Comparative study of unsupervised dimension reduction techniques for the visualization of microarray gene expression...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• DNA microarray data visualization is crucial for quality assessment and hypothesis generation.
• Principal Component Analysis (PCA) is a standard linear method for dimensionality reduction.
• The efficacy of nonlinear dimension reduction methods for microarray data remains largely unexplored.

Purpose of the Study:

• To assess the performance of PCA against six nonlinear dimension reduction techniques for visualizing DNA microarray data.
• To compare the ability of these methods to capture the underlying structure of gene expression data.

Main Methods:

• A systematic benchmark was employed, including Support Vector Machine classification, cluster validation, and noise evaluations.
• Ten microarray datasets and several simulated datasets were analyzed.
• Methods assessed include PCA, Kernel PCA, Locally Linear Embedding, Isomap, Diffusion Maps, Laplacian Eigenmaps, and Maximum Variance Unfolding.

Main Results:

• Significant differences were observed between PCA and most nonlinear methods in 2D and 3D visualizations.
• Performance became similar across methods with increased dimensions and differentially expressed genes.
• PCA and Diffusion Maps demonstrated greater robustness to noise compared to other nonlinear methods.

Conclusions:

• Locally Linear Embedding and Isomap exhibited superior performance across all tested datasets.
• These two methods are favorable alternatives to PCA for microarray data visualization, especially in low dimensions.
• Locally Linear Embedding and Isomap better preserve the underlying data structure when few genes are differentially expressed.