ABCDEFG: New AI method discovers causal gene networks from interventional data

Learning causal relationships from interventional data is a cornerstone of molecular biology—especially for reconstructing gene regulatory networks from large-scale perturbation experiments. Existing approaches, such as score-based optimization or approximate Bayesian inference, often fail to scale to thousands of variables, cannot handle unknown intervention targets, or lack uncertainty quantification. To overcome these limitations, a team from Microsoft Research and the University of Michigan introduces ABCDEFG (Amortized Bayesian Causal Discovery of Extended Factor Graphs). The method amortizes inference across many graph structures using a flexible neural architecture, allowing it to scale to graphs with thousands of nodes while maintaining exact acyclicity—a key constraint in directed acyclic graphs. It also naturally incorporates interventions whose targets are unknown, which is common in biological experiments. Critically, ABCDEFG produces a full posterior distribution over causal graphs, and its maximum a posteriori estimate provably identifies the true causal graph up to its Markov equivalence class.

On synthetic benchmarks, ABCDEFG outperforms previous score-based and approximate Bayesian methods in both accuracy and calibration of the posterior. When applied to real-world single-cell perturbation data (e.g., from CRISPR screens), the model successfully identified both established and novel gene targets regulated by growth factors, demonstrating its practical utility for discovery biology. The authors note that the method’s ability to handle unknown intervention targets and quantify uncertainty makes it particularly suitable for high-throughput biological assays where perturbation effects are noisy or only partially observed. With its combination of scalability, theoretical guarantees, and empirical performance, ABCDEFG represents a significant step forward for causal discovery in genomics and beyond.