Single-cell perturbation prediction near experimental reproducibility with GENESIS

GENESIS predicts single-cell perturbation response distributions from experimental covariates, approaching experimental reproducibility on independently repeated conditions across the Tahoe-100M and PBMC cytokine atlases.

• Why it matters: Exhaustively measuring all combinations of biological contexts, perturbations, doses, and experimental variables is infeasible, so models must generalize reliably to new, repeated conditions.
• What they did: The generative framework trains a lightweight neural network using the energy score and evaluates predictions by calibrating error against reproducibility from repeated measurements of the same condition.
• The result: GENESIS significantly outperforms prior generative models and statistical baselines, stays accurate with only a fraction of conditions or cells, recovers context-dependent transcriptional programs, and favors broader coverage over deeper profiling at equal training budgets.

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