Adapted tensor decomposition and PCA based unsupervised feature extraction select more biologically reasonable...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Unsupervised feature extraction methods like tensor decomposition and principal component analysis are established for genomic analysis.
• Existing methods face challenges including a limited number of identified genes and P-value distributions deviating from the null hypothesis.
• These limitations impact the reliability of findings in drug repositioning, biomarker identification, and disease gene discovery.

Purpose of the Study:

• To improve existing unsupervised feature extraction techniques for genomic data analysis.
• To address fundamental problems in gene identification, specifically false negatives and P-value distribution deviations.
• To enhance the biological relevance and reliability of selected differentially expressed genes.

Main Methods:

• Optimization of standard deviation to align P-value histograms with the Gaussian null hypothesis.
• Development of improved tensor decomposition and principal component analysis-based methods.
• Elimination of the need for empirical assumptions of negative binomial distributions and dispersion relations.

Main Results:

• Increased number and biological reliability of selected genes compared to previous methods.
• Achieved P-value distributions more coincident with the null hypothesis.
• Enabled selection of differentially expressed genes without requiring specific distribution assumptions.

Conclusions:

• The enhanced methods provide a more robust approach for identifying biologically meaningful genes.
• These improvements offer greater accuracy and reliability in genomic analyses.
• The study advances the field by offering a more principled method for unsupervised gene selection.