Abstract
While single-cell transcriptomics routinely generates massive datasets, turn ing raw sequencing data into reliable cell-type annotations remains messy. Subjective labeling and poor dataset quality blur population boundaries, degrading downstream analysis and breaking supervised models.This thesis tackles this annotation bottleneck systematically. First, it investigates label correctness at scale, introducing a formal detection work f low and a Label Uncertainty Learning framework that handles unverified labels probabilistically. Second, it replaces classical flat classification with a hierarchical Multi-Dimensional Cell Classification Tree. This architecture maps cells across lineage, markers, and functional states rather than forcing them into rigid categories. Finally, it defines strict reference data criteria and shares a harmonized PBMC dataset via a public platform.
Testing revealed major structural improvements over standard pipelines. Because the hierarchical system maps fluid biological dimensions, it cleanly captures ambiguous transitional cellular states that typically confuse flat classifiers. Stepping away from predefined lists solves a core limitation of traditional supervised learning: it directly enables the discovery of novel cell subtypes. Furthermore, classifiers trained solely on healthy baselines successfully flagged pathological shifts in separate disease cohorts, proving their robust real-world utility.
Ultimately, this work closes the gap between data curation and reliable algorithm deployment. It establishes a computational framework to turn massive single-cell collections into valuable biological resources with stable labels, transitioning cell annotation from an ad-hoc interpretative exercise into a resilient modeling and discovering pipeline.
| Date of Award | 18 Jul 2026 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Stuart McDonald (Supervisor) & Richard Rankin (Supervisor) |
Cite this
- Standard