Abstract
Genome scale molecular datasets are often highly structured, with many correlated measurements. This general phenomenon can be related to the underlying data generating process. In assays of mixed cell populations, such as blood, variation in cell-type proportion induces a complex correlation structure at the gene-level. Likewise, groups of genes can be co-regulated/co-expressed through shared transcription factors and signaling pathways. Many applications of gene expression analysis rely on their ability to reflect these unobserved biological processes in order to draw mechanistic conclusions. On the other hand, correlated patterns of expression may also reflect nuisance factors, such as batch effects, which interfere with correct biological interpretation. The choice of analysis method is heavily dependent on which of these factors (nuisance or interesting-biological) is believed to account for more variation and the optimal variance analysis strategy remains an open question. In this study we describe a method to infer a biologically grounded data generating model that provides estimates of underlying biological processes, including explicitly identified pathway-level and cell-type proportion effects. Specifically, we formulate a new matrix decomposition framework, PLIER (Pathway-level Information ExtractoR), that explicitly incorporates prior biological knowledge. Using simulations, we demonstrate the superiority of our method in recovering the true data generating model. Using real data, we show that our approach is able to recover interpretable biological variables, reproduce previous findings in a simplified framework, distinguish biological and technical variation, and provide additional biological insight. The PLIER method and auxiliary functions and data are compiled in the PLIER R package available at https://github.com/wgmao/PLIER.
