An explainable machine learning pipeline for prediction of antimicrobial resistance in Pseudomonas aeruginosa.

Abstract

Motivation: Prediction of antimicrobial resistance in Pseudomonas aeruginosa using machine learning and genomic sequences holds the potential to serve as comparable alternatives to laboratory based detection if not better. Additionally, model interpretability can further enhance the potential of these models paving way for their reproducibility.

Results: We have developed a machine-learning based 2-tier pipeline to predict resistance phenotype in P. aeruginosa using only genomic sequences as input in the form of k-mers. Our Decision Tree Model yields an accuracy of 79% and area under the receiver operating characteristic curve of 0.77 with a 70% specificity and 84% sensitivity. We have interpreted the model's predictions using explainable AI as an attempt to bridge the gap between computational prediction and biological insight. Through these interpretations we have gathered antibiotic specific k-mer signatures pushing phenotype towards resistance.

Availability and implementation: The curated dataset and related codes are available on request.