Machine learning to attribute the source of Campylobacter infections in the United States: A retrospective analysis of national surveillance data

Ben Pascoe, Georgina Futcher, Johan Pensar, Sion Bayliss, Evangelos Mourkas, Jessica Calland, Matthew D. Hitchings, Lavin A. Joseph, Charlotte Lane, Tiffany Greenlee, Nicolas Arning, Daniel J. Wilson, Keith A. Jolley, Jukka Corander, Martin Maiden, Craig T. Parker, Kerry K. Cooper, Erica Billig Rose, Kelli L. Hiett, Bonnie Bruce, Samuel K. SheppardJournal of Infection 2024 · DOI

TL;DR

This study leverages machine learning on genomic data from Campylobacter bacteria to pinpoint the sources of human infections in the US. The findings strongly implicate poultry as the primary culprit, offering crucial insights for public health interventions aimed at reducing illness and antimicrobial resistance.

What Problem Does This Paper Address?

Campylobacter is a leading cause of bacterial foodborne illness in high-income countries, including the United States, leading to millions of infections annually, hospitalizations, and even fatalities. Identifying the specific sources of these infections is challenging due to the pathogen's ubiquity in animal hosts and its ability to spread through various routes. Understanding these sources is critical for developing effective public health strategies to prevent infections and combat the growing threat of antimicrobial resistance (AMR). While previous studies have pointed to poultry as a major source in Europe, reliable, routine source attribution estimates for the US have been lacking.

Methods

The researchers employed a multi-pronged approach combining genomic sequencing, machine learning, and probabilistic modeling.

  • Genomic Sequencing: They sequenced a large number of Campylobacter genomes from both human infections and potential animal sources across the US.
  • Pangenome Analysis and Host-Segregating Markers: They analyzed the entire genetic repertoire (pangenome) of the bacteria to identify specific genetic variations (unitigs or genes) that are strongly associated with particular host species. This involved using random forest models and mutual information scores to pinpoint these "host-segregating markers."
  • Machine Learning for Source Attribution: A weighted random forest model was trained using these host-segregating markers to predict the most likely source of human Campylobacter infections.
  • Probabilistic Modeling (iSource): The iSource algorithm was used to probabilistically attribute each human case to a specific source reservoir.
  • Comparison with Existing Methods: The results were validated by comparing them with a core genome multi-locus sequence typing (cgMLST) based source attribution model (aiSource).
  • Antimicrobial Resistance (AMR) Identification: Genomes were screened for known AMR determinants to assess resistance patterns associated with different sources.

Datasets

The study utilized two main types of datasets:

  • Human Clinical Isolates: 8,856 Campylobacter genomes from confirmed human campylobacteriosis cases in the US, collected between 2009 and 2019.
  • Source Isolates: 16,703 Campylobacter genomes from potential sources, including:
  • Chicken (n=9,395)
    • Cattle (n=5,023)
    • Pork (n=1,175)
    • Turkey (n=497)
    • Wild birds (n=365)
    • Environmental sources (n=248)

These datasets were collected through collaborations with US public health agencies (CDC, FSIS, FDA) and supplemented with publicly available global collections. All sequence data has been uploaded to NCBI's Sequence Read Archive (SRA) and associated with specific BioProjects.

Key Insights and Ideas

  • Genomic Signatures of Host Adaptation: The study powerfully demonstrates that Campylobacter genomes carry distinct genetic signatures that reveal their host origin, even across different geographic locations. This is a fundamental principle that underpins the success of genomic source attribution.
  • Machine Learning as a Powerful Tool for Public Health: The application of machine learning, specifically random forests, proved highly effective in identifying these host-specific genetic markers and accurately attributing infection sources. This highlights the growing utility of ML in public health surveillance.
  • Poultry Dominance in US Campylobacter Infections: The consistent finding that poultry is the primary source of human Campylobacter infections in the US is a critical takeaway. This reinforces the need for targeted interventions in the poultry sector.
  • Regional Variation in Sources: While poultry is dominant nationwide, there are regional differences, with some areas showing a higher contribution from cattle or pork, likely linked to local agricultural practices.
  • Link Between Poultry and AMR: The study provides evidence that isolates attributed to chickens, particularly those from specific clonal complexes, are associated with a higher burden of antimicrobial resistance. This underscores the dual benefit of targeting poultry for both infection reduction and AMR control.

Key Results

  • Primary Source Identified: Poultry was identified as the primary source of human Campylobacter infections in the US, responsible for an estimated 68% of cases.
  • Secondary Sources: Cattle contributed approximately 28% of cases, with smaller contributions from wild birds (3%) and pork sources (1%).
  • Attribution Accuracy: The machine learning models achieved high attribution accuracy, with overall accuracy of 90% for both C. jejuni and C. coli.
  • AMR Trends: An increase in multidrug resistance was observed, particularly in isolates attributed to chickens.
  • Regional Poultry Contribution: The proportion of cases attributed to poultry sources ranged from 49% to 87% across different US Health and Human Services (HHS) regions, with 7 out of 10 regions exceeding 70%.

Limitations and Future Work

  • Representativeness of Data: While the study utilized a large dataset, the authors acknowledge that US surveillance is heavily weighted towards poultry, beef, and pork due to perceived risk. Other sources, such as companion animals, may contribute to human infections but their contribution appears small and was not extensively modeled.
  • Model Specificity: The biology of the bacteria imposes limits on attribution specificity. The model relies on the presence of host-segregating markers, which require sufficient periods of isolated evolution in a host. Frequent host switching can erode these signals.
  • Future Work: The authors suggest that ML methods for marker selection and subsequent attribution can be updated in future studies to include a broader suite of potential infection sources. Implementing routine national-scale pathogen genomic surveillance and machine learning source attribution is highlighted as a valuable strategy for monitoring infection trends and supporting interventions for Campylobacter and other pathogens.

Summary updated

Abstract

All sequence reads were uploaded and shared on NCBI's Sequence Read Archive (SRA) associated with BioProjects; PRJNA239251 (CDC / PulseNet surveillance), PRJNA287430 (FSIS surveillance), PRJNA292668 & PRJNA292664 (NARMS) and PRJNA258022 (FDA surveillance). Publicly available genomes, including reference genomes and isolates sampled worldwide from wild birds are associated with BioProject accessions: PRJNA176480, PRJNA177352, PRJNA342755, PRJNA345429, PRJNA312235, PRJNA415188, PRJNA524300, PRJNA528879, PRJNA529798, PRJNA575343, PRJNA524315 and PRJNA689604. Contiguous assemblies of all genome sequences compared are available at Mendeley data (assembled C. coli genomes doi: 10.17632/gxswjvxyh3.1; assembled C. jejuni genomes doi: 10.17632/6ngsz3dtbd.1) and individual project and accession numbers can be found in Supplementary tables S1 and S2, which also includes pubMLST identifiers for assembled genomes. Figshare (10.6084/m9.figshare.20279928). Interactive phylogenies are hosted on microreact separately for C. jejuni (https://microreact.org/project/pascoe-us-cjejuni) and C. coli (https://microreact.org/project/pascoe-us-ccoli).
BibTeX
@article{pascoe2024machine,
  author = {Ben Pascoe and Georgina Futcher and Johan Pensar and Sion Bayliss and Evangelos Mourkas and Jessica Calland and Matthew D. Hitchings and Lavin A. Joseph and Charlotte Lane and Tiffany Greenlee and Nicolas Arning and Daniel J. Wilson and Keith A. Jolley and Jukka Corander and Martin Maiden and Craig T. Parker and Kerry K. Cooper and Erica Billig Rose and Kelli L. Hiett and Bonnie Bruce and Samuel K. Sheppard},
  title = {{Machine learning to attribute the source of Campylobacter infections in the United States: A retrospective analysis of national surveillance data}},
  journal = {Journal of Infection},
  year = {2024},
  doi = {10.1016/j.jinf.2024.106265},
}
Figure