Commentary on "Rapid identification of Streptococcus and Enterococcus species using diffuse reflectance-absorbance Fourier transform infrared spectroscopy and artificial neural networks".

This is an invited review/commentary by the first and last authors of a paper that was the most cited in FEMS Microbiology Letters for 1996, presently showing in excess of 150 citations at Web of Science, and over 200 at Google Scholar. It was the first paper in which diffuse reflectance absorbance FT-IR spectroscopy was used with a supervised learning method in the form of artificial neural networks, and showed that this combination could succeed in discriminating a series of closely related, clinically relevant, Gram-positive bacterial strains. © FEMS 2017.

INTRODUCTION AND BACKGROUND

Goodacre and Kell had long been seeking phenotypic methods for what is commonly known as ‘rapid microbiology’, a term that covers microbiological methods designed to detect and speciate organisms in various samples (and often to establish their sensitivity or otherwise to antibiotics). Such methods would also be of use in taxonomy. As we stated in the opening of the paper (Goodacre et al.1996), ‘The ideal method for the examination of the relationships between bacterial strains would have minimum sample preparation, would analyse samples directly (i.e. would not require reagents), would be rapid, automated, non-invasive, quantitative and (at least relatively) inexpensive (Goodacre and Kell 1996)’.

Following one of its multiple renaissances (Rumelhart, McClelland and The PDP Research Group 1986) (a much greater one is presently in progress; LeCun, Bengio and Hinton 2015; Schmidhuber 2015), Kell had become impressed by the ability of artificial neural networks (ANNs) to ‘learn’ to analyse complex multivariate data, and to ‘generalise’ so as to be able to predict the properties of novel, unseen samples. This essentially involved an approach to multivariate calibration, whether the output was quantitative or a classification, and was recognised as a ‘supervised’ method in which the ANNs would be trained or calibrated with samples for which the answer was known. A powerful theorem (Hornik, Stinchcombe and White 1989) had proven that a suitable network (of ‘arbitrary’—hence possibly unfeasible—size) could affect any such non-linear mapping. This meant (in principle) that, given suitable data, an ANN could theoretically solve any classification or regression problem.

In ca 1991, Kell had been funded by the Biotechnology Directorate of the UK Science and Engineering Research Council, with Horizon Instruments and Neural Computer Sciences, as part of its LINK Scheme in Analytical Biotechnology, to explore the utility of a combination of pyrolysis mass spectrometry (PyMS) and ANNs to provide analyses of complex biological samples. Pyrolysis involves the thermal breakdown of materials in an inert atmosphere, and while it does not sound as though it might be very reproducible, the use of Curie-point heating meant that the pyrolysis was done at a ‘highly’ reproducible temperature and thus the only bonds to break were those labile at temperatures below the Curie point of the metal or alloy. The fragments would then be sent to a low- (unit mass-) resolution mass spectrometer and provide a pattern or fingerprint that could be analysed. Masses collected were from 51–200 mass:charge, so that the instrument effectively produced 150-dimensional data, perfect for the ANNs of the day (given the computational power then available to us). Goodacre had just finished his PhD at Bristol on the PyMS of various bacteria (e.g. Goodacre and Berkeley 1990), and accepted the postdoctoral position that came with this project. The combination proved extremely successful (∼15 publications from the project) (Goodacre and Kell 1996) and Goodacre secured a Wellcome Trust Career Development Fellowship, in the same Department as Kell in Aberystwyth, to pursue PyMS and ANNs for rapid microbiology. Some highlights of the original project included methods for detecting olive oil adulteration (Goodacre, Kell and Bianchi 1992; Goodacre, Kell and Bianchi 1993), recombinant protein expression (Goodacre et al. 1994a) and metabolite overproduction during fermentations (Goodacre et al. 1994b, 1995). The latter paper actually used linear transfer functions (Goodacre et al.1995) to enable extrapolation, now seen as a key element of the much more recent success in training ‘deep’ neural networks (Nair and Hinton 2010; Dahl, Sainath and Hinton 2013).

Based on the success of this basic strategy, Kell had recognised that another means of rapid microbiology might involve an FT-IR instrument ‘flying’ over samples held on a roughened metal plate, where the interrogating light passed through the sample, was reflected by the plate and detected above the plate (in practice, the plate moved, as in a TLC plate scanner). This was funded through another LINK scheme, now by the SERC Chemicals and Pharmaceuticals Directorate, in collaboration with Bruker Instruments UK. It became apparent to us that GSK, who were interested in screening microbial strains for high natural product titres, might like other (bio)pharmaceutical companies be interested in whole-organism phenotyping methods. Kell's long-time collaborator Jem Rowland, from the Aberystwyth Computer Science Department, brought additional numerical heft, while Bruker produced a special instrument for us (based on a TLC plate scanner), which they were subsequently able to sell (after some modifications, in particular using transmission rather than reflective optics). Our use of FT-IR in microbiology was not at all new, this having been pioneered in particular by Naumann and colleagues (Naumann, Fijala and Labischinski 1988; Helm et al.1991), but our innovations were the use of reflectance spectroscopy—which permitted extremely rapid sample changing—and ANNs (with their much greater analytical power). Subsequently (Winson et al.1997; Kell et al.1998), we contrived the name DRASTIC (Diffuse Reflectance Absorbance Spectroscopy Taking In Chemometrics) to describe this combination.

THE PAPER ITSELF

As in most of our papers using multivariate spectroscopies, it became obvious that supervised methods—in which we calibrate or train the computational (‘machine learning’) system with samples for which the answer is known—were massively more powerful than were ‘unsupervised’ or ‘clustering’ methods, which were really only fit for exploratory analyses. The basic reason is that the latter cannot discriminate signal from noise. This paper was no different. As part of his fellowship, Goodacre had secured the services of Éadaoin Timmins as a PhD student, and had teamed up with Dr Paul Rooney, Head of Microbiology at the local hospital in Aberystwyth, and was eager to try the FT-IR method on what was then a slow and difficult taxonomic problem in terms of speciating various Gram-positive pathogens in the hospital, which could then be used for epidemiology. Those chosen were Enterococcus faecalis, Streptococcus pyogenes, S. pneumoniae, S. mitis, S. bovis and E. faecium. The diffuse-reflectance method was relatively little known, though had been shown in previous work (Mitchell 1993) to generate excellent and reproducible spectra, and it did so in the paper. Even with pre-processing, principal components analysis (an unsupervised method) could not fully discriminate the strains, but when the ‘values’ of the principal components were fed into ANNs as the inputs (thereby reducing the dimensionality and speeding up the training hugely), the ANNs were fully able to discriminate these strains in ‘unseen’ samples. That was the essential finding of the paper.

WHERE ARE THEY NOW?

Goodacre and Kell both moved to the University of Manchester in 2002/2003 where they hold Chairs, and work in the Manchester Institute of Biotechnology (http://mib.ac.uk). Kell focuses on microbial dormancy Kell focuses on microbial dormancy (Kell, Potgieter and Pretorius 2015; Kell and Kenny 2016), synthetic biology (Currin et al.2014, 2015; Swainston et al.2014) and pharmaceutical drug transporters (Kell and Goodacre 2014; Kell and Oliver 2014; Kell et al.2015). Goodacre focuses on Raman spectroscopy (Ashton, Hollywood and Goodacre 2015; Muhamadali et al.2015) and an interesting variant called surface-enhanced Raman scattering (SERS) (Jarvis and Goodacre 2004; Subaihi et al.2016; Westley et al.2016), as well as mass spectrometry-based metabolomics (Goodacre et al.2004; Dunn et al.2011; Sayqal et al.2016). Rowland has retired. Timmins is a Senior Technical Officer at NUI Galway where she supports staff and students in many analytical methods including FT-IR spectroscopy. Rooney is a consultant microbiologist in Belfast City Hospital and he and RG have very recently published together assessing Raman, FT-IR and MALDI-TOF-MS for rapid subspeciation of Enterococcus faecium (AlMasoud et al.2016), illustrating that rapid discrimination within this species of bacterium is still very important, even 20 years on.

WHAT HAPPENED SUBSEQUENTLY?

We have continued to develop phenotypic methods for rapid microbiology (e.g. Davey and Kell 1996; Goodacre et al.1998), including in this journal (Goodacre, Heald and Kell 1999). Recognising the power of whole-cell phenotyping methods, which were shortly to be popularised as ‘omics‘, Goodacre and Kell went into metabolomics, with the first paper using the word ‘metabolome’ (Oliver et al.1998) (commentary; Kell and Oliver 2016) showing FT-IR spectra from the same instrument, and Goodacre being Founding Editor of the eponymous journal. Recent examples of our metabolomics work include Begley et al. (2009); Zelena et al. (2009); Dunn et al. (2011, 2015), while Goodacre has also published widely using FT-IR, Raman and SERS (e.g. reviews Ellis and Goodacre 2006; Jarvis and Goodacre 2008; Huang et al.2010; Ellis et al.2012) as well as mass spectrometry (e.g. Dunn et al.2010; Rattray et al.2014), including volatile analysis for non-invasive human infections (Fowler et al.2015).

WHY SO HIGHLY CITED?

In one sense, it is recognised that ‘analytical methods’ can be highly cited (e.g. those for protein content; Lowry et al.1951; Bradford 1976), and looking through the citing papers, most are either on metabolomics or on rapid microbiology. Arguably the main reason for the paper being highly cited was at least partly its pioneering primacy, but mainly that the paper did deliver precisely what it set out to do, i.e. it did indeed provide a method for rapid microbiology that ‘would have minimum sample preparation, would analyse samples directly (i.e. would not require reagents), would be rapid, automated, non-invasive, quantitative and (at least relatively) inexpensive’ (Goodacre et al.1996).