Learning Glycan Structure Similarity using Sequence Models

Haley Brennan

Mentor: Dr. Nathan Edwards, Department of Biochemistry and Molecular & Cellular Biology, Georgetown University Medical Center.

Date/Time: August 25th, 2026 at 1:30 PM.

Abstract: Glycans are complex, highly branched carbohydrate structures that play important roles in immune recognition, cell signaling, and disease progression. However, defining structural similarity between glycans remains difficult because traditional methods are based on the addition and removal of residues, and may miss broader structural or functional relationships. This project investigates whether transformer-based masked language models can learn meaningful representations of glycan structure directly from their IUPAC sequence representations, similar to how language models capture semantic relationships in text. A RoBERTa-style masked language modeling workflow was developed to train transformer models on glycan IUPAC sequences, compare various tokenization strategies, including byte-pair and glycan-specific sequence-aware approaches, and to evaluate how well such models capture glycan structural patterns. Specifically, the models’ representations are used to learn and predict glycan classification labels; and place glycans with similar structures close together in the learned embedding space. These analyses evaluate whether the learned embeddings capture meaningful structural similarities between glycans beyond those useful for masked token prediction. This work aims to provide a scalable framework for modeling glycan similarity using transformer-learned embeddings and may support future applications in glycobiology, including functional annotation, biomarker discovery, and glycoengineering.