Quantifying uncertainty in microbiome-based prediction using Gaussian processes with microbial community dissimilarities.

Summary: The human microbiome is closely associated with the health and disease of the human host. Machine learning models have recently utilized the human microbiome to predict health conditions and disease status. Quantifying predictive uncertainty is essential for the reliable application of these microbiome-based prediction models in clinical settings. However, uncertainty quantification in such prediction models remains unexplored. In this study, we have developed a probabilistic prediction model using a Gaussian process (GP) with a kernel function that incorporates microbial community dissimilarities. We evaluated the performance of probabilistic prediction across three regression tasks: chronological age, body mass index, and disease severity, using publicly available human gut microbiome datasets. The results demonstrated that our model outperformed existing methods in terms of probabilistic prediction accuracy. Furthermore, we found that the confidence levels closely matched the empirical coverage and that data points predicted with lower uncertainty corresponded to lower prediction errors. These findings suggest that GP regression models incorporating community dissimilarities effectively capture the characteristics of phylogenetic, high-dimensional, and sparse microbial abundance data. Our study provides a more reliable framework for microbiome-based prediction, potentially advancing the application of microbiome data in health monitoring and disease diagnosis in clinical settings. Availability and implementation: The code is available at https://github.com/asahiadachi/gp4microbiome.