STATISTICAL DECONVOLUTION METHODS FOR TRANSCRIPTOMICS DATA

Deconvolution of bulk transcriptomics data from mixed heterogeneous cell populations to each cell type is important to identify the cellular mechanism of complex trait diseases. Existing statistical approaches can be divided into two major groups: supervised and unsupervised ones. Supervised deconvolution methods require cell type-specific prior information including cell proportions, reference cell type-specific gene signatures or marker genes for each cell type, which may not be fully available in practice. Unsupervised methods, in contrast, do not use any cell type-specific prior information and are therefore not efficient for data with partial cell type-specific information. Motivated by the limitations of existing deconvolution methods, we develop semi-supervised and supervised methodologies to fully utilize partially available cell type information to infer cell proportions. We first propose a partial marker gene-guided semi-supervised method, semi-CAM, in which we extend a fully unsupervised deconvolution method, Convex Analysis of Mixtures (CAM). Compared with CAM, which does not take any prior cell type-specific information, the proposed method searches for additional novel markers while employing several known markers from partial or all cell types. In our evaluations, semi-CAM outperforms CAM by yielding more accurate cell proportion estimates when markers from partial/all cell types are available. Furthermore, when markers for all cell types are available, semi-CAM achieves similar or better accuracy than existing supervised marker-based methods. Besides cell type compositions, important covariates, such as gender and age, may greatly influence bulk gene expressions. Most of the exsiting deconvolution methods are incapable of adjusting for covariate effects. SVA is a widely used method for detecting hidden confounding variables in transcriptomics data. We propose a SVA-based approach which treats the unmeasured cell proportions as hidden variables. It deconvolutes data with top-ranked surrogate variables and cell proportions from a set of training samples, or relative cell proportions from samples with one missing cell type. We show that the SVA-based method is superior to existing deconvolution methods. In additional to observed covariates, there may exist other hidden confounding variables whose effects, if ignored, can lead to inaccurate cell proportion estimates. Finally, we extend the SVA-based deconvolution method to further adjust for hidden confounding variables.Doctor of Philosoph