KernelBiome

KernelBiome implements kernel-based nonparametric regression and classification to model compositional, sparse high-throughput sequencing data and provide interpretable supervised learning for microbiome research.

Key Features:

• Kernel-based nonparametric framework: Employs kernel-based regression and classification to capture complex signals in sparse compositional datasets and automatically adapt model complexity.
• Compositional and sparsity handling: Tailors modeling to the characteristics of compositional data, including sparsity and compositional constraints.
• Incorporation of prior knowledge: Integrates prior knowledge such as phylogenetic structure into the kernel framework.
• Interpretability and contribution analysis: Provides two novel quantities that consistently estimate average perturbation effects on the conditional mean, extending interpretability of linear log-contrast coefficients to nonparametric models.
• Kernel–distance connection and embedding: Leverages the relationship between kernels and distances to derive data-driven embeddings that augment downstream interpretation.
• Python implementation: Provided as a Python package for use within computational analysis workflows.

Scientific Applications:

• Microbiome predictive modeling: Applied to microbiome studies to model associations and predict outcomes from high-throughput sequencing compositional data.
• Benchmarking and comparative evaluation: Demonstrated on-par or improved predictive performance compared to state-of-the-art machine learning methods across 33 publicly available datasets.

Methodology:

Uses kernel-based nonparametric regression and classification; integrates prior knowledge such as phylogenetic structure; computes two novel contribution quantities that estimate average perturbation effects on the conditional mean; leverages the kernel–distance relationship to obtain data-driven embeddings and automatically adapts model complexity.