acceptodds
Under review as a conference paper at ICLR 2027

PhenTTP: Phenotype-Driven Molecular Generation with Transferable Target Priors

Abstract

Drug discovery ultimately seeks molecules that elicit desired biological responses. Phenotype-driven generation pursues this functional objective directly from cellular responses without requiring a predefined target. However, the inverse mapping from downstream cellular response to molecular structure is underdetermined. To reduce this ambiguity, we propose PhenTTP, a phenotype-driven molecular generation framework with transferable target priors. PhenTTP uses a state-transition phenotype encoder and proceeds in three stages: it aligns phenotype-derived representations with target-conditioned priors, learns phenotype-conditioned generation from phenotype-molecule pairs, and transfers target-conditioned behavior through dual-route distillation, supporting both phenotype-only and phenotype-plus-target generation. Across chemical, scaffold, cell-line, and target out-of-distribution settings, the phenotype-only route lowers Vina scores by 0.13-0.54 kcal/mol over the best phenotype-only baseline despite receiving no targets at inference; when a target is available, the target-informed route further increases Boltz-2 binder probability by 6.6-18.3 percentage points over the same baselines while maintaining phenotype correlation similar to that of the phenotype-only route and achieving high molecular validity, uniqueness, and novelty. Together, these results show that PhenTTP improves target compatibility by transferring target-conditioned molecular priors while preserving phenotype fidelity, establishing a framework that links cellular functional objectives to molecular-level constraints in generative drug discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Transcriptomics, Molecular generation, Drug discovery, Multimodal alignment

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_C5NBOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies phenotype-conditioned molecular generation when the molecular target is not necessarily known at inference time. PhenTTP transfers knowledge from a pretrained target-conditioned generator to phenotype-driven generation and supports both P-only and P+T inference from a shared model. Training combines phenotype–target alignment, phenotype-conditioned generation from phenotype–molecule pairs, and dual-route distillation. The method is evaluated on L1000 and Tahoe under four OOD settings using phenotype consistency, molecular quality, and target compatibility metrics.

Critical strengths
  1. The problem formulation is interesting and practically relevant. Treating target-conditioned knowledge as a prior for phenotype-only generation is a natural way to handle the case where the molecular target is unavailable at inference. The shared P-only/P+T model is also a useful feature when target information is available.

  2. The P-only results are encouraging. PhenTTP performs well on the phenotype metrics and outperforms the phenotype-only baselines on Vina across all four L1000 splits. Table 3 also shows that removing KD hurts the target-side results much more than removing P–T alignment.

Critical weaknesses
  1. The role of P+T is less clear. Adding the target improves PbindP_{\rm bind} relative to P-only, but P+T also restores the native target input of the pretrained generator. In Appendix Table 8, the original DrugGPT T-only model is often better than P+T on affinity, and its phenotype scores are also frequently better. I therefore do not see clearly what phenotype conditioning adds in the P+T setting.

  2. There are several design choices that are not well justified by the current ablations. For example, zPz_P is split into zCz_C and rPr_P, but there is no comparison with using only zCz_C, only rPr_P, or no split. Also, KD-only is already comparable to the full model in Table 3, and joint training is competitive with staged training. Although the Tahoe ablations show benefits from alignment and staged training in some settings, their additional value on L1000 remains unclear.

Questions for authors
  1. Why is P+T sometimes worse than the original T-only model, even on phenotype metrics? Do the authors think this comes from a phenotype–target trade-off, interference from the phenotype pathway, or limited joint supervision?

  2. Do zCz_C and rPr_P provide distinct contributions in practice? I would like to see an ablation of the two pathways to understand whether the split is necessary.

  3. How is PhenTTP implemented on ProtoBind-Diff? The DrugGPT implementation is described in much more detail. Are the intervention points and trainable modules the same across the two backbones?

  4. Does P+T perform worse than T-only for most targets, or is the overall gap mainly driven by a few targets? I am also curious whether this difference is related to how much target-annotated training data is available for different targets.

  5. Have the authors considered using additional target–molecule data during adaptation? I wonder whether the limited target-side supervision contributes to the performance gap between T-only and P+T.

6 comments
AuthorsOpenReview ↗

Response to C5NB. CW 1 and Q1: Phenotype control and the role of P+T

1. Our task is phenotype-driven molecular generation.

Our goal is to generate molecules from a requested cellular response, without requiring a known target. Native DrugGPT requires a target and does not provide a trained phenotype-conditioning interface.

We tested direct phenotype input to unmodified DrugGPT by supplying numeric expression values through its native protein prompt slot, without a target or further training. Under its native context limit, 153–188 of the 978 gene values are retained after reserving generation tokens.

OOD splitDrugGPT (P text): Valid (%)DrugGPT (P text): PRNet rΔr_\DeltaPhenTTP (P): Valid (%)PhenTTP (P): PRNet rΔr_\Delta
Chemical0.700.00495.400.578
Scaffold0.800.00595.450.570
Cell-line0.900.00595.150.552
Target0.550.00394.750.549

Each method has 2,000 requests per split. Only 59 of 8,000 DrugGPT phenotype-text outputs are valid (0.740.74\\%), all representing CN. Thus, its unmodified native interface does not provide effective phenotype-driven generation.

2. P+T adds target guidance while retaining phenotype control.

P+T is an additional function built on our phenotype-driven framework: users can specify both the desired cellular response and a target. Its value is that the phenotype continues to influence generation when the target is fixed.

We test this using the original P+T checkpoint. For 2,000 requests per split, we keep the target unchanged and replace the full phenotype input with a frozen dissimilar donor. Both outputs are evaluated against the original requested phenotype in its original cell, dose and time context.

OOD splitPRNet rΔr_\DeltaCRISP rΔr_\DeltaMAP rΔr_\Delta
Chemical0.557 / 0.5210.552 / 0.5210.531 / 0.499
Scaffold0.555 / 0.5190.538 / 0.4930.522 / 0.492
Cell-line0.539 / 0.5100.499 / 0.4730.499 / 0.470
Target0.543 / 0.5340.525 / 0.5150.521 / 0.513

Entries are matched/replaced phenotype input. The matched input gives higher correlations in all twelve evaluator-by-split point-estimate comparisons. All requests are retained, including the original invalid-output penalties. Both diagnostic experiments above use the original forward-KL configuration.

These results show that phenotype conditioning improves fulfillment of the requested response at a fixed target. T-only does not provide this control: its conditional distribution is unchanged across different phenotype requests. Its aggregate phenotype scores can nevertheless be high because target identity is associated with cellular responses.

3. Joint supervision is limited.

The main data limitation is the availability of linked phenotype–target–molecule supervision. In the constructed L1000 dataset, only 2,127 of 17,766 molecules have usable target-sequence annotations. Phenotype–molecule training can use the full pair pool, whereas joint conditioning relies on this much smaller annotated subset.

Sparse joint coverage may leave phenotype-dependent adaptation insufficiently constrained by target supervision. The resulting updates can perturb the pretrained target-conditioned prior, potentially reducing both phenotype agreement and affinity relative to T-only. This is a plausible contributor to the original P+T performance gap.

There is also a trade-off between phenotype agreement and target affinity when adding target guidance. Improving this balance with broader joint supervision remains an important direction for further work. The revised model retains rPr_P with adjustable strength, rather than removing phenotype conditioning from the provided-target route.

4. DrugGPT is a target-only reference.

DrugGPT is included as a reference for target performance, not a baseline for our phenotype-driven task. The direct task baselines are phenotype-conditioned generators. DrugGPT shows what the pretrained target expert achieves when a target is supplied.

The revised model uses controlled residual adaptation and reverse KL. Its P+T affinity results are now competitive with the DrugGPT reference:

OOD splitDrugGPT (T): Vina ↓PhenTTP (P+T): Vina ↓DrugGPT (T): PbindP_{\mathrm{bind}} ↑PhenTTP (P+T): PbindP_{\mathrm{bind}} ↑
Chemical-8.57-8.290.6110.593
Scaffold-8.56-8.430.6130.605
Cell-line-8.59-8.410.5980.603
Target-7.38-8.540.5230.500

These updated results are reported in Appendix Table 9. PhenTTP supports a broader discovery setting: generation from a desired cellular response when the target is unknown, and additional target guidance when one is available. The target-only reference measures how much target performance is retained within this phenotype-driven framework.

AuthorsOpenReview ↗

Response to C5NB. CW 2 and Q2: Pathway ablations and training strategy

1. Pathway ablations and model revision.

We compared the original model with zCz_C-only, rPr_P-only, and Shared variants. The zCz_C-only variant disables the residual pathway; rPr_P-only suppresses the phenotype-derived prompt on the P-only route; Shared replaces the separate projection heads with one shared projection.

OOD splitRouteOriginal FullzCz_C-onlyrPr_P-onlyShared
ChemicalP-only0.5780.5790.5690.572
ScaffoldP-only0.5700.5750.5670.571
Cell-lineP-only0.5520.5570.5500.550
TargetP-only0.5490.5550.5500.555
ChemicalP+T0.5570.5840.5590.562
ScaffoldP+T0.5550.5840.5590.559
Cell-lineP+T0.5390.5630.5320.538
TargetP+T0.5430.5660.5400.543

Entries are PRNet rΔr_\Delta from the original forward-KL configuration. In the zCz_C-only variant, the provided-target route uses TT alone because zTz_T replaces zCz_C and the residual pathway is disabled.

The zCz_C-only variant matches or exceeds Full in PRNet across all four OOD splits, while rPr_P-only performs below zCz_C-only throughout. Shared is also competitive with Full. These results show little additional PRNet benefit from the original residual pathway and motivated us to make its contribution adjustable through a learned scale α\alpha. Specifically, we added a global scale

α=σ(α~)∈(0,1)\alpha=\sigma(\tilde{\alpha})\in(0,1)

to the residual and conditional LoRA updates. Both conditioning interfaces are retained, while their residual contribution is adjusted during training. The scale is shared across both routes and initialized at 0.50.5.

We have adopted this residual-scaled model with reverse KL and updated the main results and component ablations accordingly in our new manuscript. The KL comparison is reported in our response to R1 Q3.

2. When alignment and staged training help.

We agree that KD-only and joint training remain competitive on L1000. To better clarify the contribution of alignment and staged training, we compare these configurations with our full model on both L1000 and Tahoe under P-only inference.

MethodL1000 (4-split avg.): PRNet rΔr_\Delta ↑L1000: VUN ↑Tahoe Cell-line OOD: PRNet rΔr_\Delta ↑Tahoe Cell-line OOD: VUN ↑
KD-only0.58294.4%0.22552.8%
Joint0.58194.2%0.38375.4%
PhenTTP (Full)0.58395.2%0.47293.3%

KD-only retains distillation but removes phenotype–target alignment, while Joint optimizes the training objectives simultaneously. The L1000 results are macro-averaged across four OOD splits.

On L1000, all three configurations achieve comparable performance, suggesting that the simpler training strategies can be sufficient for this setting. However, the differences become more pronounced on the more challenging Tahoe Cell-line OOD split. Our full model achieves substantially higher phenotype agreement (0.472 vs. 0.225/0.383) and VUN (93.3% vs. 52.8%/75.4%) than KD-only and Joint, respectively.

These results suggest that alignment and staged training become more beneficial under challenging distribution shifts, while their additional gains on L1000 are relatively modest (Tahoe: approximately 60,000 gene features vs. L1000: 978 landmark genes). The detailed results are provided in Table 3 and Appendix Tables 8 and 11 of the revised manuscript.

AuthorsOpenReview ↗

Response to C5NB. Q3 and Q4

Q3: Implementation on ProtoBind-Diff

We thank the reviewer for pointing out this missing implementation detail.

Shared design. In the full α\alpha-gated formulation, both backbones use a phenotype encoder and a split projection producing zCz_C and rPr_P. P-only supplies zCz_C to the native target-conditioning interface; P+T supplies zTz_T, while retaining the same rPr_P branch. The trainable components are the phenotype encoder, split projection, gated residual modules, conditional LoRA in the last two decoder layers, and a global scalar α=σ(α~)\alpha=\sigma(\tilde{\alpha}). The pretrained backbone and native target pathway remain frozen. Initialized to 0.50.5, α\alpha is shared across layers and routes and scales both residual and LoRA updates, without scaling the native target condition.

DrugGPT adapter. The native interface is the autoregressive protein-prompt slot: zCz_C supplies four soft prompt tokens in the model's token-embedding space, or an explicit protein prompt is used in P+T mode. The pooled rPr_P conditions an added gated residual module at the adapter slot within each Transformer block. A query mask restricts its updates to states predicting molecular tokens, preserving the native protein-prompt computation.

ProtoBind-Diff adapter. The native interface is the protein-conditioning memory of the masked-diffusion decoder. Four 1,280-dimensional zCz_C tokens replace the ESM-2 protein embeddings in P-only mode; P+T uses the native protein embeddings. Both pass through the frozen sequence projection into the existing cross-attention memory, together with the diffusion-noise token. In the full formulation, the rPr_P-conditioned adapter adds its gated residual and applicable LoRA update after each denoising decoder block, using the same shared α\alpha rule. Thus, the module roles are shared, while the native conditioning interface and residual insertion points are backbone-specific.

Q4: Target-wise comparison and training coverage

We examined paired test conditions against DrugGPT, with both models receiving the same target. We report the distribution of wins and its relationship with target-annotated training coverage below.

1. Distribution of performance differences.

For PRNet, we average scores within each target and count which model performs better. DrugGPT wins slightly more than half of the targets in Scaffold, Cell-line and Target OOD, while Combined wins slightly more than half in Chemical OOD. Thus, the target-wise win frequencies are close to split evenly.

OOD splitPRNet target wins: CombinedPRNet target wins: DrugGPTVina condition wins: CombinedVina condition wins: DrugGPTBoltz-2 condition wins: CombinedBoltz-2 condition wins: DrugGPT
Chemical60/11656/11645/10054/10046/10053/100
Scaffold55/11964/11947/10052/10043/10056/100
Cell-line246/501255/50151/10046/10053/10044/100
Target28/6032/6048/10050/10041/10057/100

Combined denotes the earlier model's provided-target route. PRNet uses all 2,000 phenotype requests per split. Affinity wins are counted over all 100 requested conditions per split; 99/99/97/98 conditions have jointly finite scores in Chemical/Scaffold/Cell-line/Target OOD, respectively. Lower Vina and higher Boltz-2 PbindP_{\mathrm{bind}} indicate a win. Ties and unscored requests award neither model a win and remain in the denominator.

These counts describe how frequently each model performs better. They do not quantify whether a few large target-specific differences dominate the aggregate gap. The affinity entries are condition-level comparisons, rather than target-win counts.

2. Relationship with target-annotated training coverage.

We relate each target's PRNet difference, defined as Combined minus DrugGPT, to its number of distinct target-annotated training molecules using Spearman correlation. The results are ρ=−0.053\rho=-0.053 (p=0.559p=0.559) for Chemical, ρ=−0.045\rho=-0.045 (p=0.607p=0.607) for Scaffold, and ρ=−0.066\rho=-0.066 (p=0.158p=0.158) for Cell-line OOD.

We find no clear monotonic association between target-specific training coverage and the PRNet difference in these splits. For Target OOD, every held-out target has zero target-specific training coverage, so the correlation is undefined.

Reviewer_YTJ1OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces PhenTTP, a phenotype-driven generative model for molecules. The model encodes the biological target and transcriptomic perturbation and uses them as conditions for LoRA adaptation of the DrugGPT model, which generates the molecular structure. The model works in two modes: the first combines targets and transcriptomic perturbations, and the second operates without the target when its structure is unavailable. During training, target and perturbation embeddings are first aligned, then phenotype-based generation is trained, and finally target-guided distillation is performed. The experiments confirm that PhenTTP can generate molecules with the correct predicted phenotype as well as molecules that dock well to their targets.

Critical strengths
  1. The proposed method combines expression profiles with protein target structures as conditional signal for generating molecules. The generator is a frozen backbone trained with a conditional LoRA, where the conditioning signal can be the combination of transcriptomic perturbation and target structure or just perturbation. This is the core contribution of this paper. The experiments show that the combined conditioning improves predicted binding affinity to the targets of interest.
  2. The proposed model performs strongly compared with other generators across different OOD splits. I appreciate that the Authors evaluated the model from different perspectives: chemical, scaffold, cell-line, and target.
Critical weaknesses
  1. The paper's clarity could be improved. Different terms are used in Figure 1 and in the main text. The phenotype encoder (Sections 3.2 and 3.3) is not shown in the figure, and neither is zPz_P. What exactly is zPz_P, and is it based on both pre- and post-perturbation, as suggested by Equation 3? In the main text, FF, EΔE_\Delta, and ee are not defined. The losses in Equations 6 and 7 should be explained in more detail or deferred to the appendix if they are not considered the main contribution. Currently, it is impossible to understand what these losses are and what their inputs are without checking the appendix. The code is not provided yet, and reproducing the model from the paper's description would be difficult.
  2. The model results are mixed but very interesting. Training on both perturbations and target structures appears to worsen performance on phenotype-based metrics while improving predicted binding affinity. As a consequence, it is still better to train a phenotype-only or target-only model to optimize specific tasks, and no positive transfer is observed. The ablation study confirms this observation further, showing that P-T alignment is unnecessary. No confidence intervals are provided, but the results are sometimes better with and sometimes without this step, so these differences are probably not statistically significant. Finally, the performance of DrugGPT (the model backbone) is presented only in the appendix, and this model outperforms the proposed PhenTTP in predicted binding affinity.
  3. The docking experiment presented in Figure 2 is unclear. The presentation could be improved by not repeating the reference compound in the second row on the left, which I assume shows the same pose in the same pocket. The right and wrong targets should be side by side for easier comparison. Beyond presentation, this figure shows only one example compound for one target, and it provides no quantitative comparison for the negative control experiment. Are the blue and red compounds the same? It is difficult to assess that because of the different rotations, and atom types are not annotated here. Adding a molecular structure in the corner would make this comparison clearer. The difference in docking scores between HTR1B and HTR3A could also be influenced by how these structures are prepared, and no information is provided on the protein preparation protocol or the average docking scores of active compounds for these targets.
Questions for authors
  1. Have you considered other learning strategies than multi-stage training? Maybe the benefits of the P-T alignment are lost because the model forgets what it learned. Would it make sense to train these objectives in parallel and verify if the alignments are unlearned in the following steps?
  2. What does it mean that “PhenTTP in these tables is the DrugGPT instantiation”? Do you mean that is the backbone used?
10 comments
AuthorsOpenReview ↗

Response to YTJ1. CW 1 and CW 3

CW 1: Figure 1, notation, losses, and reproducibility

We thank the reviewer for these clarity and reproducibility comments. We have revised Figure 1 and Sections 3.2–3.5 so that the encoder path, zPz_P, and the Stage 1–2 losses can be read without the appendix.

zPz_P is the output of the phenotype encoder EPE_P. It is a state-transition latent computed from both the unperturbed profile XpreX_{\mathrm{pre}} and the perturbed profile XpostX_{\mathrm{post}}, together with XΔ=Xpost−XpreX_\Delta=X_{\mathrm{post}}-X_{\mathrm{pre}} and dose–time covariates cc; it is not a post-treatment snapshot. In the revised manuscript, this is Eq. (2), after the bidirectional coupling of the two states in Eq. (1). The split projection in Eq. (3) then maps zPz_P to zCz_C and rPr_P. We define the previously undefined maps in the main text: EΔE_\Delta projects XΔX_\Delta, ee embeds the covariates, and FF is the fusion map. Figure 1b now labels zPz_P as the output of the transcriptome, cross-attention, and fusion path; Section 3.2 names that path EPE_P. The two projection outputs zCz_C and rPr_P remain as in the original figure.

We keep Eqs. (6) and (7) in the main text and now state their inputs and roles there. Equation (6) is the Stage 1 alignment loss on triples DP,T\mathcal{D}_{P,T}: cosine pulls pooled zCz_C toward the matched frozen zTz_T, NCE separates target identities, and the margin term pushes zCz_C away from a mismatched target T−T^-; residual adapters and α\alpha are not updated in this stage. Equation (7) is the Stage 2 generation loss on pairs DP\mathcal{D}_P: P-only molecular NLL, a contrast that holds the molecule fixed and prefers the paired phenotype over a context-matched negative P−P^-, and a moment term that matches batch statistics of zCz_C to native target encodings. Full expansions remain in Appendix A.3.

Code, checkpoints, splits, and training configs are at https://anonymous.4open.science/r/PhenTTP-85EA/.

CW3: Revised docking figure and quantitative controls

We thank the reviewer for these helpful suggestions. We agree that the original Figure 2 could be clearer. We have revised Figure 2 in the main text and added further docking examples, quantitative negative-control results, and methodological details in Appendix E. We address each point below.

1. Repeated reference compound.

Yes, the two reference panels showed the same SB-216641 pose in the same HTR1B pocket. We repeated it to facilitate comparison across rows, but agree that this was unnecessary. The revised figure now shows the reference only once.

2. Intended and control targets side by side.

We have placed the same P-only generated molecule docked to HTR1B and HTR3A in adjacent panels, explicitly labeled “Intended target” and “Control target.” The P+T result is shown separately for clarity.

3. Additional examples and quantitative negative controls.

We have added two additional cases, CA6/CHRNA7 and TOP2B/CA2, associated with boldine and ciprofloxacin, respectively (Appendix Figure 3). Each shows the reference compound once and the same generated molecule docked to both receptors.

For quantitative evaluation, we further tested 100 held-out conditions with distinct reference molecules and intended targets. Control targets were selected using negative direct-binding annotations (Ki/Kd≥10,μMK_i/K_d \geq 10\\,\mu\mathrm{M}). Among 97 complete pairs, the intended target was favored in 67/97 cases by Vina and 78/97 by Boltz-2. These results provide broader quantitative support while indicating that the docking examples illustrate predicted target preference rather than establishing general target selectivity.

4. Molecular identity, orientation, and atom types.

Yes, the blue and red compounds represent the same generated molecule, independently docked to different receptors. We agree that the original presentation made this difficult to assess. We have revised the visualization and added 2D molecular structures and heteroatom labels to make the comparison clearer.

5. Protein preparation and known-active compound scores.

We agree that receptor preparation and target-specific docking-score distributions may affect the comparison. We have now provided the protein and ligand preparation protocols in Appendix E, including the AlphaFold structures, pocket identification, and docking settings.

We additionally evaluated ten experimentally known active compounds per receptor using the same docking setup. Their average Vina scores and Boltz-2 binding probabilities are reported in Appendix Table 6, providing target-specific reference distributions. We also clarify the structural limitation of the monomeric HTR3A model and interpret the docking results accordingly.

AuthorsOpenReview ↗

Response to YTJ1. CW 2: Task, P+T control, and DrugGPT reference (1/2)

1. Our task is phenotype-driven molecular generation.

Our goal is to generate molecules from a requested cellular response without requiring a known target. Native DrugGPT requires a target and does not provide a trained phenotype-conditioning interface.

We tested direct phenotype input to unmodified DrugGPT by supplying numeric expression values through its native protein prompt slot, without a target or further training. Under its native context limit, 153–188 of the 978 gene values fit after reserving generation tokens.

OOD splitDrugGPT (P text): Valid (%)DrugGPT (P text): PRNet rΔr_\DeltaOriginal PhenTTP (P): Valid (%)Original PhenTTP (P): PRNet rΔr_\Delta
Chemical0.700.00495.400.578
Scaffold0.800.00595.450.570
Cell-line0.900.00595.150.552
Target0.550.00394.750.549

Each method has 2,000 requests per split. Only 59 of 8,000 DrugGPT phenotype-text outputs are valid (0.740.74\\%), all representing CN. Its unmodified native interface therefore does not provide effective phenotype-driven generation. Separately optimizing a target-only model does not address this task.

2. P+T adds target guidance while retaining phenotype control.

P+T is an additional function of our phenotype-driven framework: users can specify both the desired cellular response and a target. Its contribution is that phenotype continues to influence generation when the target is fixed.

We test this using the original P+T checkpoint. For 2,000 requests per split, we keep the target unchanged and replace the full phenotype input with a frozen dissimilar donor. Both outputs are evaluated against the original requested phenotype in its original cell, dose and time context.

OOD splitPRNet rΔr_\DeltaCRISP rΔr_\DeltaMAP rΔr_\Delta
Chemical0.557 / 0.5210.552 / 0.5210.531 / 0.499
Scaffold0.555 / 0.5190.538 / 0.4930.522 / 0.492
Cell-line0.539 / 0.5100.499 / 0.4730.499 / 0.470
Target0.543 / 0.5340.525 / 0.5150.521 / 0.513

Entries are matched/replaced phenotype input. The matched input gives higher correlations in all twelve evaluator-by-split point-estimate comparisons. All requests are retained, including the original invalid-output penalties. These two diagnostic experiments use the original forward-KL configuration.

Thus, receiving the requested phenotype improves query fulfillment while the explicit target remains unchanged. T-only does not provide this control: its conditional distribution is unchanged across different phenotype requests. Its aggregate phenotype scores can nevertheless be high because target identity is associated with cellular responses.

3. DrugGPT is a target-only reference.

DrugGPT is included as a reference for target performance, not a baseline for our phenotype-driven task. The direct task baselines are phenotype-conditioned generators. DrugGPT shows what the pretrained target expert achieves when a target is supplied.

We have updated the main results using controlled residual adaptation and reverse KL. The revised P+T affinity results are competitive with the DrugGPT reference:

OOD splitDrugGPT (T): Vina ↓PhenTTP (P+T): Vina ↓DrugGPT (T): PbindP_{\mathrm{bind}} ↑PhenTTP (P+T): PbindP_{\mathrm{bind}} ↑
Chemical-8.57-8.290.6110.593
Scaffold-8.56-8.430.6130.605
Cell-line-8.59-8.410.5980.603
Target-7.38-8.540.5230.500

These updated comparisons are reported in Appendix Table 9. PhenTTP supports generation from a desired cellular response when the target is unknown, and additional target guidance when one is available. The target-only reference shows how much target performance is retained within this broader phenotype-driven framework.

AuthorsOpenReview ↗

Response to YTJ1. CW 2: Task, P+T control, and DrugGPT reference (2/2)

4. Alignment helps more in the challenging Tahoe setting.

We agree that simpler configurations are competitive on L1000, which uses 978 landmark-gene features. For the original L1000 alignment ablation, the paired 95% bootstrap interval for the PRNet difference, No alignment minus Full, is [−0.002,+0.007][-0.002,+0.007], using 2,000 target–molecule cluster resamples. Its additional average PRNet benefit is therefore unclear in that configuration.

The revised model shows the same small L1000 difference: PRNet is 0.5830.583 for Full and 0.5820.582 for KD-only. Here, KD-only retains distillation but removes initial pre-alignment and the later alignment losses. We will add 95% confidence intervals alongside the revised PRNet ablation results in Appendix Table 7.

The benefit is clearer on Tahoe, whose single-cell profiles cover approximately 60,000 gene features. On Tahoe Cell-line OOD, removing alignment reduces PRNet from 0.4720.472 to 0.2250.225 and VUN from 93.393.3\\% to 52.852.8\\%. We therefore retain alignment in the full framework for this more challenging setting, while recognizing that simpler configurations can perform well on L1000.

5. Joint conditioning faces a data bottleneck in this domain.

We acknowledge the observed phenotype–affinity trade-off: adding target guidance can improve predicted affinity while reducing phenotype agreement. Learning to satisfy both requirements depends on having linked phenotype–target–molecule data.

Such joint data are scarce in this domain. Phenotype measurements and protein–ligand activity records are largely available separately. After linking the available target annotations to L1000, only 2,127 of 17,766 molecules have usable target-sequence annotations. Consequently, the data supporting joint phenotype–target conditioning cover a much smaller set of molecules than phenotype–molecule supervision.

This limited joint coverage can constrain P+T performance and contribute to its remaining gap from a target-only specialist. P+T adds the ability to follow a phenotype request alongside a target, but learning this joint relationship requires data that remain limited in current phenotype-driven discovery resources. Expanding this coverage is an important direction for improving the balance between phenotype agreement and target compatibility.

Reviewer_P1EqOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a novel molecular generation framework called PhenTTP. PhenTTP uses Phenotype as a condition. Therefore, they are able to generate molecules without a target condition by generating molecules directly from gene expressions and cellular perturbation profiles. The major problem with phenotype-conditioned generation is that it's challenging to capture molecular interactions and constraints. Therefore, PhenTTP transfers knowledge from target-conditioned generative models (or representation models) into a phenotype-conditioned generator.

They proposed a 3-step training process. Stage 1 is for representation alignment by using a contrastive alignment objective. Stage 2 is per-token supervised fine-tuning for molecular GPT models. (conditioned on Phenotype) with regularization. Stage 3 is forward KL distillation from target-conditioned teacher to phenotype-conditioned student. Stage 3's objective can be interpreted to align phenotype conditional distribution into mixture of target conditional distribution.

Experiments are done mainly on the L1000 benchmark; PhenTTP showed promising performance in both IID and OOD settings. Ablation studies are properly done.

Critical strengths
  • This paper has clear novelty: transferring a protein) target-conditional prior into a phenotype-conditioned generator. Subsequent techniques have empirically enabled this transfer and seem to achieve promising performance (I'm not an expert on the benchmark).

  • Techniques (for transferring prior knowledge) are well designed. Representation alignment and moment-matching regularization are technically sound (although they are not novel); the authors seem to have designed the techniques carefully.

Critical weaknesses
  • Methods are not principled from a probabilistic point of view. For example, in stage 3, I think the target distribution becomes a mixture of target-conditioned distributions; there could be a potential collapse problem in theory. Therefore, theoretical analysis and a clear definition of the probabilistic target are recommended.
Questions for authors
  1. I'm curious about a simpler version of PhenTTP. Current algorithms are too complex (e.g., do we really need stage 1? Could we integrate stages 2 and 3?). Is there any possibility of reducing components while keeping performance? (This is an open question)

  2. Did you consider reverse KL distillation or JSD at stage 3?

5 comments
AuthorsOpenReview ↗

Response to P1Eq. Q1: Probabilistic Target and Teacher Mixture (1/2)

We thank the reviewer. The observation about the mixture is correct, and it led us to state Stage 3 as an explicit probabilistic program. A new appendix adds Definition 1, Lemma 1, and Propositions 1--2 stated below.

The probabilistic target. Stage 3 estimates the data conditionals pDP(M∣P)p_{\mathcal{D}_P}(M\mid P) and pDP,T(M∣P,T)p_{\mathcal{D}_{P,T}}(M\mid P,T) by constrained maximum likelihood over the phenotype modules ϕ\phi, with the backbone θ0\theta_0 frozen:

\max_{\phi}\quad & \rho\\,\mathbb{E}_{\mathcal{D}_P}\big[\log p_\theta(M\mid P)\big] +(1-\rho)\\,\mathbb{E}_{\mathcal{D}_{P,T}}\big[\log p_\theta(M\mid P,T)\big]\\\\ \text{s.t.}\quad & d_r\le\varepsilon, \qquad \delta_i\ge\gamma . \end{aligned}\tag{1}$$ Eliminating nonnegative slacks turns these constraints into the implemented hinges $\tfrac12[d_r-\varepsilon]_+^2$ and $[\gamma-\delta_i]_+$, recovering $\mathcal{L}^{(3)}$ exactly. The teacher $q_{T,t}=p_0(\cdot\mid m_{ < t},T)$ therefore enters only through a *feasible set*, never as the estimand. If the data conditional is feasible, it is the population optimum of (1), so Stage 3 is consistent for it; if the KL constraint binds, the optimum is the information projection of the data conditional onto $\\{d_r\le\varepsilon\\}$, with the induced bias controlled by $\varepsilon$. **What the mixture is.** Let $s_P=p_\theta(\cdot\mid m_{ < t},P)$ and let $\nu(T\mid P,m_{ < t})$ be the conditional law induced by triple sampling and comparable-token weighting. **Lemma 1.** *For every phenotype and prefix,* $$\mathbb{E}_{T}\\,D_{\mathrm{KL}}\\!\big(q_{T,t}\\,\big\\|\\,s_P\big) = C+D_{\mathrm{KL}}\\!\big(\bar q_P\\,\big\\|\\,s_P\big), \qquad \bar q_P=\sum_T\nu(T\mid P,m_{ < t})\\,q_{T,t},$$ *with $C$ independent of $\phi$. Hence the minimizer of the hinge-free distillation term is $\bar q_P$, and it varies with $P$ unless $\nu$ is independent of $P$.* The mixture is thus conditioned on *phenotype and prefix*; it is not a single distribution obtained by pooling all targets, and assigning two phenotypes with different mixtures a common student distribution is strictly suboptimal for this term. **The distillation term is a budget, not an attractor.** Lemma 1 concerns the hinge-free term. The implemented term behaves differently. **Proposition 1.** *Wherever $d_r < \varepsilon$, both $\mathcal{L}_{\mathrm{proj}}$ and its gradient vanish. Consequently, $\bar q_P$ is the* center *of the feasible region $\\{d_r\le\varepsilon\\}$ and not a point the objective moves toward: inside the budget, the solution is fixed by the likelihood and margin terms alone.* This is the main reason the mixture does not induce collapse here. Even in the worst case for Lemma 1, where $\nu$ barely depends on $P$ so the mixtures across phenotypes nearly coincide, the distillation term is *neutral* rather than collapse-inducing: it bounds deviation from the chemical prior without rewarding agreement with it. Exact agreement with the frozen teacher is not required by the soft KL budget, whether using forward KL or reverse KL. The arithmetic-mixture decomposition in Lemma 1 applies specifically to forward KL; we compare both divergence directions in Question 3. The P-only route is additionally anchored to data by $\mathcal{L}_{\mathrm{replay}}$ on the pair set.
AuthorsOpenReview ↗

Response to P1Eq. Q1: Probabilistic Target and Teacher Mixture (2/2)

Phenotype-invariant solutions are excluded. We can now state the collapse question formally.

Definition 1. A model is phenotype-invariant if pθ(⋅∣m<t,P)p_\theta(\cdot\mid m_{ < t},P) does not depend on PP for almost every prefix. Write M0\mathcal{M}_0 for the set of such ϕ\phi. Collapse is the event that the Stage 3 optimum lies in M0\mathcal{M}_0.

With the phenotype contrast written as

δi=ℓi(Pi−,Ti)−ℓi(Pi,Ti),LctrP,T=1∣VP∣∑i∈VP[γP−δi]+,\delta_i=\ell_i(P_i^-,T_i)-\ell_i(P_i,T_i), \qquad \mathcal{L}_{\mathrm{ctr}}^{P,T} =\frac{1}{|\mathcal{V}_P|}\sum_{i\in\mathcal{V}_P}\big[\gamma_P-\delta_i\big]_+ ,

the following holds.

Proposition 2. On M0\mathcal{M}_0 every δi=0\delta_i=0, so LctrP,T=γP\mathcal{L}_{\mathrm{ctr}}^{P,T}=\gamma_P exactly, the maximal value of the hinge, and the Stage 3 objective on M0\mathcal{M}_0 carries a constant excess of λctrγP\lambda_{\mathrm{ctr}}\gamma_P. Moreover, at any ϕ∈M0\phi\in\mathcal{M}_0 with dr<εd_r < \varepsilon the hinge is strictly active, so the objective is differentiable there with contrast gradient −λctr∇ϕδ-\lambda_{\mathrm{ctr}}\nabla_\phi\delta; any direction along which the conditioning pathway is locally sensitive to the phenotype, ∇ϕδ≠0\nabla_\phi\delta\neq0, strictly decreases the objective at first order while the budget remains slack. Hence M0\mathcal{M}_0 contains no local minimizer of the Stage 3 objective.

Proposition 2 bears directly on our optimization path. Since WrW_r is zero-initialized, the P+T route starts inside M0\mathcal{M}_0, reproducing p0(⋅∣T)p_0(\cdot\mid T) exactly; at that point the full penalty λctrγP\lambda_{\mathrm{ctr}}\gamma_P is paid, and it can be reduced only by making the generator phenotype-dependent. The target-switch margin excludes target-invariant solutions symmetrically at fixed phenotype.

Empirical corroboration. Adding dual distillation to the alignment-enabled model increases PRNet, CRISP and MAP rΔr_\Delta from 0.544/0.523/0.5120.544/0.523/0.512 to 0.562/0.546/0.5310.562/0.546/0.531 and VUN from 83.383.3\\% to 92.692.6\\% in Table 3; the paired 95% bootstrap interval for the PRNet difference, No-KD minus Full, is [−0.024,−0.014][-0.024,-0.014] under target--molecule cluster resampling of the archived predictions. In the full forward-KL P+T checkpoint, holding the molecule and the explicit target fixed, the matched phenotype attains lower length-normalized NLL than a frozen dissimilar donor in 64.1564.15--68.5068.50\\% of comparisons across the four OOD splits. Re-generating 2,0002{,}000 requests per split with the target fixed and the phenotype input replaced by that donor (seed 2026082920260829, unchanged decoding), and scoring both outputs against the originally requested phenotype in its original cell, dose, and time context with the original invalid-output penalties retained, gives matched / replaced:

OOD splitPRNet rΔr_\DeltaCRISP rΔr_\DeltaMAP rΔr_\Delta
Chemical0.557 / 0.5210.552 / 0.5210.531 / 0.499
Scaffold0.555 / 0.5190.538 / 0.4930.522 / 0.492
Cell-line0.539 / 0.5100.499 / 0.4730.499 / 0.470
Target0.543 / 0.5340.525 / 0.5150.521 / 0.513

The matched input is favored in all twelve evaluator-by-split comparisons, and the ordering follows the analysis: the gap is smallest on the target split, where the explicit target carries most of the conditioning signal, and largest on the chemical and scaffold splits, where the phenotype must supply what the target prompt does not.

AuthorsOpenReview ↗

Response to P1Eq. Q2: Simpler training schedules

Yes. Our experiments support simpler training schedules with competitive phenotype performance on L1000. PhenTTP uses a single frozen molecular backbone with the same phenotype encoder, projection, and adapters throughout. The three stages organize supervision and optimization within this model. Their roles are summarized below.

StageSupervisionObjectivePurpose
1P--T annotationsLalign=Lcos+LNCE+12Lmargin\mathcal L_{\mathrm{align}} =\mathcal L_{\mathrm{cos}}+\mathcal L_{\mathrm{NCE}} +\tfrac12\mathcal L_{\mathrm{margin}}Initialize phenotype-derived prompts in the native target space.
2P--M pairsLNLLP+λctrLctrP+λmomLmom\mathcal L_{\mathrm{NLL}}^P +\lambda_{\mathrm{ctr}}\mathcal L_{\mathrm{ctr}}^P +\lambda_{\mathrm{mom}}\mathcal L_{\mathrm{mom}}Learn phenotype-conditioned generation from the larger pair set.
3P--T--M triples and P--M replay(1−ρ)Ltriple+ρLreplay(1-\rho)\mathcal L_{\mathrm{triple}} +\rho\mathcal L_{\mathrm{replay}}Transfer target-conditioned knowledge while retaining phenotype supervision.

Here, Ltriple\mathcal L_{\mathrm{triple}} combines P+T likelihood, phenotype and target-switch contrasts, dual-route KD, representation alignment, and moment regularization. Replay continues the P-only pair objectives with the Stage-3 weights in Appendix A.3. Stage 1 aligns phenotype-derived prompts with the target space that the frozen generator was pretrained to use. This gives phenotype-conditioned generation a starting point that fits the generator's existing target interface. The amount and quality of P--T annotations may limit the benefit of this initialization. Stage 3 also includes alignment, so it can partly compensate when Stage 1 is skipped.

Our new No Stage 1 ablation removes only the 5,000 pre-alignment updates, preserving subsequent objectives and training budgets. L1000 phenotype scores remain close: all eight split/route paired 95% PRNet difference intervals include zero, using 2,000 target--molecule cluster bootstrap resamples of single-seed predictions. Some affinity outcomes favor retaining Stage 1: Chemical P-only Vina changes from −8.26-8.26 to −7.88-7.88 upon removal, and Cell-line P-only PbindP_{\mathrm{bind}} from 0.3310.331 to 0.2910.291. We also tested joint training, which optimizes all three stage objectives together from the start. It provides a simpler training schedule and remains competitive on L1000 (Appendix Table 7).

SettingSimplificationFullSimplified
L1000 mean, P-onlyNo Stage 10.5620.562
L1000 mean, P+TNo Stage 10.5490.548
L1000 mean, P-onlyJoint training0.5620.565
Tahoe Cell-line OOD, P-onlyJoint training0.4620.361
Tahoe Cell-line OOD, P-onlyNo P--T alignment0.4620.206

Entries report PRNet rΔr_\Delta; L1000 means average the reported scores across four OOD splits. The Tahoe alignment-off ablation removes both Stage 1 and the Stage-3 alignment losses. Tahoe comparisons use the reported configuration with the phenotype residual adapter disabled.

On Tahoe Cell-line OOD, staged training achieves PRNet correlation of 0.4620.462 and VUN of 91.591.5\\%, compared with 0.3610.361 and 71.471.4\\% for joint training (Appendix Table 10). Removing P--T alignment also lowers PRNet correlation to 0.2060.206. These results show a clearer benefit from alignment and staged training in this challenging setting (Tahoe: approximately 60,000 gene features vs. L1000: 978 landmark genes). In contrast, the modest additional benefit of Stage 1 on L1000 may partly reflect sparse phenotype--target supervision, with usable target annotations available for only 2,127 of 17,766 molecules. We therefore retain the full staged framework for its robustness in challenging settings, while agreeing with the reviewer that simpler training strategies are viable and may be preferable for less challenging tasks.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Updated manuscript and thanks to the reviewers

We thank all three reviewers for their constructive feedback and for recognizing the motivation and potential of our phenotype-driven generation framework. Their suggestions have helped us improve the model, strengthen the evaluation, and clarify the presentation.

We have updated the main manuscript and the corresponding experimental results. The revised model uses reverse KL for distillation and introduces a learnable scale α\alpha to control the contribution of the phenotype residual pathway rPr_P. The main results and component ablations have been updated accordingly.

Our individual replies explain these changes and present the additional experiments addressing each reviewer's questions. We appreciate the reviewers' guidance throughout this revision.

AuthorsOpenReview ↗

Rebuttal Summary and Key Revisions

We sincerely thank all three reviewers for their constructive feedback. Throughout the rebuttal, we have conducted extensive additional experiments, revised the model, and substantially improved the manuscript to address the main concerns raised during the review process.

First, we strengthened the methodological foundations and revised the model. We clarified the probabilistic formulation of Stage 3 and analyzed the potential collapse concern. Following the reviewers' suggestions, we systematically compared forward KL, reverse KL, and JSD, and adopted reverse KL based on its consistent improvements in phenotype agreement. We also introduced learnable scaling of the phenotype residual pathway, motivated by new component ablations. The revised model and its updated experimental results are incorporated into the manuscript.

Second, we provided additional evidence addressing the key experimental concerns. In particular, fixed-target phenotype-swap experiments demonstrate that P+T genuinely responds to phenotype conditioning rather than relying solely on target information. Importantly, we repeated this experiment with the revised model and confirmed that the requested phenotype outperforms a swapped phenotype in all 12 evaluator-by-split comparisons. We also investigated alternative training strategies and the role of phenotype–target alignment, showing that simpler configurations are competitive on L1000, while alignment and staged training provide clearer benefits on the challenging Tahoe Cell-line OOD setting. Additional ablations, uncertainty analyses, and quantitative docking controls further strengthen the evaluation.

Third, we improved clarity, reproducibility, and the interpretation of our contributions. We revised the framework illustration, methodological descriptions, and docking visualizations, and provided implementation resources. We also clarified that phenotype-driven generation is our primary objective, while P+T provides optional target guidance when mechanistic information is available. We explicitly discussed the phenotype–affinity trade-off and the limitations of current phenotype–target–molecule supervision.

We are encouraged by the reviewers' feedback following these revisions. Reviewer P1Eq explicitly stated that their concerns were resolved and maintained a positive score. Reviewer C5NB confirmed that our additional phenotype-swap experiments addressed their main concern and indicated that they would raise their overall rating. Reviewer YTJ1 also acknowledged that the manuscript had become clearer, the docking presentation had improved significantly, and most of their comments had been resolved. We subsequently provided further discussion and evidence regarding their remaining concerns about target conditioning and data limitations.

Overall, we believe the rebuttal has substantially strengthened both the technical formulation and empirical support for PhenTTP, while making its contributions and limitations clearer. We are grateful for the reviewers' constructive engagement and respectfully hope that the Area Chair will consider these revisions and the reviewers' updated feedback in the final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.