Causal Structure Learning Proves Unreliable on Experimental Gene Regulatory Network Data

A benchmark evaluation indicates that existing causal structure learning algorithms face substantial practical limitations when inferring gene regulatory networks from experimental perturbation data, despite performing well under ideal synthetic conditions. Evaluated against synthetic single-cell perturbation benchmarks and experimental data from a clustered regularly interspaced short palindromic repeats screen, the algorithms demonstrated a clear gap between their theoretical potential and their real-world application in cellular biology.

Causal structure learning has emerged as a promising approach to studying cellular gene regulation, aiming to provide deeper mechanistic insights than purely association-based methods. The deployment of these techniques is motivated by theoretical groundwork establishing that interventions improve the identifiability of causal structures, prompting researchers to apply causal structure learning algorithms to scenarios containing interventional information.

To assess how effectively existing algorithms leverage the information revealed by targeted interventions, the benchmark evaluated a suite of causal structure learning algorithms across distinct metrics tailored to synthetic ground truth and real biological data. On synthetic single-cell perturbation data, the algorithms achieved accurate recovery of gene regulatory networks under favourable benchmark conditions characterized by strong interventions, large sample sizes, and low measurement noise.

In contrast, algorithm performance on experimental CRISPR screen data remained unreliable, as no reliable ground truth was available and the inferred networks had to be assessed indirectly using observed perturbation effects and curated regulatory annotations. Technical noise, biological noise, and algorithmic scalability emerged as primary limiting factors, while the benchmark provides foundational insight along with an accessible software package to apply modern causal structure learning to custom datasets and foster future methodological development.

The investigation focuses on the premise that causal structure learning can deliver deeper mechanistic insights into gene regulation in cells than purely association-based methods. Because theoretical groundwork indicates that interventions improve the identifiability of causal structure, the study examines whether existing algorithms successfully leverage interventional information in practice. By systematically evaluating a suite of causal structure learning algorithms across distinct testing scenarios, the researchers assessed how effectively these computational tools translate theoretical identifiability into accurate gene regulatory network inference.

A central component of the benchmark design is the separation of evaluation metrics tailored specifically to synthetic ground truth and real biological data. When tested on synthetic single-cell perturbation data where ground truth was fully known, the suite of algorithms proved capable of achieving accurate recovery of gene regulatory networks. However, this success depended on highly favourable benchmark conditions, requiring strong interventions, large sample sizes, and low measurement noise to reliably learn the underlying regulatory structures.

Moving beyond synthetic settings to experimental clustered regularly interspaced short palindromic repeats screen data introduced severe evaluation challenges. Because no reliable ground truth exists for real biological systems, the researchers evaluated the inferred gene regulatory networks indirectly by measuring them against observed perturbation effects and curated regulatory annotations. Under these real experimental conditions, algorithm performance remained unreliable, with technical noise, biological noise, and algorithmic scalability serving as the critical limiting factors that hinder practical gene network inference.

These findings illustrate an evident gap between the theoretical potential of causal structure learning and its practical application to cellular gene regulatory networks. By detailing where algorithms succeed and where they falter, the benchmark provides essential groundwork for future methodological development. To encourage further exploration and assist other researchers in evaluating these approaches, the authors provided an accessible software package designed to apply modern causal structure learning algorithms directly to custom datasets.

Subject of Research: Benchmarking interventional causal structure learning algorithms for gene regulatory network inference

Article Title: Comparison of interventional causal structure learning algorithms for gene regulatory network inference

Article References: Sprengel, J. L., & Velten, B. (2026). Comparison of interventional causal structure learning algorithms for gene regulatory network inference. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06728-1

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06728-1

Keywords: causal structure learning, gene regulatory networks, interventional data, CRISPR screens, single-cell perturbation, algorithm benchmark, synthetic ground truth, experimental perturbation data, measurement noise, biological noise, algorithmic scalability, perturbation effects, regulatory annotations, gene regulation, targeted interventions, causal inference software