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

Predicting responses to compounds absent from training requires connecting molecular structure to cellular effects. We present Single-Cell Transcriptomic Response Estimation After Treatment (scTREAT), a latent diffusion model conditioned on cellular background, molecular structure, and dose, with four-state training for separate state and drug guidance. On condition-level pseudobulk profiles derived from the Tahoe-100M single-cell atlas, scTREAT attains the best value on 11 of 12 metrics for unseen drugs (UD), including +36.11%/+34.21% relative gains in DEG logFC-Spearman/Pearson over the second-best model, and 10 of 12 metrics for unseen cell line–drug combinations (UC). Ablations examine the contributions of molecular conditioning and guidance. These retrospective results support response prediction for held-out drugs and motivate its study for experimental prioritization.