LRAcluster

LRAcluster applies a low-rank approximation-based integrative probabilistic model to perform dimensionality reduction and unsupervised clustering of high-dimensional cancer multi-omics data for identification of molecular subtypes.

Key Features:

• Low-rank integrative probabilistic model: Employs a low-rank approximation-based integrative probabilistic model to jointly model diverse omics datasets.
• Shared principal subspace identification: Identifies a shared principal low-dimensional subspace across diverse omics datasets for joint analysis.
• Dimensionality reduction: Produces an effective low-dimensional representation of high-dimensional cancer data to reduce complexity prior to clustering.
• Convex low-rank regularized likelihood: Optimizes a convex low-rank regularized likelihood function to ensure efficient and stable model fitting.
• Unsupervised clustering: Performs unsupervised clustering within the identified low-dimensional subspace to discover candidate molecular subtypes.
• Computational performance: Exhibits faster execution times and improved clustering outcomes compared to existing methods in tests on various datasets.

Scientific Applications:

• Molecular subtype discovery: Identification of molecular subtypes from cancer multi-omics datasets.
• Pan-cancer analysis (TCGA): Applied to large-scale cancer multi-omics data from The Cancer Genome Atlas (TCGA) for pan-cancer analyses, showing that cancers of different tissue origins generally form independent clusters except squamous-like carcinomas.
• Single cancer-type subtyping: Evaluation of subtyping capabilities within single cancer types across different omics data types.

Methodology:

Finds a shared principal subspace via a low-rank approximation-based integrative probabilistic model by optimizing a convex low-rank regularized likelihood function, followed by unsupervised clustering in the resulting low-dimensional subspace.