Finally, we apply DeepDive to islets (Sander lab), and dive into beta-cell subtypes linked to T2D. We id TFs differentially active between subtypes and after virtually perturbing disease status. TFs with counterfactual support more often have genetic and functional support! 4/6

Next, we turn counter-factual prediction and attempt to use the disentangled representations to predict the epigenome in an unseen covariate combination by virtually perturbing the observed covariates in real data. DeepDive outperforms baselines and achieves high accuracy. 3/6

To evaluate disentanglement, we analyzed collinear covariates in both simulated and real datasets. We found that regular statistical methods fail, as signal from one covariate leaks into another, but DeepDive accurately resolves which signal belongs to which covariate. 2/6

Excited to share DeepDive, a generative model for disentangling covariate effects in snATAC-seq developed by the amazing @andreasfm.bsky.social. DeepDive enables deeper understanding of a dataset by separating the effects of known and unknown covariates. Here are some highlights. 1/6