PathME

PathME integrates gene expression, miRNA expression, DNA methylation, and copy number variations (CNVs) to produce pathway-specific patient scores for robust multi-omics patient clustering and molecular subtype identification.

Key Features:

• Multi-Omics Data Integration: Integrates gene expression, miRNA expression, DNA methylation, and CNV data into a unified representation for patient-level analysis.
• Pathway-Based Dimensionality Reduction: Reduces high-dimensional omics data to pathway-specific score profiles that summarize pathway activity per patient.
• Sparse Denoising Autoencoder Framework: Applies multi-modal sparse denoising autoencoders to denoise data and learn sparse latent representations across omics modalities.
• Sparse Non-Negative Matrix Factorization: Uses sparse non-negative matrix factorization in combination with autoencoders to perform dimensionality reduction and factorization of pathway-associated signals.
• Interpretability and Biological Relevance: Maps latent features to pathway scores to identify pathways characteristic of specific patient clusters.
• Impact Disentanglement: Disentangles the contributions of individual omics features to pathway scores to assess each modality's impact.

Scientific Applications:

• Patient Clustering in Cancer: Clusters patients across cancer datasets based on integrated multi-omics pathway profiles.
• Molecular Subtype Identification: Identifies disease subtypes characterized by specific pathway activity patterns to inform precision medicine research.
• Post-hoc Biological Validation: Enables downstream analyses linking identified clusters to somatic mutations and clinical data.

Methodology:

Combines multi-modal sparse denoising autoencoders with sparse non-negative matrix factorization and pathway-based scoring to generate pathway-specific patient score profiles and disentangle per-omics contributions.