A machine learning framework to determine geolocations from metagenomic profiling.

Area of Science:

• Microbiology
• Bioinformatics
• Machine Learning

Background:

• Microbial communities in environmental samples are often geolocation-specific.
• Microbiome abundance profiles can serve as unique identifiers for sample origins.
• Metagenomic data analysis offers potential for geographic profiling.

Purpose of the Study:

• To develop and evaluate a machine learning framework for determining sample geolocations from metagenomic data.
• To predict the geographical origins of microbial samples using microbiome fingerprints.
• To assess the framework's performance on multi-source microbiome data.

Main Methods:

• Feature extraction from metagenomic abundance profiles.
• Training prediction models using logistic regression with L2 normalization.
• Developing a novel approach using biological coordinates and Kriging interpolation for predicting origins in unsampled locations.

Main Results:

• Achieved 86% prediction accuracy for known city locations, averaged over 100 random training/validation splits.
• Successfully assigned high probabilities to the true origins of testing samples from previously unsampled cities.
• Demonstrated the framework's robustness in predicting geographic origins using microbiome data.

Conclusions:

• The machine learning framework effectively predicts the geographic origin of metagenomic samples.
• The method shows promise for identifying sample locations even when training data from those specific locations is unavailable.
• This approach advances the field of metagenomic forensics and geolocation prediction.