scPPDM: A Diffusion Model for Single-Cell Drug-Response Prediction

TL;DR

scPPDM introduces a diffusion-based framework for predicting single-cell drug responses directly from scRNA-seq by operating in a shared latent space with dual-condition channels (pre-perturbation state and drug dose). It employs a non-concatenative GD-Attn mechanism to fuse state and drug information, plus four-state training and factorized guidance that maps dose to guidance strength, enabling transparent what-if analyses. Across Tahoe-100M UC/UD splits, it achieves state-of-the-art performance on logFC recovery, correlation, explained variance, and DE-overlap, with notable improvements in unseen-drug generalization. The work enables interpretable, dose-controlled in-silico perturbation predictions that can reduce experimental burden and support rapid compound scoring for drug discovery and precision medicine.

Abstract

This paper introduces the Single-Cell Perturbation Prediction Diffusion Model (scPPDM), the first diffusion-based framework for single-cell drug-response prediction from scRNA-seq data. scPPDM couples two condition channels, pre-perturbation state and drug with dose, in a unified latent space via non-concatenative GD-Attn. During inference, factorized classifier-free guidance exposes two interpretable controls for state preservation and drug-response strength and maps dose to guidance magnitude for tunable intensity. Evaluated on the Tahoe-100M benchmark under two stringent regimes, unseen covariate combinations (UC) and unseen drugs (UD), scPPDM sets new state-of-the-art results across log fold-change recovery, delta correlations, explained variance, and DE-overlap. Representative gains include +36.11%/+34.21% on DEG logFC-Spearman/Pearson in UD over the second-best model. This control interface enables transparent what-if analyses and dose tuning, reducing experimental burden while preserving biological specificity.

Figures (4)

• Figure 1: Overview of scPPDM. Top: In a shared VAE latent space, VP diffusion denoises $z_t$ (time-embedded) to $\hat{z}_0$ and decodes to $\hat{x}$. Bottom: Conditions comprise $z_{\text{pre}}{=}E(x_{\text{pre}})$ and a dose-fused drug vector $\tilde{z}_{\text{drug}}$ from Morgan fingerprints + MLP Rogers2010ECFPLandrumRDKit. They are fused into a token $c$ and injected non-concatenatively via GD-Attn (\ref{['sec.3.2']}), yielding stable training and interpretable state/drug control.
• Figure 2: Decomposable, dose-controlled inference. At denoising step $t$, the model produces $\epsilon$ under both/no-pre/no-drug and forms $\hat{\epsilon}_t=(1+s_p+s_d)\epsilon_{\text{both}}-s_p\epsilon_{\text{no-pre}}-s_d\epsilon_{\text{no-drug}}$, where $s_d$ is mapped from dose. The combined guidance updates $z_t\!\to\!z_{t-1}$; decoding $D(\hat{z}_0)$ gives the predicted post-perturbation $\hat{x}$.
• Figure 3: Visualization of predictions. Left: UMAP of model-predicted post-perturbation transcriptomes colored by cell line, showing that clusters align with cell-line identity. Right: Violin plots for CVCL-0069/1097/1715 showing gene-wise logFC distributions (Predicted vs. Truth), indicating strong agreement in direction and magnitude.
• Figure 4: Ablations on UC (a) and UD (b). We evaluate five factors (A–E; see text). Trends are consistent: conditioning is required, GD-Attn (\ref{['sec.3.2']}) and dual-knob guidance improve performance, structure priors enable UD generalization, and mapping dose$\!\rightarrow\!s_d$ preserves cross-dose monotonicity.