Discriminating cell line specific features of antibiotic‐resistant strains of Escherichia coli from Raman spectra via machine learning analysis

Abstract

While Raman spectroscopy can provide label‐free discrimination between highly similar biological species, the discrimination is often marginal, and optimal use of spectral information is imperative. Here, we compare two machine learning models, an artificial neural network and a support vector machine, for discriminating between Raman spectra of 11 bacterial mutants of Escherichia coli MDS42. While we find that both models discriminate the 11 bacterial strains with similarly high accuracy, sensitivity and specificity, it is clear that the models form different class boundaries. By extracting strain‐specific (and function‐specific) spectral features utilized by the models, we find that both models utilize a small subset of high intensity peaks while separate subsets of lower intensity peaks are utilized by only one method or the other. This analysis highlights the need for methods to use the complete spectral information more effectively, beginning with a better understanding of the distinct information gained from each model.

Recent Progresses in Machine Learning Assisted Raman Spectroscopy

“…[ 176 ] Zahn et al. [ 177 ] used ANN and SVM models to distinguish the Raman spectra of 11 E. coli strains, respectively. They experimentally demonstrated that both methods could potentially analyze complex spectral datasets for biomedical applications, with SVM showing better results on small datasets and ANN with relatively complex structures showing better results on large datasets.…”

“…KPCA extracts the nonlinear features of the raw data and evaluates and discriminates individual bacterial cells at the serotype level by a DT algorithm with classification accuracy in the range of 87.1%-95.8%. [176] Zahn et al [177] used ANN and SVM models to distinguish the Raman spectra of 11 E. coli strains, respectively. They experimentally demonstrated that both methods could potentially analyze complex spectral datasets for biomedical applications, with SVM showing better results on small datasets and ANN with relatively complex structures showing better results on large datasets.…”

“…They experimentally demonstrated that both methods could potentially analyze complex spectral datasets for biomedical applications, with SVM showing better results on small datasets and ANN with relatively complex structures showing better results on large datasets. [177] Nakar et al [178] first used Raman spectroscopy to successfully differentiate between resistant and sensitive strains of E. coli without exposure to antibiotics. Previous experiments revealed that resistant strains had a higher ratio of nucleic acid to protein.…”

Recent advances of Raman spectroscopy for the analysis of bacteria

“…CNN is suitable for multi‐classification analysis for in‐depth mining and interpretation of Raman spectra. A study compared two machine learning models (an artificial neural network and a SVM) for the discrimination between 11 bacterial mutants of E. coli with Raman spectroscopy 73 . Both models were found to have similarly high sensitivity, specificity, and accuracy.…”

Toward Gene-Correlated Spatially Resolved Metabolomics with Fingerprint Coherent Raman Imaging

“…With sufficient sensitivity to chemical changes, this specificity is ample to identify subtle phenotype changes. Indeed, individual Raman spectra, with acquisition averaged over entire cells, have been used to distinguish differentiated and pluripotent cells, − to identify activated immune cells, − and to discriminate between bacterial subtypes. − Even at a spectral resolution of 30 cm –1 , rapid bacterial phenotype sorting is possible …”