Tensor-Decomposition-Based Unsupervised Feature Extraction Applied to Prostate Cancer Multiomics Data.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• The large p small n problem, where features vastly outnumber samples, is a significant challenge in multiomics data analysis.
• Existing feature selection methods often struggle with the high dimensionality and limited sample size characteristic of these datasets.

Purpose of the Study:

• To propose and evaluate a novel tensor decomposition (TD)-based unsupervised feature extraction (FE) formalism for multiomics datasets.
• To demonstrate the superiority of TD-based FE over conventional supervised and unsupervised methods in addressing the large p small n problem.

Main Methods:

• Application of tensor decomposition (TD) for unsupervised feature extraction (FE) on multiomics data.
• Comparison of TD-based FE with methods including random forest, ANOVA, penalized linear discriminant analysis, non-negative matrix factorization, and principal component analysis (PCA).

Main Results:

• TD-based unsupervised FE significantly outperformed conventional methods on synthetic and real multiomics datasets.
• Genes identified by TD-based FE were enriched for known tissue-related and transcription factor genes.
• The method demonstrated superior feature selection and biological relevance.

Conclusions:

• TD-based unsupervised FE is a powerful and effective approach for the large p small n problem in multiomics data integration.
• This method provides biologically reliable gene selection, advancing multiomics data analysis.
• This study represents the first application of TD-based unsupervised FE to diverse multiomics measurements.