MDITRE: Scalable and Interpretable Machine Learning for Predicting Host Status from Temporal Microbiome Dynamics.

Area of Science:

• Microbiome research
• Computational biology
• Machine learning applications in health

Background:

• Longitudinal microbiome studies are crucial for understanding human health and disease mechanisms.
• Existing computational tools for microbiome time-series analysis are limited in scalability and interpretability.
• There is a need for efficient methods to link dynamic microbiome changes to host status.

Purpose of the Study:

• To develop an open-source software package, MDITRE, for analyzing longitudinal microbiome data.
• To implement a novel deep-learning method for deriving interpretable rules from microbiome time-series data.
• To enable prediction of host status based on temporal microbiome patterns.

Main Methods:

• Developed Microbiome Differentiable Interpretable Temporal Rule Engine (MDITRE), an open-source software package.
• Utilized deep-learning technologies for efficient rule derivation from longitudinal microbiome data.
• Validated MDITRE using semi-synthetic datasets and a large compendium of 16S rRNA amplicon and metagenomics data.

Main Results:

• MDITRE performs comparably to or better than uninterpretable machine learning methods on microbiome time-series data.
• MDITRE is orders of magnitude faster than previous interpretable techniques.
• MDITRE's graphical user interface facilitates biologically meaningful interpretations linking microbiome dynamics to host phenotypes.

Conclusions:

• MDITRE overcomes limitations of existing computational methods for microbiome time-series analysis.
• The software enables analysis of larger, more complex datasets with interpretable outputs.
• MDITRE has the potential to drive new insights into microbiome-host interactions and aid diagnostic test development.