ETD

• GALA Pathway
• glycosylation
• machine learning

Protein O-GalNAc glycosylation influences protein folding, secretion, cell adhesion, and signalling. Under normal conditions, it is initiated in the Golgi apparatus by a family of glycosyltransferases, the GALNTs enzymes. In several cancers, including pancreatic ductal adenocarcinoma part of this system is relocalised to the endoplasmic reticulum. This changes the set of residues that can be glycosylated, allowing modification of sites that would normally remain untouched. Such aberrant glycosylation has been associated with tumour invasion.
Glycoproteomics can identify thousands of modified residues in a single experiment, but it cannot determine where each modification was installed. As a result, sites generated by the relocalised enzymes cannot be distinguished directly from those produced by the canonical Golgi pathway.
My thesis addresses this problem using a machine learning approach. Glycosylation sites identified by mass spectrometry were characterised using structural features, together with their local sequence context, represented by a protein language model. These features were then used to train a model that ranks observed sites according to how strongly they resemble sites associated with the relocalised condition. The resulting score provides a way to identify residues that are more characteristic of aberrant glycosylation in cancer.