Latent Process Decomposition of High-Dimensional Count Data

We present a novel approach to probabilistically model high-dimensional count data in an unsupervised way using a three-level hierarchical Bayesian model. Its application is ex-plored in the context of next-generation sequencing data for the purpose of identifying subsets of genes with consistent expression patterns, and that explain a large portion of variability. Each sample is modeled as a finite mixture of Poisson random variables over an underlying set of latent variables that are assumed to correspond to biological functions. Each biological function is further modeled as an infinite mixture over an underlying set of biological function probabilities. We call this model Latent Process Decomposition (LPD). It combines ideas from ma-chine learning and resampling-based methods, and uses a computationally efficient variational method for parameter estimation. The performance of LPD is investigated in both simulated and real data settings to demonstrate that it is a useful modular and extensible tool for iden-tifying interesting genes for further exploration. LPD is implemented as an R/Bioconductor package called themes.