Node-Level Mechanism-Shift Detection in Linear Non-Gaussian Causal Networks: A Bootstrap-Based Framework with Application to the Human Gut Microbiome

As microbiome research increasingly seeks to identify true ecological shifts, transitioning from associational to causal approaches is essential. However, detecting structural changes across independent networks remains challenging due to the absence of established biological ground truths and the small, imbalanced sample sizes typical of microbiome cohorts. To address this, we extend an existing network comparison framework to enable node-level mechanism-shift detection under the direct linear non-Gaussian acyclic model (DirectLiNGAM). We evaluate the Naive, Bootstrap, and Relative sample size Bootstrap Stability (RSBS) estimators across extensive synthetic discovery runs and a semi-synthetic, batch-corrected human gut microbiome cohort. Our results demonstrate that resampling-based estimation consistently outperforms a single-fit Naive baseline by trading marginal recall for substantial precision gains. On both semi-synthetic and synthetic data, standard Bootstrap is optimal for comparing datasets of equal size, whereas RSBS is the only estimator that reliably handles imbalanced cohorts. This advantage strengthens as network dimensionality increases and persists under authentic compositional noise. Navigating this complex and emerging research area is currently constrained by limitations in data quantity and quality, scarce biological knowledge, and a lack of dedicated software. To address these critical gaps, we provide the complete benchmark pipeline and synthetic data generators as an open-access Python package, causal-comparator, to support node-level mechanism-shift detection across systems biology applications.