Dimension reduction with redundant gene elimination for tumor classification

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Gene expression data analysis is crucial for tumor classification.
• High-dimensional gene expression data from microarrays pose challenges for traditional classifiers due to limited samples and numerous genes.
• Redundant features obscure effective dimension reduction in microarray datasets.

Purpose of the Study:

• To develop an effective dimension reduction strategy for tumor classification using gene expression data.
• To address the challenge of redundant features in high-dimensional microarray data.
• To improve the generalization performance of classifiers by enhancing dimension reduction.

Main Methods:

• A novel metric, DIScriminative Contribution (DISC), was developed to measure feature redundancy by evaluating linear classifiers for each gene.
• DISC considers label information, unlike standard linear correlation, to assess the redundancy of discriminative abilities between features.
• The REDISC (Redundancy Elimination based on Discriminative Contribution) algorithm was proposed to eliminate redundant genes prior to feature extraction.

Main Results:

• The REDISC algorithm effectively reduces feature redundancy by eliminating genes before feature extraction.
• Experimental results on two microarray datasets demonstrate that REDISC improves the performance of dimension reduction.
• The proposed DISC metric offers a more accurate assessment of feature redundancy compared to linear correlation.

Conclusions:

• Eliminating redundant genes before feature extraction enhances tumor classification compared to methods using only feature extraction.
• Supervised redundant gene elimination, as implemented in REDISC, is superior to unsupervised methods like linear correlation coefficients.
• The REDISC algorithm offers a reliable approach to improve classifier generalization performance in high-dimensional gene expression data analysis.