Structure-based neural network protein-carbohydrate interaction predictions at the residue level

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in drug discovery

Background:

• Carbohydrate-protein interactions are vital for cellular processes like recognition and differentiation.
• Predicting these interactions is challenging due to a lack of reliable computational tools.
• Existing methods for identifying carbohydrate-binding sites are limited.

Purpose of the Study:

• To develop novel deep learning models for predicting carbohydrate-binding sites on proteins.
• To compare the performance of different deep learning architectures for this task.
• To assess the utility of these models on predicted protein structures.

Main Methods:

• Developed two deep learning models: a 3D-UNet (CAPSIF:V) and an equivariant graph neural network (CAPSIF:G).
• Trained and evaluated models on predicting non-covalent carbohydrate-binding sites.
• Tested model performance on experimentally determined and AlphaFold2-predicted protein structures.

Main Results:

• Both CAPSIF:V and CAPSIF:G models outperformed previous methods for carbohydrate-binding site prediction.
• CAPSIF:V achieved higher accuracy (Dice score 0.597, MCC 0.599) compared to CAPSIF:G (Dice score 0.543, MCC 0.538).
• CAPSIF:V showed equivalent performance on experimentally determined and AlphaFold2-predicted protein structures.

Conclusions:

• The CAPSIF models offer reliable computational tools for predicting carbohydrate-binding sites.
• CAPSIF:V is a highly effective model, performing well even on predicted protein structures.
• These models can enhance glycan-docking protocols for predicting protein-carbohydrate complex structures.