Year: 2025
Type: dissertation
Abstract: This thesis investigates whether hierarchical clustering combined with large language models (LLMs) can improve cell type annotation in single-cell RNA sequencing (scRNA-seq) data. To this end, a pipeline named GPTree was developed. It implements recursive clustering followed by LLM-based interpretation of cluster marker genes, mimicking the hierarchical approach commonly used in manual annotation. GPTree was compared to GPTCelltype, an LLM-based annotation method, and SingleR, a reference-based tool, across multiple human tissue datasets. The performance was evaluated using accuracy scoring and match type percentages. Parameter sensitivity analysis confirmed the stability of the clustering algorithm. The results show that GPTree achieves similar or improved accuracy compared to GPTCelltype, particularly at broad cell type levels, and outperforms SingleR in most cases. These findings suggest that recursive clustering combined with LLM-based annotation offers a biologically intuitive and effective approach for automated cell type identification.
Source: Aaltodoc (Aalto University)
Author: Kreivi, Nea
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7113445440