CauTrigger models biological state transitions through hierarchical causal decomposition

A hierarchical deep latent-variable model called CauTrigger outperformed existing methods in identifying candidate causal triggers and predicting system-level responses to perturbations across synthetic and real-world datasets with multilayer regulatory structures. The computational study, published in Nature Methods, reports that the method can reconstruct hierarchical regulatory relationships from observational data while separating causal factors from potentially spurious associations.

Biomedical measurements often contain relationships among multiple regulatory layers, but an observed association does not necessarily identify the variable responsible for a change in system state. CauTrigger was designed to address this causal-discovery problem by combining hierarchical latent-variable modeling with do-calculus. Its recursive decomposition is intended to distinguish causal and spurious latent factors, allowing the model to represent an underlying regulatory hierarchy rather than treating all measured relationships as equivalent.

The researchers evaluated CauTrigger on two connected tasks: identifying variables that act as causal triggers in multilayer systems and predicting the effects of perturbing those variables. Synthetic datasets with known regulatory structures enabled comparisons with baseline methods, while real single-cell multi-omics and gut microbiome datasets tested the framework in biological applications. Across these evaluations, CauTrigger reported stronger performance than existing approaches for both causal-trigger identification and perturbation-effect prediction.

In an application to single-cell multi-omics data from type 2 diabetes, the framework dissected regulation involving transcription factors, cyclic AMP response elements and target genes, described as TF–CRE–TG regulation. CauTrigger prioritized candidate causal triggers and used in silico perturbations to generate testable hypotheses about biological state transitions. These results are model-based prioritization and prediction from observational data, not experimental confirmation that any particular candidate trigger causes a clinical outcome.

The method was also applied to a cross-omic analysis linking gut microbiota with the fecal host proteome in data associated with autism spectrum disorder. This analysis modeled relationships between microbial and host layers and sought to identify candidate triggers of transitions within the combined system. The two biological applications show how CauTrigger can be used to formulate causal hypotheses from complex measurements, but the supplied study evidence does not establish intervention effectiveness or clinical benefit. The framework instead provides an interpretable, computational route for prioritizing candidate triggers and predicting responses that can be tested in subsequent experiments.

Subject of Research: Causal discovery and perturbation prediction in multilayer biological systems

Article Title: CauTrigger: deciphering biological system state transitions by hierarchical causal decomposition

Article References: Zhang, C., Cai, D., Aihara, K., & Chen, L. (2026). CauTrigger: deciphering biological system state transitions by hierarchical causal decomposition. Nature Methods. https://doi.org/10.1038/s41592-026-03274-9

Image Credits: AI Generated

DOI: 10.1038/s41592-026-03274-9

Keywords: CauTrigger, causal inference, hierarchical modeling, deep learning, biological state transitions, perturbation prediction, single-cell multi-omics, type 2 diabetes, gut microbiome, autism spectrum disorder, deciphering, biological