Researchers developed a machine learning pipeline that can predict with 90 to 95 percent accuracy whether bacterial strains will cause disease in poultry, potentially helping the industry prevent costly infections caused by Enterococcus cecorum bacteria.
- Machine learning pipeline achieves 90-95% accuracy in predicting disease-causing poultry bacteria strains
- Enterococcus cecorum causes billions of dollars in annual losses across the poultry industry
- Researchers identified a 20-gene signature associated with pathogenic strains for laboratory testing
- Analysis examined 259 bacterial genomes, including isolates from Arkansas poultry operations and public databases
- Open-source code available on GitHub has been downloaded 400-500 times and adopted by other research groups
A machine learning pipeline that can predict with roughly 90 to 95 percent accuracy whether a bacterial strain is likely to cause disease in poultry could help the industry catch costly infections before they spread.
The work published in Frontiers in Microbiology targeted Enterococcus cecorum, a bacterium responsible for billions of dollars in annual losses across the poultry industry.
The research team analyzed 259 bacterial genomes, including 34 isolates from Arkansas poultry operations and 225 publicly available genomes drawn from the National Center for Biotechnology Information database. Of those, 110 were classified as commensal, or harmless, and 95 as pathogenic.
Aranyak Goswami, an assistant professor with the Center for Agricultural Data Analytics at the Arkansas Agricultural Experiment Station, the research arm of the University of Arkansas Division of Agriculture, led the project in collaboration with Cobb Genetics and funding from the Arkansas Research Alliance.
Mapping gene interactions
Machine learning helped the researchers identify patterns in gene order — dubbed “genetic neighborhoods” by the researchers – to help determine whether the bacteria were pathogenic or nonpathogenic to poultry.
Genomic sequencing helped to identify antimicrobial resistance genes and metabolic pathways, while a tool called Island Viewer detected pathogenicity islands, or regions of the genome associated with disease-causing potential. From there, the team applied a technique called cassette-to-vector analysis, adapted from natural language processing, to study how genes interact with one another rather than examining them in isolation.
"We are not looking at just one particular gene, but we are looking at all the interactions between the genes," Goswami said.
That data was fed into two widely used predictive machine learning tools, which Goswami said were benchmarked against existing methods to confirm their reliability. The analysis also produced a 20-gene signature associated with pathogenic strains, giving researchers a starting point for laboratory testing through PCR and other diagnostic methods.
The underlying code is modular and publicly available on GitHub, where it has logged an estimated 400 to 500 downloads.
"You know the value of a pipeline when it gets adopted in a different system," Goswami said, adding that a separate research group that has applied the method to studying social behavior in bees.
