Abstract

The aim of this study was to investigate whether women with polycystic ovary syndrome (PCOS) had a unique metabolomic profile that was different from controls and to assess the feasibility of a definitive study. Twelve women with PCOS and 10 healthy women  as controls had measurements of demographic and anthropometric data, venepunctures and assays on plasma samples for metabolomic profiles using hydrogen-1, nuclear magnetic resonance (1H NMR) spectroscopy. There did not appear to be any clear differences between the metabolomic profiles of women with PCOS compared with controls when the NMR spectra were visually inspected and initial principal component analysis showed only a subtle differentiation between the two groups which was spread over three principal components. However, ‘supervised’ data analysis in the form of partial least-squares discriminant analysis (PLS-DA) and non-parametric univariate analysis allowed a stable PLS-DA model to be built, which appeared to differentiate between the two groups in a robust manner. Peak assignments for those spectral regions which appeared to differentiate between control and PCOS were consistent with amino acids (arginine, lysine, proline, glutamate and histidine), organic acids (citrate) and potentially lipids (CH2–CH2–C=C) with significant decreases noted in the levels of citrulline, lipid (CH2–CH2–C=C), arginine, lysine, ornithine, proline, glutamate, acetone, citrate and histidine in PCOS compared with controls. Women with PCOS may have a unique metabolomic finger print and a definitive study is feasible. These findings may enable sample size calculations for confirmatory studies and stimulate further research using metabolomics to improve the understanding and management of PCOS.

Introduction

Unraveling the fundamental basis of the polycystic ovary syndrome (PCOS) remains a mystery despite the fact that it affects 5–10% of premenopausal women (Azziz et al., 2004). Although progress has been made in the treatment of PCOS there remain a few challenges. Its key features remain chronic anovulatory infertility, hyperandrogenism and unique ovarian ultrasound features as agreed at a consensus conference in 2003 (The Rotterdam Group, 2004). Other features include menstrual disorders, hirsutism, acne or male-pattern hair loss (Setji and Brown, 2007), an adverse metabolic profile (Apridonidze et al., 2005) including insulin resistance and long-term health risks of type II diabetes, endometrial cancer and increased risk factors for cardiovascular disease (Krentz et al., 2007; Shaw et al., 2008; Glueck et al., 2009).

The exact cause(s) of PCOS, however remains uncertain. Several hypotheses have been previously advanced including aberrations in the hypothalamo–pituitary ovarian axis, insulin resistance, genetics, fetal programming, obesity and a combination of environmental and genetic factors. However, no single unifying hypothesis has been identified. Predicting which  women with obesity, infertility or hirsutism will respond to their respective treatments also remains a challenge as not all women with PCOS will respond to life-style advice to lose weight, ovulation induction drugs or anti-androgens. Predicting which  young women with PCOS will go on to develop diabetes or endometrial cancer also remains a challenge.

Developments in the field of ‘omics’ technology such as genomics, proteomics and metabolomics provide an opportunity to address some of these challenges. ‘Omic’ approaches allow the detection several molecules in one single experiment in contrast to traditional approaches/assays which detect one gene, protein or metabolite in each experiment. Data from these experiments are then integrated and analyzed using a ‘systems’ approach in collaboration with experts in bio-informatics to identify novel pathways and biomarkers to improve the management of women with PCOS.

The use of genomic and proteomic technologies for the study of the PCOS has  previously been reported (Urbanek et al., 1999; Hughes et al., 2006; Luque-Ramirez et al., 2006; Choi et al., 2007; Ma et al., 2007; Matharoo-Ball et al., 2007; Corton et al., 2008; Goodarzi, 2008; Atiomo et al., 2009; Insenser et al., 2010; Chen et al., 2011). These reports identified several potential biomarkers for PCOS and studies to validate and integrate them are ongoing. However, there have had  been no previously published studies on metabolomic biomarkers in PCOS before this study commenced. Metabolomics offers advantages over other ‘omic’ investigations such as greater sensitivity, a knowledge base that comes from years of biochemical research, the relatively small size of the number of endogenous molecules relative to the number of genes, mRNA species or proteins and the fact that it is the fastest system to react to stimuli or to change, enabling the most current view possible of the organism (Reo, 2002).

This aim of this pilot study was to investigate whether women with PCOS had a unique metabolomic profile that was different from weight and age matched control women to enable the sample size calculation and assessment of the feasibility of a definitive study.

Materials and Methods

Sample collection

Institutional review board approval was obtained for this study from the Nottingham local research ethics committee. This was a cross-sectional study in which experiments were performed on stored samples prospectively collected from women who participated in a previously published study of proteomic biomarkers in the serum of women with PCOS (Matharoo-Ball et al., 2007) conducted at the Queen Medical Centre in Nottingham, UK. Patient characterization for recruitment, sample collection have been previously published (Matharoo-Ball et al., 2007) but briefly, blood samples from 12 women with PCOS and 10 from healthy women acting as controls were collected. Initial metabolomic analysis was based on comparison of eight PCOS and eight healthy controls with the remaining four PCOS and two controls used as external validation samples.

All the subjects had been diagnosed with PCOS, using the Rotterdam criteria (The Rotterdam Group, 2004) if they presented with two or more of olio-/an ovulation, polycystic ovaries and clinical and/or biochemical sign of hyperandrogenism. Control samples were recruited as part of the same study. They were classed as healthy based on the control subjects having regular 21- to 35-day menstrual cycles and no more than one of ultrasound evidence of polycystic ovaries or evidence of hyperandrogenism. Women with any of the following were excluded from the study. Thyroid disease, current pregnancy, delivery or miscarriage occurring within the preceding 3 months, a recent surgery (within 3 months), history of myocardial infarction, use of aspirin or heparin, sex steroid therapy, a history of haematological disease, malignancy or liver disease, hyperprolactinaemia, a positive synacthen test following raised 17 hydroxyprogesterone levels, a history of thrombosis and recent viral illness.

Nine PCOS women were Caucasian, one Afro-Caribbean, one South Asian and one recorded as ‘other’. Seven of the controls were Caucasian and three were Afro-Caribbean.

Participants were asked not to eat from 12 midnight the day before the blood samples were collected. All blood samples were collected between 8.19 and 10.50 a.m. Fasting blood samples were collected in lithium heparin tubes centrifuged at 4000 rpm for 10 min at 4°C and plasma aliquotted into smaller fractions. Samples were then snap-frozen in liquid nitrogen and stored without delay at −80°C until analysis. Samples were removed from the freezer and allowed to thaw in a refrigerator overnight prior to analysis.

Nuclear magnetic resonance (1H NMR) spectroscopy

Plasma (0.5 ml) was deproteinized by mixing with acetonitrile (1 ml) and vortexed (30 s). The mixture was allowed to stand for 5 min and then centrifuged for 3 min. The entire supernatant was taken, evaporated to dryness under a nitrogen stream and reconstituted in phosphate-buffered saline (0.1 M, pH 7.4) prepared with deuterium oxide (D2O). Samples were placed into 5 mm o.d. (outer diameter) nuclear magnetic resonance (NMR) tubes in preparation for analysis by NMR spectroscopy. All spectra were acquired on a Bruker Avance 400 spectrometer operating at 400.13 MHz 1H observation frequency and equipped with a 5-mm quadruple nuclei probe equipped with z-axis gradients. All spectra were measured with 298-k internal probe temperature. In order to suppress the large water signal, spectra were acquired using a presaturation solvent suppression pulse sequence (NOESYPRESAT) with a relaxation delay of 1.5 s, during which the water resonance was selectively irradiated. Typically, 256 transients were collected into 64-k data points with a spectral width of 6000 Hz. Prior to Fourier transform, the exponential line broadening of 0.5 Hz was applied to free induction decay, which were also zero-filled by a factor of 2. Spectra were manually phase and baseline corrected, and chemical shifts were referenced to the primary reference standard, sodium trimethylsilyl [2,2,3,3-2H] propionate at δ 0.00 in TopSpin version 3.0 (Bruker GmbH, Germany).

1H NMR data preprocessing

Data in the spectral region δ9.0–0.5 were reduced to ASCII format using AMIX (Analysis of MIXtures, version 3.9.11, Bruker GmbH). Each NMR spectrum was reduced to 352 discrete regions of equal width (0.02 ppm) and the integral of each region was determined. The resulting table of intensity information was exported to Microsoft Excel 2010. The region δ4.5–5.5 was removed from all data, because it contains the large water peak. Each remaining region was then normalized to the total spectral area (block normalization) to minimize the effects of concentration differences between the samples. Peaks were assigned to specific compounds by comparison with previously published papers (Ala-Korpela, 1995; Nicholson  et al., 1995) in combination with an in-house spectral database.

Multivariate data analysis

Multivariate analysis was carried out using SIMCA-P version 12 (Umetrics, Umeå, Sweden). Two types of scaling were used, mean-centring (subtracting the calculated average of a variable from the data so that the mean for each variable is 0) and mean-centring followed by autoscaling (division of each variable by the standard deviation for that variable). The use of mean-centred, but not autoscaled, data in basic multivariate analyses such as principal component analysis (PCA) often results in an emphasis on perturbation of the metabolites that are present in high concentrations, whereas autoscaled data convert each variable to a standard variable with equal weight and are more sensitive to the changes in the levels of minor metabolites. For the purpose of illustrating the results of this study, autoscaling of the data is reported.

PCA was used for initial visualization of the 1H NMR profiles where PCA reduces the dimensionality of a data set, which has a large number of variables, while maintaining as much variation within the data as possible. Next, a supervised method, partial least-squares discriminant analysis (PLS-DA), was performed to maximize the separation between samples from different treatments. The PLS-DA model was validated using random subset cross-validation, where eight control and eight PCOS samples were used in the analysis, divided up into six control samples and six PCOS samples being used to create a training set, while two control samples and two PCOS samples were used as a test set. Ten iterations of PLS-DA models were generated using different random combinations of test and training samples (in the split indicated above) to evaluate the stability of the model. The remaining four PCOS samples and two control samples were not included in the initial PLS-DA models as these were used as external validation samples after variable selection. After initial PLS-DA, variables with a variable importance in projection (VIP) score of >1.2, a % coefficient of variation (CV) of <20 and which showed consistent changes across age and BMI matched pairs of samples, were selected and PLS-DA was repeated on a smaller subset of variables. Internal cross-validation was carried out as described previously and, in addition, the six samples which were not used in the variable selection process were used to externally validate the model.

Partial least-squares regression (PLS) was applied to explore correlations between the 1H NMR spectra of blood plasma and both age and BMI. Random subset cross-validation was used to evaluate the model stability where three-quarters of the samples were used to create a training set, while one-quarter of the samples was used to create a test set and the procedure was repeated ten times.

Univariate data analysis

Univariate analysis was carried out using Matlab 2009a (The MathWorks, Natick, MA, USA). The Wilcoxon rank sum test (a.k.a. Mann–Whitney U-test) was performed on the subset of spectral ‘buckets’ that was selected during PLS-DA, and the P-values were subsequently corrected for multiple comparison using the false discovery rate correction (Benjamini and Hochberg, 1995). A P-value of <0.05 was accepted as statistically significant.

Results

Demographic data

Demographic data from women with and without PCOS are summarized in Table I. There were no significant differences in age, BMI, waist–hip ratio and the other baseline clinical and endocrine profiles in women in the PCOS group compared with the control groups except for a higher free androgen index (FAI) and sex hormone-binding globulin (SHBG).

Table I

Demographic data in healthy women and PCOS women.

AgeEthnicityDays since LMPBMIWCHCWHRFGTSHBGFAIInsulinGlucose
PCOS 
 P1 37 49 45 144 143 1.01 1.7 21 8.10 10.3 4.8 
 P2 32 35 35 100 108 0.93 13 2.3 26 8.85  5.9 
 P3 27 83 21 65 88 0.74 2.5 25 10.00 3.3 4.8 
 P4 21 36 103 119 0.87 2.5 17 14.71 15.1 4.8 
 P5 35 15 36 103 116 0.89 19 2.0 27 7.41  6.0 
 P6 22 14 37 107 112 0.96 14 2.2 25 8.80 14.4 5.2 
 P7 22 541 27 97 105 0.92 3.3 30 11.00 5.6 4.9 
 P8 26 263 37 108 118 0.92 10 1.9 34 5.59 20.2 4.6 
 P9 24 180 29 97 110 0.88 17 3.3 33 10.00 7.0 5.1 
 P10 21 28 38 86 110 0.78 16 1.6 22.86 8.1 5.2 
 P11 31 22 73 104 0.70 15 0.900 47 1.91 3.5  
 P12 28 28 36 117 122 0.96 14 3.4 24 14.17 11.2 6.3 
 Mean 27.17   33.17 100.00 112.92 0.88 11.75 2.30 26.33 10.28 9.87 5.24 
 SD 5.54   7.06 20.28 13.02 0.09 5.10 0.76 9.74 5.25 5.48 0.57 
Controls 
 C1 20 15 22 69 92 0.75 1.5 81 1.85 10.0 4.8 
 C2 31 20 67 91 0.74 1.8 93 1.94  4.6 
 C3 22 19 66 80.5 0.82 1.4 66 2.12 2.4 4.7 
 C4 39 31 47 122 134 0.91 3.0   13.3 4.6 
 C5 34 13 31 97 120 0.81 64 2.66 
 C6 40 26 91 110 0.83      
 C7 25 16 32 41.5 47 0.88 14 19 7.37 
 C8 19 11 25 62.5 91 0.69 14 2.0 49 4.08 3.6 5.2 
 C10 22 28 104 79 1.32 2.0 2.4 74 3.24 3.6 4.6 
 C12 32  20 58 81 0.72 10 2.8 64 4.38   
Mean 28.4   26.995 77.8 92.55 0.84545 6.3 63.75 3.45414 6.8 4.875 
SD 7.82   8.3625 24.631 24.3087 0.18037 4.9677 0.60 22.31 1.84758 4.05 0.31 
T-test (P0.67   0.0748 0.031 0.02079 0.58354 0.0201 0.342 <0.001 0.00249 0.23 0.12 
AgeEthnicityDays since LMPBMIWCHCWHRFGTSHBGFAIInsulinGlucose
PCOS 
 P1 37 49 45 144 143 1.01 1.7 21 8.10 10.3 4.8 
 P2 32 35 35 100 108 0.93 13 2.3 26 8.85  5.9 
 P3 27 83 21 65 88 0.74 2.5 25 10.00 3.3 4.8 
 P4 21 36 103 119 0.87 2.5 17 14.71 15.1 4.8 
 P5 35 15 36 103 116 0.89 19 2.0 27 7.41  6.0 
 P6 22 14 37 107 112 0.96 14 2.2 25 8.80 14.4 5.2 
 P7 22 541 27 97 105 0.92 3.3 30 11.00 5.6 4.9 
 P8 26 263 37 108 118 0.92 10 1.9 34 5.59 20.2 4.6 
 P9 24 180 29 97 110 0.88 17 3.3 33 10.00 7.0 5.1 
 P10 21 28 38 86 110 0.78 16 1.6 22.86 8.1 5.2 
 P11 31 22 73 104 0.70 15 0.900 47 1.91 3.5  
 P12 28 28 36 117 122 0.96 14 3.4 24 14.17 11.2 6.3 
 Mean 27.17   33.17 100.00 112.92 0.88 11.75 2.30 26.33 10.28 9.87 5.24 
 SD 5.54   7.06 20.28 13.02 0.09 5.10 0.76 9.74 5.25 5.48 0.57 
Controls 
 C1 20 15 22 69 92 0.75 1.5 81 1.85 10.0 4.8 
 C2 31 20 67 91 0.74 1.8 93 1.94  4.6 
 C3 22 19 66 80.5 0.82 1.4 66 2.12 2.4 4.7 
 C4 39 31 47 122 134 0.91 3.0   13.3 4.6 
 C5 34 13 31 97 120 0.81 64 2.66 
 C6 40 26 91 110 0.83      
 C7 25 16 32 41.5 47 0.88 14 19 7.37 
 C8 19 11 25 62.5 91 0.69 14 2.0 49 4.08 3.6 5.2 
 C10 22 28 104 79 1.32 2.0 2.4 74 3.24 3.6 4.6 
 C12 32  20 58 81 0.72 10 2.8 64 4.38   
Mean 28.4   26.995 77.8 92.55 0.84545 6.3 63.75 3.45414 6.8 4.875 
SD 7.82   8.3625 24.631 24.3087 0.18037 4.9677 0.60 22.31 1.84758 4.05 0.31 
T-test (P0.67   0.0748 0.031 0.02079 0.58354 0.0201 0.342 <0.001 0.00249 0.23 0.12 

SD, standard deviation; Ethnicity (Caucasian = 1, afro Caribbean = 2, South Asian =3; Other =4); LMP, last menstrual period; BMI, body mass index; WC, waist circumference in centimetres; HC, hip circumference in centimetres; WHR, waist–hip ratio; FG, Ferriman–Galwey score; T, testosterone; SHBG, sex hormone-binding globulin; FAI, free androgen index.

Table I

Demographic data in healthy women and PCOS women.

AgeEthnicityDays since LMPBMIWCHCWHRFGTSHBGFAIInsulinGlucose
PCOS 
 P1 37 49 45 144 143 1.01 1.7 21 8.10 10.3 4.8 
 P2 32 35 35 100 108 0.93 13 2.3 26 8.85  5.9 
 P3 27 83 21 65 88 0.74 2.5 25 10.00 3.3 4.8 
 P4 21 36 103 119 0.87 2.5 17 14.71 15.1 4.8 
 P5 35 15 36 103 116 0.89 19 2.0 27 7.41  6.0 
 P6 22 14 37 107 112 0.96 14 2.2 25 8.80 14.4 5.2 
 P7 22 541 27 97 105 0.92 3.3 30 11.00 5.6 4.9 
 P8 26 263 37 108 118 0.92 10 1.9 34 5.59 20.2 4.6 
 P9 24 180 29 97 110 0.88 17 3.3 33 10.00 7.0 5.1 
 P10 21 28 38 86 110 0.78 16 1.6 22.86 8.1 5.2 
 P11 31 22 73 104 0.70 15 0.900 47 1.91 3.5  
 P12 28 28 36 117 122 0.96 14 3.4 24 14.17 11.2 6.3 
 Mean 27.17   33.17 100.00 112.92 0.88 11.75 2.30 26.33 10.28 9.87 5.24 
 SD 5.54   7.06 20.28 13.02 0.09 5.10 0.76 9.74 5.25 5.48 0.57 
Controls 
 C1 20 15 22 69 92 0.75 1.5 81 1.85 10.0 4.8 
 C2 31 20 67 91 0.74 1.8 93 1.94  4.6 
 C3 22 19 66 80.5 0.82 1.4 66 2.12 2.4 4.7 
 C4 39 31 47 122 134 0.91 3.0   13.3 4.6 
 C5 34 13 31 97 120 0.81 64 2.66 
 C6 40 26 91 110 0.83      
 C7 25 16 32 41.5 47 0.88 14 19 7.37 
 C8 19 11 25 62.5 91 0.69 14 2.0 49 4.08 3.6 5.2 
 C10 22 28 104 79 1.32 2.0 2.4 74 3.24 3.6 4.6 
 C12 32  20 58 81 0.72 10 2.8 64 4.38   
Mean 28.4   26.995 77.8 92.55 0.84545 6.3 63.75 3.45414 6.8 4.875 
SD 7.82   8.3625 24.631 24.3087 0.18037 4.9677 0.60 22.31 1.84758 4.05 0.31 
T-test (P0.67   0.0748 0.031 0.02079 0.58354 0.0201 0.342 <0.001 0.00249 0.23 0.12 
AgeEthnicityDays since LMPBMIWCHCWHRFGTSHBGFAIInsulinGlucose
PCOS 
 P1 37 49 45 144 143 1.01 1.7 21 8.10 10.3 4.8 
 P2 32 35 35 100 108 0.93 13 2.3 26 8.85  5.9 
 P3 27 83 21 65 88 0.74 2.5 25 10.00 3.3 4.8 
 P4 21 36 103 119 0.87 2.5 17 14.71 15.1 4.8 
 P5 35 15 36 103 116 0.89 19 2.0 27 7.41  6.0 
 P6 22 14 37 107 112 0.96 14 2.2 25 8.80 14.4 5.2 
 P7 22 541 27 97 105 0.92 3.3 30 11.00 5.6 4.9 
 P8 26 263 37 108 118 0.92 10 1.9 34 5.59 20.2 4.6 
 P9 24 180 29 97 110 0.88 17 3.3 33 10.00 7.0 5.1 
 P10 21 28 38 86 110 0.78 16 1.6 22.86 8.1 5.2 
 P11 31 22 73 104 0.70 15 0.900 47 1.91 3.5  
 P12 28 28 36 117 122 0.96 14 3.4 24 14.17 11.2 6.3 
 Mean 27.17   33.17 100.00 112.92 0.88 11.75 2.30 26.33 10.28 9.87 5.24 
 SD 5.54   7.06 20.28 13.02 0.09 5.10 0.76 9.74 5.25 5.48 0.57 
Controls 
 C1 20 15 22 69 92 0.75 1.5 81 1.85 10.0 4.8 
 C2 31 20 67 91 0.74 1.8 93 1.94  4.6 
 C3 22 19 66 80.5 0.82 1.4 66 2.12 2.4 4.7 
 C4 39 31 47 122 134 0.91 3.0   13.3 4.6 
 C5 34 13 31 97 120 0.81 64 2.66 
 C6 40 26 91 110 0.83      
 C7 25 16 32 41.5 47 0.88 14 19 7.37 
 C8 19 11 25 62.5 91 0.69 14 2.0 49 4.08 3.6 5.2 
 C10 22 28 104 79 1.32 2.0 2.4 74 3.24 3.6 4.6 
 C12 32  20 58 81 0.72 10 2.8 64 4.38   
Mean 28.4   26.995 77.8 92.55 0.84545 6.3 63.75 3.45414 6.8 4.875 
SD 7.82   8.3625 24.631 24.3087 0.18037 4.9677 0.60 22.31 1.84758 4.05 0.31 
T-test (P0.67   0.0748 0.031 0.02079 0.58354 0.0201 0.342 <0.001 0.00249 0.23 0.12 

SD, standard deviation; Ethnicity (Caucasian = 1, afro Caribbean = 2, South Asian =3; Other =4); LMP, last menstrual period; BMI, body mass index; WC, waist circumference in centimetres; HC, hip circumference in centimetres; WHR, waist–hip ratio; FG, Ferriman–Galwey score; T, testosterone; SHBG, sex hormone-binding globulin; FAI, free androgen index.

1H NMR spectra of PCOS and healthy women

An NMR spectrum representative of one woman with PCOS's plasma is shown in Fig. 1. The spectrum comprises signals of sugars (present in regions 3.2–5.4 ppm) and also organic acids, such as tricarboxylic acid cycle intermediates and amino acids. Aromatic compounds give resonances in the region ∼6.0–9.0 ppm, and in the case of the samples collected in this study, the peaks visible in this region are aromatic amino acids, which, as would be expected, are present in relatively low concentrations in comparison with sugars. On initial visual inspection, the spectra of healthy and women with PCOS were very similar, but the spectra were too complex to compare by eye and therefore multivariate data analysis was used to ease visualization of the data.

Figure 1

Typical 1D NOESYPRESAT 1H NMR spectrum of plasma from a PCOS patient, with peak assignments

Multivariate data analysis

Figure 2 shows PCA scores plot of the first three principal components, where PC1 accounts for 36.7% of the total spectral variance, PC2 accounts for 24.7% of the total spectral variance and PC3 accounts for 10.3% of the total spectral variance. Discrete separation between seven controls and the PCOS samples, with one control being indistinct from PCOS samples, was observed. Comparison was made with clinical data to investigate why sample C5 plotted with the PCOS samples; however, no distinct reason could be found and it must hence be assumed that this represents variation in normal, healthy individuals. However, because separation was only achieved using a combination of three PCs, PLS-DA was used to optimize separation between the classes and to obtain an easier to interpret model. Initial PLS-DA validated using random subset cross-validation, resulted in a model based on two latent variables (LV) and where R2X = 0.348, R2Y = 0.60, Q2 = 0.234. Random subset cross-validation showed that 85% of control samples were correctly classified, whereas 60% of PCOS samples were correctly classified (Table II). This indicated that while there appeared to be some correlation between NMR spectral profiles and disease status, this correlation was somewhat masked by the presence of other (biological) variation in the data. The range of ages and BMIs of individuals in the study was large at 19–40 and 19–47, respectively. PLS regression was used to explore the possibility that age- and BMI-related variation was present in the data and this hypothesis was confirmed. Therefore, variable selection was used to remove variables from the model that were affecting the model stability. Variables with a VIP coefficient of >1.2 and whose VIP coefficient had a %CV of <20% based on the random subset cross-validation routine applied previously were initially selected as being the variables which allowed the best differentiation between the two sample groups, while being the most stable. Then, the Wilcoxon rank-sum test was applied to samples from a subset of age and BMI matched pairs of five women with PCOS and five controls to further identify the changes in metabolites between groups.

Table II

PLS-DA model using random subset cross-validation, where six control samples and six PCOS samples were used to create a training set while two control samples and two PCOS samples were used as a test set.

 % Correctly classified training set
% Correctly classified test set
ControlPCOSControlPCOS
Model 1 (full data set) 95 83 85 60 
Model 2 (reduced data set) 98 88 90 90 
 % Correctly classified training set
% Correctly classified test set
ControlPCOSControlPCOS
Model 1 (full data set) 95 83 85 60 
Model 2 (reduced data set) 98 88 90 90 
Table II

PLS-DA model using random subset cross-validation, where six control samples and six PCOS samples were used to create a training set while two control samples and two PCOS samples were used as a test set.

 % Correctly classified training set
% Correctly classified test set
ControlPCOSControlPCOS
Model 1 (full data set) 95 83 85 60 
Model 2 (reduced data set) 98 88 90 90 
 % Correctly classified training set
% Correctly classified test set
ControlPCOSControlPCOS
Model 1 (full data set) 95 83 85 60 
Model 2 (reduced data set) 98 88 90 90 
Figure 2

Metabolic differences in blood plasma is highlighted in the PCA score plot. Blood plasma collected from women with PCOS is denoted P and from controls is denoted C. The scores plot shows the first three principal components, where PC1 accounts for 34.7% of the total spectral variance, PC2 accounts for 24.7% of the total spectral variance and PC 3 accounts for 10.3% of the total spectral variance.

PLS-DA was then repeated using only the subset of metabolites detailed in Table III. The resultant PLS-DA model was validated with random subset cross-validation and also externally validated using an additional four PCOS samples and two control samples, which had not been used in any of the analyses thus far. The model was based on one LV and R2X = 0.766, R2Y = 0.634, Q2 = 0.562. Cross-validation showed that 90% of control samples were correctly classified and 90% of PCOS samples were also correctly classified (Table II). Validation from external samples showed that 100% of controls were correctly classified and 75% of PCOS samples were correctly classified. Peak assignments for those spectral regions which appeared to differentiate between control and PCOS were consistent with amino acids (arginine, lysine, proline, glutamate and histidine), organic acids (citrate) and potentially lipids (CH2–CH2–C=C). Interestingly, glucose levels were observed to be ∼1.2 times greater in PCOS subjects than controls, but with high levels of variability and with P = 0.095.

Table III

Data analysis of 1H NMR spectra of deproteinised blood plasma: PCOS versus controls.

Spectral ‘Buckets’NMR chemical shift (multiplicity)Possible peak assignmentVIPMean fold change (PCOS/control)P-valueFDR corrected
1.39 1.40 (m) Unassigned 1.77 ± 0.19 0.91 0.008 0.035 
1.41 2.05 ± 0.21 0.016 0.035 
1.53 1.57 (m) Citrullene 1.82 ± 0.20 0.91 0.032 0.037 
1.55 Lipid (CH2–CH2–C=C) 1.89 ± 0.20 0.016 0.035 
1.57     1.94± 0.19   0.056 0.058 
1.59 1.83 ± 0.26 0.032 0.037 
1.61 1.65 (m) Arginine 1.89 ± 0.35 0.88 0.008 0.035 
1.63 Lysine 1.91 ± 0.24 0.008 0.035 
1.65 1.71± 0.21 0.095 0.095 
1.67 1.84 ± 0.17 0.016 0.035 
1.83 1.83 (m) Citrulline 1.89 ± 0.21 0.86 0.032 0.037 
Ornithine 
2.05 2.04 (m) Proline 1.80 ± 0.25 0.86 0.032 0.037 
2.09 2.10 (m) Proline 1.75 ± 0.21 0.89 0.032 0.037 
2.11 1.86 ± 0.20 0.032 0.037 
2.17 2.17 (m) Glutamate 1.63 ± 0.25 0.93 0.016 0.035 
2.21 2.23 (s) Acetone 1.91 ± 0.24 0.86 0.016 0.035 
2.27 2.29 (m) Proline 1.71 ± 0.19 0.86 0.032 0.037 
2.29 Glutamate 1.85 ± 0.24 0.016 0.035 
2.31 1.99 ± 0.21 0.016 0.035 
2.37 Glutamate 1.79 ± 0.25 0.89 0.032 0.037 
2.69 2.68 (dd) Citrate 1.37 ± 0.20 0.87 0.032 0.037 
7.81 7.83 (s) Histidine 1.29 ± 0.24 0.38 0.048 0.052 
Spectral ‘Buckets’NMR chemical shift (multiplicity)Possible peak assignmentVIPMean fold change (PCOS/control)P-valueFDR corrected
1.39 1.40 (m) Unassigned 1.77 ± 0.19 0.91 0.008 0.035 
1.41 2.05 ± 0.21 0.016 0.035 
1.53 1.57 (m) Citrullene 1.82 ± 0.20 0.91 0.032 0.037 
1.55 Lipid (CH2–CH2–C=C) 1.89 ± 0.20 0.016 0.035 
1.57     1.94± 0.19   0.056 0.058 
1.59 1.83 ± 0.26 0.032 0.037 
1.61 1.65 (m) Arginine 1.89 ± 0.35 0.88 0.008 0.035 
1.63 Lysine 1.91 ± 0.24 0.008 0.035 
1.65 1.71± 0.21 0.095 0.095 
1.67 1.84 ± 0.17 0.016 0.035 
1.83 1.83 (m) Citrulline 1.89 ± 0.21 0.86 0.032 0.037 
Ornithine 
2.05 2.04 (m) Proline 1.80 ± 0.25 0.86 0.032 0.037 
2.09 2.10 (m) Proline 1.75 ± 0.21 0.89 0.032 0.037 
2.11 1.86 ± 0.20 0.032 0.037 
2.17 2.17 (m) Glutamate 1.63 ± 0.25 0.93 0.016 0.035 
2.21 2.23 (s) Acetone 1.91 ± 0.24 0.86 0.016 0.035 
2.27 2.29 (m) Proline 1.71 ± 0.19 0.86 0.032 0.037 
2.29 Glutamate 1.85 ± 0.24 0.016 0.035 
2.31 1.99 ± 0.21 0.016 0.035 
2.37 Glutamate 1.79 ± 0.25 0.89 0.032 0.037 
2.69 2.68 (dd) Citrate 1.37 ± 0.20 0.87 0.032 0.037 
7.81 7.83 (s) Histidine 1.29 ± 0.24 0.38 0.048 0.052 

NMR spectral regions in which changes correlated with health/disease status were observed in the study. Tentative identifications of peaks are given based on comparison against a metabolite spectral database. Italicized variables indicate those where P > 0.05, but represent part of a peak where other regions of the same peak show P < 0.05.

FDR, false discovery rate.

Table III

Data analysis of 1H NMR spectra of deproteinised blood plasma: PCOS versus controls.

Spectral ‘Buckets’NMR chemical shift (multiplicity)Possible peak assignmentVIPMean fold change (PCOS/control)P-valueFDR corrected
1.39 1.40 (m) Unassigned 1.77 ± 0.19 0.91 0.008 0.035 
1.41 2.05 ± 0.21 0.016 0.035 
1.53 1.57 (m) Citrullene 1.82 ± 0.20 0.91 0.032 0.037 
1.55 Lipid (CH2–CH2–C=C) 1.89 ± 0.20 0.016 0.035 
1.57     1.94± 0.19   0.056 0.058 
1.59 1.83 ± 0.26 0.032 0.037 
1.61 1.65 (m) Arginine 1.89 ± 0.35 0.88 0.008 0.035 
1.63 Lysine 1.91 ± 0.24 0.008 0.035 
1.65 1.71± 0.21 0.095 0.095 
1.67 1.84 ± 0.17 0.016 0.035 
1.83 1.83 (m) Citrulline 1.89 ± 0.21 0.86 0.032 0.037 
Ornithine 
2.05 2.04 (m) Proline 1.80 ± 0.25 0.86 0.032 0.037 
2.09 2.10 (m) Proline 1.75 ± 0.21 0.89 0.032 0.037 
2.11 1.86 ± 0.20 0.032 0.037 
2.17 2.17 (m) Glutamate 1.63 ± 0.25 0.93 0.016 0.035 
2.21 2.23 (s) Acetone 1.91 ± 0.24 0.86 0.016 0.035 
2.27 2.29 (m) Proline 1.71 ± 0.19 0.86 0.032 0.037 
2.29 Glutamate 1.85 ± 0.24 0.016 0.035 
2.31 1.99 ± 0.21 0.016 0.035 
2.37 Glutamate 1.79 ± 0.25 0.89 0.032 0.037 
2.69 2.68 (dd) Citrate 1.37 ± 0.20 0.87 0.032 0.037 
7.81 7.83 (s) Histidine 1.29 ± 0.24 0.38 0.048 0.052 
Spectral ‘Buckets’NMR chemical shift (multiplicity)Possible peak assignmentVIPMean fold change (PCOS/control)P-valueFDR corrected
1.39 1.40 (m) Unassigned 1.77 ± 0.19 0.91 0.008 0.035 
1.41 2.05 ± 0.21 0.016 0.035 
1.53 1.57 (m) Citrullene 1.82 ± 0.20 0.91 0.032 0.037 
1.55 Lipid (CH2–CH2–C=C) 1.89 ± 0.20 0.016 0.035 
1.57     1.94± 0.19   0.056 0.058 
1.59 1.83 ± 0.26 0.032 0.037 
1.61 1.65 (m) Arginine 1.89 ± 0.35 0.88 0.008 0.035 
1.63 Lysine 1.91 ± 0.24 0.008 0.035 
1.65 1.71± 0.21 0.095 0.095 
1.67 1.84 ± 0.17 0.016 0.035 
1.83 1.83 (m) Citrulline 1.89 ± 0.21 0.86 0.032 0.037 
Ornithine 
2.05 2.04 (m) Proline 1.80 ± 0.25 0.86 0.032 0.037 
2.09 2.10 (m) Proline 1.75 ± 0.21 0.89 0.032 0.037 
2.11 1.86 ± 0.20 0.032 0.037 
2.17 2.17 (m) Glutamate 1.63 ± 0.25 0.93 0.016 0.035 
2.21 2.23 (s) Acetone 1.91 ± 0.24 0.86 0.016 0.035 
2.27 2.29 (m) Proline 1.71 ± 0.19 0.86 0.032 0.037 
2.29 Glutamate 1.85 ± 0.24 0.016 0.035 
2.31 1.99 ± 0.21 0.016 0.035 
2.37 Glutamate 1.79 ± 0.25 0.89 0.032 0.037 
2.69 2.68 (dd) Citrate 1.37 ± 0.20 0.87 0.032 0.037 
7.81 7.83 (s) Histidine 1.29 ± 0.24 0.38 0.048 0.052 

NMR spectral regions in which changes correlated with health/disease status were observed in the study. Tentative identifications of peaks are given based on comparison against a metabolite spectral database. Italicized variables indicate those where P > 0.05, but represent part of a peak where other regions of the same peak show P < 0.05.

FDR, false discovery rate.

Discussion

The results showed that there did not appear to be any clear differences between the metabolomic profiles of women with PCOS compared with controls when the NMR spectra were visually inspected and initial PCA showed only a subtle differentiation between the two groups which was spread over three principal components. However, ‘supervised’ data analysis in the form of PLS-DA and non-parametric univariate analysis allowed a stable PLS-DA model to be built, which appeared to differentiate between the two groups in a robust manner. Significant decreases were noted in the levels of citrulline, lipid (CH2–CH2–C=C), arginine, lysine, ornithine, proline, glutamate, acetone, citrate and histidine in PCOS compared with controls.

Some of the metabolites found appeared to match (glutamate, arginine and citrate) with a recently published study on metabonomics and PCOS (Liye et al., 2012), while some were different. In the cross-sectional study of women with PCOS and controls (Liye et al., 2012), significant decreases in the levels of amino acids (leucine, isoleucine, methionine, glutamine and arginine), citrate, choline and glycerophosphocholine/phosphocholine and increases in the levels of lactate, dimethylamine, creatine and N-acetyl glycoproteins were observed in PCOS patients compared with the controls. The differences in metabolite changes in both studies may have been the result of differential protein binding, which can complicate the analysis of blood plasma, where protein binding of metabolites in whole plasma may be misinterpreted as concentration changes rather than changes in the degree of their interactions with macromolecules (Daykin et al., 2012).

These data are preliminary and will need to be validated in a larger study to ensure reproducibility; however, it was a very useful proof of concept study which has demonstrated that the extension of the application of omic studies in PCOS to metabolomics is feasible. Some of the metabolites identified as different in women with PCOS have roles that may be associated with the spectrum of clinical problems associated with PCOS and may therefore not be incidental findings. PCOS is associated with insulin resistance/prediabetes (Apridonidze et al., 2005) and obesity. Although not statistically significant, women with PCOS in our study had a higher body mass index, fasting insulin and glucose. Perturbations in amino acid levels have been previously implicated in prediabetic states and diabetes (Fiehn et al., 2010). Fiehn et al., archived plasma samples derived from BMI and age-matched overweight to obese type 2 diabetic (n= 44) and non-diabetic (n= 12) African-American women. In addition to many unidentified small molecules, specific metabolites that were increased significantly in subjects with type-2 diabetes were leucine, 2-ketoisocaproate, valine, cystine, histidine, 2-hydroxybutanoate, long-chain fatty acids and carbohydrate derivatives. Similarly, raised acetate and acetoacetate levels are associated with states of insulin deprivation that may occur in women with PCOS who are insulin resistant (Foster, 1967).

Our study however showed that citrulline, lipid (CH2–CH2–C=C), arginine, lysine, ornithine, proline, glutamate, acetone, citrate and histidine were decreased in PCOS compared with controls. It is difficult at this stage to advance a single unifying hypothesis to explain this finding. However, lower amino acid levels have recently been linked with type 2 diabetes (Mihalik et al., 2012). One recently advanced hypothesis is that activity of gut microflora in PCOS is increased which may lead to obesity, diabetes, cardiovascular disease and fatty liver (Liye et al., 2012). Glutamate depression as found in our study has also been associated with negative effects on the functional integrity of the gut (O'Dwyer et al., 1989) and immunosuppresion (Kafkewitz and Bendich, 1983) which may lead to increased activity of gut microflora. This hypothesis will however need to be tested in future studies.

Our findings if subsequently validated may improve the understanding and management of women with PCOS by, for example, providing suitable biomarkers that can be used in clinical practice to provide early detection of which subgroups of women with PCOS are more to develop type 2 diabetes so that early preventive measures can be introduced. They may also help with risk stratification of long-term risk of cardiovascular disease and other long-term health conditions associated with PCOS, but the biomarkers will need to be validated in a prospective longitudinal cohort study.

The strengths of our study are in its originality and the rigorous way in which the metabolomic experiments were conducted. It has demonstrated for the first time that metabolomic approaches are feasible in women with PCOS in our setting and it has provided some pilot data with which it will be possible to conduct a sample size calculation for a definitive study. It has also provided some metabolite data on women with PCOS which will be useful in future systematic reviews. On the other hand, since the metabolome is dynamic, it is plausible to assume that several variables, such as the phase of the menstrual cycle in controls, could have acted as potential confounders. It would have ideal to precisely control for this although it would not have been practical in women with PCOS because of their irregular menstrual cycles. Future studies should ideally collect blood samples from a larger number of control women at different time points in their menstrual cycle. It would also have been ideal to get samples from a larger number of women. We explored using samples from our previously published proteomic study (Matharoo-Ball et al., 2007), but no further stored samples were available for analysis. However, we were able to obtain stored NMR spectral data on de-proteinized plasma from 12 women with PCOS and 10 controls as presented in this study. We are currently in the process of addressing the issue of the sample size by recruiting more women. We debated the merits of merging the data from newly recruited women with the current data if and when these data become available and it was felt that this might compromise the scientific validity of the study given the fact that the metabolome is dynamic and that several variables including analysis at different time points could act as potential confounders. There have however been precedents set where novel primary studies on women with PCOS using ‘omic’ techniques have been published with fewer sample sizes than ours with one study recruiting three PCOS and one control (Ma et al., 2007) and another five PCOS and four normal women (Wood et al., 2003).

It is known that ethnicity can affect the metabolic profile and so ideally it is best to use patients of one ethnicity. However, given the size of our cohort, this was not possible. We do not however think ethnicity affected the metabolic profile in this cohort as the PCA (Fig. 2) showed that there was only one outlier (control C5) who happened to be Caucasian. As seven of the controls were Caucasian and three were Afro-Caribbean, we do not think different ethnicities affected the PCA in this cohort.

In summary, this study evaluated metabolomic profiles in the plasma of women with PCOS compared with controls. Peak assignments for those spectral regions which appeared to differentiate between control and PCOS were consistent with amino acids (arginine, lysine, proline, glutamate, histidine), organic acids (citrate) and potentially lipids (CH2–CH2–C=C) with significant decreases noted in the levels of citrulline, lipid (CH2–CH2–C=C), arginine, lysine, ornithine, proline, glutamate, acetone, citrate and histidine in PCOS compared with controls. Although the data will obviously need to be validated in larger populations, this study has been a useful proof of concept study and will enable sample size calculation which will be of benefit to other scientists. The study has also provided some data to contribute to future systematic reviews of studies in this field. In addition, the ability to obtain and integrate metabolomic with proteomic and genomic data in women with PCOS will increase the potential of building a more complete picture of the syndrome in a systems approach. Large multidisciplinary collaborations of basic scientists, clinicians, computer scientists and mathematicians will be required to make sense of all the new data generated and translate the information into improved patient care.

Authors’ roles

W.A. conceived of study, supervised patient recruitment, data collection, data evaluation, drafting, editing and approving the final version of this paper for submission. C.D. supervised the metabolomic experimental work and contributed to drafting, editing and approving the final version of this paper for submission.

Funding

This work was supported by Departmental research funds.

Conflict of interest

None declared.

Acknowledgements

We would like to thank all the women who participated in the study, Dedi Hanwar who performed the metabolomic experimental work and wrote the first draft of the experimental work for the paper. We also thank Dr Robert Layfield for help with sample storage, Dr Florian Wulfert (Sheffield Hallam University) for carrying out the univariate statistics and Professor David Barrett for his comments on the manuscript.

References

Ala-Korpela
M
1H NMR spectroscopy of human blood plasma
Prog Nucl Magn Reson Spectrosc
1995
, vol. 
27
 (pg. 
475
-
554
)
Apridonidze
T
Essah
PA
Iuorno
MJ
Nestler
JE
Prevalence and characteristics of the metabolic syndrome in women with polycystic ovary syndrome
J Clin Endocrinol Metab
2005
, vol. 
90
 (pg. 
1929
-
1935
)
Atiomo
W
Khalid
S
Parameshweran
S
Houda
M
Layfield
R
Proteomic biomarkers for the diagnosis and risk stratification of polycystic ovary syndrome: a systematic review
BJOG
2009
, vol. 
116
 (pg. 
137
-
143
)
Azziz
R
Woods
KS
Reyna
R
Key
TJ
Knochenhauer
ES
Yildiz
BO
The prevalence and features of the polycystic ovary syndrome in an unselected population
J Clin Endocrinol Metab
2004
, vol. 
89
 (pg. 
2745
-
2749
)
Benjamini
Y
Hochberg
Y
Controlling the false discovery rate: a practical and powerful approach to multiple testing
J R Stat Soc B
1995
, vol. 
57
 (pg. 
289
-
300
)
Birdsall
MA
Farquhar
CM
White
HD
Association between polycystic ovaries and extent of coronary artery disease in women having cardiac catheterization
Ann Intern Med
1997
, vol. 
126
 (pg. 
32
-
35
)
Chen
ZJ
Zhao
H
He
L
Shi
Y
Qin
Y
Shi
Y
Li
Z
You
L
Zhao
J
Liu
J
, et al. 
Genome-wide association study identifies susceptibility loci for polycystic ovary syndrome on chromosome 2p16.3, 2p21 and 9q33.3
Nat Genet
2011
, vol. 
43
 (pg. 
55
-
59
)
Choi
BC
Kim
YS
Kim
MS
Chung
MK
Choi
CH
Baek
KH
Identification of over expressed proteins by proteomic analysis using human follicular fluids derived from polycystic ovary syndrome (PCOS) patients
Fertil Steril
2007
, vol. 
88
 pg. 
S180
 
Córton
M
Botella-Carretero
JI
López
JA
Camafeita
E
San Millán
JL
Escobar-Morreale
HF
Peral
B
Proteomic analysis of human omental adipose tissue in the polycystic ovary syndrome using two-dimensional difference gel electrophoresis and mass spectrometry
Hum Reprod
2008
, vol. 
23
 (pg. 
651
-
661
)
Daykin
CA
Wulfert
F
Bro
R
Data handling for interactive metabolomics: tools for studying the dynamics of metabolome-macromolecule interactions
Metabolomics
2012
, vol. 
8
 (pg. 
52
-
63
)
Fiehn
O
Garvey
WT
Newman
JW
Lok
KH
Hoppel
CL
Adams
SH
Plasma metabolomic profiles reflective of glucose homeostasis in non-diabetic and type 2 diabetic obese African-American women
PLoS One
2010
, vol. 
5
 pg. 
e15234
 
Foster Daniel
W
Studies in the ketosis of fasting
J Clin Invest
1967
, vol. 
46
 (pg. 
1283
-
1296
)
Glueck
JC
Morrison
JA
Goldenberg
N
Wang
P
Coronary heart disease risk factors in adult premenopausal white women with polycystic ovary syndrome compared with healthy female population
Metab Clin Exp
2009
, vol. 
58
 (pg. 
714
-
721
)
Goodarzi
MO
Looking for polycystic ovary syndrome genes: rational and best strategy
Sem Rep Med
2008
, vol. 
26
 (pg. 
5
-
13
)
Hughes
C
Elgasim
M
Layfield
R
Atiomo
W
Genomic and post-genomic approaches to polycystic ovary syndrome-progress so far: mini review
Hum Reprod
2006
, vol. 
21
 (pg. 
2766
-
2775
)
Insenser
M
Martínez-García
MA
Montes
R
San-Millan
JR
Escobar-Morreale
HF
Proteomic analysis of plasma in the polycystic ovary syndrome identifies novel markers involved in iron metabolism, acute-phase response, and inflammation
J Clin Endocrinol Metab
2010
, vol. 
95
 (pg. 
3863
-
3870
)
Kafkewitz
D
Bendich
A
Enzyme-induced asparagine and glutamine depletion and immune system function
Am J Clin Nutr
1983
, vol. 
37
 (pg. 
1025
-
1030
)
Krentz
AJ
von Muhlen
D
Barrett-Connor
E
Searching for polycystic ovary syndrome in postmenopausal women: evidence of a dose-effect association with prevalent cardiovascular disease
Menopause
2007
, vol. 
14
 (pg. 
284
-
292
)
Liye
Sun
Weihong
Hu
Qiao
Liu
Qinfang
Hao
Bo
Sun
Qi
Zhang
Sha
Mao
Jie
Qiao
Xianzhong
Yan
Metabonomics reveals plasma metabolic changes and inflammatory marker in polycystic ovary syndrome patients
J Proteome Res
2012
, vol. 
11
 (pg. 
2937
-
2946
)
Luque-Ramirez
M
San Millan
JL
Escobar-Morreale
HF
Genomic variants in polycystic ovary syndrome
Clin Chim Acta
2006
, vol. 
366
 (pg. 
14
-
26
)
Ma
X
Fan
L
Meng
Y
Hou
Z
Mao
YD
Wang
W
Ding
W
Liu
JY
Proteomic analysis of human ovaries from normal and polycystic ovarian syndrome
Mol Hum Reprod
2007
, vol. 
13
 (pg. 
527
-
535
)
Matharoo-Ball
B
Hughes
C
Lancashire
L
Tooth
D
Ball
G
Creaser
C
Elgasim
M
Rees
R
Layfield
R
Atiomo
W
Characterization of biomarkers in polycystic ovary syndrome (PCOS) using multiple distinct proteomic platforms
J Proteome Res
2007
, vol. 
6
 (pg. 
3321
-
3328
)
Mihalik
SJ
Michaliszyn
SF
de las Heras
J
Bacha
F
Lee
S
Chace
DH
DeJesus
VR
Vockley
J
Arslanian
SA
Metabolomic profiling of fatty acid and amino acid metabolism in youth with obesity and type 2 diabetes: evidence for enhanced mitochondrial oxidation
Diabetes Care
2012
, vol. 
35
 (pg. 
605
-
611
)
Nicholson
JK
Foxall
PJD
Spraul
M
Farrant
RD
Lindon
JC
750 MHz 1H and 1H-13C NMR spectroscopy of human blood plasma
Anal Chem
1995
, vol. 
67
 (pg. 
793
-
811
)
O'Dwyer
ST
Smith
RJ
Hwang
TL
Wilmore
DW
Maintenance of small bowel mucosa with glutamine-enriched parenteral nutrition
JPEN J Parenter Enteral Nutr
1989
, vol. 
13
 (pg. 
579
-
585
)
Reo
NV
NMR-based metabolomics
Drug Chem Toxicol
2002
, vol. 
25
 (pg. 
375
-
382
)
Setji
TL
Brown
AJ
Polycystic ovary syndrome: diagnosis and treatment
Am J Med
2007
, vol. 
120
 (pg. 
128
-
132
)
Shaw
LJ
Bairey Merz
CN
Azziz
R
Stanczyk
FZ
Sopko
G
Braunstein
GD
Kelsey
SF
Kip
KE
Cooper-Dehoff
RM
Johnson
BD
, et al. 
Postmenopausal women with a history of irregular menses and elevated androgen measurements at high risk for worsening cardiovascular event–free survival: results from the National Institutes of Health–National Heart, Lung, and Blood Institute sponsored Women's Ischemia Syndrome Evaluation
J Clin Endocrinol Metab
2008
, vol. 
93
 (pg. 
1276
-
1284
)
The Rotterdam ESHRE/ASRM-Sponsored PCOS Consensus Workshop Group
Revised 2003 consensus on diagnostic criteria and long-term health risks related to polycystic syndrome (PCOS)
Hum Reprod
2004
, vol. 
19
 (pg. 
41
-
47
)
Urbanek
M
Legro
RS
Driscoll
DA
Thirty-seven candidate genes for polycystic ovary syndrome: strongest evidence for linkage is with follistatin
Proc Natl Acad USA
1999
, vol. 
96
 (pg. 
8573
-
8578
)
Wood
JR
Nelson
VL
Ho
C
Jansen
E
Wang
CY
Urbanek
M
McAllister
JM
Mosselman
S
Strauss
JF
III
The molecular phenotype of polycystic ovary syndrome (PCOS) theca cells and new candidate PCOS genes defined by microarray analysis
J Biol Chem
2003
, vol. 
278
 (pg. 
26380
-
26390
)