Machine-learning assisted antibiotic detection and categorization using a bacterial sensor array

Abstract

With the extensive global use of antibiotics, the problems associated with environmental and food antibiotic residues have significantly increased, necessitating new methods for rapid detection and categorization of compounds with antibiotic activity. In an answer to this need, we report a new platform, bacterial array solid-phase assay (BacSPA), based on monitoring the responses of 15 stress-responsive Escherichia coli sensor strains. These bioreporters, genetically modified by fusing bioluminescence (luxCDABE) reporter genes upstream of stress-induced gene promoters, were inoculated on solidified agar slabs individually mixed with 11 different antibiotics, belonging to 7 mode of action classes. The antibiotic-induced bioluminescence by the different strains generated a distinct response pattern for each antibiotic class. This luminescence pattern was monitored by timelapse photography, and a machine learning algorithm, Multiclass Decision Forest, was applied to train categorization models that either identified the compound or categorized its class. The best model displayed a 65% accuracy for compound identification and 90% for class classification, within three hours of exposing the sensor array to the tested compound. The method also effectively categorized antibiotics at different concentrations: the trained model categorized eight antibiotics at concentrations ranging from 125 ppb to 1000 ppb, with accuracies mostly higher than 70%. The method was further successfully applied for categorizing antibiotics not included in the training. With a more extensive future database, encompassing a broader range of existing antibiotics, this method may be turned into a powerful tool for detecting and categorizing both known and new antibiotic residues in food or environmental samples.