Description

Abstract: Our microbiota, or the trillions of micro-organisms living on and within us, constitute extremely rich ecosystems with critical functions for human health. Major perturbations of the microbiome, such as diet or therapeutic drugs (e.g., antibiotics) may result in dysbiosis, which has been associated with many human disorders, including infections, inflammatory bowel disease, neurological diseases, autoimmunity/allergies, diabetes, and malignancies. Recent microbiome studies have powerfully demonstrated the ability to go beyond association-finding, to drill down into molecular mechanisms, facilitated by culture-based systems or animal models (e.g., gnotobiotic mice). Other recent exciting trends are the investigation of bacteriotherapies, or live commensal bacteria, for tackling diseases that are difficult to treat with conventional therapies, and the development of microbiome-based diagnostics/prognostics. However, understanding our microbiota and ultimately harnessing its properties to improve human health is challenging, given the scale, complexity, temporal/multi-modal nature, and noisiness of microbiome datasets. In this talk I will describe novel computational approaches my lab is developing to tackle these challenges, including: (1) probabilistic machine learning models that learn dynamical systems from longitudinal microbiome data to support rational design of bacteriotherapies, e.g., for C. difficile and autoimmune diseases, (2) deep learning methods that use metabolomics and sequencing data to infer fully human-interpretable predictive rules for improved diagnostics, e.g., for C. difficile recurrence, (3) generative deep learning methods for discovering relationships between the microbiome, human immune system, and infections, e.g., tuberculosis, (4) generative deep learning methods for characterizing the spatial organization of the microbiome over time and dissecting how perturbations such as diet impact microbial interactions.