Machine learning reveals time-varying microbial predictors with complex effects on glucose regulation

Abstract

The incidence of type 2 diabetes (T2D) has been increasing globally and a growing body of evidence links type 2 diabetes with altered microbiota composition. Type 2 diabetes is preceded by a long pre-diabetic state characterized by changes in various metabolic parameters. We tested whether the gut microbiome could have predictive potential for T2D development during the healthy and pre-diabetic disease stages. We used prospective data of 608 well-phenotyped Finnish men collected from the population-based Metabolic Syndrome In Men (METSIM) study to build machine learning models for predicting continuous glucose and insulin measures in a shorter (1.5 year) and longer (4.5 year) period. Our results show that the inclusion of gut microbiome improves prediction accuracy for modelling T2D associated parameters such as glycosylated hemoglobin and insulin measures. We identified novel microbial biomarkers and described their effects on the predictions using interpretable machine learning techniques, which revealed complex linear and non-linear associations. Additionally, the modelling strategy carried out allowed us to compare the stability of model performances and biomarker selection, also revealing differences in short-term and long-term predictions. The identified microbiome biomarkers provide a predictive measure for various metabolic traits related to T2D, thus providing an additional parameter for personal risk assessment. Our work also highlights the need for robust modelling strategies and the value of interpretable machine learning.

Machine learning applications in microbial ecology, human microbiome studies, and environmental monitoring

“…In order to glean biologically meaningful data from these ML methods, it may be important to consider the choice of model with preference toward more interpretable algorithms as well as novel methods for interpreting models such as permutational approaches. Microbial ecology studies that demonstrate model transparency are limited to reporting single feature to response interaction or are overburdened by investigating feature contributions to each observation for accumulated local explanations of modeling procedures [48] , [76] , [106] , [107] .…”

“…In practice, ML can perform surprisingly well on datasets that are sampled from and represent messy real-world systems, such as the human body, soil, and water [48] , [49] , [50] and demonstrates superiority over traditional multivariate statistics in analyzing metagenomic data. In addition to these benchmarks, there is an increase in the development of microbiome-specific 'pipelines' that have user-friendly ML implementation and can be accessed through web-interfaces, the statistical compute language R [51] , or Python [52] .…”

Wholegrain fermentation affects gut microbiota composition, phenolic acid metabolism and pancreatic beta cell function in a rodent model of type 2 diabetes

“…Although there are still doubts about the actual causal role of the gut microbiota in T2D (Gurung et al, 2020), accumulating findings support its potential to predict related metabolic outcomes (Rothschild et al, 2018;Aasmets et al, 2021) and indicate that its manipulation in T2D patients or animal models has positive implications for a range of metabolic markers including blood glucose (Cani et al, 2008;Everard et al, 2011;Murphy et al, 2013;Kreznar et al, 2017), with cascading impacts on liver, adipose tissue and muscle (Zhao et al, 2018;Adeshirlarijaney and Gewirtz, 2020). Direct modulation of pancreatic beta cell function may furthermore play a significant role in the central regulation of glucose metabolism, either by gut metabolites directly affecting insulin release from beta cells in the pancreatic islet or by enhancing beta cell survival in response to glucolipotoxicity and inflammation linked to the development of T2D.…”