This dataset was generated by GSK using an UPLC system and a hybrid electrospray LTQ/Orbitrap mass spectrometric platform (UPLC-LTQ/Orbitrap-MS) to determine the human metabolic profiles associated with food intake and exercise.
Nine healthy male subjects provided plasma samples. They underwent 4 regimens:
- Regimen A: food+, exercise-
- Regimen B: food-, exercise+
- Regimen C: food+, exercise+
- Regimen D: food-, exercise-
The subjects' plasma were taken at 8 time points (TP) during each regimen. All volunteers were placed on a controlled diet 24 hours prior to the start of each regimen. The first set of plasma samples (TP 0) were taken from individuals in the morning at 8 am after overnight fasting. Thereafter, two groups of volunteers received breakfast (volunteers following regimens A and C; food+ intervention), whilst the other two groups did not consume food at this time (volunteers following regimens B and D; food- intervention). The next set of plasma samples were taken at 9 am (TP 1) and 12 noon (TP 4). Volunteers following the exercise+ intervention (regimens B and C) then exercised for 30 min at 60% of maximum heart rate level, whilst the volunteers following exercise- intervention (regimens A and D) did not undertake any physical activity. The fourth set of plasma samples was taken at 1 pm (TP 5). After this time point, volunteers of all regimens followed the same schedule for food and drink intake, and additional samples were taken at 4 pm (TP 8), 8 pm (TP 12), 11 pm (TP 16) and at 8 am the following day (TP 24).
After completing the first regimen, each volunteer followed a controlled diet for the next 24 hours before starting another randomly selected regimen. After completion of this second regimen, all volunteers were discharged and then re-admitted after two weeks to undergo the remaining two regimens. Each of the regimens was again preceded by a 24 hour controlled diet.
Two identical sets of plasma samples were prepared in randomized order from a total of 288 samples which were collected from the 9 volunteers undergoing 4 regimens at 8 time points. Two 100 μl aliquots were taken from each sample for the preparation of duplicate samples for analysis by the UPLC-LTQ/Orbitrap-MS instrument in ESI+ and ESI- modes. In addition, 40 μl from each sample were used to prepare pooled QC samples. All samples were produced in accordance to the standard preparation protocol described in Section 2.3.
Each identical set of plasma samples was split into 4 analytical blocks and analysed together with a standard metabolite mix, blank and pooled QC samples by the UPLC-LTQ/Orbitrap-MS system (Section 2.5.2). The first set of samples was analysed in ESI- mode, whilst ESI+ mode was used in the analysis of the second set of samples. Ten pooled QC samples were analysed at the beginning of each analytical block. Thereafter, a pooled QC sample was analysed after every fourth plasma sample (Appendix 4).
To avoid any possible bias introduced by the order in which the samples were collected, prepared or analysed, three independent randomizations were performed for choosing the regimen sequence for each volunteer as well as sample preparation and analysis sequence.
All data files were transformed into netCDF format and deconvolved using XCMS software (Section 2.7).
This was the R code used by Eva:
xset <- xcmsSet(step=0.02,snthresh=3,mzdiff = 0.05)
grp <- group(xset,bw=10,mzwid = 0.05)
pc.tmp = ""
an <- annotate(grp, cor_eic_th=0, cor_exp_th=0,na.ok=TRUE)
write.csv(an$annotated,'data0210mzwid05diff05.csv')
How to organise the data files in directories for xcms processing?
- The data produced from the four analytical blocks of plasma within each ionization mode were combined and normalised to the peak areas of pooled QC samples.
- Features showing low reproducibility were removed from each data set (Section 2.8).
- Both normalized data matrices originating from the two ionization modes were joined into a single matrix and the data normalized to TP 0. This normalization procedure enables to partially compensate for the metabolic differences among the volunteers and to identify the changes in metabolome that are relative to the baseline.
This is done using the MATLAB scripts written by Dave Broadhurst
A two-step process was used for the pre-treatment of data originating from the analysis of QC samples. Firstly, any feature vector of UPLC-MS feature responses for all QC samples with more than 40% of its data consisting of missing values was removed. This step was undertaken due to the fact that vectors which are missing over 40% of values show poor repeatability across the experiment and would not represent the actual distribution in normal samples. QC samples are biologically identical and therefore it would be expected that identical results (within an expected error range) would be observed. Secondly, for all the UPLC-MS feature vectors that pass through the first step of pre-treatment, the relative standard deviation (RSD) is calculated as a ratio of population standard deviation to population mean of intensity responses. Feature vectors with RSDs higher than 20% in more than 60% of the samples were considered to show poor repeatability for biologically identical samples and were removed.
Features from QC samples that passed the pre-treatment process are expected to be highly reproducible. Analogous features were taken from data originating from normal samples of serum/plasma and the measured intensity response was normalised to the intensities of corresponding QC features. The process of normalisation involved the division of the median of feature intensity responses measured for QC samples with the intensity response of each feature for a normal sample. These pre-treated data were then submitted for further statistical analysis.
Does the MATLAB scripts written by Dave B do the above?
Is there a R script equivalent?
The subsequent process of data analysis is summarized in Figure 6.1. Statistical analyses using multivariate and univariate methods on pre-treated data were performed using the tools provided in the MATLAB software (version 7.0). The statistical analyses were defined in the MATLAB programming language and these scripts were provided by Dr David Broadhurst. Principal Component Analysis (PCA) (Section 1.4.2.3.5.2) was applied for the reduction of data dimensionality and the first 30 Principal Components were used for Canonical Variate Analysis (Section 1.4.2.3.5.3) The optimal number of Principal Components was selected by random permutation cross-validation, where the model was repeatedly built on 70% of the total sample set (training set) and tested by projection of remaining 30% of samples through the model (test set). For each iteration of the cross-validation process, training/test samples were randomly assigned after stratification by regimen class.
The entire data matrix was then split on the basis of time points to produce four subsets of data corresponding to time points 1, 4, 5 and 8. Furthermore, data from TP 5 and TP 8 were further split on food intervention resulting in a further four subsets of data; TP 5 and TP 8 subsets for food+ intervention and TP 5 and TP 8 subsets for food- intervention. Kruskal-Wallis analysis (Section 1.4.2.3.5.1) together with the calculation of the area under ROC curve (Section 1.4.2.3.3) were performed on each of the six data subsets (TP 1, TP 4, TP 5 food+, TP 8 food+, TP 5 food- and TP 8 food-). The influence of food was examined by undertaking two sets of univariate non-parametric analysis on data from TP 1 and TP 4. A null hypothesis was postulated for the TP 1 and 4 univariate tests by presuming there is no difference in the metabolic profiles of samples from volunteers with or without food intake prior to TP 1 sample collection. The effects of exercise were examined by performing univariate tests on the remaining four sets of data from TP 5 and TP 8. For these cases, a null hypothesis was postulated which presumed there is no difference between samples from volunteers from exercise+ versus exercise- interventions
Putative metabolite annotation was obtained for features with p<0.01 significance levels from the six Kruskal-Wallis tests using the process described in Section 4.2.4. As discussed in detail in Chapter 4, a single metabolite separated by the UPLC system into a single chromatographic peak can be detected by MS as several features corresponding to various adducts, fragments and/or isotopes. Such features originating from one compound were grouped together based on tR similarities, Pearson correlation coefficient values (Section 1.4.2.2.4) and adduct/fragment/isotope annotations. The p-values for all features that passed the data pre-treatment step were added to the results from the automated part of putative metabolite annotation. The semi-automated part of the putative identification process was performed by calculation of Pearson correlation coefficient values and manual retrieval of putative metabolite identification.