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

Molecular-science workflows depend on predictive and generative models that transfer across unseen compounds and cellular contexts. We introduce Single-Cell Transcriptomic Response Estimation After Treatment (scTREAT), a latent diffusion model for single-cell drug-response prediction. scTREAT represents pre-perturbation state and dose-conditioned molecular structure as separate channels, injects them through GD-Attn, and uses four-state training for factorized guidance. On Tahoe-100M, scTREAT achieves the best value on 10 of 12 metrics for unseen covariate combinations (UC) and 11 of 12 metrics for unseen drugs (UD), including +36.11%/+34.21% gains in DEG logFC-Spearman/Pearson on UD over the second-best model. Targeted ablations isolate the contributions of molecular conditioning, GD-Attn, factorized guidance, and dose calibration, providing a well-benchmarked generative component for molecular-science workflows.