HSIC Lasso

HSIC Lasso performs supervised nonlinear feature selection using the Hilbert–Schmidt Independence Criterion to identify biomolecular variables associated with biological outcomes in high-dimensional genomic and transcriptomic datasets.

Key Features:

• Nonlinear Feature Selection: A supervised, nonlinear feature selection algorithm that captures complex input–output relationships and functions as a convex variant of minimum redundancy maximum relevance (mRMR) to improve parsimony and computational efficiency.
• Block HSIC Lasso: A block-wise, model-free variant that retains more biologically relevant information and avoids the pitfalls of non-convex optimization.
• Data-Type Versatility: Validated on gene-expression microarrays, single-cell RNA sequencing, and genome-wide association studies (GWAS), with strong performance on synthetic and real-world datasets.

Scientific Applications:

• Single-cell RNA sequencing (mouse hippocampus): Identification of genes associated with brain development and function that contribute to neuronal differences in single-cell datasets.
• Biomarker discovery and genomic data analysis: Feature selection in complex biological datasets where nonlinear interactions are prevalent, supporting biomarker detection and exploration of genomic signals.

Methodology:

Applies the Hilbert–Schmidt Independence Criterion within a Lasso-like convex optimization framework and a block HSIC Lasso variant for model-free, supervised nonlinear feature selection.