Machine Learning-Based Prediction of Antimicrobial Resistance in ESKAPEE Pathogens Using Genomic Data

Machine Learning-Based Prediction of Antimicrobial Resistance in ESKAPEE Pathogens Using Genomic Data

Konstantinos Daniilidis1, Anargyros Skoulakis1,2, Stefanos Digenis1,2, Artemis G. Hatzigeorgiou1,2

1 DIANA-Lab, Department of Computer Science and Biomedical Informatics, University of Thessaly, Lamia, Greece, 

2 Hellenic Pasteur Institute, Athens, Greece

∗Correspondence to: arhatzig@uth.gr, skulakis@gmail.com

Background

Antimicrobial resistance (AMR) is a global health issue that poses significant threats to modern medicine. Traditional methods for antimicrobial susceptibility testing are often slow, low-throughput, and require microbial cultures. Advances in next-generation sequencing (NGS), bioinformatics tools, and machine learning (ML) models offer potential solutions for efficient AMR identification. In this study, we analyzed the genomes of 18,916 assemblies of ESCAPEE pathogens (Enterococcus faecium, Staphylococcus aureus, Klebsiella pneumoniae, Acinetobacter baumannii, Pseudomonas aeruginosa, Enterobacter spp, and Escherichia coli) from the NDARO, BV-BRC, and CDC-NARMS databases, along with their corresponding antibiograms, to develop ML models capable of predicting AMR phenotypes from genomic data.

Results

We established a pipeline to identify all proteins associated with antimicrobial resistance within each assembly. We also extracted the 300 bp DNA region upstream of the transcription start site for each of these proteins, as well as the rRNA genes (5s, 16s, and 23s) from each assembly. By integrating data from protein and DNA sequences, we constructed various ML models to predict the AMR phenotype for the antibiotic gentamicin and Amoxicillin-Clavulanic Acid. We employed multiple encoding techniques (k-mers, one-hot encoding, and position-specific scoring matrix [PSSM]) alongside several ML algorithms (SVM, Logistic Regression, Random Forest, XGBoost,  and multilayer perceptron [MLP]). Random Forest algorithm using k-mer encoding of size 3 achieved the highest performance, with an Accuracy of 98.39% for gentamicin and 97.18% for Amoxicillin-Clavulanic Acid, surpassing previous efforts documented in the literature for phenotype prediction.

Conclusions

Leveraging extensive online datasets and advanced ML algorithms, we can develop models for rapid and accurate prediction of AMR phenotypes across a wide range of antibiotics. While this abstract presents results only for gentamicin and Amoxicillin-Clavulanic Acid, our goal is to create models for 90 additional antibiotics and develop a comprehensive suite for predicting AMR phenotypes from bacterial genomic data.