mixKernel

mixKernel integrates heterogeneous omics datasets using multiple-kernel methods to enable unsupervised exploratory analysis and interpretable feature selection.

Key Features:

• Multiple Kernel Framework: Employs a multiple-kernel approach to integrate various datasets obtained from the same samples for comprehensive exploratory analyses.
• Meta-Kernel Computation: Implements strategies to compute a meta-kernel either by achieving consensus among datasets or by preserving their original topological structures.
• Kernel PCA and Visualization: Uses kernel Principal Component Analysis to visualize sample similarities in a non-linear space across multiple data sources.
• Feature Selection Methods: Implements feature selection formulated as a non-convex optimization with an ℓ1 penalty, solved using proximal gradient descent, to select relevant and less redundant features.
• Interpretability Enhancements: Provides a generic procedure to link kernel PCA results back to the original data to improve interpretability of components.

Scientific Applications:

• Multi-omics Systems Biology: Integrates multi-omics datasets such as TARA Oceans and The Cancer Genome Atlas (TCGA) to retrieve known findings and reveal new sample structures.
• Environmental Microbiomes: Analyzes environmental microbiome data exemplified by the TARA Oceans expedition to characterize community structure.
• Cancer Genomics: Applies to cancer genomics datasets such as TCGA to explore sample relationships across molecular modalities.
• Epidemiological Analysis: Identifies associations such as key governmental measures influencing Covid-19 dynamics.

Methodology:

Computational methods include multiple-kernel integration, meta-kernel computation via consensus or topology-preserving strategies, kernel PCA, feature selection formulated as a non-convex optimization with an ℓ1 penalty solved by proximal gradient descent, and a procedure to map kernel PCA results back to original features.