Multi-level gene/MiRNA feature selection using deep belief nets and active learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Selecting discriminative genes/miRNAs is crucial for disease classification and reducing data dimensionality.
• Traditional methods often select features individually, neglecting their collective performance.
• Deep learning offers advanced data representation for improved class discrimination.

Purpose of the Study:

• To propose a novel multi-level feature selection (MLFS) approach for identifying informative genes and miRNAs from expression profiles.
• To leverage deep learning and active learning for enhanced feature selection.
• To extend the MLFS approach for miRNA selection by incorporating gene-miRNA biological relationships.

Main Methods:

• Developed a novel multi-level feature selection (MLFS) approach integrating deep learning and active learning.
• Applied MLFS to gene and miRNA expression profiles for feature selection.
• Extended MLFS to incorporate biological relationships between genes and miRNAs for miRNA selection.

Main Results:

• The MLFS approach significantly improved classification performance across multiple cancer types.
• Achieved performance gains of 9% in hepatocellular carcinoma, 6% in lung cancer, and approximately 10% in breast cancer (F1-measure).
• Demonstrated superior F1-measure compared to classical feature selection methods and recent related work.

Conclusions:

• The proposed MLFS approach effectively enhances disease classification by selecting more discriminative genes and miRNAs.
• Integrating deep learning and active learning provides a powerful framework for feature selection in bioinformatics.
• The extended MLFS approach shows promise for miRNA-based diagnostics and therapeutics by considering gene-miRNA interactions.