Ensemble Feature Learning of Genomic Data Using Support Vector Machine.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Identifying key genes for patient classification is vital in bioinformatics.
• Ensemble and bagging methods, like Random Forest, excel at gene selection and classification.
• Support Vector Machines (SVMs) have seen limited application in gene selection.

Purpose of the Study:

• To introduce an ensemble SVM-Recursive Feature Elimination (ESVM-RFE) method for effective gene selection.
• To leverage ensemble and bagging concepts for improved feature ranking and selection accuracy.
• To address imbalanced datasets in gene selection using bootstrap sampling.

Main Methods:

• Developed ESVM-RFE by combining ensemble/bagging with SVM-Recursive Feature Elimination (RFE).
• Utilized bootstrap samples to build multiple SVM models, generating aggregated feature rankings.
• Implemented a backward elimination strategy based on aggregated rankings for feature selection.
• Constructed nearly balanced bootstrap samples to handle imbalanced datasets.

Main Results:

• ESVM-RFE significantly improved classification performance on five microarray datasets.
• Achieved an average 9% higher accuracy than SVM-RFE and 5% higher than Random Forest on a childhood leukemia dataset.
• Exploration of selected genes using Singular Value Decomposition (SVD) revealed significant data clusters.

Conclusions:

• ESVM-RFE offers a robust approach for gene selection, enhancing classification accuracy.
• The method effectively addresses challenges posed by imbalanced datasets.
• Selected genes demonstrate biological significance through clustering analysis.