Predicting tuberculosis drug resistance with machine learning-assisted Raman spectroscopy

Tuberculosis (TB) is the world’s deadliest infectious disease, with 1.5 million annual deaths and half a million annual infections. Rapid TB diagnosis and antibiotic susceptibility testing (AST) are critical to improve patient treatment and to reduce the rise of new drug resistance. Here, we develop a rapid, label-free approach to identify Mycobacterium tuberculosis (Mtb) strains and antibiotic-resistant mutants. We collect over 20,000 single-cell Raman spectra from isogenic mycobacterial strains each resistant to one of the four mainstay anti-TB drugs (isoniazid, rifampicin, moxifloxacin and amikacin) and train a machine-learning model on these spectra. On dried TB samples, we achieve > 98% classification accuracy of the antibiotic resistance profile, without the need for antibiotic co-incubation; in dried patient sputum, we achieve average classification accuracies of ~ 79%. We also develop a low-cost, portable Raman microscope suitable for field-deployment of this method in TB-endemic regions.