Assessing the Polygenic Risk between Anxiety and Gut Microbiota Using Machine Learning

There is growing evidence that changes in the gut microbiome are linked to anxiety. Some studies have evaluated the association between gut microbiota and anxiety using either two-sample mendelian randomization or polygenic risk score (PRS). However, they only used linear regression to fit the results of PRS analysis, which could lead to under-fitting of the estimation model and consequently achieve imprecise conclusions. In this study, we performed both the PRS analysis and the two-sample mendelian randomization analysis between anxiety and gut microbiotas, as well as potential pathways involved in the connection between them. In addition to the Lasso regression performed as the initial assessment tool, we implemented four machine learning models (decision tree, random forest, XGBoost, and lightGBM) to fit the PRS values of each gut microbiota and pathway. A statistical significance test was performed for the feature importance generated by these four machine learning models to find the statistically significant important (p-value <0.05) gut microbiota and pathway to the polygenic risk of anxiety. Five kinds of gut microbiota were detected and four of them were reported by experimental literature research, which means that our strategy of analysis provides critical insights into the association between anxiety and gut microbiotas as well as pathways. This work combines the machine learning methods’ nonlinear fitting ability with disease association and provides new insights into understanding the results of PRS analysis.