Improving DirectLiNGAM for high-dimensional microbiome data: roots screening and eBIC based model selection

Identifying causal relationships from observational data is a central challenge in gut microbiome research, where complex, multivariate interactions shape host health and disease. These data are typically high-dimensional and sample-limited, creating substantial obstacles for causal discovery and motivating the development of methods tailored to this regime. In this study, we address this challenge by focusing on DirectLiNGAM and introducing two complementary methodological improvements designed to facilitate its practical application in microbiome data. Specifically, we propose two extensions to the DirectLiNGAM algorithm targeting prior knowledge extraction via roots screening and model selection via the integration of the extended BIC criteria. Together, these contributions extend the applicability of DirectLiNGAM to microbiome systems without altering the core modeling assumptions of the method. We validated the proposed methodology through a rich set of numerical experiments on synthetic data and demonstrate its application on a real biological dataset. This work supports the wider adoption of LiNGAM-based approaches for causal discovery in systems biology and related domains.