A blocking strategy to improve gene selection for classification of gene expression data.

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• High-dimensional microarray gene expression data presents challenges for machine learning due to the large number of features relative to samples.
• Traditional feature selection methods are computationally intensive and prone to errors in such datasets.
• Feature selection is crucial for effective classification and requires robust evaluation of gene subsets.

Purpose of the Study:

• To propose an original blocking strategy to improve feature selection in high-dimensional microarray gene expression data.
• To enhance the reliability and accuracy of gene subset selection by aggregating validation outcomes from multiple learning algorithms.
• To address the computational hardness and error-proneness of feature selection in datasets with many features and few samples.

Main Methods:

• Interpreting feature selection as a stochastic optimization task to find gene subsets with optimal classification generalization.
• Developing a novel blocking strategy that pairs validation outcomes from multiple learning algorithms to assess gene subsets.
• Comparing the proposed blocking strategy against conventional wrapper methods using 16 cancer expression datasets and six different classifiers.

Main Results:

• The proposed blocking strategy significantly improved the performance of conventional forward selection methods.
• Improvements in feature selection accuracy were observed independently of the classification algorithm used post-selection.
• Biological validation using PubMed abstracts and Gene Ontology analysis supported the enhanced accuracy and quality of selected gene subsets.

Conclusions:

• The developed blocking strategy offers a more robust and accurate approach to feature selection in high-dimensional gene expression data.
• Aggregating validation from multiple algorithms via blocking mitigates issues arising from limited sample sizes.
• This method holds promise for improving the discovery of biologically relevant genes from expression datasets.