acceptodds
Under review as a conference paper at ICLR 2027

GxEvo: Evolutionary Search for Transcriptome-Conditioned Molecular Design

Abstract

Transcriptome-based drug design (TBDD) seeks molecules that induce a desired transcriptional response. This is a one-to-many inverse problem: a single desired response can correspond to multiple molecular scaffolds, motivating the generation of structurally diverse, response-aligned candidates. Existing TBDD generators, however, struggle to jointly achieve high structural similarity to known actives and scaffold novelty, and their promising outputs do not directly guide subsequent proposals. We introduce GxEvo, a population-based evolutionary framework that uses candidate evaluations to guide further generation under a fixed transcriptional objective. GxEvo combines a condition-specific seed generator, a forward response model trained with response prediction and contrastive learning for candidate discrimination, and a learned mutation operator conditioned on both the parent molecule and the target condition. Across ten gene-knockdown targets and three held-out drug–cell splits, GxEvo achieves the highest mean structural similarity to reference compounds among the evaluated methods. On knockdown targets, it generates 23.8 structurally hit-like unseen scaffolds absent from the task corpus per 100 valid molecules, compared with 3.9 for the strongest prior method under its standard sampling protocol. The advantage persists under matched candidate budgets, while controlled substitutions of existing response predictors and variation operators demonstrate the benefits of the proposed forward model and mutation operator. These results establish evolutionary search as an effective computational approach to generating novel, hit-like molecular hypotheses for subsequent experimental validation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Transcriptome-Based Drug Design, Molecular Generation, Evolutionary Search, AI4Science

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_3oz3OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript proposes an evolutionary framework (GxEvo) for transcriptome-conditioned molecular generation. Instead of generating molecules independently from a transcriptional condition, GxEvo iteratively refines a population of candidates using a learned forward response model and a parent-conditioned mutation operator. The model combines response prediction with contrastive learning to provide a fitness signal for evolutionary selection. The experimental results on ten gene-knockdown targets and three held-out drug–cell splits show improvements in structural similarity and novel hit scaffold generation. The authors also conduct candidate-budget comparisons and ablation studies to evaluate the contributions of the proposed components.

Critical strengths
  • First, this manuscript addresses a meaningful limitation of existing transcriptome-conditioned molecular generators: independently generated candidates cannot directly guide subsequent proposals. The proposed population-based search provides a reasonable way to refine promising molecules. I particularly appreciate the candidate-budget comparison in Table 2, where GxEvo outperforms independent sampling from its own seed generator under the same proposal budget.

  • In addtion, the authors they compare GxEvo with seven existing TBDD methods and evaluate both gene-knockdown targets and held-out drug–cell conditions. The comparisons with alternative forward response models and mutation operators are also useful. In particular, removing InfoNCE reduces novel hit scaffold yield from 23.83 to 14.08, supporting the importance of candidate discrimination in the evolutionary process.

Critical weaknesses
  • The main knockdown results use structural similarity to known actives as a proxy for hit-likeness, while the evolutionary fitness is calculated using predicted transcriptional responses. Although Appendix J provides an additional evaluation using an external response predictor and measured target signatures, it still relies on computationally predicted responses for generated molecules. Therefore, the results do not directly establish that the generated molecules would induce the desired transcriptional changes or exhibit the intended biological activity.

  • Also, GxEvo evaluates 1,500 candidates to return a population of 100 molecules. According to Appendix C, the full search takes approximately 401 seconds per target, compared with 3–8 seconds for several baseline generators and approximately 70 seconds for CURE. Although the authors provide candidate-budget controls, the substantial difference in wall-clock cost limits the practical efficiency advantage of the method. A comparison under matched computational time would help clarify this trade-off.

Questions for authors
  • Since the evolutionary fitness is based on a learned response predictor, how can the authors ensure that repeated optimization does not exploit inaccuracies in this model? Appendix J provides an external response-model evaluation, but could the authors further analyze whether the agreement between the search fitness and independently predicted response alignment changes across evolutionary generations?

  • GxEvo uses a substantially larger computational budget than most baseline generators. Could the authors provide a comparison under matched wall-clock time, or discuss how the performance changes when the number of generations or evaluated candidates is reduced? This would help clarify the practical trade-off between generation quality and computational cost.

5 comments
AuthorsOpenReview ↗

Author Response (1/3)

We thank the reviewer for the positive assessment and the helpful questions. We address them below with new generation-wise analyses and expanded-generation-budget comparisons on the same settings as the paper.

Weakness 1 - Complementary Response-Based Evaluation

GxEvo has higher mean knockdown-response alignment than all seven published generators under both XPert and TranSiGen, extending the response-based evidence beyond Appendix J's held-out conditions. The analysis extends Appendix J's evaluation to the knockdown populations, using two separately trained predictors, XPert and TranSiGen, that take no part in the search. Their separation from the search avoids directly reusing its fitness scores, although shared data sources and correlated model errors remain possible, and experimental measurements remain necessary to establish the effects of the generated molecules.

We score predicted responses against published Level-5 CGS knockdown signatures, with known actives, off-target actives and random molecules as references. The search conditions on Level-3-derived targets (Appendix B), while this score uses Level-5 CGS signatures, which add plate-relative standardization and shRNA consensus, so the change is one of preprocessing rather than a held-out experiment. TRIOMPHE and GxVAEs also use Level-5 profiles for conditioning, and GxEvo still ranks above them. Since the predictors' outputs and the CGS signatures are not calibrated to each other, we read these scores as connectivity across representations rather than as prediction error.

Mean predicted-response alignment against Level-5 CGS signatures, averaged over all ten knockdown targets:

RowXPertTranSiGen
REFERENCE: the target's own ExCAPE actives.1097.1018
GxEvo.1017.1030
TRIOMPHE.0875.0518
CURE.0785.0367
NULL: the other nine targets' actives.0754.0495
Gx2Mol.0750.0543
GexMolGen.0728.0443
FAME.0723.0331
SmilesGEN.0719.0285
GxVAEs.0710.0276
FLOOR: random ZINC-250k.0669.0300

GxEvo's mean score also exceeds the off-target and random controls under both predictors and is comparable to the known-actives reference. To assess ranking utility, we test whether the same score ranks a target's real ExCAPE actives above random ZINC molecules, with no generated molecules involved. XPert meets the AUROC threshold of .60 on 6 of 10 targets and TranSiGen on 7 of 10. Within each predictor's passing subset, GxEvo's advantage holds across every target and baseline. Our reply to Q1 tracks both search fitness and independent evaluation across generations. The knockdown evaluation and per-target predictor checks are now reported in Appendix J.

AuthorsOpenReview ↗

Author Response (2/3)

Question 1 - Fitness and Independent Evaluation Across Generations

GxEvo improves both structural alignment and independently predicted response alignment over the knockdown search. Our reward approximates transcriptional responses, so repeated optimization may exploit prediction errors and increase fitness without improving independently evaluated alignment. The MOLLEO control illustrates this risk: under the same reward, its fitness increases while mean maximum Morgan similarity to known actives peaks at generation 2 and then falls below initialization.

To reduce reliance on surrogate scores alone, GxEvo couples reward-based selection with a learned mutation operator conditioned on both the parent and target response. The operator is trained on LINCS molecule pairs using the child's measured response, providing a response-informed proposal prior alongside the reward. We tracked fitness and independent structural alignment to compare the resulting search trajectories; the structural metric is unavailable to the reward in either method. We also scored the saved GxEvo populations with TranSiGen at eleven generation checkpoints, using W1's Level-5 reference and scoring procedure over all ten targets.

GenerationGxEvo rewardGxEvo MorganGxEvo TranSiGen alignmentMOLLEO rewardMOLLEO Morgan
0.1522.3318.0627.1487.3294
2.2312.3514.0788.2390.3813
4.2613.3583.0879.2791.3681
12.3073.3705.0992.3437.3296
20.3237.3713.1030.3729.3132

GxEvo retains improved structural alignment through generation 20, while TranSiGen response alignment increases from .0627 to .1030. This contrast is consistent with response-informed variation helping limit the structural drift observed with MOLLEO under the same reward on this benchmark. GxEvo's final novel hit scaffold yield is 23.83 per 100 valid molecules, compared with 12.60 at MOLLEO's generation-2 Morgan peak. GxEvo thus combines sustained improvement in population similarity with greater scaffold breadth.

One MOLLEO seed and four GxEvo seeds. TranSiGen does not participate in search; its endpoint scores reproduce W1's evaluation and retain the same interpretation.

The held-out conditions tell a less uniform story. There we evaluated Appendix J's PRnet-based response-specificity lift at eleven points during search. This compares predicted responses with the requested measured response and other measured responses in the same cell; PRnet does not participate in search. Values use Table 24's normalization, with random ZINC at 0 and the ground-truth drug at 100.

SplitLift at gen 0Peak liftLift at gen 20Net changeSpearman(fitness, lift)
RANDOM83.696.6 (gen 14)95.9+12.3+0.83
SCAFFOLD92.892.8 (gen 0)91.7-1.1+0.11
CELL83.990.4 (gen 4)87.4+3.5-0.31

Independent response-specificity lift finishes above initialization on random (+12.3) and cell (+3.5), while remaining approximately flat on scaffold (-1.1). On the cell split, alignment peaks around generations 4-6 and then declines while fitness continues to rise, a pattern consistent with late-stage reward exploitation, although the final score remains above initialization. Each Spearman coefficient is computed across the eleven population-average generation checkpoints (0, 2, ..., 20). A practical mitigation is to choose the stopping generation on separate validation conditions using saturation of an external response metric, then fix that generation before evaluating held-out conditions. This question improved our analysis of the search. We have added the generation-wise results and discussed validation-based stopping in Appendix I; selecting a stopping generation on separate validation conditions remains a proposed mitigation rather than an evaluated stopping rule.

AuthorsOpenReview ↗

Author Response (3/3)

Weakness 2 / Question 2 - Computational Budget and Earlier Stopping

GxEvo retains higher mean similarity and novel hit scaffold yield when baselines receive an expanded generation allowance, while earlier stopping offers a practical way to reduce its cost. For the baseline comparison, we allocate GxEvo's full search time to candidate generation and then apply our reward model to select the best 100 molecules, as in Table 2.

We measured generation throughput for five baselines and GxEvo's total search time in the same session on one otherwise-unused GPU, then estimated the candidate counts that fit this generation allowance. Reward scoring and selection are additional to the baseline generation allowance, giving the baselines more total computation than the reference search budget. The resulting draw budgets are 65x-5,696x larger than the 1,500-draw control, reaching up to 8.5M candidates per target. For GexMolGen and CURE, Appendix C-based cost estimates imply generation allowances of 526 and 782 draws; we retain their larger 1,500-draw controls.

Mean similarity to the target's known actives, averaged over ten MCF7 knockdown targets:

MethodPublished1,500 draws + rewardExpanded generation allowance + rewardCandidate draws in expanded control
TRIOMPHE.2049.2117.2053905,119
GxVAEs.2779.2714.28137,468,264
FAME.2716.2737.290797,253
SmilesGEN.2810.2944.29718,543,677
Gx2Mol.2990.3136.30631,071,204
GexMolGen.2946.3063-1,500 (retained control)
CURE.2842.2620-1,500 (retained control)
GxEvo-.3713.37131,500

Increasing the sampling budget beyond 1,500 yields limited additional improvement. Despite a 65x-5,696x increase in draws under the expanded generation allowance, mean similarity improves by at most .017 and decreases for some baseline configurations. Novel hit scaffold yield reaches at most 6.8 per 100 molecules, compared with 23.83 for GxEvo. Paired over the ten targets, GxEvo exceeds every baseline evaluated with the expanded generation allowance and reward selection on both metrics. Gx2Mol's best-of-population similarity also decreases from .748 at 1,500 draws to .488 at 1.07M draws, compared with .617 for GxEvo.

Earlier stopping provides a complementary way to reduce GxEvo's cost. The population stored at each generation is the output that a run stopped at that point would return. The following checkpoints use the same four seeds as Tables 1-2; intermediate times are estimated with the Appendix C cost model, using the measured seed-only and full-search times as endpoints.

GenerationCandidate drawsTime (s)Mean similarity ↑NHS/100 valid ↑
010029.4.331812.64
6520140.9.362019.65
12940252.5.370522.50
201,500401.3.371323.83

Generation 6 requires 35% of the full wall-clock time while retaining 76% of the mean-similarity gain and 63% of the novel hit scaffold gain. Mean similarity plateaus around generations 12-14. Together, the expanded-baseline controls and generation-wise results show that GxEvo's quality advantage persists with greater baseline compute, while earlier stopping retains most of the improvement at substantially lower search cost. We have added both analyses to Appendix C and summarize the cost trade-off in Section 4.3.

We thank the reviewer again for the questions, and we are happy to discuss further.

Reviewer_vb1HOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents GxEvo, an evolutionary framework for Transcriptome-Based Drug Design (TBDD). Given a cellular basal expression profile and a desired transcriptional response, the method aims to generate candidate molecules capable of inducing the corresponding transcriptional changes. GxEvo combines a condition-specific molecular generator, a forward response prediction model, and a learned mutation operator to iteratively optimize candidate molecules. Across ten gene-knockdown targets, GxEvo achieves a mean Morgan similarity of 0.371 to known active compounds and a Novel Hit Scaffolds (NHS) score of 23.83 per 100 valid molecules, outperforming most existing baselines.

Critical strengths

The experimental evaluation is comprehensive and well organized, covering multiple evaluation settings, baseline methods, and ablation studies. In particular, the authors include random molecules from ZINC and ChEMBL, as well as compounds active against other targets, as additional controls. These comparisons strengthen the experimental analysis and are commendable.

The results in Table 4 are also noteworthy. The proposed response model demonstrates clear improvements over random selection and existing prediction models. I particularly appreciate the use of Top-k retrieval and ranking-based metrics to evaluate the discriminative ability of the response model. These metrics offer a useful perspective beyond conventional regression-based evaluation.

Critical weaknesses
  1. The main evaluation metrics do not adequately reflect the actual objective of transcriptome-based drug design.

The most significant concern is that the main Knockdown Benchmark does not directly verify whether the generated molecules can induce the desired transcriptional response. In Section 4.1, the authors primarily evaluate generated molecules based on their Morgan similarity to known active compounds associated with the corresponding genes. However, structural similarity does not necessarily imply similar biological activity, let alone the ability to reproduce a specific transcriptional response.

Consequently, improvements in structural similarity or scaffold novelty cannot be interpreted as convincing evidence of improved TBDD performance. In my view, even an evaluation using a strong, independently trained transcriptomic response prediction model would be more directly relevant than structural similarity alone, although such predictions would still require careful validation. While the authors provide additional evaluation using an external predictor, the main conclusions remain heavily dependent on indirect metrics.

  1. Important biological factors affecting transcriptional responses are insufficiently considered.

Drug-induced transcriptional responses are highly dependent on dosage, treatment duration, and cellular context. However, the current data aggregation strategy and fitness formulation may obscure these biologically important factors. In particular, aggregating responses across different experimental conditions and evaluating candidates at a fixed dose may not adequately capture dose-dependent or context-specific transcriptional effects.

The authors should clarify how these choices affect the biological interpretation of the optimization objective and investigate whether the reported improvements remain consistent across different doses and cellular conditions.

  1. The experimental coverage and dataset diversity remain insufficient to establish broad generalizability.

Although the paper includes several evaluation settings, the main gene-knockdown experiments cover only ten targets. This limited target coverage makes it difficult to assess whether the proposed method generalizes to more diverse biological mechanisms and transcriptional perturbations.

Furthermore, broader pharmacogenomic and perturbation datasets could provide additional evidence of biological relevance. For example, GDSCv1/v2 contain complementary drug-response measurements, while transcriptomic perturbation resources beyond the currently used L1000/LINCS data could support more direct evaluation of transcriptional responses. The authors should discuss the applicability of these resources, their limitations, and the extent to which the current experimental results generalize beyond the evaluated targets and datasets.

Questions for authors

see weak.

Overall, the paper presents an interesting framework with substantial experimental effort, and the proposed evolutionary search strategy shows promising computational results. However, my primary concern is the mismatch between the stated objective of transcriptome-conditioned molecular design and the metrics used to establish its effectiveness. Structural similarity to known active compounds is insufficient to demonstrate that generated molecules can reproduce the desired transcriptional responses. I currently assign a score of 2. I would be willing to reconsider this rating if the authors provide stronger evidence of biological relevance, particularly through more convincing response-based evaluations, and adequately address the concerns regarding experimental conditions and generalizability.

5 comments
AuthorsOpenReview ↗

Author Response (1/3)

We thank the reviewer for the detailed and constructive review. We address the three weaknesses below with new experiments, including response-based evaluations with predictors that take no part in the search and an expanded panel of forty additional knockdown targets.

Weakness 1 - Response-Based Evaluation

GxEvo achieves lower PRnet-MSE on held-out conditions and higher mean XPert/TranSiGen alignment on knockdown targets than the evaluated published generators. These response-based comparisons complement structural similarity and are evaluated with predictors that do not participate in search.

Specifically, Table 3 already includes a response-based metric: PRnet-MSE compares each generated molecule's predicted response with the condition's measured response, using a model that takes no part in our search or reward. GxEvo achieves the lowest PRnet-MSE on all three splits. Among the published baselines, Gx2Mol comes closest:

SplitGxEvoGx2Mol
Random.00888.00929
Scaffold.00957.00981
Cell.01754.01775

The absolute differences are small, but the paired comparisons favor GxEvo in all 18 baseline-by-split tests (Table 12).

Appendix J and Table 24 ask whether the predicted response is specific to the requested condition. For a molecule generated for condition i, we compare its MSE to that condition's measured response (diag) with its average MSE to ten other measured responses in the same cell line (off). A positive off-minus-diag margin means the predicted response is closer to the requested response than to the alternatives. Table 24 normalizes this margin between random ZINC molecules (0) and the condition's own ground-truth drug (100), with all reference drugs evaluated through the same PRnet protocol. GxEvo reaches 95.9, 91.7 and 87.4 on the random, scaffold and cell splits, respectively, whereas a real drug assigned to the wrong condition reaches -2.7, 2.1 and 13.5. These are normalized specificity margins, not percentages of experimentally reproduced response.

We use two separately trained predictors, XPert and TranSiGen, which like PRnet take no part in the search, and compare generated populations with known actives, off-target actives and random molecules against published Level-5 CGS knockdown signatures. The generated molecules remain computational hypotheses requiring experimental validation.

Mean predicted-response alignment against Level-5 CGS signatures, averaged over all ten knockdown targets:

RowXPertTranSiGen
REFERENCE: the target's own ExCAPE actives.1097.1018
GxEvo.1017.1030
TRIOMPHE.0875.0518
CURE.0785.0367
NULL: the other nine targets' actives.0754.0495
Gx2Mol.0750.0543
GexMolGen.0728.0443
FAME.0723.0331
SmilesGEN.0719.0285
GxVAEs.0710.0276
FLOOR: random ZINC-250k.0669.0300

GxEvo has higher mean alignment than all seven published generators and both displayed controls under both predictors, and is comparable to the known-actives reference. Not every target gives an equally informative score, so we check each predictor per target: can the alignment score separate known ExCAPE actives from random ZINC molecules, at AUROC >= .60? XPert passes on 6 of 10 targets and TranSiGen on 7 of 10, and within each passing subset GxEvo exceeds all seven baselines on every target. The full-panel means above retain all ten targets and apply no such filter. This evaluation scores against Level-5 CGS signatures, whereas the search conditions on Level-3-derived targets (Appendix B). The two differ in preprocessing, as Level-5 adds plate-relative standardization and shRNA consensus. TRIOMPHE and GxVAEs also use Level-5 profiles for conditioning, and GxEvo still ranks above them. Since the two representations are not calibrated to each other, we read these scores as connectivity across representations rather than as prediction error, and the known actives as a reference level rather than a biological ground truth.

Together, the held-out condition-specificity analysis and the two additional knockdown predictors support GxEvo's response alignment beyond structural similarity. We have added the knockdown evaluation and per-target predictor checks to Appendix J and summarize the response-based results in Sections 4.2 and 4.5.

AuthorsOpenReview ↗

Author Response (2/3)

Weakness 2 - Dose, Exposure Time and Cellular Context

GxEvo retains its mean performance advantage across the tested scoring doses and measured-dose strata; the exposure-time and cell-wise analyses provide a more detailed view of robustness across conditions. Two of the three factors are already modeled in the submitted version. Cellular context enters through the target cell's basal profile, which conditions both the reward model and the proposal models (Section 3), and dose enters the reward model as the log-dose FiLM input of Equation 2, with the fixed scoring dose specified in Appendix A. Exposure time is the factor that is not, and we treat the three separately below.

We first examine cellular context, which is represented in the model rather than aggregated away. The cell split holds out an entire tissue (Appendix G), where GxEvo achieves the highest Morgan and Fraggle similarity and the lowest PRnet-MSE in Table 3. We additionally decomposed the original held-out results by cell line. Among the eighteen cell-by-split groups with at least eight conditions, seventeen favor GxEvo over Gx2Mol, the strongest baseline in aggregate on each split; group sizes range from 8 to 134.

We next examine dose. Dose enters the reward model as a log-dose FiLM conditioning input, while the original protocol scores every candidate at d_eval = 10 μM. We repeated the knockdown search at 1 μM and with a dose-marginal objective that averages predicted responses over 0.1, 1 and 10 μM before computing fitness. All rows below use the same ten targets and two seeds, with the 1,500-candidate budget and other settings unchanged; the 10 μM reference reuses the corresponding existing runs.

ConfigurationMean similarity ↑NHS/100 valid ↑
GxEvo (10 μM).374 ± .00823.15 ± 1.63
w/ 1 μM scoring.369 ± .00517.90 ± 2.97
w/ dose-marginal scoring.349 ± .00420.25 ± .35

Mean ± sample SD across two seed-level averages over ten targets.

GxEvo outperforms all seven 1,500-candidate baselines from Table 2 on both mean similarity and novel hit scaffold yield under all three scoring settings.

This sweep tests sensitivity to the dose used during optimization. Complementarily, we asked whether the original held-out advantage persists across conditions measured at different dose ranges. The table gives the smallest paired mean Morgan advantage over the six Table 3 baselines, with condition counts in parentheses:

Median measured doseRandomScaffoldCell
<10 μM+.0496 (83)+.0396 (85)+.0532 (92)
10 μM+.0294 (98)+.0070 (97)+.0174 (82-83)
>10 μM+.0275 (19)+.0181 (18)+.0131 (25)

GxEvo's mean advantage remains positive in every dose stratum on all three splits.

Unlike dose and cellular context, however, exposure time is not retained as a model input under the benchmark's compound-cell aggregation. We recovered exposure-time annotations for every training condition from its original wells; the measured durations are 3, 6, 24 and 48 h, and roughly a third of conditions pool more than one of them.

To test whether this aggregation affects the reported comparisons, we applied the same recovery to the held-out conditions and stratified them by whether their measured wells share a single duration. As above, each entry is the smallest paired mean Morgan advantage over the six Table 3 baselines, with condition counts in parentheses:

Measured exposure timeRandomScaffoldCell
Single duration+.0365 (132)+.0246 (137)+.0431 (176-177)
Mixed durations+.0377 (68)+.0244 (63)-.0069 (23)

GxEvo's advantage is positive over every baseline in five of the six groups and significant in all five, with a largest p of .004. In the remaining group all six differences fall within ±.007 and none is significant, the smallest p being .19. Those 23 conditions all come from one held-out cell line and are matched by no evaluated method, every arm scoring near .10 within a total spread of .009, whereas that line's 111 single-duration conditions give GxEvo +.060 to +.106 over the same baselines. The advantage also persists in the single-duration strata, although this analysis does not isolate a causal effect of exposure-time mixing. Explicit time conditioning would require reconstructing time-resolved training conditions and retraining the reward model.

These comments helped clarify the presentation. The revised manuscript states the aggregation and fixed-dose protocol in Sections 3–4 and Appendix A, and reports the dose-sensitivity, exposure-time, and cell-wise analyses in Appendix Q.

AuthorsOpenReview ↗

Author Response (3/3)

Weakness 3 - Target Coverage and Generalization Beyond the Benchmark

We constructed forty new knockdown targets from raw LINCS L1000 data, each with an MCF7 shRNA consensus signature, a reference set of known actives, and a distinct annotated mechanism, giving forty Drug Repurposing Hub mechanism classes. Conditioning inputs were rebuilt from L1000 and six of the seven baselines were evaluated without retraining; FAME's release supports only the original ten genes. GxEvo and its resampling control each use 1,060 candidate draws per target and seed over twenty generations, with two seeds; the ten-target reference below is evaluated under the same configuration.

Absolute similarity is not comparable between the two panels, because the new reference sets are smaller: a median of 622 known actives against 2,202 for the original ten. We therefore also score each method against a floor that shares this property: actives assigned to the other thirty-nine targets, including related kinases. That floor is .257 on the new panel against .334 on the original ten, and GxEvo's margin over it is +.043 here against +.026 there.

GxEvo exceeds the evaluated baselines on both primary structural metrics:

Forty new targetsMean similarity ↑NHS/100 valid ↑
GxEvo0.30010.93
budget-matched resampling (ours)0.2734.20
Gx2Mol0.2460.21
GexMolGen0.2350.88
SmilesGEN0.2310.11
GxVAEs0.2270.21
CURE0.2270.02
TRIOMPHE0.1840.19

NHS is per 100 valid molecules. CURE shows the repaired-conditioning variant; both were evaluated.

Paired by target, GxEvo's mean-similarity advantage over the evaluated baseline arms ranges from +0.054 to +0.116, and its NHS advantage ranges from +10.05 to +10.91 per 100 valid molecules, with significant differences in all comparisons.

Because the baselines retain their published operating points and therefore smaller candidate budgets, our primary controlled comparison is with our own budget-matched resampling control, which GxEvo significantly exceeds on mean similarity, hit rate and novel hit scaffold yield. Pooling all fifty targets under the common 1,060-draw, two-seed configuration gives mean similarity .312 versus .287, hit rate .161 versus .079 and NHS/100 valid 12.86 versus 6.65 for GxEvo and resampling, respectively; target-level paired differences reach significance on all three metrics.

Training/reference overlap for the new targets has a median of 0.64% of reference actives; excluding those actives changes GxEvo's mean similarity from .300 to .299 and NHS/100 valid from 10.93 to 10.80, with its advantage over resampling still significant on all three metrics.

We also applied W1's signature-connectivity analysis to all forty targets with the same predictors and references. We applied the same per-target validity check as in W1: AUROC >= 0.60 for ranking known ExCAPE actives above random ZINC molecules, without using generated molecules. The table reports the fifteen targets passing the check under both predictors; as in W1, the known actives provide a reference level rather than an upper bound.

Population or referenceXPertTranSiGen
GxEvo.2301.2049
REFERENCE: the target's own ExCAPE actives.2008.1364
budget-matched resampling.1937.1386
GxEvo seed only, generation 0.1767.1092
NULL: the other thirty-nine targets' actives.1765.1268
TRIOMPHE.1612.0945
Gx2Mol.1525.1027
CURE.1509.0969
GexMolGen.1439.0624
SmilesGEN.1335.0548
GxVAEs.1298.0549
FLOOR: random ZINC-250k.1292.0596

On this subset both predictors place GxEvo above every evaluated generator, the off-target null, the random floor, the generation-0 population and budget-matched resampling, while every baseline falls at or below the null. Paired signed-rank and sign tests support the baseline differences on the common fifteen-target subset. Across all forty targets, GxEvo retains higher mean alignment than every baseline under both predictors and exceeds each baseline on 60-75% of targets.

A leave-one-out control on GDSCv1/v2 found only a weak association between structural similarity and measured MCF7 sensitivity (Spearman rho = .159), and its viability endpoint does not assess reproduction of a transcriptional response. Beyond LINCS, sci-Plex and similar single-cell perturbation resources could support external evaluation after aligning gene coverage and normalization. This suggestion strengthened the evaluation.

We have added the forty-target results to Section 4.5 and included the panel construction, overlap audit, per-target results, and structural and response-based evaluations in Appendix P, together with the discussion of GDSC and sci-Plex.

We thank the reviewer again for the detailed comments, and we are happy to discuss further.

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

GxEvo evolves a population of molecules toward a target transcriptional response using a conditional seed generator, a learned mutation operator, and a response-prediction reward. It improves structural similarity to reference compounds and the yield of unseen, structurally hit-like scaffolds in the reported computational benchmarks.

Critical strengths

The study addresses the one-to-many nature of transcriptome-conditioned design: it evaluates the quality and scaffold breadth of a returned population, rather than only its closest molecule. The knockdown comparisons include useful controls. Matching the 1,500-candidate budget and reward-based selection does not close the gap; GxEvo also exceeds independent resampling from its own seed generator. The authors audit several reproduction and metric issues, including baseline conditioning, validity, and the novelty reference corpus.

Critical weaknesses

The main evaluation relies on proxies. Structural similarity to known actives and predicted transcriptional responses do not establish that the generated molecules have the desired biological effect. Experimental validation would strengthen the central claim. The held-out comparisons do not match the search budget. GxEvo evaluates many more candidates than the generation baselines in Table 3. The budget-matched control is limited to the knockdown setting, leaving the source of the held-out gains unclear. The methodological novelty is moderate. The framework combines established evolutionary search and molecular modeling components. Its distinct algorithmic contribution beyond this integration could be stated more clearly.

Questions for authors

Can the authors validate any generated molecules experimentally? Do the held-out results persist under matched candidate budgets?

4 comments
AuthorsOpenReview ↗

Author Response (1/3)

We thank the reviewer for the positive assessment and the constructive comments. We address the three points below with new experiments on the knockdown benchmark and the held-out splits.

Weakness 1 / Question 1 - Experimental Validation and Response-Based Evaluation

We agree that experimental validation is needed to establish the biological effects of generated molecules. We do not currently have access to wet-lab validation, and our Conclusion and Ethics statement explicitly describe the generated molecules as hypotheses requiring experimental validation because both the reward and evaluation metrics are computational proxies.

On the other hand, to assess transcriptional alignment more directly within this setting, we conducted an additional response-based evaluation of the knockdown populations. The additional evaluation uses two separately trained response predictors, XPert and TranSiGen, neither of which provides the reward or participates in the search. We compare their predicted responses with published consensus knockdown signatures (Level-5 CGS), using the same scoring procedure for generated molecules, known actives and control compounds. This provides a complementary computational comparison against a differently processed knockdown reference; the underlying LINCS resource is shared with the search data.

Mean predicted-response alignment against Level-5 CGS signatures, averaged over all ten knockdown targets:

RowXPertTranSiGen
REFERENCE: the target's own ExCAPE actives.1097.1018
GxEvo.1017.1030
TRIOMPHE.0875.0518
CURE.0785.0367
NULL: the other nine targets' actives.0754.0495
Gx2Mol.0750.0543
GexMolGen.0728.0443
FAME.0723.0331
SmilesGEN.0719.0285
GxVAEs.0710.0276
FLOOR: random ZINC-250k.0669.0300

Under this supplementary alignment score, GxEvo has a higher mean than all seven published generators, the off-target null and the random floor under both predictors. Its mean score is comparable to that of the target's known actives, which serve as a reference rather than an upper bound on this metric. The predictor check (AUROC >= .60 for known actives versus random ZINC) passes on 6 of 10 targets under XPert and 7 under TranSiGen. The table retains all ten targets; GxEvo exceeds all seven baselines on every target in each predictor's passing subset. These comparisons provide response-based support for GxEvo's advantage beyond structural similarity, and we have added the evaluation and its per-target predictor checks to Appendix J and summarize the response-based result in Section 4.2.

AuthorsOpenReview ↗

Author Response (2/3)

Weakness 2 / Question 2 - Budget Matching for Held-Out Conditions

GxEvo retains higher mean Morgan and Fraggle similarity than the four evaluated published baselines on all splits under a matched candidate budget and shared reward model. We conducted the comparison on the same 200 held-out conditions per split. For each of four baselines, we draw 1,500 candidates per condition and use our reward model to select the best 100 distinct molecules. The baselines therefore receive both GxEvo's candidate budget and its learned response model for selection. This comparison covers four held-out baselines; GexMolGen and CURE are excluded because of their substantially higher generation costs.

Mean Morgan similarity to the ground-truth compound:

MethodCandidatesSelectionRandomScaffoldCell
TRIOMPHE100original protocol.1482.1142.1167
TRIOMPHE1,500our reward.1411.1184.1210
GxVAEs100original protocol.1506.1707.1398
GxVAEs1,500our reward.1653.1849.1526
SmilesGEN100original protocol.1501.1762.1310
SmilesGEN1,500our reward.1813.1903.1538
Gx2Mol100original protocol.1838.2087.1467
Gx2Mol1,500our reward.2097.2162.1724
Our seed generator100no selection.2072.2374.1568
Our independent resampling1,500our reward.2205.2366.1754
GxEvo1,500our reward.2226.2340.1847

GxEvo leads the four published baselines on every split even when they use our reward to select from the same candidate budget, and paired tests over the 200 conditions support its advantage over TRIOMPHE, GxVAEs and SmilesGEN. Against Gx2Mol, the Morgan difference is significant only on the scaffold split.

Our reward also improves candidates proposed by another generator. Within the same Gx2Mol 1,500-candidate pools, reward-based selection of distinct molecules changes Morgan similarity by +.0164, +.0031 and +.0195 relative to uniform selection on the random, scaffold and cell splits, respectively. Paired tests support these selection gains on all three splits, with the largest improvements on random and cell.

The seed-only and independent-resampling rows identify the sources of the held-out result: evolutionary refinement adds +.0154 and +.0279 mean Morgan similarity over the seed population on the random and cell splits, while the three arms lie within .004 of one another on the scaffold split, where our conditional generator alone already exceeds every published baseline. Conditional generation and reward-guided refinement therefore contribute on different splits, and the full framework retains the highest mean Morgan similarity on the random and cell splits together with the highest Fraggle similarity over the published baselines on all three.

The matched-budget populations also differ on the remaining evaluation metrics:

Method, 1,500 candidates + our rewardFraggle similarityPRnet-MSENoveltyScaffold novelty
GxEvo.4555 / .4693 / .4241.00888 / .00957 / .01754.908 / .908 / .951.414 / .404 / .501
Gx2Mol.4431 / .4551 / .4060.00894 / .00949 / .01732.458 / .427 / .461.118 / .092 / .124
SmilesGEN.4187 / .4326 / .3861.00937 / .01008 / .01811.351 / .392 / .393.096 / .087 / .102
GxVAEs.4038 / .4263 / .3871.00993 / .01054 / .01848.987 / .984 / .986.721 / .741 / .760
TRIOMPHE.3578 / .3072 / .3275.01131 / .01272 / .01999.985 / .995 / .991.860 / .895 / .848

Each entry gives the random / scaffold / cell split. Higher Fraggle similarity, novelty and scaffold novelty are better; lower PRnet-MSE is better.

GxEvo also retains the highest Fraggle similarity on every split and achieves the lowest PRnet-MSE on random, while Gx2Mol combined with our reward obtains slightly lower errors on scaffold and cell.

GxVAEs and TRIOMPHE achieve high molecular and scaffold novelty, while GxEvo provides higher Morgan and Fraggle similarity on every split. Against Gx2Mol and SmilesGEN, which are closer on structural similarity, GxEvo achieves both higher structural similarity and substantially higher molecular and scaffold novelty. For Gx2Mol, molecular and scaffold novelty remain .427-.461 and .092-.124 after reward-based selection, compared with .908-.951 and .404-.501 for GxEvo. Together, these comparisons show GxEvo's strength in combining reference alignment with novel molecular scaffolds under the same candidate budget and reward model, and we have added this held-out budget-matched control to Appendix F and summarize it in Section 4.4.

AuthorsOpenReview ↗

Author Response (3/3)

Weakness 3 - Distinct Algorithmic Contribution

To our knowledge, GxEvo is the first framework to formulate TBDD as population-based evolutionary optimization. The TBDD baselines in our comparison generate candidates independently from a transcriptional condition; GxEvo instead uses selected candidates to guide subsequent proposals, explicitly refining a population under the same target response.

This formulation requires more than combining an existing generator, response predictor and molecular evolutionary algorithm. The transcriptional effect of an unmeasured candidate must be predicted with enough molecule-specific resolution for selection, while variation must propose useful changes toward a high-dimensional, cell-dependent response. Predictive fidelity alone does not ensure candidate discrimination, and generic molecular variation does not incorporate the target transcriptional condition.

We address these requirements through two task-specific designs. Our forward model combines response prediction with a standard contrastive objective to preserve molecule-specific distinctions for fitness-based ranking. Our parent-aware mutation operator is trained on LINCS molecule pairs, using the child's measured response as the conditioning target. Together, these components connect selection and proposal generation through the same transcriptional objective.

The reward and operator substitutions in Tables 5 and 6, together with the contrastive-loss ablation, test why these adaptations are needed within the same search framework. An additional 2x2 comparison across four search seeds and ten knockdown targets examines their interaction:

learned operatorindependent resampling
our reward.3713 / 23.83.3436 / 16.48
random selection.3413 / 11.60.3295 / 12.38

Mean similarity / novel hit scaffolds per 100 valid molecules; four seeds and ten targets, matching the main benchmark.

The full combination achieves the highest mean similarity and scaffold yield in the 2x2 comparison. The learned operator's scaffold-yield benefit emerges under reward-based selection, supporting the interaction between our fitness and variation mechanisms. Under random selection the operator has no fitness signal to direct its proposals, so parent-conditioned variation and independent resampling perform comparably; the two mechanisms are therefore effective in combination rather than individually. Our contribution is therefore the population-search formulation of TBDD and the task-specific fitness and variation mechanisms that make this formulation effective. The revised manuscript includes the 2x2 comparison in Appendix I and summarizes its result in Section 4.6.

We thank the reviewer again for the comments, and we are happy to discuss further.

Dear Reviewer WTrv,

As the author–reviewer discussion period is approaching its end, we would greatly appreciate it if you could take a moment to review our rebuttal and additional experimental results. In response to your three points, we added a budget-matched comparison on the held-out splits, a response-based evaluation with two predictors that take no part in the search, and a 2x2 comparison isolating the reward and the variation operator; all three are included in the revised manuscript.

We hope our responses have addressed your concerns. If you have any remaining questions or concerns, we would be happy to provide further clarification before the discussion period closes.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.