Principal component analysis based unsupervised feature extraction applied to budding yeast temporally periodic gene...

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Principal Component Analysis (PCA) based unsupervised feature extraction (FE) is a novel technique applied to bioinformatics.
• Its efficacy in biomarker identification and gene screening is established, but conditions for success and mechanisms of superiority over supervised methods remain unclear.

Purpose of the Study:

• To investigate the application and effectiveness of PCA-based unsupervised FE in a well-studied organism, budding yeast.
• To elucidate the conditions and mechanisms underlying its performance compared to conventional methods.

Main Methods:

• PCA-based unsupervised FE was applied to gene expression profiles of the yeast metabolic cycle (YMC) and yeast cell division cycle (YCDC).
• Performance was compared against sinusoidal fitting and popular unsupervised clustering algorithms.
• Methodological differences and necessary conditions were analyzed by comparing PCA-based FE with fittings to artificial periodic profiles.

Main Results:

• PCA-based unsupervised FE outperformed sinusoidal fitting for YMC and YCDC.
• It enabled feasible biological term enrichment for YMC without assuming periodicity.
• It identified novel periodic profiles in YMC and a concise set of 37 genes for YCDC, where sinusoidal fitting failed.
• PCA-based unsupervised FE was more successful than four popular unsupervised clustering algorithms for YMC.

Conclusions:

• PCA-based unsupervised FE is a valuable and effective unsupervised method for analyzing biological processes like YMC and YCDC.
• This study explains how unsupervised methods, without pre-defined criteria, can outperform supervised methods that rely on human-defined criteria.