PH.D DEFENCE - PUBLIC SEMINAR

Predicting Microbial Interactions with Modelling Approaches

Speaker
Mr Li Chenhao
Supervisor
Dr Wong Lim Soon, Kithct Chair Professor, School of Computing
Dr Nagarajan, Niranjan, Adjunct Associate Professor, School of Computing


  15 May 2019 Wednesday, 02:00 PM to 03:30 PM

 Executive Classroom, COM2-04-02

Abstract:

Computational modelling represents an attractive avenue for scalable data-driven analysis of microbial community function and dynamics. The practicability of utilising in silico modelling to directly learn ecological models is now conceivable with the vast amount of publicly available microbial community profiling data. Despite the promise, progress has been relatively muted as existing model inference algorithms relied on absolute abundances rather than the relative measurements generated with high-throughput microbial profiling.

We introduce a new algorithm for learning generalised Lotka-Volterra models (gLVMs) from longitudinal microbial profiling data by coupling Biomass Estimation and model inference in an Expectation-Maximization-like algorithm (BEEM). We show that BEEM outperforms existing methods for inferring gLVMs, while simultaneously eliminating the need for absolute abundances as input. BEEM's application to previously inaccessible public datasets (due to lack of information on absolute abundances) allowed us for the first time to construct ecological models of microbial communities in the human gut on a per individual basis, revealing personalised dynamics and keystone species.

For cross-sectional microbial community profiles, correlation based strategies have been the most widely used approach to inferring microbial interactions. However, our benchmarking evaluations showed that correlation based methods have varied performance for predicting interactions. To better infer interactions and construct models from cross-sectional data, we developed an extension of BEEM (BEEM-static). BEEM-static improves inference accuracy by automatically identifying samples that are close to steady states for training. In addition to interaction predictions, BEEM-static also enable instantaneous growth inference for each species member of the community. BEEM-static outperforms correlation based methods for modelling cross-sectional data, substantially improving the prediction accuracy of directed interactions.