acceptodds
Under review as a conference paper at ICLR 2027

ReachDiff: Future-Feasibility-Guided Search for Constrained Cyclic Peptide Design

Abstract

Designing cyclic peptides requires several coupled requirements to be satisfied simultaneously, including chemical validity, cyclization-topology realization, and closure geometry. Their intersection defines a substantially narrower design space than any individual requirement, making efficient exploration particularly important. Most generative pipelines nevertheless spend the majority of sampling computation before determining whether a completed molecule satisfies the joint feasibility criteria. We introduce **ReachDiff**, a masked-diffusion approach that predicts future feasibility, defined as the probability that a partial molecular state will satisfy the benchmark joint-feasibility objective when generation continues with the frozen base model. A lightweight predictor scores competing partial states once during generation. The highest-scoring states receive the remaining continuation budget, after which standard denoising resumes without further guidance. Across head-to-tail, disulfide, and isopeptide-lactam cyclization, ReachDiff improves joint feasibility from 59.61% to 77.59% under an equal terminal-evaluation budget, and the advantage is retained under matched recorded model work. Future feasibility is predictable early in denoising, and controlled analyses attribute the gain to state-dependent ranking and global competition. ReachDiff also collects verifier-positive trajectories more efficiently, which can be reused for denoising alignment to improve search-free generation. Together, these results support future-feasibility-guided allocation as an effective use of finite inference compute.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cyclic peptide generation, future-feasibility prediction, inference-time compute allocation, constrained molecular generation, test-time search

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

ReachDiff trains a lightweight predictor on a frozen diffusion generator’s intermediate representations to estimate whether a generated molecule will ultimately satisfy three conditions: chemical validity, the desired cyclization topology, and feasible closure geometry. It then uses a single global selection step to decide which partial generation trajectories should receive additional computation. Successful trajectories can optionally be used to fine-tune the generator, which is subsequently evaluated with ordinary diffusion and no search. The paper reports improved feasible-molecule discovery under both equal terminal-evaluation budgets and matched total-computation budgets.

Critical strengths
  1. Simple, practical design. A single selection step improves constrained generation without modifying the generator during search.
  2. Useful controlled experiments. Random and shuffled-score controls, global versus per-parent selection, and independent structural checks provide evidence that the learned ranking contributes to the gains.
Critical weaknesses
  1. The contribution beyond existing value-guided diffusion is insufficiently established. SVDD (2408.08252) is closely related and cited, but no direct comparison is reported.
  2. Robustness is insufficiently characterized. The results establish gains for a fixed feasibility objective and sampling configuration, but leave sensitivity to hyperparameters, changed constraints, and predictor reuse unclear.
Questions for authors
  1. Comparison with SVDD. SVDD-MC already learns intermediate-state values from terminal rewards. With binary feasibility rewards, its prediction target closely resembles ReachDiff’s. Could the authors compare against SVDD-MC using the same generator, verifier, training data, and total inference budget? This would clarify whether the contribution comes from the predictor, global competition, or reduced guidance frequency.

  2. Hyperparameter sensitivity. What motivates six survivors, up to twelve candidates, five training snapshots, and the 0.70/0.65 mask thresholds? Please provide targeted sensitivity results under matched compute. The synchronized-scoring control addresses timing variation, but does not establish robustness to substantially different scoring times or branching budgets.

  3. Feasibility versus candidate quality. The paper appropriately reports both feasibility ratios and unique discovery counts. However, binary supervision cannot distinguish barely feasible products from stronger candidates. How do gains change across stricter thresholds, and what are the quality distributions among passing products? The decrease from 77.59% benchmark feasibility to 21.1% strict feasibility makes this distinction consequential.

  4. Changing objectives and total cost. If thresholds change or new evaluation dimensions are added, can the existing predictor retain useful rankings, or must trajectories be relabeled and the predictor retrained? Please include the cost of collecting the 6,000 trajectories, evaluating labels, and training the predictor, and indicate when this initial investment is recovered through more efficient generation.

  5. Contribution of generator alignment. The aligned model uses ordinary diffusion without the original search procedure. Since other positive-trajectory collectors also improve generation, and ReachDiff alignment reaches 68.52% versus 67.80% for TR2-D2 alignment, is the main downstream benefit cheaper trajectory acquisition rather than a distinctly better-aligned generator?

2 comments
AuthorsOpenReview ↗

Response to Reviewer zBfP (1/2)

Dear Reviewer zBfP,

We thank the reviewer for the constructive feedback and for recognizing the practical design and controlled experiments of ReachDiff. We have conducted additional experiments to clarify its methodological contribution, robustness, and computational efficiency.

W1 & Q1. Comparison with SVDD-MC and Methodological Contribution

To directly address the relationship with value-guided diffusion, we implemented SVDD-MC with stepwise Monte Carlo guidance and evaluated it using the same frozen generator, verifier, training data, and recorded-work budget across five seeds and nine conditions.

MethodUnique Class-Feasible ↑Unique Strict-Feasible ↑
Independent Rejection1,841476
SVDD-MC (M=5)496152
One-shot Per-parent (same MSE critic)1,589421
One-shot Global (same MSE critic)2,015526
ReachDiff2,024546

ReachDiff achieves 4.08× more unique class-feasible and 3.59× more strict-feasible discoveries than SVDD-MC under matched inference computation. SVDD-MC obtains higher feasibility among completed outputs but incurs substantial overhead from repeated candidate expansion and value evaluation. ReachDiff instead makes one global allocation decision and resumes ordinary denoising, translating more computation into successful discoveries.

To isolate the allocation strategy, we further compared methods using exactly the same MSE-trained critic. Global allocation improves class-feasible discovery from 1,589 to 2,015 (+26.8%; paired gain +85.2 per seed, 95% CI [75.4, 95.6]) and strict-feasible discovery from 421 to 526 (+24.9%). Its performance also closely matches the original ReachDiff (2,015 vs. 2,024).

Additional frequency controls reinforce this finding: under a common recorded-work budget, one-, two-, and three-stage allocation yield 3,161, 2,413, and 2,387 unique feasible products, respectively.

These results identify the principal contribution of ReachDiff: efficient global allocation of continuation computation, rather than the value-prediction objective itself.

W2 & Q2. Hyperparameter Sensitivity and Robustness

Our default configuration uses six survivors, up to twelve candidates, and one global allocation stage. The timing is motivated by early future-feasibility predictability: AUROC increases from 0.798 at 10% diffusion progress to 0.881 at 30%, with limited subsequent improvement. The 0.70/0.65 mask thresholds balance informative scoring with sufficient remaining computation.

We additionally performed a matched-work sensitivity analysis varying survivor count, candidate expansion, and scoring time on the original validation split.

ConfigurationUnique Feasible / Query ↑
B=6, K=2, mask 0.904.778
B=6, K=2, mask 0.705.000
B=6, K=2, mask 0.505.000
B=6, K=2, mask 0.304.111
B=3, K=2, mask 0.702.519
B=12, K=2, mask 0.703.889
B=6, K=4, mask 0.705.259
Independent rejection5.259

The results characterize the trade-off between candidate exploration, scoring time, and completion cost. Later scoring and larger intermediate search budgets can reduce completed discoveries under fixed computation. On this small validation panel, differences relative to rejection sampling are not statistically resolved; we retain the original test configuration.

The five training snapshots represent mask levels 0.9, 0.7, 0.5, 0.3, and 0.1 sampled from each trajectory, providing supervision across denoising. Snapshot-count sensitivity remains to be evaluated separately.

We also replicated the analysis using an independently frozen 5k generator checkpoint. ReachDiff retains 36.41% and 35.97% improvements over random and shuffled-score allocation, respectively, supporting robustness across generator checkpoints.

Q3. Feasibility versus Candidate Quality

We extended the evaluation to stricter topological and geometric requirements, keeping the generator, predictor, and search procedure unchanged.

The strict-topology verifier additionally enforces exact residue counts, terminal/cap annotations, positional anchors, linkage atoms, and the intended ring bond. Under matched recorded computation, ReachDiff discovers 149.4 unique strict-feasible products per seed, compared with 128.6 for independent rejection sampling (+16.2%; paired 95% CI for the difference [6.0, 39.4]).

We further evaluated saved conformers using progressively tighter MMFF94-based criteria covering bond/closure geometry, steric clashes, angular deviations, and energy thresholds.

AuthorsOpenReview ↗

Response to Reviewer zBfP (2/2)

Structural CriterionIndependent SamplingReachDiff
Original physical feasibility42.78%55.56%
Tighter physical feasibility42.50%55.56%
Stringent physical feasibility8.89%13.89%

Under the most stringent criteria, ReachDiff achieves a 56.3% relative improvement (+5.00 percentage points; 95% CI [1.11, 8.33]).

We also examined the distributions of energy and structural deviations among evaluated conformers, retaining all failed structural evaluations in the overall feasibility denominator. These additional criteria were excluded from predictor training.

The results demonstrate that ReachDiff's discovery advantage persists under substantially stricter structural criteria without predictor retraining or changes to inference.

Q4. Changing Objectives, Predictor Reuse, and Total Cost

We evaluated whether the original predictor remains informative when the objective changes from benchmark feasibility to strict topology, and compared it with a strict-label predictor trained using cached intermediate-state features.

MetricFrozen PredictorStrict-Label Retraining
Strict-label AUROC ↑0.69790.7214
Within-condition ranking accuracy ↑0.67400.7237
Top-quarter strict-feasible yield ↑0.21950.2610
Brier score ↓0.39230.1251

The original predictor retains useful ranking ability under the stricter objective, reaching 0.6979 AUROC and 0.6740 within-condition ranking accuracy without retraining. Strict-label adaptation further improves the point estimates, although the paired intervals for the ranking improvements overlap zero.

Importantly, retraining the lightweight head using cached features required only 6.11 CPU seconds, without additional generator rollouts or backbone updates.

We also audited the original predictor's data and computational costs. The 6,000 trajectories produce 30,000 intermediate observations, with recorded collection and overlapping-labeling jobs totaling 2,588.25 shard-seconds.

On a separate matched-online-work panel, ReachDiff improves class-feasible discovery from 1,505 to 1,638 and strict-feasible discovery from 403 to 460 relative to independent rejection.

A collection-component timing extrapolation suggests amortization at approximately 10,716 class-feasible or 2,126 strict-feasible unique discoveries. These are conditional estimates; the original critic-training timing was not separately recorded, preventing a complete end-to-end break-even estimate.

These findings show that the learned predictor can retain useful information across related constraints, while objective-specific adaptation can be achieved efficiently through cached-feature reuse.

Q5. Contribution of Generator Alignment

Our additional analysis clarifies that ReachDiff's primary downstream benefit is more efficient acquisition of verifier-positive trajectories for generator alignment.

We compared three collectors using 384 positive trajectories and an identical subsequent alignment schedule.

CollectorVerifier Calls ↓Acquisition Work ↓Aligned Joint@1 ↑
Independent Sampling7621.1261 × 10^1566.87%
TR2-D2-style7921.3581 × 10^1567.80%
ReachDiff5700.9508 × 10^1568.52%

Compared with TR2-D2-style collection, ReachDiff requires 28.03% fewer verifier calls and approximately 29.99% less recorded acquisition work.

All positive-trajectory sources improve the standalone generator over the base Joint@1 of 60.80%. ReachDiff achieves the highest point estimate (68.52%), while its +0.72 percentage-point difference over TR2-D2 is not statistically resolved (95% CI [−4.32, 5.56]).

Thus, the principal contribution is acquiring useful denoising trajectories at lower computational cost, enabling effective alignment through the same standard training objective. The aligned generator subsequently operates with ordinary one-pass diffusion, without search or critic evaluation.

These additional experiments strengthen the evidence for ReachDiff's methodological contribution and practical efficiency. The SVDD-MC comparison and same-critic controls establish one-shot global allocation as the key driver of improved compute efficiency, while stricter structural evaluations and predictor-transfer experiments demonstrate its usefulness beyond the original feasibility objective. Together with more efficient positive-trajectory acquisition, these results support ReachDiff as an effective approach to allocating limited diffusion computation for constrained cyclic-peptide discovery.

Reviewer_o33DOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed ReachDiff, a discrete masked diffusion inference optimization method for de novo cyclic peptide design. To address the computational waste caused by multiple restrictions in cyclic peptide design, the authors trained a lightweight predictor to assess final "reachablility" of mid-state diffusion outputs, and re-allocated computing budgets by beam search, allowing high efficiency of cyclic peptide generation when denoting peptide length and cyclization mode.

Critical strengths

The paper discussed and verified that success or failure signals are already known in early diffusion steps. It replaced best-of-N search with beam search empowered by a trained lightweight confidence predictor, significantly outperforms the baselines with same computational costs.

Critical weaknesses
  1. The generation conditions are too weak and might deviate from practical drug development scenarios, if only based on cyclization mode and chain length. And it lacks capacity of target-conditioned pockets design or motif scaffolding or simple conformation folding based on sequence.
  2. The geometric metics are too lenient, and conformation generation relies significantly on RDKit energy minimization, lacking assessment of energy distribution or other physical or chemical metrics.
  3. Innovation is limited and lacks direct comparison with classic value guidance methods whose workflow is similar to itself such as SVDD.
Questions for authors
  1. Is this method capable of transmitting to more practical cyclic peptide design tasks such as pocket-conditioned and motif-constrained design?
  2. How about the physical and chemical properties of generated conformations? Please provide more practical metrics and employ a bioforce field more suitable for cyclic peptide.
7 comments
AuthorsOpenReview ↗

Response to Reviewer o33D (1/2)

Dear Reviewer o33D,

We thank the reviewer for the constructive feedback and for recognizing the effectiveness of early feasibility prediction and compute-efficient search. We have conducted additional experiments to strengthen the practical relevance, structural evaluation, and methodological contribution of ReachDiff.

W1 & Q1. Practical Design Constraints and Generalizability

We appreciate the reviewer's suggestion to investigate more demanding cyclic-peptide design settings.

The current benchmark evaluates three cyclization chemistries—head-to-tail, disulfide, and isopeptide-lactam—under joint chemical, topological, and geometric constraints. To examine more demanding requirements, we extended the evaluation to strict topology, additionally enforcing exact residue counts, terminal/cap annotations, positional anchors, correct linkage atoms, and the intended ring bond.

Under matched recorded computation across five seeds and nine conditions, ReachDiff discovers 149.4 unique strict-feasible products per seed, compared with 128.6 for independent rejection sampling. This represents a 16.2% improvement, with a paired difference of +20.8 products per seed (95% CI [6.0, 39.4]).

Importantly, these additional constraints were not included in the original predictor's training objective, yet the discovery advantage persists without modifying the generator or search procedure.

We further investigated adaptation to changed objectives. When evaluated against strict-feasibility labels, the original frozen predictor achieves 0.6979 AUROC and 0.6740 within-condition ranking accuracy. Retraining a lightweight head using cached representations improves these point estimates to 0.7214 and 0.7237, respectively, requiring only 6.11 CPU seconds and no additional generation. This illustrates how the existing representation can be reused when the acceptance criteria change.

Regarding motif- and pocket-conditioned design, ReachDiff's allocation mechanism is compatible in principle with richer conditional generators. It operates on intermediate generative states and terminal feasibility estimates, so the same allocation strategy can be applied when the underlying generator supports additional structural or target-specific conditions. The current experiments establish this mechanism for topology-conditioned generation and stricter structural criteria.

These findings support the extensibility of future-feasibility-guided allocation to increasingly constrained molecular design tasks.

W2 & Q2. Geometric Evaluation and Physical Quality

To further assess the structural quality of generated cyclic peptides, we conducted additional evaluations using progressively stricter MMFF94-based molecular-mechanics criteria, covering closure/bond geometry, nonlocal steric clashes, angular deviations, and energy thresholds.

All methods were evaluated using the same frozen structural protocol, and failed conformer construction or relaxation attempts were retained in the overall feasibility denominator.

Structural CriterionIndependent SamplingReachDiff
Original physical feasibility42.78%55.56%
Tighter physical feasibility42.50%55.56%
Stringent physical feasibility8.89%13.89%

ReachDiff consistently achieves higher structural feasibility across all three evaluation levels. Under the most stringent criteria, it achieves a 56.3% relative improvement (+5.00 percentage points; 95% CI [1.11, 8.33]).

We additionally analyzed the distributions of molecular-mechanics energy and structural deviations among successfully evaluated conformers, including nonlocal distances and geometry violations. These analyses complement the all-output feasibility measurements by characterizing the structural properties of generated candidates.

Notably, the relative advantage increases from 29.9% under the original physical criteria to 56.3% under the stringent criteria, demonstrating that ReachDiff remains effective as structural acceptance becomes more restrictive.

These results extend our evaluation beyond the original geometric thresholds. The present physical analysis uses MMFF94 rather than a peptide-specific biomolecular force field; accordingly, our conclusions concern computational structural consistency rather than thermodynamic stability or biological activity.

AuthorsOpenReview ↗

Response to Reviewer o33D (2/2)

W3. Methodological Novelty and Comparison with SVDD

We thank the reviewer for highlighting the relationship between ReachDiff and existing value-guided diffusion methods. To directly address this concern, we implemented SVDD-MC with stepwise Monte Carlo value guidance and evaluated it using the same frozen generator, feasibility verifier, training data, and inference-compute budget across five seeds and nine conditions.

MethodUnique Class-Feasible ↑Unique Strict-Feasible ↑
Independent Rejection1,841476
SVDD-MC (M=5)496152
One-shot Per-parent (same MSE critic)1,589421
One-shot Global (same MSE critic)2,015526
ReachDiff2,024546

ReachDiff achieves 4.08× more unique class-feasible and 3.59× more unique strict-feasible discoveries than SVDD-MC under matched inference computation.

SVDD-MC achieves higher feasibility among completed samples through repeated value-guided candidate selection, but its stepwise expansion incurs substantial computational overhead. ReachDiff performs a single global allocation and then completes selected trajectories through ordinary denoising, enabling more efficient use of the available inference budget.

To further isolate the methodological contribution, we conducted a same-critic comparison, using exactly the same MSE-trained value predictor for global and per-parent allocation.

Global allocation increases unique class-feasible discovery from 1,589 to 2,015 (+26.8%), with a paired improvement of +85.2 products per seed (95% CI [75.4, 95.6]). Strict-feasible discovery similarly increases from 421 to 526 (+24.9%).

Furthermore, one-shot global allocation with the MSE critic achieves nearly identical class-feasible discovery to the original ReachDiff (2,015 versus 2,024).

These experiments establish that ReachDiff's main methodological contribution lies in the global allocation of continuation computation, rather than the particular value-prediction objective. By exploiting future-feasibility information through a single cohort-level decision, ReachDiff achieves substantially higher feasible-discovery efficiency than repeated local value guidance under the evaluated compute budget.

We thank the reviewer again for the valuable suggestions. The additional experiments strengthen the evidence for ReachDiff's effectiveness under stricter molecular constraints, its structural consistency, and its computational advantages over existing value-guided diffusion. The direct SVDD-MC comparison and same-critic controls establish global continuation-budget allocation as the principal methodological contribution, while the stricter topology and geometric evaluations demonstrate that the discovery gains extend beyond the original feasibility criteria. Together, these findings support ReachDiff as an efficient and practical approach to constrained cyclic-peptide generation.

Thank you very much for your detailed rebuttal and the effort put into the additional experiments. I appreciate the newly added comparison with SVDD-MC, the same-critic global vs. per-parent controls, and the stricter topology and MMFF94 checks, which help clarify how the proposed allocation strategy behaves.

However, after carefully reviewing your responses, I feel that my main concerns regarding practical applicability and physical/chemical validation remain largely unresolved for the current submission:

  1. Practical design tasks (W1, Q1): While the strict-topology evaluation adds useful structural checks, the generation setup is still limited to length and cyclization-class conditions without target/pocket conditioning, motif scaffolding, or functional property objectives. As such, it remains somewhat distant from practical cyclic-peptide drug discovery scenarios.
  2. Physical quality, biomolecular force fields, and physicochemical metrics (W2, Q2): I appreciate your candid acknowledgment that the expanded physical evaluation still relies on RDKit/MMFF94 rather than a peptide-specific biomolecular force field, and that quantitative conformational energy distributions and practical physicochemical/drug-likeness metrics are not yet included. Moreover, the sharp drop in pass rate to 13.89% under the Stringent physical feasibility criterion (vs. 8.89% for independent sampling) suggests that a large fraction of generated candidates still struggle under stricter geometric/energetic checks.
  3. Methodological positioning under matched compute (W3): While ReachDiff outperforms SVDD-MC (M=5) (2,024 vs. 496), SVDD-MC (M=5) falls well behind plain Independent Rejection (1,841) under the same compute budget due to stepwise branching overhead. Against compute-matched Independent Rejection, the gain in unique discoveries is more modest (~10%, 2,024 vs. 1,841, before factoring in the upfront cost of collecting 6,000 trajectories and training the critic), and a simple MSE critic with one-shot global selection (2,015) performs similarly to ReachDiff (2,024).

That said, I find the core empirical observation—that partial diffusion states already carry useful feasibility signals for global compute allocation—genuinely interesting. I strongly encourage the authors to extend this framework in future work to more realistic cyclic-peptide design tasks (such as pocket-conditioned binding or motif-constrained generation) and to incorporate peptide-appropriate biomolecular force fields (e.g., AMBER, CHARMM, Rosetta, or GFN2-xTB) alongside comprehensive physicochemical and energetic evaluations.

Given that the above concerns remain open in the current manuscript, I will maintain my original rating (Rating: 1), and I hope these suggestions are helpful for further strengthening this line of work.

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

The paper targets applications in cyclical peptide design, a task which is inherently positioned within a constrained subspace. To enable efficient exploration of this space, the paper introduces ReachDiff, a masked diffusion approach which predicts final state feasibility during the sampling process. This prediction acts as an early signal, which filters unsuccessful designs prior to investing the entire budgeted inference-time compute. Empirical results demonstrate that designs which pass these filters report a higher feasibility rate across various classes of cyclical peptides.

Critical strengths
  • Motivation: The application is well motivated; design pipelines result is significant overhead, generally relying on rejection sampling. Hence, reducing the required rollout time through this early exiting condition seems potentially significant for accelerated discovery.
  • Clarity: The paper is easy enough to follow, and the ideas are presented in a logical way.
Critical weaknesses
  • Evaluation Bias: The headline results in Table 1 (a, b) are misleading since the number of generated samples is matched. The baselines do not get to benefit from filtering / rejection sampling in the same was as ReachDiff, providing an advantage to the proposed methods in the primary results. Table 1 (c) is a more fair comparison, matching the budget given to the baselines, reporting a more modest improvement as compared to the previous tables.
  • Baselines: It would be useful to supplement the current evaluation with comparison to best-of-N sampling, rejection sampling, tree-based guidance [1], and search-based approaches [2]. This would provide a more fair comparison, considering that the current method exploits a future-feasibility prediction that is not accessible to any of the compared baselines. The evaluation weakly supports that the method improves of unguided diffusion, but it does not support that it improves over value guided diffusion approaches.
  • Novelty: In my opinion this remains relatively weak. The key difference from existing rejection sampling pipelines is the time in the generation process when this filtering begins. While this is a practical time to evaluate, due to the potential compute savings, the technical contribution is not altogether different from standard rejection sampling or filtering operations.

[1] Guo, Yingqing, et al. "Training-free guidance beyond differentiability: Scalable path steering with tree search in diffusion and flow models." Advances in Neural Information Processing Systems 38 (2026): 73343-73384.

[2] Ta, Huu Binh, et al. "Search-Augmented Masked Diffusion Models for Constrained Generation." arXiv preprint arXiv:2602.02727 (2026).

Questions for authors
  • These constraints create an interesting proof-of-concept, but they seem to only capture simple first-order properties, not representative of real “hit rates” for these cyclical peptide design. I’m curious how this method would scale under a stricter (and expanded) set of filters / constraints. Have any experiments been conducted on these harder sets of filters? Is there more potential to accelerate discovery when capturing a comprehensive set of filters or is this subset sufficient?
  • What is “matched recorded work”? Have your experiments captured things like wall time, total FLOPs, etc.?
  • It would be great if the authors could supplement Table 1 (a, b) with a compute matched version of these metrics — essentially, let the baselines models perform post-hoc filtering to match the number of samples provided by ReachDiff.
4 comments
AuthorsOpenReview ↗

Response to Reviewer MwXF (1/2)

Dear Reviewer MwXF,

We thank the reviewer for recognizing the motivation and clarity of ReachDiff and for the constructive suggestions. We have conducted additional experiments to strengthen the evaluation fairness, establish the contribution beyond existing value-guided diffusion, and examine robustness under stricter molecular constraints.

W1 & Q2 & Q3. Evaluation Fairness, Computational Accounting, and Post-Hoc Rejection Sampling

We appreciate the reviewer's suggestion to evaluate discovery efficiency under matched computational budgets while allowing baselines to perform additional sampling and post-hoc filtering.

Our matched recorded work metric accounts for estimated model computation throughout inference, including backbone execution, attention operations, intermediate feature extraction, and critic evaluation. All intermediate branching, discarded candidates, and unfinished trajectories are charged to the computational budget, providing a more comprehensive comparison than matching completed samples alone.

We also audited wall-clock measurements. On a separate three-seed panel, independent sampling and ReachDiff recorded 3,095.79 s and 2,973.73 s, respectively. Since these runs used shared hardware, our controlled comparisons rely on recorded model-operation estimates, distinguished from hardware-profiled FLOPs.

To directly address the evaluation concern, we conducted an additional compute-matched comparison across five paired seeds and nine generation conditions, assigning each method an identical budget of 1.1764 × 10^15 estimated operations per seed. All methods share the same frozen generator, verifier, and molecular deduplication protocol. Independent sampling retains every verifier-passing product, with all rejected, invalid, and duplicate attempts included in the budget.

MetricIndependent + RejectionPepTune-styleReachDiff
Completed terminal samples4,2904,1703,672
Unique class-feasible / seed ↑490.6496.6542.2
Unique strict-feasible / seed ↑128.6123.6149.4

Remarkably, ReachDiff discovers more unique feasible cyclic peptides despite completing fewer terminal samples. Compared with independent rejection sampling, ReachDiff achieves 10.5% more class-feasible discoveries (+51.6 per seed; 95% CI [39.6, 64.2]) and 16.2% more strict-feasible discoveries (+20.8 per seed; 95% CI [6.0, 39.4]). Against PepTune-style search, strict-feasible discovery further improves by 20.9% (+25.8 per seed; 95% CI [17.2, 34.4]).

All three paired confidence intervals exclude zero. These results demonstrate that ReachDiff's advantage extends beyond equal-terminal comparisons: early feasibility-guided allocation enables more productive molecular discovery under the same inference-compute budget, even when competing methods are allowed additional sampling and post-hoc rejection.

W2 & W3. Value-Guided Baselines and Methodological Novelty

To directly investigate the contribution beyond existing value-guided diffusion, we implemented SVDD-MC with stepwise Monte Carlo guidance and evaluated it using the same frozen generator, feasibility verifier, training data, and recorded-work budget across five seeds and nine conditions.

The following results are obtained on a separate compute-matched evaluation panel:

MethodUnique Class-Feasible ↑Unique Strict-Feasible ↑
Independent Rejection1,841476
SVDD-MC (M=5)496152
One-shot Per-parent (same MSE critic)1,589421
One-shot Global (same MSE critic)2,015526
ReachDiff2,024546

ReachDiff achieves 4.08× more unique class-feasible and 3.59× more unique strict-feasible discoveries than SVDD-MC under matched inference computation. While SVDD-MC achieves higher feasibility among completed outputs, its repeated candidate expansion and value evaluation incur substantial computational overhead. ReachDiff instead performs a single global allocation and completes selected trajectories through ordinary denoising, achieving substantially higher discovery efficiency.

Importantly, ReachDiff's contribution goes beyond moving rejection earlier in the sampling process. Rather than filtering completed molecules or repeatedly selecting candidates within individual trajectories, ReachDiff uses predicted future feasibility to allocate a shared continuation budget globally across competing partial states.

To isolate this contribution from the predictor itself, we compared global and per-parent allocation using exactly the same MSE-trained critic. Global allocation increases class-feasible discovery from 1,589 to 2,015 (+26.8%; paired gain +85.2 per seed, 95% CI [75.4, 95.6]). Strict-feasible discovery similarly increases from 421 to 526 (+24.9%).

AuthorsOpenReview ↗

Response to Reviewer MwXF (2/2)

Furthermore, one-shot global allocation using the MSE critic closely matches the original ReachDiff (2,015 versus 2,024 class-feasible discoveries), demonstrating that the primary efficiency gain arises from the global allocation strategy rather than the specific value-prediction objective.

Additional guidance-frequency controls reinforce this finding: under a common recorded-work budget, one-, two-, and three-stage allocation yield 3,161, 2,413, and 2,387 unique feasible products, respectively.

Together, these results establish the methodological value of combining early future-feasibility estimation with one-shot global competition, providing substantially greater feasible-discovery efficiency than repeated local value guidance under limited inference computation.

Q1. Robustness under Stricter and Expanded Design Constraints

To investigate whether ReachDiff remains effective beyond the original feasibility criteria, we conducted additional evaluations under stricter topological and geometric constraints, keeping the generator, predictor, and search procedure unchanged.

Strict-topology evaluation. Beyond the original topology-class criterion, we enforce exact residue counts, terminal/cap annotations, positional anchors, correct linkage atoms, and formation of the intended ring bond. These requirements were not included in the original predictor's training objective.

Under matched recorded computation, ReachDiff discovers 149.4 unique strict-feasible products per seed, compared with 128.6 for independent rejection sampling (+16.2%; paired difference +20.8, 95% CI [6.0, 39.4]). The improvement over PepTune-style search reaches 20.9%.

Stringent geometric evaluation. We further evaluated saved conformations under progressively tighter MMFF94-based structural criteria, including closure/bond tolerances, steric clashes, angular deviations, and energy thresholds.

Structural CriterionIndependent SamplingReachDiff
Original physical feasibility42.78%55.56%
Tighter physical feasibility42.50%55.56%
Stringent physical feasibility8.89%13.89%

On this separate fixed-output evaluation panel, ReachDiff achieves a 56.3% relative improvement under the most stringent criteria (+5.00 percentage points; 95% CI [1.11, 8.33]). None of these additional criteria were used during predictor training or search.

Notably, the relative structural-feasibility advantage increases from 29.9% under the original criteria to 56.3% under the most stringent criteria, even as the absolute acceptance rates decrease. This suggests that early feasibility-guided allocation can be particularly beneficial when increasingly restrictive constraints make successful terminal outcomes harder to obtain.

The potential benefit of expanded filtering depends on whether constraint satisfaction can be predicted from intermediate states. Our results demonstrate this advantage for increasingly stringent chemical-topology and structural criteria, while experimental hit rates additionally depend on target-specific biological properties.

Overall, ReachDiff maintains its discovery advantage under substantially more demanding structural requirements without retraining the predictor or modifying the inference procedure.

We sincerely thank the reviewer for the valuable suggestions. The additional experiments strengthen the evidence for ReachDiff's computational efficiency, robustness, and methodological contribution. The direct SVDD-MC comparison, same-critic allocation controls, and compute-matched rejection-sampling evaluation demonstrate that one-shot global allocation uses limited inference computation more effectively, while the stricter structural evaluations show that these advantages extend beyond the original feasibility criteria. Together, these findings support ReachDiff as an efficient approach to constrained cyclic-peptide discovery.

Thank you for the clarifications and additional results. The provided results have directly addressed my concern regarding a matched evaluation setup.

My remaining concerns center around the selected baselines and novelty. I'll summarize below:

  • Additional baselines. Could the authors please clarify how their baselines connect to the tree-based and search-based guidance approaches cited in my review? The added SVDD-MC baseline is interesting to include, but it differs from the baselines which were suggested in my review. It also remains unclear whether SVDD-MC has been tuned, given the results only report M=5.

  • Conceptual novelty. Connected to the previous point, the overlap with tree-based and search-based methods is not clarified in the rebuttal. I also am not sure I understand claims like "ReachDiff's contribution goes beyond moving rejection earlier in the sampling process." While I agree that moving rejection earlier in the sampling process enables better allocation of the compute budget -- and indeed the empirical results support this is useful -- it doesn't imply that the methodology is novel or different than early-state rejection sampling.

  • Constraint complexity. While the expanded constraint sets are interesting, they do not imply "experimental hit rates on target-specific biological properties". This is a very important caveat, as these hit rates are likely dramatically more difficult to predict than structural criteria -- and ReachDiff may not be directly applicable to improving real hit rates. Hence, it remains largely unproven whether the efficiency results actually translate to improved downstream efficiency.

Thus, while I appreciate the additional results and more fair evaluation, several of my key concerns remain unaddressed.

AuthorsOpenReview ↗

Response to Reviewer MwXF — Follow-up Clarifications

Dear Reviewer MwXF,

We thank the reviewer for the thoughtful follow-up and for acknowledging that our additional experiments address the evaluation fairness concern. We would like to further clarify the relationship to existing methods and the methodological contribution of ReachDiff.

1. Relationship to TreeG, SearchDiff, and SVDD-MC

ReachDiff belongs to the broader family of search-guided diffusion methods, but differs in how future-feasibility information is used to allocate inference computation.

TreeG [1] employs iterative tree search with candidate evaluation and selection, including global top-A selection. SearchDiff [2] integrates constraint-aware search into the reverse denoising process. In contrast, ReachDiff uses a learned, policy-conditional terminal-feasibility predictor to make a single global allocation of the remaining continuation budget, after which ordinary frozen denoising resumes without further intervention.

Our guidance-frequency experiments support this design: under matched recorded work, one-, two-, and three-stage allocation yield 3,161, 2,413, and 2,387 unique feasible products, respectively.

Regarding SVDD-MC, M=5 follows the official implementation's default configuration. We additionally evaluated M=2 under the same recorded-work budget:

MethodUnique Class-Feasible ↑Unique Strict-Feasible ↑
SVDD-MC (M=5)496152
SVDD-MC (M=2)752195
ReachDiff2,024546

Although reducing M improves SVDD-MC's discovery efficiency, ReachDiff maintains a substantial advantage. These comparisons highlight the benefit of one-shot allocation over repeated stepwise value guidance in our constrained-generation setting.

2. Clarification of Methodological Novelty

We appreciate the reviewer's distinction between early rejection and methodological novelty. While intermediate-state selection and global top-K search are established primitives, ReachDiff contributes a future-feasibility-guided allocation strategy that explicitly distributes a fixed continuation budget across competing partial trajectories.

Rather than simply rejecting completed candidates, ReachDiff estimates their likelihood of successful completion under the frozen generator and reallocates computation before full generation.

Crucially, our same-critic experiments isolate the contribution of the allocation strategy. Using exactly the same MSE predictor and candidate construction, global allocation improves unique class-feasible discovery by 26.8% over per-parent selection (2,015 versus 1,589; paired 95% CI [75.4, 95.6]). The MSE-based global variant also closely matches the original ReachDiff (2,015 versus 2,024).

These results demonstrate that the efficiency improvement primarily comes from the allocation strategy rather than a specialized predictor-training objective. We therefore position ReachDiff as a compute-efficient approach within search-guided diffusion, supported by controlled mechanism studies and matched-budget discovery improvements.

3. Constraint Complexity and Downstream Applicability

We appreciate the distinction between computational feasibility and experimental biological hit rates. Our study focuses on the joint satisfaction of chemical validity, cyclization topology, and closure geometry, without equating these criteria with target-specific biological activity.

Importantly, ReachDiff's advantage persists under stricter evaluations: 16.2% more unique strict-feasible discoveries under matched computation and a 56.3% relative improvement under stringent MMFF94-based structural screening.

These results establish that early feasibility-guided allocation remains effective beyond the original structural criteria. Extending this benefit to biological hit rates would depend on the predictability of target-specific outcomes from intermediate generative states.

Closing Remarks

We sincerely appreciate the reviewer's careful assessment. Our central contribution is establishing how early future-feasibility information can guide one-shot global allocation of finite diffusion computation to improve constrained molecular discovery. The SVDD-MC comparisons, same-critic controls, and matched-budget results provide direct empirical support for this contribution.

If our additional experiments and clarifications have addressed your remaining concerns, we would be very grateful if you could consider increasing your rating accordingly. Thank you again for your time and thoughtful consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.