Identification of human pathogens in soil by virulence gene-based machine learning method

Soils are important reservoirs of human pathogenic bacteria that can spread to humans through various pathways. Metagenomics enables high-throughput pathogen identification by mapping sequencing reads to known pathogen genomes. However, this approach has several limitations, e.g., sequence assembly is time-consuming, and reliance on reference databases may overlook potential pathogens lacking close genomic matches. Here, we developed a novel, virulence factor (VF) based machine learning method using the K-Nearest Neighbors model (VF-KNN) for identifying human pathogenic bacteria from soil metagenomes. Through learning the VF features of pathogenic and non-pathogenic bacteria, VF-KNN could achieve the desired performance in soil pathogen identification (AUC: 0.95, Accuracy: 0.85). Model prediction accuracy (0.95) was further validated using 61 pathogenic strains isolated from soil. For the top 15 most frequent soil pathogens, the prediction accuracy was >0.90 ​at 0.4X-1.0X genome coverage. VFs contributing significantly to pathogen identification were associated with regulation, effector delivery, motility, etc. By using VF-KNN, the averaged abundance of total potential pathogens in topsoils across China was 0.44% (<i>n</i> ​= ​336), predominantly concentrated in the eastern coastal provinces. Compared with the conventional method based on a predefined pathogen list, VF-KNN identified 28% more potential pathogenic species, including some newly reported but not in the predefined list (e.g., <i>Mycolicibacterium cosmeticum</i>). Agricultural land exhibited significantly higher pathogen abundance and diversity than the other land types. This newly developed VF-KNN method is applicable for pathogen identification in broader environments.