Machine learning and metagenomics reveal shared antimicrobial resistance profiles across multiple chicken farms and abattoirs in China

China's high antimicrobial use in poultry necessitates improved AMR surveillance. Antibiotic use alters the gut resistome, and microbial communities influence AMR phenotypes. Environmental factors like temperature and stress can affect colonization and AMR transfer. Current culture-based approaches may miss ARG diversity. Machine learning (ML) and big data offer tools for precision poultry farming. This study aimed to develop a metagenomic-based surveillance method for Chinese livestock farming, considering resource limitations, using *E. coli* as an indicator species and exploring the impacts of environmental factors and antimicrobial protocols on microbiomes.

Previous research has shown that in-feed antibiotics affect the swine intestinal microbiome, and that *Salmonella typhimurium* can enhance antibiotic tolerance in *E. coli*. The resilience of the intestinal microbiota influences health and disease. Studies have also linked environmental conditions (temperature, humidity) and bacterial species abundance to ARG presence and broiler infections. While AMR surveillance is crucial in non-healthcare settings, it's not widely adopted. ML and big data mining show promise in advancing precision poultry farming. Culture-based approaches combined with whole-genome sequencing (WGS) and ML predict genomic characteristics linked to AMR. However, focusing solely on WGS of individual pathogens may not capture the diversity of resistomes and ARGs.

Ten large-scale commercial poultry farms in three Chinese provinces were sampled over 2.5 years. Samples included pooled faeces and feathers from live birds, barn floor litter, carcasses, processing line materials, wastewater, and outdoor soil. Environmental sensor data (temperature and humidity) were collected. DNA extraction and library construction were performed, followed by Illumina Novaseq 6000 sequencing. Bioinformatics analysis involved read processing, host DNA removal, assembly using MEGAHIT, and taxonomic classification with MetaPhlAn. The resistome was analyzed using BLASTn against the CARD database. Mobile ARGs were identified by analyzing proximity to mobile genetic elements (MGEs). *E. coli* isolates were cultured from a subset of samples and their AMR profiles were characterized. Machine learning models (logistic regression, linear SVM, RBF SVM, extra tree classifier, random forest, AdaBoost, XGBoost) were developed to predict *E. coli* resistance based on metagenomic features (ARG counts and relative abundances of microbial species). Regression models examined correlations between features and temperature/humidity. Antibiotic use was also considered as a potential bias.

The study found 661 different MGE-ARG combinations (potentially mobile ARGs), featuring 195 unique ARGs. Chicken faeces had the highest number of potentially mobile ARGs. 145 MGE-ARG combinations were found in both bird and environmental samples from the same farm, with 46 containing clinically relevant ARGs. *bla*_NDM-5 and *qnrS1* were found in multiple sample types. Analysis of 170 *E. coli* isolates revealed resistance to various antibiotics (1% to 98%). Machine learning identified a core set of 419 features (186 microbial species and 233 ARGs) strongly predicting *E. coli* resistance to 10 antibiotics (AUC > 0.90). 233 ARGs belonged to β-lactams, aminoglycosides, and MLSB classes. 66 ARGs predicted resistance to more than three antibiotics; *aphA6*, *vat(A)*, and *vgb(A)* predicted resistance to eight. 28 microbial species, including *Arcobacter*, *Acinetobacter*, and *Sphingobacterium*, predicted resistance to five antibiotics. 130 ARGs and 48 microbial species correlated with humidity; 39 ARGs and 20 species correlated with temperature. Ten mobile ARGs correlated with *E. coli* resistance and temperature/humidity; 67 correlated with *E. coli* resistance and humidity. Tetracycline, lincosamide, or aminoglycoside use was associated with altered counts for 21 ARGs and 20 microbial species. Evolutionary analysis suggested recent branching of isolates within individual farms, with much earlier MRCAs between different farms.

This study demonstrated the use of metagenomics and ML to reveal complex correlations between microbial communities, resistomes, and AMR in a large-scale setting. The findings show that monitoring a wider range of pathogens than just *E. coli* is more effective for AMR surveillance. The correlation between AMR features and environmental variables (temperature, humidity) offers opportunities for developing novel monitoring solutions, especially in LMICs. The co-localization of AMR genes, potentially facilitated by MGEs, may play a significant role in AMR selection. The observed correlations with antibiotic use highlight the impact of farm practices on AMR development.

This study's metagenomic and ML-based approach provides a comprehensive view of AMR dynamics in Chinese chicken farms. The identified correlations between microbial communities, ARGs, environmental factors, and antibiotic use suggest strategies for improving AMR surveillance and control. Future research could focus on expanding the study to other indicator species, including human samples, and standardizing metagenomic methodologies for wider adoption.

The study focused on *E. coli* as an indicator species and did not include human samples. The generalizability of the findings to other geographical locations and livestock types remains uncertain. Variations in farm practices and environmental controls could influence the results. The study was restricted to a specific region of China and did not necessarily consider the whole country.