kernInt: A Kernel Framework for Integrating Supervised and Unsupervised Analyses in Spatio-Temporal Metagenomic...

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Next-generation sequencing (NGS) enables microbiome quantification and analysis of spatial-temporal variations.
• Supervised learning is increasingly used for microbiome-based phenotype prediction, but a gap exists with unsupervised ecological analyses.
• Both supervised and unsupervised methods face challenges with compositional NGS data and integrating spatial-temporal dimensions.

Purpose of the Study:

• To propose a unified kernel framework for microbiome analysis, bridging supervised and unsupervised approaches.
• To develop methods for integrating spatial and temporal data within this framework.
• To enable the retrieval of microbial signatures (taxa importances) through this unified approach.

Main Methods:

• Development of a kernel framework integrating supervised and unsupervised microbiome analyses.
• Definition of two compositional kernels: Aitchison-RBF and compositional linear.
• Integration of spatial data using multiple kernel learning and longitudinal data using specific kernels.
• Transformation of non-compositional beta-dissimilarity measures into kernels.

Main Results:

• The proposed kernel framework successfully unifies supervised and unsupervised microbiome analyses.
• Demonstrated integration of spatial and temporal data within the framework.
• Successfully retrieved microbial signatures (taxa importances) across different datasets.
• The framework and case studies are available in the R package 'kernInt'.

Conclusions:

• The kernel framework provides a common ground for diverse microbiome analyses, enhancing interpretability and integration of complex data.
• The approach effectively handles compositional data and integrates spatial-temporal information.
• The 'kernInt' package offers a practical tool for researchers to apply this unified framework to their microbiome studies.