Structural modifications of known antibiotic scaffolds have kept the upper hand on resistance, but we are on the verge of not having antibiotics for many common infections. Mechanism-based discovery assays reveal novelty, exclude off-target liabilities, and guide lead optimization. For that, we developed a fast and automatable protocol using high-throughput Fourier-transform infrared spectroscopy (FTIRS). Metabolic fingerprints of Staphylococcus aureus and Escherichia coli exposed to 35 compounds, dissolved in dimethyl sulfoxide (DMSO) or water, were acquired. Our data analysis pipeline identified biomarkers of off-target effects, optimized spectral preprocessing, and identified the top-performing machine learning algorithms for off-target liabilities and mechanism of action (MOA) identification. Spectral bands with known biochemical associations more often yielded more significant biomarkers of off-target liabilities when bacteria were exposed to compounds dissolved in water than DMSO. Highly discriminative models distinguished compounds with predominant off-target effects from antibiotics with well-defined MOA (AUROC > 0.87, AUPR > 0.79, F1 > 0.81), and from the latter predicted their MOA (AUROC > 0.88, AUPR > 0.70, F1 > 0.70). The compound solvent did not affect predictive models. FTIRS is fast, simple, inexpensive, automatable, and highly effective at predicting MOA and off-target liabilities. As such, FTIRS mechanism-based screening assays can be applied for hit discovery and to guide lead optimization during the early stages of antibiotic discovery.
Keywords: Fourier-transform infrared spectroscopy (FTIR); high-throughput screening; machine learning; mechanism of action; off-target liabilities.
© 2021 Wiley Periodicals LLC.