PIntMF: Penalized Integrative Matrix Factorization method for multi-omics data.

Area of Science:

• Genomics and Bioinformatics
• Statistical Modeling
• Computational Biology

Background:

• Multi-omics analyses integrate diverse biological data levels for deeper insights.
• Matrix factorization (MF) is a powerful unsupervised method for clustering individuals and identifying key variables.
• Existing MF methods may lack constraints for optimal biological interpretation and variable selection.

Purpose of the Study:

• To introduce Penalized Integrative Matrix Factorization (PIntMF), an enhanced MF model for multi-omics data analysis.
• To incorporate sparsity, positivity, and equality constraints for improved clustering and interpretability.
• To provide criteria for selecting the optimal number of latent variables.

Main Methods:

• Developed PIntMF, an MF model with Lasso penalization for sparsity and equality constraints for normalization.
• Implemented automatic tuning of sparsity parameters using the glmnet package.
• Compared PIntMF against state-of-the-art integrative methods using synthetic and real multi-omics datasets.

Main Results:

• PIntMF successfully identified relevant clusters and variables in both correlated and uncorrelated simulated data.
• Application to real datasets (Diet and cancer) revealed interpretable clusters linked to clinical data.
• PIntMF demonstrated superior performance in clustering and variable selection compared to existing methods.

Conclusions:

• PIntMF is an efficient and robust tool for extracting patterns and clustering samples from multi-omics data.
• The method facilitates the discovery of novel biological insights and potential biomarkers.
• An R package for PIntMF is publicly available for broader scientific application.