Interpretable machine learning framework reveals novel gut microbiome features in predicting type 2 diabetes

Abstract

New interpretable machine-learning analytic framework identifies a combination 55 of microbes consistently associated with type 2 diabetes risk across three 56 independent cohorts involving 9111 participants 57 • Faecal microbiota transplantation from humans to germ-free mice demonstrates a 58 causal role of the identified combination of microbes in the type 2 diabetes 59 development 60 • Body shape could modify the gut microbiome-diabetes relationship 61 4 Abstract 62Gut microbiome targets for type 2 diabetes (T2D) prevention among human cohorts 63 have been controversial. Using an interpretable machine learning-based analytic 64 framework, we identified robust human gut microbiome features, with their optimal 65 threshold, in predicting T2D. Based on the results, we constructed a microbiome risk 66 score (MRS), which was consistently associated with T2D across 3 independent 67 Chinese cohorts involving 9111 participants (926 T2D cases). The MRS could also 68 predict future glucose increment, and was correlated with a variety of gut microbiota-69 derived blood metabolites. Faecal microbiota transplantation from humans to germ-70 free mice demonstrated a causal role of the identified combination of microbes in the 71 T2D development. We further identified adiposity and dietary factors which could 72 prospectively modulate the MRS, and found that body fat distribution may be the key 73 factor modulating the gut microbiome-T2D relationship. Taken together, we proposed 74 a new analytical framework for the investigation of microbiome-disease relationship. 75 The identified microbiota may serve as potential drug targets for T2D in future. 77Type 2 diabetes (T2D) is a complex disorder influenced by both host genetic and 78 environmental factors (1), and its prevalence is rising rapidly in both developed and 79 developing countries (2). Gut microbiome is considered as a modifiable 80 environmental factor, which plays an important role in the development of T2D (3-7). 81The research interest to identify gut microbiome-related treatment/prevention target is 82 emerging recently (8). Although there are a few human studies investigating the 83 association of gut microbiome with T2D in the past few years, the results are 84 inconsistent, and the causality is lacking (9). So far, there are sparse human evidence 85 robustly linking specific gut microbiome features to T2D. 87Machine learning has been widely used in biomedical fields in recent years (10). 88However, its application in the clinical setting is still limited as their predictions are 89 usually difficult to interpret. Of note, with the methodology development in the past 90 few years, interpretable algorithms could unlock the traditional "black box" of 91 machine learning results (11). The integration of the new algorithms with large-scale 92 gut microbiome data have the potential to radically unveil the relationship between 93 gut microbiome and T2D. Yet, no such investigation has been done. 95Therefore, in the present study, we aimed to identify robust h...

Machine Learning Reveals Time-Varying Microbial Predictors with Complex Effects on Glucose Regulation

“…Most of the current studies describing the role of bacteria in diabetes have been case-control studies, with diabetes being a binary trait defined by setting a cutoff to some continuous glucose measure ( 3 , 4 , 12 ). Type 2 diabetes, however, is a disease preceded by a long-lasting prediabetic state, and the development of the disease is a continuous process ( 13 ).…”

“…Recently, Gou et al ( 12 ) used a similar interpretable machine learning strategy and found bacteria that effectively differentiated type 2 diabetes cases from healthy controls in the Chinese population. Additionally, they built a microbiome risk score (MRS) and showed the causal role of identified bacteria on diabetes development after fecal microbiota transplantation to mice.…”

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

“…Similarly to the work by Gou et al (12), the main reasons behind these inconsistencies are likely study design and population structure. We are not aware of any population with similar follow-up period and where microbiome data is available and oral glucose tolerance test has been carried out at the baseline and at the follow-up.…”

“…Most of the current studies describing the role of bacteria in diabetes have been case-control studies with diabetes being a binary trait defined by setting a cut-off to some continuous glucose measure (3,4,12). Type 2 diabetes however is a disease preceded by a long-lasting prediabetic state and the development of the disease is a continuous process (13).…”

3MCor: An Integrative Web Server for Metabolome-Microbiome-Metadata Correlation Analysis

“…For g_Streptococcus, it has been reported that diabetes is the most frequently identi ed risk factors in invasive Streptococcus agalactiae (GBS) infections [53]. Recent machine learning studies have found that f_Mogibacteriaceae is one of the markers that can effectively predict the risk of DM [54]. Ruminococcus.spp was positively associated with DM in many studies [55][56][57][58][59].…”