Rapid SERS-machine learning-enabled virulence profiling of Acinetobacter baumannii for environmental surveillance

Acinetobacter baumannii (A. baumannii) is a critical multidrug-resistant pathogen capable of environmental dissemination through water and veterinary sources. Rapid analytical methods to identify high-risk strains are essential for One Health surveillance. Here, we report a proof-of-concept surface-enhanced Raman spectroscopy and machine learning (SERS-ML) framework for rapid, label-free virulence-associated profiling of A. baumannii. Interestingly, whole-cell SERS spectra from 20 environmental and veterinary isolates (10 virulent, 10 avirulent) with silver nanoparticles (AgNPs) colloid yielded reproducible biochemical fingerprints in the 400-1800 cm^-1 region. Particularly, virulent isolates exhibited enhanced spectral features corresponding to nucleic acids, proteins, and lipids, reflecting elevated metabolic activity, membrane complexity, and biofilm formation. To address replicate-level data leakage, we evaluated classification models across three validation schemes. While naïve spectrum-level cross-validation yielded inflated accuracies, strain-blocked GroupKFold and leave-one-strain-out (LOSO) cross-validations provided realistic accuracy of 78-88%, with partial least-squares discriminant analysis (PLS-DA) achieving optimal performance under LOSO (87.9% accuracy, receiver operating characteristic-area under the curve (ROC-AUC) = 0.959). Moreover, confounding audits further revealed that geographic and host metadata are partially embedded within spectral signatures. Overall, this study highlights SERS-ML as a promising analytical technique for bacterial risk profiling, while emphasizing the critical necessity of strain-aware validation strategies in spectroscopic machine learning.