Assessing reproducibility of matrix factorization methods in independent transcriptomes

• DOI: 10.1093/bioinformatics/btz225
• Publication Date: 2019-04-01T11:36:14Z
• Authors: Laura Cantini; Ulykbek Kairov; Aurélien de Reyniès; Emmanuel Barillot; François Radvanyi; Andrei Zinovyev
• Keywords: Benchmark (surveying); Identification

Matrix factorization (MF) methods are widely used in order to reduce dimensionality of transcriptomic datasets the action few hidden factors (metagenes). MF algorithms have never been compared based on between-datasets reproducibility their outputs similar independent datasets. Lack this knowledge might a crucial impact when generalizing predictions made study others.We systematically test several collected from same cancer type (14 colorectal, 8 breast and 4 ovarian datasets). Inspired by concepts evolutionary bioinformatics, we design novel framework Reciprocally Best Hit (RBH) graphs benchmark for ability produce generalizable components. We show that particular protocol application component analysis (ICA), accompanied stabilization procedure, leads significant increase reproducibility. Moreover, signals detected through method more interpretable than those other standard methods. developed user-friendly tool performing Stabilized ICA-based RBH meta-analysis. apply methodology colorectal (CRC) which 14 can be collected. The resulting graph maps landscape interconnected associated biological processes or technological artifacts. These as clinical biomarkers robust tumor-type specific signatures tumoral cells microenvironment. Their intensities different samples shed light mechanistic basis CRC molecular subtyping.The construction is available http://goo.gl/DzpwYp.Supplementary data at Bioinformatics online.