Amortized Bayesian Causal Discovery of Extended Factor Graphs

Yichen Gu, Yuxuan Song, Weizhou Qian, Yixin Wang, Joshua WelchSubmission to ICLR 2026 · Rejected

Abstract

Learning a causal graph from interventional data is a challenging problem with broad applications. For example, a key goal in molecular biology is to identify gene regulatory networks from large-scale perturbation data. An ideal approach for this task should scale to thousands of nodes, incorporate interventions even when their targets are unknown, quantify uncertainty, and offer identifiability guarantees. Previous approaches using score-based optimization or approximate Bayesian inference do not meet all of these criteria. To address these limitations, we developed Amortized Bayesian Causal Discovery of Extended Factor Graphs (ABCDEFG). Our method guarantees exact acyclicity; scales to graphs with thousands of nodes; and incorporates interventions with known or unknown targets. Additionally, ABCDEFG estimates a posterior distribution whose maximum identifies the true causal graph up to an equivalence class. ABCDEFG achieves state-of-the-art accuracy on simulated datasets, estimating a well-calibrated posterior distribution while outperforming previous score-based and approximate Bayesian methods. Our approach also identifies known and novel genes targeted by growth factors in large-scale single-cell perturbation data.
BibTeX
@misc{gu2026amortized,
  author = {Yichen Gu and Yuxuan Song and Weizhou Qian and Yixin Wang and Joshua Welch},
  title = {{Amortized Bayesian Causal Discovery of Extended Factor Graphs}},
  year = {2026},
  eprint = {2607.22934},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.22934},
}

Loading related content...

Figure