Assessing reproducibility of matrix factorization methods in independent transcriptomes.

Area of Science:

• Bioinformatics
• Genomics
• Computational Biology

Background:

• Matrix factorization (MF) methods are crucial for reducing transcriptomic data dimensionality into latent factors (metagenes).
• The reproducibility of MF outputs across independent datasets has not been systematically evaluated, impacting the generalizability of findings.
• Understanding this reproducibility is vital for reliable transcriptomic data analysis and biomarker discovery.

Purpose of the Study:

• To systematically compare widely used MF methods for their between-dataset reproducibility on transcriptomic data.
• To develop and apply a novel framework using Reciprocally Best Hit (RBH) graphs for benchmarking MF methods.
• To identify MF approaches that yield generalizable and interpretable biological signals.

Main Methods:

• Systematic testing of MF methods on multiple cancer transcriptomic datasets (colorectal, breast, ovarian).
• Development of an RBH graph framework inspired by evolutionary bioinformatics to benchmark reproducibility.
• Application of a stabilization procedure to independent component analysis (ICA) for enhanced performance.

Main Results:

• A specific protocol of ICA with stabilization significantly increased between-dataset reproducibility.
• The stabilized ICA method produced more interpretable biological signals compared to standard MF methods.
• The developed tool facilitates Stabilized ICA-based RBH meta-analysis for identifying robust transcriptomic signatures and biomarkers.

Conclusions:

• Stabilized ICA-based RBH meta-analysis offers a robust approach for reproducible transcriptomic data analysis.
• This methodology can identify clinically relevant biomarkers and tumor-specific transcriptomic signatures.
• The findings provide insights into the mechanistic basis of cancer molecular subtyping and facilitate biomarker discovery.