Simultaneous elucidation of antibiotic mechanism of action and potency with high-throughput Fourier-transform infrared (FTIR) spectroscopy and machine learning

• Published in: Applied microbiology and biotechnology 2021-02, Vol.105 (3), p.1269-1286
• Main Authors: Ribeiro da Cunha, Bernardo, P. Fonseca, Luis, Calado, Cecília
• Format: Article
• Language: English
• Subjects: Algorithms; Anti-Bacterial Agents - pharmacology; Antibiotic discovery; Antibiotics; Antiinfectives and antibacterials; Antimicrobial activity; Antimicrobial potency; Biomedical and Life Sciences; Biotechnology; Cell density; Data analysis; Discriminant analysis; Fourier transform infrared spectroscopy; Fourier transforms; Gaussian process; Infrared spectroscopy; Learning algorithms; Least-Squares Analysis; Life Sciences; Low concentrations; Machine Learning; Mechanism of action (MOA); Methods and Protocols; Microbial Genetics and Genomics; Microbiology; Observations; Optimization; Physiological aspects; Regression analysis; Spectroscopy; Spectroscopy, Fourier Transform Infrared; Spectrum analysis
• ISSN: 0175-7598; 1432-0614
• DOI: 10.1007/s00253-021-11102-7

Este trabalho foi financiado pelo Concurso Anual para Projetos de Investigação, Desenvolvimento, Inovação e Criação Artística (IDI&CA) 2017 do Instituto Politécnico de Lisboa. Código de referência IPL/2017/DrugsPlatf/ISEL The low rate of discovery and rapid spread of resistant pathogens have made antibiotic discovery a worldwide priority. In cell-based screening, the mechanism of action (MOA) is identified after antimicrobial activity. This increases rediscovery, impairs low potency candidate detection, and does not guide lead optimization. In this study, high-throughput Fourier-transform infrared (FTIR) spectroscopy was used to discriminate the MOA of 14 antibiotics at pathway, class, and individual antibiotic level. For that, the optimal combinations and parametrizations of spectral preprocessing were selected with cross-validated partial least squares discriminant analysis, to which various machine learning algorithms were applied. This coherently resulted in very good accuracies, independently of the algorithms, and at all levels of MOA. Particularly, an ensemble of subspace discriminants predicted the known pathway (98.6%), antibiotic classes (100%), and individual antibiotics (97.8%) with exceptional accuracy, and similar results were obtained for simulated novel MOA. Even at very low concentrations (1 mu g/mL) and growth inhibition (15%), over 70% pathway and class accuracy was achieved, suggesting FTIR spectroscopy can probe the grey chemical matter. Prediction of inhibitory effect was also examined, for which a squared exponential Gaussian process regression yielded a root mean square error of 0.33 and a R-2 of 0.92, indicating that metabolic alterations leading to growth inhibition are intrinsically reflected on FTIR spectra beyond cell density.