Identifying cell populations in flow cytometry data with Bayesian hierarchical models

Gaussian mixture models are useful for clustering many kinds of data. They are a natural tool to use for the cell population identification problem in flow cytometry data, where multidimensional measurements on individual blood cells are to be partitioned into populations. To obtain comparable results across samples, ideally one would like to superimpose measurements before fitting the mixture model, but this is impeded by biological variation between blood samples and technical variation between runs of the flow cytometer that impact the location and shape of cell populations. I will present a computational pipeline with a Bayesian hierarchical model that we have constructed to lend strength between samples, while allowing variation in location and shape. Another important step in the pipeline is the quality control, where diagnostics are computed to ensure that the obtained populations correspond to plausible biological cell populations.