pipette
ESEspañol

Resistome-based machine learning source attribution of a superbug

C. Mendez, D. Vargas-Peralta, L. A. Rendon, S. Castillo-Ramirez

Preprint

In the authors' words

Many bacterial human pathogens infect not only humans but also other species, and source attribution has been very useful for studying transmission among the different sources. Yet, critical pathogens such as Acinetobacter baumannii have not been studied using source attribution approaches. Here, we conducted machine source attribution using the resistome of almost 1200 A. baumannii genomes from four sources: humans, companion animals , soil, and white storks. Of note, no other study has used the resistome as input data before. First, we selected the best model among 168 parameter combinations, which was Random Forest, with 5-fold cross-validation and chi2 for feature selection. The model shows a reasonably good accuracy with an F1 weighted score = 0.744, yet class accuracy was dependent on the source. Whereas for companion animals and humans, 93% and 86% of the predictions were correct, respectively; the other two categories (soil and white stork) show correct predictions below 50%. Yet, a core phylogeny and ST assignment reveal that the wrong predictions are associated with multi-host lineages. Finally, we identify the most important ARGs contributing to each class prediction and, in the case of companion animals, tetgenes, which confer resistance to tetracycline, were among the top contributors. Our study shows the resistome's potential for developing models to predict source associations. More broadly, this study shows how integrating machine learning and phylogenomics can help us understand multi-host transmission dynamics.

Main resultLimitation the authors admit

Appeared: Thursday, November 19. bioRxiv. Preprint, not yet peer-reviewed.

DOI: 10.64898/2026.11.16.766933