Exploring Perturbation Effects on Transcriptional Dynamics with ContrastiveBiVI

Ashenafee Anteneh Mandefro, Ethan Weinberger, Addie Woicik, Tal Ashuach

Confirmation: I have read and agree with the workshop's policy on behalf of myself and my co-authors.

Track: long paper (4–8 pages excluding references)

Keywords: single-cell RNA sequencing (scRNA-seq), CRISPR perturbations (Perturb-seq), contrastive latent variable models (cLVMs), biophysical modelling, generative modelling, transcriptional dynamics

TL;DR: ContrastiveBiVI disentangles perturbation-specific transcriptional kinetics in single-cell CRISPR screens by combining contrastive latent variable models with biophysical models of transcription.

Abstract: CRISPR-Cas9-based genetic screens combined with single-cell transcriptomic profiling have emerged as a powerful tool for investigating the relationships between genotype and cellular phenotypes. However, most computational analyses of these screens solely focus on comparisons of observed measurements of mature mRNA abundance between groups of cells, and ignore the underlying biophysical phenomena (e.g., mRNA splicing dynamics) that lead to the observed data. Towards explicitly modelling such phenomena in single-cell RNA sequencing data, a recent line of work has proposed simultaneously considering nascent and mature mRNA count levels and relating the two via biophysically plausible models of transcription formalized by chemical master equations. Yet, applying these models directly to perturbation screening datasets is not straightforward, as perturbation-induced variations of interest are often confounded by uninteresting "background" variation shared with control cells. To remedy this issue, here we propose ContrastiveBiVI, a generative model that combines ideas from the biophysical modelling literature with so-called contrastive latent variable models, which explicitly disentangle perturbation-induced variations from non-perturbation-related variations into separate sets of latent factors. Applied to a publicly available CRISPR activation dataset, we find that our method successfully recovers perturbation-induced variations and facilitates the exploration of perturbation-induced changes in transcription and splicing kinetics. An open-source implementation of our model is available at https://github.com/insitro/contrastive_bivi.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Submission Number: 76