- 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
- Simple, practical design. A single selection step improves constrained generation without modifying the generator during search.
- 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
- The contribution beyond existing value-guided diffusion is insufficiently established. SVDD (2408.08252) is closely related and cited, but no direct comparison is reported.
- 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
-
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.
-
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.
-
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.
-
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.
-
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
Response to Reviewer zBfP (2/2)
| Structural Criterion | Independent Sampling | ReachDiff |
|---|---|---|
| Original physical feasibility | 42.78% | 55.56% |
| Tighter physical feasibility | 42.50% | 55.56% |
| Stringent physical feasibility | 8.89% | 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.
| Metric | Frozen Predictor | Strict-Label Retraining |
|---|---|---|
| Strict-label AUROC ↑ | 0.6979 | 0.7214 |
| Within-condition ranking accuracy ↑ | 0.6740 | 0.7237 |
| Top-quarter strict-feasible yield ↑ | 0.2195 | 0.2610 |
| Brier score ↓ | 0.3923 | 0.1251 |
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.
| Collector | Verifier Calls ↓ | Acquisition Work ↓ | Aligned Joint@1 ↑ |
|---|---|---|---|
| Independent Sampling | 762 | 1.1261 × 10^15 | 66.87% |
| TR2-D2-style | 792 | 1.3581 × 10^15 | 67.80% |
| ReachDiff | 570 | 0.9508 × 10^15 | 68.52% |
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.
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.
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.
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.