Modeling Dynamic Cross-Correlation in Spatial Transcriptomic Data

Thomas M. Menino Convention & Exhibition Center 

In living tissues, mechanisms such as ligand-receptor signaling, immune synapse formation, and coordinated tissue repair depend on physical proximity or localized arrangements of cells in order to function effectively. To gain insight into these spatial cellular phenomena, we utilize a Bayesian spatial deep Gaussian process factor model that allows for the modeling of intracellular as well as intercellular gene association. Differences between cell types are also captured by the structure of the model and can be analyzed using the posterior samples of the MCMC.


The model is made scalable to thousands of cells and genes through the use of the SPDE formulation of the Matérn covariance kernel as well as through random Fourier features.

Sampling is carried out using a partially collapsed Gibbs sampler with conjugate updates for the uncollapsed portion and no U-turn sampling used for the collapsed parameters.

SPDE

Spatial Factor Model

Deep Gaussian Process

Cell type differences 

Section on Statistics in Genomics and Genetics