A workflow for bacterial metabolic fingerprinting and lipid profiling: application to Ciprofloxacin challenged Escherichia coli

J. William Allwood1,4*†, Haitham AlRabiah1*, Elon Correa1, Andrew Vaughan1, Yun Xu1, Mathew Upton2,5, Royston Goodacre1, 3

1 Manchester Institute of Biotechnology, School of Chemistry, University of Manchester, 131 Princess Street, Manchester, M1 7DN, U.K.

2School of Translational Medicine, Microbiology and Virology Unit, Stopford Building, University of Manchester, Oxford Road, Manchester, M13 9PL, U.K.

3Manchester Centre for Integrative Systems Biology, Manchester Institute of Biotechnology, School of Chemistry, University of Manchester, 131 Princess Street, Manchester, M1 7DN, U.K.

4School of Biosciences, University of Birmingham, Edgbaston,Birmingham,B15 2TT, U.K.

5 School of Biomedical and Healthcare Sciences, Plymouth University, Drake Circus, Plymouth, PL4 8AA, U.K.

* J. William Allwood and Haitham Alrabiah have contributed equally to this publication

† Corresponding author:

Supplementary Methods

1.1 Preparation of Escherichia coli inoculates

The pathogenic E.coli isolates were cultured inLysogenybroth (LB):10g tryptone (Formedia, Hunstanton, U.K.), 5g yeast extract (Amersham Life Sciences, Cleveland, U.S.A.) and 10g sodium chloride (Fisher Scientific Ltd., Loughborough, U.K.), dissolved in 1L of deionised water and autoclaved (121°C, 15 min and 15psi). Each 49 mL flask culturewas inoculated with 1mL of bacterial stock (20% [v/v] glycerol) and incubated at 37°C, in a shaking incubator at 200rpm for 24 h. The overnight cultures (1mL) were diluted with 49mL fresh media and further incubated at 37°C, 200rpm for 1 h. These cultures were washed and diluted with physiological saline (0.9% [w/v] NaCl) to 0.5 Mcfarland standard at optical density (OD) 600nm using a Biomate 5 spectrophotometer (Thermo, Hemel Hempstead, U.K.) and used as experimental inoculates (Figure 1).

1.2 Determining Ciprofloxacin minimal inhibitory concentration (MIC)

Bacterial growth curves were measured using an OD 600nm in a Bioscreen spectrophotometer (Labsystems, Basingstoke, U.K.). The Bioscreen was run at the following settings: 10 min preheating, incubation temperature 37°C, continuous medium shake, measurement interval 10 min and 24 h total experiment. 180 µL of LB was inoculated with 10µL of the experimental inoculate and 10 µL of Ciprofloxacin hydrochloride solution, ranging from 100 to 0.0025 mg/L, in Bioscreen plates. The control samples consisted of 180 µL LB with 10µL of experimental inoculate and 10 µL sterile distilled water then incubated for 24 h as described above.

1.3Escherichia coliculture and antibiotic challenge for metabolic fingerprinting and lipid profiling

18 mL of LB medium was inoculated with 1 mL of experimental inoculate foreach of the four isolates in 100 mL flasks and incubated for 18 h at 37 °C and 200 rpm. The E. coli cultures were challenged with 1 mL of a 20x stock of Ciprofloxacin hydrochloride with 160 and 173 receiving a final dosing concentration of 0.02 mg/L and isolates 161 and 171 receiving a final dosing concentration of 0.3 mg/L. For the control, non-treated samples, the Ciprofloxacin hydrochloride solution was replaced with an equal volume of sterile distilled water (Figure 1). For each E. coli isolate, six biological replicate cultures were prepared, three were challenged with Ciprofloxacin hydrochloride and three were used as controls. The selected concentrations of antibiotic were determined according to the previously established MIC's. After antibiotic challenge the cultures were incubated for a further 18 h at 37ºC and 200 rpm prior to sample collection (Figure 1).

1.4Escherichia colisample collection and quenching for metabolic fingerprinting and lipid profiling

The sample collection and quenching of metabolism followed the procedures developed by Winder et al. (2008) (Figure 1). From each culture, 15 mL was collected and quenched in 30 mL of 60% cold methanol (-48ºC, chilled on dry ice) and mixed quickly. The quenched culture was centrifuged for 10 min at 5000g and -9ºC. The supernatant was removed rapidly, the remaining bacterial pellets were further centrifuged for 2 min andthe residual supernatant removed. At this stage the quench supernatant may be sampled for the purpose of assessing metabolite leakage which is particularly relevant to the analysis of the polar metabolome. The bacterial pellets were stored at -80ºC until metabolite extraction and UHPLC-MS analysis were performed. For FT-IR spectroscopy, a 1 mL sub sample of each bacterial culture was collected and spotted directly on to the FT-IR plate and dried at 50ºC for 30 mins following the method of AlRabiah et al. (2013). A further 1 mL from each culture were collected for OD 600nm determination, the OD measurement provides a factor for sample normalisation to account for differences between the cell densities obtained for the various replicate cultures (Figure 1).

1.5 Fourier Transform Infrared (FT-IR) spectroscopy

A 96-well zinc selenideFTIR plate (Bruker Ltd., Coventry, U.K.) was solvent washed and air-dried at room temperature.20 µL from each culture was spotted and the plate oven dried at 40 ºC for 45 min. High throughput screening (HTS) FT-IR spectroscopic analysis was carried out using a Bruker Equinox 55 infrared spectrometer (Bruker Ltd., Coventry, U.K.) equipped with an HTX™ module according to the method of Winder et al.(2006). FT-IR absorbance spectra were recorded directly from the dried cell biomass in transmission mode. A background spectrum was collected for each measurement from ablank reference well. All spectra were obtained in the 4000–600 cm-1range, and 64 scans were acquired at 4 cm-1 resolution. Spectral acquisition and background subtraction were performed using OPUS software (Bruker Ltd., U.K.) and data directly exported to the R statistical package ( for chemometric analysis.

1.6 Sample extraction for UHPLC-MS lipid profiling

The bacterial cell pellets were suspended in 1 mL of freshly prepared and pre-chilled (-20oC for minimum of 24h) HPLC grade methanol : chloroform solution (1:1). The samples were thoroughly vortexed and shaken for 15 mins. After which, 0.5 mLof coldHPLC grade water was added to the sample extract. The samples were mixed and centrifuged at13,363 xg for 3 mins to aid phase separation.The upper polar phase was removed. To a clean 2 mL microcentrifuge tube,150 uL of the lower non-polar phase was recovered, centrifuged at 13,363 xg and 100 uL recovered to a further clean 2 mL microcentrifuge tube, thus removing any particulate matter. The non-polar phase samples were dried under a nitrogen gas stream prior to transfer to -80°C storage.

1.7 UHPLC-MS analysis

All samples were analysed on the Accela UHPLC system (Thermo-Fisher Ltd. Hemel Hempsted, U.K.) coupled to an electrospray LTQ-Orbitrap XL hybrid mass spectrometry system (ThermoFisher, Bremen, Germany). For UHPLC-MS analysis the samples were reconstituted in 8:2 HPLC grade methanol:water (100ul per OD600nm of 0.32: See Table S1), vortex mixed and centrifuged for 15 minutes at 13,363 xg. A quality control (QC) sample was prepared by combining an equal volume of each extract and vortex mixing thoroughly. Each sample extract was transferred to a single analytical vial with 200 µL fixed insert, capped, stored in the autosampler at 5 °C and analysed within 48 h of reconstitution in both positive and negative ESI modes. The samples were analysed over two separate analytical blocks (respective of ESI polarity), each completely randomised. UHPLC separations were performed with a method identical to that previously described by Dunn et al. (2011) and Wedge et al. (2011). Briefly, 5ul of extract was injected onto a Hypersil GOLD UHPLC C18column (length 100mm, diameter 2.1mm, particle size 1.9µm, Thermo-Fisher Ltd. Hemel Hempsted, U.K.), the UHPLC was operated at a flow rate of 400 µL.min-1, the column was maintained at a temperature of 50°C. The solvent A (HPLC water and 0.1% formic acid) and solvent B (HPLC methanol and 0.1% formic acid) gradient programme was as follows: ESI+ 100% A held for 1 min, 0-100% B over 11 min, 100% B held for 8 min, returning to 100% A over 2 min (total run time of 22 min); ESI- 100% A held for 1 min, 0-100% B over 16 min, 100% B held for 5 min, returning to 100% A over 2 min (total run time of 24 min). The Thermo LTQ-Orbitrap XL MS system was operated under Xcalibur software (Thermo-Fisher Ltd. Hemel Hempsted, U.K.) and precisely following the method described in Wedege et al., (2011). Prior to sample analysis, the LTQ-Orbitrap MS was tuned to optimise detection of ions in the m/z 100-1000 range and calibrated according to the manufacturers predefined methods in both ESI polarities. Data were acquired in the Orbitrap mass analyser operating at a mass resolution of 30,000 (FWHM defined at m/z 400) and a scan speed of 0.4 s. For each analytical block, initially 20 injections of QC sample were performed for column conditioning, after which 5 injections of experimental samples were made, followed by a QC injection. This was repeated until all samples within the block were analysed. Finally three QC injections were made at the end of the block run. Prior to and in-between each analytical block the ESI ion tube and spray deflector were cleaned using 8:2 HPLC grade methanol:water acidified with 1 % formic acid and ultrasonication for 15 min.

1.8 Processing of raw UHPLC-MS profilesand lipid identification

The UHPLC-MS raw data profiles were first converted into a NetCDF format within the Xcalibur software's file conversion programme. Peak deconvolution was performed using the freely available XCMS software ( as described previously (Dunn et al., 2008; Wedge et al., 2011).The XCMS deconvolution results in the production of an MS Excel based XY matrix of mass spectral features (with related accurate m/z and retention time variable pairs) x sample, with peak area inputted where the mass spectral feature was detected in each sample. The data matrix was next signal corrected so that peaks failing quality control (greater than 20% RSD within QC samples across the analytical run) were removed. Each sample was then normalised totheir sum peak area (i.e. TIC normalisation). The putative identification of lipid features detected on the UHPLC-MS platform was performed applying the PUTMEDID-LCMS set of workflows as previously described (Brown et al., 2011). As different lipids can be detected with the same accurate m/z (for example, isomers with the same molecular formula), multiple identifications can be observed for a single lipid feature. A single lipid can also be detected as multiple features, as different types of ion (for example, protonated and sodiated ions). A table of the putative lipid assignments is available in the supplementary information (Table S2).

1.9 Principal component - discriminant function analysis (PC-DFA)

Discriminant function analysis (DFA) is a supervised technique that discriminates groups using a priori knowledge of class membership. The algorithm works to maximise between-group variance and minimise within-group variance (Varmuza and Filzmoser, 2009). In the present work pre-calculated Principal Components (PCs) were used as inputs for DFA and the results were validated by 1000 bootstrap cross-validations. Bootstrap is a re-sampling technique that can be applied as cross-validation to estimate the prediction performance of a model. The basic idea of this method is to select randomly, with replacement, N samples from a set containing exactly N samples. All selected samples, including the repetitions, are then used as training set and the non-selected samples are used as test set (Efron, 1981). One can think of this as having all samples analysed (N = ||X|| for our case) in a bag. A single sample is then taken out of the bag randomly and its number noted – this sample now forms part of the training data, and the sample is placed back into the bag. This random sample picking process is repeated until ||X||samples are in the training set. Some samples will be used multiple times, and on average 63.2% of all of the samples will have been selected for training, the remaining 36.8% are used as the test data. PC-DFA scores and loadings plots, as well as the associated loadings weights for each FT-IR spectral wavelength or UHPLC-MS feature (RT-m/z pair), were computed using the R statistics package (

1.10Feature selection

To identify the most significant metabolites generated by this study we combined the power of two well-known statistical methods. The first is a partial least squares regression (PLS) algorithm and the second is an analysis of variance (ANOVA) statistical test. PLS is a supervised learning method that relates a set of independent variables X (metabolites) to a set of dependent variables Y (the identity of the samples). PLS projects the X and Y variables into sets of orthogonal latent variables, scores of X and scores of Y, so that the covariance between these two sets of latent variables is maximised (Martens and Næs, 1992). The purpose of PLS is to build a linear model Y = XB + E, where B is a matrix of regression coefficients and E represents the difference (error) between observed and predicted Y values (Vinziet al., 2010). The size of the absolute value of the coefficient for each independent variable represents the influence of that variable on the prediction or dependent variable. The higher the absolute value of the coefficient is, the higher the influence of the variable. ANOVA is a statistical method that examines the difference in means across multiple groups of interest (Hairet al., 2005). It is a generalisation of a t-test to more than two groups with the advantage that ANOVA reduces the chances of committing a type I error when comparing more than two groups. For each of the treatment types investigated we compared the control versus treatment samples to determine the top 50 most significantly changed metabolites using the following approach. Firstly, a PLS algorithm was applied to the data and the metabolites were sorted by the magnitude of their corresponding PLS coefficient values from the highest to the lowest. Secondly, the ANOVA statistical test was applied to all metabolites. Finally, starting with the metabolite with the highest PLS coefficient we choose the top 50 metabolites whose p-value computed by ANOVA was less than 0.05. All calculations were again computed using the R statistics package (

Citations

AlRabiah, H., Correa, E., Upton, M. and Goodacre, R. (2013) High-throughput phenotyping of uropathogenicE. coli isolates with Fourier transform infrared spectroscopy. Analyst138: 1363-1369.

Brown, M., Wedge, D., Goodacre, R., Kell, D.B., Baker, P.N., Kenny, L.C., Mamas, M.A., Neyses, L. & Dunn, W.B. (2011) Automated workflows for accurate mass-based putative metabolite identification in LC/MS-derived metabolomic datasets.Bioinformatics 27: 1108-112.

Dunn, W.B., Broadhurst, D., Begley, P., Zelena, E., Francis-McIntyre, S., Anderson, N., Brown, M., Knowles, J.D., Halsall, A., Haselden, J.N., Nicholls, A.W., Wilson, I.D., The Husermet consortium, Kell, D.B., Goodacre, R. (2011). Procedures for large-scale metabolic profiling of serum and plasma using gas chromatography and liquid chromatography coupled to mass spectrometry.Nature Protocols 6: 1060-1083.

Efron, B. (1981) Nonparametric estimates of standard error: The jackknife, the bootstrap and other methods. Biometrika68: 589-599.

Hair J.F., Black B., Babin B., Anderson R.E., Tatham R.L. (2005) Multivariate Data Analysis. Pearson Education 6th ed.

Martens, H. and Naes, T. (1992).Multivariate Calibration.Wiley Press.

Varmuza, K. and Filzmoser, P. (2009) Introduction to Multivariate Statistical Analysis in Chemometrics. CRC Press.

Vinzi E.V., Chin W.W., Henseler J. and Wang H. (Eds.) (2010) Handbook of Partial Least Squares. Springer.

Wedge, D.C., Allwood, J.W., Dunn, W.B., Vaughan, A.A., Simpson, K., Brown, M., Priest, L., Blackhall, F.H., Whetton, A.D., Dive, C. & Goodacre, R. (2011) Is serum or plasma more appropriate for inter-subject comparisons in metabolomic studies? An assessment in patients with small-cell lung cancer.Anal. Chem.83: 6689-6697.

Winder, C.L., Gordon, S.V., Dale, J., Hewinson, R.G. & Goodacre, R. (2006) Metabolic fingerprints of Mycobacterium boviscluster with molecular type: implications for genotype-phenotype links. Microbiology152: 2757-2765.