2023-09-13

Mapping perturbations across contexts

Presentation by Izzy Grabski

In scRNA-seq, we want to learn differential expression between two biological conditions. What are the causal drivers as opposed to secondary correlated genes? Genetic perturbation screens can help: knock down using CRISPR and understand the effects of knocked down genes: deconvolute the differences in terms of perturbations.

\(\mathbf{D}_0\): baseline condition. \(\mathbf{D}_1\): list of perturbations that have taken place.

\[\mathbf{Y}_0 = \Phi f_0 + \mathbf{E}_0\] \[\mathbf{Y}_k = \Phi f_k + \Lambda_k \mathbf{I}_k + \mathbf{E}_k\]

CPA-Perturb-seq: perturb the same locations across multiple cells. The target is to find which cells are perturbed. This can be done in different cell lines (e.g. HEK and K562). The idea is to learn the \(\Lambda\) for different cell lines and perform regression,

• What is the relation between \(\mathbf{D}\) and \(\mathbf{Y}\)?
• Is the noise Gaussian in the gene expression model?
• Instead of constraining \(\delta\), what happens if we try learning \(\Lambda\) from \(\mathbf{Y}_k - \mathbf{Y}_0\)?