Exploring feature limitations in antimicrobial resistance prediction: machine learning and deep learning in A. baumannii
Open Access
- 28 June 2026
- journal article
- research article
- Published by Springer Nature in Scientific Reports
- Vol. 16 (1)
Abstract
Antimicrobial resistance phenotype prediction (AMRPP) provides a computational alternative to conventional susceptibility testing. While most studies focus on single antibiotics, systematic multi-antibiotic modeling and optimal strain representation remain underexplored. Acinetobacter baumannii (A. baumannii), designated by the World Health Organization as a critical priority pathogen, still lacks a comprehensive computational investigation. This study introduces a multi-antibiotic framework for AMRPP in A. baumannii, in which strains are encoded by gene presence/ absence (GPA) profiles. We evaluate whether these features contain sufficient predictive information for classical machine learning (ML) models or require advanced deep learning (DL) architectures. To enhance the representational informativeness and mitigate the high dimensionality of GPA features, four strategies are applied: pathway-based filtering using KEGG, statistical selection by information content (IC), unsupervised reduction via PCA, and model-driven explainability selection with SHAP. Moreover, strong ML baselines: support vector machines, random forests, and extreme gradient boosting (XGB), are benchmarked against a custom DL model, TripSimAcin-AMR, a three-subnetwork siamese neural network tailored for limited data. According to the results, Data-driven representations (PCA, IC, SHAP) outperform KEGG-based filtering, achieving an overall performance of 92.64–93.16%. TripSimAcin-AMR and XGB yield comparable accuracy, but TripSimAcin-AMR achieves higher specificity (85.17% vs. 78.77%). From a clinical perspective, this high specificity is particularly crucial, as false-positive resistance predictions can mislead clinicians to avoid effective first-line antibiotics and unnecessarily escalate to broad-spectrum or last-line therapies. The best configuration, TripSimAcin-AMR with IC-based features, reaches 94.2% accuracy, 94.2% AUC-ROC, 97.0% sensitivity, and 91.3% specificity, underscoring the potential of enriched GPA representations for robust AMR prediction.Keywords
This publication has 52 references indexed in Scilit:
- Inference of antimicrobial resistance (AMR) from a whole genome database outperforming AMR gene detectioniScience, 2025
- AMRLearn: Protocol for a machine learning pipeline for characterization of antimicrobial resistance determinants in microbial genomic dataSTAR Protocols, 2025
- Integrating MALDI-TOF Mass Spectrometry with Machine Learning Techniques for Rapid Antimicrobial Resistance Screening of Foodborne Bacterial Pathogens2024
- Predicting Relative Risk of Antimicrobial Resistance using Machine Learning MethodsIFAC-PapersOnLine, 2022
- Global burden of bacterial antimicrobial resistance in 2019: a systematic analysisThe Lancet, 2022
- A review of artificial intelligence applications for antimicrobial resistanceBiosafety and Health, 2021
- A Deep Learning Approach to Antibiotic DiscoveryCell, 2020
- MALDI-TOF/MS identification of species from the Acinetobacter baumannii (Ab) group revisited: inclusion of the novel A. seifertii and A. dijkshoorniae speciesClinical Microbiology and Infection, 2017
- Deep learningNature, 2015
- Educational interventions to improve antibiotic use in the community: report from the International Forum on Antibiotic Resistance (IFAR) colloquium, 2002The Lancet Infectious Diseases, 2004