Evidence of different metabolic phenotypes in humans

Results and Discussion

In this study, we collected ≈40 urine samples of 22 healthy individuals (11 females and 11 males for a total of 873 samples) with ages in the 25–50 range over a period of 3 months. The corresponding one-dimensional ¹H NMR spectra were measured on a spectrometer operating at 600 MHz proton Larmor frequency following standardized procedures [see supporting information (SI) Methods]. The spectra were first normalized to the NMR signal intensity of the CH₃-group of creatinine to compensate for the large variations in urine metabolite concentrations and then segmented into N consecutively integrated spectral regions (buckets) of fixed width. If not stated differently, the bucketing ranges were set to a 0.5- to 9.5-ppm chemical-shift range, and the bucket width was set to 0.02 ppm. The 4.5- to 6.0-ppm chemical-shift region was left out of the analysis to remove the effects of variations in the suppression of the water resonance and variations in the urea signal caused by partial cross-solvent saturation through solvent-exchanging protons.

The raw dataset was then analyzed by applying standard statistical methods or combinations thereof, as detailed in Statistical Methods and SI Methods. On the descriptive level, principal component analysis (PCA), hierarchical cluster analysis (HCA), and canonical analysis (CA) were used, where multivariate analysis of variance (MANOVA) was used to determine the dimensionality of the relevant CA subspaces. On the predictive level, classification rules were derived on the basis of soft independent modeling of class analogy (SIMCA) and K-nearest neighbor (K-NN) methods. Several classification approaches for multiclass problems were used combining statistical techniques, such as a multi-model SIMCA classification (MM-SIMCA), a PCA subspace K-NN classification (PCA/K-NN), and a PCA/CA subspace K-NN classification (PCA/CA/K-NN). The identification of an individual person was performed either from a single new sample/spectrum (single vote) or from a set of new samples/spectra according to the majority vote—i.e., majority rule classification (MRC). The predictivity of the various classification methods was assessed by using test-set validation (TSV) combined with a Monte Carlo (MC) approach.

The first objective of the analysis was to ascertain whether the spectral line-patterns of the individual samples carry unique features that are donor-specific. For this purpose, the individual objects of the data matrix were labeled according to donor identity. For dimension reduction, the data were projected into a PCA subspace explaining 99.9% of the variance in the data. The respective score submatrix was used as input for MANOVA to obtain the dimensionality of the multivariate group means, which resulted to be 21 (with all p values <10⁻⁸). The score matrix was projected into the respective 21-dimensional subspace with maximum group discrimination provided by CA. The results are illustrated in Fig. 1, where a convex hull for each donor, rather than the 37–41 individual data points enclosed, is reported in the space of the first three canonical variables.

It is immediately clear (Fig. 1) that a very strong individual characterization is possible, allowing even discrimination with respect to individual donors [even if not looked for, some “natural” gender-clustering is also obtained (see SI Fig. 6)]. It should be noted that no separation according to donors can be found when looking at simple PCA scores plots (SI Fig. 7).

Although a limited overlap within a few pairs of clusters is present in Fig. 1, it should be kept in mind that the discriminating subspace is 21-dimensional according to the MANOVA output, so that a 3-dimensional plot is not sufficient to provide the complete picture. This becomes apparent from the results of HCA applied to the canonical variables, where the complete 21-dimensional CA subspace is analyzed. The dendrogram in Fig. 2 illustrates how, with the only exception of two spectra, all of the remaining spectra are clearly clustered according to donor identity. It is striking to notice how intergroup distances are by far larger than intragroup distances.

It should be noted that the intergroup separation found by simple PCA in combination with HCA is rather low, although some clustering begins to appear (SI Fig. 8). This is due to the fact that the principal components are aligned along the directions of maximum variance, which do not necessarily coincide with the directions of maximum intergroup discrimination.

The separation in the PCA/CA subspace found among the data of individual donors (Fig. 2) is impressive. However, this analysis is on a purely descriptive level. The significance of these findings ultimately depends on the possibility of correctly predicting the donors from unknown urine samples not included in any model. It is shown below that an excellent predictive power can be indeed achieved. To this end, we have applied MC/TSV to assess the predictivity of three classification approaches—i.e., MM-SIMCA, PCA/K-NN, and PCA/CA/K-NN—using 1,000 MC steps for each. In each of the MC steps, the dataset was randomly divided into a model set (30 spectra per donor) and a test set (remaining spectra ranging from 7 to 11 depending on the total amount of samples available from each donor).

The results are summarized in Fig. 3, where a, c, e, and g refer to results obtained by individual classification of test-set spectra (single vote), and b, d, f, and h refer to MRC results (majority vote). The PCA/CA/K-NN approach, which is the most powerful, permits a correct classification of the individual spectra at a probability >0.95 for the spectra of all donors (15 at a probability >0.99) (Fig. 3e and SI Table 1). When using a group of spectra and MRC for donor identification instead of a single spectrum, the overall results are further improved (cf. Fig. 3 a with b, c with d, and e with f). Also in this case, the PCA/CA/K-NN method is the most powerful because this approach achieves a correct classification for 100% of the MC runs (Fig. 3f and SI Table 1). In summary, the results of Fig. 3f clearly demonstrate that a natural metabolic phenotype characteristic of each individual can be defined. Moreover, it is stable and invariant over at least a period of 2–3 months (the time over which the urine samples were collected).

A legitimate question is whether a similar result could be obtained not by fingerprinting but rather by quantitating specific metabolites, and how large their number should be to obtain comparable results. We have provisionally addressed this point by a limited metabolite quantitation in our dataset. Up to 40 different metabolites could be identified, and 12 of them were clearly apparent in all spectra (see SI Methods). Surprisingly, donor identification is largely possible already when using the intensities of these 12 metabolites as input rather than using a complete bucket table. When using the PCA/CA/K-NN model tested by using MRC (Fig. 3h), 7 donors are always recognized correctly, 10 donors are identified correctly in 95% of the cases, and 5 donors are recognized correctly in at least 80% of the 1,000 MC runs.

It should be noted that the probability of correct classification from the concentrations of the 12 metabolites from single samples by PCA/CA/K-NN is always lower, and in some cases significantly lower (see SI Table 1) with respect to MRC, ranging from 41.2% for donor BG (81.1% in MRC) to 96.8% for donor AR (100% in MRC). We recall that when using the complete spectra for classification from BG and AR rather than just 12 metabolite concentrations, the probability of correct single-sample classification increases up to 97.3% and 99.8%, respectively. This illustrates that the donor-specific metabolic fingerprint is a real multiparametric fingerprint, in which the whole detectable metabolome contributes. The more spectroscopic descriptors used in the data analysis, the better the identification of unique personal spectroscopic fingerprints.

It is also important to estimate the minimum number of spectra required to obtain donor-specific fingerprints. Essentially, it is necessary to define a learning curve indicating the percentage of correct donor identification as a function of the number of available model spectra per donor. Considering the powerful performance of the PCA/CA/K-NN model, we calculated the learning curve for this particular model. The MC/TSV approach (1,000 runs) was used, and the model sets were constituted by 5, 10, 15, 20, 25, and 30 spectra, respectively, randomly selected out of the 37–41 spectra for each donor. At variance with the previous case, the dimensionality of the respective PCA subspaces needed to be defined between 98% and 99.9% explained variance for numerical reasons. The predictive performance of each model was tested again by using for each donor both a single spectrum (Fig. 4a) and an MRC performed on the basis of seven spectra randomly selected from those not in the model set (Fig. 4b). It can be noted that in the MRC case, only 15 spectra are needed to obtain a prediction at a probability very close to 1 for each subject (median = 1, mean = 0.9994).

The present analysis has revealed, by subject discrimination on a very high predictive level, that in each urine spectrum there are highly donor-specific traits, providing a fingerprint characteristic of the donor that is consistent over time—at least over the collection time (2–3 months) used. Such donor-specific traits can in principle be reconstructed from an existing sample set coming from a fixed set of donors. To illustrate this concept, the original dataset (bucket table) and a reconstructed bucket table are shown in Fig. 5 a and b, respectively. For reconstruction of each bucket spectrum in Fig. 5b, its representation in the 21-dimensional discriminating subspace described above (the dimension that was obtained, as already indicated, by MANOVA output) was back-projected into the original space. Comparison between original and reconstructed datasets confirms that the procedure had just a “cleaning” effect without creating any new artificial features.

It is even possible to obtain, by applying a similar procedure (see SI Methods), median reconstructed spectra for individual donors (Fig. 5c). Such spectra may be viewed as representing an image of the individual metabolic fingerprint (somewhat related to “metabotype”) and by which it is possible to assess easily, even from visual inspection, donor-specific spectral differences. For example, it is possible to note that subject AU is characterized by persistently low intensity of the doublet at ≈7.36 ppm and of the two triplets at ≈7.45 and 7.74 ppm (i.e., very low levels of hippuric acid despite the absence of antibiotic treatment), whereas subject BI presents higher intensity of the multiplet at ≈1.73 ppm (higher lysine levels) than the other individuals. A graphical representation of the variability of each spectral point and of its discrimination ability is given in SI Fig. 9. However, identification of just one particular metabolite for discrimination of a single person seems to be the absolute exception (see SI Fig. 10). When analyzing the concentration levels of the 12 selected metabolites, apart from donor AU, no single metabolite can be used for unique identification of any other donor in the panel. Instead, the unique fingerprints are coded in the levels of multiple markers. For example, donor AZ is unique because of a combination of high levels in seven metabolites—i.e., 3-hydroxyisovalerate, citrate, glycine, isoleucine, lactate/threonine, leucine, and valine. In another example, median concentrations of donor AO with respect to creatinine, 3-hydroxyisovalerate, alanine, citrate, dimethylamine, glycine, isoleucine, lactate/threonine, leucine, and valine are consistently at the lower end within the panel. In both examples, the combined concentration values are quite unique, leading to a high single-sample correct classification rate even on the basis of the 12 selected metabolites (AZ, 90.4%; AO, 87.8%), and in the MRC approach AZ and AO can be correctly recognized in practically 100% of the cases. Unfortunately, available metadata are not sufficient for any conclusion on the origin of this particular metabolic finding.

Conclusions

So far, metabonomic studies have concentrated on metabolic “snapshots” of individual samples, which were ostensibly influenced by uncontrolled variables. The possibility, demonstrated here, of eliminating the daily noise by multiple sample collection hints at a new perspective in metabonomic studies, based on the definition and identification of an individual metabolic fingerprint constituted by the invariant part of multiple samples of a single subject.

Although the genotype surely is a unique characteristic of each and every living organism, the existence of a truly individual metabolic phenotype still needs to be assessed. We report experimental evidence of the possibility of recognizing with 100% probability a subject within a group of individuals (none of which being subjected to particular imposed treatment or diet) by the metabolic fingerprint of one of its biofluids. This fingerprint is linked to the genotype but probably also to the general lifestyle and persisting environmental factors. It seems, however, to be independent from random daily variation. NMR confirms itself as a valuable tool for obtaining such fingerprints.

It is interesting to consider that a subject could be already characterized by only using the data from 12 metabolites, and identified with a confidence not much lower than in the case of the analysis performed on the complete fingerprint, provided that MRC is used. If one considers that in biofluids there are hundreds (cerebrospinal fluid) to thousands (urine) of metabolites, there is a strong possibility of being able to distinguish individual fingerprints even in a much larger panel study. From the investigation of the 12 metabolites, we conclude that a significant part of the individual metabolic fingerprint seems to be “coded” in the concentration levels and ratios of the metabolites rather than “coded” just by their pure presence or absence. It is the strength of NMR to provide access to multiparametric metabolic fingerprints in a fast, untargeted, and highly reproducible manner providing precise quantitative information even on the smallest concentration changes of multiple metabolites at the same time, despite a large dynamic range covering several orders of magnitude.

Characterizing individual metabolic fingerprints may allow researchers to (i) better plan personalized therapy and nutrition; (ii) perform studies of pharmacometabonomics to better predict and assess drug efficacy and toxicity; (iii) follow phenotype changes as a function of disease progression, possibly leading to earlier diagnosis and prognosis; (iv) perform cost-effective screenings on large human populations; and (v) address how possible long-term changes may be related to aging, because the metabolic fingerprint found is consistent over at least the collection period (≈2–3 months).

Statistical Methods

Supplementary Material