Protein structural domain-disease association prediction based on heterogeneous networks

Area of Science:

• Protein structure and function analysis
• Computational biology and bioinformatics
• Genomics and disease mechanisms

Background:

• Protein domains are fundamental units of structure, function, and evolution.
• Multiple domains in large proteins enable complex cellular functions.
• Domain dysfunction is linked to human diseases, necessitating identification of disease-related domains.

Purpose of the Study:

• To develop a computational method for predicting associations between protein domains and human diseases.
• To enhance understanding of the molecular mechanisms underlying complex diseases.
• To identify key protein domains implicated in disease pathogenesis.

Main Methods:

• Construction of a global heterogeneous information network integrating domains, proteins, and diseases.
• Extraction of topological features using meta-paths between domain and disease nodes.
• Training and evaluation of a binary classifier using the XGBoost algorithm.

Main Results:

• The XGBoost-based binary classifier achieved an Area Under Curve (AUC) score of 0.94 in cross-validation.
• The developed method significantly outperformed other machine learning algorithms in predicting domain-disease associations.
• The model demonstrates strong predictive performance on independent test sets.

Conclusions:

• A powerful method for predicting potential domain-disease associations has been developed.
• Topological features and meta-path analysis are effective for this prediction task.
• Integrating multi-omic data offers future potential for optimizing predictive accuracy.