acceptodds
Under review as a conference paper at ICLR 2027

DECO: Decoupled Conditional Generation for Transcriptomics-Driven Drug Design

Abstract

Transcriptomics-driven drug design seeks to generate molecules that induce desired cellular responses without requiring predefined molecular targets. A central challenge is that transcriptomic responses and molecular structures inhabit fundamentally different representation spaces, making it nontrivial to connect biological conditions to chemical generation. Existing approaches bridge these modalities through direct conditioning, cross-modal alignment, or coupled representation learning. Our analysis further reveals a weak and non-unique relationship between transcriptomic-response similarity and chemical similarity. Guided by this observation, we introduce DECO (DEcoupled COnditional generation), which learns chemical representations independently and models desired responses as conditional distributions over the resulting latent space. To make this decoupled space responsive to transcriptomic conditions, DECO combines global transition features for adaptive modulation with gene-level transition tokens for cross-attention. A generation-centered objective explicitly promotes condition dependence while retaining cross-modal alignment only as auxiliary semantic supervision. A prediction-error-aware decoder improves robustness to generator-induced latent deviations while preserving reconstruction fidelity. Across standard, cell-line OOD, and scaffold OOD splits, DECO achieves the best paired structural similarity and predicted transcriptomic consistency among evaluated methods, reducing Best@10 response MSE by 19.0–43.2% relative to the best-performing baseline in each split while maintaining 94.5–96.1% molecular validity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular generative models, transcriptomics-driven drug discovery

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_8stJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript proposes a framework for transcriptomics-driven molecular generation, called DECO. The authors first show that transcriptomic response similarity has very little correlation with molecular structural similarity, and use this observation to motivate learning the chemical latent space independently of transcriptomic information. The method combines a pretrained molecular VAE with conditional latent diffusion, using both global and gene-level transcriptomic features. It also introduces auxiliary losses to encourage condition usage and a decoder adaptation strategy to handle errors in generated latents. The experimental results on LINCS L1000 show improvements over TransGEM, XMolRL, and CURE in structural similarity and PRNet-predicted response consistency, including cell-line and scaffold OOD settings.

Critical strengths
  • I like that the method is motivated by an actual observation in the data. The weak correlation between transcriptomic response similarity and chemical similarity is interesting, and it provides a reasonable motivation for not forcing the two modalities into a shared representation. I think this makes the overall approach more convincing than simply introducing another conditional diffusion architecture.

  • DECO outperforms the compared methods on both structural similarity and predicted response consistency across all three splits. I also appreciate the ablations on global and gene-level conditioning, decoder adaptation, and auxiliary objectives. These results suggest that the improvements are not coming from a single component alone.

Critical weaknesses
  • One of my main concerns is that the response metric compares PRNet predictions for the generated molecules against PRNet predictions for the reference molecules, rather than directly comparing them with the desired experimental transcriptomic response. I understand the motivation for using a frozen predictor, but I think this distinction matters. A generated molecule could receive a good score because PRNet considers it similar to the reference molecule, without necessarily reproducing the desired cellular response. Since transcriptomics-driven generation is the main goal of the paper, I would like to see stronger evidence that the improvement goes beyond agreement with this particular predictor.

  • In additon, the authors argue that weak response–structure correspondence motivates learning the chemical latent space independently. I find this reasonable, but DECO also introduces dual-scale conditioning, additional training losses, and decoder adaptation. The current ablations show that these components help, but they do not directly compare decoupled and coupled representation learning under the same generative architecture. I therefore find it difficult to tell whether the gains mainly come from decoupling or from the other improvements. This is important because decoupling is presented as the central contribution of the paper.

Questions for authors
  • In Appendix E.2, the response metrics compare PRNet predictions for generated and reference molecules. Why not also evaluate the generated molecules against the experimentally observed target transcriptomic profiles? Could the authors clarify how well the PRNet-based improvements correlate with agreement with the actual desired responses? This would materially affect my confidence in the main experimental conclusions.

  • Could the authors compare DECO with a coupled representation-learning variant using the same VAE, diffusion architecture, conditioning modules, and decoder adaptation? I would be interested to know whether independent chemical representation learning still provides an advantage when the other components are controlled for. This would help me judge the importance of the paper's central design choice.

2 comments

We thank the reviewer for the constructive feedback.

Q1. Response-consistency evaluation

Following prior work, we compare PRNet predictions of generated and reference molecules to reduce the influence of prediction errors shared by the same surrogate model.

We acknowledge that this evaluation relies on a predictive model and does not constitute experimental validation of the generated molecules' biological activity.

Q2. Controlled evaluation of decoupling

We compared Decoupled and Coupled training, initialized from the same pretrained molecular VAE and retaining the core diffusion and conditioning architecture.

Decoupled keeps its chemical representation fixed during conditional training, whereas Coupled jointly optimizes the chemical representation and conditional generation. We also adapted the Coupled decoder to its own latent prediction errors after joint training to reduce differences in decoder robustness.

Table 2. Decoupled versus Coupled.

SplitModelValidity ↑Fraggle@10 ↑Morgan@10 ↑
TestDecoupled0.95780.56690.2780
Coupled0.82190.50780.1985
Cell OODDecoupled0.94530.49530.1746
Coupled0.83440.47060.1638
Scaffold OODDecoupled0.96090.47660.1602
Coupled0.80940.48120.1507

Table 3. Quality-constrained conditional coverage (10 samples, PRNet MSE threshold = 10⁻³).

SplitModelSuccess Rate@10 ↑Successful Scaffolds@10 ↑
TestDecoupled22.97%2.250
Coupled13.75%1.375
Cell OODDecoupled16.33%1.602
Coupled15.08%1.500
Scaffold OODDecoupled30.23%2.875
Coupled16.95%1.680

Success requires validity, ChemPass, and surrogate-response MSE ≤ 10⁻³; 128 identical conditions per split.

Decoupled achieves higher validity and Morgan similarity across all splits, and higher average quality-constrained coverage, although the Cell OOD difference is small and Coupled slightly leads in Scaffold OOD Fraggle@10.

Joint optimization also contracts chemical latent variance from 27.94 to 0.66 on 1,024 held-out molecules. These observations support the practical value of the decoupled strategy under our evaluated training setting, beyond the original component ablations.

Reviewer_8stJOpenReview ↗

After rebuttal

Thank you for the authors' responses. Could the authors answer the weaknesses that I concerned.

Until now, I think that this manuscript need improve more on the paper quality. So I keep my current rating.

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

This paper presents DECO, a decoupled conditional generation framework for transcriptomics-driven drug design. DECO independently learns a molecular latent space and models the conditional distribution of molecular latents given a desired transcriptomic transition.

Critical strengths
  1. The dual-scale conditioning mechanism is well designed. By combining global and gene-level conditions, the model can capture both transcriptomic changes and gene-resolved responses.
  2. The experimental results are strong and consistent across standard, cell-line OOD, and scaffold OOD settings. DECO achieves substantially lower surrogate-predicted response errors while maintaining high molecular validity.
  3. The ablation studies support the usefulness of several proposed components.
Critical weaknesses
  1. The architectural novelty of DECO appears limited. As described in Appendix B, it relies on an LSTM-based VAE for molecular representation, while other components such as pooling and attention use largely conventional architectures. Thus, it is unclear how much of the reported improvement comes from the proposed decoupling strategy itself rather than the combination of established components.
  2. On page 3, line 161, the paper states that “Figure 2 summarizes the three stages of DECO: independent chemical representation learning, dual-scale conditional latent generation, and prediction-error-aware decoder adaptation.” However, Figure 2 only presents (a) Dual-scale Conditioning and Latent Diffusion and (b) Prediction-Error-Aware Decoding. The first stage, independent chemical representation learning, is not shown in the figure, which makes the description of Figure 2 inconsistent with its actual content. The authors should clarify this discrepancy or revise Figure 2 accordingly.
  3. The caption of Figure 2 states that “Global pooled features condition latent diffusion through AdaLN, while gene-level features [HCTL, HCP, ∆H], with ∆H = HCP − HCTL, are injected through cross-attention.” However, the role of AdaLN is unclear, as it is not explained anywhere in the main text. Moreover, the pipeline shown in Figure 2 indicates that [HCTL, HCP, ∆H] are first fed into an MLP and then injected into the cross-attention module. This also appears inconsistent with Eq. (4), which defines the computation as Cgene=fgene([HCTL,HCP,HCP−HCTL])C_{gene} = f_{gene}([H_{CTL}, H_{CP}, H_{CP} - H_{CTL}]). The authors should clarify the role of AdaLN and explain the apparent discrepancy between the architecture illustrated in Figure 2 and the computation described in Eq. (4).
  4. On page 4, line 203, the paper states that “..., and an adaptive fusion module followed by an MLP produces the global condition.” However, Eq. (3) only describes the global pooling operation and does not provide the formulation or computation details of the adaptive fusion module. In addition, the function fglobalf_{global} is not sufficiently defined or explained. The authors should provide the missing formulation and clarify the role and implementation of fglobalf_{global} to make the proposed global conditioning mechanism reproducible.
  5. Several mathematical notations and definitions in the equations require clarification. In Eq. (5), the notation ϵ∼N(0,I)\epsilon \sim \mathcal{N}(0, I) is ambiguous. While a standard normal distribution is typically written as N(0,1)\mathcal{N}(0, 1), the meaning and dimensionality of II in N(0,I)\mathcal{N}(0, I) are not defined. In addition, μ\mu in Eq. (2) is not sufficiently explained. Eq. (6) does not define E(z0,s),tE_{(z_0, s),t}, while τ\tau in Eq. (7) is also undefined. Furthermore, Eq. (8) appears to be missing a closing parenthesis after the first summation term, and the base of the loglog is not specified. These issues should be addressed to improve the mathematical clarity and reproducibility of the proposed method.
  6. The biological and practical utility of the generated molecules remains largely unvalidated experimentally. The experiments do not include experimental transcriptomic validation, target-binding assays, phenotypic assays, or synthesis validation. Therefore, the reported results should primarily be interpreted as computational evidence of conditional molecular generation rather than evidence of effective drug design.
Questions for authors
  1. What is the key technical contribution of DECO beyond existing components? Given the use of an LSTM-based VAE and conventional pooling and attention modules, could the authors provide ablations that isolate the contribution of the proposed decoupling strategy and prediction-error-aware decoder adaptation?
  2. Could the authors clarify the mathematical formulation and implementation of the conditioning mechanism? In particular, how do AdaLN, the adaptive fusion module, and fglobalf_{global} operate, and how should Eq. (4) be reconciled with Figure 2? Please also define the missing symbols and operations in Eqs. (2) and (5)-(8).
  3. How well do the experiments support the claim of effective drug design? Since no experimental transcriptomic validation, target-binding assays, phenotypic assays, or synthesis validation are provided, could the authors discuss the limitations of the current computational evaluation and provide evidence that the generated molecules have biological and practical utility?
1 comment

We thank the reviewer for the careful reading and concrete suggestions. We have checked the implementation and prepared revisions to the mathematical descriptions and figure references.

W1. Technical contribution and controlled evaluation

DECO proposes a data-motivated, generation-centered modeling strategy, separating chemical representation learning from transcriptome-conditioned distribution estimation. Dual-scale conditioning and prediction-error-aware decoding support this formulation.

We compared Decoupled and Coupled training, initialized from the same pretrained molecular VAE and sharing the core conditional generation architecture. Decoupled keeps the chemical representation fixed, while Coupled jointly optimizes the chemical representation and conditional generator. We also adapted the Coupled decoder to its own latent prediction errors after joint training.

Decoupled retains higher molecular validity across Test, Cell-line OOD, and Scaffold OOD (0.958/0.945/0.961 vs. 0.822/0.834/0.809).

Table 1. Quality-constrained conditional coverage (threshold = 10⁻³, 10 samples per condition).

SplitModelSuccess Rate@10 ↑Successful Scaffolds@10 ↑
TestDecoupled22.97%2.250
Coupled13.75%1.375
Cell OODDecoupled16.33%1.602
Coupled15.08%1.500
Scaffold OODDecoupled30.23%2.875
Coupled16.95%1.680

Success requires validity, ChemPass, and surrogate-response MSE ≤ 10⁻³. Results use 128 matched conditions per split; scaffold counts are per-condition averages.

Decoupled achieves higher quality-constrained coverage, particularly on Test and Scaffold OOD. Joint training also contracts total chemical latent variance from 27.94 to 0.66. The original ablations separately evaluate prediction-error-aware decoder adaptation; this comparison further examines the decoupled training strategy.

W2. Figure 2

We have clarified that Figure 2 illustrates only dual-scale conditional generation and decoder adaptation. Independent chemical VAE pretraining is described separately in Section 3.2.

W3. AdaLN and gene-level conditioning

We verified the implementation and clarified the conditioning pathway in Section 3.3 and Appendix B.

The global condition combines with the timestep embedding to produce AdaLN scale, shift, and residual-gating parameters for the cross-attention and feed-forward sublayers.

For gene-level conditioning, control, perturbed, and difference features are concatenated and processed by a projection MLP (192→256→256). The resulting 978 gene tokens provide cross-attention keys and values, while the projected noisy molecular latent provides the query. Eq. (4) and the Figure 2 caption now clarify this pathway.

W4. Adaptive fusion

Revised Eq. (3) specifies feature-wise gated fusion of two 64-dimensional attention-pooled state vectors. The control vector, perturbed vector, gated mixture, and element-wise product are concatenated into 256 dimensions, projected to a 512-dimensional phenotype representation, and then mapped to the 256-dimensional global condition. Details are provided in Appendix B.

W5. Mathematical notation

We clarified latent dimensions and posterior notation (Eq. 2), Gaussian noise (Eq. 5), denoiser mapping and timestep indexing (Eq. 6), auxiliary projection and similarity definitions (Eq. 7), and positive sets, parentheses, and the logarithm convention (Eq. 8). These corrections do not alter the implementation.

W6. Experimental validation

We agree that the results establish computational feasibility rather than experimentally verified molecular activity. Molecular validity, chemical-quality metrics, and surrogate-predicted response consistency do not replace wet-lab transcriptomic, binding, phenotypic, or synthesis validation. We will clarify these limitations and avoid overstating biological efficacy.

Reviewer_F2NiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies transcriptomics-driven molecular generation, aiming to design molecules that induce desired cellular responses specified by control-to-perturbed gene-expression transitions. Motivated by an empirical analysis showing a weak association between transcriptomic-response similarity and molecular structural similarity, the authors propose DECO, a decoupled conditional generation framework.

Specifically, DECO first learns a molecular latent space independently of transcriptomic supervision and subsequently performs conditional latent diffusion using both global transition representations and gene-level features. The framework further incorporates auxiliary semantic alignment and prediction-error-aware decoder adaptation to improve conditional generation and molecular validity.

Experiments on LINCS L1000 demonstrate improvements over the evaluated baselines in paired structural similarity and surrogate-predicted transcriptomic consistency.

Critical strengths
  1. Motivation is supported by an empirical analysis. I appreciate the quantitative analysis of the relationship between transcriptomic-response similarity and molecular structural similarity. The weak correlation observed across matched experimental contexts provides useful empirical evidence for the non-uniqueness of the inverse mapping from biological responses to chemical structures.
  2. Encouraging empirical results and meaningful component analyses. DECO achieves favorable results across standard and OOD evaluation settings. The ablation studies also provide evidence for the effectiveness of gene-level conditioning and decoder adaptation.
Critical weaknesses

1. Limited methodological novelty relative to existing work, particularly CURE.

My primary concern is that the central motivation and several important methodological elements substantially overlap with prior work, particularly CURE (Xu et al., ICML 2026).

CURE has already formulated transcriptomics-driven molecular design as an inherently ill-posed, one-to-many inverse problem, explicitly recognizing that structurally distinct molecules may induce similar transcriptomic responses. It also adopts a distributional generation perspective rather than assuming a unique molecular solution. Therefore, while DECO's quantitative response–structure correlation analysis is valuable, the underlying observation is not conceptually new.

Furthermore, DECO's bidirectional transcriptomic interaction blocks share considerable architectural similarity with CURE's TFE-I module, which also employs cross-attention and self-attention to extract perturbation representations from pre- and post-perturbation states. Independently learned molecular representations followed by transcriptome-conditioned latent diffusion also have precedents, such as GLDM (Wang et al., 2024).

2. Insufficient evidence to support the central decoupling hypothesis.

The weak global correlation between response similarity and structural similarity does not necessarily imply that cross-modal alignment imposes harmful constraints. The paper does not directly demonstrate that such alignment compromises the modeling of the underlying multimodal conditional distribution. Moreover, the comparison with CURE simultaneously changes the molecular representation, diffusion backbone, conditioning architecture, training objectives, and decoding procedure. Consequently, the observed performance improvements cannot be attributed specifically to the proposed decoupling strategy.

A more convincing investigation would compare coupled and decoupled training within a controlled architecture, isolating the effects of freezing the chemical encoder and modifying the alignment objective. Such evidence is particularly important because decoupling is presented as the central contribution of the paper.

3. Insufficient evidence to fully substantiate the empirical advantages over prior work and the claimed functional consistency.

Reported low paired Morgan similarity and QED values of CURE differ substantially from those reported in the original CURE paper. The experimental protocols differ, so these numbers are not directly comparable. Nevertheless, the discrepancy is sufficiently large that a more detailed explanation of baseline adaptation, training configurations, and evaluation settings is important for interpreting the claimed improvements.

Additionally, the primary response-consistency evaluation relies on a frozen PRNet surrogate. An evaluation using an additional independent response predictor would strengthen the conclusions.

Questions for authors
  1. Can the authors more directly isolate and validate the contribution of decoupling?
  2. Could the authors clarify the implementation and reproduction protocol for CURE?
  3. Can the authors provide stronger evidence of functional consistency and one-to-many generation?
1 comment

We thank the reviewer for the constructive feedback.

W1. Methodological novelty

We acknowledge that CURE has discussed the one-to-many inverse problem, while GLDM has explored pretrained molecular representations with conditional latent diffusion. We do not claim novelty for these individual concepts.

Our contribution is a data-motivated, generation-centered framework. Our matched-context analysis reveals weak global response–structure correspondence (Spearman ρ = 0.012), motivating us to learn chemical representations independently and focus transcriptomic supervision on the conditional distribution of molecular solutions. Dual-scale conditioning, condition-usage supervision, and prediction-error-aware decoding jointly support this modeling strategy.

W2. Controlled evaluation of decoupling

We compared Decoupled (DECO) with a Coupled variant, both initialized from the same pretrained molecular VAE and retaining the core diffusion and conditioning architecture.

Decoupled keeps the chemical representation fixed during conditional training, whereas Coupled jointly optimizes the chemical representation and conditional generation with reconstruction and KL regularization. To reduce differences in decoding robustness, we also adapted the Coupled decoder to its own latent prediction errors after joint training.

Table 1. Decoupled versus Coupled.

SplitModelValidity ↑Fraggle@10 ↑Morgan@10 ↑
TestDecoupled0.95780.56690.2780
Coupled0.82190.50780.1985
Cell OODDecoupled0.94530.49530.1746
Coupled0.83440.47060.1638
Scaffold OODDecoupled0.96090.47660.1602
Coupled0.80940.48120.1507

We further evaluated quality-constrained conditional coverage using 10 samples and 128 matched conditions per split.

Table 2. Quality-constrained coverage (PRNet MSE threshold = 10⁻³).

SplitModelSuccess Rate@10 ↑Successful Scaffolds@10 ↑
TestDecoupled22.97%2.250
Coupled13.75%1.375
Cell OODDecoupled16.33%1.602
Coupled15.08%1.500
Scaffold OODDecoupled30.23%2.875
Coupled16.95%1.680

Success requires validity, ChemPass, and surrogate-response MSE ≤ 10⁻³. Scaffold counts are distinct qualifying scaffolds per condition.

Decoupled maintains higher validity and Morgan similarity across all splits, together with higher quality-constrained coverage, although the Cell OOD margin is small. Coupled slightly outperforms Decoupled on Scaffold OOD Fraggle@10.

Joint optimization also contracts total latent variance from 27.94 to 0.66 on 1,024 held-out molecules. These observations support the practical value of our decoupled strategy under the evaluated setting.

W3. CURE reproduction

We appreciate the question about differences from CURE's originally reported results. Our baseline follows the publicly released CURE implementation under our data splits and evaluation protocol.

The paper describes joint global–local alignment, whereas the released scripts perform global alignment followed by fingerprint-based local alignment initialized from the global checkpoint. We contacted the authors for clarification and corresponding training configurations, but the complete joint-training implementation and pretrained weights were unavailable to us.

We will clarify the reproduction version and protocol. This observation concerns the reproducibility of the public implementation, not the validity of CURE's originally reported results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.