Pawl, E. (STAT) – Flexible and Scalable Mixtures of Experts for Oceanographic Flow Cytometry Data

Flow cytometry is a valuable technique in microbial research used to measure the optical properties of single-celled organisms at high throughput. Oceanographers often deploy flow cytometers on research cruises in order to study the characteristics of phytosynthetic microbes—called phytoplankton—in regions and times with diverse environmental conditions. Because cytometers cannot distinguish between subpopulations, researchers typically cluster observations into subpopulations and subsequently analyze cluster characteristics. This two-stage workflow is often manual, difficult to reproduce, and fails to account for uncertainty in cluster assignments when relating subpopulation behavior to environmental conditions. To address these shortcomings, statistical mixture models are gradually being introduced as alternatives to manual flow cytometry data analysis. However, existing models either cannot use covariates or make restrictive assumptions about the relationships between cluster characteristics and covariates. Additionally, they are designed to analyze individual cruises and consequently characterize local, rather than global, patterns in phytoplankton behavior. We propose to develop computationally efficient mixtures of experts which account for the complex dependency structures in oceanographic flow cytometry data. In this framework, cells are probabilistically assigned to latent subpopulations, while cluster-specific regressions relate each subpopulation’s optical properties and relative abundance to environmental conditions. Our first project develops a mixture of random weight neural network experts which can estimate arbitrary nonlinear regressions at low computational cost, without a priori specification of functional forms. In the second project, we develop a variational Bayesian mixture of experts which automatically selects variables without requiring cross-validation for hyperparameter selection. The final project incorporates spatial and temporal dependence, allowing joint inference on data collected from multiple research cruises conducted at different locations and times.

Event Host: Ethan Pawl, Ph.D. Student, Statistical Science

Advisors: Sangwon Hyun & Paul Parker

Zoom- https://ucsc.zoom.us/j/96353239941?pwd=a4PJ94EMSD6D0SJ75S3WYzrPbYsBtn.1

Passcode- 244463