Machine Learning analysis of the human infant gut microbiome identifies influential species in type 1 diabetes

• Published in: Expert systems with applications 2021-12, Vol.185, p.115648, Article 115648
• Main Authors: Fernández-Edreira, Diego, Liñares-Blanco, Jose, Fernandez-Lozano, Carlos
• Format: Article
• Language: English
• Subjects: Body parts; Design of experiments; Diabetes; Disease; Feature selection; Generalized Linear Model; Infants; Machine Learning; Metagenomics; Microbiota; Random forest; T1D
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2021.115648

Diabetes is a disease that is closely linked to genetics and epigenetics, yet mechanisms for clarifying the onset and/or progression of the disease have sometimes not been fully managed. In recent years and due to the large number of recent studies, it is known that changes in the balance of the microbiota can cause a high battery of diseases, including diabetes. Machine Learning (ML) techniques are able to identify complex, non-linear patterns of expression and relationships within the data set to extract intrinsic knowledge without any biological assumptions about the data. At the same time, mass sequencing techniques allow us to obtain the metagenomic profile of an individual, whether it is a body part, organ or tissue, and thus identify the composition of a given microbe. The great increase in the development of both technologies in their respective fields of study leads to the logical union of both to try to identify the bases of a complex disease such as diabetes. To this end, a Random Forest model has been developed at different taxonomic levels, obtaining results above 0.80 in AUC for families and above 0.98 at species level, following a strict experimental design to ensure that results are compared under equal conditions. It is identified how, in infants, the species Bacteroides uniformis, Bacteroides dorei and Bacteroides thetaiotaomicron are reduced in the microbiota of those with T1D, while, the populations of Prevotella copri increase slightly and that of Bacteroides vulgatus is much higher. Finally, thanks to the more specific metagenomic signature at species level, a model has been generated to predict those seroconverted patients not previously diagnosed with diabetes but who have expressed at least two of the autoantibodies analysed. [Display omitted] •A Machine Learning methodology is proposed for the modelling of metagenomic data and the diagnosis of T1D infants.•A new metagenomic signature was obtained highly correlated with T1D diagnosis.•Machine Learning algorithms are able to obtain great results with sparse metagenomic data.