Large-Scale Differentiable Causal Discovery of Factor Graphs
Romain Lopez, Jan-Christian Hütter, Jonathan K. Pritchard, Aviv Regev
Abstract
A common theme in causal inference is learning causal relationships between observed variables, also known as causal discovery. This is usually a daunting task, given the large number of candidate causal graphs and the combinatorial nature of the search space. Perhaps for this reason, most research has so far focused on relatively small causal graphs, with up to hundreds of nodes. However, recent advances in fields like biology enable generating experimental data sets with thousands of interventions followed by rich profiling of thousands of variables, raising the opportunity and urgent need for large causal graph models. Here, we introduce the notion of factor directed acyclic graphs (f -DAGs) as a way to restrict the search space to non-linear low-rank causal interaction models. Combining this novel structural assumption with recent advances that bridge the gap between causal discovery and continuous optimization, we achieve causal discovery on thousands of variables. Additionally, as a model for the impact of statistical noise on this estimation procedure, we study a model of edge perturbations of the f -DAG skeleton based on random graphs and quantify the effect of such perturbations on the f -DAG rank. This theoretical analysis suggests that the set of candidate f -DAGs is much smaller than the whole DAG space and thus may be more suitable as a search space in the high-dimensional regime where the underlying skeleton is hard to assess. We propose Differentiable Causal Discovery of Factor Graphs (DCD-FG), a scalable implementation of f -DAG constrained causal discovery for high-dimensional interventional data. DCD-FG uses a Gaussian non-linear low-rank structural equation model and shows significant improvements compared to state-of-the-art methods in both simulations as well as a recent large-scale single-cell RNA sequencing data set with hundreds of genetic interventions. 10 2 10 3 10 4 Number of nodes 10 -3 10 -2 10 -1 10 0 10 1 Penalty gradient runtime (s/it) Tr-exp Power Spectral Radius Tr-exp Factor Spectral Factor 10 2 10 3 10 4 Number of nodes 10 -3 10 -2 10 -1 Likelihood gradient runtime (s/it) Linear-dense Linear-LR MLP Dense MLP Factor 10 2 10 3 10 4 Number of nodes 10 -2 10 -1 10 0 Total runtime (s/it) NOTEARS NOBEARS NOTEARS-LR DCDI DCD-FG DCD-FG (linear) A. B. C.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
19 top-tier papers have cited it since. Know what they found before your reviewers do.
Lune can draw on
the full paper · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
- DynGFN: Towards Bayesian Inference of Gene Regulatory Networks with GFlowNets cites Gradient-Based Neural DAG Learning
- On the Role of Sparsity and DAG Constraints for Learning Linear DAGs cites Gradient-Based Neural DAG Learning
- Constraint-Free Structure Learning with Smooth Acyclic Orientations cites Gradient-Based Neural DAG Learning
- Constraint-Free Structure Learning with Smooth Acyclic Orientations cites Differentiable Causal Discovery from Interventional Data
- Constraint-Free Structure Learning with Smooth Acyclic Orientations cites Beware of the Simulated DAG! Causal Discovery Benchmarks May Be Easy to Game
- DiBS: Differentiable Bayesian Structure Learning cites Gradient-Based Neural DAG Learning
- DiBS: Differentiable Bayesian Structure Learning cites On the Role of Sparsity and DAG Constraints for Learning Linear DAGs
- DiBS: Differentiable Bayesian Structure Learning cites Differentiable Causal Discovery from Interventional Data
- Stable Differentiable Causal Discovery cites Gradient-Based Neural DAG Learning
- Stable Differentiable Causal Discovery cites On the Role of Sparsity and DAG Constraints for Learning Linear DAGs
- Stable Differentiable Causal Discovery cites Differentiable Causal Discovery from Interventional Data
- Stable Differentiable Causal Discovery cites Beware of the Simulated DAG! Causal Discovery Benchmarks May Be Easy to Game
- Differentiable Causal Discovery from Interventional Data cites Gradient-Based Neural DAG Learning
- ProDAG: Projected Variational Inference for Directed Acyclic Graphs cites On the Role of Sparsity and DAG Constraints for Learning Linear DAGs
- ProDAG: Projected Variational Inference for Directed Acyclic Graphs cites DiBS: Differentiable Bayesian Structure Learning
- ProDAG: Projected Variational Inference for Directed Acyclic Graphs cites Differentiable Causal Discovery from Interventional Data
- Beware of the Simulated DAG! Causal Discovery Benchmarks May Be Easy to Game cites Gradient-Based Neural DAG Learning
- Beware of the Simulated DAG! Causal Discovery Benchmarks May Be Easy to Game cites On the Role of Sparsity and DAG Constraints for Learning Linear DAGs
- Beware of the Simulated DAG! Causal Discovery Benchmarks May Be Easy to Game cites Differentiable Causal Discovery from Interventional Data
- Interpretable Neural ODEs for Gene Regulatory Network Discovery under Perturbations cites Differentiable Causal Discovery from Interventional Data
Cited by top-tier papers19
- DynGFN: Towards Bayesian Inference of Gene Regulatory Networks with GFlowNetsLazar Atanackovic, Alexander Tong, Bo Wang, Leo J. Lee et al.NeurIPS 2023 · 37 citations
- Stable Differentiable Causal DiscoveryAchille Nazaret, Justin Hong, Elham Azizi, David M. BleiICML 2024 · 29 citations
- Constraint-Free Structure Learning with Smooth Acyclic OrientationsRiccardo Massidda, Francesco Landolfi, Martina Cinquini, Davide BacciuICLR 2024 · 10 citations
- Interpretable Neural ODEs for Gene Regulatory Network Discovery under PerturbationsZaikang Lin, Sei Chang, Aaron Zweig, Minseo Kang et al.ICML 2026 · 8 citations
- ProDAG: Projected Variational Inference for Directed Acyclic GraphsRyan Thompson, Edwin V. Bonilla, Robert KohnNeurIPS 2025 · 6 citations
Builds on7
- Gradient-Based Neural DAG LearningSébastien Lachapelle, Philippe Brouillard, Tristan Deleu, Simon Lacoste-JulienICLR 2020 · 337 citations
- On the Role of Sparsity and DAG Constraints for Learning Linear DAGsIgnavier Ng, AmirEmad Ghassami, Kun ZhangNeurIPS 2020 · 306 citations
- Differentiable Causal Discovery from Interventional DataPhilippe Brouillard, Sébastien Lachapelle, Alexandre Lacoste, Simon Lacoste-Julien et al.NeurIPS 2020 · 295 citations
- Beware of the Simulated DAG! Causal Discovery Benchmarks May Be Easy to GameAlexander G. Reisach, Christof Seiler, Sebastian WeichwaldNeurIPS 2021 · 213 citations
- DiBS: Differentiable Bayesian Structure LearningLars Lorch, Jonas Rothfuss, Bernhard Schölkopf, Andreas KrauseNeurIPS 2021 · 144 citations
Related papers
- Differentiable Cyclic Causal Discovery Under Unmeasured ConfoundersMuralikrishnna G. Sethuraman, Faramarz FekriNeurIPS 2025 · 5 citations
- PACER: Acyclic Causal Discovery from Large-scale Interventional DataRamon Viñas Torné, Sílvia Fàbregas Salazar, Soyon Park, Ivo Alexander Ban et al.ICML 2026
- Scalable and Flexible Causal Discovery with an Efficient Test for AdjacencyAlan Nawzad Amin, Andrew Gordon WilsonICML 2024 · 4 citations
- Identifying biological perturbation targets through causal differential networksMenghua Wu, Umesh Padia, Sean H. Murphy, Regina Barzilay et al.ICML 2025
- Interventional Causal Discovery in a Mixture of DAGsBurak Varici, Dmitriy Katz, Dennis Wei, Prasanna Sattigeri et al.NeurIPS 2024 · 10 citations
