- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper introduces PREFGATEGFN, which maps a user preference vector to multiplicative gates on features used by GFlowNet action scorers, with optional preference-dependent action temperatures. It evaluates preference response separately from candidate quality, using exact small grids, matched molecular controls, fragment generation, and synthesis-aware reaction generation.
Critical strengths
-
The conceptual distinction between feature gating and a positive scalar temperature is clear.
-
The paper measures the response of a frozen policy to changed preferences rather than inferring controllability from Pareto coverage. It also reports cases where HN-GFN or MOGFN-PC is more responsive, which makes the interpretation more credible.
-
The study includes exact-distribution controls, a matched seven-arm conditioning comparison, and full-budget normalization/curriculum ablations.
Critical weaknesses
-
The central empirical attribution remains unresolved. Table 1 compares complete fragment systems with different preference schedules, normalization, and controller freezing. In the within-gated full-budget ablation, normalization alone raises hypervolume from 0.668 to 0.885, while adding the tested curriculum lowers its mean to 0.809. The same normalization has not been applied to MOGFN or HN-GFN. Thus the strongest coverage gain is evidence for the training recipe, not specifically for preference gating.
-
The matched architecture controls are not decisive in favor of gates. On the fixed 512-molecule library, gates plus temperature have L1 error 0.262 versus 0.270 for concatenation plus temperature (Table 3), a small difference relative to variation. The matched short-budget fragment responses are explicitly inconclusive. A full-budget comparison that changes only the preference interface would better establish the proposed method’s distinctive value.
-
The 14-objective results use one training seed per target; the reported bars vary over five targets, not independent model trainings. This limits the robustness of the 133-versus-69 unique-hit comparison. The target-level counts and repeated training seeds would clarify whether gains are broad or driven by a few favorable runs.
Questions for authors
-
Can the authors run MOGFN and HN-GFN with the same dynamic normalization, curriculum, update schedule, loss, and evaluation budget used for PREFGATEGFN?
-
How does the normalization-only gated checkpoint respond to direct preference interventions? It has the best fragment hypervolume, but preference response is shown for the original both-on checkpoint.
-
Why is the measured 14-objective preference response smaller than MOGFN-PC’s in Appendix H.2, and how sensitive is that comparison to the unequal controller schedules?
-
How variable are the 14-objective distinct qualifying counts across independently trained seeds for each target?
3 comments
We also tested a gradual shift from QED toward sEH, keeping the SA weight fixed at 0.3. Each preference uses 512 attempts per training seed:
| Weights (sEH, QED, SA) | sEH/8 | QED | Normalized SA |
|---|---|---|---|
| (0.0, 0.7, 0.3) | 0.4658 ± 0.0446 | 0.7681 ± 0.0167 | 0.8295 ± 0.0176 |
| (0.2, 0.5, 0.3) | 0.5250 ± 0.0454 | 0.7168 ± 0.0504 | 0.8550 ± 0.0108 |
| (0.3, 0.4, 0.3) | 0.5985 ± 0.0532 | 0.6375 ± 0.1029 | 0.8717 ± 0.0123 |
| (0.35, 0.35, 0.3) | 0.6687 ± 0.0614 | 0.5219 ± 0.1489 | 0.8756 ± 0.0108 |
| (0.4, 0.3, 0.3) | 0.7435 ± 0.0532 | 0.3827 ± 0.1630 | 0.8670 ± 0.0145 |
| (0.5, 0.2, 0.3) | 0.8013 ± 0.0458 | 0.2817 ± 0.1385 | 0.8552 ± 0.0177 |
| (0.7, 0.0, 0.3) | 0.8335 ± 0.0442 | 0.2416 ± 0.1091 | 0.8364 ± 0.0201 |
As the preference shifts from QED toward sEH, the normalization-only checkpoints show increasing mean sEH and decreasing mean QED, demonstrating the requested trade-off. They also achieve higher mean QED and normalized SA than the curriculum-only checkpoints at every evaluated preference. These results show that the higher-coverage normalization-only configuration retains preference responsiveness in the evaluated region.
Q3: 14-objective preference response and controller schedules
Thank you for highlighting this distinction. The original Appendix H.2 comparison used different effective controller and preference schedules, so its response gap cannot be attributed solely to the conditioning mechanism. Our new evaluation uses common sampling settings.
At a budget of 64 valid outputs per request and sampling temperature 1, the standardized response across the four H.2 focus objectives—docking, hERG, DILI and bioavailability—is 0.164 ± 0.093 for PrefGateGFN, versus 0.195 ± 0.029 for MOGFN-PC. Extending the same sweep to all 14 focus objectives gives 0.102 ± 0.054 versus 0.122 ± 0.006, respectively. Response measures the score change from focused weight 0.04 to 0.40, standardized by the pooled score SD over weights 0.04, 0.12 and 0.40.
Despite this smaller response, PrefGateGFN produces a higher mean number of qualifying candidates. For the four H.2 focused requests at weight 0.4, it generates 10.18 ± 2.33 unique molecules satisfying all 14 criteria per 64 valid outputs, compared with 8.80 ± 1.31 for MOGFN-PC. The remaining weight is shared equally across the other 13 objectives. Qualification requires the physicochemical criterion, all twelve ADMET filters and docking ≤ −8 kcal/mol; duplicates consume the sampling budget.
A larger preference response measures greater sensitivity to changing weights; it does not guarantee more molecules satisfying the requested constraints. Roy et al. (2023) (Sections 3.3–4.2) likewise distinguish controllability from quality and coverage: scalarized preference conditioning can concentrate samples near the extremes of concave Pareto fronts while poorly covering intermediate trade-offs. Our results support evaluating response alongside absolute qualifying yield, while leaving the contribution of controller schedules unresolved.
Q4 / W3: Variation across independently trained seeds
Thank you for asking us to distinguish these sources of variation. The original bars show variation across targets. We now report three training seeds per method, with three sampling repeats per seed at the original fixed preference.
| Target | MOGFN-PC | PrefGateGFN |
|---|---|---|
| ADRB2 | 33.6 ± 9.6 | 38.6 ± 12.3 |
| ALDH1 | 207.3 ± 20.8 | 247.0 ± 42.8 |
| ESR agonist | 110.0 ± 11.0 | 93.9 ± 15.8 |
| ESR antagonist | 170.3 ± 13.3 | 189.0 ± 33.9 |
| FEN1 | 17.4 ± 9.5 | 31.7 ± 12.5 |
Entries count unique molecules satisfying all 14 criteria per 1,000 valid outputs. We average the three sampling repeats within each training seed, then report mean ± SD across three training seeds. Preferences are 0.4 for docking, 0.2 for the physicochemical objective and 0.4/12 for each ADMET objective. Qualification requires the physicochemical criterion, all twelve ADMET filters and docking ≤ −8 kcal/mol; failed or nonfinite predictions cannot qualify, and duplicates consume the budget.
PrefGateGFN has higher mean qualifying yield on four of five targets: ADRB2, ALDH1, ESR antagonist and FEN1. MOGFN-PC leads on ESR agonist, making the remaining target dependence explicit.
Thank you for the additional experiments. The full-budget fragment comparison addresses my concern about matched training conditions, and the preference sweep shows that the normalization-only checkpoints retain useful steering. My main concern on how much of the observed gain can be attributed to gating still holds. Under the shared recipe, the mean HV advantage over MOGFN-PC is small.
I appreciate the stronger evidence for the method’s practical utility, but the incremental benefit of the proposed gating mechanism remains uncertain. I therefore lean toward maintaining my original score.
Thank you for the detailed questions about matched training conditions, preference interventions and training-seed variation. The completed fragment comparisons demonstrate the benefits of the proposed interface under the requested controls. We also report the available independent-seed results for the reaction setting.
Q1 / W1 / W2: Matched conditions and architectural attribution
We completed 36 fragment models across the tested settings, with 15,000 updates per model and three training seeds per condition. Within each comparison, all interfaces share dynamic normalization, controller update/freeze schedule, loss and evaluation budget, as well as the backbone and preference bank. With normalization and learned temperatures enabled and curriculum disabled:
The three matched interfaces are newly trained controlled implementations, separate from the original complete-system checkpoints in Table 1. MOReinforce and Routing by Reaching use the same 960,000 training attempts per seed and evaluation budgets, with method-specific training recipes. The shared-recipe comparison concerns MOGFN-PC, HN-GFN and PrefGateGFN. Routing by Reaching’s parameter count sums its three specialists.
Joint hits count distinct canonical molecules meeting all three cutoffs (0.50, 0.25, 0.60) on sEH/8, QED and (10−SA)/9 in 512 equal-preference attempts. Repeated scores are averaged before thresholding; duplicates and invalids consume the attempt budget. Entries are mean ± SD across three training seeds.
Under this shared normalization-on, curriculum-off recipe, PrefGateGFN has the highest mean HV and distinct joint yield among the three interfaces, while using about 73% fewer parameters than HN-GFN.
Q2: Direct interventions on normalization-only checkpoints
We evaluated the requested normalization-only checkpoints directly, using frozen checkpoints from the three original training seeds. We changed only the preference at evaluation and measured whether each focused request produces more distinct molecules meeting the requested property's cutoff than equal-weight conditioning. Using the cutoffs defined above, we obtain:
Weights are ordered as (sEH, QED, SA). Each entry is mean ± SD across three seeds, with 512 attempts per preference. Joint-qualified means that all three cutoffs are met. Focusing on sEH or QED increases its qualifying count in every seed. The SA cutoff is already nearly saturated at equal weights, leaving little room for a higher qualifying count; it is therefore uninformative about an additional SA benefit. The fall in joint yield makes the trade-off from prioritizing one property explicit.
We also evaluated the three focused requests on the separate matched normalization-only models, with three seeds per interface and the same counting rules:
PrefGateGFN has the highest mean qualifying count for QED and SA; HN-GFN leads for sEH. The direct interventions on the original checkpoints show that the higher-coverage normalization-only configuration retains useful steering: focusing on sEH or QED increases its qualifying count in every seed. The matched comparison additionally supports useful candidate yield under the same normalization-only recipe. The seed variation is reported explicitly.
[continued in the next comment]