acceptodds
Under review as a conference paper at ICLR 2027

PhAME: Phenotype-Aware Molecular Editing via Latent Diffusion

Abstract

Small-molecule drug discovery requires simultaneous optimization of numerous properties of candidate molecules. These properties can be investigated through the analysis of high-dimensional biological signatures, such as cell morphology and transcriptomic perturbations, which provide a rich perspective on the underlying biological mechanisms. However, existing generative methods, which use those signatures for optimization, fail to meet two key requirements: providing precise guidance toward desired phenotypic signatures while maintaining structural proximity to a known hit. We introduce PhAME (Phenotype-Aware Molecular Editing), a latent diffusion framework that overcomes this challenge by recasting molecular optimization as editing in the latent space of a pretrained graph-based VAE. Our central contribution is a compositional classifier-free guidance scheme with two independent scales, one for the target signature (phenotype or molecular property) and one for similarity to the seed structure, allowing practitioners to control the tradeoff between these two objectives. Empirical evaluations across diverse benchmarks, including docking score optimization and multimodal phenotypic generation, demonstrate that PhAME achieves state-of-the-art results while maintaining high chemical validity and novelty.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular Generation, Latent Diffusion Models, Controllable Generation, Phenotype-Conditioned Generation, Drug Discovery

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qmbHOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PhAME, a latent-diffusion framework for molecular editing conditioned on molecular properties or biological phenotypes. PhAME operates in the latent space of a pretrained graph VAE and uses separate guidance scales for the target signature and structural anchoring to a seed molecule. The method is evaluated on logP optimization, docking-score optimization, transcriptomic-guided generation, and Cell-Painting-guided generation.

Critical strengths
  1. The problem formulation is interesting and practically relevant. Combining phenotype-conditioned generation with hit-to-lead-style molecular editing is a useful setting, and the separation of target and structural guidance is intuitive.

  2. The logP experiments provide useful evidence for the proposed guidance mechanism. The comparison with standard CFG and the additional ablations suggest that c-CFG provides useful control over target optimization and structural preservation.

Critical weaknesses
  1. The main biological editing setting is not evaluated as directly as the logP task. The transcriptomic experiment uses seed molecules, but target relevance is mainly measured by similarity to known ligands. The main Cell Painting experiment is de novo generation, while the editing results are mostly qualitative recovery examples. I think the paper demonstrates molecular editing clearly on logP and biological conditioning clearly on the transcriptomic and Cell Painting tasks, but the combination of the two is less directly quantified.

  2. Some key experiments do not clearly isolate the source of the reported gains. The docking experiments claim a new SOTA while omitting post-MOOD methods such as GEAM [1] and f-RAG [2], which report stronger results on the same five-target docking benchmark. In the transcriptomic experiment, retrieval and editing are evaluated together, making it difficult to tell how much each contributes. The curriculum ablation also changes several training factors at once, so it is unclear whether the improvement comes from the curriculum itself.

Questions for authors
  1. Can the authors provide quantitative results for the editing setting illustrated in Figure 1 on either the transcriptomic or Cell Painting task, reporting both target relevance and similarity to the input seed?

  2. What do the retrieved transcriptomic seeds already achieve before editing? How much improvement does PhAME provide over these seeds across the full set of targets?

  3. Why are post-MOOD docking methods such as GEAM [1] and f-RAG [2] omitted from Table 2 despite the paper's SOTA claim? Both report stronger results on the same five-target docking benchmark.

  4. Does the curriculum still improve docking performance when the training budget is controlled? Also, which percentile schedule was actually used in Tables 8–9?

  5. What evidence supports the claims about MoA-aware chemistry and structure–activity relationships given that the Cell Painting evaluation is based on structural proxies? Could the authors also clarify the statement that only PhAME exceeds chance in all three representation spaces? The GFlowNet means in Table 11 appear to exceed the 14.3% chance level in all three spaces.

References:

[1] Lee et al., Drug Discovery with Dynamic Goal-aware Fragments, ICML 2024.

[2] Lee et al., Molecule Generation with Fragment Retrieval Augmentation, NeurIPS 2024.

6 comments

W1/Q1. Quantitative evaluation of biological molecular editing.

We thank the Reviewer for highlighting the need for a more direct quantitative evaluation of biological molecular editing. We agree that the original manuscript demonstrated editing more directly for logP than for the biological tasks. To address this gap, we added a Cell Painting editing benchmark and expanded the transcriptomic editing evaluation, jointly reporting target relevance, similarity to the input seed, and molecular validity. These additions are summarized in GR1 and GR2 of our General Response, with full results in Appendices G and E.

W2/Q3. Docking comparisons under oracle-free and oracle-guided settings.

We thank the Reviewer for highlighting this omission. An important distinction, also emphasized in MOOD (Appendix E.2, Property Optimization), concerns the use of expensive oracle calls during generation. PhAME learns from a fixed, pre-scored dataset and subsequently generates candidates through classifier-free-guided denoising. Like MOOD, PhAME does not require docking-oracle calls during sampling. Unlike MOOD, PhAME also avoids gradients from a separately trained property predictor, using classifier-free guidance instead. GEAM and f-RAG use target-property feedback to adapt their search. Their stronger docking scores are obtained in a different generation regime, which should be made explicit in the comparison. Nevertheless, this distinction does not justify an unrestricted docking SOTA claim, particularly because our table already includes oracle-guided baselines: REINVENT, MORLD, FREED, FREED-QS.

Following the Reviewer’s comment, we have divided the results in Table 3 into oracle-free and oracle-guided generation methods. Among the evaluated oracle-free methods, PhAME achieves the best mean docking score on all five targets and the highest mean hit ratio on four of the five targets.

For oracle-guided generation, we use the baseline results reported in GEAM and additionally compare with f-RAG. We evaluate PhAME in an editing setting with a lightweight random forest surrogate that chooses candidates for the next optimization round. PhAME + RF achieves competitive results compared with GEAM and f-RAG. It achieves the highest hit ratio on all five targets, and the best mean docking score on FA7, while falling behind f-RAG on the remaining targets and remaining close to GEAM.

We have also compared computational runtimes. The PhAME + RF pipeline produces and docks 3,000 selected candidates in less than two hours on an RTX 4090, including surrogate-model training, PhAME training, molecule generation, and docking. In our tests, GEAM required approximately 2.5 hours for the corresponding task, while f-RAG reports approximately two hours per docking task in its appendix.

For transcriptomic-guided editing, please see the “Transcriptomic-guided editing” subsection of GR2 in our General Response.

For the curriculum ablation, please also see our response to Q4.

Q2. Improvement over transcriptomic retrieval.

PhAME improves the maximum Tanimoto similarity to known binders over the retrieved seeds by 5.3%-73.0% across targets (+30.4% on average).

TargetSeed maximum TanimotoPhAME (mean ±\pm SD)Relative gain
AKT10.3380.459 ±\pm 0.054+35.8%
AKT20.2670.462 ±\pm 0.077+73.0%
AURKB0.3580.463 ±\pm 0.046+29.3%
CTSK0.3190.336 ±\pm 0.026+5.3%
EGFR0.3680.457 ±\pm 0.051+24.2%
HDAC10.3420.453 ±\pm 0.075+32.5%
MTOR0.3510.410 ±\pm 0.063+16.8%
PIK3CA0.2690.370 ±\pm 0.036+37.5%
SMAD30.6400.803 ±\pm 0.188+25.5%
TP530.4300.590 ±\pm 0.235+37.2%
Mean0.3680.480+30.4%

Q4: Curriculum ablation

At a matched training budget, curriculum learning achieves the best mean hit rate on four of five targets and the best mean docking score on three of five targets. We compared it with two Stage 1-only baselines: No Curriculum preserves the original learning-rate schedule and optimizer resets, whereas Extended Stage 1 continues training at the Stage 1 learning rate without resetting AdamW. All variants use the same total number of optimizer updates and batch size. The gains are particularly pronounced for FA7 and BRAF. For FA7, curriculum increases the mean hit rate from 3.14-3.29% to 6.37%, approximately doubling performance. For BRAF, it increases the hit rate from 3.14-3.52% to 6.10%.

Docking Score – lower is better

TargetCurriculumNo CurriculumExtended Stage 1
PARP1−11.234 ±\pm 0.077−11.423 ±\pm 0.066−11.393 ±\pm 0.066
FA7−9.143 ±\pm 0.058−8.800 ±\pm 0.100−8.754 ±\pm 0.174
5HT1B−11.270 ±\pm 0.140−11.163 ±\pm 0.184−11.155 ±\pm 0.131
JAK2−10.433 ±\pm 0.109−10.425 ±\pm 0.072−10.482 ±\pm 0.152
BRAF−11.089 ±\pm 0.144−10.637 ±\pm 0.392−10.606 ±\pm 0.266

Hit Rate (%) – higher is better

TargetCurriculumNo CurriculumExtended Stage 1
PARP111.27 ±\pm 0.8413.67 ±\pm 0.7814.04 ±\pm 0.99
FA76.37 ±\pm 0.653.29 ±\pm 0.453.14 ±\pm 0.87
5HT1B13.58 ±\pm 1.9612.93 ±\pm 2.2613.15 ±\pm 3.14
JAK211.89 ±\pm 1.1211.49 ±\pm 2.0611.30 ±\pm 2.55
BRAF6.10 ±\pm 1.163.52 ±\pm 1.163.14 ±\pm 1.34

We apologize for the inconsistent reporting. The experiments underlying Tables 8–9 used the 50th → 99th → 99.8th percentile schedule for all five targets. The reference to the 90th percentile in Appendix D was incorrect, as were the FA7 and BRAF threshold values reported in Table 9. We have corrected these entries and clarified the actual schedule in the revised manuscript.

Q5. Scope of MoA claims and clarification of chance baselines

Thank you for highlighting these points. PhAME achieves the best or tied-best performance in 20 of the 21 combinations of metrics and representation spaces in Table 11 (in revised Table 17). This consistency across ECFP, CLOOME, and InfoAlign supports the robustness of the computational results and suggests that the gains are not specific to a single representation. It does not, however, directly establish the biological effects of generated molecules.

Previous work has demonstrated supervised MoA prediction from experimentally measured Cell Painting profiles extracted with CellProfiler (Way, Gregory P., et al). These approaches require morphological measurements for the compounds being assessed. Such measurements are unavailable for our novel generated molecules, and obtaining them would require new Cell Painting experiments. As detailed in GR1 of our General Response, the added Cell Painting evaluation uses measured profiles for recovered library compounds, providing phenotype-level evidence for that subset of generations. We have therefore revised the framing of these claims to reflect the scope of the evidence. Specifically, we have replaced the statements that the gains “reflect genuine MoA-aware chemistry” and demonstrate “a true understanding of structure-activity relationships” with statements that “the results suggest that the gains are not specific to a single embedding” and “the improvements suggest robustness to the choice of evaluation representation.” This distinguishes consistency across computational evaluations from experimental confirmation of biological activity. The Reviewer is also correct that GFlowNet’s mean Top-1 Cluster Accuracy exceeds the 14.3% chance level in all three representation spaces, reaching 21.2% in CLOOME, 17.1% in InfoAlign, and 17.0% in ECFP. Our original statement was incorrect when interpreted as a claim about Top-1 Cluster Accuracy alone, and we have corrected it.

The distinction supported by Table 11 (in revised Table 17) concerns performance across all evaluated metrics and representation spaces. For example, GFlowNet’s mean kNN Accuracy @1 in ECFP is 12.7%, below the corresponding chance baseline of 14.3%. The chance baseline is (1/7) for Top-1 Cluster Accuracy and kNN classification accuracy. We have therefore revised the statement to “Among the evaluated methods, PhAME is the only method whose reported mean exceeds the corresponding chance baseline for every evaluated metric in all three representation spaces.”

Way, Gregory P., et al. "Morphology and gene expression profiling provide complementary information for mapping cell state." Cell systems 13.11 (2022): 911-923.

Reviewer_qmbHOpenReview ↗

Response to the rebuttal

Thanks for the detailed response and the additional experiments. The new docking comparisons and the matched-budget curriculum ablation address my concerns about those experiments. The Cell Painting results also make the trade-off clearer. I’m still unsure how often joint guidance helps with both editing goals on the same request, so I’ll keep my overall rating.

I have one question about the existing Cell Painting results. Among requests that could be scored under both joint guidance and no guidance, how often did joint guidance improve phenotype specificity without reducing similarity to the starting molecule?

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

This paper proposes PhAME, a latent diffusion framework for phenotype-conditioned molecular generation and editing. The method combines a pretrained DGVAE, InfoAlign molecular representations, structure-based molecular pair construction, and compositional classifier-free guidance (c-CFG). Its main methodological contribution is the use of two independent guidance scales to control target property or phenotype matching and structural preservation relative to a seed molecule.

The authors evaluate PhAME on MOSES logP optimization, ZINC250k docking optimization, LINCS L1000 transcriptomics-conditioned generation, and JUMP Cell Painting phenotype-conditioned generation. The paper includes comparisons with existing methods and several ablation studies.

Critical strengths
  1. A meaningful application problem. Combining target cellular phenotypes with structural preservation around a seed molecule is relevant to molecular optimization. The paper presents a unified framework supporting several types of conditioning signals.

  2. Competitive experimental results. PhAME achieves the highest novel hit ratio on four out of five docking targets and substantially outperforms existing baselines in ECFP and InfoAlign evaluations on Cell Painting. The paper also provides ablations on CFG, structural encoders, molecular pairing strategies, and guidance scales.

Critical weaknesses

1. Limited methodological novelty and insufficient positioning against prior molecular editing work.

The method primarily integrates existing DGVAE, InfoAlign, and latent diffusion components, with two independently weighted CFG directions. The cellular morphology and transcriptomic inputs are obtained from existing feature extraction pipelines rather than a newly developed biological representation learning mechanism. While the model learns phenotype-conditioned molecular generation, it does not introduce a new model of cellular response mechanisms.

Furthermore, the Introduction provides an incomplete picture of prior molecular editing research. Optimizing molecular properties while preserving similarity to a seed molecule is already a well-established problem. Although Mol-CycleGAN and PURE are cited and compared, other relevant approaches, such as Modof and moleculeSTM, Prompt-MolOpt are not sufficiently discussed. The more directly related PhenoMoler is cited but not included in the corresponding empirical comparisons. Consequently, the paper does not clearly establish its incremental contribution over existing molecular optimization and phenotype-guided editing methods.

2. Insufficient direct evidence for the central dual-guidance mechanism and phenotype-guided editing capability.

The paper emphasizes jointly controlling phenotype matching and structural preservation. However, the main docking de novo experiments (Table 9) and quantitative Cell Painting experiments (Appendix F) both use (w_a=0). The authors do provide dual-guidance ablations on logP (Table 17), as well as CFG comparisons and guidance-scale sweeps on Cell Painting (Tables 13, 14, and 18). Nevertheless, the joint benefit of both guidance components has not been sufficiently demonstrated quantitatively in high-dimensional phenotype-guided editing of seed molecules.

Moreover, the Cell Painting evaluation relies primarily on structural similarity to reference compounds with known MoAs, rather than directly verifying the cellular responses of generated molecules. Although the drug-recovery experiments show successful examples, they involve a total of 100,000 generated candidates without systematic recovery-rate or sampling-efficiency analysis. These limitations weaken the evidence supporting the paper's central claims.

Questions for authors

Q1. What is the key methodological distinction between PhAME and existing structure-preserving molecular optimization methods such as Modof and moleculeSTM,Prompt-MolOpt? Could the authors provide comparisons on compatible tasks or explain why direct comparisons are unsuitable? Why was the cited PhenoMoler not included in the empirical evaluation?

Q2. Could the authors quantitatively compare full c-CFG, phenotype-only guidance, structure-only guidance, and standard CFG on a seed-based L1000 or Cell Painting editing task, reporting target matching, structural similarity, and validity? This would directly affect the assessment of the main methodological contribution.

Q3. Is there independent evidence of phenotype matching beyond structural similarity to known compounds? For the drug-recovery experiments in Appendix G, could the authors report recovery rates under different sampling budgets and clarify the overlap between the Cell Painting evaluation references and training data?

4 comments

W1. Clarifying Our Methodological Contributions and Positioning.

We thank the Reviewer for the detailed critique and agree that our positioning against prior molecular editing work was insufficient. We address both the novelty concern and the positioning gap below.

Guidance mechanism. We acknowledge that the compositional classifier-free guidance decomposition in Eq. 6 follows the multi-conditional CFG formulation introduced in vision: InstructPix2Pix (Brooks et al., CVPR 2023, Eq. 3) applies two independent scales to image and text conditions with a similar algebraic structure, and related formulations appear in Composable Diffusion (Liu et al., ECCV 2022) and Composer (Huang et al., ICML 2023). We have cited these works in the revision. Our contribution is not the CFG formula itself but its instantiation in a molecular-phenotype setting, which required solving four problems with no direct analogue in vision:

  1. Non-orthogonal guidance directions. The InfoAlign structural anchor zalignz_\text{align} is a phenotype-aware embedding, not a plain fingerprint. Because both the target condition cc and the anchor zalignz_\text{align} carry biological information (the former explicitly, the latter through InfoAlign's phenotype-structure pretraining), the two guidance directions Δc\Delta_c and Δa\Delta_a are not guaranteed to be orthogonal. The interaction ablation (Table 26) is therefore a genuine empirical finding: adding an explicit interaction term wcaΔcaw_{ca}\Delta_{ca} produces no measurable improvement up to wca=3w_{ca}=3, supporting the additive formulation, but this outcome could not have been assumed a priori.

  2. Two complementary distance-from-seed mechanisms. Partial noising at noise level t∗t^{\ast} determines where exploration begins in latent space; the structural-guidance scale waw_a controls how strongly the denoising trajectory is pulled back toward the seed. These two knobs are not redundant as t∗t^{\ast} sets exploration radius while waw_a modulates directional pull within that radius. In image editing, the seed is the noised image and these roles collapse; in molecular editing with a frozen VAE latent space, they are independently useful (Fig. 3).

  3. Validity under frozen decoding. Linear combinations of ϵ\epsilon-directions must decode to chemically valid molecules through a frozen DGVAE decoder that was never trained on such combinations. Table 24 shows that c-CFG(6,3) retains 71% validity vs. 42% at CFG(3) with matched target success, demonstrating that the two-scale decomposition navigates the validity landscape more effectively than single-scale guidance.

  4. Data construction. Beyond the guidance mechanism, the training-pair construction contributes independently: Table 23 shows nearest-neighbor pairing (27.2% Cluster Accuracy) substantially outperforms matched molecular pairs (16.5%) and random pairing (18.1%) on Cell Phenotype, a +65% relative gain over MMP. This result is specific to the phenotype-divergence threshold design and is not implied by the guidance formulation.

Positioning. We added a positioning table (Table 1) covering 13 methods across four paradigm groups (property editing, docking, phenotype de novo, phenotype editing) and two feature dimensions (conditioning signal, generation paradigm). We also expanded Section 2 with paragraphs on: (i) structure-preserving molecular optimization (Modof, Prompt-MolOpt, MoleculeSTM), (ii) compositional guidance prior art in generative modeling, and (iii) PhenoMoler as the closest prior work in phenotype-conditioned editing. See our response to Q1 for per-method distinctions.

W2. Quantitative evidence for dual-guidance editing.

We thank the Reviewer for highlighting the need for quantitative evidence of joint structural and phenotypic control during molecular editing. As detailed in GR1 and GR2 of our General Response, we have addressed this gap by adding a Cell Painting editing benchmark evaluated using measured profiles, expanding the transcriptomic editing evaluation, and providing per-seed drug-recovery statistics.

Q1. Methodological distinctions and comparison with prior work.

We thank the Reviewer for suggesting these relevant baselines and address them below by clarifying their methodological differences and compatibility with our evaluation tasks, alongside the newly added empirical comparison with our reproduction of PhenoMoler.

Modof (Chen et al., NMI 2021): it is scalar-property conditioning only (logP, QED). So, no possibility to have a high-dimensional phenotype input (Table 1); cannot participate in phenotype evaluations.

Prompt-MolOpt (Wu et al., NMI 2024): also scalar/binary properties, and requires one model per task. There is no inference-time trade-off control. It is comparable on scalar tasks, but has the same phenotype limitation as Modof.

MoleculeSTM (Liu et al., NMI 2023): text-conditioned and it requires a natural-language description of the target. This works for text-describable properties ("increase solubility") but not for a 978-D L1000 profile or a CellProfiler centroid, which are numerical vectors with no text equivalent. Also requires per-molecule gradient optimization for its λ\lambda trade-off, vs. PhAME's amortized wcw_c, waw_a. Comparable on text-describable scalar tasks, but infeasible on phenotype tasks.

PhenoMoler (Song and Liu, arXiv 2025): the closest prior work, performing L1000-conditioned molecular editing using a chemistry LLM. We did not include it in the original submission because its pretraining code was not publicly available at the time. For the revision, we have reconstructed the pretraining procedure using code fragments recovered from the repository’s commit history and added our reproduction of PhenoMoler to the MCF7 benchmark, evaluated using the same metrics as the other methods (Table 4 and Figure 5). PhAME achieves higher mean maximum Tanimoto similarity to known ligands than our PhenoMoler reproduction across 9 of 10 evaluated protein targets, together with higher QED and lower SA scores.

Q2. Quantitative Comparison of Guidance Variants in Phenotype-Guided Editing

We thank the Reviewer for this suggestion. As detailed in GR2 of our General Response, we now compare full c-CFG, phenotype-only guidance, structure-only guidance, and standard CFG on the seed-based L1000 task, reporting target relevance, seed similarity, and validity. The results show that the model benefits from combining phenotype and structural guidance.

Q3. Phenotype evidence, drug recovery, and training-data overlap.

We thank the Reviewer for these questions.

Phenotype evidence, recovery statistics, and sampling budget. As detailed in GR1 of our General Response, the new evaluation uses measured Cell Painting profiles, providing evidence beyond structural similarity, with phenotype specificity exceeding the random-library baseline across all twelve guidance settings. We also report per-seed drug-recovery statistics across three initial noise levels at a fixed budget of 100,000 candidates per seed and noise level.

Training-data overlap and exclusion protocol. Regarding training-data overlap, all 56 Cell Painting evaluation references (eight compounds from each of seven MoA classes) belong to ChEMBL2K. Following InfoAlign’s exclusion protocol, PhAME’s training pairs use only CP-JUMP molecules outside ChEMBL2K, excluding these references and their associated profiles. The drug-recovery experiments follow the same exclusion protocol: neither target drugs nor seed molecules appear in PhAME’s training pairs, and each target drug’s profile is excluded from its MoA conditioning centroid. We have clarified these exclusions in Appendices F and G.

Liu, Gang, et al. "Learning molecular representation in a cell." International Conference on Learning Representations. Vol. 2025. 2025.

Thank you for your response. After considering your rebuttal, I have decided to keep my score unchanged.

We thank the Reviewer for taking the time to read our rebuttal. Since the discussion period is still open, we wanted to briefly summarize the new evidence added in the revision, in case any of it is useful for your assessment:

  • Positioning (Q1). We added a positioning table and expanded Section 2. As PhenoMoler’s pretraining code was not released, we reconstructed its pretraining from the repository’s commit history and included it as a baseline on the MCF7 benchmark. PhAME reaches higher maximum Tanimoto similarity to known ligands on 9 of 10 targets, with higher QED and lower SA (Table 4, Figure 5).
  • Guidance variants (Q2). We now compare full c-CFG, phenotype-only, structure-only, and standard CFG on seed-based L1000 and Cell Painting editing, reporting target relevance, seed similarity, and validity (GR2; Appendices E, G).
  • Phenotypic evidence (Q3). We added a Cell Painting editing benchmark scored on measured CellProfiler profiles, per-seed drug-recovery statistics across sampling budgets, and a clarified exclusion protocol confirming no overlap between evaluation references and training pairs (Appendices F, G).

We would also flag one analysis that appeared late in a parallel thread and may be easy to miss: a per-request paired comparison on Cell Painting, scored only on seed–target requests recoverable under both settings. At guidance scales (5, 4), joint guidance improved phenotype specificity without reducing seed similarity in 50.9 ± 3.8% of requests, against 27.6 ± 3.1% for unguided-vs-unguided runs under identical scoring and 18.4 ± 0.8% for phenotype-only guidance. Reviewer qmbH, who raised this concern, indicated that the analysis largely addressed it and updated their rating accordingly. We will move this table into the main text.

More broadly, we see the contribution of this work less as a single benchmark gain than as evidence that phenotype and structure guidance can be composed at inference time, with a tunable trade-off between phenotypic specificity and proximity to a starting compound. This is a control that, to our knowledge, prior phenotype-conditioned generators do not expose, and one that we validate against measured Cell Painting profiles rather than structural proxies alone.

We fully respect that these additions may not change your overall assessment. If there is a specific concern that remains open, or an analysis that would be more convincing than the ones we ran, we would genuinely welcome knowing, and we would be glad to run it within the remaining discussion time. Either way, we are grateful for the review, which has made the paper stronger.

Reviewer_DauQOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary
  • The authors perform guided molecular editing via a generative model, using a simple but expressive formulation of guided or conditional generation. Specifically, they generalize CFG to consider two guidance directions: staying close to a "seed" molecule anchor and matching a target signature (some desired phenotypic measure or molecular property). They use partial noising to control exploration.
  • This formulation is quite general, and the authors apply it to many tasks (logP optimization, docking, transcriptomic conditioning, and Cell Painting–conditioned generation).
Critical strengths
  • The topic and application area is very important, both because of the practical utility of the task (drug design) and also the unique inherent challenges of the problem (discreteness, gigantic search space).
  • I like the simplicity, flexibility, and generality of the approach, which is easy to understand yet applicable across a diverse array of tasks.
  • The experimental results are quite extensive and appear to be state-of-the-art on multiple tasks (in some cases with considerable improvements).
Critical weaknesses
  1. A somewhat conceptual concern I have is the relation between structure and function, as well as proxies for them, with regards to the underlying task.

The space of chemicals is not only discrete, but the mapping from structure to phenotype could also be very discontinuous (meaning small structure changes causing large functional changes), no? Is there good evidence that similar structures would provide the desired results (biological relevance)? Indeed, the training pairs deliberately combine structural similarity with phenotypic dissimilarity, if I understood correctly.

E.g., the transcriptomic and cell painting tasks use structural proxies. So, concretely, given that the paper claims to do phenotype-controlled molecular optimization, it is not fully clear if such similarities indicate the requested phenotype has been met.

  1. There is not much in the way of methodological novelty in the guidance mechanism, as people have explored many similar decompositions/variations of CFG (e.g., in computer vision, for image/video/3d guidned generation). However, the data and training aspect (e.g., the curriculum for docking) also seems important and useful, and the authors show empirically that their model functions well after training.
  • Minor:
    • Regarding table 1, it is true that the authors' method has the best balance (and a controllable balance between sim and tgt), but I had one concern: Pure and DiGress seem better at each individual measure (also, at matched tgt, Digress has very slightly better sim, in table 7). It also was not clear to me why sim was important, how to intuitively understand whether a particular sim value was sufficiently high.
    • The paper may be hard to understand for non-specialists.
Questions for authors
  1. See my question in Weakness 1.

  2. One unclear aspect is how a user of this system could/should choose w_a and w_c. I am not sure how the values were arrived at for most of the tasks presented.

  3. On a related note, the authors utilize (in their words) "structural proximity" to the seed molecule, but you actually use the InfoAlign embedding, which seems a bit more functionally oriented in nature, to my understanding. Is this better than using a more structural approach? I did see H.2, but I still don't have a good understanding of the properties of InfoAlign vs other options, especially in the context of the other experiments.

  4. Is the guidance weight balance something we can optimize for a task, potentially? If we are using a fixed value per task, would it be better to train a specialized model for that case, instead of a flexible one for which we then fix the weights?

Minor

  • Regarding proxies like QED and SA, which again seem a bit removed from more direct measures (like target binding affinity or phenotypic impact). Is there a concern regarding identifying good candidates being limited by such proxies? Or are they fundamentally necessary to make the search space tractable? The more physics-based docking evaluation seems like a less limited form of assessment.
  • On the Cell Painting experiments, the authors seem to use w_a=0, meaning no structural proximity term in the CFG, and no dual guidance. Is this because there is no principled way to select a seed?
5 comments

W1/Q1. Concern about the structure-function relationship.

We thank the Reviewer for raising this important concern about the relationship between molecular structure and biological function. Below, we clarify the interpretation of our evaluations.

The role of structural similarity in the editing task. We agree that small structural changes can cause large biological changes. Such activity cliffs also occur in Cell Painting data (Sanchez et al.). We deliberately pair structurally similar molecules with different measured phenotypes because this is the task we aim to learn: changing the biological response while retaining a related molecular structure. Structural similarity is therefore an editing constraint, not a guarantee of phenotypic similarity.

Interpretation of the existing evaluation metrics (Appendices E and F). We also agree that structural similarity alone cannot establish whether generated molecules induce the requested phenotype. For transcriptomic generation, we follow GxVAEs and SmilesGEN in measuring Tanimoto similarity to known ligands of the target proteins. This enables comparison with prior methods but remains an indirect measure of biological relevance. For Cell Painting, we evaluate across ECFP, CLOOME, and InfoAlign to assess whether the results hold across different representations. ECFP measures structural similarity; CLOOME and InfoAlign use representations learned from cellular morphology, and from morphology and gene expression, respectively. These representations allow us to assess similarity to compounds with the target MoA using both structural and biologically informed features. Consistent performance across these spaces supports the robustness of the computational results, but does not directly confirm the biological effects of generated molecules.

Additional evidence beyond structural similarity. As detailed in GR1 of our General Response, the new Cell Painting evaluation assesses phenotype alignment using measured CellProfiler profiles for recovered library compounds, rather than inferring it from structural similarity. This provides evidence beyond structural proxies, while confirming the responses induced by novel generated compounds requires new transcriptomic or Cell Painting assays.

Sanchez, Maxime et al. “Large scale compound selection guided by cell painting reveals activity cliffs and functional relationships.” Communications biology vol. 9,1 225. 13 Jan. 2026

W2. Limited algorithmic novelty.

We agree with the Reviewer that CFG with two independent scales is not novel per se: related formulations appear in InstructPix2Pix (Brooks et al., CVPR 2023), Composable Diffusion (Liu et al., ECCV 2022), and Composer (Huang et al., ICML 2023), now explicitly cited in the revision. Our contribution lies in porting this formulation to the molecular-phenotype domain, where the non-trivial aspects are the phenotype-aware InfoAlign anchor (whose non-orthogonality to the target direction makes the interaction ablation in Table 26 a genuine finding), the complementary roles of partial-noising timestep t∗t^* and structural-guidance scale waw_a, and the maintenance of chemical validity through a frozen DGVAE decoder. Beyond the guidance mechanism, we highlight the training-pair construction as an independent contribution: nearest-neighbor pairing with a phenotypic-divergence threshold yields 27.2% Cluster Accuracy vs. 16.5% for matched molecular pairs on the Cell Phenotype benchmark (Table 23), a +65% relative improvement that is specific to our data-curation design.

Q2. How should a user choose wcw_c and waw_a?

For the benchmarks these are tuned hyperparameters: logP (6, 3), docking 2.0–5.0 per target with wa=0w_a=0 for de novo and 1.0–4.0 for editing, transcriptomic (1, 1), cell phenotype (34, 0) de novo and (14, 1) for editing. The sweeps are in Table 8 and Figure 3 (logP), Tables 19 and 20 (cell phenotype) and Table 12 (docking). The rule is the same throughout: raise wcw_c until the target metric stops improving, since past that point validity drops without a gain (wc=34w_c=34 is the peak for the phenotype task, wc=40w_c=40 is worse), then set waw_a by how much of the seed must be kept. Where the seed carries no information about the target, wa=0w_a=0 is correct; in the de novo phenotype benchmark the anchor is sampled at random, so any wa>0w_a>0 hurts.

The more important point is that neither value needs tuning in use. Both act at inference, so one trained model produces proposals at several settings and the chemist picks. This matters because a lead compound is already good at properties the model does not condition on, and those are unlikely to survive a large structural change; re-establishing them costs wet-lab work. How close the proposals must stay therefore depends on the campaign, not on us. Section 5.1 of the revision now says this, and Figure 9 shows one seed across both scales, where the substructure shared with the seed shrinks as waw_a falls.

Q3. InfoAlign vs. a structural measure for anchoring

We agree with the Reviewer, and our wording was loose. The revision drops "structural proximity": the abstract now says the model stays close to a known hit, the introduction says proximity is enforced in an InfoAlign embedding that reflects cellular response, and Section 4.1 states that Tanimoto/ECFP is used for evaluation only.

The three options compared in Appendix H.2 are: ECFP, which describes a molecule by the small connected subgraphs around each atom and scores two molecules by the fraction they share, the standard measure in medicinal chemistry, purely structural, blind to whether a change matters biologically; the DGVAE latent, which encodes what the decoder needs to rebuild the molecule; and InfoAlign, trained with an information bottleneck to keep the structural information that predicts cellular response.

In Figure 14, InfoAlign gives the best target success for a marginal loss in similarity, DGVAE preserves similarity but reaches lower target success, and ECFP is weakest on both axes. Note that Sim is itself an ECFP similarity, so the comparison is set up in the ECFP anchor's favour and it still performs worst. Appendix H.2 now states this.

Q4. Task-specific guidance optimization.

Thank you for this question. Flexibility is a central design goal of PhAME. Adjustable inference-time guidance allows the model to match or approach the individual strengths of specialized methods. The main benefit is that end users, such as chemists, can steer generation by adjusting two inference-time guidance scales, without retraining the model or requiring extensive machine-learning expertise. This allows users to adjust the trade-off according to their goals and assess whether the generated outcomes meet their expectations. Yes, the guidance balance can be optimized for a specific task, either manually or automatically. Given a task-specific objective based on NTS, QED, SA, a weighted combination of metrics, or another oracle score, the guidance scales wcw_c, waw_a and the noise level could be tuned using standard hyperparameter optimization methods, such as Bayesian optimization or grid search.

A specialized model may be suitable when only one fixed objective matters. PhAME’s advantage is that users can reuse the same trained model for different objectives by choosing appropriate guidance scales for each task, without retraining. Whether a specialized model would perform better on a particular task would require an empirical comparison.

WMinor1. Balancing target and similarity compared with baselines.

We thank the Reviewer for recognizing PhAME’s ability to balance these objectives.

  • PURE achieves high similarity but only 45% target success. The task requires both reaching the requested property class and remaining close to the seed; PURE misses the first objective for most samples.
  • As the Reviewer correctly observes, Table 8 in Appendix C shows that PhAME reaches target success of 1.00 at wcw_c=25, waw_a=0, with similarity 0.12, compared with 0.13 for DiGress - a difference within standard deviation. We do not claim superiority at this operating point. Our contribution is the ability to adjust the trade-off within a single trained model. At the default configuration, PhAME achieves a higher combined NTS score - defined as novelty × target success × similarity - of 0.30, compared with 0.26 for PURE and 0.13 for DiGress.
  • Similarity measures whether property editing preserves a structural relationship to the seed. The appropriate level depends on the user’s objective: larger waw_a encourages more conservative edits, whereas smaller waw_a allows broader exploration around the seed molecule. We do not assume a universally sufficient similarity threshold.
  • The new figure in Appendix C shows molecules generated from the same seed under different guidance scales. At wa=0w_a = 0, the shared structure is largely limited to a benzene ring. At larger waw_a, several candidates retain a larger portion of the starting structure while replacing its sulfonamide-containing side group, consistent with the contribution of this group to lower logP (Fujita et al., 1964). This example also illustrates that similarity values around 0.4 can correspond to preservation of a recognizable part of the seed structure.

Fujita, Toshio., Iwasa, Junkichi., & Hansch, Corwin. (1964). A New Substituent Constant, π, Derived from Partition Coefficients. Journal of the American Chemical Society

WMinor2. Readability for non-specialists

We have added plain-language sentences where each term first appears:

  • Section 4.1: two sentences on what ECFP and Tanimoto similarity compute. Molecules are scored by the fraction of small connected subgraphs they share; the measure is purely structural and says nothing about biological activity.
  • Start of Section 5: two sentences on why proximity to the seed is an objective. A seed has already been optimized for properties that are not modeled, and those are unlikely to carry over to a structurally distant molecule.
  • Section 5.1 and Section 5.2: definitions of logP, docking score, QED, and SA.
  • Section 5.4: one sentence on what a Cell Painting profile is and what conditioning on one asks for.

We are glad to add more if the Reviewer can point to a passage that is still opaque.

QMinor1: Role of QED and SA as proxies.

Thank you for raising this distinction. We agree that QED and SA cannot establish target binding or phenotypic efficacy. They assess complementary properties: QED estimates drug-likeness, while SA estimates synthetic accessibility. In the docking benchmark, QED and SA serve as evaluation filters following the established protocol, while the objective is the docking score. We acknowledge that fixed filters can exclude biologically valuable candidates, so our docking results should be interpreted within these benchmark constraints. However, PhAME’s generation procedure does not depend on these filters. In controlled logP editing, transcriptomic-guided generation, and Cell Painting-guided generation, QED and SA are reported as molecular-quality diagnostics alongside task-specific performance measures and do not determine task success. For context, Bickerton et al. (2012), who introduced QED, reported a mean score of 0.67 for compounds judged attractive starting points by medicinal chemists. PhAME’s mean QED (higher is better) is from 0.674 to 0.74 across the main configurations in controlled logP editing, transcriptomic-guided generation, and Cell Painting-guided generation.
 Ertl and Schuffenhauer (2009) describe SA scores above 6 as indicating difficult synthesis. For PhAME’s main configurations in these three tasks, mean SA scores (lower is better) range from 2.94 to 3.23. PhAME achieves favorable scores on both metrics relative to these reference values.

Bickerton, G. R., Paolini, G. V., Besnard, J., Muresan, S., & Hopkins, A. L. (2012). Quantifying the chemical beauty of drugs. Nature Chemistry, 4(2), 90–98. Ertl, P., & Schuffenhauer, A. (2009). Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions. Journal of Cheminformatics, 1, 8.

QMinor2: Seed selection in Cell Painting experiments.

The choice of waw_a in the Cell Painting experiments primarily reflects the inference setting, rather than a difficulty in selecting a seed. PhAME supports two inference modes. In de novo mode, the input specifies only the desired phenotypic effect, and the goal is to generate new molecules without a seed. Structural guidance is therefore disabled (waw_a=0), and generation relies solely on property guidance. In editing mode, the goal is to modify an existing molecule, which serves as the seed. PhAME starts from a partially noised latent representation of that molecule and uses structural guidance (wa>0w_a>0) to encourage similarity to it while guiding generation toward the desired effect. As noted above, this further highlights PhAME’s flexibility: a single parameter waw_a controls whether structural guidance is applied alongside effect guidance. As detailed in GR2 of our General Response, we added a seed-based Cell Painting editing benchmark that jointly evaluates phenotype alignment, seed similarity, and validity (Appendix G.2).

Comments on the paper

4 comments

Dear Area Chair and Reviewers,

We sincerely thank the reviewers for the time and care they devoted to our paper and for their constructive feedback. Their comments have helped us improve the quality and clarity of the manuscript. We address all comments in our detailed responses. Two concerns recur across the reviews, and we discuss them together below. Reviewer-specific questions are addressed in the individual replies. Table and figure numbers refer to the revised manuscript.

GR1. Does the evidence show that generated molecules achieve the requested phenotype?

Our transcriptomic task evaluates target relevance using Tanimoto similarity to known ligands. Our Cell Painting task evaluates similarity to reference compounds with the target mode of action (MoA), using ECFP fingerprints and biologically informed representations such as CLOOME and InfoAlign. We agree that structural similarity should serve as an editing constraint rather than direct evidence of phenotypic alignment. To address this concern, we added two complementary evaluations in Appendix G:

  • Editing evaluated using measured Cell Painting profiles (Appendix G.2, Figure 12). We use 600 seed–target pairs spanning 379 targets, drawn from ChEMBL2K. When a generated molecule exactly matches a library compound, we evaluate it using that compound's measured CellProfiler profile. This allows us to assess phenotypic alignment using existing experimental measurements for the recovered compounds. Phenotype specificity measures similarity to the target profile relative to other profiles. Mean specificity exceeds the random-library baseline (−0.0018 ±\pm 0.0014) in all twelve guidance settings, ranging from 0.0175 to 0.0365, with the highest value at waw_a=2, wcw_c=20. For comparison, library compounds sharing the target's MoA achieve a specificity of 0.0277 ±\pm 0.0053.

  • Drug recovery (Appendix G.1, Table 21). For each of three drugs: aspirin, clofibrate, and dyphylline, we start from six structural analogues and sample 100,000 candidates per seed and starting noise level. The target drug is recovered from 15 of 18 seeds at T=400T=400, 14 of 18 at T=500T=500, and 6 of 18 at T=700T=700. Recovery depends strongly on the seed, as detailed in Table 21. Under the same setup, a GFlowNet baseline did not recover any of the target drugs from the tested seeds from Figures 16, 17, 18, including when using alternative reward functions designed to favor their recovery.

A limitation of this evaluation is that measured profiles exist only for compounds in the library, so we can score only generated molecules that match a library compound. Direct confirmation that newly generated compounds induce the intended cellular responses requires laboratory experiments, such as transcriptomic or Cell Painting assays. We also welcome suggestions for alternative or complementary evaluations that could strengthen the evidence for phenotypic alignment.

GR2. Is joint structural and phenotypic guidance quantified in seed-based biological editing?

Transcriptomic-guided editing. We expanded the quantitative evaluation of transcriptomic-guided editing. Appendix E reports the full results, including guidance-scale sweeps from 0 to 3 for both c-CFG and standard CFG (Figure 11, Table 16). The key results are summarized below.

Guidance methodScaleMean similarity to seed moleculeMaximum similarity to known bindersValidity (%)
CFG(1)0.13 ±\pm 0.000.44 ±\pm 0.0369.3 ±\pm 5.2
Structure-only(0,1)0.14 ±\pm 0.010.46 ±\pm 0.0367.9 ±\pm 6.0
Phenotype-only(1,0)0.11 ±\pm 0.000.41 ±\pm 0.0164.5 ±\pm 6.0
c-CFG(1,1)0.13 ±\pm 0.000.48 ±\pm 0.0369.9 ±\pm 5.0

Among the settings shown, c-CFG with guidance scales wcw_c=1 and waw_a=1 achieves the highest maximum similarity to known binders and the highest validity. Phenotype-only guidance achieves a maximum binder similarity of 0.41, while structure-only guidance (wcw_c=0, waw_a=1) achieves the highest mean similarity to the seed. These results show that combining phenotype and structural guidance improves target relevance without compromising molecular validity or structural similarity to the seed.

Cell Painting-guided editing. The new benchmark evaluates editing from seed molecules under joint structural and phenotypic guidance.

Guidance methodScaleMean similarity to seed moleculePhenotype specificityValidity (%)
No guidance(0,0)0.1787 ±\pm 0.00010.0214 ±\pm 0.002485.84 ±\pm 0.01
Structure-only(0,4)0.2348 ±\pm 0.00000.0257 ±\pm 0.005984.04 ±\pm 0.01
Phenotype-only(20,0)0.1467 ±\pm 0.00010.0315 ±\pm 0.005976.94 ±\pm 0.01
c-CFG(20,2)0.1655 ±\pm 0.00000.0365 ±\pm 0.002976.69 ±\pm 0.00
c-CFG(20,4)0.1835 ±\pm 0.00010.0346 ±\pm 0.006175.73 ±\pm 0.01

This setting allows us to quantify both objectives within the same editing task. At a fixed phenotype guidance scale of wcw_c=20, increasing structural guidance from waw_a=0 to waw_a=4 raises mean Tanimoto similarity to the seed monotonically from 0.1467 through 0.1655 to 0.1835. At a fixed structural guidance scale of waw_a=4, increasing phenotype guidance from wcw_c=0 to wcw_c=20 raises mean phenotype specificity from 0.0257 to 0.0346. The highest specificity in the sweep, 0.0365 ±\pm 0.0029, occurs at waw_a=2, while waw_a=4 trades a small amount of specificity for higher seed similarity. Stronger phenotype guidance also consistently reduces seed similarity and molecular validity. c-CFG(20,4) improves on the unguided model in both seed similarity (0.1835 vs 0.1787) and specificity (0.0346 vs 0.0214), at a cost of about 10 points of validity. Compared with the unguided model, c-CFG at (wcw_c, waw_a)=(20,4) achieves higher mean seed similarity (0.1835 vs. 0.1787) and phenotype specificity (0.0346 vs. 0.0214), at the cost of approximately 10 percentage points in molecular validity. These results quantify the trade-offs among structural preservation, phenotypic alignment, and validity, using existing experimental measurements for recovered library compounds.

Summary of revisions

UpdateLocation
Added a Cell Painting editing benchmark evaluated using measured profilesAppendix G.2; Figure 12
Added drug-recovery statisticsAppendix G.1; Table 21
Expanded the L1000 editing evaluationAppendix E; Figure 11; Tables 15–16
Expanded the comparison with prior work and added the PhenoMoler baselineTable 1; Section 2; Table 4; Figure 5
Separated docking results into oracle-free and oracle-guided settings; added GEAM and f-RAG; added a curriculum ablation with a matched budgetTable 3; Section 5.2; Table 13
Added plain-language explanations and improved wordingAbstract; Introduction; Sections 4.1, 5, 5.1, 5.2, and 5.4
Area_Chair_xfKCOpenReview ↗

Requesting reviewer engagement

Dear Reviewers,

Thank you for your reviewing service. The authors have provided detailed rebuttals to the comments. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your recommendations. Thank you for your continued engagement.

Thank you,

AC

Dear Area Chair and Reviewers,

We thank the Area Chair for handling our submission and the Reviewers for their time and constructive feedback. The revised manuscript incorporates the experiments and analyses added during the discussion. Changes are marked in blue.

  • Related work and baselines. We expanded the discussion of molecular editing and compositional guidance and added a positioning table, and included our reproduction of PhenoMoler. (Sec. 2: Table 1; Sec. 5.3: Table 4, Figure 5).
  • Paired analysis of joint guidance. We added a per-request comparison of joint, phenotype-only, and structure-only guidance against unguided generation (Appendix G.2: Table 22).
  • Cell Painting editing. We added a seed-based editing benchmark for Cell Painting using measured CellProfiler profiles for recovered library compounds. (Appendix G.2: Figure 12).
  • Transcriptomic editing. We added a retrieval-only baseline and guidance-scale sweeps comparing joint, phenotype-only, structure-only, and standard classifier-free guidance on L1000 (Appendix E: Figure 11, Tables 15–16).
  • Docking comparisons and controls. We separated oracle-free and oracle-guided comparisons, added GEAM and f-RAG, and compared de novo generation with seed-based editing. We also added a matched-budget curriculum ablation and corrected the reported percentile schedule (Sec 5.2: Table 3; Appendix D: Tables 9, 12, and 13).
  • Drug recovery. We added recovery statistics for aspirin, clofibrate, and dyphylline. We also clarified the training-data exclusions (Appendix F; Appendix G.1: Table 21).
  • Visual example. We also added a visual example illustrating how the two guidance scales affect edits to the same seed molecule (Appendix C: Figure 9).

We believe these additions address the concerns raised during the discussion and strengthen the paper by clarifying how our approach relates to prior work, evaluating joint guidance more thoroughly, and introducing additional baselines and controls across our experiments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.