Human microbiomes and their interactions with various body systems have been linked to a wide range of diseases and lifestyle variables. To understand these links, citizen science projects such as the American Gut Project (AGP) and Human Food Project (HFP) have provided large open source datasets for microbiome investigation. In this talk, I will present our recent work that leverages such open source datasets by learning a microbial “language” model using techniques originally developed for Natural Language Processing (NLP). Our microbial “language” model is trained in a self-supervised fashion to capture the interactions among different microbial species (taxa) and the common compositional patterns in forming microbial communities, much like the language model in NLP trained to capture word interactions and the grammatical patterns in natural texts. Importantly, the learned model allows for individual bacterial species to be interpreted and represented differently in different contexts of the microbial environment and produces a representation of a sample by collectively interpreting different bacteria species in the sample and their interactions as a whole. To demonstrate the power of our model, we show that our sample representation leads to improved prediction performance compared to baseline representations consistently across multiple tasks including predicting disease states and diet patterns. We also show that the learned representation, coupled with a simple ensemble strategy, can produce highly robust models that can generalize well to microbiome data independently collected from different populations. Finally, I will present some interpretation results that help understand our model and its behavior.
Xiaoli Fern is an associate professor of Computer Science at Oregon State University. She received her Ph.D. (2005) in computer engineering from Purdue University and her M.S. (2000) and B.S. (1997) degrees from Shanghai Jiao Tong University. Dr. Fern is broadly interested in applied machine learning and data mining, where she draws inspirations from practical challenges from real-world applications to develop new methods and new understanding for machine learning. Her current research focuses on self-supervised learning from large and complex data with application areas spanning from material characterization and design, to learning “rules of life” from microbial and metabolomics data. Her work is sponsored by NSF, DARPA, DOE and USDA.