Graph Random Forest: A Graph Embedded Algorithm for Identifying Highly Connected Important Features.

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Random Forest (RF) is a robust machine learning algorithm effective for classification and regression, particularly in biological applications with limited sample sizes.
• Gene expression datasets often present a high-dimensional challenge, with significantly more features (genes) than samples (p >> n).
• Standard RF for feature selection can yield scattered important genes, contradicting the biological principle of functional gene network consistency.

Purpose of the Study:

• To develop an enhanced feature selection method that incorporates biological network topology into the Random Forest algorithm.
• To identify functionally connected important genes by leveraging known biological networks.
• To improve the interpretability of feature selection in high-dimensional biological data.

Main Methods:

• Introduction of Graph Random Forest (GRF), a novel algorithm that integrates external biological network information during forest construction.
• GRF identifies important features that form highly connected sub-graphs within the biological network.
• Validation through simulation experiments and application to real-world RNA-seq datasets (non-small cell lung cancer and human embryonic stem cells).

Main Results:

• GRF achieves classification accuracy comparable to standard RF.
• The selected important genes identified by GRF exhibit high connectivity within biological networks, forming interpretable sub-graphs.
• The method demonstrates effectiveness in identifying biologically relevant features.

Conclusions:

• Graph Random Forest (GRF) offers an effective approach for feature selection in biological data by incorporating network topology.
• GRF enhances the biological interpretability of selected features through identification of connected gene sub-graphs.
• The proposed method is a valuable addition to graph-based classification and feature selection techniques in bioinformatics.