acceptodds
Under review as a conference paper at ICLR 2027

DualEvo: Chemical-Space-Aware Online Evolution for Dual-Target Molecular Generation

Abstract

Dual-target molecular generation seeks molecules that satisfy activity objectives for two biological targets while maintaining useful exploration of chemical space. Existing generative and feedback-learning methods can enrich high-value candidates, but continual generation also requires deciding how evaluated experience is accumulated and which structures are reused for updating. We introduce DualEvo, a chemical-space-controlled online evolution framework that combines fixed dual-target evaluation, cumulative experience archiving, reward-guided filtering, scaffold-balanced selection, and lightweight LoRA updates. Reward ranks candidate value, whereas scaffold balancing controls the structural composition of the high-value update set. Across six tasks, DualEvo attains the highest six-task mean unique predicted-success yield at the final 500-sample readout; relative to Reward-only, per-task means are higher on five of six tasks and successful-scaffold coverage is broader. Larger fixed-checkpoint readouts show more distinct successful molecules and scaffolds despite a lower successful-event frequency. Saved LoRA states can be reused for independent sampling and as fixed proposal models in downstream search.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Dual-target Molecule Generation, de novo drug design, Reward-guided Online Adaptation, Parameter-efficient Fine-tuning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vnHtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies dual-target molecular generation, focusing on two coupled questions in continual online generation: how evaluated experience should accumulate across stages, and which high-value structures should be reused for parameter updates. The authors propose DualEvo, which organizes molecular generation, fixed dual-target evaluation, cumulative experience archiving, experience selection, and lightweight LoRA updates into a continual loop. The core design separates two functions: reward determines which archived molecules are high-value for the task, whereas scaffold balancing determines how structurally concentrated the selected update set is. The Reward-only condition simply selects the top N records from the high-value pool; DualEvo instead groups that pool by Bemis-Murcko scaffold, preserves reward order within each group, orders groups by their best record, and then selects one record from each nonempty group per round until the same update budget is filled. Experiments show the effectiveness of the modules design.

Critical strengths
  1. The paper provides thorough ablations of the components it proposes. Experience accumulation, reward direction , scaffold balancing, and the learning interface are each isolated and tested, with all conditions sharing the same base checkpoint, training exposure, and online budget so that one factor varies at a time.
Critical weaknesses
  1. All comparisons in this paper are ablations of the authors’ own design. The four conditions in Table 1 (Frozen, One-shot, Reward-only, DualEvo) come from the same pipeline, and REINFORCE-LoRA is also a control built by the authors rather than a published method. No external method is run, which is not sufficient to judge the contribution.

  2. The paper claims discovery, but supports it only with diversity measures (successful-scaffold count and Top-10 concentration), never comparing the successful molecules or scaffolds against the ChEMBL actives used to train the predictors. Diversity and novelty are not interchangeable: a set of ~1970 scaffolds consisting entirely of established cores for these target classes would score well on both reported measures and contain no discovery at all. We note that the nearest-neighbor Tanimoto analysis in Appendix O is computed against the authors’ own experience archive, which addresses self-repetition rather than novelty relative to known actives.

Questions for authors

See weakness.

2 comments
AuthorsOpenReview ↗

Response to Reviewer vnHt

We thank the reviewer for the suggestions on external algorithm comparisons and comparison with known active compounds. The common-query-budget benchmark and ChEMBL reference analysis show higher unique predicted-success yield and broader coverage of reference-unseen scaffolds with DualEvo.

1. External-method benchmark

All methods use the same fixed dual-target oracle and property gates. We report the fraction of distinct molecules meeting the predicted-success criteria among the first 2,000 distinct canonical-molecule oracle queries, averaged over five seeds per task.

MethodGSK3B–JNK3RORC–DHODH
DualEvo53.62%71.89%
PolyTamGen-DPO19.72%27.05%
CombiMOTS, author-supplied fragment priors13.39%20.34%
MARS10.45%14.98%
REINVENT4 + Diversity Filter7.87%22.32%

Values are mean unique predicted-success yields over five seeds. REINVENT4 uses its native IdenticalMurckoScaffold filter, with results aggregated from the original logs under the same distinct-query definition. Each method retains its own model and optimization configuration. The DualEvo query stream includes candidates generated by the final fixed model.

Under the common evaluation criteria and query budget, DualEvo achieves the highest unique predicted-success yield among the listed methods on both tasks.

Revised manuscript: Appendix L, p. 25, Table S11b.

2. Structural novelty relative to known ChEMBL actives

The reference set is the union of ChEMBL training and validation active molecules for the two targets. Novelty is defined relative to this set, using canonical SMILES for molecule matching and Bemis–Murcko scaffolds for scaffold matching.

Six-task meanReward-onlyDualEvoRelative increase
Predicted-success Murcko scaffolds unseen in the reference active set1,522.61,958.328.6%
Murcko scaffolds successful under both the original evaluator and PSICHIC-XL and unseen in the reference active set1,227.31,594.930.0%

Results cover six tasks and three seeds, with 5,000 independent samples per final fixed model per seed.

Consensus success requires the same valid canonical molecule to pass the original QED thresholds for both targets, the SA upper bound, and the logP interval, and to meet both target-specific thresholds under each of the original ECFP and PSICHIC-XL evaluators. All four activity comparisons must pass, using scores at least equal to the corresponding threshold selected by maximum validation MCC. Molecules are deduplicated by canonical SMILES. Unseen consensus scaffolds are distinct nonempty Murcko scaffolds from this intersection that are absent from the reference active set.

DualEvo increases reference-unseen successful scaffolds by 28.6%, with a 30.0% increase retained in the consensus-success intersection. These results demonstrate broader coverage of predicted-success structures beyond the known target-active reference set.

Revised manuscript: Appendix O, p. 29, Table S14b.

AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewer,

Thank you for your time and consideration. We would appreciate hearing whether our responses and revised manuscript address your concerns. We would be happy to clarify any remaining questions.

Best regards, The Authors

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

DualEvo adapts a pretrained molecular generator to dual-target objectives through repeated generation, fixed-predictor evaluation, cumulative experience storage, and LoRA updates. Its main contribution is scaffold-balanced selection from a high-reward archive. Across six target pairs, this improves structural coverage and the yield of distinct predicted-success molecules, with the clearest gains at larger sampling budgets.

Critical strengths

The Reward-only comparison isolates the effect of scaffold-balanced selection within the proposed framework. At 5,000 samples, mean unique predicted-success yield increases from 48.4% to 55.2%, and successful-scaffold count increases from 1,533 to 1,970. These results support a useful improvement in structural coverage.

The paper distinguishes repeated successful samples from distinct successful molecules. Reporting both reveals a meaningful trade-off: scaffold balancing lowers predicted-success frequency while increasing distinct discoveries. This is more informative than reporting hit rate alone.

Critical weaknesses

The contribution beyond existing replay and diversity-control methods remains insufficiently established. REINVENT4, Augmented Memory, and Saturn already address closely related issues through experience reuse, scaffold filtering, or memory management. The main experiments establish the benefit of balancing over reward-only selection, but do not establish its advantage over these existing mechanisms. The external results in Appendices K/L mix optimization streams and post-training readouts with different call and candidate budgets. They therefore do not resolve this concern. At 500 samples, the mean gain over Reward-only is 2.5 percentage points, with one task declining and only three seeds per condition.

Essential implementation details are missing. Section 3.1 and Appendix A do not define the paired-tree token vocabulary, serialization rules, or reconstruction procedure D(x)D(x). It is therefore unclear what individual tokens encode and how generated sequences become molecules. Other unresolved details include which sequence is retained when several sequences decode to the same canonical molecule, the stated deterministic tie-breaking rule, and the token reduction used in the autoregressive loss. These choices determine the actual training data and objective, so they matter for reproducibility.

Questions for authors

Could you provide one complete molecule-to-paired-trees-to-tokens-to-molecule example, together with the token grammar and decoding algorithm? Is this representation inherited from an existing implementation, and where is it documented?

For the external baselines in Appendices K/L, what are the total optimization budgets before evaluation, and which replay and diversity-control settings were enabled? Is there a matched-budget comparison that separates the benefit of scaffold round-robin selection from existing scaffold filters or memory-management schemes?

How are duplicate molecular encodings and reward ties resolved? Is sequence cross-entropy summed or averaged over valid tokens before averaging across sequences, and is optimizer state carried across stages?

Is there existing evidence from an independent activity evaluator or applicability-domain analysis that the gains persist beyond the predictors used for optimization? Fresh sampling tests the learned distribution, but does not independently validate predicted activity. I acknowledge the stated limitation and am not requesting wet-lab experiments.

3 comments
AuthorsOpenReview ↗

Response to Reviewer MBRZ — Part 1/2

We thank the reviewer for the detailed questions on paired-tree encoding, deduplication and loss reduction, external budgets, and diversity control.

1. Paired-tree representation, vocabulary, and decoding

The representation comprises a pharmacophore tree and a BRICS fragment tree with a shared ordered topology and aligned leaves, separated by <smi:>. The pharmacophore tree encodes structural semantics, and molecular reconstruction uses the BRICS fragment tree.

ItemContent
Input canonical SMILESCC(=O)Nc1ccc(O)cc1
Pharmacophore tree<<HBA><<HBD/HBA><HBD/HBA_Pi>>>
Separator<smi:> (ID 5)
BRICS fragment tree<<[1*]C(C)=O><<[2*]N[3*]><[4*]c1ccc(O)cc1>>>
Decoded canonical SMILESCC(=O)Nc1ccc(O)cc1

The complete sequence contains 52 token IDs:

78,50,78,48,30,80,80,5,78,78,91,69,63,92,83,61,83,62,79,88,80,78,78,91,70,63,92,87,91,71,63,92,80,78,91,72,63,92,93,69,93,93,93,61,88,62,93,93,69,80,80,80
ItemDefinition
Vocabulary size100 tokens (IDs 0–99)
IDs 0–5<pad>, <sos>, <unk>, <eos>, <mask>, <smi:>
IDs 6–58Pharmacophore and state labels, including <HBA>, <HBD/HBA>, and <HBD/HBA_Pi>
IDs 59–99Fragment-text characters, structural symbols, and multi-character tokens such as Br and Cl
Tree serializationRecursively parse nesting with < and >, preserving child order; <smi:> separates the trees
Molecular reconstructionExtract BRICS leaves from the fragment tree, reconnect numbered attachment sites, remove dummy atoms, and reconstruct and canonicalize with RDKit

The complete example and decoding description appear in Appendix A, p. 12, “Encoding and reconstruction.”

2. Deduplication, reward ties, and training objective

The archive deduplicates molecular encodings by canonical SMILES. Among multiple records, it deterministically retains the highest-ranked complete sequence, ordered by reward, success and validity flags, higher QED, lower SA, and then earlier stage and generation order. Training inputs include the <sos> prefix, with no additional <eos> supervision target. Cross-entropy and KL are each averaged over valid target tokens within each sequence and then over the selected sequences, with KL weight beta = 0.05. Each stage inherits the preceding stage's LoRA weights and initializes a fresh AdamW optimizer, without carrying optimizer state across stages.

These details appear in Appendix E, p. 17, “Sequence selection and loss.”

AuthorsOpenReview ↗

Response to Reviewer MBRZ — Part 2/2

3. External budgets, REINVENT settings, and controlled selection ablation

Appendices K/L report native optimization budgets and candidate discovery under common oracle-query budgets.

Experiment or settingBudget and configuration
Original Saturn runsApproximately 500–527 oracle queries
Original REINVENT4 runs600 RL proposals
Added external comparisonSame dual-target evaluator; 500 / 2,000 distinct canonical-molecule oracle queries; five seeds
Added native REINVENT4 runs60 training steps, batch size 50; IdenticalMurckoScaffold, bucket size 25, minimum score 0.4
Controlled Murcko-filter ablationSame generator and online framework as DualEvo; bucket size 10, minimum score 0.5; 2,000 online proposals

The common query budget counts distinct canonical molecules actually evaluated. The DualEvo query stream includes candidates generated by the final fixed model. The controlled ablation holds the generator, reward, cumulative archive, and update procedure fixed to directly compare the two experience-selection rules.

TaskCapacity filter: unique successes / scaffoldsDualEvo: unique successes / scaffolds
GSK3B–JNK32,152 / 1,4222,444 / 1,714
RORC–DHODH2,935 / 1,6293,223 / 2,193

These are rounded means over five seeds, with 5,000 independent samples from each final fixed model. Scaffold round-robin yields more distinct successful molecules and scaffolds than the tested capacity filter within the same online-learning framework.

Revised manuscript: Appendix K, p. 24 (native budgets); Appendix L, p. 25, Table S11b (common query budgets); Appendix E, p. 17, Table S4b (controlled ablation).

4. Independent evaluation and novelty relative to ChEMBL actives

We independently rescored final fixed-generator samples across six tasks using PSICHIC-XL, which was not used in the original online optimization. Successful structures were compared against the union of training and validation active molecules for the two targets in ChEMBL.

Consensus success requires the same valid canonical molecule to pass the original QED thresholds for both targets, the SA upper bound, and the logP interval, and to meet both target-specific thresholds under each of the original ECFP and PSICHIC-XL evaluators. All four activity comparisons must pass, using scores at least equal to the corresponding threshold selected by maximum validation MCC. Molecules are counted once by canonical SMILES. Unseen consensus scaffolds are distinct nonempty Murcko scaffolds from this intersection that are absent from the reference active set.

Six-task meanReward-onlyDualEvo
PSICHIC-XL unique predicted-success yield39.18%46.86%
PSICHIC-XL successful scaffolds1,323.31,764.2
Original-evaluator successful scaffolds unseen in the ChEMBL reference1,522.61,958.3
Consensus-success scaffolds unseen in the reference1,227.31,594.9

Results use three seeds and 5,000 independent samples per fixed model per seed. DualEvo covers more successful scaffolds under independent scoring and more reference-unseen scaffolds within the consensus-success intersection (Appendix O, p. 29, Table S14b).

AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewer,

Thank you for your time and consideration. We would appreciate hearing whether our responses and revised manuscript address your concerns. We would be happy to clarify any remaining questions.

Best regards, The Authors

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

This paper proposes DualEvo, an online evolution framework for dual-target molecular generation. The method combines cumulative experience archiving, reward-guided filtering, scaffold-balanced selection, and LoRA updates to improve the quality and diversity of generated molecules. On five of six dual-target tasks, DualEvo achieves higher unique predicted-success yield than Reward-only, with broader chemical scaffold coverage.

Critical strengths

The study addresses an important problem in dual-target molecular design. It integrates cumulative experience archiving, reward-guided selection, scaffold-balanced selection, and LoRA updates into an online evolution framework, with support for saving and reusing adapted generator states. The method improves chemical-space coverage. In the 5,000-sample evaluation, DualEvo increases unique predicted-success yield from 48.4% to 55.2% and successful scaffolds count from 1,533.3 to 1,969.7, demonstrating improved structural diversity.

Critical weaknesses

The main evaluation relies on insufficiently validated activity predictors. The ECFP-based classifiers lack reported independent test performance, raising concerns about potential reward hacking. Without systematic independent evaluation, the reliability of the predicted activity and chemical-space coverage gains remains uncertain. The methodological novelty and baseline comparisons are limited. The framework mainly combines established techniques, and scaffold round-robin is not compared with existing Diversity Filters. The main experiments only compare variants of the same base generator, while direct comparisons with representative dual-target methods are missing.

Questions for authors

Can the authors report independent test performance (e.g., PR-AUC and MCC) for the activity predictors and assess whether the predicted-success and diversity gains persist under independent activity predictors or systematic molecular docking evaluations? Do the results hold against representative dual-target generation methods under comparable oracle budgets and consistent evaluation criteria? Does scaffold round-robin offer clear advantages over existing diversity-control strategies, such as REINVENT's Diversity Filter?

3 comments
AuthorsOpenReview ↗

Response to Reviewer miKS — Part 1/2

We thank the reviewer for the constructive suggestions on activity evaluation, external baselines, and diversity control. The revised manuscript adds independent rescoring, comparisons under a common oracle-query budget, and a controlled comparison with scaffold filtering.

1. Activity-predictor performance and independent rescoring

All 12 classifiers use Bemis–Murcko scaffold splits, with thresholds selected by maximum validation MCC. We report class proportions alongside PR-AUC, MCC, sensitivity, and specificity to characterize both ranking performance and classification at the selected thresholds (Appendix B, p. 13, Table S1c).

TargetClassifierThresholdPos.Neg.Pos. (%)PR-AUCMCCSens.Spec.
ABL1RBF-SVM0.70001611790.40.9930.5070.8140.941
SRCRBF-SVM0.52002335481.20.9760.6400.9530.648
FGFR1RF0.56002981694.90.9980.4901.0000.250
KDRRBF-SVM0.84007395593.10.9930.5320.9540.636
EGFRRF0.65007878790.00.9890.4680.9420.540
ERBB2RBF-SVM0.60501732587.40.9610.4620.9480.480
ESR1RBF-SVM0.68502423587.40.9910.7710.9710.800
ESR2RF0.49501141687.70.9760.4240.9040.563
GSK3BLR0.45602103685.40.9730.3560.9860.222
JNK3LR0.4467932678.20.9920.8030.9570.846
RORCLR0.57477592297.20.9990.6691.0000.455
DHODHRF0.67831321888.00.9890.6610.9550.722

RF denotes random forest and LR logistic regression. The positive fraction is P/(P+N), specificity is TN/(TN+FP), and sensitivity is TP/(TP+FN). RORC's PR-AUC of 0.999 should be interpreted alongside its 97.2% positive prevalence. FGFR1 and GSK3B have specificities of 0.250 and 0.222, indicating limited rejection of inactive molecules at the selected thresholds.

We therefore independently rescored final fixed-generator samples using PSICHIC-XL (Koh et al., 2024), which models protein–ligand interactions from protein sequences and ligand information and was not used in the original online optimization.

Six-task meanReward-onlyDualEvoIncrease
PSICHIC-XL unique predicted-success yield39.18%46.86%7.68 pp
PSICHIC-XL successful Murcko scaffolds1,323.31,764.2440.9
Consensus-success scaffolds unseen in the ChEMBL reference1,227.31,594.9367.6

Each fixed model was sampled 5,000 times per seed across six tasks and three seeds. The reference comprises the union of training and validation actives for the two targets.

Consensus success requires a valid molecule to pass the original QED, SA, and logP gates and all four activity thresholds: both targets under the original ECFP classifiers and both targets under PSICHIC-XL. Thresholds are selected by validation MCC, and scores must be at least the corresponding threshold. Molecules are deduplicated by canonical SMILES. Unseen consensus scaffolds are distinct nonempty Murcko scaffolds absent from the reference active set.

DualEvo's gains thus persist under independent rescoring and within the intersection of the two evaluators, supporting broader coverage of predicted-success structures (Appendix O, p. 29, Table S14b).

Reference: Koh et al. (2024), “Physicochemical graph neural network for learning protein–ligand interaction fingerprints from sequence data,” Nature Machine Intelligence. https://doi.org/10.1038/s42256-024-00847-1

AuthorsOpenReview ↗

Response to Reviewer miKS — Part 2/2

2. Online adaptation with PSICHIC and KarmaDock feedback

We also used sequence-based PSICHIC and structure-based KarmaDock (Zhang et al., 2023) as alternative online feedback signals within the same cumulative-archive, scaffold-selection, and LoRA-update framework.

FeedbackTaskUnique predicted-success yield, Frozen → adaptedSuccessful scaffolds, Frozen → adapted
PSICHICFGFR1–KDR34.87% → 68.33%172.0 → 314.3
PSICHICEGFR–ERBB249.33% → 74.27%240.0 → 350.0
PSICHICESR1–ESR251.67% → 78.73%251.3 → 355.7
KarmaDockGSK3B–JNK323.93% → 70.60%114.7 → 314.7
KarmaDockFGFR1–KDR16.13% → 40.13%77.3 → 174.3

These representative tasks have complete paired results over three seeds. Each run uses 2,000 online proposals followed by 500 final fixed-generator samples, paired with the Frozen generator at the same seed. Success uses the feedback model's validation-selected thresholds and the original property gates.

Both feedback types improve unique predicted-success yield and scaffold coverage over the paired Frozen generator, supporting adaptation with sequence- and structure-based evaluators (Appendix E, p. 17, Table S4c).

3. Representative external methods under a common query budget

All candidates are evaluated using the same fixed dual-target oracle and property gates. The table reports unique predicted-success yield at 2,000 distinct canonical-molecule oracle queries, averaged over five seeds per task.

MethodGSK3B–JNK3RORC–DHODH
DualEvo53.62%71.89%
PolyTamGen-DPO19.72%27.05%
CombiMOTS, author-supplied fragment priors13.39%20.34%
MARS10.45%14.98%
REINVENT4 + native Diversity Filter7.87%22.32%

Each method retains its own model and optimization configuration. The DualEvo query stream includes candidates generated by the final fixed model. REINVENT4 results are computed from native run logs using the same distinct-query definition.

DualEvo achieves the highest unique predicted-success yield among these external methods on both tasks under the common evaluation and query budget (Appendix L, p. 25, Table S11b).

4. Controlled comparison with Diversity Filter

To isolate the selection rule, we replaced scaffold round-robin with a REINVENT-style Murcko capacity filter while holding the generator, original reward, cumulative archive, LoRA updates, and 2,000-proposal online budget fixed. The filter uses bucket size 10 and minimum score 0.5.

TaskFilter: unique successesDualEvo: unique successesFilter: successful scaffoldsDualEvo: successful scaffolds
GSK3B–JNK32,1522,4441,4221,714
RORC–DHODH2,9353,2231,6292,193

Results are rounded means over five seeds, with 5,000 independent samples from each final fixed generator. Within the same online-learning framework, scaffold round-robin selects training experience that yields more distinct successful molecules and scaffolds than the tested capacity filter (Appendix E, p. 17, Table S4b).

Reference: Zhang et al. (2023), “Efficient and accurate large library ligand docking with KarmaDock,” Nature Computational Science. https://doi.org/10.1038/s43588-023-00511-5

AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewer,

Thank you for your time and consideration. We would appreciate hearing whether our responses and revised manuscript address your concerns. We would be happy to clarify any remaining questions.

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.