acceptodds
Under review as a conference paper at ICLR 2027

Navigating Multiobjective Drug Discovery Trade-offs with Preference-Gated GFlowNets

Abstract

Multi-objective molecular generation faces a double challenge: finding molecules that survive demanding screens and changing what is sampled when a scientist changes priorities. A broad Pareto frontier alone does not establish such control. We introduce PrefGateGFN, a GFlowNet that uses preference-dependent feature gates to modulate molecular construction-action scores. Gates can change action rankings, while optional learned temperatures adjust sampling concentration; dynamic reward normalization addresses coverage as a separate training choice. We examine these roles using exact-grid distributions, matched molecular controls, and preference interventions on frozen policies. In a three-objective fragment task, the full system achieves higher mean top-10 reward, diversity, and hypervolume than the evaluated MOGFN and HN-GFN systems. A three-seed ablation raises mean hypervolume from 0.668 with normalization and curriculum disabled to 0.885 with normalization alone. In a comparison of complete systems with different training recipes, PrefGateGFN achieves substantial yield under a demanding fourteen-objective screen across five protein targets: an average of 133 distinct molecules per 1,000 draws satisfying docking, physicochemical, and twelve predicted ADMET criteria, outperforming MOGFN-PC by factors of 1.93 in qualifying molecules and 2.10 in qualifying scaffolds. Together with its measured preference response, these results support steering molecular generation toward diverse, computationally screened candidates for experimental follow-up.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI-aided drug discovery, multiobjective optimization, gflownets

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_xeyeOpenReview ↗
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
  1. The conceptual distinction between feature gating and a positive scalar temperature is clear.

  2. 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.

  3. The study includes exact-distribution controls, a matched seven-arm conditioning comparison, and full-budget normalization/curriculum ablations.

Critical weaknesses
  1. 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.

  2. 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.

  3. 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
  1. Can the authors run MOGFN and HN-GFN with the same dynamic normalization, curriculum, update schedule, loss, and evaluation budget used for PREFGATEGFN?

  2. 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.

  3. 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?

  4. How variable are the 14-objective distinct qualifying counts across independently trained seeds for each target?

3 comments

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:

MethodParameters (M)HVDistinct joint hits / 512
MOReinforce3.3640.2671 ± 0.19090.3 ± 0.6
MOGFN-PC3.390.8739 ± 0.0876267.0 ± 211.4
HN-GFN12.710.8562 ± 0.0485377.3 ± 62.8
PrefGateGFN3.450.8913 ± 0.0426403.3 ± 72.2
Routing by Reaching10.0920.6411 ± 0.0054171.3 ± 21.0

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:

Preference (sEH, QED, SA)Joint-qualified / 512sEH-qualified / 512QED-qualified / 512SA-qualified / 512
Equal weights (1/3, 1/3, 1/3)414.0 ± 2.6475.0 ± 46.9450.3 ± 48.5511.7 ± 0.6
(1, 0, 0)204.7 ± 169.4511.0 ± 1.0227.0 ± 155.9478.3 ± 22.0
(0, 1, 0)209.3 ± 63.0264.3 ± 70.0512.0 ± 0.0397.3 ± 27.5
(0, 0, 1)284.0 ± 138.6466.7 ± 21.1314.0 ± 162.8511.3 ± 1.2

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:

MethodsEH (1, 0, 0)QED (0, 1, 0)SA (0, 0, 1)
PrefGateGFN491.3 ± 28.1469.3 ± 56.7318.0 ± 41.6
MOGFN-PC359.3 ± 262.7326.0 ± 237.4167.7 ± 133.9
HN-GFN504.3 ± 6.4314.3 ± 121.89.0 ± 3.6

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]

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/8QEDNormalized SA
(0.0, 0.7, 0.3)0.4658 ± 0.04460.7681 ± 0.01670.8295 ± 0.0176
(0.2, 0.5, 0.3)0.5250 ± 0.04540.7168 ± 0.05040.8550 ± 0.0108
(0.3, 0.4, 0.3)0.5985 ± 0.05320.6375 ± 0.10290.8717 ± 0.0123
(0.35, 0.35, 0.3)0.6687 ± 0.06140.5219 ± 0.14890.8756 ± 0.0108
(0.4, 0.3, 0.3)0.7435 ± 0.05320.3827 ± 0.16300.8670 ± 0.0145
(0.5, 0.2, 0.3)0.8013 ± 0.04580.2817 ± 0.13850.8552 ± 0.0177
(0.7, 0.0, 0.3)0.8335 ± 0.04420.2416 ± 0.10910.8364 ± 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.

TargetMOGFN-PCPrefGateGFN
ADRB233.6 ± 9.638.6 ± 12.3
ALDH1207.3 ± 20.8247.0 ± 42.8
ESR agonist110.0 ± 11.093.9 ± 15.8
ESR antagonist170.3 ± 13.3189.0 ± 33.9
FEN117.4 ± 9.531.7 ± 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.

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

The paper introduces PrefGateGFN, which uses preference-dependent feature gates and optional temperatures to condition molecular construction. It evaluates distribution matching on grids, molecular quality and preference response in fragment generation, and computational screening yield in synthesis-aware generation. On the fourteen-objective screen, the system produces a mean of 133 distinct qualifying molecules per 1,000 draws across five targets. A separate fragment ablation identifies reward normalisation as the main contributor to improved coverage among the training components tested.

Critical strengths
  1. The evaluation tests what happens when preferences change in a frozen policy, rather than treating a broad frontier as evidence of control. Exact distribution checks and the gate/temperature factorial make this distinction concrete (Section 4.3; Appendix C).

  2. The screening analysis counts distinct qualifying molecules within a fixed draw budget, penalising duplication. The fourteen-objective results across five targets are promising, and the authors appropriately distinguish predicted screening success from experimental biological validation.

Critical weaknesses
  1. The incremental benefit of gating remains uncertain at a full molecular training budget. Within the gated model, enabling normalisation with curriculum off raises mean hypervolume from 0.668 to 0.885; the same ablation is not run for MOGFN or HN-GFN (Appendix E.1). The matched 400-update fragment study is inconclusive, and the expressive conditional models have close finite-library errors (Appendix C). The full reaction comparisons also differ in controller updates and preference schedules. These are acknowledged limitations, but leave the architectural contribution less convincing than the complete-system results.

  2. The connection between improved coverage and useful steering is not fully tested. Preference sweeps are missing for the strongest fragment coverage configuration, which uses normalisation without curriculum (Appendix E.1). The separate five-target reaction sweep gives mean responses of 0.055 and 0.438 for PrefGateGFN and MOGFN-PC, respectively, under different controller schedules (Appendix H.2). These are response magnitudes, not calibration errors, so the smaller value does not itself establish worse control. Evidence that the strongest configurations combine quality with reliable steering would strengthen the paper.

Questions for authors
  1. Evidence request: Are full-budget comparisons available with the same normalisation and training schedule applied to the alternative conditioners, reporting both molecular quality and preference response? These would clarify the contribution of gating.

  2. Clarification: What criterion defines useful molecular steering beyond response magnitude? Supporting sweeps for the normalisation-only fragment checkpoints, with variation across seeds, would help assess the claimed combination of coverage and control.

2 comments

Thank you for the questions about coverage and practical preference control.

Q1/W1 — Full-budget comparison under matched training conditions

We compared molecular quality and preference response for MOGFN-PC, HN-GFN, and PrefGateGFN in full-budget fragment runs. The three methods shared the backbone, preference sampling, loss, controller schedule, and evaluation budget. All runs used dynamic reward normalization and learned temperatures, with the preference curriculum disabled (15,000 updates per model, 64 trajectories per update).

MethodHypervolumeDistinct joint hits / 512Gain over equal weights
MOGFN-PC0.874 ± 0.088267.0 ± 211.40.041 ± 0.024
HN-GFN0.856 ± 0.049377.3 ± 62.80.047 ± 0.034
PrefGateGFN0.891 ± 0.043403.3 ± 72.20.042 ± 0.041

All tables report mean ± SD across three training seeds. Both sEH and SA are normalized scores. Joint hits count distinct molecules meeting all three cutoffs (sEH ≥ 0.50, QED ≥ 0.25, and SA ≥ 0.60) within 512 attempts at equal weights. Invalid and duplicate outputs count toward the attempt budget.

Gain over equal weights is the difference in mean score between requested and equal-weight conditioning, with both sets of outputs scored under the requested weights. We averaged this difference over 40 requests, using 512 attempts per conditioning setting and seed for each request. PrefGateGFN had the highest mean hypervolume and joint yield. HN-GFN had the highest mean gain, followed by PrefGateGFN, with small differences in mean gain relative to seed-to-seed variation.

Q2/W2 — Useful steering with normalization alone

Steering. We assess useful steering by asking whether a focused preference produces more distinct molecules meeting the requested property cutoff than equal weights, within the same 512 attempts. We first vary the preference on the three original normalization-only checkpoints while holding each checkpoint fixed. The last column counts molecules that satisfy all three cutoffs.

Preference (sEH, QED, SA)sEH-qualified / 512QED-qualified / 512SA-qualified / 512Joint-qualified / 512
Equal weights (1/3, 1/3, 1/3)475.0 ± 46.9450.3 ± 48.5511.7 ± 0.6414.0 ± 2.6
(1, 0, 0)511.0 ± 1.0227.0 ± 155.9478.3 ± 22.0204.7 ± 169.4
(0, 1, 0)264.3 ± 70.0512.0 ± 0.0397.3 ± 27.5209.3 ± 63.0
(0, 0, 1)466.7 ± 21.1314.0 ± 162.8511.3 ± 1.2284.0 ± 138.6

Focusing on sEH or QED increased the corresponding qualifying count in every seed. Nearly all attempts already yielded distinct SA-qualified molecules at equal weights, leaving little room for improvement on that count. Mean joint yield fell from 414.0 under equal weights for each of the single-objective preferences.

We also swept seven preferences on the original normalization-only checkpoints, using 512 attempts per preference and seed. We held the SA weight at 0.3 and shifted weight from QED to sEH.

Weights (sEH, QED, SA)sEHQEDSA
(0.0, 0.7, 0.3)0.466 ± 0.0450.768 ± 0.0170.830 ± 0.018
(0.2, 0.5, 0.3)0.525 ± 0.0450.717 ± 0.0500.855 ± 0.011
(0.3, 0.4, 0.3)0.599 ± 0.0530.638 ± 0.1030.872 ± 0.012
(0.35, 0.35, 0.3)0.669 ± 0.0610.522 ± 0.1490.876 ± 0.011
(0.4, 0.3, 0.3)0.744 ± 0.0530.383 ± 0.1630.867 ± 0.015
(0.5, 0.2, 0.3)0.801 ± 0.0460.282 ± 0.1390.855 ± 0.018
(0.7, 0.0, 0.3)0.834 ± 0.0440.242 ± 0.1090.836 ± 0.020

Mean sEH increased from 0.466 to 0.834, and mean QED decreased from 0.768 to 0.242. Both means moved in the expected direction across all seven tested preferences.

Qualifying candidate yield across models. We evaluated the three focused preferences on the matched models from Q1. Each column counts distinct molecules that meet the named property's cutoff within 512 attempts.

MethodsEH (1, 0, 0)QED (0, 1, 0)SA (0, 0, 1)
MOGFN-PC359.3 ± 262.7326.0 ± 237.4167.7 ± 133.9
HN-GFN504.3 ± 6.4314.3 ± 121.89.0 ± 3.6
PrefGateGFN491.3 ± 28.1469.3 ± 56.7318.0 ± 41.6

PrefGateGFN yielded the most qualifying molecules on average for the QED and SA requests, and HN-GFN led for the sEH request.

Dear Reviewer rjSW,

We would like to follow up as the author–reviewer discussion period will close in less than 10 hours.

In response to your feedback, we have added experiments, statistical analyses, and clarifications. We hope these additions address your concerns.

If you feel that our responses have addressed your concerns, we would be grateful if you could reconsider your evaluation. If any questions remain, we would be happy to provide further clarification.

Thank you for your time and consideration.

Best regards, The Authors

Reviewer_F4BUOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a preference-conditioned GFlowNet architecture that uses learned multiplicative gates to modulate the features used for action scoring, with a separate optional temperature. The experiments demonstrate superior performance on a multiple-objective molecule generation task.

Critical strengths
  1. This method sounds technical and interesting. Applying GFlowNet for sampling molecules that satisfy multiple objectives sounds novel and useful.
  2. The extensive ablation study supports the novelty of the proposal components.
Critical weaknesses

Although I am not an expert in the GFlowNets domain, I have several concerns about the presentation and experiments.

  1. Some important baselines are missing, like Goal-conditioned GFlowNets [1] (2023), [2](ICML 2026). Also, the original paper of MOGFN compares with a lot of baselines, but this paper didn’t include them. More detailed discussion and comparison will make the paper convincing.
  2. Some ablation studies are conducted on the toy dataset (like sec 4.1), it would be better to demonstrate the importance of the components on the real dataset.

[1]Goal-conditioned GFlowNets for Controllable Multi-Objective Molecular Design [2]Routing by Reaching: Composition of Pre-trained GFlowNets for Multi-Objective Generation Minor:

  1. The introduction is too long; too many details about the proposed model. 2.Some baselines should add reference in the intro section like HN-GFN. The method part I think need to make it better understand. It is hard to read as an ML paper for a broad audience. Also, an algorithm summary and an architecture Schematic figure would also be helpful.
Questions for authors

I was wondering if the author could add more baselines.

2 comments

Thank you for the suggestions on baselines, molecular ablations, and presentation.

Q1/W1 — Additional baselines and goal conditioning

Goal-conditioned GFlowNets generate candidates for a specified region in objective space (Roy et al., 2023). Our fragment benchmark evaluates generation given weights that express the relative importance of the objectives. For a direct comparison under those inputs, we evaluated Routing by Reaching and MOReinforce.

In our evaluation, Routing by Reaching combines three separately trained single-objective GFlowNets at inference time. MOReinforce uses reinforcement learning to maximize the expected preference-weighted reward. Each baseline used 960,000 training attempts per seed, with Routing by Reaching splitting that budget across its three specialists.

We compare the new baselines with the original MOGFN, HN-GFN, and PrefGateGFN checkpoints. For each method and seed, we used 128 draws per preference across the same 128 preferences. Values are mean ± SD across three training seeds. The parameter count for Routing by Reaching includes all three specialists.

PolicyParameters (M)HVMean top-10 rewardTop-10 diversity
MOReinforce3.3640.267 ± 0.1910.657 ± 0.1240.000 ± 0.000
MOGFN3.2510.615 ± 0.0040.659 ± 0.0030.595 ± 0.007
HN-GFN4.5600.669 ± 0.0170.729 ± 0.0050.596 ± 0.009
Routing by Reaching10.0920.641 ± 0.0050.714 ± 0.0020.584 ± 0.005
PrefGateGFN3.3620.809 ± 0.0700.829 ± 0.0220.648 ± 0.025

PrefGateGFN achieved the highest mean hypervolume, top-10 reward, and top-10 diversity among the five methods. Each method retained its own training recipe, so these results compare the complete systems under a common training-attempt budget and evaluation protocol.

Goal conditioning and feature gating address complementary aspects of generation. A goal direction specifies where to sample in objective space, and feature gates determine how the conditioning signal affects action scores. A goal direction could therefore serve as input to the gate controller.

W2 — Ablations on molecular generation

We tested the components on the fragment generation task. In the full-budget training ablation, we changed one setting at a time from a reference model with normalization and learned temperatures enabled and curriculum disabled. We trained these models separately from the original checkpoints in Q1. Values are mean ± SD across three training seeds.

PrefGateGFN training settingHVMean top-10 score
Reference: normalization on, curriculum off, learned temperatures0.891 ± 0.0430.864 ± 0.045
Remove normalization0.654 ± 0.0050.704 ± 0.007
Add curriculum0.886 ± 0.0270.874 ± 0.006
Remove learned temperatures0.880 ± 0.0130.861 ± 0.002

Removing normalization produced the largest drop, lowering mean HV by 0.237 and mean top-10 score by 0.160. Adding curriculum or removing learned temperatures produced smaller changes.

To test how the trained model uses each pathway, we held the checkpoints fixed and set the gates, temperatures, or both to their equal-preference outputs. We evaluated 40 preferences with 512 draws per preference and seed. Score decreases are paired against the responsive setting across the three training seeds.

Evaluation settingMean top-10 scoreMean decreasePaired 95% interval
Responsive gates and temperatures0.901——
Fixed gates0.8950.006[−0.019, 0.032]
Fixed temperatures0.8950.006[−0.003, 0.015]
Fixed gates and temperatures0.8880.013[−0.032, 0.058]

The mean top-10 score decreased in each intervention, but all three descriptive intervals include zero. With three seeds, the size of these effects remains uncertain.

W3 — Presentation

We revised the presentation as suggested, including shortening the introduction and citing HN-GFN there. In the Methods section, we clarified how preferences generate gates and how those gates affect action scores, and added a training algorithm summary. We will add an architecture schematic in the next revision.

Thanks for the detailed response! My main concerns were addressed, and I raise my score to 3.

Reviewer_F1prOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how user preferences can directly influence multi-objective molecular generation. PrefGateGFN uses a preference vector to produce feature-wise gates that adjust action scores during molecular construction, with an optional temperature controlling sampling concentration. The model is trained using a GFlowNet objective and combined with reward normalization and reaction constraints. Experiments on enumerable grids, a fixed molecular library, fragment-based generation, and synthesis-aware generation evaluate distribution fitting, preference response, and candidate quality. The full system improves reward, coverage, or joint screening yield on the main molecular tasks. The appendix also examines the effects of normalization, curriculum learning, and different conditioning mechanisms.

Critical strengths
  1. The paper clearly distinguishes candidate coverage from preference control. It also explains the different roles of gating and temperature: gating can change the action ordering, while temperature mainly changes sampling concentration. Changing preferences with a frozen policy and observing the outputs is a more direct test of the proposed interface than showing Pareto fronts alone.

  2. The ablations in Appendices C and E are substantial. They include matched controls for feature pathways and temperature, as well as normalization and curriculum ablations under the full training budget. The authors report close and unfavorable results rather than hiding them, and they are explicit about the contribution of normalization and the limits of comparisons across conditioners.

Critical weaknesses
  1. The independent benefit of gating over other conditioners remains unclear. The main molecular experiments compare complete systems trained with different recipes. Appendix E shows that normalization accounts for much of the coverage improvement, but the same normalization was not applied to the other conditioners. The matched experiments in Appendix C show that gating is effective for conditioning, but more expressive conditioners perform similarly, and the short-budget generation results do not establish a stable ranking. The evidence therefore supports the full system more strongly than the claimed advantage of the gating architecture itself.

  2. The recommended configuration for practical use is not clear enough. The original configuration gives both good candidate quality and useful preference responses. In Appendix E, the normalization-only configuration has better quality metrics but no corresponding preference sweep. It is therefore unclear whether its quality gains come at the cost of weaker preference control.

Questions for authors
  1. If the same reward normalization and training schedule were applied to the other conditioners, how much of the gating advantage would remain? Under a shared training recipe, which existing result provides the strongest evidence for an independent benefit of gating? What scope do the authors think is appropriate for claims about architectural advantages?

  2. The new normalization-only configuration in Appendix E has not been evaluated with a preference sweep. Which configuration do the authors recommend in practice? Could they provide preference-response results for this configuration to show whether its quality improvement preserves controllability?

  3. The magnitude of preference responses and calibration to the requested targets do not always appear to agree across tasks. What metric do the authors recommend for evaluating preference control? Could they provide a consistent evaluation procedure so that larger output changes are not automatically interpreted as better control?

2 comments

Thank you for the thoughtful questions about the contribution of gating and preference control.

Q1/W1 — Gating under matched training conditions

We trained MOGFN-PC, HN-GFN, and PrefGateGFN under the same full-budget fragment protocol, with dynamic reward normalization and learned temperatures enabled and curriculum disabled. We matched the backbone, preference sampling, loss, controller schedule, and evaluation budget. (Training budget: 15,000 updates, 64 trajectories per update.)

MethodHypervolumeDistinct joint hits / 512
MOGFN-PC0.874 ± 0.088267.0 ± 211.4
HN-GFN0.856 ± 0.049377.3 ± 62.8
PrefGateGFN0.891 ± 0.043403.3 ± 72.2

Both tables report mean ± SD across three training seeds. Both sEH and SA are normalized scores. Joint hits count distinct molecules meeting all three cutoffs (sEH ≥ 0.50, QED ≥ 0.25, and SA ≥ 0.60) within 512 generation attempts at equal weights.

This matched comparison provides the most direct evidence for the architectural contribution. PrefGateGFN has the highest observed mean hypervolume and joint yield, with substantial seed-to-seed variation. We scope the claim to coverage and candidate yield under this fragment training protocol.

Q2/W2 — Recommended configuration and preference sweep

We recommend normalization without curriculum when coverage is the priority in the fragment task. We evaluated the three original normalization-only checkpoints using 512 attempts per preference and seed. We held the SA weight at 0.3 and shifted weight from QED to sEH.

Weights (sEH, QED, SA)sEHQEDSA
(0.0, 0.7, 0.3)0.466 ± 0.0450.768 ± 0.0170.830 ± 0.018
(0.2, 0.5, 0.3)0.525 ± 0.0450.717 ± 0.0500.855 ± 0.011
(0.3, 0.4, 0.3)0.599 ± 0.0530.638 ± 0.1030.872 ± 0.012
(0.35, 0.35, 0.3)0.669 ± 0.0610.522 ± 0.1490.876 ± 0.011
(0.4, 0.3, 0.3)0.744 ± 0.0530.383 ± 0.1630.867 ± 0.015
(0.5, 0.2, 0.3)0.801 ± 0.0460.282 ± 0.1390.855 ± 0.018
(0.7, 0.0, 0.3)0.834 ± 0.0440.242 ± 0.1090.836 ± 0.020

As we shifted weight from QED to sEH, mean sEH increased from 0.466 to 0.834 and mean QED decreased from 0.768 to 0.242. Both means moved in the expected direction across all seven tested preferences. Quantifying any change in steering strength would require evaluating the original configuration with both normalization and curriculum at the same seven preferences.

Q3 — Evaluating preference control

We recommend a fixed-budget preference sweep. For each training seed, hold the checkpoint fixed, vary only the preference weights, and report mean property scores at each weight. The direction and consistency of the resulting curves show whether the model follows the requested trade-off, while the absolute scores and variation across seeds show what it achieves and how reliably. Preference weights specify relative priorities. When the target distribution can be computed, as in our grid experiment, we assess calibration using the L1L_1 error between the learned and target distributions.

Dear Reviewer F1pr,

We would like to follow up as the author–reviewer discussion period will close in less than 10 hours.

In response to your feedback, we have added experiments, statistical analyses, and clarifications. We hope these additions address your concerns.

If you feel that our responses have addressed your concerns, we would be grateful if you could reconsider your evaluation. If any questions remain, we would be happy to provide further clarification.

Thank you for your time and consideration.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.