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

Zhaokang Liang; Shuyang Zhuang; Xiaoran Jiao; Weian Mao; Hao Chen

q-bio.QMOct 8, 2025

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

Authors: Zhaokang Liang, Shuyang Zhuang, Xiaoran Jiao, Weian Mao, Hao Chen

Organizations: Zhejiang University, China · Massachusetts Institute of Technology, USA

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.

Figures & tables

Figure 1: Overview of scTREAT. Top: Expression is encoded in a shared autoencoder latent space; the denoising path summarizes neural noise prediction and reverse-sampling updates to $\hat{z}_{0}$ , followed by decoding to $\hat{x}$ . Bottom: The matched-control representation $z_{\text{pre}}{=}E(x_{\text{pre}})$ and dose-fused drug vector $\tilde{z}_{\text{drug}}$ from Morgan fingerprints + MLP ( Rogers & Hahn, 2010 ; Landrum et al., ) form a condition token $c$ injected through GD-Attn ( Section 3.2 ).

Figure 2: Conditional denoising from $z_{T}$ . Three noise predictions from a four-state-trained denoiser are combined with separate state/drug guidance for each reverse-sampler update to $z_{t-1}$ . Decoding the final $\hat{z}_{0}$ gives the predicted expression.

Model	logFC-P	logFC-S	$\Delta$ -P	$\Delta$ -S	$EV_{\mathrm{median}}$
scTREAT	0.81	0.69	0.55	0.51	0.73
chemCPA	0.70	0.46	0.59	0.44	0.70
PRnet	0.61	0.58	0.35	0.27	0.51
STATE	0.77	0.69	0.50	0.50	0.69
scGPT	0.46	0.39	0.39	0.35	0.41
Linear	0.50	0.51	0.48	0.41	0.20

Table 1: UC results. Higher is better; bold and underline mark the best and second-best distinct values, with ties sharing formatting. Correlations and median explained variance (

EV

) are reported for all genes and the Top-200 genes by observed absolute logFC (DEG). P and S denote Pearson and Spearman correlations. Overlap@200 and Overlap@1000 denote the corresponding DEG-accuracy metrics.

Model	logFC-P	logFC-S	$\Delta$ -P	$\Delta$ -S	$EV_{\mathrm{median}}$
scTREAT	0.67	0.58	0.50	0.49	0.66
chemCPA	0.42	0.44	0.53	0.43	0.59
PRnet	0.59	0.48	0.33	0.25	0.42
Linear	0.43	0.49	0.46	0.40	0.18
MLP	0.40	0.36	0.43	0.34	0.32
$\overline{\mathrm{Context}}$	0.31	0.32	0.42	0.39	0.40

Table 2: UD results. Methods without unseen-drug support are omitted. Higher is better; bold and underline mark the best and second-best distinct values, with ties sharing formatting. Metric abbreviations and the DEG subset follow Table 1 .

Figure 3: Ablations on UC (a) and UD (b). Conditioning, GD-Attn, and separate guidance support predictive performance; molecular structure supports UD generalization; dose-dependent guidance improves selected metrics. Variants: (A) no conditioning; (B) GD-Attn replaced by Concat+MLP; (C) molecular structure replaced by Drug-ID; (D) single-channel CFG; (E) constant drug-guidance strength; (F) concatenation of condition and latent tokens before attention. Full configurations are in Appendix F .

Figure 4: Qualitative response prediction. Predicted profiles separate by cell line in UMAP space (left), while predicted and observed gene-wise logFC distributions agree for representative cell lines (right).

Panel of Figure 5 — Figure 5: Biological context. Left: KRAS GSEA curves share a broad shape but differ in their dominant signed deviations. Right: drug representations show MoA-associated grouping with overlap between some classes.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Exploratory GSEA comparison for HALLMARK_PI3K_AKT_MTOR_SIGNALING. Predicted and observed curves share broad shape features but differ in amplitude and dominant signed deviation.

Single-Cell Perturbation Modeling

Single-cell perturbation modeling predicts how genetic or chemical interventions change cellular states across cells and contexts.

See all 61 papers

scDEFT: A deep learning framework for drug-effect prediction and counterfactual reasoning
Sep 9, 2026Murthy DevarakondaSingle-Cell Perturbation Modeling Clinical Outcome Prediction
Causal Data AI (causaldata.ai)
PerturbRx: Learning Treatment-Conditioned Latent Transitions for Patient Drug Response Prediction
Aug 21, 2026Yoshitaka Inoue, Minoh Jeong, Alfred Hero +2Cancer Drug Response Prediction Predictive Representation Learning
Department of Computer Science, University of Minnesota, Minneapolis, MN, USA · Computational Biology Branch, National Library of Medicine, Bethesda, MD, USA · Therapeutic Development Branch, National Library of Medicine, Bethesda, MD, USA +2
Differentiable Retrieval-Augmented Generation for Predicting Cellular Responses to Gene Perturbation
Mar 7, 2026Andrea Giuseppe Di Francesco, Andrea Rubbi, Rishabh Jain +1Single-Cell Perturbation Modeling Retrieval-Augmented Generation
University of Cambridge Wellcome Sanger Institute

Sep 9, 2026J/K move · Enter open · S save

scDEFT: A deep learning framework for drug-effect prediction and counterfactual reasoning

Murthy Devarakonda

Causal Data AI (causaldata.ai)

Single-Cell Perturbation Modeling Clinical Outcome Prediction Counterfactual Learning Single-Cell Transcriptomics

Open paper AskPDF arXiv

Longitudinal single cell atlases now capture matched pre treatment and post treatment states from responders and non responders, presenting an opportunity to mechanistically explain why two patients on the same drug diverge. We introduce scDEFT (single cell Drug EFfect Transducer), which treats a drug as a conditioning operator on cell representations, enabling prediction and explanation. In scDEFT, feature wise linear modulation produces drug conditioned cell latents, learned under abundant per cell supervision and then frozen. Two independent heads aggregate those latents over shared transcriptional neighborhoods to predict drug induced state change and responder status. A backward stage ranks the latent dimensions by how strongly they separate responders from non responders and maps them to genes under a cell composition control. On a harmonized inflammatory bowel disease atlas of 1.16 million cells, three cohorts and two drug classes, scDEFT predicts state change at 45% of the baseline to reproducibility ceiling headroom and stratifies responders before treatment at AUROC 0.70, where standard predictors remain at chance. These predictions and the drivers behind them support target and co target nomination, patient stratification, and counterfactual prediction of unseen drug cohort effects.

Abstract

Single-Cell Perturbation Modeling

Explore similar work

scDEFT: A deep learning framework for drug-effect prediction and counterfactual reasoning

PerturbRx: Learning Treatment-Conditioned Latent Transitions for Patient Drug Response Prediction

Differentiable Retrieval-Augmented Generation for Predicting Cellular Responses to Gene Perturbation