← Back to portfolio

Glyde

A domain-aware, topology-biased glycan language model for viral receptor binding

0.96 AUROC Binding prediction
~380K params Lightweight transformer
Topology-aware Not sequence-only

Overview

Glyde is a transformer architecture designed specifically for glycans, incorporating biological structure directly into attention and pooling. Instead of treating glycans as flat sequences, the model encodes branching patterns, linkage chemistry, and terminal motifs to better capture viral binding behavior.

Key Idea

Standard language models rely purely on statistical patterns. Glyde introduces inductive bias through topology-aware features, allowing the model to prioritize biologically meaningful structures during training and inference.

Paper

If the embedded viewer does not load, open the PDF directly.