Single-cell Raman spectroscopy combined with deep learning for antimicrobial resistance detection in Mycobacterium abscessus.

Mycobacterium abscessus (M. abscessus) is a rapidly growing nontuberculous mycobacterium that poses a serious therapeutic challenge due to its intrinsic and acquired resistance to multiple antibiotics. Conventional antimicrobial susceptibility testing (AST) is slow and labor-intensive, underscoring the urgent need for rapid, culture-free alternatives. Here, we present a phenotype-based approach that integrates deuterium-labeled single-cell Raman spectroscopy with deep learning for the rapid discrimination of antibiotic-resistant M. abscessus. Cells were incubated in medium containing 50% D2O under exposure to clarithromycin (CLA) or linezolid (LZD), and metabolic activity was quantified via the CD ratio. Following optimization, assay conditions were established as 24 h incubation with inhibitory concentrations of 4 mg/L for CLA and 16 mg/L for LZD. A convolutional neural network (CNN) model was developed to analyze the full-range Raman spectra (400-4000 cm-1) acquired from 14 clinical isolates, encompassing both the fingerprint region and the CD band. The CNN demonstrated superior classification accuracy compared to a support vector machine (SVM) model, reaching 96.13% for CLA and 97.12% for LZD, and reliably distinguished resistant from susceptible phenotypes within 24 h. This study establishes a rapid, label-free platform for detecting antimicrobial resistance in M. abscessus based on metabolic phenotyping. The presented approach could be adapted to streamline resistance profiling of clinical isolates, aiding in timely therapeutic decision-making.