Systematic evaluation of supervised machine learning for sample origin prediction using metagenomic sequencing data.

Area of Science:

• Microbial Ecology
• Bioinformatics
• Machine Learning

Background:

• Metagenomic sequencing reveals microbial patterns useful for sample origin prediction.
• Machine learning models accurately predict sample origin when origins are pre-sampled.
• The 2019 CAMDA challenge datasets were used to assess prediction methods.

Purpose of the Study:

• Evaluate the influence of technical, analytical, and machine learning approaches on sample origin prediction.
• Assess the accuracy of predicting novel sample origins.
• Compare regression and classification models for origin prediction.

Main Methods:

• Compared 16S rRNA amplicon and shotgun sequencing, and metagenomic analytical tools (Kraken2, Bracken).
• Employed Lasso-regularized multivariate regression for geographic coordinate prediction.
• Utilized Leave-1-city-out and 10-fold cross-validation to assess model robustness.
• Developed a strategy based on prediction ambiguity for novel origin inference.

Main Results:

• Shotgun sequencing with Kraken2/Bracken showed higher detection sensitivity for microbial abundance.
• Prediction errors were significantly higher in Leave-1-city-out validation, indicating challenges with novel origins.
• Regression and classification models performed comparably on known origins but struggled with new ones.
• Including data from different sequencing protocols increased prediction error.

Conclusions:

• Metagenomics enables accurate sample origin prediction for known origins.
• Predicting novel origins remains a significant challenge for both regression and classification models.
• Prediction ambiguity analysis offers a strategy to identify samples from new origins.
• Sequencing techniques, protocols, and analytical/machine learning methods impact prediction accuracy.