Is bagging effective in the classification of small-sample genomic and proteomic data?

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Genomics

Background:

• Bagging is frequently applied to gene-expression and protein-abundance mass spectrometry data classification.
• The justification is often improved performance of unstable, overfitting classifiers in small-sample scenarios.

Purpose of the Study:

• To determine if bagging sufficiently improves classifier performance to surpass stable, nonoverfitting classifiers for small-sample genomic and proteomic data.
• To evaluate the practical utility of ensemble methods in this context.

Main Methods:

• Empirical study using publicly available genomic and proteomic datasets.
• Feature selection using t-test and RELIEF.
• Comparison of bagging with stable classifiers like linear discriminant analysis and 3-nearest neighbors.

Main Results:

• Bagging improved unstable classifiers (CART, neural networks) but not enough to outperform stable classifiers (linear discriminant analysis, 3-nearest neighbors).
• Ensemble methods did not significantly enhance the performance of already stable classifiers.

Conclusions:

• While bagging enhances unstable classifiers, it does not provide a practical advantage over stable classifiers for small-sample genomic and proteomic data analysis.
• The study questions the widespread applicability of bagging for these specific data types and sample sizes.