Feature selection and nearest centroid classification for protein mass spectrometry.

Area of Science:

• Proteomics
• Biomedical data analysis
• Computational biology

Background:

• Mass spectrometry is vital for disease diagnosis and biomarker discovery.
• High dimensionality in mass spectrometry data (curse of dimensionality) limits direct application of machine learning.
• Feature selection is essential to reduce dimensionality for effective classification.

Purpose of the Study:

• To evaluate various feature selection techniques for proteomic mass spectrometry data.
• To compare the performance of filter, wrapper, and embedded feature selection methods.
• To identify robust feature selection algorithms for reliable disease diagnosis.

Main Methods:

• Nearest centroid classifier combined with multiple feature selection algorithms.
• Tested univariate filters (t-test, K-S test, P-test), wrapper methods (SFS, modified SBS), and embedded methods (shrinkage, novel boosting).
• Utilized stratified cross-validation with internal leave-one-out cross-validation for evaluation.

Main Results:

• Sequential forward selection and boosted feature selection demonstrated consistent performance across five cancer datasets.
• Some previously reported state-of-the-art algorithms showed inconsistent results when evaluated across multiple datasets.
• The study highlights the importance of robust evaluation procedures for feature selection methods.

Conclusions:

• Many popular feature selection algorithms perform poorly when rigorously evaluated using stratified cross-validation.
• Algorithm evaluation must be conducted on multiple datasets with consistent cross-validation for statistically sound conclusions.
• Robust feature selection is key for reliable biomarker identification and disease diagnosis using mass spectrometry.