Latent Process Decomposition Of High-Dimensional Count Data

Motivation: Next-generation sequencing (NGS) technologies have become the preferred way of exploring a genome. These data are high-dimensional discrete counts with correlated variables (e.g., genes). We present a novel latent factor model for high-dimensional count data, Latent Process Decomposition (LPD-C), that accounts for the correlations among genes and models the biological hypothesis that genes work in groups (e.g., pathways), which are referred to as processes. LPD-C is a two stage unsupervised approach for grouping genes into a pre-specified number of clusters, and for selecting genes that belong to these clusters with high probability. The first stage of LPD-C uses a variational Bayes approach for efficient estimation of its parameters. The second stage of LPD-C selects genes grouped as gene-subsets using empirical Bayes hypothesis testing. Results: The performance of LPD-C is explored using simulated and publicly available NGS data, compared with existing approaches, and shown to be a useful and extensible framework for identifying genes suitable for further exploration. Although we apply LPD-C in a genomic context, it can be used for any high-dimensional count data. Availability: R code for fitting LPD-C is available from the authors on request. Contact: [email protected]