Knowledge-based gene expression classification via matrix factorization.

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Machine learning, specifically matrix decomposition (ICA, NMF), offers advanced tools for gene expression profile analysis.
• These methods extract key features like expression modes (ICA) or metagenes (NMF), indicating regulatory processes.
• Extracted features aid in classifying gene expression datasets and grouping genes for pathway analysis.

Purpose of the Study:

• To apply unsupervised matrix factorization techniques, including ICA and sparse NMF, to microarray datasets.
• To evaluate the ability of these methods to identify relevant biological signatures and marker genes.
• To assess the utility of extracted gene sets for classifying biological samples.

Main Methods:

• Unsupervised matrix factorization (Independent Component Analysis and sparse Non-negative Matrix Factorization).
• Application to microarray datasets of human peripheral blood cells (monocytes to macrophages).
• Utilized leave-one-out and random forest cross-validation for assessing classification accuracy.

Main Results:

• ICA and sparse NMF successfully identified relevant signatures in component matrices.
• Informative sets of marker genes were extracted from gene expression profiles.
• Classification of datasets into diagnostic categories (monocytes vs. macrophages, healthy vs. Niemann Pick C disease) was achieved using gene sets.

Conclusions:

• Matrix factorization techniques are powerful tools for extracting biologically relevant signatures from gene expression data.
• The joint discriminative power of marker gene sets, rather than single genes, is crucial for accurate classification.
• These methods facilitate diagnostic classification and functional gene grouping in biological research.