acceptodds
Under review as a conference paper at ICLR 2027

Generator Disagreement Acquisition for Sample-Efficient Molecular Discovery

Abstract

Goal-directed molecular design often emphasizes the best molecule found, whereas practical discovery campaigns require multiple promising candidates under a limited budget of expensive oracle evaluations. In parent-conditioned search, however, a parent's observed reward need not reflect its ability to generate valuable offspring. This motivates selecting parents using predicted offspring reward distributions rather than their own scores alone. We introduce Generator Disagreement Acquisition (GDA), a closed-loop framework that allocates oracle queries at the parent level. GDA samples offspring from each candidate parent and combines the mean and dispersion of their surrogate-predicted rewards to select which parents supply candidates for oracle evaluation. This dispersion, termed generator disagreement, measures variation in predicted rewards across offspring rather than disagreement among predictors for a fixed molecule. Oracle feedback updates both the surrogate and the parent pool. We evaluate GDA on six tasks across DOCKSTRING and GEAM. All methods start from the same 512 labeled molecules and receive 800 additional oracle evaluations. With random-forest ensembles, GDA improves mean top-1% hit yield over a screened genetic algorithm using the same proposal operator and surrogate family by 38-61% on DOCKSTRING and 16-33% on GEAM. It also improves hit yield over Augmented Memory by 27-69% on DOCKSTRING, while Augmented Memory remains competitive on two GEAM tasks. These results support parent-level acquisition from predicted offspring distributions as a sample-efficient strategy for identifying multiple high-scoring candidates for downstream evaluation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular discovery, generative molecular design, black-box optimization, acquisition functions, sample efficiency.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_yEX5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how to discover multiple high-reward molecules under a limited evaluation budget. At each round, GDA generates and screens offspring for every molecule in an evaluated parent memory. It scores each eligible parent using the mean and sample standard deviation of surrogate-predicted rewards across its surviving offspring. Acquisition and structural coverage select parents; a separate combination of predicted reward, ensemble spread, and novelty selects one child per chosen parent.

Critical strengths

S1. The paper addresses a meaningful problem: discovering multiple promising molecules under a limited budget of expensive evaluations. This objective reflects practical discovery campaigns, where several candidates are needed for downstream screening and validation. S2. The paper is clearly presented and well organized. Its progression from motivation and method to experiments makes the proposed approach and supporting results easy to follow.

Critical weaknesses

W1. The paper's technical novelty is not sufficiently clear. PopVar predicts simulated offspring performance [1], and Osthushenrich et al. combine offspring means and standard deviations to rank crosses [2]. The authors should explain how GDA extends these breeding principles, which difficulties in molecular search require new designs, and why its score is preferable to simpler alternatives. W2. The external baseline comparison does not adequately cover recent methods. The six comparison entries include only three external algorithms, Graph GA (2019), REINVENT (2017), and Augmented Memory (2024), alongside two internal controls and Random. Comparisons with suitable recent methods such as GenMol, Saturn, or SEGO are needed to establish GDA's competitiveness[3-5]. References [1] PopVar: A Genome-Wide Procedure for Predicting Genetic Variance and Correlated Response in Biparental Breeding Populations. [2] Genomic selection of crossing partners on basis of the expected mean and variance of their derived lines. [3] GenMol: A Drug Discovery Generalist with Discrete Diffusion. [4] Sample-efficient generative molecular design using memory manipulation. [5] Sample Efficient Generative Optimization for Molecular Design.

Questions for authors

Q1. What is new in GDA compared with the progeny-selection principles in [1,2], and what difficulty in molecular search does it address? Q2. Can you justify the baseline selection and compare GDA with suitable recent methods, prioritizing Saturn and one of GenMol or SEGO where feasible?

8 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for recognizing the practical importance of discovering multiple promising molecules within a limited evaluation budget, and for finding the paper's organization and progression clear. We particularly appreciate the breeding references and the specific baseline suggestions. They help us explain the contribution more precisely and test its competitiveness against recent methods. To help the reviewer locate the changes, we have highlighted new and revised text in blue in the updated manuscript.

1. Relationship to PopVar and offspring-based selection in breeding.

We thank the reviewer for highlighting this relevant connection. PopVar predicts progeny performance through simulation and genomic prediction, while Osthushenrich et al. derive progeny means and variances to guide the selection of crossing partners. These studies establish a connection between offspring distributions and parental value.

GDA develops a parent-level acquisition framework for molecular discovery that connects this distributional information to the allocation of expensive oracle evaluations. Its contribution lies in coordinating parent selection, child acquisition, and adaptive memory using the offspring that survive molecular screening. The central decision is which parents should supply the limited number of molecules evaluated in each round, and how their observed outcomes should guide subsequent search. Three aspects of this setting shape the method:

  • The usable offspring distribution depends on the operator and screening. Graph crossover/mutation or fragment recombination proposes molecular structures, some of which are invalid, fail feasibility filters, or have already been queried. GDA estimates reward statistics on the surviving proposals. The interface requires samples and surrogate evaluations, allowing the same acquisition rule to operate with different parent-conditioned graph operators.
  • Parent acquisition and expensive child evaluation are separate decisions. Many offspring can be predicted cheaply, but a selected parent contributes at most one child to a small oracle batch. GDA specifies both which parents supply candidates and how to choose their children using predicted reward, ensemble spread, and novelty. Deduplication handles offspring shared by several parents.
  • The search updates after each evaluated batch. New labels refit the surrogate and update a bounded parent memory. The implementation combines offspring-based acquisition, structural coverage, and descendant-informed memory quality so these decisions can be repeated under an explicit evaluation cap.

The methodological contribution is the joint design of these decisions within a closed-loop molecular search procedure. GDA scores parents through their screened offspring distributions, selects individual children for evaluation, and uses the resulting feedback to update both predictions and the parent memory. The standard expected-maximum bound motivates the rollout-dependent exploration scaling. Controlled ablations and comparisons with maximum and top-kk parent scores assess the empirical value of the resulting acquisition rule. We have sharpened the related-work discussion and contribution statement to make this connection to breeding research and GDA's specific algorithmic contribution clear.

AuthorsOpenReview ↗

Rebuttal by Authors

2. Is the score preferable to simpler alternatives?

We have tested parent selection by the maximum predicted offspring reward and by the mean of the top four predictions. Both retain GDA's generation, screening, child selection, coverage, and memory-update rules. The following are distinct top-1% hits, mean ± standard error across five matched seeds:

TaskGDAOffspring maxTop-4 mean
F2379.0 ± 18.6303.6 ± 59.5316.4 ± 33.3
JAK2186.6 ± 28.9204.2 ± 28.8183.2 ± 34.3
PPAR403.6 ± 32.1376.6 ± 30.7325.2 ± 10.9
FA7461.4 ± 39.0400.4 ± 66.2436.8 ± 30.2
BRAF495.0 ± 24.5529.6 ± 23.6500.8 ± 31.7
PARP1373.2 ± 44.6382.8 ± 38.9423.4 ± 41.9

GDA has the highest mean of these three rules on F2, PPAR, and FA7. The alternatives are competitive on the remaining tasks, so we present GDA as an effective choice within this framework rather than a uniformly superior functional form. The paired gain over Top-4 mean on PPAR is 78.4 ± 24.4 hits, positive in all five seeds.

The more direct test of the dispersion component is c=0c = 0 with all other rules retained. Keeping the bonus yields paired gains of 127.8 ± 31.7 on F2, 119.6 ± 40.3 on PPAR, and 106.0 ± 34.5 on FA7, each positive in all five seeds. The other three tasks have smaller average gains with greater seed uncertainty. The revised manuscript now shows the disagreement ablation for all six tasks. This evidence supports the usefulness of offspring dispersion in several molecular search settings while preserving the distinction between a design motivation and an empirical comparison.

AuthorsOpenReview ↗

Rebuttal by Authors

3. Recent external baselines: Saturn, GenMol, and GEAM.

Following the reviewer's suggested priority, we have completed Saturn and GenMol on all six tasks with five seeds per method. We also evaluated the GEAM method and the sequential-restart control on all six tasks.

All methods receive the same 512 initial labels and a cap of 800 online evaluations. Below, OB@100 is the mean cumulative number of calls needed to obtain 100 distinct top-1% hits, measured at batch completion over successful seeds. Parentheses give success counts below five. A dash means no seed reaches the count.

MethodF2 hitsJAK2 hitsPPAR hitsOB@100: F2 / JAK2 / PPAR
GDA379.0 ± 18.6186.6 ± 28.9403.6 ± 32.1810 / 1050 / 807
Saturn292.2 ± 23.8225.2 ± 18.8351.2 ± 32.81014 / 1072 / 942
GenMol66.4 ± 7.058.0 ± 10.491.6 ± 8.1— / — / 1256 (2/5)
GEAM method60.0 ± 5.043.4 ± 6.162.2 ± 12.2— / — / —
Augmented Memory223.8 ± 43.8142.6 ± 18.5319.0 ± 13.91118 / 1194 (4/5) / 987
REINVENT53.8 ± 7.532.6 ± 3.067.8 ± 14.2— / — / 1280 (1/5)
Graph GA111.0 ± 21.885.8 ± 7.0188.2 ± 17.11172 (2/5) / 1304 (2/5) / 1054

GDA reaches 100 hits with lower mean oracle burden than every external baseline in this table on all three tasks, and reaches the count in every seed. It also has the highest final mean yield on F2 and PPAR. On JAK2, it reaches 100 sooner than Saturn, while Saturn achieves a higher final yield. This distinction is useful for campaigns with different stopping requirements. Both results are reported in the revised manuscript.

The GEAM-target comparison shows the same early-discovery strength. On FA7/BRAF/PARP1, GDA reaches 100 hits in 751/721/806 cumulative calls, versus Saturn's 861/814/875, with all seeds successful. Final mean yields are 461.4/495.0/373.2 for GDA and 368.8/511.2/440.8 for Saturn.

The newly completed GEAM-method and GenMol results on the GEAM targets are:

MethodFA7 hitsBRAF hitsPARP1 hitsOB@100: FA7 / BRAF / PARP1
GDA461.4 ± 39.0495.0 ± 24.5373.2 ± 44.6751 / 721 / 806
GEAM method2.2 ± 0.98.6 ± 1.34.8 ± 1.7— / — / —
GenMol51.8 ± 5.797.2 ± 9.670.8 ± 5.8— / 1240 (3/5) / —

Hit yields are means ± standard errors across five seeds. GDA has higher mean hit yield than both added baselines on every GEAM target and reaches 100 hits in all five seeds. GenMol reaches 100 hits on BRAF in three seeds, while the GEAM method does not reach 100 on any target. These comparisons evaluate the methods under the stated initialization and evaluation protocols.

Implementation and comparison scope. Saturn uses our implementation of its Mamba-based replay and memory-manipulation procedure. GEAM and GenMol use their released implementations adapted to our oracle and budget. GEAM's fragment scorer and initial vocabulary use the shared 512 labels. GenMol uses its pretrained model with an initial fragment vocabulary derived from those labels and a warmup scaled to the 800-call budget. We have documented these choices in the appendix. This comparison uses the same initial molecules and the stated evaluation protocols.

Finally, the recorded budget profiles help explain the value of oracle burden. At 200 online calls, GDA obtains 55.6/23.6/58.0 hits on F2/JAK2/PPAR, compared with Saturn's 14.6/13.4/23.2. The lead narrows or changes by the endpoint on some tasks. We now report oracle burden in the main results, retain final hit yield, and include the budget profiles in the appendix. The profiles describe completed-batch prefixes of the recorded searches. They do not establish a causal explanation for the learning dynamics.

We have updated the related-work discussion, contribution statements, baseline descriptions, and results accordingly. These new experimental results clearly demonstrate that GDA makes offspring-distribution-based parent acquisition operational for screened molecular search and provides strong early discovery performance under the tested budgets, with controlled evidence showing where its disagreement term contributes.

Reviewer_L5UKOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes the Generator Disagreement Acquisition (GDA) framework for generating multiple desirable molecules with high property scores. GDA generates children molecules from candidate parents, and select the parents based on the mean and variance of the surrogate-predicted rewards achieved by the corresponding children. In addition, GDA considers structural coverage when selecting the candidates. Experiments on six tasks across DOCKSTRING and GEAM demonstrate that GDA achieve higher performance than several representative baselines.

Critical strengths
  1. This paper is well-motivated: finding multiple molecules with desirable properties is meaningful in some real-world scenarios.

  2. The idea of evaluating parents based on the predicted reward distribution of their children instead of solely considering child molecules is insightful.

  3. The experiemental evaluation is comprehensive: in addition to performance comparison, the experiments also examine component contributions and the effects of surrogate models, and other key factors.

Critical weaknesses
  1. The benefit of parent-based acquisition is not completely isolated from other factors. Though GDA is compared with pooled candidate selection, the two methods also differ in their memory-update strateges. So the comparison cannot demonstrate that the improvement all comes from the proposed acquisition method.

  2. The rationale for encouraging high dispersion is not fully convincing. A large dispersion can also reflect poor lower-tail candidates, rather than always meaning a good results.

  3. There is no experimental results demonstrate that the molecules generated by GDA possess higher structural diversity than representative baselines.

  4. Comparisons with strong and closely related methods such as GEAM and Saturn are missing.

Questions for authors

Have the authors considered a simple sequential-restart baseline that repeatedly searches for a high-scoring molecule while excluding previously found ones, under the same total oracle budget?

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for recognizing the practical motivation of finding multiple promising molecules, the insight in evaluating parents through their offspring's predicted rewards, and the breadth of the experimental analysis. The questions about attribution, structural diversity, and restarting are helpful because they ask what a discovery campaign actually gains from the method. We address each below with additional controls and a more precise account of the claims. To help the reviewer locate the changes, we have highlighted new and revised text in blue in the updated manuscript.

Results below use five matched seeds, 512 shared initial labels, and a cap of 800 additional evaluations. Hit yield counts distinct online top-1% hits. Values after ± are standard errors. Oracle burden includes initialization and is measured at batch completion.

1. Separating acquisition from memory updates.

The submitted manuscript already included a controlled test of the disagreement term in Figure 4, “GDA with and without the disagreement bonus. Setting c=0c = 0 retains the mean offspring score and the other selection components.” in Section 4.3 (RQ2). This experiment sets c=0c = 0 while retaining GDA's proposal operator, surrogate, coverage, child-selection rule, and memory-update rule. The F2, JAK2, and PPAR results below were already reported in that section.

This control isolates the contribution of disagreement within GDA. We agree that the separate comparison with Pooled children evaluates complete search policies, since their memory-update rules also differ. We have made this distinction more explicit so that the pooled comparison is not read as attributing the entire gain to acquisition alone.

Building on the submitted DOCKSTRING ablation, we have now added the same control on the three GEAM tasks in the revised manuscript:

TaskPaired hit-yield gain from retaining disagreementSeeds favoring GDA
F2+127.8 ± 31.75/5
JAK2+22.4 ± 37.63/5
PPAR+119.6 ± 40.35/5
FA7+106.0 ± 34.55/5
BRAF+0.2 ± 24.32/5
PARP1+2.6 ± 66.93/5

The submitted ablation shows gains in every seed on F2 and PPAR, and the new GEAM experiments extend this evidence to FA7. The magnitude of the benefit varies by task. We also tested ζ=1ζ = 1 on F2: removing the coverage term gives 369.0 ± 34.5 hits versus GDA's 379.0 ± 18.6. This is a joint change to parent-selection and memory coverage, so it does not isolate memory alone. We do not attribute the entire GDA–Pooled gap to acquisition or treat these separate ablations as an additive decomposition.

2. Could dispersion reward a poor lower tail?

Yes. A large standard deviation alone does not establish a better upper tail, and the population bound does not guarantee that the empirical score orders parents correctly. Our rationale is that GDA screens multiple offspring cheaply and chooses one child from each selected parent for an expensive evaluation. This makes the opportunity to select a good child relevant, rather than requiring every generated child to be good. The mean term also remains part of the parent score.

To check how selection behaves, we examined successful, distinct online queries within the paper's budget and located each child's prediction within its proposing parent's predicted offspring distribution. For parents with nonzero dispersion, the mean within-run fraction of queries above the parent's predicted mean is 98.4% on F2, 91.3% on JAK2, and 97.6% on PPAR. Their mean standardized positions are +1.45, +1.35, and +1.36, respectively. These observations show that the child-selection stage generally chooses the upper part of the sampled predictions. They do not prove that every dispersed parent has a better oracle-reward distribution.

We state this distinction explicitly in the revised manuscript. Child selection also includes ensemble spread and novelty, so it does not need to choose the largest predicted mean. The controlled c=0c = 0 results above provide the empirical evidence for the bonus. We have additionally evaluated maximum and top-4 parent scores on all six tasks. They are useful alternatives, with task-dependent rankings, and are included in the revision.

AuthorsOpenReview ↗

Rebuttal by Authors

3. Structural diversity of the discovered molecules.

Our primary objective is to find many distinct molecules above a reward threshold with few oracle evaluations. This objective does not require GDA to achieve the highest structural diversity, but it is possible that distinct molecules can still be closely related analogues. GDA includes coverage and novelty terms to encourage exploration, without guaranteeing greater diversity in the final hit set. To assess structural breadth alongside discovery efficiency, we have therefore added scaffold-level hit yield, which counts distinct Bemis–Murcko scaffolds among the top-1% hits.

TaskGDASaturnAugmented Memory
F2171.4 ± 62.0175.4 ± 4.8149.4 ± 20.1
JAK254.0 ± 11.5110.8 ± 7.977.8 ± 5.6
PPAR257.6 ± 43.0195.0 ± 20.3184.4 ± 34.1
FA7281.0 ± 108.7144.0 ± 12.593.4 ± 9.9
BRAF116.2 ± 26.0235.0 ± 11.9181.8 ± 9.8
PARP1117.2 ± 29.3196.8 ± 12.1201.0 ± 20.0

Among the methods shown, GDA achieves the highest mean scaffold-level hit yield on PPAR and FA7. Its F2 mean is close to Saturn's, although variation across seeds is substantial. Saturn or Augmented Memory achieves higher scaffold yields on the remaining tasks. These results show that GDA's hit-yield gains can extend across multiple scaffolds, while the structural breadth depends on the task. We have reported this alongside discovery efficiency and retained the internal-diversity comparisons in the appendix. Our claim is efficient discovery of high-reward candidates, without assuming that higher molecule-level hit yield implies greater structural diversity.

4. Adding GEAM, Saturn, and GenMol.

We have completed Saturn, GEAM, and GenMol on all six tasks with five seeds per method. The DOCKSTRING comparisons are shown below, followed by the GEAM-task results.

MethodF2 hitsJAK2 hitsPPAR hitsOB@100: F2 / JAK2 / PPAR
GDA379.0 ± 18.6186.6 ± 28.9403.6 ± 32.1810 / 1050 / 807
Saturn292.2 ± 23.8225.2 ± 18.8351.2 ± 32.81014 / 1072 / 942
Augmented Memory223.8 ± 43.8142.6 ± 18.5319.0 ± 13.91118 / 1194 (4/5) / 987
GEAM method60.0 ± 5.043.4 ± 6.162.2 ± 12.2— / — / —
GenMol66.4 ± 7.058.0 ± 10.491.6 ± 8.1— / — / 1256 (2/5)

OB@100 averages the cumulative calls to reach 100 hits over successful seeds. Parentheses show success counts below five. A dash means no seed reached 100. GDA and Saturn reach the count in every seed on all six tasks.

GDA reaches 100 hits sooner than these external baselines on all three DOCKSTRING tasks and has higher final mean yield on F2 and PPAR. JAK2 illustrates why both metrics matter: GDA reaches 100 sooner, whereas Saturn accumulates more hits by the end. On FA7/BRAF/PARP1, GDA's mean OB@100 is 751/721/806, versus Saturn's 861/814/875.

The newly completed GEAM-method and GenMol results on the GEAM targets are:

MethodFA7 hitsBRAF hitsPARP1 hitsOB@100: FA7 / BRAF / PARP1
GDA461.4 ± 39.0495.0 ± 24.5373.2 ± 44.6751 / 721 / 806
GEAM method2.2 ± 0.98.6 ± 1.34.8 ± 1.7— / — / —
GenMol51.8 ± 5.797.2 ± 9.670.8 ± 5.8— / 1240 (3/5) / —

Hit yields are means ± standard errors across five seeds. GDA has higher mean hit yield than both added baselines on every GEAM target and reaches 100 hits in all five seeds. GenMol reaches 100 hits on BRAF in three seeds, while the GEAM method does not reach 100 on any target. These comparisons evaluate the methods under the stated initialization and evaluation protocols.

The comparisons adapt the baselines to our shared objective, feasibility filters, initialization, and evaluation cap. Saturn uses our implementation of its Mamba/replay/filtering procedure. GEAM and GenMol use their released implementations with an oracle interface. In the original GEAM setup, the fragment scorer is trained on ZINC250k and used to construct the initial fragment vocabulary. Our adaptation builds these components using labels from the 512 initial molecules. This substantially reduces the task-specific training data available to GEAM and may contribute to its lower hit yields under our limited-label protocol. GenMol uses its pretrained model, a vocabulary built from those labels, and a budget-scaled warmup. We have documented these adaptations in the appendix, since the results concern this limited-label protocol rather than unrestricted performance of the original methods.

AuthorsOpenReview ↗

Rebuttal by Authors

5. Sequential search with restarts.

We have implemented and evaluated this suggested control on all six tasks. It uses the screened-GA search with memory resets every 12 rounds, retains accumulated labels and the surrogate, and excludes previously queried identities through the shared screening rule. At a reset, it excludes the outgoing parent memory and current batch from subsequent parent reseeding and preferentially draws from unused initial molecules. Thus, it tests repeated parent-memory restarts under the same total budget, rather than independent runs with fresh labels.

TaskSequential restart hitsGDA hits
F247.0 ± 6.0379.0 ± 18.6
JAK238.6 ± 4.6186.6 ± 28.9
PPAR75.8 ± 6.6403.6 ± 32.1
FA7142.2 ± 5.7461.4 ± 39.0
BRAF224.6 ± 22.0495.0 ± 24.5
PARP1120.4 ± 7.3373.2 ± 44.6

All entries report mean ± standard error across five seeds. Full per-seed results are now included in the appendix. GDA achieves higher mean yield on every task. This supports retaining and adapting productive parents under the tested budget, while leaving open whether a different restart schedule could improve the control. We have now added this as a named baseline and described its reset policy explicitly.

We thank the reviewer for the detailed suggestions and we believe the revision makes the paper much stronger. We have clarified what the c=0c = 0 ablation tests and how it differs from the comparison with Pooled children. We also report scaffold-level hit yield and results for the new baselines, including sequential restarts.

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

GDA is a parent-selection rule for generative molecular search under a fixed oracle budget. A normal method picks which molecule to mutate from its own observed reward. GDA instead generates offspring from every parent in memory, scores them with a cheap surrogate ensemble, and selects parents by the mean plus the dispersion of their offspring's predicted rewards. The dispersion term is called generator disagreement. It measures variation across one parent's offspring. A second rule then picks one child per selected parent, using predicted reward, ensemble spread and novelty. The evaluation uses six tasks from DOCKSTRING and GEAM, with 512 shared initial labels and 800 oracle calls.

Critical strengths

The core idea is new and clearly stated, and the separation of offspring dispersion (σ_g) from ensemble uncertainty about one molecule (σ_r) is a real distinction that Table 1 positions correctly against prior work. The control chain with Pooled children, GA+screen, c = 0 and full GDA, where the c = 0 ablation isolates the actual contribution. Appendices A and B scope the theory, including a counterexample where the score falls below the sampled maximum (0.651 against 1). The shift from best-found reward to hit yield and oracle burden is a genuine methodological contribution.

Critical weaknesses
  • The simplest baseline is missing. Proposition 1 motivates mean + κ·σ as an estimate of the maximum of K offspring. If the maximum is what matters, then use the maximum of the predicted rewards directly, or a top-k mean. This needs no new machinery. It is never tested, so the paper cannot show that mean plus dispersion beats a plain maximum. Please add both baselines.
  • The disagreement term helps on some tasks only. The c = 0 ablation gives 127.8 ± 31.7 hits on F2 and 119.6 ± 40.3 on PPAR, better in every seed. On JAK2 it gives 22.4 ± 37.6, better in three of five seeds. The three GEAM tasks are never ablated. So the main claim holds on two tasks, is null on one, and is untested on three. The abstract gives one percentage range and hides this. Please ablate the GEAM tasks and give per-task numbers in the main text.
  • The hyperparameters are fixed and never swept: c = 1, K = 16, ζ = 0.7, γ̃ = 1, J = 64, B = 8.
  • Hit yield ignores structural redundancy. It counts distinct molecules above a threshold. One good scaffold plus 400 close analogs scores very well. A parent-level rule that reuses a good parent benefits from this. Scaffold counts appear in Appendix J, but not next to the main results. Please report scaffold-level hit yield too.
  • The surrogate dependence is not explained. RF beats GNN on every task and every rule. What is authors' hypothesis for this behaviour?
  • The budget never varies: 512 initial labels and 800 online calls. The paper is about sample efficiency, so budget scaling will be interesting to review and include in the paper.
  • Limitations section is missing.
Questions for authors

Please see weaknesses.

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for recognizing the clarity of the core idea, the distinction between offspring dispersion and ensemble uncertainty, and the practical value of evaluating hit yield and oracle burden. We also appreciate the careful reading of the controls and the scope of the theory. The questions help distinguish which parts of the method are supported by controlled evidence and which remain design choices. To help the reviewer locate the changes, we have highlighted new and revised text in blue in the updated manuscript.

Below, hit yield means distinct online top-1% hits. Results are means ± standard errors across five matched seeds. All comparisons use 512 shared initial labels and a cap of 800 additional evaluations. Oracle burden includes initialization and is measured at batch completion.

1. Maximum and top-kk parent scores.

We have evaluated both requested alternatives on all six tasks. They replace only the parent score: Offspring max uses the largest predicted offspring reward. Top-4 mean averages the four largest, or all survivors if fewer than four remain. Generation, screening, coverage, child selection, and the memory-update rule are retained.

TaskGDAOffspring maxTop-4 mean
F2379.0 ± 18.6303.6 ± 59.5316.4 ± 33.3
JAK2186.6 ± 28.9204.2 ± 28.8183.2 ± 34.3
PPAR403.6 ± 32.1376.6 ± 30.7325.2 ± 10.9
FA7461.4 ± 39.0400.4 ± 66.2436.8 ± 30.2
BRAF495.0 ± 24.5529.6 ± 23.6500.8 ± 31.7
PARP1373.2 ± 44.6382.8 ± 38.9423.4 ± 41.9

GDA achieves the highest mean hit yield among the three rules on F2, PPAR, and FA7, showing gains across both benchmark families. On PPAR, its paired improvement over Top-4 mean is 78.4 ± 24.4 hits, with gains in every seed. On F2 and FA7, GDA also exceeds both alternatives in mean hit yield. Offspring max leads on JAK2 and BRAF, while Top-4 mean leads on PARP1, indicating that the most effective parent score depends on the task.

The proposition motivates the offspring-count scaling, and these controlled comparisons assess the resulting empirical parent-selection rule. Together with the disagreement ablation below, they provide evidence that using offspring mean and dispersion can improve discovery under a fixed oracle budget. We have reported the per-task results and paired differences to make the settings where GDA provides the greatest benefit clear.

2. Does the disagreement term help on the GEAM tasks?

We have completed the c=0c = 0 ablation on all three GEAM tasks. This control retains GDA's other components and removes only the disagreement bonus. The paired change in final hit yield is:

TaskGDA minus c=0c = 0Seeds favoring GDA
F2+127.8 ± 31.75/5
JAK2+22.4 ± 37.63/5
PPAR+119.6 ± 40.35/5
FA7+106.0 ± 34.55/5
BRAF+0.2 ± 24.32/5
PARP1+2.6 ± 66.93/5

The new GEAM ablation extends the submitted DOCKSTRING evidence: retaining the disagreement bonus adds more than 100 top-1% hits on average on F2, PPAR, and FA7, with improvements in every seed on each task. Because the control retains the other GDA components, these gains directly demonstrate the value of the disagreement term within the framework. The remaining tasks show smaller mean gains and greater variation across seeds, identifying where the bonus has a more limited observed effect. We have added relevant figures in the revised manuscript showing the disagreement ablation on all six tasks, highlighting these controlled gains separately from comparisons between complete methods.

AuthorsOpenReview ↗

Rebuttal by Authors

3. Hyperparameter sensitivity.

The submitted Method already distinguishes coefficients computed from the current search state from settings chosen for the experiment. In particular, the child-acquisition coefficients follow explicit formulas and are not tuned separately for each task. We have made this distinction clearer alongside the new sensitivity results.

We have added sweeps of the disagreement multiplier cc on F2 and JAK2 and the shared quality–coverage weight ζ\zeta on F2. Results are mean ± standard error across five seeds, with the submitted defaults unchanged.

Taskc = 0c = 0.5c = 1, defaultc = 2
F2251.2 ± 31.9320.8 ± 40.9379.0 ± 18.6374.8 ± 36.3
JAK2164.2 ± 10.5215.6 ± 26.0186.6 ± 28.9214.0 ± 34.5

Every tested positive value of cc gives higher mean hit yield than c=0c=0 on both tasks. F2 achieves similar mean yields at c=1c=1 and c=2c=2, while JAK2 achieves its highest observed mean at c=0.5c=0.5. These results support the usefulness of disagreement across the tested positive weights, without requiring a task-specific choice of the best observed setting.

For F2, ζ=0.5\zeta=0.5, 0.70.7, and 1.01.0 yield 304.6±26.1304.6\pm26.1, 379.0±18.6379.0\pm18.6, and 369.0±34.5369.0\pm34.5 hits, respectively. The same ζ\zeta controls coverage in both parent selection and memory updates. The submitted SA ablation additionally examines feasibility requirements, with increased proposal-generation allowances at stricter cutoffs.

Why some coefficients were specified by formula.

  • Parent exploration scaling. The standard expected-maximum bound motivates the 2log⁡K\sqrt{2\log K} dependence in κ=c2log⁡K\kappa=c\sqrt{2\log K}. The new cc sweep directly tests the strength of this bonus while holding the actual rollout count fixed, separating acquisition weighting from changes in the sampled candidate pool.

  • Child ensemble-spread weighting. We compute βt=2log⁡max⁡{Nt,1}\beta_t=\sqrt{2\log\max\{N_t,1\}} from the number of surviving offspring in each round. Thus, its value adapts to screening outcomes without a separately tuned multiplier. As stated in the submitted appendix, this specifies a ranking rule rather than a calibrated confidence guarantee.

  • Novelty scaling. We recompute γt=γ~ med⁡(σr)/med⁡(novt)\gamma_t=\tilde\gamma\,\operatorname{med}(\sigma_r)/\operatorname{med}(\mathrm{nov}_t) each round, subject to the numerical safeguard described in the Method. The default γ~=1\tilde\gamma=1 matches the median novelty bonus to the median unweighted ensemble spread. This provides an explicit scale convention across tasks, while βt\beta_t separately weights ensemble spread.

We held K=16K=16, J=64J=64, and B=8B=8 fixed to keep proposal effort, memory capacity, and evaluation cadence comparable across the acquisition-weight sweeps. These are computational design choices, and their sensitivity is not established by the cc and ζ\zeta experiments. The novelty multiplier γ~\tilde\gamma also remains fixed. We have clearly separated formula-defined coefficients, fixed computational settings, and directly tested parameters in the revision.

4. Structural redundancy and scaffold-level hit yield.

We agree that distinct molecule counts alone cannot establish structural breadth. We have therefore computed distinct Bemis–Murcko scaffolds among the top-1% hits, using the same scaffold convention across both benchmarks, and have reported these counts alongside the main results.

TaskGDASaturnAugmented Memory
F2171.4 ± 62.0175.4 ± 4.8149.4 ± 20.1
JAK254.0 ± 11.5110.8 ± 7.977.8 ± 5.6
PPAR257.6 ± 43.0195.0 ± 20.3184.4 ± 34.1
FA7281.0 ± 108.7144.0 ± 12.593.4 ± 9.9
BRAF116.2 ± 26.0235.0 ± 11.9181.8 ± 9.8
PARP1117.2 ± 29.3196.8 ± 12.1201.0 ± 20.0

GDA's gains extend to higher mean scaffold-level hit yield than these two baselines on PPAR and FA7. Its F2 mean is close to Saturn's, with substantial seed variation. Saturn and Augmented Memory provide greater scaffold breadth on JAK2, BRAF, and PARP1. Thus, the practical choice depends on whether a campaign prioritizes rapid accumulation of high-scoring molecules or broader scaffold coverage. We have retained internal diversity as a complementary measure and included per-seed scaffold counts, so neither structural concentration nor variability is hidden by a mean.

AuthorsOpenReview ↗

Rebuttal by Authors

5. Why might RF outperform GNN surrogates?

We suspect that the limited training data favor RF’s use of Morgan fingerprints. Each run starts with 512 labeled molecules and adds at most 800 labels. RF learns the reward mapping from a fixed molecular representation, while the tested GNN must also learn its representation from these data. This additional learning requirement may make accurate reward prediction more difficult for the GNN. Prediction errors can then affect both the ranking of individual candidates and the offspring statistics used to select parents.

RF and GNN also visit different molecules during search, so differences in their observed offspring dispersion cannot directly establish which surrogate predicts more accurately. Evaluating their prediction errors on the same molecules would help test this explanation. We have added this discussion to RQ3. Although RF performs better overall in these experiments, GDA achieves higher mean hit yield than Pooled children with both surrogate families on all three DOCKSTRING tasks.

6. Budget scaling.

We have evaluated the completed-batch prefixes of the existing runs at 200, 400, 600, and 800 online calls. This measures how quickly the recorded searches accumulate hits, with initialization fixed. It is not a separate experiment that changes the initial training set or retunes each budget.

For example, at 200 online calls GDA finds 55.6, 23.6, and 58.0 hits on F2, JAK2, and PPAR, compared with Saturn's 14.6, 13.4, and 23.2. The early advantage also holds on FA7, BRAF, and PARP1: 79.0, 94.4, and 64.4 versus 42.8, 51.8, and 34.2. Later in the runs, Saturn achieves higher final yield on JAK2, BRAF, and PARP1. The full profiles therefore identify where GDA is most useful: obtaining an initial set of high-scoring molecules with few evaluations.

Consistent with this profile, GDA reaches 100 hits with lower mean oracle burden than the tested external baselines on all six tasks. We now report oracle burden in a main-paper table alongside final yield. The alternative parent-score controls remain visible, including cases where they reach 100 hits sooner than GDA.

7. Limitations and paper changes.

We thank the reviewer for encouraging a clearer discussion of limitations. Given the page limit, we incorporated this discussion into the Conclusion, where the submitted manuscript already notes task-dependent gains, trade-offs involving structural diversity, and reliance on computational molecular objectives.

We agree that these limitations should be more explicit. We have briefly expanded the Conclusion to discuss the structural concentration. Detailed sensitivity and budget analyses now appear in the appendix. This keeps the limitations visible in the main paper while making room for the additional experiments requested by the reviewers.

AuthorsOpenReview ↗

Follow-up on our rebuttal and manuscript revisions

Dear Reviewer KjZF,

Thank you again for your positive and constructive feedback.

Our rebuttal addresses the requested parent-score comparisons, GEAM ablations, hyperparameter sensitivity, scaffold-level hit yield, surrogate behavior, budget dependence, and limitations.

Since the author-reviewer discussions are about to close soon, could you please review our rebuttal and let us know whether it addresses all your concerns?

Thank you for your time and effort.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.