Predicting protein-carbohydrate binding sites: a deep learning approach integrating protein language model embeddings...

Area of Science:

• Biochemistry
• Computational Biology
• Machine Learning

Background:

• Protein-carbohydrate interactions are vital for numerous biological processes, including inflammation and cell adhesion.
• Experimental methods for identifying these interactions are often costly and time-consuming.
• Computational approaches are needed to efficiently predict non-covalent carbohydrate binding sites.

Purpose of the Study:

• To develop a deep learning model for predicting non-covalent protein-carbohydrate binding sites.
• To explore and integrate sequence-based, structural, and protein language model features.
• To create a robust and accurate computational tool for identifying these binding sites.

Main Methods:

• Developed DeepCPBSite, an ensemble deep learning model based on the ResNet+FNN architecture.
• Utilized datasets from RCSB, UniProt, and CASP, incorporating sequence and structural features.
• Employed protein language model embeddings, feature selection techniques, and SHAP analysis for interpretability.
• Compared structural features from AlphaFold and ESMFold predictions.

Main Results:

• DeepCPBSite achieved 78.7% balanced accuracy and 59.6% sensitivity on the TS53 dataset.
• The model outperformed existing state-of-the-art methods in F1, MCC, and AUPR scores.
• SHAP analysis provided insights into the contribution of structural features based on protein organism information.

Conclusions:

• DeepCPBSite demonstrates superior performance in predicting non-covalent protein-carbohydrate binding sites.
• The model offers a valuable computational resource for biological research.
• Integrating diverse features and advanced deep learning architectures enhances prediction accuracy.