High performance Legionella pneumophila source attribution using genomics-based machine learning classification.

Area of Science:

• Microbial genomics
• Epidemiology
• Machine learning applications in public health

Background:

• Legionnaires' disease outbreak control relies on rapid source identification of *Legionella pneumophila*.
• Standard genomic methods face challenges due to *L. pneumophila*'s complex ecology and population structure.
• Accurate source attribution is critical for effective public health interventions.

Purpose of the Study:

• To develop and validate a machine learning (ML) approach for precise geographical source attribution of Legionnaires' disease outbreaks.
• To compare the performance of the ML method against conventional phylogenomic and allelic distance-based classification approaches.
• To leverage genomic variation data for improved outbreak investigation support.

Main Methods:

• Development of ML classification models using 534 *L. pneumophila* genome sequences, including 149 linked to 20 outbreaks.
• Cross-validation framework utilizing environmental *L. pneumophila* genomes for model training.
• Comparative analysis with conventional phylogenomic trees and core genome multi-locus sequence typing (MLST).

Main Results:

• The ML approach demonstrated higher accuracy in geographical source attribution compared to standard methods.
• Models achieved high predictive sensitivity and specificity, with 13 out of 20 outbreaks showing no false positives or negatives.
• The ML method outperformed phylogenomic and core genome MLST approaches in agreement with epidemiological data.

Conclusions:

• The developed ML approach offers a powerful tool for accurate geographical source attribution of Legionnaires' disease outbreaks.
• This method effectively utilizes genomic variation, outperforming traditional techniques.
• The ML approach has significant potential to enhance public health surveillance and expedite outbreak control efforts.