A random forest model based on core genome allelic profiles of MRSA for penicillin plus potassium clavulanate susceptibility prediction

Treatment failure of methicillin-resistant Staphylococcus aureus (MRSA) infections remains problematic in clinical practice because therapeutic options are limited. Penicillin plus potassium clavulanate combination (PENC) was shown to have potential for treating some MRSA infections. We investigated the susceptibility of MRSA isolates and constructed a drug susceptibility prediction model for the phenotype of the PENC. We determined the minimum inhibitory concentration of PENC for MRSA (n=284) in a teaching hospital (SRRSH-MRSA). PENC susceptibility genotypes were analysed using a published genotyping scheme based on the mecA sequence. mecA expression in MRSA isolates was analysed by qPCR. We established a random forest model for predicting PENC-susceptible phenotypes using core genome allelic profiles from cgMLST analysis. We identified S2-R isolates with susceptible mecA genotypes but PENC-resistant phenotypes; these isolates expressed mecA at higher levels than did S2 MRSA (2.61 vs 0.98, P<0.05), indicating the limitation of using a single factor for predicting drug susceptibility. Using the data of selected UK-sourced MRSA (n=74) and MRSA collected in a previous national survey (NA-MRSA, n=471) as a training set, we built a model with accuracies of 0.94 and 0.93 for SRRSH-MRSA and UK-sourced MRSA (n=287, NAM-MRSA) validation sets. The AUROC of this model for SRRSH-MRSA and NAM-MRSA was 0.96 and 0.97. Although the source of the training set data affects the scope of application of the prediction model, our data demonstrated the power of the machine learning approach in predicting susceptibility from cgMLST results.