ExPert infers causal perturbations from cell states across biological and clinical contexts

ExPert infers which perturbation produced an observed cell state, learning 1 prototype per perturbation across 12 contexts from 8,964 perturbations and 19.5 million cells.

• Why it matters: Existing pooled single-cell CRISPR models predict responses forward, but the inverse problem—identifying the perturbation behind a measured state—remains difficult.
• What they did: ExPert (Extracting Perturbations) embeds cell states into a shared latent space, ranks 19,895 candidate genes per query, and is trained on perturbations with over half the candidates absent from training.
• The result: It more than doubled held-out-context recall over the best method with 1,000-fold less compute, transfers without retraining to drugs and combinations, and in alcohol-associated liver disease and primary mitochondrial disease enabled experimentally supported driver identification.

AI summary of the abstract · may contain errors · not medical advice · read the paper.