Breast cancer subclass analytics from glycoenzyme gene expression using machine learning and bioinformatics methods

Truter, Mia

Breast cancer subclass analytics from glycoenzyme gene expression using machine learning and bioinformatics methods

Thesis / Dissertation

2026

Publisher

University of Cape Town

Publisher

University of Cape Town

Department

Department of Chemistry

Faculty

Faculty of Science

Abstract

Glycosylation, a post-translational modification that covalently attaches carbohydrate structures to proteins and lipids, is essential for almost all cellular processes, including cell-to-cell communication, response to cellular signalling, and maintaining the balance between self-renewal and differentiation. Aberrant expression of the glycoenzymes that perform glycosylation is a hallmark of cancer, but little is known regarding how glycosylation is functionally integrated with the broader tumorigenic landscape. The clinical and phenotypic heterogeneity of breast cancer suggests activation of multiple tumorigenic pathways, each likely accompanied by distinct glycosylation aberrations that co-operate with other biochemical pathways to drive tumorigenesis. This research investigates the benefit of utilising a glycosylation-perspective to understand breast cancer. Glycosylation signatures within traditional breast cancer subtypes are explored and contextualised with accompanying functions, demonstrating that glycosylation detects nuances in breast cancer that traditional classifications cannot. Following this, unsupervised glycosylation-based breast cancer subtype discovery is undertaken, and machine learning models are built, optimised, and benchmarked for this purpose. The expression of glycoenzymes is used as input to the Growing Hierarchical Self-Organising Representation Map (GHSORM), an unsupervised clustering method designed specifically for cancer subtyping. The discovered subtypes undergo feature discovery using the Recursive Feature Elimination with Strikeout (RFES) wrapper to identify glycoenzymes that are important for subtyping. Using bioinformatics methods, the subtypes are functionally interrogated, specifically in the context of the immune response and the epigenome. A gene regulatory network is constructed that highlights distinct functional roles for glycosylation in each subtype. Through a deep learning approach, drugs are predicted that specifically inhibit the biochemical functions associated with aberrant glycosylation in each subtype. The pervasive role of glycosylation in the mechanistic functions of cells enables a panoramic view of breast cancer, disentangling biochemical nuances that drive the phenotypic heterogeneity associated with the disease. Thus, glycosylation can be harnessed for personalised diagnostic and therapeutic approaches in breast cancer.

Keywords

glycosylation

machine learning

Reference:

Collections

PhD / Doctoral

Full item page