Zoonoticus: A machine learning model for genomic prediction of zoonotic bacterial strains using virulence, resistance, and mobile genetic elements.

Abstract

Zoonotic diseases continue to rise globally, yet no existing genomic tool integrates virulence, antimicrobial resistance (AMR), and mobile genetic elements to predict zoonotic potential. Here, we present Zoonoticus, a machine learning-based model that classifies bacterial strains as zoonotic or non-zoonotic using whole-genome data. The model was developed using a curated reference database of 37,291 genes across 40 functional categories, encompassing virulence factors, AMR genes, and integrative and conjugative elements (ICEs). More than 5000 bacterial genomes from zoonotic (Brucella abortus, Mycobacterium bovis, Escherichia coli O157:H7, Leptospira interrogans, Salmonella enterica) and non-zoonotic species (Histophilus somni, Mycoplasma bovis, Streptococcus uberis, Trueperella pyogenes, Corynebacterium pseudotuberculosis) were analysed using stringent BLASTn thresholds (≥80 % coverage, ≥95 % identity). A scoring matrix integrating virulence, resistance, and mobility markers was developed to assign tiered threat levels. A Random Forest classifier, trained on binary gene-presence matrices, achieved > 90 % accuracy in differentiating zoonotic from non-zoonotic strains. Model generalizability was validated using 585 Pasteurellaceae genomes, demonstrating high concordance with previously published gene detection results. Notably, the analysis revealed widespread ICE acquisition among non-zoonotic strains, highlighting the evolutionary potential for future zoonotic emergence. Zoonoticus provides the first reproducible, scalable framework that predicts zoonotic risk directly from bacterial genomes, offering a valuable tool for early surveillance, risk assessment, and public health preparedness.