Development and validation of a random forest algorithm for source attribution of animal and human Salmonella Typhimurium and monophasic variants of S. Typhimurium isolates in England and Wales utilising whole genome sequencing data

Abstract

Source attribution has traditionally involved combining epidemiological data with different pathogen characterisation methods, including 7-gene multi locus sequence typing (MLST) or serotyping, however, these approaches have limited resolution. In contrast, whole genome sequencing data provide an overview of the whole genome that can be used by attribution algorithms. Here, we applied a random forest (RF) algorithm to predict the primary sources of human clinical Salmonella Typhimurium (S. Typhimurium) and monophasic variants (monophasic S. Typhimurium) isolates. To this end, we utilised single nucleotide polymorphism diversity in the core genome MLST alleles obtained from 1,061 laboratory-confirmed human and animal S. Typhimurium and monophasic S. Typhimurium isolates as inputs into a RF model. The algorithm was used for supervised learning to classify 399 animal S. Typhimurium and monophasic S. Typhimurium isolates into one of eight distinct primary source classes comprising common livestock and pet animal species: cattle, pigs, sheep, other mammals (pets: mostly dogs and horses), broilers, layers, turkeys, and game birds (pheasants, quail, and pigeons). When applied to the training set animal isolates, model accuracy was 0.929 and kappa 0.905, whereas for the test set animal isolates, for which the primary source class information was withheld from the model, the accuracy was 0.779 and kappa 0.700. Subsequently, the model was applied to assign 662 human clinical cases to the eight primary source classes. In the dataset, 60/399 (15.0%) of the animal and 141/662 (21.3%) of the human isolates were associated with a known outbreak of S. Typhimurium definitive type (DT) 104. All but two of the 141 DT104 outbreak linked human isolates were correctly attributed by the model to the primary source classes identified as the origin of the DT104 outbreak. A model that was run without the clonal DT104 animal isolates produced largely congruent outputs (training set accuracy 0.989 and kappa 0.985; test set accuracy 0.781 and kappa 0.663). Overall, our results show that RF offers considerable promise as a suitable methodology for epidemiological tracking and source attribution for foodborne pathogens.