Estimating Genotype-Tissue Specific Gene Expression Via Promoter Motifs, Genomic Proximity, And Hybrid Deep Learning
Year: 2026
Type: dissertation
Abstract: Understanding how genotype-tissue specific gene expression is regulated remains a central challenge in genomics and medicine. Insights into these mechanisms can reveal disease origins and guide targeted therapies. This dissertation develops computational methods for analyzing genotype-tissue expression (GTEx) data to identify key regulatory elements and reliably estimate missing or incomplete expression profiles, thereby reducing experimental burden. We introduce hybrid machine learning approaches to address three connected challenges: (1) defining biologically meaningful promoter boundaries; (2) identifying genomic factors influencing expression similarity; and (3) estimating missing genotype-tissue specific profiles.Chapter 3 systematically investigates how genomic proximity and gene orientation shape expression similarity using large-scale GTEx data. We show that nearby genes tend to share expression patterns regardless of orientation, except for oppositely oriented nested genes. This similarity is strongest within clusters of up to three neighbors or within 20-60 kb. We also observe that shorter chromosomes exhibit greater local expression similarity. These findings motivate using neighboring genes to estimate expression, including for less-characterized noncoding genes and potentially for other species. Chapter 4 addresses the challenge of arbitrary promoter definitions, which may omit essential regulatory elements or include excessive noise. We introduce a data-driven strategy that infers optimal promoter lengths using expression similarities among promoter-sharing gene pairs, and further refines these thresholds using candidate cis-regulatory elements and transcription factor binding data. The resulting promoter boundaries better capture true regulatory structure. Chapter 5 leverages proximity-based expression similarity to estimate genome-wide expression from a minimal set of reference genes. We formulate reference gene selection as a set-cover problem and develop a dynamic programming based vicinity set cover algorithm that identifies an optimal and compact reference set while minimizing distances between reference and target genes. This approach reduces both experimental and computational effort and generalizes to other coverage problems. Chapter 6 integrates these insights into a hybrid framework that combines promoter motifs (CNNs), tissue correlations (transformers), and genomic features (tree-based regression). Our method reduces estimation errors by ~34% at short (<5 kb) and ~30% at long (>100 kb) distances compared with distance-only baselines, and generalizes to noncoding genes and data-limited settings.
Source: TSpace (University of Toronto)
Author: Dong, Jiahong
Institution: University of Toronto
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7203958703
