acceptodds
Under review as a conference paper at ICLR 2027

MolDeriveX: Multi-Objective Optimization via Adaptive Structural Inheritance

Abstract

Multi-objective molecular optimization requires translating competing property objectives into effective structural edits. A key challenge is to identify and reuse substructures that support different properties within the current molecule, while updating this structural guidance as optimization proceeds. We introduce MolDeriveX, a graph optimization framework based on adaptive structural inheritance. It identifies property-specific anchors by combining isolated-fragment scores with changes in parent-level property scores after fragment deletion. These anchors guide Pareto-based graph optimization through soft retention penalties, favoring the retention of property-supporting regions while keeping them editable. Selected high-quality descendants become new structural references, with anchors recomputed to guide subsequent search. Across five multi-objective evaluated tasks, MolDeriveX achieves the highest hypervolume on four and the highest Top-100 joint score on all four target activity tasks. In a challenging, real-world-motivated antibiotic derivatization case study involving three single-parent optimization tasks, MolDeriveX achieves the highest hypervolume for predicted antibacterial activity and parent-molecule similarity among the evaluated methods.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular optimization, AI for Sciences, Drug discovery, Multi-objective optimization

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

This paper presents MolDeriveX, a multi-objective molecular optimization framework based on adaptive structural inheritance. It identifies property-specific structural anchors by combining isolated-fragment scores with changes in parent-level property scores after fragment deletion. These anchors guide Pareto-based differentiable graph editing through soft retention penalties, while lineage-aware exploration and descendant promotion allow structural references and anchors to evolve during search.

Critical strengths

(1) The paper addresses a relevant challenge of balancing structural preservation and exploration in multi-objective molecular optimization. (2) The framework integrates context-informed anchor selection, soft retention, and lineage updates while keeping protected regions editable. (3) Evaluation spans six tasks and eight baselines, with activity-task improvements supported by hypervolume, elite-candidate quality, and joint-success metrics.

Critical weaknesses

(1) The paper does not sufficiently substantiate its central contribution of adaptive structural inheritance. Why converting heterogeneous property-specific evidence into a shared modification cost is sufficient to provide the context-adaptive guidance motivating the method?

(2) The paper fails to sufficiently justify the optimization properties of its central update rule. EPO weights are computed from the property losses and their gradients, whereas the actual update additionally includes the structural penalty gradient. Consequently, the original EPO coordination properties do not directly carry over. The retention gradient can cancel the property-optimization gradient, producing a stationary point that is not Pareto-stationary for the original property losses. The paper neither analyzes when this cancellation occurs nor provides conditions on the penalty strength, gradient relationships, or step size that characterize how retention changes the optimization behavior.

(3) The notation, definitions of key variables, and update rules are incompletely specified, leaving the mathematical description insufficiently connected. Equation (3) defines A as edge logits and P as the corresponding edge probabilities, whereas Appendix A directly uses A, described as an adjacency matrix, in graph convolution without clarifying the mapping between the two. Equation (8) introduces node weights without adequately explaining how they are constructed or how they change during continuous editing. Equation (11) does not define how population-level feedback updates the base preference or fully report the parameter settings used. Equation (12) does not specify its computation and update rules.

Questions for authors
  1. What capability distinguishes adaptive structural inheritance from existing graph editing with structural guidance?
  2. In Equation (5), retaining only the largest connected component can discard additional regions outside the selected fragment. How is their influence distinguished from the selected fragment’s contribution?
  3. How is ρ selected in Equation (7)?
  4. Does graph convolution use P = sigmoid(A) or the raw logits A? How are the node weights in Equation (8) constructed and updated during continuous editing?
  5. Why do MolSearch’s reported final HV values on all four activity tasks match the initial-population values in Table 11 (0.6948, 0.5688, 0.5614, and 0.3597)?
  6. Why does QED–logP HV differ between Tables 10–11 and Table 20 (0.9470 versus 0.9456), and diversity between Tables 18 and 20 (0.8459 versus 0.8550)?
12 comments

Thank you for your careful review and constructive feedback. Your comments helped us clarify how MolDeriveX works and strengthen the experimental evaluation. We address each comment below.

Q1 & Weakness 1: What capability distinguishes adaptive structural inheritance from existing graph editing with structural guidance?

Response: MolDeriveX carries property-specific anchors across generations and can update which regions receive protection.Before each edit, MolDeriveX matches inherited anchors to the current parent and applies a soft penalty to changes in the matched regions. When a qualifying descendant is promoted to the lineage root, anchors extracted from that molecule can guide subsequent edits. This allows the search to protect a newly introduced motif associated with higher predicted activity while optimizing other properties.

Single-parent derivatization.F-CReM uses local chemical context to select compatible fragment replacements around a specified parent. We used this configuration for antibiotic derivatization. Applying it to the general benchmark would require an additional policy for selecting and advancing a population across multiple objectives. For each antibiotic parent, we evaluated the trade-off between predicted activity against A. baumannii and similarity to the starting molecule. We also measured diversity within the modified regions after excluding the parent core.

Table 12. Hypervolume (HV) and diversity within modified regions in single-parent derivatization.

Starting moleculeF-CReM HVMolDeriveX HVF-CReM diversityMolDeriveX diversity
Linezolid0.340.690.420.64
Clindamycin0.540.690.570.65
Erythromycin0.670.790.50.62

Population-based multi-objective optimization. InversionGNN uses property gradients to guide edits to the current molecule. Fixed L1 uses the same soft structural penalty and decoding protection as the full MolDeriveX configuration and retains the initial anchor set throughout the run. On the four-objective task, mean final HV was 0.7064 for MolDeriveX and 0.6514 for Fixed L1, a gain of 0.0550.

Table 13. Final HV across three search seeds, reported as mean ± sample standard deviation.

ObjectivesInversionGNNFixed L1MolDeriveX
GSK3β–JNK30.7538 ± 0.01560.8145 ± 0.00460.8696 ± 0.0246
QED–SA–GSK3β0.6291 ± 0.02960.8697 ± 0.00510.8813 ± 0.0038
QED–SA–GSK3β–JNK30.3258 ± 0.04930.6514 ± 0.05060.7064 ± 0.0127

Property-specific evidence determines which regions are protected. We select anchors separately for each objective, match them to the current parent, and combine the matches into a single mask over editable graph coordinates. The shared modification cost applies to the masked coordinates, while EPO continues to balance the individual property losses. Rematching anchors to each parent and reselecting them after lineage-root promotion allow the protected region to change as the search progresses.

Under the common budget, refreshing property-selected anchors increased the mean match rate for current-root anchors on both tasks relative to the fixed policy. Its effect on final hypervolume (HV) varied by task. The refreshed policy achieved the highest mean final HV on QED–SA–GSK3β, and the fixed policy achieved the highest mean final HV on QED–SA–GSK3β–JNK3.

Table 10. Final HV and full anchor-match rates under a budget of 5,000 online full-evaluation equivalents (FEO) per run. Values are mean ± sample standard deviation across three paired search seeds.

TaskAnchor policyFinal HVOriginal anchor match (%)Current-root anchor match (%)
QED–SA–GSK3βRandom, resampled0.8742 ± 0.010366.68 ± 9.2769.34 ± 3.01
QED–SA–GSK3βFixed, property-selected0.8783 ± 0.007264.46 ± 4.0864.46 ± 4.08
QED–SA–GSK3βRefreshed, property-selected0.8802 ± 0.004365.59 ± 5.3272.61 ± 2.62
QED–SA–GSK3β–JNK3Random, resampled0.6569 ± 0.027769.45 ± 7.7268.60 ± 5.83
QED–SA–GSK3β–JNK3Fixed, property-selected0.6745 ± 0.025273.71 ± 1.3473.71 ± 1.34
QED–SA–GSK3β–JNK3Refreshed, property-selected0.5997 ± 0.076168.70 ± 8.1575.94 ± 4.20

Q2: In Equation (5), retaining only the largest connected component can discard additional regions outside the selected fragment. How is their influence distinguished from the selected fragment’s contribution?

Response:

When fragment deletion leaves multiple connected components, Eq. (5) captures both the effect of deleting the selected fragment and the effect of retaining only the largest remaining component. We examined 400 selected initial anchors from 100 starting molecules in one four-objective run. Retaining only the largest component discarded additional atoms in 313 of 400 cases (78.25%).

Table 14. Frequency and extent of additional atom loss after fragment deletion.

Anchor objectiveDeletions leaving multiple componentsMean share of remaining atoms discarded
QED61/10019.80%
SA62/10014.00%
GSK3β94/10039.10%
JNK396/10038.80%

Means include all 100 anchors per objective. Additional atom loss was counted as zero when the remainder stayed connected.

To quantify the effect on QED, we used RDKit to compare the score of the full disconnected remainder with that of its largest component. Of the 61 QED-anchor deletions that left multiple components, 30 had a full remainder that RDKit could sanitize and an available largest-component score. Across these 30 cases, the median absolute difference between the two QED scores was 0.143, and retaining only the largest component reversed the sign of the deletion effect in 4 of 30 cases. For example, one parent had a QED of 0.732. After deletion, QED rose to 0.790 when all remaining components were retained and fell to 0.640 when only the largest component was retained.

We have clarify that Eq. (5) uses this combined effect as a heuristic for ranking anchors and report the QED sensitivity check in the appendix.

Q3 & Weakness 2: How is ρ selected in Equation (7)?

Response:

Structural penalty settings. We manually set the base value of ρ for each task. A larger ρ increases the cost of edits to matched anchor regions. When enabled, the search-wide stagnation rule lowers ρ if overall search progress stalls. Table 15 lists the settings used in the main experiments.

Table 15. Base structural penalty coefficients and their values under search-wide stagnation.

TaskBase ρρ when overall search progress stalls
QED–SA20.5
GSK3β–JNK31.50.5
QED–SA–GSK3β40.5
QED–SA–JNK333
QED–SA–GSK3β–JNK30.360.12

When an individual lineage stalls and lineage-level relaxation is enabled, the default rule caps that lineage’s current ρ at 1.0. The four-objective configuration uses 0.12 for the stalled lineage.

EPO computes the property weights from the property losses and their gradients. The editor then adds the weighted structural penalty gradient. We agree that the original EPO guarantee applies to the EPO update alone.

On the editable coordinates, the combined gradient can vanish when

gEPO=−ρgstr,g_{\mathrm{EPO}} = -\rho g_{\mathrm{str}},

even if gEPO≠0g_{\mathrm{EPO}} \neq 0. A sufficient condition to rule out exact cancellation at a given step is

ρ∥gstr∥<∥gEPO∥.\rho\lVert g_{\mathrm{str}}\rVert < \lVert g_{\mathrm{EPO}}\rVert.

This condition concerns the combined gradient at a single editing step. Table 16 reports the paired changes in final hypervolume (HV) across three search seeds after enabling the structural penalty.

Paired penalty ablation. In a follow-up experiment, we held the anchors and decoding rules fixed and compared ρ = 0 with a shared schedule that started at 0.36 and dropped to 0.12 when overall search progress stalled. Under this schedule, the penalty had little effect on final hypervolume (HV) for QED–SA, increased it in all three QED–SA–GSK3β runs, and decreased it in all three four-objective runs.

Table 16. Paired change in final HV, calculated as the HV under the shared penalty schedule minus the HV at ρ = 0 for each seed.

TaskMean change ± sample SDDirection across three seeds
QED–SA+0.000002 ± 0.0000061 increase, 1 decrease, 1 tie
QED–SA–GSK3β+0.0081 ± 0.00783 increases
QED–SA–GSK3β–JNK3−0.0214 ± 0.00553 decreases

We have reported the task-specific settings, stagnation rules, and paired results in the appendix.

Q4 & Weakness 3: Does graph convolution use P = sigmoid(A) or the raw logits A? How are the node weights in Equation (8) constructed and updated during continuous editing?

Response:

Adjacency used during continuous editing. Every graph convolution uses P=sigmoid⁡(A)P = \operatorname{sigmoid}(A) as its adjacency input, where AA contains the raw edge logits and the sigmoid is applied elementwise. The convolution computes PHW+bPHW + b before applying the hidden-layer activation.

Pooling weights. At each editing step, we recompute the pooling weights in Eq. (8) from the current PP. We first compute the clipped row sum

q(v)=min⁡(1,∑uPv,u).q(v) = \min(1, \sum_u P_{v,u}).

We then assign node weights according to Table 17. For each original leaf ℓ\ell, p(ℓ)p(\ell) denotes its original neighbor and e(ℓ)e(\ell) denotes its candidate extension node.

Table 17. Node weights used for pooling during continuous editing.

Editing pathInternal nodeOriginal leaf ℓ\ellExtension node e(ℓ)e(\ell)
Fewer than three rings11q(e(ℓ))q(e(\ell))
Three or more rings1Pℓ,p(ℓ)P_{\ell,p(\ell)}q(e(ℓ))Pℓ,p(ℓ)q(e(\ell)) P_{\ell,p(\ell)}

Edge updates and decoding. Only leaf–extension edge logits retain their updates between editing steps. The logits for edges between original leaves and their original neighbors remain at their initial value of zero. In the three-or-more-ring path, each original leaf therefore has a constant pooling weight of 0.5, while its extension node’s weight can change with q(e(ℓ))q(e(\ell)). The decoder samples leaf deletions after continuous editing is complete.

When population-level feedback is enabled, Eq. (11) computes a deficit for each objective from its score quantile in the evaluated pool. It also measures how often each objective has the lowest score. These signals adjust the fixed reference preference. The algorithm clips and normalizes the resulting weights, then blends them with the previous generation’s preference. The parent-level rule further adjusts this preference using the current parent’s own deficits.

We have corrected the continuous-editing convolution in the appendix from AHW+bAHW + b to PHW+bPHW + b and explicitly describe the pooling weights and edge mask.

Q5: Why do MolSearch’s reported final HV values on all four activity tasks match the initial-population values in Table 11 (0.6948, 0.5688, 0.5614, and 0.3597)?

Response:

We verified that the archived MolSearch outputs for each of the four activity tasks contained exactly the 100 starting molecules. The hypervolume (HV) of each returned set therefore equaled that of the corresponding starting population.

We reran MolSearch with three search seeds per task using the same starting populations. Each run evaluated 5,000 new molecules, and the 100 starting molecules were scored separately. Stage 1 scores rule-generated proposals before filtering them, so discarded proposals also count toward the evaluation budget. Every rerun exhausted the budget in Stage 1, before Stage 2 began. On the four activity tasks, no newly evaluated molecule had entered the retained set at the cutoff, so the rerun exports again contained only the starting molecules. The archived exports establish what the original runs returned. The evaluation counts and stopping stage come from the reruns.

Table 11. MolSearch HV across three search seeds. Initial HV is computed over the 100 starting molecules. Each seed column reports HV over the 5,000 newly evaluated molecules, excluding the starting molecules.

TaskInitial HVSeed 1 HVSeed 2 HVSeed 3 HV
GSK3β–JNK30.69480.71490.71490.7149
QED–SA–GSK3β0.568810.447880.447880.44788
QED–SA–JNK30.561380.457030.457030.45703
QED–SA–GSK3β–JNK30.35970.137840.137840.13784
QED–SA0.592510.944870.944870.94487

All three seeds evaluated identical sets of molecules on both three-objective tasks and on the four-objective task. On GSK3β–JNK3 and QED–SA, the sets varied across seeds but yielded identical nondominated objective vectors and therefore identical HVs.

We have presented the MolSearch results, evaluation protocol, and diagnostics in the appendix. The main text will focus on the other comparisons.

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

This paper proposes MolDeriveX, a multi-objective molecular optimization framework based on adaptive structural inheritance. It identifies property-specific anchors using isolated fragment scores and parent-level deletion effects, incorporates soft retention penalties into Pareto-guided graph editing, and dynamically updates structural references. The method is evaluated on six molecular optimization benchmarks and three single-parent antibiotic derivatization tasks, reporting strong optimization performance.

Critical strengths
  1. Strong empirical performance across multiple metrics. MolDeriveX achieves the best HV on four of the five main tasks and the best Top-100 joint score on all four activity-related tasks. On the four-objective benchmark, HV improves from 0.6281 to 0.7103, while joint hits increase from 22 to 381 compared with the strongest baseline (Tables 1–2 and 6). These improvements are substantial.

  2. Meaningful structure-aware optimization framework. Combining parent-context fragment attribution, soft structural retention, and lineage updates offers a useful form of structural guidance beyond objective-level optimization. The existing ablations provide preliminary evidence that context-aware anchoring, structural retention, and lineage management contribute to performance (Table 3).

Critical weaknesses
  1. The fairness and reliability of the experimental comparisons are not sufficiently established. Equation (5) requires property-oracle evaluations of isolated fragments and deletion complements, and anchors are recomputed after descendant promotion. However, Algorithm 1 explicitly deducts the evaluation budget only for candidate molecules, leaving unclear whether these additional oracle calls are included in the 5,000-query budget. If excluded, the methods may not be compared under equivalent information budgets. Furthermore, MolSearch reports exactly the same final HV as the initial population on four activity-related tasks (Tables 1 and 11), raising questions about its effective search execution. The evaluation also lacks independent repeated runs.

  2. The specific contribution of adaptive structural inheritance remains insufficiently isolated. Prior work has explored property-related molecular rationales, Pareto-guided graph editing, and iterative reference updates. The key novelty should therefore lie in context-aware anchor selection and adaptive inheritance. However, the current ablations do not compare against size-matched random anchors or directly isolate static from dynamically refreshed anchors. Consequently, the gains cannot be clearly attributed to structural inheritance rather than generic regularization or population selection. Moreover, Equation (6) merges property-specific anchors into a shared retention penalty, limiting property specificity during editing. In Appendix I.1, only 1.18% of four-objective proposals directly Pareto-dominate their parents, further motivating a more careful separation of local editing effects from population selection.

Questions for authors
  1. Oracle budget: Are oracle calls for fragment scoring, deletion-complement evaluation, and anchor refreshing fully counted within the 5,000-query budget? Please provide a detailed breakdown of actual calls and cache usage.

  2. Mechanism validation: Can the authors compare random, static, and dynamically refreshed anchors under identical editing, population selection, and query budgets, and report both anchor retention rates and optimization performance?

  3. Experimental reliability: Why does MolSearch achieve exactly the initial-population HV on four activity-related tasks? Can the authors clarify its termination conditions, actual query usage, and provide results across multiple independent random seeds?

  4. Training protocol: What training molecules and labels were used for the differentiable predictor hϕh_\phi, and how are these training data related to the evaluation oracles?

6 comments

Thank you for your careful review and constructive feedback. Your comments helped us clarify how MolDeriveX works and strengthen the experimental evaluation. We address each comment below.

Q1 & Weakness 1: Oracle budget: Are oracle calls for fragment scoring, deletion-complement evaluation, and anchor refreshing fully counted within the 5,000-query budget? Please provide a detailed breakdown of actual calls and cache usage.

Response:

Original evaluation budget. The 5,000-evaluation limit counted each newly generated molecule scored as a search candidate, including candidates that were later discarded. We scored the 100 starting molecules separately. In MolDeriveX, fragment and deletion-complement scoring for initial anchor selection and subsequent anchor refreshing fell outside the 5,000-candidate count.

Property scores also guide intermediate decisions. MARS scores proposed edits before deciding whether to accept them, and RetMol uses scores to select exemplars and update its retrieval library. MolDeriveX uses fragment and deletion-complement scores to select and refresh anchors.

Follow-up experiment with a unified oracle budget. The original records provide only a partial breakdown of oracle calls and cache usage. In a separate four-objective follow-up run with anchor refreshing enabled, we counted all online oracle calls, including initialization and anchor-related scoring, within a single budget. We define one full-evaluation equivalent (FEO) as four single-property oracle calls on this task.

Table 8. Actual oracle calls in the four-objective follow-up run.

Scoring stepActual single-property callsFEO used
Initial population400100
Initial anchor selection — fragments1,040260
Initial anchor selection — deletion complements3,648912
Initial anchor selection — parent and auxiliary scores20050
Anchor refreshing — fragments369
Anchor refreshing — deletion complements1,028257
Anchor refreshing — parent and auxiliary scores7218
Candidate scoring during search13,5763,394
Total20,0005,000

Initial anchor selection used 1,222 FEO (24.4%), and anchor refreshing used 284 FEO (5.7%) of the run’s 5,000-FEO budget. In the same run, 57.2% of recorded scoring requests were served from the cache and incurred no additional oracle calls. The matched-budget anchor-policy comparison in Table 10 uses this same accounting.

We have clarify the original candidate-counting rule in the main text and provide the available per-run breakdowns of oracle calls and cache usage in the supplement.

Q2 & Weakness 2: Mechanism validation: Can the authors compare random, static, and dynamically refreshed anchors under identical editing, population selection, and query budgets, and report both anchor retention rates and optimization performance?

Response:

We compared MolDeriveX with Fixed L1 across three search seeds on each of three tasks. Both configurations use a soft structural penalty and decoding protection. Fixed L1 retains the initial anchors throughout the search, whereas MolDeriveX can refresh them when a qualifying descendant becomes the new lineage root. MolDeriveX achieved higher final hypervolume (HV) in all nine paired comparisons. On the four-objective task, its mean final HV was 0.7064, compared with 0.6514 for Fixed L1, and 69.29% of its initial anchor queries matched the generated candidates on average.

Table 9. Final HV and initial-anchor retention across three search seeds.

TaskFixed L1 final HVMolDeriveX final HVMolDeriveX initial-anchor retention
GSK3β–JNK30.8145 ± 0.00460.8696 ± 0.024682.27% ± 2.24%
QED–SA–GSK3β0.8697 ± 0.00510.8813 ± 0.003864.33% ± 7.30%
QED–SA–GSK3β–JNK30.6514 ± 0.05060.7064 ± 0.012769.29% ± 7.01%

We also compared three anchor policies using the same initial population, editor, and population selection rules, with a common budget of 5,000 full-evaluation equivalents (FEO) per run, including anchor scoring. The fixed policy retained the initial property-selected anchors throughout the search. The refreshed policy reselected property-associated anchors after lineage-root promotion. The random policy sampled size-matched anchors and resampled them after promotion. For each candidate, we checked for full matches to the original property-selected anchors of its initial ancestor and to the anchors assigned to its current root.

Table 10. Final HV and anchor retention under a common 5,000-FEO budget.

TaskAnchor policyFinal HVOriginal property anchors matchedCurrent anchors matched
QED–SA–GSK3βSize-matched random, resampled0.8742 ± 0.010366.68% ± 9.27%69.34% ± 3.01%
QED–SA–GSK3βFixed, property-selected0.8783 ± 0.007264.46% ± 4.08%64.46% ± 4.08%
QED–SA–GSK3βRefreshed, property-selected0.8802 ± 0.004365.59% ± 5.32%72.61% ± 2.62%
QED–SA–GSK3β–JNK3Size-matched random, resampled0.6569 ± 0.027769.45% ± 7.72%68.60% ± 5.83%
QED–SA–GSK3β–JNK3Fixed, property-selected0.6745 ± 0.025273.71% ± 1.34%73.71% ± 1.34%
QED–SA–GSK3β–JNK3Refreshed, property-selected0.5997 ± 0.076168.70% ± 8.15%75.94% ± 4.20%

Under the common budget, refreshed property-selected anchors had the highest mean match rate for current-root anchors on both tasks. On QED–SA–GSK3β, the three policies had similar mean final HVs, ranging from 0.8742 to 0.8802. On the four-objective task, the fixed policy achieved the highest mean final HV (0.6745), compared with 0.5997 for the refreshed policy. We will report the per-seed HVs, anchor match counts, and budget details in the supplement.

Q3: Experimental reliability: Why does MolSearch achieve exactly the initial-population HV on four activity-related tasks? Can the authors clarify its termination conditions, actual query usage, and provide results across multiple independent random seeds?

Response:

We verified that the archived MolSearch outputs for each of the four activity tasks contained exactly the 100 starting molecules. This explains why their reported hypervolumes (HVs) equaled the initial HVs.

We reran MolSearch with three search seeds per task, using the same starting populations as in the original benchmark. Each run evaluated 5,000 new molecules. The 100 starting molecules were scored separately. Stage 1 scores rule-generated proposals before filtering them, so discarded proposals also count toward the evaluation budget. Every rerun exhausted the budget during Stage 1, before Stage 2 began. On the four activity tasks, no newly evaluated molecule had entered the retained set at the cutoff, so the rerun exports again contained only the starting molecules. The archived exports establish what the original runs returned. The evaluation counts and stopping stage were recorded in the reruns.

Table 11. MolSearch HV over the 5,000 newly evaluated molecules per seed. The initial HV is shown for reference. Each seed’s HV excludes the 100 starting molecules.

TaskInitial HVSeed 1 HVSeed 2 HVSeed 3 HV
GSK3β–JNK30.69480.71490.71490.7149
QED–SA–GSK3β0.568810.447880.447880.44788
QED–SA–JNK30.561380.457030.457030.45703
QED–SA–GSK3β–JNK30.35970.137840.137840.13784
QED–SA0.592510.944870.944870.94487

On both three-objective tasks and the four-objective task, all three seeds evaluated identical sets of molecules. On GSK3β–JNK3 and QED–SA, the sets varied across seeds but yielded identical nondominated objective vectors and therefore identical HVs.

We will present the MolSearch results, evaluation protocol, and diagnostics in the appendix. The main text will focus on the other comparisons.

Q4: Training protocol: What training molecules and labels were used for the differentiable predictor (h_\phi), and how are these training data related to the evaluation oracles?

Response:

We trained nine differentiable GNN predictors offline on labeled PubChem molecules. Each GNN learned to predict the properties listed below. The source file contains 107,044,436 SMILES. The training set and test set are divided in a 9:1 ratio.

The QED, SA, GSK3β, and JNK3 labels came from the same underlying scorers used in the benchmark. We evaluated QED, SA, and logP with RDKit, kinase activity with the pretrained PyTDC oracles, and A. baumannii activity with the pretrained Chemprop model of Liu et al [1]. During search, the GNN weights stayed fixed and its gradients guided molecular edits. The evaluation oracles scored the decoded molecules directly, and we used those scores to compute the reported hypervolume. We will document the label sources and training protocol in the supplementary methods.

[1] Liu, G., Catacutan, D. B., Rathod, K., et al. (2023). Deep learning-guided discovery of an antibiotic targeting Acinetobacter baumannii. Nature Chemical Biology, 19(11), 1342–1350. DOI 10.1038/s41589-023-01349-8

Reviewer_zjcqOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This article proposes a multi-objective molecular optimization framework MolDeriveX based on adaptive structural inheritance. It uses fragment scores and deletion effects to identify anchors of specific attributes, incorporates soft structure preservation into DST based EPO graph editing, and updates structural references through lineage based search. The experimental reports on molecular optimization benchmarks and antibiotic derivatization tasks indicate an improvement in both overcapacity and combined performance scores.

Critical strengths

This article addresses the practical issues related to the preservation and modification of equilibrium structures in multi-objective molecular optimization. The results of the report are encouraging, especially in terms of the four objective benchmarks and the task of single parent antibiotic derivatization.

Critical weaknesses
  1. This article assumes that explicit structural inheritance is necessary, but does not specify the specific failures of existing methods to indicate why this mechanism is needed. The proposed method largely combines existing DST based editing, EPO, and heuristic search strategies, therefore its innovation is limited. More importantly, no controlled baseline of DST+EPO or comparison between fixed anchor and adaptive anchor was provided to determine the independent contribution of structural genetics.

  2. Anchor point selection depends on heuristic fragment scores and deletion effects, without sufficient reason to prove its reliability or the size of the selected anchor point. The handling of partially modified anchors is also unclear. In addition, the Oracle evaluation used for anchor extraction may not be included in the budget of the report, resulting in a lack of independent operational statistical data. In the four objective setting, the weak discriminative ability of Pareto dominance further raises the question of whether performance improvement mainly comes from additional population selection heuristics.

  3. The method heavily relies on the number of evaluations and the gradient of the evaluator, while most methods (such as Docking) do not have gradients in the evaluator.

Questions for authors

Q1. What are the specific limitations of existing multi-objective molecular optimization methods and the need for clear structural inheritance? Why is standard DST+EPO optimization not enough?

Q2. Core contribution. Why is there no comparison between DST+EPO, fixed anchor optimization, and adaptive anchor optimization? Without these comparisons, how can we determine the benefits of adaptive structural inheritance?

Q3. Anchor selection. What is the reason for using fragment deletion effect to identify important structures? How to determine the anchor size range (3-18 atoms) and the number of anchors for each target?

Q4. Partial inheritance. What happens when an editor only modifies a portion of the anchor point? If the original substructure can no longer be accurately matched, how will the remaining structural information be inherited?

Q5. Pareto optimization. Given the weak advantage relationship in the four objective optimization, how much of the selection pressure actually comes from Pareto ranking rather than quality quotas, diversity, and other heuristic methods?

Q6. Experimental reliability. Is the Oracle evaluation used for segment scoring and anchor point recalculation calculated within a budget of 5000 evaluations? Why haven't the results of multiple independent runs been reported? Especially in this type of search task

6 comments

Q1 & Weakness 1: What are the specific limitations of existing multi-objective molecular optimization methods and the need for clear structural inheritance? Why is standard DST+EPO optimization not enough?

Response :Thank you for your careful review. We summarize three limitations of existing multi-objective optimization methods below.

  1. Latent-space optimization only implicitly links search directions to molecular structures and their properties. A change in the latent vector provides no direct indication of which substructures will change after decoding or how those changes will affect each property. This makes it difficult to choose a search direction that improves one property while preserving a structural region associated with another. MLPS constrains the latent-space search through trust regions and leaves individual substructures unconstrained. In our experiments, MLPS increased HV on QED–SA from 0.5925 to 0.9375. On the four-objective task, its HV decreased from 0.3597 to 0.1479. Hgraph assembles molecules from motifs and achieved higher final HV than MLPS on the four-objective task (0.6281 vs. 0.1479), suggesting that explicit structural information can aid optimization. Hgraph's motifs guide molecular assembly and leave the choice of which property-associated motifs to retain across successive edits implicit.

  2. Discrete edit operators limit which property trade-offs the search can reach. From a given parent molecule, MARS, MolSearch, and GA+DL can only propose modifications supported by their respective edit operators. Under a limited oracle budget, promising trade-offs may remain unexplored if reaching them requires a sequence of edits. InversionGNN (DST+EPO) uses gradients to guide the search within a relaxed local neighborhood defined by the underlying discrete edits. On the four-objective task, InversionGNN achieved a final HV of 0.2777, compared with 0.5740 for MARS. In this setting, continuous guidance within a local neighborhood yielded lower final HV.

  3. Molecule-level guidance leaves substructure retention implicit. HN-GFN conditions candidate sampling on objective preferences, and RetMol uses retrieved molecules to guide refinement. On the four-objective task, their final HVs were 0.3426 and 0.6054, respectively, compared with 0.7103 for MolDeriveX. Both methods guide candidate generation at the molecular level, leaving implicit which property-associated substructures of the current parent should be retained across successive edits.

Building on DST's continuous optimization of structural edits, MolDeriveX identifies property-associated substructures and carries them forward as anchors. A soft penalty guides edits toward other regions while allowing the optimizer to trade off anchor retention against predicted property gains.

InversionGNN uses EPO to balance predicted property losses during DST editing. Consider two edits with similar predicted gains in synthetic accessibility and similar predicted GSK3β activity. If one edit alters a substructure previously associated with high predicted activity, standard DST+EPO assigns no additional cost to that change. MolDeriveX carries the substructure forward as an anchor and applies a soft penalty to changes in the matched region, giving the optimizer an explicit reason to preserve it.

Q2 & Weakness 1: Core contribution. Why is there no comparison between DST+EPO, fixed anchor optimization, and adaptive anchor optimization? Without these comparisons, how can we determine the benefits of adaptive structural inheritance?

Response: Thank you for suggesting this comparison. To assess the contribution of adaptive structural inheritance, we evaluated MolDeriveX alongside five baselines and ablations on tasks with two, three, and four objectives, using three search seeds per task.

  • InversionGNN (DST+EPO) uses EPO to balance predicted property losses during DST editing.
  • DST uses the differentiable scaffolding tree to optimize a scalar objective that averages the property scores.
  • PZ extracts and fixes the initial anchors, with both the soft structural penalty and decoding protection disabled.
  • L0 uses the same fixed anchors as PZ and enables decoding protection.
  • Fixed L1 enables both decoding protection and the soft structural penalty and uses the initial anchors throughout the search.
  • MolDeriveX enables both mechanisms and can refresh the anchors when a qualifying descendant becomes the new lineage root.

Table 1. Final hypervolume (HV) for DST-based baselines and anchor configurations, reported as mean ± sample standard deviation across three search seeds.

TaskInversionGNNDSTPZL0Fixed L1MolDeriveX
GSK3β–JNK30.7538 ± 0.01560.8024 ± 0.00760.7992 ± 0.01100.7952 ± 0.01110.8145 ± 0.00460.8696 ± 0.0246
QED–SA–GSK3β0.6291 ± 0.02960.7109 ± 0.00310.8534 ± 0.03770.8602 ± 0.00230.8697 ± 0.00510.8813 ± 0.0038
QED–SA–GSK3β–JNK30.3258 ± 0.04930.4765 ± 0.00640.6214 ± 0.04490.6314 ± 0.06080.6514 ± 0.05060.7064 ± 0.0127

Fixed L1 provides the most direct test of adaptive inheritance because it uses the same decoding protection and soft structural penalty as MolDeriveX, with the anchors fixed throughout the search. MolDeriveX achieved higher final HV than Fixed L1 for every seed on each task. The mean gains were 0.0551, 0.0115, and 0.0550 for the two-, three-, and four-objective tasks, respectively. The gains varied most across seeds on the four-objective task, ranging from 0.0064 to 0.1233. InversionGNN serves as an external DST+EPO baseline.

Q3.1 & Weakness 2: What is the reason for using the fragment deletion effect to identify important structures?

Response: The deletion effect measures how the parent molecule’s score changes when a fragment is removed. We use it alongside the isolated fragment score to rank anchors. This use of local perturbations follows the rationale of matched molecular pair analysis and substructure masking [1, 2]. In Table 2, the isolated fragment had a QED of 0.548, and removing it lowered the parent’s QED from 0.948 to 0.543, a decrease of 0.405.

Table 2. QED of the parent molecule, isolated fragment, and connected post-deletion remainder.

StructureFull SMILESQED
Parent moleculeCN1C=C(c2sccc2CF)C(=O)C1(N)c1ccccc10.948
Isolated fragmentCc1sccc1CF0.548
Molecule scored after deletionCN(C)C(N)(C=O)c1ccccc10.543

The two scores showed limited rank agreement, including when the post-deletion remainder stayed connected (Table 3). Across 400 selected initial anchors from 100 parents in one four-objective run, 313 deletions produced multiple components. In those cases, the evaluator scores the largest remaining component, so the measured effect can also reflect the loss of other components. We therefore use the deletion effect as a parent-specific ranking signal.

Table 3. Rank correlation between isolated-fragment QED and the signed deletion effect.

QED anchor subsetnSpearman’s ρ
Both scores available470.271
Both scores available and remainder connected150.31

Q3.2 & Weakness 2: How did you determine the anchor size range (3–18 atoms) and the number of anchors per objective?

Response: We used a shared heuristic across objectives: 3–18 atoms per anchor and at most one anchor per objective. The lower bound excludes single atoms and atom pairs. The upper bound and the cap on anchor count limit how much of the molecule is selected for protection. On QED–SA–GSK3β, this default achieved the highest final HV on both seeds among the four settings tested (Table 4).

Table 4. Sensitivity of final HV to anchor size and count on QED–SA–GSK3β.

Anchor size (atoms)Max. anchors per objectiveSeed 1 HVSeed 2 HVMean HVΔ vs. default
3–18 (default)10.883980.879480.88173—
2–1810.844430.874210.85932−0.02241
3–1820.874460.877940.8762−0.00553
3–1210.878720.873180.87595−0.00578

References

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.