which makes the argument unconvincing.
  2. Lacking report on checkpoint memory overhead.
Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Title: Author Response To Reviewer (1/2)

Dear Reviewer weYk,

We sincerely appreciate your valuable feedback and provide point-wise responses below. Key changes are highlighted in blue\color{blue}{\text{blue}} in the revised manuscript.


W1: The argument 'intermediate states are unreliable' lacks experimental support, which makes the argument unconvincing.

We thank the reviewer for requesting supporting evidence. We agree that our original statement about intermediate-state evaluation was too broad and have narrowed it. Across ten ProtDBench targets, completed-candidate AF2 scores predicted continuation quality more strongly than the tested inexpensive structural proxy at all three branch depths. Intermediate AF2 after sequence design was also predictive. These results support using completed-candidate fitness to rank saved prefixes in this setting. They do not establish that intermediate evaluators are generally unreliable.

We evaluated 16 saved prefixes at depths 0.70, 0.80, and 0.90 for each of ten ProtDBench targets and three seeds. Each prefix’s continuation quality was the mean AF2 ipTM of four fresh continuations, excluding the original completed rollout used as a ranking signal. The prefixes comprised the entire generation-0 population before fitness-based parent selection. The same 16 trajectories supplied prefixes at all three depths. Spearman correlations were computed within target–seed–depth cells using average ranks for ties, averaged over seeds, and weighted equally across targets. Paired differences were formed within cells using identical prefixes and continuation outcomes. Unadjusted 95% percentile CIs used 20,000 target-level bootstrap draws (seed 0), preserving pairing and nested observations. All prefixes had the required four scored continuations, included regardless of quality-filter status.

Table 1. Predictive performance on saved PRES-NPG prefixes.

Ranking signal or paired contrast0.700.800.90
Intermediate structural proxy: Spearman ρ+0.015−0.045−0.016
Intermediate AF2 after sequence design: Spearman ρ+0.183+0.444+0.528
Completed-rollout AF2: Spearman ρ+0.299+0.425+0.613
Completed AF2 − structural proxy: Δρ [95% CI]+0.285 [+0.181, +0.403]+0.470 [+0.286, +0.654]+0.629 [+0.475, +0.785]

Completed-rollout AF2 scores predicted continuation AF2 quality positively at every tested depth, reaching a mean ρ of 0.613 at depth 0.90. Their advantage over the structural proxy was supported at all three depths, with every paired confidence interval excluding zero. Intermediate AF2 scores after sequence design also predicted continuation quality and had a numerically higher mean correlation than completed-rollout AF2 at depth 0.80. PRES obtains a useful ranking signal by reusing completed-candidate evaluations already required for final retention, avoiding additional intermediate sequence-design and verifier calls.

QGF [1] provides related evidence from reinforcement learning: guidance using critic gradients at noisy intermediate actions produced high predicted Q-values but poor task performance, whereas guidance at estimated denoised actions performed better. This supports caution when applying evaluators to intermediate states, although it does not directly test completed-rollout versus intermediate signals for prefix selection.

W2: Lacking report on checkpoint memory overhead.

We thank the reviewer for highlighting this omission. We now report this overhead in Table 2. In the measured generation process, checkpointing increased peak GPU memory by 1.3–4.3 MiB, with no further increase from branching or PRES-NPG in these runs. CPU changes were small (−4.5 to +23.9 MiB) and inconsistent, preventing attribution to checkpointing. Parent selection added at most 10.6 MiB to driver memory and took ≤0.09 s per run. PRES-NPG checkpoint storage was 15.69 MiB per run at the median. These measurements cover three targets and one seed.

These measurements characterize incremental checkpoint overhead in the generation process during one denoising batch. They do not establish the peak memory requirement of the complete generation, sequence-design, and scoring pipeline. Checkpoint storage is reported separately from CPU and GPU memory.

Table 2. Memory and checkpoint storage.

ConfigurationPeak GPU (GiB)Peak CPU (GiB)Files/runMiB/fileMiB/run
Independent guided sampling, checkpoints off2.84 (1.90–5.71)3.82 (3.81–3.82)0—0
Independent guided sampling, checkpoints on2.84 (1.90–5.72)3.82 (3.81–3.83)61.74 (1.03–3.45)10.46 (6.18–20.70)
Random-parent branching2.84 (1.90–5.72)3.82 (3.81–3.82)91.74 (1.03–3.45)15.69 (9.27–31.05)
PRES-NPG2.84 (1.90–5.72)3.82 (3.81–3.84)91.74 (1.03–3.45)15.69 (9.27–31.05)
AuthorsOpenReview ↗

Title: Author Response To Reviewer (2/2)

Median (minimum–maximum) over TrkA, VEGFA and TNFa. Seed 30011, 4 × 16 candidates, 400 denoising steps, bf16 and one RTX A6000 per run. GPU and CPU peaks are the generation process’s CUDA-allocator and resident-set peaks during one denoising batch, respectively.

These measurements characterize incremental checkpoint overhead in the generation process during one denoising batch. They do not establish the peak memory requirement of the complete generation, sequence-design, and scoring pipeline.

[1] Test-Time Gradient Guidance of Flow Policies in Reinforcement Learning. NeurIPS, 2026.


We appreciate the reviewer’s insight, which has helped us refine both our positioning and the clarity of our claims.

Thank you for your response. I will keep my original score of 3.

Dear Reviewer weYk,

We respect your assessment and sincerely appreciate your thoughtful feedback, which has helped us strengthen the clarity and experimental support of our work.

Thank you again for your time and careful review.

Best regards,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions and Evidence for the Area Chair and Reviewers

Dear Area Chair and Reviewers,

We sincerely thank the reviewers for their constructive feedback. In response, we have substantially strengthened the manuscript with new controlled ablations, expanded cross-verifier evaluations, alternative search baselines, computational cost analyses, and a systematic investigation of structural diversity. Below, we summarize the key new evidence and revisions.

1. New controlled ablations directly validate the core methodological contributions.

We added a controlled comparison between Normalized Proxy Guidance (NPG) and magnitude-calibrated unnormalized guidance, holding the generator, proxy, guidance schedule, and evaluation budget fixed. NPG consistently improves both retained-set and whole-pool interface confidence under AF2 and held-out Protenix-mini evaluation, with confidence intervals supporting all reported improvements. dSk8(W1/Q1)

We also directly isolated fitness-guided parent selection from random-parent branching. PRES achieves consistently higher predicted quality under both AF2 and held-out Protenix-mini, demonstrating that completed-fitness-based selection contributes beyond gradient guidance and final candidate filtering. dSk8 (W1) czSb(W2/Q2)

2. Expanded evaluations provide stronger evidence of cross-verifier generalization and robustness.

We expanded the BindCraft evaluation with additional independent seeds, strengthening the statistical evidence for fitness-guided parent selection and resolving uncertainty in the original comparison. dSk8(W2) czSb(W2/Q2)

Importantly, held-out Protenix-mini and Boltz-2 evaluations demonstrate that improvements extend beyond the AF2 predictor used for search and final selection. The improvements persist in both retained candidates and complete generation pools, providing evidence that the gains are not solely attributable to final filtering. dSk8(W2/Q2)

We further reran the complete evolutionary search using Protenix-mini for parent selection. The resulting search retains an advantage over independent guided sampling, although its performance is sensitive to the choice of scorer. Experiments with Proteina-Complexa additionally support the applicability of PRES-NPG across different frozen generative frameworks. dSk8(Q4)

3. Additional experiments clarify the benefits of completed-rollout selection and prefix reuse.

We directly compared intermediate structural proxies, intermediate AF2 evaluations, and completed-rollout fitness for predicting the quality of future continuations. Completed-rollout fitness provides a substantially stronger ranking signal than the inexpensive intermediate structural proxy, supporting the design choice of selecting saved prefixes using observed terminal outcomes. weYk(W1)

We also introduced comparisons against re-noising, Diffusion-ES, and Diffusion Tree Sampling-inspired search adaptations under matched candidate budgets. PRES-NPG consistently achieves higher whole-pool predicted quality under both AF2 and held-out Protenix-mini, supporting the effectiveness of completed-fitness-driven prefix reuse. czSb(W1/Q1)

4. New computational and diversity analyses clarify practical trade-offs.

We added checkpoint-memory measurements and storage overhead analysis, as well as measured end-to-end runtime results that distinguish denoising-step savings from actual computational efficiency. A further timing-model-adjusted comparison supports an advantage over independent guided sampling under the evaluated elapsed-time budgets, including a stronger single-invocation baseline that avoids repeated initialization. weYk(W2) czSb(W3/Q3) dSk8(Q5)

We also directly investigated the quality–diversity trade-off. Although PRES-NPG improves unconstrained predicted confidence and total passing yield, its performance under a strict structural-diversity constraint is benchmark-dependent. Per-target analyses and branching diagnostics identify concentration within existing structural modes as an important limitation. dSk8 (W3/Q3, follow-up)

AuthorsOpenReview ↗

Summary of Revisions and Evidence for the Area Chair and Reviewers

To address this, we evaluated broader parent selection and explore-first, diversity-aware polishing. The exploratory results show that reallocating more computation to independent exploration can substantially improve structural diversity while maintaining competitive predicted confidence. We explicitly distinguish this alternative from the original configuration and clarify the scope of our conclusions. dSk8 (W3/Q3, follow-up)

Finally, we strengthened the theoretical discussion, algorithmic specification, experimental reporting, and reproducibility details.


We appreciate the reviewers' suggestions, which have substantially strengthened the manuscript. The expanded evidence directly supports the contributions of normalized guidance and fitness-guided prefix reuse, demonstrates improvements across independent predictors and generative frameworks, and provides a more rigorous understanding of computational efficiency and structural diversity.

We respectfully ask the Area Chair and reviewers to consider these additional experiments, methodological controls, and clarifications in their final assessment.

Thank you for your time and consideration.

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

Related papers

Discussion (0)

Sign in to comment.