An Adaptive Unsupervised Feature Selection Algorithm Based on MDS for Tumor Gene Data Classification.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate tumor classification and personalized treatment rely on identifying key genes from high-dimensional gene expression data.
• Unsupervised feature selection methods are crucial for discovering biologically significant gene subsets.
• Existing methods often overlook the global data structure, focusing only on local patterns.

Purpose of the Study:

• To develop an adaptive, unsupervised feature selection scheme that considers both global and local data structures.
• To improve the identification of key genes for enhanced tumor classification and treatment decisions.

Main Methods:

• Proposed a novel scheme reconstructing high-dimensional data into a low-dimensional space while preserving feature distance.
• Employed the ℓ2,1-norm within a local manifold structure-learning framework for gene selection.
• Developed an effective algorithm to solve the associated optimization problem.

Main Results:

• The proposed method demonstrated effectiveness in comparative experiments on real tumor datasets.
• The scheme successfully embedded gene selection within a local manifold structure-learning framework.
• Selected genes maintained essential global and local structures of the original data.

Conclusions:

• The developed unsupervised feature selection method offers a significant improvement over classical approaches.
• This technique enhances the ability to identify critical genes for tumor analysis and clinical applications.
• Preserving both global and local data structures is vital for robust gene selection in bioinformatics.