HyperPrior

HyperPrior integrates biological prior knowledge into a hypergraph-based semi-supervised learning framework to improve classification of gene expression and array-based comparative genomic hybridization (arrayCGH) data for cancer-related analyses.

Key Features:

• Hypergraph-Based Framework: Models samples or features with hypergraphs to capture higher-order relationships beyond pairwise graphs.
• Semi-Supervised Learning Approach: Leverages both labeled and unlabeled data to improve classification accuracy.
• Iterative Optimization Process: Employs a two-step iterative method that alternates between optimizing sample labeling and feature weighting under constraints derived from biological prior knowledge.
• Biological Knowledge Integration: Incorporates interactions among cancer-related genes within protein-protein interaction (PPI) networks for gene expression data and accounts for spatial proximity along chromosomes for arrayCGH data, recognizing that nearby probes often share events like amplifications or deletions.
• Consistent Feature Weighting: Imposes consistent weighting on correlated genomic features to enhance interpretability and relevance within the graph-based learning framework.

Scientific Applications:

• Cancer classification: Applied to classify cancer samples using gene expression and arrayCGH datasets.
• Biomarker identification: Identified discriminative chromosomal regions and subnetworks within PPI networks associated with cancer-related genomic elements.
• Benchmarking versus baseline methods: Demonstrated competitive performance compared to Support Vector Machines (SVMs) and other baseline methods that utilize similar prior knowledge.

Methodology:

Constructs a hypergraph with nodes representing samples or features and hyperedges informed by biological priors; iteratively refines the model via alternating optimization of sample labeling and feature weighting while enforcing consistency with known biological interactions and spatial arrangements.