Supervised pattern recognition techniques (O-PLS-DA) for investigation

Finally, a supervised pattern recognition approach was carried out to elucidate a set of metabolites for each person, which is descriptive for this volunteer. For this approach, O-PLS-DA analysis was performed, where each volunteer was compared to all others. This resulted in 7 different O-PLS-DA models, from which the scores plots are shown in Figure 32. The model parameters from every model revealed good

models (R2Y= 74.7% - 92.6%) and good predictive ability (Q2Y=59.6% - 85.8%) describing the inter-individual variations (Table 9). The high Q2-value suggests a statistical significance.

Figure 32 O-PLS-DA analysis scores plots with all urine samples, using 1 predictive and 1 orthogonal component for each of the 7 models, and plotting the cross-validated (Tcv) and orthogonal (TYosc) scores. Colour-code: (A) volunteer 1, (B) volunteer 2, (C) volunteer 4, (D) volunteer 5, (E) volunteer 6, (F) volunteer 7, (G) volunteer 8. Open circles: all other volunteers, for each model.

Table 9 Model parameters, R2Y and Q2Y for O-PLS-DA models discriminating the urinary spectral profiles of each volunteer to all other volunteers. For each model, 1 predictive and 1 orthogonal component was used.

O-PLS-DA model R2Y Q2Y

volunteer 1 vs. all others 81.7% 70.8% volunteer 2 vs. all others 89.5% 82.5% volunteer 4 vs. all others 74.7% 59.6% volunteer 5 vs. all others 89.1% 82.4% volunteer 6 vs. all others 88.7% 80.2% volunteer 7 vs. all others 84.1% 71.0% volunteer 8 vs. all others 92.6% 85.8%

All resonances with a correlation coefficient greater than 0.2 were regarded as important6, selected as a heuristic cut-off. Consistent differences in a volunteer will result in a relatively higher correlation coefficient. Some of the metabolite resonances known to discriminate samples in the models are related to slight differences in food choice, e.g. use of chewing gum (attributed to the detection of mannitol in volunteer 1b) and consumption of tea instead of coffee (unknown metabolite with resonances at δ 4.51 (s), 6.98 (d), 7.38 (t) in volunteer 2). Also, inter- individual differences in responses to food were observed, such as a lack of taurine excretion after animal protein by volunteer 1b and higher excretion of the unknown compound with a chemical shift at δ 2.90 (t) after fish and beef consumption from volunteer 6. While some urinary metabolites were present in relatively similar concentration across individuals, other metabolites showed a high degree of inter- individual variation. Among these are high creatinine levels in volunteer 8, high glycine levels of volunteer 4, and low hippurate levels in volunteer 5. A summary of all metabolite changes is shown in Table 10.

6 _{Based on small group size as a heuristic cut-off after visual inspection. Orientation at n=25 samples} and p<0.01, correlation coefficient is 0.15. For future analysis including larger datasets standards published by Chadeau-Hyam et al. will be used: Chadeau-Hyam, M., et al., Metabolic Profiling and the

Table 10 Summary of inter-individual differences between 7 volunteers, as revealed by 7 independent O-PLS-DA analyses. Each independent analysis compared 1 volunteer vs. all other volunteers. (# indicates the direction of change, (+) indicates a relatively higher metabolite concentration in this person, (-) a relatively lower concentration). Only resonances assigned to a metabolite are shown here, a full table is listed in the appendix.

metabolite chemical shift multiplicity assignment r2 r#

Vo lu nte e r 1 ( V 1) p-cresol-sulphate 2.35 s CH3 0.35 + phenylacetylglutamine 7.35 t H3,H5 0.37 + indoxylsulphate 7.50 d H7 0.23 + N-methylnicotinamide 8.97 m H6 0.21 + trigonelline 8.84 t H2 0.18 + taurine 3.42 t CH2NH2 0.2 - mannitol 3.84 dd CH2OH 0.29 + N-acetylated metabolite 2.02 s CH3 0.27 - glutamine 2.45 m gammaCH2 0.39 - citrate 2.54 d CH2(ii) 0.34 - DMA 2.77 s CH3 0.23 - TMA 2.86 s CH3 0.66 + V 2 creatinine 4.05 s CH2 0.52 + 2-hydroxy-isobutyrate 1.36 s CH3 0.36 - glutamine 2.45 m gammaCH2 0.22 + V o lu nte e r 4 hippurate _trigonelline 7.55 _9.12 t _s H4 _CH2 0.3 _0.28 +₊ creatine 3.93 s CH2 0.21 + glycine 3.57 s CH 0.55 + N-acetylated metabolite 2.03 s CH3 0.28 + V 5 hippurate 7.55 t H4 0.22 + formate 8.45 s CH 0.2 + 3-hydroxy-isovalerate7 1.27 s CH3 0.35 + Vo lu nte e r 6 2-hydroxy-isobutyrate 1.36 s CH3 0.3 - alanine 1.48 d CH3 0.27 + glutamine 2.14 m betaCH2 0.33 + citrate 2.54 d CH2(ii) 0.25 + DMG8 2.93 s CH3 0.28 +

V 7 2-hydroxy-isobutyrate _{N-acetylated metabolite} 1.36 _2.05 s _s CH3 _CH3 0.3 _0.21 +₊

V 8 alanine 1.48 d CH3 0.35 + citrate 2.54 d CH2(ii) 0.64 + DMA 2.72 s CH3 0.43 + creatinine 4.05 s CH2 0.62 +

7 _{and various unassigned aromatic resonances} 8 _{and various unassigned aromatic resonances}

4.5.5.1 Investigation of the reproducibility within individual’s metabolic space

To test the reproducibility and investigate temporal changes in the individual metabolic phenotype after a few months, one volunteer (volunteer 1) repeated the study after 3 months.

O-PLS-DA analyses were carried out comparing (i) volunteer 1a with all other volunteers, (ii) volunteer 1b with all others and (iii) volunteer 1a with 1b. All three m0dels had a good predictive ability: (i) R2Y=85.3%, Q2Y=75.8%, (ii) R2Y=81.7%, Q2Y=70.8%, (iii) R2Y=93.0%, Q2Y=74.7%. The analyses revealed that volunteer 1 had consistent differences in taurine, TMA, DMG, trigonelline, δ 1.12 (d) and metabolite with chemical shifts at 1.69 (m), 1.79 (m), 1.99 (m) and N-acetylated compounds (δ 2.04 m) compared to the other volunteers, regardless of time of sampling. In addition, temporal differences in the urine profiles of volunteer 1 (i.e. differences between 1a and 1b) included excretion of relatively higher levels of p-cresol-sulphate, phenylacetylglutamine, N-methylnicotinamide of volunteer 1b, and lower amounts of several energy metabolism-related metabolites (citrate, glycine, glutamine, glutamate, alanine) and excretion of mannitol (correlation coefficient r2 >0.2). Mannitol is likely to result from occasional consumption of chewing gum or food products where mannitol was used as sweetener and is also naturally found in some foods [212, 213]. Occasional consumption of chewing gum was reported by volunteer 1b as a change compared to the first trial. Phenylacetylglutamine, p-cresol-sulphate, and N-methylnicotinamide are all known to be influenced by the gut microbiota [214, 215]. Unfortunately, the clinical trial did not include collection of faecal samples and it was therefore not possible to investigate hypothesised changes in the gut microbiota in this study.

4.5.5.2 Investigation of effects of diurnal variation

An O-PLS-DA analysis was carried out to differentiate urinary metabolite profiles from morning, lunch-time, evening and bed-time collections to assess the inter-day contribution to the overall variance in the data set. The model goodness and

predictive ability were fairly weak with a R2Y=0.37 and Q2Y=0.17. No clustering was observed for morning samples and no consistent changes were observed with respect to the evening samples. However, urine samples collected before lunchtime tended to have increased coffee metabolites (δ 4.40 (s), 6.65 (m), 8.78 (d)), and increased resonances with chemical shifts at δ 2.21 (s), 2.48 (d), and hippurate (δ 3.97 (d), 7.55 (t), 7.64 (t), 7.84 (d)) excretion was decreased. Samples collected at bedtime had relatively higher concentrations of metabolites related to animal protein consumption and wine and grapes consumption.

In summary, all known metabolites were related to the dietary pattern. Slupsky et al. found several metabolites to be different between urine samples collected during the morning and the afternoon [216]. These were creatinine, mannitol, dimethylamine, N-methylnicotinamide, xylose and acetone. Their analysis was based on 60 subjects and each person provided two spot urine samples and no dietary standardisation was made. Gavaghan et al. reported diurnal differences in rat urine, comparing day and night samples [217]. Especially TCA cycle intermediates were different, suggesting that urinary excretion profiles reflect differences in metabolic activity between day and night. In this study, only a few day and night differences were observable. It is possible that the strong inter-individual diurnal differences and the dietary regimen masked possible diurnal variations.