New machine learning model reveals unexpected protein-sugar interactions

A team of GlycoNet researchers at the University of Alberta has contributed to a major step forward in computational glycomics with a new study published in Science Advances, introducing MCNet, a machine learning model that predicts how strongly proteins (lectins) interact with sugars (glycans). Accurate prediction of these interaction strengths is central to understanding immune recognition, cell–cell communication and pathogen binding, yet has been difficult to achieve at scale.

MCNet stands out for several reasons. It incorporates lectin concentration directly into the model, an element most approaches ignore. This allows the system to capture how the same protein behaves under different experimental conditions. MCNet also uses an atom-level representation of each glycan, which enables it to evaluate “mirror-form” sugars that are rare in nature and absent from typical training data.

“The most significant outcome is that the model made predictions we could then validate experimentally. It isn’t just generating numbers; it’s showing behaviour we can confirm in the lab,” says Eric Carpenter, the first author on the paper and a computer analyst at the University of Alberta.

That validation proved critical. The researchers verified several unexpected interactions, including binding between Aleuria aurantia lectin (AAL) and L-glucose (the mirror-form of glucose). The results show that MCNet can extend well beyond the 500–600 glycans used during training, opening the door to exploring thousands of additional structures.

“Mirror image glycans have been frequently considered as ‘extremely rare in nature’ but their rarity is an artefact of a human-centric focus on glycomics,” adds Ratmir Derda, the corresponding author on this manuscript. “While indeed mirror image glycans aren’t commonplace in humans and higher mammals, a closer look at microorganisms in three kingdoms of life reveals that these amazing cells employ nearly 100 diverse mirror image glycans to build their biomolecules. The supporting information in the Science Advances paper provides the first glimpse to the prevalence of such mirror image glycans in nature. A full systematic analysis of mirror image building blocks will be coming out shortly.”

This work highlights how integrated computational and experimental efforts can broaden the reach of glycomics and illuminate how biomolecules might behave in alternative or emerging biological systems.

The study was led by the lab of Ratmir Derda, a GlycoNet investigator and professor at the University of Alberta, and involved collaboration with Lara Mahal, another GlycoNet investigator and professor at the University of Alberta, on experimentation and sugar synthesis. GlycoNet and its Services intégrés also provided funding and support for the research.