NMTF

NMTF performs Bayesian Semi-Nonnegative Matrix Tri-Factorization (BSNMTF) to identify pathways associated with cancer phenotypes, including molecular sub-types and treatment outcomes, from real-valued gene expression data containing both positive and negative values.

Key Features:

• Semi-Nonnegative Factorization: Allows one factor matrix (the centroid matrix) to be real-valued so each centroid can represent up-regulation or down-regulation of genes within a pathway.
• Bayesian Framework: Employs structured spike-and-slab priors on the centroid matrix informed by existing pathways and gene-gene interaction (GGI) networks.
• Network-informed Gene Inclusion: Permits inclusion of genes not initially annotated to a pathway if they interact with pathway member genes in the GGI network.
• Uncertainty Quantification: Uses a full Bayesian approach with variational inference to estimate posterior distributions rather than single point estimates.
• Clinical Dataset Validation: Applied to datasets including The Cancer Genome Atlas (TCGA) gastric cancer and metastatic gastric cancer immunotherapy clinical-trial data to identify biologically and clinically relevant pathways.
• Prognostic Biomarker Identification: Identifies pathways validated as prognostic biomarkers that stratify patients with distinct survival outcomes across independent datasets.

Scientific Applications:

• Pathway discovery in cancer: Identification of pathways associated with cancer molecular sub-types and treatment outcomes using real-valued molecular data.
• Biomarker discovery and stratification: Discovery of predictive and prognostic pathway-level biomarkers for patient stratification and treatment response analysis.

Methodology:

Semi-nonnegative matrix tri-factorization within a Bayesian framework; structured spike-and-slab priors informed by pathways and gene-gene interaction (GGI) networks; variational inference for posterior distribution updates.