Excited to presenting the results of my internship at insitro on modeling pooled CRISPR screening data tomorrow at the #NeurIPS2024 AI for New Drug Modalities workshop (4:30-5:10pm, West Meeting Room 109/110)! Paper: arxiv.org/abs/2411.08072 (1/n) - ThreadSky

ethanweinberger.bsky.social • 74 days ago

Excited to presenting the results of my internship at insitro on modeling pooled CRISPR screening data tomorrow at the #NeurIPS2024 AI for New Drug Modalities workshop (4:30-5:10pm, West Meeting Room 109/110)!

Paper: https://arxiv.org/abs/2411.08072

(1/n)

Comments

ethanweinberger.bsky.social•74 days ago

Analyzing pooled genetic screening data (e.g. from Perturb-seq) is known to be challenging, due to confounding sources of variation shared with control cells (e.g. cell cycle phase) that may obscure novel perturbation-induced phenotypes (e.g. new gene programs).

(2/n)

ethanweinberger.bsky.social•74 days ago

For example, in data from Norman et al., 2019 (https://science.org/doi/10.1126/science.aax4438), 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)

ethanweinberger.bsky.social•74 days ago

Previous work (e.g. ContrastiveVI; https://nature.com/articles/s41592-023-01955-3) has shown that such issues can be mitigated by learning representations that explicitly disentangle confounding variations shared with controls from "salient" variations uniquely present in perturbed cells.

(4/n)

ethanweinberger.bsky.social•74 days ago

Yet, while ContrastiveVI does lead to better recovery of perturbation-induced variations, it leaves a major source of noise unaccounted for.

(5/n)

ethanweinberger.bsky.social•74 days ago

In particular, the original model does not attempt to account for variable guide efficiency (i.e., some cells being labeled with a perturbation but not actually experiencing the corresponding perturbation's effects).

(6/n)

ethanweinberger.bsky.social•74 days ago

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)

Comments

Posting Rules

Reply