End-to-End Deep Learning and Subgroup discovery approaches to learn from metagenomics data

Technological advances have made high-resolution sequencing of genetic material possible at ever lower cost. In this context, the human microbiome (considered as our second "genome") has demonstrated its great capacity to stratify various human diseases. As a "super-integrator" of patient status, the gut microbiota is set to play a key role in precision medicine. Omics biomarkers identification has become a major goal of metagenomics processing, as it allows us to understand the microbial diversities that induce the patient stratification. There remain many challenges associated with mainstream metagenomics pipelines that are both time consuming and not stand-alone. This prevents metagenomics from being used as "point-of-care" solutions, especially in resource-limited or remote locations. Indeed, state-of-the-art approaches to learning from metagenomics data still relies on tedious and computationally heavy projections of the sequence data against large genomic reference catalogs. In this thesis, we address this issue by training deep neural networks directly from raw sequencing data building an embedding of metagenomes called Metagenome2Vec. We also explore subgroup discovery algorithms that we adapt to build a classifier with a reject option which then delegates samples, not belonging to any subgroup, to a supervised algorithm. Several datasets are used in the experiments to discriminate patients based on different diseases (colorectal cancer, cirrhosis, diabetes, obesity) from the NCBI public repository. Our evaluations show that our two methods reach high performance comparable to the state-of-the-art, while being respectively stand-alone and interpretable.