GlyNet: a multi-task neural network for predicting protein-glycan interactions.

Area of Science:

• Computational biology
• Glycoscience
• Biochemistry

Background:

• Understanding glycan-protein interactions is crucial for advances in medicine, including diagnostics, therapeutics, vaccines, transfusion, and organ transplantation.
• Existing models often provide binary classifications, limiting the depth of information on binding strengths.

Purpose of the Study:

• To develop GlyNet, a novel computational model for predicting the relative binding strengths between mammalian glycans and glycan-binding proteins.
• To provide a more quantitative and detailed prediction of protein-glycan interactions compared to existing methods.

Main Methods:

• Developed GlyNet, a multi-output regression model trained on relative fluorescence units (RFUs) from 599 glycans on Consortium for Functional Glycomics arrays.
• The model predicts continuous interaction strengths for 352 glycan-binding proteins across various concentrations.
• Extrapolated predictions to untested glycans based on learned RFU values.

Main Results:

• GlyNet accurately predicts relative binding strengths between mammalian glycans and 352 glycan-binding proteins.
• The model outputs continuous values for 1257 interactions per glycan input, offering more detail than binary classification.
• A GlyNet classifier, using a simple threshold on continuous outputs, outperformed standard binary classification models.

Conclusions:

• GlyNet is the first multi-output regression model for predicting protein-glycan interactions.
• The model serves as a benchmark for quantitative computational glycobiology.
• GlyNet facilitates the development of more sophisticated tools for understanding and utilizing glycan-protein interactions in biomedical applications.