Using graph convolutional neural networks to learn a representation for glycans.

Area of Science:

• Computational Biology
• Glycobiology
• Machine Learning

Background:

• Glycans, diverse and nonlinear biological sequences, are crucial in biological processes but poorly understood computationally.
• Existing computational methods for glycan analysis have limitations in leveraging available information and linking sequences to functions.

Purpose of the Study:

• To develop a novel computational framework, SweetNet, for understanding glycobiology.
• To create a method that explicitly models the nonlinear nature of glycans and maps sequences to representations.
• To improve the prediction of glycan properties and their association with organismal and viral functions.

Main Methods:

• Developed SweetNet, a graph convolutional neural network utilizing graph representation learning.
• Implemented a framework to generate unique representations for any given glycan sequence.
• Applied machine learning techniques for glycan-focused prediction tasks, including viral glycan binding.

Main Results:

• SweetNet demonstrated superior performance compared to existing computational methods in predicting glycan properties.
• Learned glycan representations from SweetNet were found to be predictive of organismal phenotypic and environmental properties.
• The model successfully predicted viral glycan binding, offering potential for discovering viral receptors.

Conclusions:

• SweetNet provides a robust computational approach to address the complexities of glycan sequences.
• The learned representations offer new insights into the functional roles of glycans in biological systems.
• This work facilitates advancements in understanding glycan functions and aids in the discovery of viral receptors.