A novel method for feature selection based on molecular interactive effect network.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate identification of disease biomarkers is crucial for understanding complex diseases.
• Analyzing molecule interactions can enhance biomarker discovery.
• Existing feature selection methods may not fully capture complex biological relationships.

Purpose of the Study:

• To propose a novel feature selection method, DCG-Net, that incorporates molecule interactive effects.
• To develop a metric (DCG) for measuring pairwise feature interactions relevant to disease.
• To evaluate DCG-Net's performance against established feature selection techniques.

Main Methods:

• Developed Distance Correlation Gain-Network (DCG-Net) using a novel Distance Correlation Gain (DCG) index.
• DCG measures interactive effects between features for physiological and pathological changes.
• Employed a greedy search strategy to identify informative feature modules with high statistical dependence on disease outcome.

Main Results:

• DCG-Net demonstrated superior performance in classification accuracy, sensitivity, and specificity compared to eight other methods across ten datasets.
• Applied to a lung adenocarcinoma dataset, DCG-Net identified key metabolites involved in important pathways with enhanced discrimination ability.
• The method effectively detected molecular interactions, highlighting their importance in biomarker identification.

Conclusions:

• DCG-Net provides an effective approach for feature selection by integrating molecular interaction information.
• Incorporating molecular interactions aids in identifying robust biomarkers for complex diseases.
• The DCG index is versatile, applicable to both discrete and continuous variables for analyzing biological data.