scGen documentation

scGen Architecture

Introduction

scGen is a generative model to predict single-cell perturbation response across cell types, studies and species (Nature Methods, 2019). scGen is implemented using the scvi-tools framework.

Getting Started

What you can do with scGen:

  • Train on a dataset wih multiple cell types and conditions and predict the the perturbation effect on the cell type which you only have in one condition. This scenario can be extended to multiple species where you want to predict the effect of a specific species using another or all the species.

  • Train on a dataset where you have two conditions (e.g. control and perturbed) and predict on second dataset with similar genes.

  • Remove batch effect on labeled data. In this scenario you need to provide cell_type and batch labels to the method. Note that batch_removal does not require all cell types to be present in all datasets (batches). If you have dataset specific cell type it will preserved as before.

  • We assume there exist two conditions in you dataset (e.g. control and perturbed). You can train the model and with your data and predict the perturbation for the cell type/species of interest.

  • We recommend to use normalized data for the training. A simple example for normalization pipeline using scanpy:

  • We further recommend to use highly variable genes (HVG). For the most examples in the paper we used top ~7000 HVG. However, this is optional and highly depend on your application and computational power.

Reproducing paper results

In order to reproduce paper results visit here.

References

Lotfollahi, Mohammad and Wolf, F. Alexander and Theis, Fabian J. “scGen predicts single-cell perturbation responses.” Nature Methods, 2019 pdf.