A Machine Learning Approach in Predicting Antimicrobial Resistance (AMR) in Escherichia Coli (E. Coli)

Antimicrobial resistance (AMR) is a rapidly growing threat to public health, therefore rapid and accurate AMR diagnostic methods are needed. But, traditional culture-based antimicrobial susceptibility testing (AST) has many limits, such as being time-consuming and can be costly. Traditional AST methods are based on culturing bacteria and measuring their growth in the presence of different antibiotics. However, this process can be time-consuming in the case of many types of bacteria; an example is mycobacterium tuberculosis which can up to multiple take months to grow the bacteria in the first place. Additionally, the results of traditional AST can be affected by different growth conditions and AST may not be able to detect all types of antibiotic resistance. Rather than using traditional AST, my machine learning approach could be a faster, more efficient method, by allowing us to identify patterns in the whole genome sequences of the bacterial isolates. I studied the ability of four traditional machine learning models: Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosted Decision Trees (GBDT) to predict AMR in E. coli isolates using publicly available data on GitHub. I selected these four machine learning models because of their ability to handle binary classification tasks and generally their wide use in literature. I focused on four antibiotics: ceftazidime (CTZ), gentamicin (GEN), piperacillin-tazobactam (TZP), and amoxicillin-clavulanic acid (AMC). The data used in my study is genetic data collected from whole genome sequencing (WGS) in the lab [1]. WGS sequencing can compete with traditional techniques in speed, accuracy, and cost to better control infectious diseases and outbreaks. The four algorithms were trained with traditional AST results as their labels. The accuracies of all four models in my study varied from 73% to 99%. The ROC curves and AUC values of the different models range from 0.74 to 0.93 for CTZ, AMC, and GEN, showing that the models can effectively predict AMR. The ROC curves and AUC values for TZP were outliers, ranging from 0.48 to 0.72. In general, my results show that AMR can be predicted computationally rather than through traditional AST. This can help healthcare workers choose which antibiotic would be useful to effectively treat a patient. Furthermore, this method can be implemented on various drugs and predict resistance in different types of bacteria.