[MS16] Complex Biological Dynamics and SOTA Mathematical Methods

[MS16-04] Deciphering cell state transitions by hierarchical causal decomposition for multi-perturbation predictions

*Chengming Zhang¹, Deyu Cai², Kazuyuki Aihara¹, Luonan Chen³ (1. The University of Tokyo (Japan), 2. ShanghaiTech University (China), 3. Chinese Academy of Sciences (China))

Keywords:

State transition,Hierarchical causal decomposition,Perturbation prediction,Causal information flow

Deciphering and manipulating cell state transitions remain fundamental challenges in biology, as master regulators, e.g. some transcription factors (TFs), orchestrate these processes by regulating downstream effectors through hierarchical interactions. However, existing methods often struggle to accurately identify causal drivers within these layers of regulation and predict transcriptional responses to their perturbations. To address these challenges, we developed CauTrigger, a deep learning framework that uses hierarchical causal decomposition to identify master regulators or TFs driving state transitions from an initial state to a target state based on observed gene expression profiles. Specifically, by explicitly inferring hierarchical regulatory relations with a do-calculus causal framework, CauTrigger disentangles causal drivers from spurious associations and offers mechanistic insights into regulatory cascades. Additionally, it predicts transcriptional responses to combinatorial perturbations, enabling a deeper understanding of regulatory mechanisms. Benchmarking on diverse datasets demonstrates that CauTrigger outperforms traditional methods in driver identification and perturbation outcome prediction. Applied to single-cell and spatial transcriptomics datasets, CauTrigger not only validated known drivers and perturbation predictions but also uncovered novel insights through real data, offering a versatile tool for not only achieving accurate cell state transitions but also revealing their hierarchical causal/regulatory mechanisms.