but also provides predictions as to which cells with a given perturbation label were "truly" perturbed, which may reduce noise in further downstream analyses.

(8/n)

To account for this issue, we augmented the original ContrastiveVI model with a richer mixture model prior on the salient latent space that accounts for variability in guide effect. Our new model not only can isolate perturbation-induced effects from confounders,

(7/n)

For example, in data from Norman et al., 2019 (science.org/doi/10.1126/...), we find that variations due to cell cycle (left) strongly confound standard exploratory data analysis workflows, leading to poor separation of known perturbation-induced gene programs (right).

(3/n)