acceptodds
Under review as a conference paper at ICLR 2027

scTREAT: Latent Diffusion for Single-Cell Drug-Response Prediction

Abstract

Generating drug-induced transcriptional responses requires a model to account for cellular background together with molecular identity and dose. We present scTREAT, the first diffusion framework to factorize state and drug guidance for single-cell drug-response prediction. A shared latent expression space connects baseline and response, while Guided Decomposable Attention (GD-Attn) injects a baseline representation and a dose-conditioned molecular embedding. Independent channel dropout exposes four conditioning states during training; inference combines three conditional predictions with separately specified state and drug coefficients, and dose modulates drug-guidance strength. On condition-level pseudobulk profiles derived from the Tahoe-100M single-cell atlas, scTREAT achieves the best value on 10 of 12 metrics for unseen cell line–drug combinations (UC) and 11 of 12 for unseen drugs (UD). Ablations examine condition injection, molecular representation, and guidance design. On UD, the relative gains over the second-best model reach 36.11% and 34.21% for DEG logFC-Spearman and logFC-Pearson. These results demonstrate the effectiveness of combining shared latent expression modeling with explicit state and drug conditioning for response prediction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell transcriptomics, drug response prediction, latent diffusion models, molecular conditioning, classifier-free guidance, unseen-drug generalization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_fQGUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Authors have proposed scTREAT which uses latent diffusion to predict pseudobulk experssion profiles after drug conditioning.

They use a frozen scimilarity AE to define the latent space and the conditions are fused into 1 token before being injected into the denoiser.

Training uses an infoNCE regularizer. Inference uses CFG with separate state and drug representation coefficient.

Eval is done on unseen pairs and seen drugs benchmark. The authors report best values on 10-11 metrics against 8 baselines.

Ablations are also reported with qualitative analysis using UMAP representation and gene set enrichment.

Critical strengths

the datasets are clean and leakage is managed properly. DMSO controls are plate matched and unseen drugs held out set is across all doses and plates.

pathway analysis is correct in showing partial agreement.

Critical weaknesses

the metrics benchmarks are old. cell-eval 2 package has been released by ARC which calculate correctly calibrated metric values.

methods like CellFlow and scDFM are not reported which are current state of the art.

simple baselines like control + mean training drug difference within cell lines is not reported.

all results seem to be single runs. i did not see the word random seed or multiple seeds anywhere. always report mean + standard deviation on multiple seeds. confidence intervals are also important for baseline comparison and ablations since some wins are 0.01 to 0.02 margins.

cross attn to single condition token reduces to an input independent learned bias since softmax over one key equals 1. please explain the "non concatenative attention injection" claim in more detail for the reader.

I believe the FiLM layer is doing the same job as cross attention in here since output is condition dependent vector added to every token. please do the ablation and show results.

The biggest issue is that the contribution and novelty is extremely narrow. Latent diffusion and multi condition compositional guidance is not new, and just applying it to perturbation modelling with fixed dose to strength map is not the level of novelty for an ICLR paper.

Questions for authors
  1. please reported calibrated metrics (maybe from cell-eval2) like perturbation discrimination, DE overlap/score, MAE relative to a mean-response baseline with multiple seeds and confidence interval. a control + mean delta baseline is also important. this will invoke more confidence for the reader.

  2. what do ablations of L align or L info and of FiLM only conditioning (no cross-attention) show, given that cross attention to one token cannot depend on the latent?

  3. How were sp, s0, alpha, gamma chosen, on which split, and how many zT draws per condition enter the metrics?

i believe this paper would be good for a computational biology venue rather than ICLR. the field of perturbation modelling has moved really quickly in the past year and incremental methods may show better numbers without any true ML novelty.

Reviewer_g2cZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

scTREAT predicts post-treatment gene expression profiles from pre-treatment expression, drug structure, and dosage using a SCimilarity-based autoencoder coupled with a latent diffusion model. The autoencoder is fine-tuned on training-set expression profiles and then frozen, providing a shared latent space for pretreatment and post-treatment profiles. Morgan fingerprints are embedded through an MLP and modulated by dosage using Feature-wise Linear Modulation (FiLM). The resulting drug representation is fused with the pre-treatment latent representation and incorporated into the denoiser through Guided Decomposable Attention (GD-Attn), which uses self-attention, cross-attention, and FiLM.

During training, Gaussian noise is added to the post-treatment latent representation, and a shared denoiser learns to predict the noise under four conditioning settings: both state and drug information present, state information removed, drug information removed, or both removed. At inference, three noise predictions, joint, drug-only, and state-only, are combined through factorized classifier-free guidance to separately weight state and drug guidance, with drug-guidance strength determined by dosage. Evaluated on condition-level pseudobulk profiles from Tahoe-100M under unseen cell line–drug combination (UC) and unseen drug (UD) settings, scTREAT achieves the best reported values on 10 of 12 UC metrics, including a tie, and 11 of 12 UD metrics.

Critical strengths

A major strength of this work is its systematic evaluation under two complementary generalization settings, unseen combinations (UC), where complete cell line–drug pairs are held out, and unseen drugs (UD), where entire drugs are excluded from training. These settings assess distinct practical challenges, predicting unmeasured responses to known compounds and extending predictions to new compounds. Strong performance under these challenging settings suggests that scTREAT could predict post-treatment gene expression in real-world situations where experimental measurements are limited. This supports its potential applicability to tasks such as drug screening, where predicting responses to untested compounds or cell line–drug combinations could help prioritize experiments.

Another major strength of this work is its flexible integration of pretreatment expression, molecular structure, and dose via separately weighted cell-state and drug guidance. Four-state training enables these two guidance terms to be weighted independently during inference, while dosage adjusts both the drug representation and the strength of drug guidance. This design provides control over the relative emphasis given to gene expressions versus drug properties when predicting post-treatment expression, while effectively incorporating concentration effects. The component ablations confirm the efficacy of this architecture, showing that separate guidance weights and dose-dependent modulation give performance gains across multiple evaluation metrics.

Critical weaknesses

scTREAT does not explicitly incorporate prior information about gene-gene interaction networks. Drug-induced transcriptional responses arise not only from interactions between a compound and its molecular targets, but also from downstream effects propagated through signaling and regulatory networks. Although the model may implicitly learn some of these relationships from expression data, the reported gene-level metrics do not directly establish whether it preserves coordinated responses between genes across treatments. Evaluating whether predicted gene pairs change together as they do in the observed data would strengthen the biological validation. Incorporating or evaluating predictions against interaction information from resources such as STRING could further help assess whether the predicted responses are consistent with known biological relationships.

Also, the evaluation does not assess generalization to unseen cell lines. Although the UC and UD splits test unobserved cell line–drug combinations and unseen drugs, respectively, neither setting holds out entire cell lines. Drug responses can vary substantially across cellular contexts, and performance on cell lines represented during training does not establish whether the model can predict responses in new ones. scTREAT uses pretreatment expression to represent the initial cellular state. Therefore, an evaluation that holds out entire cell lines would help determine whether this representation supports generalization to new cellular backgrounds and clarify the model’s applicability when treatment-response measurements are unavailable for a target cell line.

Questions for authors

How were the seven cell lines shown in Figure 4 selected from the 50 cell lines in the dataset? Please clarify the selection criteria or whether similar qualitative results were observed for the remaining cell lines.

How is the fusion function (F) in Equation (7) defined? Please specify how the pretreatment latent representation and the dose-modulated drug representation are combined.

Does using a diffusion-based approach provide additional benefits over directly predicting post-treatment expression with a deterministic model? Could this be evaluated using the same encoder and decoder, an appropriately adapted GD-Attn backbone, and the same drug and dose encoding? Such a comparison would help determine whether the improvements arise from the diffusion-based formulation or from the underlying architecture and conditioning design, which the existing regression baselines do not isolate.

Could the comparison be extended to include additional relevant perturbation-prediction models? For example, CSG²A (Bang et al., Bioinformatics, 2024) predicts drug-induced gene expression from basal expression, molecular structure, dose, and treatment duration using condition-specific gene–gene attention informed by protein–protein interaction networks. Comparing scTREAT with its transcriptomic prediction component under matched data splits and evaluation metrics would help assess the proposed method against an approach that explicitly models gene relationships.

Reviewer_ok6xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes scTREAT, a single-cell drug response prediction framework based on latent diffusion models. By sharing a latent expression space, utilizing Guided Decomposable Attention (GD-Attn), and employing four-state conditional training, this method enables the joint modeling of the cell’s initial state, drug molecular structure, and dose information. On the Tahoe-100M dataset, scTREAT achieved strong performance in prediction tasks involving unseen cell line–drug combinations (UC) and unseen drugs (UD).

Critical strengths
  1. The research question has practical significance. This study focuses on predicting transcriptomic responses under unobserved drug interventions, while taking into account cellular state, molecular structure, and dose information, and thus has clear biomedical application value.

  2. The experimental results are highly competitive. scTREAT achieved the best results on 10 out of 12 metrics in the UC task and 11 out of 12 metrics in the UD task, demonstrating particularly outstanding performance in predicting differentially expressed genes in the absence of observed drug treatments.

  3. The experimental design is comprehensive. The authors compared multiple methods, including chemCPA, PRnet, and BioLORD, and analyzed the impact of key design elements on predictive performance through conditional injection, molecular representation, and ablation of guidance strategies.

Critical weaknesses

1. The independent innovativeness of the core method requires further clarification.

This paper primarily combines existing latent diffusion models, FiLM, Cross-Attention, and Classifier-Free Guidance (CFG) techniques. The four-state training and the state–drug dual-condition guidance approach have some merit, but they essentially represent a combination of multi-condition diffusion guidance mechanisms adapted to the task. The paper has not sufficiently explained what new algorithmic contributions this method offers compared to existing hybrid CFG and conditional diffusion frameworks. We recommend further clarifying the technical differences and including targeted comparative experiments using diffusion backbones with the same architecture.

2. The conditional interaction mechanism of GD-Attn lacks a sufficient theoretical explanation.

Section 3.2 notes that GD-Attn fuses cell state and drug information into a single conditional token and injects latent features via Cross-Attention. However, when Cross-Attention contains only a single key-value token, the Softmax weights are always 1, and the attention output cannot dynamically select conditional information based on different queries. Therefore, this operation is closer to injecting the same conditional representation into each latent token than to true selective cross-modal attention. Although the model can still model interactions through subsequent FiLM and residual modules, the unique advantages of GD-Attn over simple conditional projection or additive modulation remain unclear.

3. The experimental validation does not fully cover the single-cell prediction objectives proposed in the paper.

Although this paper uses the Tahoe-100M single-cell dataset, Section 4.1 and Appendix C.2 indicate that the authors first averaged single-cell expression across the same cell lines, drugs, doses, and experimental plate conditions, and then performed predictions at the pseudobulk expression level. Therefore, the current results primarily demonstrate the model’s ability to predict condition-averaged transcriptomic responses but have not yet validated its capacity to model single-cell heterogeneity, differences among cell subpopulations, and response distributions. Furthermore, the paper lacks an analysis of performance stability across multiple random seeds and does not separately ablate the alignment and contrastive learning losses introduced by Eqs. (10)–(12), leaving the independent contributions of different design elements unclear.

Questions for authors

See in Weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.