Supervised learning and model analysis with compositional data.

Area of Science:

• Microbiome research
• Machine learning
• Bioinformatics

Background:

• Supervised learning is crucial for high-throughput sequencing data analysis, particularly in microbiome research.
• Existing methods struggle with compositional and sparse data, lacking interpretability or the ability to capture complex signals.
• Linear log-contrast models and black-box machine learning methods have limitations in handling these data characteristics.

Purpose of the Study:

• To introduce KernelBiome, a kernel-based nonparametric framework for regression and classification of compositional data.
• To address the challenges of sparsity and compositionality in microbiome data analysis.
• To provide a framework that can incorporate prior knowledge, such as phylogenetic structure.

Main Methods:

• KernelBiome utilizes a kernel-based nonparametric approach tailored for sparse compositional data.
• The framework incorporates prior knowledge, like phylogenetic relationships, to enhance analysis.
• It captures complex signals, including the zero-structure, and adaptively manages model complexity.

Main Results:

• KernelBiome demonstrates competitive or superior predictive performance across 33 microbiome datasets compared to state-of-the-art methods.
• Novel interpretation quantities are proposed, extending the interpretability of linear models to nonparametric settings.
• The framework's connection between kernels and distances facilitates interpretability and enables data-driven embeddings.

Conclusions:

• KernelBiome offers a powerful and interpretable solution for analyzing sparse compositional microbiome data.
• The framework enhances predictive accuracy while providing novel insights into component contributions.
• KernelBiome is available as an open-source Python package, promoting wider adoption and further research.