Generative Models of Biological Variations in Bulk and Single-cell RNA-seqMao, Weiguang (2020) Generative Models of Biological Variations in Bulk and Single-cell RNA-seq. Doctoral Dissertation, University of Pittsburgh. (Unpublished)
AbstractThe explosive growth of next-generation sequencing data enhances our ability to understand biological process at an unprecedented resolution. Meanwhile organizing and utilizing this tremendous amount of data becomes a big challenge. High-throughput technology provides us a snapshot of all underlying biological activities, but this kind of extremely high-dimensional data is hard to interpret. Due to the curse of dimensionality, the measurement is sparse and far from enough to shape the actual manifold in the high-dimensional space. On the other hand, the measurements may contain structured noise such as technical or nuisance biological variation which can interfere downstream interpretation. Generative modeling is a powerful tool to make sense of the data and generate compact representations summarizing the embedded biological information. This thesis introduces three generative models that help amplifying biological signals buried in the noisy bulk and single-cell RNA-seq data. In Chapter 2, we propose a semi-supervised deconvolution framework called PLIER which can identify regulations in cell-type proportions and specific pathways that control gene expression. PLIER has inspired the development of MultiPLIER and has been used to infer context-specific genotype effects in the brain. In Chapter 3, we construct a supervised transformation named DataRemix to normalize bulk gene expression profiles in order to maximize the biological findings with respect to a variety of downstream tasks. By reweighing the contribution of hidden factors, we are able to reveal the hidden biological signals without any external dataset-specific knowledge. We apply DataRemix to the ROSMAP dataset and report the first replicable trans-eQTL effect in human brain. In Chapter 4, we focus on scRNA-seq and introduce NIFA which is an unsupervised decomposition framework that combines the desired properties of PCA, ICA and NMF. It simultaneously models uni- and multi-modal factors isolating discrete cell-type identity and continuous pathway-level variations into separate components. The work presented in Chapter 2 has been published as a journal article. The work in Chapter 3 and Chapter 4 are under submission and they are available as preprints on bioRxiv. Share
Details
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|
![[feed]](/images/feed-icon-32x32.png){kind=icon}