Learning Causal Representations of Single Cells via Sparse Mechanism Shift Modeling

This repository contains an implementation of the sparse VAE framework applied to single-cell perturbation data, as descibed in "Learning Causal Representations of Single Cells via Sparse Mechanism Shift Modeling".

Overview of the sparse VAE framework applied to single-cell perturbation data. (A) Input data are gene expression profiles of cells under different genetic or chemical perturbations (colors), as well as the intervention label. (B) A schematic of the generative model, and the causal semantics of the sparse VAE (C) Three method outputs. (i) identification of target latent variables, encoded as a causal graph between the interventions and latent variables; (ii) a disentangled latent model for which individual latent variables are more likely to be interpreted as the activity of a relevant biological process; and (iii) the generalization of the generative model to unseen interventions (e.g., for latent target identification).

User guide

Installation

Download or clone this repository. Then from inside the folder simply run:

pip install -e . 

Example

An example script for the sandbox can be found in entry_points/demo.py. The code for reproducing the real data analysis can be found in entry_points/run_real_data_replogle_wandb.py.

References

@article{svae+,
  title={Learning Causal Representations of Single Cells via Sparse Mechanism Shift Modeling},
  author={Lopez, Romain and Tagasovska, Natasa and Ra, Stephen and Cho, Kyunghyun and Pritchard, Jonathan K. and Regev, Aviv },
  journal={Conference on Causal Learning and Reasoning},
  year={2023},
}