Integrating multi-covariate disentanglement with counterfactual analysis on synthetic data enables cell type discovery and counterfactual predictions

Stathis Megas, Arian Amani, Antony Rose, Olli Dufva, Kian Shamsaie, Hesam Asadollahzadeh, Krzysztof Polanski, Muzlifah Haniffa, Sarah Amalia Teichmann, Mohammad Lotfollahi
June, 2025
PDF Cite Code Project DOI bioRxiv

Abstract

CellDISECT (Cell DISentangled Experts for Covariate counTerfactuals) is a powerful causal generative model that enhances single-cell analysis by disentangling variations to separate covariate variations at test time, learning to make accurate counterfactual predictions, achieving flexible fairness through expert models for each latent space, and capturing both covariate-specific information and novel biological insights. *These authors contributed equally.

Type
Preprint
Publication
bioRxiv
Deep Learning Single-Cell Genomics Generative Models Causal Representation Learning Disentanglement Counterfactuals Virtual Cell Perturbation Prediction
Arian Amani
Arian Amani
Machine Learning Scientist

I build AI systems for drug discovery and virtual cells — bridging deep generative models, single-cell biology, and production ML to accelerate therapeutic discovery.