Computational estimation of cell types and their dynamics in health and disease

Abstract

Changes in cell type composition are fundamental to understanding human disease mechanisms. While single-cell omics technologies enable unprecedented resolution in cellular profiling, their widespread clinical application is limited by technical biases and cost constraints. Measurements on bulk tissue specimens, though more robust and cost-effective, lack cell-type resolution. This creates a need for computational methods that can bridge this gap. At its core, cell type deconvolution represents a semi-blind source separation problem, where the goal is to estimate both the mixing proportions and source signals from mixture measurements, given any partial information about the sources from reference data.

This dissertation includes DISSECT, a novel deep semi-supervised learning algorithm for robust cell type deconvolution. DISSECT addresses key limitations in existing approaches by integrating information from both single-cell references and bulk data, effectively handling domain shifts between reference and target datasets. Through comprehensive benchmarking across multiple experimental settings and modalities (including bulk RNA sequencing (RNA-seq), proteomics, and spatial transcriptomics), we demonstrate DISSECT's superior performance in predicting both cell type proportions and cell type-specific expression profiles, with reduced dependency on reference selection.

We used cell type deconvolution to study two distinct diseases: antineutrophil cytoplasmic antibody-associated glomerulonephritis (ANCA-GN) and glioblastoma (GBM). In ANCA-GN, deconvolution of single-cell and spatial transcriptomics data from 34 patients and 8 controls revealed specific T helper cell accumulation patterns associated with inflammation. Computational drug prediction based on this information identified ustekinumab as a potential therapeutic agent, which showed promising results in four patients with poor prognosis under standard treatment. In glioblastoma, we used cell type deconvolution to analyze DNA methylation patterns across multiple cohorts of GBM patients, identifying two distinct and temporally stable GBM groups associated with better prognostic value than established molecular subtypes, particularly in predicting response to surgical intervention.

This dissertation makes several key contributions to bioinformatics, immunology, and neuroscience: (1) a cell type deconvolution framework that advances the state-of-the-art in source separation for biological data, (2) an integrative analysis of immune cell type-specific signals to guide therapeutic decisions in ANCA-GN, and (3) identification of clinically relevant and stable GBM subgroups based on deconvolved cell type-specific signals.