📊 statistics

Amortized Bayesian Causal Discovery of Extended Factor Graphs

This paper introduces ABCDEFG, a scalable amortized Bayesian method that guarantees exact acyclicity and handles unknown intervention targets to accurately discover causal graphs and quantify uncertainty in large-scale datasets, such as gene regulatory networks.

Original authors: Yichen Gu, Yuxuan Song, Weizhou Qian, Yixin Wang, Joshua Welch

Published 2026-07-28

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Yichen Gu, Yuxuan Song, Weizhou Qian, Yixin Wang, Joshua Welch

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a massive mystery: how does a complex system work? In the world of biology, this system is a cell, and the "suspects" are thousands of genes that talk to each other. Sometimes, scientists can just watch the cell naturally (observational data), but to really understand the cause-and-effect, they need to poke it. They might turn a gene on or off, or add a drug, and see what happens (interventional data). The goal is to draw a map, called a causal graph, showing exactly who influences whom. However, drawing this map is incredibly hard. The system is huge, the data is noisy, and sometimes the scientists don't even know exactly which gene they poked, only that they poked something. Existing methods are like detectives who either get lost in the crowd because there are too many people, or they guess the map without being able to say how sure they are.

This paper introduces a new detective tool called ABCDEFG (Amortized Bayesian Causal Discovery of Extended Factor Graphs). Think of ABCDEFG as a super-smart, high-tech mapmaker that doesn't just guess the connections; it builds a map that is mathematically guaranteed to have no time-travel loops (acyclic), can handle thousands of suspects at once, and is excellent at figuring out who was poked even when the detective didn't see the poke clearly. It uses a clever trick called "extended factor graphs," which is like organizing the suspects into small, manageable teams (factors) to make the puzzle easier to solve. The researchers tested this tool on simulated data and found it was more accurate than previous methods, especially when the relationships were complex or the targets were unknown. They even tried it on real data from thousands of cells, successfully finding known gene relationships and discovering new ones.

The Detective's Dilemma: Mapping the Cell

To understand why ABCDEFG is a big deal, let's look at the problem it's solving. Imagine a giant, chaotic party where thousands of people (genes) are chatting. You want to know who is influencing whom. If you just stand in the corner and listen (observational data), you might see two people laughing together, but you don't know if one told a joke to the other, or if they both heard a funny noise from the DJ. To find the truth, you need to intervene. You might whisper a secret to one person and see how the conversation changes.

In biology, scientists do this by using "perturbations"—like turning a gene off or adding a drug. But here's the catch: in a real lab, you might have 30,000 genes and millions of cells. Old methods of mapping these relationships are like trying to draw a map of a whole city by hand while running a marathon; they get too slow, or they make mistakes like drawing a road that leads back to where it started (a cycle), which is impossible in a cause-and-effect chain. Furthermore, many methods can't handle the "unknowns." What if you added a drug, but you don't know exactly which gene it hit first? Old tools often get confused or fail completely in these situations.

The ABCDEFG Solution: The Team-Based Map

The authors of this paper built a new method called ABCDEFG. Instead of trying to draw a direct line between every single gene and every other gene (which is a mess), ABCDEFG uses a concept called Extended Factor Graphs.

Imagine the genes are players in a massive game of tag. Instead of trying to track every single tag between every pair of players, ABCDEFG groups the players into "teams" (called factors). The genes interact with these teams, and the teams interact with each other. This is a bit like realizing that instead of 10,000 people talking to 10,000 people, there are actually 10 distinct conversation circles, and people are just jumping between them. By focusing on these "teams," the math becomes much simpler and faster.

The "Extended" part of the name is the secret sauce for handling the unknowns. When a scientist adds a drug but doesn't know the exact target, ABCDEFG treats the drug as a special "intervention node" that connects to these teams. It doesn't need to know the exact gene the drug hit; it just needs to know which "team" the drug influenced. This allows the system to figure out the hidden targets while simultaneously drawing the map of the genes.

How It Works: The "No-Loop" Guarantee

One of the biggest headaches in causal discovery is avoiding loops. In a real cause-and-effect chain, A causes B, and B causes C. But A cannot cause C and then C causes A; that's a time-travel paradox. Many computer algorithms try to fix this by adding a "penalty" if they accidentally draw a loop, but this is like trying to stop a car by gently tapping the brakes while driving at 100 mph—it often fails.

ABCDEFG is different. It builds the map in a way that makes loops impossible by design. It organizes the genes and teams into a specific order, like stacking blocks from bottom to top. You can only connect a block to the ones below it. This means the computer never has to waste time checking for loops because the structure itself guarantees the map is a straight line of cause and effect.

The Results: Simulations and Real Cells

The researchers tested ABCDEFG in two ways. First, they created fake data (simulations) where they knew the exact answer. They pitted ABCDEFG against the best existing tools. In these tests, ABCDEFG consistently found the correct map more often, especially when the data was messy or when the intervention targets were unknown. It also did a great job of telling the user how confident it was in its answers, which is crucial when dealing with noisy biological data.

Then, they took it to the real world. They used data from a massive experiment involving 31,475 cells and 1,000 genes, where the cells were treated with 46 different combinations of growth factors (molecules that tell cells what to do). The goal was to see which growth factors influenced which genes. ABCDEFG successfully identified known relationships and even found new ones that other methods missed. It also did a better job of predicting what would happen in new, unseen experiments compared to the other tools.

What It Doesn't Do (Yet)

While ABCDEFG is a powerful new tool, the authors are careful to point out its limits. It assumes the relationships form a straight line without loops, but in real biology, some genes do form feedback loops (where A influences B, and B influences A). If the real world has these loops, ABCDEFG might not capture them perfectly. Also, the method relies on the idea that the genes can be grouped into a small number of "teams" (factors). If the biological system is too complex to be broken down this way, the map might be a bit fuzzy. Finally, while the math proves the method works under certain conditions, the authors note that they haven't fully explored how it behaves when there is very little data available.

In short, ABCDEFG is a major step forward for scientists trying to map the complex machinery of life. It offers a way to handle huge datasets, deal with unknown targets, and provide a clear, loop-free map of how genes influence one another, all while telling us how sure we can be about the results. It's not a magic wand that solves every biological mystery, but it's a very sharp new tool for the detective's kit.

Technical Summary: Amortized Bayesian Causal Discovery of Extended Factor Graphs (ABCDEFG)

Problem Statement

Learning causal graphs from interventional data is a critical challenge in domains such as molecular biology, particularly for uncovering gene regulatory networks from large-scale perturbation data. Existing methods face a "trilemma" where they typically fail to satisfy three simultaneous requirements:

Scalability: Handling thousands of nodes (e.g., >20,000 genes) and millions of samples.
Intervention Handling: Incorporating interventions even when their specific targets are unknown (e.g., drug treatments).
Uncertainty Quantification: Providing a posterior distribution to quantify uncertainty, which is vital for noisy biological data.

Current score-based methods often lack uncertainty quantification and struggle with unknown intervention targets, while approximate Bayesian methods generally fail to scale to the size of modern single-cell datasets. Furthermore, many existing approaches do not guarantee acyclicity or provide strong identifiability guarantees for the recovered graph.

Methodology: ABCDEFG

The authors propose Amortized Bayesian Causal Discovery of Extended Factor Graphs (ABCDEFG), a framework that integrates variational inference with a novel graph representation to address the aforementioned limitations.

1. Extended Factor Graphs (f-DAGs)

The core innovation is the representation of causal structures using Extended Factor Directed Acyclic Graphs (f-DAGs).

Structure: An f-DAG consists of variable nodes ( $V$ ), factor nodes ( $F$ ), and edges only between variables and factors. This structure allows the adjacency matrix of the causal graph to be viewed as a Boolean factorization ($A = UV$), effectively acting as a low-rank approximation of the full causal DAG.
Extensions for Unknown Targets: To handle unknown intervention targets, the authors introduce Extended f-DAGs, which include intervention nodes ( $\Xi$ ) connected to factor nodes. This allows the model to learn which factors are affected by an intervention without knowing the specific downstream gene targets a priori.
Acyclicity Guarantee: Unlike differentiable causal discovery methods (e.g., NOTEARS) that use penalty terms to enforce acyclicity, ABCDEFG constructs graphs that are acyclic by construction. This is achieved by establishing a total order of factors and inserting nodes and interventions into partitions defined by this order.

2. Generative Model and Variational Inference

The method employs a Bayesian framework to estimate the posterior distribution over causal graphs, $q(G; \Lambda)$ .

Parametric Models: The variational distribution $q(G; \Lambda)$ is parameterized using either independent Bernoulli random variables (via Gumbel-Softmax) or a Sum-Product Network (SPN). The SPN allows for modeling complex, correlated joint distributions over edge probabilities, a feature absent in methods assuming edge independence.
Amortized Inference: The framework utilizes a Variational Autoencoder (VAE) architecture. Instead of learning separate encoders/decoders for each variable, it uses a single feed-forward neural network to amortize the conditional distributions. Causal relations are injected via masking operations (Hadamard products) using the sampled node-to-factor ( $U$ ), factor-to-node ( $V$ ), and intervention-to-factor ( $W$ ) matrices.
Objective: The method optimizes the Evidence Lower Bound (ELBO), which balances the likelihood of the data under the interventional model against the KL divergence from a prior. The objective is:
$\mathcal{L}(q) = \sum_{k=1}^{n_I} \mathbb{E}_{p^{(k)}(X|G^*)} \left[ \mathbb{E}_{q(G;\Lambda)} [\log p^{(k)}_\Phi(X|G)] \right] - \text{KL}(q(G;\Lambda) || p(G))$

3. Identifiability Guarantees

The paper provides theoretical guarantees regarding the identifiability of the recovered graph:

I-Markov Equivalence: The authors prove that the Maximum A Posteriori (MAP) estimate of the causal graph is I-Markov equivalent to the true causal DAG. This means the method recovers the true graph up to the equivalence class defined by the set of interventions.
f-DAG Identifiability: Under specific structural conditions (e.g., sufficient v-structures and unique parent/child relationships for factors), the method guarantees that the underlying f-DAG (and thus the factorization) is identifiable up to a permutation of factors.

Key Contributions

Novel Parametric Model: Introduction of a generative process for sampling acyclic extended factor graphs that explicitly models intervention nodes and unknown targets.
Variational Framework: Development of a variational Bayesian approach for causal discovery that handles both known and unknown intervention targets while scaling to large graphs.
Correlated Edge Modeling: Integration of Sum-Product Networks (SPNs) to model complex joint distributions over causal edges, moving beyond the assumption of independent edges.
Theoretical Guarantees: Proof that the MAP estimate recovers the true causal graph up to an I-Markov equivalence class, and conditions under which the underlying f-DAG is identifiable.
Empirical Validation: Demonstration of state-of-the-art performance on simulated datasets and real-world single-cell perturbation screens.

Experimental Results

Simulation Studies

The authors evaluated ABCDEFG on 24 simulated datasets involving linear/non-linear effects, independent/joint edge probabilities, and hard/soft interventions.

Performance: ABCDEFG achieved state-of-the-art accuracy (measured by Structural Hamming Distance and F1 score), particularly on graphs with non-linear causal effects and jointly distributed edge probabilities (modeled by SPNs).
Unknown Targets: In scenarios with unknown intervention targets, ABCDEFG successfully inferred both the causal graph and the intervention targets, outperforming methods that could not incorporate unknown targets or failed to scale.
Bayesian Comparison: Compared to other Bayesian methods (BaCaDi, DECI, ProDAG), ABCDEFG demonstrated superior F1 scores and lower SHD, while also providing better-calibrated posterior distributions. It was also significantly faster than other Bayesian approaches.

Real-World Application

The method was applied to a large-scale single-cell perturbation screen involving 34,469 genes and 31,475 cells treated with 46 combinations of 14 growth factors.

Gene Regulatory Networks: ABCDEFG successfully identified known and novel gene targets of growth factors.
Target Recall: The method achieved a recall of 0.325 (Basic) and 0.376 (SPN) for growth factor-to-gene interactions, significantly outperforming a random baseline (0.196).
Data Reconstruction: On held-out intervention combinations, ABCDEFG achieved lower Mean Squared Error (MSE) in data reconstruction compared to DCDFG and SDCD.

Significance and Claims

The paper claims that ABCDEFG fills a critical gap in causal discovery by enabling scalable, uncertainty-aware Bayesian inference from interventional data with unknown targets.

Scalability: The method scales to graphs with thousands of nodes (demonstrated up to 1,000 nodes in simulations and 1,000 highly variable genes in real data), overcoming the limitations of previous Bayesian approaches.
Robustness: By using extended factor graphs, the method offers robustness to noise and provides a structured way to model complex dependencies via SPNs.
Practical Utility: The application to single-cell data demonstrates the method's ability to recover biologically meaningful gene regulatory networks and perturbation targets, offering a tool for understanding cellular development and homeostasis.

The authors acknowledge limitations, noting that the acyclicity assumption may not hold for all gene regulatory networks (which can contain cycles) and that the low-rank f-DAG approximation may be insufficient for very high-rank true graphs. However, they posit that the framework provides a promising foundation for future extensions into limited data regimes and more complex biological systems.