Facing the diagnosis of hyperandrogenic polycystic ovary syndrome by a novel combination of circulating biomarkers

Abstract

The work-up of diagnosing a woman with polycystic ovary syndrome (PCOS) is still challenged by the lack of standardization of methods and threshold values used to establish the presence of its individual diagnostic criteria. This study aimed to address if a combination of circulating biomarkers suitable for clinical practice accurately discriminated among patients with polycystic ovary syndrome (PCOS) and non-hyperandrogenic women. The study included 253 premenopausal patients diagnosed with PCOS by state-of-the-art methodology and 71 non-hyperandrogenic controls with regular menses paired by body mass-index. We analyzed two circulating micro ribonucleic acids (miR-142-3p and miR-598-3p) that showed a strong association with PCOS in a previous validation study, and proton nuclear magnetic resonance spectroscopy metabolomic profiling of serum. Diagnostic accuracy was assessed by logistic regression models and receiver operating characteristics (ROC) curve analyses. Circulating miR-142-3p was under-expressed in patients with PCOS compared to controls. However, miR-598-3p expression was similar between patients with PCOS and non-hyperandrogenic controls. Patients with PCOS showed higher fasting circulating concentrations of alanine, leucine, isoleucine, glutamic acid, threonine, acetate, and lactate, and lower concentrations of glutamine and acetone, compared to controls. A binomial logistic regression [χ²(10): 94.2, P < 0.001] correctly classified patients with PCOS or non- hyperandrogenic individuals in 81% of cases. Only four biomarkers reached statistical significance in that model: miR-142-3p, isoleucine, acetate, and anti-müllerian hormone. The overall discriminatory ability of these variables was further addressed by ROC curve analyses. The model’s area under the ROC curve had a confidence interval from 0.826 to 0.917. However, the model’s performance was worse in identifying the subset of patients with non-hyperandrogenic PCOS. Patients with PCOS a particular pattern of circulating biomarkers comprised of miRNAs and low-molecular weight metabolites: compared to control women, PCOS patients consistently underexpressed circulating miR-142-3p and presented with higher circulating concentrations of isoleucine and acetate. When added to follicular-phase anti-müllerian hormone levels, this circulating biomarkers combination showed an excellent diagnostic performance in identifying hyperandrogenic PCOS, although its performance to detect normoandrogenic phenotypes was lower. 

Introduction

Despite a recent international consensus on assessment and management of the polycystic ovary syndrome (PCOS)¹, the process of diagnosing a woman with PCOS in clinical practice still presents with unsolved challenges. For instance, the cut-off values for defining hirsutism using clinical scores are controversial^1,2. Furthermore, the widespread implementation of reliable assays for androgen measurement − i.e.: mass spectrometry (MS)-based methods – is still pending in most routine laboratories worldwide^3–5, despite the fact that the use of less accurate immunoassays carries a certain risk of misdiagnosis^6,7. In the same vein, the clinical definition of ovulatory dysfunction varies between scientific societies^1,8, and establishing the presence of oligo-ovulation accurately may be time-consuming in women with anovulatory but regular menstrual cycles. Not to mention that the most accurate diagnostic marker for polycystic ovarian morphology (PCOM) – namely follicle number per ovary⁹– suffers from poor reproducibility among ultrasound (US) operators even if counting is performed in real time, and may lead to an overdiagnosis of PCOM¹⁰. In this regard novel automated systems may enhance accuracy¹¹, if available. In short, considering the very large prevalence of PCOS reported worldwide, and those of its associated comorbidities, a timely and proper diagnosis becomes of utmost importance in terms of public-health burden¹².

Searching for biomarkers to properly identify women with PCOS may contribute to solve present diagnostic limitations in the clinical setting. We recently reported that some circulating micro ribonucleic acids (miRNAs or miR), associated with androgen excess, could discriminate among women with PCOS and non- hyperandrogenic control women, especially when they were combined with other clinical and biochemical biomarkers¹³. Moreover, we have also reported that women with PCOS share with men an unfavorable intermediate metabolomic profile – mainly characterized by aromatic and increased branched chain amino acids (BCAA) as measured by proton nuclear magnetic resonance spectroscopy (¹H-NMRS) – that is not present in non-hyperandrogenic control women¹⁴.

Hence, certain miRNAs and metabolites might serve as biomarkers for PCOS, yet such an approach is still in need of a few technical updates, as follows. Firstly, our earlier reports relied on immunoassays to measure sex hormones. Although broadly used in clinical practice worldwide, these assays lack the sensitivity, specificity, and overall accuracy needed to assess circulating androgens in women with certainty⁷. Secondly, despite being a commonly used metabolomic platform, H-NMRS sensitivity and spectral resolution for low concentration analytes is limited, particularly for metabolites that bind to serum/plasma proteins, such as aromatic amino acids, or when macromolecular signals or other high concentration species are present¹⁵. In order to overcome these issues, several approaches have been developed, including protein precipitation and ultrafiltration methods, or signal subtraction of high-concentration metabolites such as glucose¹⁵. However, sample processing may hinder the translation of these novel biomarkers into clinical practice.

Thus, the rationale behind the present study was to identify a panel of biomarkers able to deal with these unsolved crucial questions, opening up the future possibility of including them into a one-step high through-output integrative multianalytical platform. With this general objective in mind, we aimed: (i) to confirm our previous results improving the quality of our studies by increasing their sample size and using a gold-standard method such as liquid chromatography–tandem mass spectrometry (LC–MS/MS) to measure circulating androgens; and (ii) to identify a selected combination of biomarkers able to discriminate accurately among women with PCOS and control women with regular menses, avoiding pretreatment of serum samples before their submission to a fine-tuned¹ H-NMRS platform, and therefore, facilitating its implementation for clinical practice.

Materials and methods

Subjects

Since 2002, we offer to sign a written informed consent form allowing us to include her data in a database for research purposes, including this study, to every woman attending our Reproductive Endocrinology clinic at the Hospital Universitario Ramón y Cajal from Madrid. We here used fasting serum/plasma samples ⸺ obtained between 8 and 10 AM ⸺ and clinical data from 340 consecutive premenopausal women referred because of symptoms of functional androgen excess or hyperandrogenemia between 2006 and 2022 (the flow of recruitment is detailed in Fig. 1). Two hundred fifty-three of them were eventually included since samples showing signs of hemolysis were excluded (see below). The present study was comprised of women with a diagnosis of PCOS by state-of-the-art methodology including LC-MS/MS assays for measurement of circulating sex steroids and assessment of PCOM by sonography or anti-müllerian hormone (AMH) immunoassay⁷. In detail, a diagnosis of PCOS required the presence of at least two out of three criteria: clinical and/or biochemical androgen excess, ovulatory dysfunction (OD), and PCOM, fulfilling accepted current consensus for PCOS diagnosis¹. Clinical hyperandrogenism was defined by hirsutism, i.e.: a modified Ferriman-Gallwey score ≥ 8¹⁷. For LC–MS/MS measurements and AMH assay, biochemical hyperandrogenism and upper limit of normality, respectively, were established by the presence of values > 95^th percentile of a sample of control women as detailed below. World Health Organization group II OD (https://www.ncbi.nlm.nih.gov/books/NBK327781/) was defined by the presence of more than six cycles longer than 36 days in the previous year, absence of menstruation for three consecutive months, or by two consecutive luteal phase progesterone concentrations below 17 nM in women presenting with regular menses, also fulfilling current consensus diagnostic criteria¹. Ovarian US examination was conducted in those women who presented with only hyperandrogenism or OD. On the contrary, in women meeting both criteria, US was not systematically performed because they had already met a PCOS diagnosis¹. PCOM was evaluated by gynecologists according to routine clinical practice based on the European Society of Human Reproduction and Embryology/American Society for Reproductive Medicine 2003 criteria¹⁸;hence, those examinations also fulfilled 2023 criteria for clinical practice¹. We systematically excluded women with other etiologies of hyperandrogenism or OD including drugs, virilizing tumors, hyperprolactinemia, thyroid dysfunction, functional hypothalamic amenorrhea, non-classic congenital adrenal hyperplasia, and Cushing’s syndrome, among others^1,16. Likewise, patients had no history of oophorectomy or hysterectomy, nor had received treatment with hormonal contraceptives, antiandrogens, or insulin sensitizers for at least 6 months before sampling.

Fig. 1.

Flow of study recruitment. BMI, body mass index; LC-MS/MS, liquid chromatography-tandem mass spectrometry; NCCAH, non-classic congenital adrenal hyperplasia; PCOS, polycystic ovary syndrome; WHO, world health organization.

As controls, we included a group of 91 premenopausal female volunteers composed of hospital’s staff, and overweight or obese women seeking advice solely for weight loss at our Department. From them, only the samples from 71 women, that did not show signs of hemolysis, were included in the study. Controls presented with no signs of hyperandrogenism in their physical examination and reported regular menses, and were assessed by the same methods used to diagnose PCOS in patients. Control women were similar in terms of body mass index (BMI) to the study population of hyperandrogenic women⁷. Controls had no history of oophorectomy or hysterectomy, nor had received treatment with hormonal contraceptives, antiandrogens, or insulin sensitizers for at least 6 months before sampling either.

Considering miRNA levels can be influenced by hemolysis, we checked plasma for this frequent event by measuring the expression of miR-23a-3p and miR-451a in all samples¹⁹. Samples suggesting hemolysis because of a difference in cycle quantification of more than 7.5 (Cq_miR−23a – Cq_miR−451) were excluded from subsequent analyses. As stated below, the samples from 253 women with PCOS and 71 controls were finally included in the present study (Table 1).

Table 1.

Anthropometric, sex steroids and insulin sensitivity profile of women with PCOS and control women.

	Control women			Women with PCOS			P
	(n = 71)			(n = 253)
Age, years	29	±	6	27	±	6	0.031
Body mass index, kg/m2	26.6	±	7.5	27.2	±	7.6	0.199
Obesity (BMI ≥ 30 kg/m2), n (%)	22 (31)			84 (33)			0.725
Waist circumference, cm	82	±	16	83	±	18	0.468
Waist-to-hip ratio	0.80	±	0.12	0.81	±	0.09	0.154
Insulin sensitivity index	7.0	±	3.5	6.3	±	4.7	0.342
Anti-müllerian hormone, pM	22.7	±	15.7	50.4	±	37.5	< 0.001
LH-to-FSH ratio	1.2	±	1.4	1.5	±	1.0	0.085
Total testosterone, nM	0.9	±	0.3	1.2	±	0.6	< 0.001
Sex hormone-binding globulin, nM	55	±	26	45	±	27	0.014
Calculated free testosterone, pM	14.1	±	5.4	24.1	±	12.9	< 0.001
Androstenedione, nM	4.6	±	1.4	6.7	±	2.7	< 0.001
DHEA-S, µM	4.8	±	1.9	6.6	±	3.3	< 0.001
Estradiol, pM	220	±	198	151	±	114	0.002
Estrone, pM	166	±	93	185	±	107	0.093
PCOS phenotype
Classic	-			184 (73)			-
Ovulatory	-			13 (5)			-
Normoandrogenic	-			56 (22)			-

BMI, body mass index; DHEA-S; dehydroepiandrosterone-sulphate; FSH, follicle-stimulating hormone; LH, luteinizing hormone; PCOS, polycystic ovary syndrome. Data are means ± SD or counts (percentages). A general linear model was used for comparisons adjusting by age. Free testosterone was calculated by Vermeulen’s formula using a default albumin level of 43 g/L⁶²

The above-mentioned research database was approved by the local Ethics Committee from Hospital Universitario Ramón y Cajal (Date of approval: 7-March-2002; Reference number: 12/02). The informed consent was revised and approved again by the same local Institutional Review Board on 2022, March 10.

Assays

The technical specifications assays used in measuring sex steroids are detailed elsewhere⁷. Serum AMH concentrations were measured using an automated immunochemiluminescent assay on a Cobas e601 ^® analyser (Elecsys ^®, Roche Diagnostics, Germany). The assay limits of detection and lower limit of quantification were 0.07 and 0.21 pM, respectively. The intra-assay and inter-assay coefficients of variation were < 4%. The limit above the measuring range was 164 pM.

The relative quantification of plasma miR-142-3p and miR-598-3p was assayed as described earlier^13,20. RNA was extracted from 200 µl plasma using the miRNeasy Serum/plasma Advanced kit (Qiagen #217204, Hilden, Germany). One µg MS2 bacteriophage carrier RNA (Roche #10165948001, Mannheim, Germany) was added during the lysis step. RNA was reverse transcribed into cDNA using the miRCURY LNA RT kit (Qiagen #339340, Maryland, USA). The miRCURY LNA RNA spike-in kit (Qiagen #339390, Maryland, USA) was used for quality control of the RNA isolation (UniSp2 and UniSp5) and cDNA synthesis (UniSp6) steps according to the manufacturer’s instructions to monitor both yield and possible contamination by inhibitors. The expression of mature miRNAs in plasma was determined with a real-time and miRNA-specific polymerase chain reaction (PCR) using the miRNA LNA SYBR green PCR kit (Qiagen #339347, Maryland, USA) in a LightCycler ^® 480 II, LightCycler^® 480 Software release 1.5.0 [(Version 1.5.0.39); https://elabdoc-prod.roche.com/eLD/web/global/en/products/3.8.1.4.4.8] (Roche Diagnostics GmbH, Mannheim, Germany). To select the most stable endogenous reference miRNAs we used previous miRNA expression data from our laboratory, selecting miR-222-3p and miR-484 as endogenous controls^13,20. The geometric mean of the expression of these two reference miRNAs was used as a normalizing factor to control for different input amounts of RNA, minimize analytic variability, and to obtain reliable and reproducible results. For relative quantification we used the comparative threshold cycle method²¹. A difference in quantification cycle (ΔCq) was calculated by subtracting the mean of the Cq values of the reference miRNAs from the Cq of the target miRNA. ΔCq data expression was transformed logarithmically before further analysis with the following equation (expression = log2^−ΔCq; arbitrary units). All PCR assays were performed by triplicate for each sample, including target and reference miRNAs, spike-ins and hemolysis controls. The three spike-ins were consistent across all plasma samples, indicating adequate RNA extraction, reverse transcription and RNA quality.

Proton nuclear magnetic resonance spectrometry metabolomics profiling

Serum samples were shipped to Biosfer Teslab (Metabolomic Plarform Biosfer Teslab, CIBERDEM, 43206 Tarragona, Spain) in dry ice for the quantification of low-molecular weight (LMW) metabolites by ¹H-NMRS. Serum samples (200 µL) were diluted with 50 µL deuterated water and 300 µL of 50 mM, pH 7.4 phosphate buffer solution before analysis. ¹H-NMR spectra were recorded at 300 K on a Bruker Avance III 600 spectrometer (Bruker Biospin, Rheinstetten, Germany), operating at a proton frequency of 600 MHz. One-dimensional¹H pulse experiments were carried out using 1D Carr-Purcell-Meiboom-Gill (CPMG) spectra. Sixty-four transients were collected into 64k data points for each spectrum. The acquired spectra were phased, baseline-corrected and referenced before performing the automatic metabolite profiling of the spectra dataset through and improved adaptation of Dolphin software²². The software used in the current analysis offers enhanced deconvolution capabilities and higher degree of automation compared to that used in our previous reports^14,23.

Sample size

Setting a two-sided alpha value of 0.05 and a β value of 0.20, introducing 238 positive and 65 negative cases in our final regression model would allow us to detect as significant a difference of 0.08 points in a receiving operating characteristics (ROC) area under the curve value (AUC_ROC) compared with a hypothetical AUC_ROC value of 0.75 (null hypothesis), which is considered as fair-acceptable [MedCalc Software Ltd. Cálculo del tamaño de la muestra: Área bajo la curva ROC. https://www.medcalc.org/es/calc/sample-size-area-under-roc-curve.php (Version 23.4.5; accessed January 17, 2026)].

Statistics

The results were expressed as means ± standard deviations (SD) and 95% confidence intervals (CI) (lower limit; upper limit), or counts (percentages), as appropriate. We tested the normality of the distribution of continuous variables using the Kolmogorov–Smirnov test. Logarithmic transformation was applied to ensure normality as needed. Between group comparisons of continuous variables were performed by Student’s t tests, Mann-Whitney’s U tests or univariate general linear models as a function of the homogeneity of variances, distribution, and adjustment by covariates. Since control women were slightly older compared with those women with PCOS, age was included as a covariate in the analyses. ROC curve analysis was used to assess the diagnostic performance of miRNAs and selected metabolites derived from¹ H-NMRS.

We also analyzed whether the addition of clinical and biochemical variables to miRNAs and metabolites could improve the accuracy of ROC curves. With this goal, we implemented the following steps:

• i)We excluded those variables used for the definition of PCOS (androgens, SHBG, and hirsutism) to avoid spurious improvements in diagnostic performance. Follicular phase AMH levels were not excluded since these concentrations were needed to establish a definitive diagnosis of PCOS in only five women.
• ii)We assessed the overall accuracy for the diagnosis of PCOS of miRNA, LMW metabolites, and clinical and biochemical variables other than sex steroids, with significant differences between PCOS and control women, using ROC curve analyses.
• iii)We performed a binomial logistic regression introducing, within a single model, women status (control vs. PCOS) as dependent variable and those variables showing an AUC_ROCabove 0.5 as independent variables in the previous step. Linearity of continuous variables with respect to the logit of dependent variables was assessed via the Box-Tidwell procedure²⁴. Multicollinearity was evaluated by constructing a correlation matrix among predictors. We defined the lack of a strong association by a coefficient below 0.8. Then, we implemented forward and backward stepwise methods for automated model building using the likelihood-ratio test to identify the best predictors of the model.
• iv)Lastly, we combined those variables retained by the logistic regression model as significant predictors of PCOS to build a final predictive model. The relative contribution of each predictor to the model was evaluated by computing likelihood ratio tests. A five-fold cross validation was used to assess the internal validity of the model’s predictive performance on unseen data. This procedure randomly splits our dataset into 5 roughly equal parts (folds), and then the model is trained on 4 folds while one fold is left to test it, averaging the prediction errors across all folds to get a robust estimate of the model’s performance on new data. Pairwise comparisons of AUCs of the validation set from the full model compared to reduced models were performed using DeLong’s test.

We displayed sensitivity, specificity, positive predictive and negative predictive values derived from the model’s classification table using a probability threshold of 0.5 to define the presence of PCOS. In addition, the point of the ROC curve showing best discrimination (Youden index) was used to provide the cut-off values showing optimal combinations of sensitivity and specificity.

No imputation method was used to manage missing values. SPSS Statistics 22.0 (SPSS Ibérica, Madrid, Spain) and the R language and Environment for Statistical Computing [R Core Team. (2025) R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing. https://www.r-project.org/] were used for the analyses. A P value < 0.05 was considered statistically significant.

Results

The distribution of the various PCOS diagnostic criteria among PCOS phenotypes is described in detail in Table 1.

Circulating miR-142-3p was underexpressed in plasma samples of women with PCOS compared with control women (Fig. 2, upper panel). On the contrary, miR-598-3p expression was similar between women with PCOS and controls (Fig. 2, lower panel). The results ofH-NMRS metabolomics profiling in women with PCOS and control women are shown in Fig. 3. Women with PCOS presented with higher fasting circulating concentrations of alanine, leucine, isoleucine, glutamic acid, threonine, acetate and lactate, and lower concentrations of glutamine and acetone, compared with control women.

Fig. 2.

Relative expression of each miRNA in control women and women with polycystic ovary syndrome. Data are expressed as means and 95% CI for the mean of log2^−ΔCq. For comparisons, data were submitted to a general linear model introducing age as a covariate.

Fig. 3.

Proton nuclear magnetic resonance spectrometry metabolomic profilings in control women and women with PCOS. Data are shown as means (95% CI) of their µM concentrations. A general linear model was used for comparisons, introducing age as a covariate. The figures above the X-axis indicate available results, or in other words, those metabolites that did not fall below the detection threshold and were quantifiable with sufficient accuracy.^*P < 0.05 for the differences between women with PCOS and controls.

Then, we assessed the overall accuracy for the diagnosis of PCOS of miR-142-3p, LMW metabolites with significant differences between PCOS and control women, LH-to-FSH ratio, and AMH using ROC curve analyses. Among LMW metabolites, only alanine, leucine, isoleucine, threonine, acetate, glutamate, and lactate resulted into statistically significant AUC_ROC above 0.5 (Fig. 4). Introducing these biomarkers as independent variables within a single model, we performed a binomial logistic regression with women status (control vs. PCOS) as dependent variable. Based on the Box-Tidell procedure, all continuous independent variables were found to be linearly related to the logit of the dependent variable. A strong association between predictors was not detected, ruling out significant multicollinearity. No significant outliers were identified in the model. The logistic regression model was statistically significant (Nagelkerke’s R² = 0.47; χ² (10) = 94.2; P < 0.001) and correctly classified 81% of cases. Sensitivity was 85.9%, specificity 60.0%, positive predictive value was 91.0%, and negative predictive value was 47.4%. Of the 10 predictive variables tested, only four were statistically significant: miR-142-3p, isoleucine, acetate, and AMH. Forward and backward stepwise methods for automated model building using the likelihood-ratio test retained these same four variables as the best predictors of the model. The changes – percentage increase compared with previous step (Δ) – in Nagelkerke’s R² corresponding to each step were 0.24 for AMH; 0.35 for AMH and acetate (Δ48.1%); 0.41 for AMH, acetate and miR-142-3p (Δ14.7%); and 0.45 for AMH, acetate, miR-142-3p, and isoleucine (Δ11.8%) (Fig. 5). The likelihood ratio test revealed that the relative contributions of each variable to the model discrimination were LR χ² = 59; P < 0.001 for AMH, LR χ² = 16; P < 0.001; LR χ² = 13; P < 0.001 for mi-142-3p, isoleucine and LR χ² = 11; P < 0.001 for acetate. The model equation of the four parameters was as follows: log[P(PCOS)/1-P(PCOS)] = −1.178–3.287(mir-142-3p) + 0.448(AMH) + 0.030(acetate) + 0.069(isoleucine), where P(PCOS) represents the probability of having PCOS.

Fig. 4.

Receiver-operating characteristics (ROC) curve analyses for mi-142-3p, low-molecular weight metabolites, and biochemical variables with statistically significant differences between women with PCOS and controls. The area under the ROC curve [AUC_ROC (95% CI)] indicates the probability of each biomarker of correctly identifying PCOS, hyperandrogenemia, or ovulatory dysfunction, respectively. The reference line in red corresponds to an AUC_ROC of 0.5 that suggests no discrimination.

Fig. 5.

Receiver operating characteristic (ROC) curve analysis displaying the diagnostic accuracy of diverse biomarkers combinations including miR-142-3p, isoleucine, acetate, and anti-müllerian hormone. The area under the ROC curve [AUC (95%CI)] indicates the ability of the models for discriminating between control women and women with PCOS. N = number of subjects included in each model after excluding missing values. The diagonal reference line in red corresponds to an AUC of 0.5 that suggests no discrimination. The red point represents the best diagnostic accuracy according to Youden’s J statistics. The horizontal green dotted line corresponds to 90% sensitivity. The vertical blue dotted line corresponds to 90% specificity. Pairwise comparisons of the full model and reduced models were conducted using DeLong’s tests.

We also conducted a sensitivity analysis by excluding the five patients in whom PCOM criteria were based only on AMH values, in order to avoid spurious associations. In this subset of women, the logistic regression model was also statistically significant (Nagelkerke’s R² = 0.45; χ² (4) = 88.6; P < 0.001) and correctly classified 83% of cases. Sensitivity was 86.3%, specificity 65.1%, positive predictive value was 82.4%, and negative predictive value was 49.1%.

The overall discriminatory ability of the binomial logistic regression model was addressed by a new ROC analysis introducing as a test variable the predicted value for the model of four independent variables (Fig. 5). The AUC ranged from 0.826 to 0.917, which is considered as an excellent/outstanding discrimination²⁵. Pairwise comparisons of the full model and reduced models with fewer predictors using DeLong’s test are also shown in Fig. 5. In other words, removal of any single biomarker led to a significant drop in AUC, confirming that each one provides unique diagnostic information. The best discrimination point according to Youden’s index (Fig. 5, red point) yielded a 65% sensitivity, 95% specificity and predicted probability of 0.874. Cut-off values at this point were 0.61 log2^−ΔCq for miR-142-3p, 24 µM for isoleucine, 65 µM for acetate, and 23.0 pM for AMH. Cut-off values of 0.61 log2^−ΔCq for miR-142-3p, 24 µM for isoleucine, 65 µM for acetate, and 23.0 pM for AMH had 65% sensitivity and 95% specificity for diagnosing PCOS. For a 90% sensitivity (Fig. 5, dotted green line), the highest specificity value (55%) was obtained with cut-offs of 0.84 log2^−ΔCq for miR-142-3p, 22 µM for isoleucine, 25 µM for acetate, and 31.6 pM for AMH. The five-fold cross validation of the full model showed an AUC of 0.859 (95% CI: 0.812; 0.907), consistent with its performance on the full dataset, thereby supporting the internal validity of the predictive model (Fig. 6).

Fig. 6.

Receiver operating characteristic (ROC) curve analysis displaying a five-fold cross validation used to assess the internal validity of the model’s predictive performance on unseen data. The area under the ROC curve [AUC (95%CI)] indicates the ability of the models for discriminating between control women and women with PCOS. The blue line represents the average ROC curve of the five models derived from internal validation. The reference line in red corresponds to an AUC of 0.5 that indicates no discrimination.

We also analyzed the performance of the model analyzing separately non-obese and obese women. The model correctly classified 82% and 90% of cases, respectively (Table 2). When the analysis was conducted separately in women with hyperandrogenic phenotypes, non-ovulatory phenotypes, and non-hyperandrogenic phenotypes – after excluding five women in who PCOM diagnosis was only based on their AMH concentrations, as stated above –the model correctly classified 82%, 84%, and 73% of cases.

Table 2.

Performance of the discrimination model comprised of four biomarkers in several subset of women and PCOS phenotypes.

Subset of women	χ2(4)	P	R 2	AUCROC (95% CI)	Sensitivity	Specificity	PPV	NPV
Non-obese (BMI < 30 kg/m 2 )	68.0	< 0.001	0.428	0.861 (0.804; 0.918)	92.6	45.7	85.7	63.6
Obese (BMI ≥ 30 kg/m 2 )	45.0	< 0.001	0.611	0.936 (0.883; 0.990)	96.1	66.7	96.0	66.7
Hyperandrogenic phenotypes	101.8	< 0.001	0.496	0.883 (0.838; 0.927)	91.8	54.7	85.3	54.7
Non-ovulatory phenotypes	106.7	< 0.001	0.473	0.877 (0.833; 0.922)	93.3	50.0	86.8	68.1
Non-hyperandrogenic phenotypes	42.5	< 0.001	0.417	0.827 (0.753; 0.901)	64.0	79.7	71.1	73.9

AUC_ROC, area under the receiver-operating characteristic curve; BMI, body mass index; NPV, negative predictive value; PCOS, polycystic ovary syndrome; PPV, positive predictive value; R², Nagelkerke’s coefficient. The subset of women with hyperandrogenic phenotypes includes those individuals with classic and ovulatory PCOS. Non-ovulatory phenotypes are classic PCOS and non-hyperandrogenic PCOS.

Discussion

The first objective in this study sought to support our previous results showing the ability of some circulating miRNAs to discriminate between women with PCOS and control women, as well as to confirm a differential profile of LMW metabolites in women with PCOS compared to their control counterparts^7,13,14,20. After phenotyping our study subjects using state-of-the-art methodology⁷, we partially corroborate earlier findings on circulating miRNAs. In consonance with our previous study¹³, miR-142-3p was underexpressed in plasma of women with PCOS compared to control women, showing an overall accuracy for the diagnosis of PCOS almost identical to that previously reported. From a pathogenic viewpoint, miR-142-3p appears to downregulate the phosphatase and tensin homolog protein (PTEN) pathway²⁶. In turn, PTEN negatively regulates the Phosphoinositol 3 Kinase (PI3K)/Protein Kinase B (AKT) pathway, which is involved in cell proliferation, apoptosis, glucose metabolism and insulin signaling, inflammation, and oxidative stress²⁷, as well as in primordial follicle activation²⁸. Furthermore, miR-142-3p potentially targets vascular endothelial growth factor (VEGF)-related pathways¹³. In conceptual agreement, an aberrant ovarian angiogenesis appears to be a contributor to PCOS pathophysiology²⁹. Of note, miR-142-3p was a predictor for ovulatory dysfunction but not for hyperandrogenemia among our study subjects.

On the other hand, the results of the current study do not support our previous research endorsing an association between miR-598-3p and PCOS^13,20, which might be spurious according to current data. It is possible that the improvement in our methodology for measuring androgens in PCOS may lie beneath this difference, being important not only for clinical practice but also for experimental and translational research⁷. It should be noted that: (i) circulating androgens were measured by immunoassays in the study population of our previous reports^13,20; (ii) miR-598-3p correlated positively with several circulating androgens in those studies¹³, but not in the current report (data not shown); and (iii) in our own experience, a significant number of women with hyperandrogenic PCOS phenotypes according to immunoassays no longer presented with androgen excess when assessed by LC-MS/MS, and in some cases, even did not accomplish PCOS diagnostic criteria anymore⁷.

These present findings do confirm our earlier report supporting an abnormal intermediate metabolism in women with PCOS¹⁴. Elevated circulating BCAA levels are consistently associated to insulin resistance and obesity in animal and human studies³⁰. In spite of the question of causation or association is still unsolved, increased circulating BCAA derived from dietary sources may activate mammalian target of rapamycin complex 1 (mTORC1) and ribosomal protein S6 kinase β1 (S6K1) signaling, leading to insulin resistance by serine phosphorylation of insulin receptor substrates 1 and 2³⁰. Such a mechanism has been proposed to explain insulin resistance in PCOS³¹. In turn, insulin resistance might increase the rate of BCAA generation from protein degradation. Alternatively, impaired BCAA metabolism induced by dysfunctional adipose tissue, lipotoxicity, and/or oxidative stress would accumulate toxic metabolites causing mitochondrial dysfunction, and consequently, insulin resistance, β-cell dysfunction, and apoptosis³⁰. Finally, increased levels of BCAA are associated with an altered microbiome – which has an enriched potential for their biosynthesis –, and such intestinal microorganisms have been found to be deprived of genes encoding inward transporters for these amino acids in insulin-resistant individuals³².

The role of circulating acetate levels is poorly studied and controversial in women with PCOS, with small studies showing either increased fasting concentrations³³ or no differences with respect to control women^34,35. Our study, including a larger study population, showed higher circulating acetate concentrations at fasting in women with PCOS compared with their control counterparts. Acetate is a short-chain fatty acid (SCFA) derived from intestinal fermentation of indigestible foods³⁶. Its circulating concentrations depend on both production and absorption rates in the gut, which in turn are related to microbiome composition, intestinal function, and dietary patterns³⁶. While microbiote-generated acetate and other SCFAs such as butyrate and propionate appear to promote metabolic benefits as suggested by studies in animal models, gut or fecal SCFA might be increased in obesity-associated altered gut microbiome^36,37, as well as in fecal samples of women with PCOS showing a positive correlation with circulating androgen levels³⁸. Furthermore, a potential positive effect on glucose homeostasis may be blunted in hyperinsulinemic subjects³⁹. Of note, even our study was not designed to address the effect of obesity on metabolomics parameters such as acetate, exploratory analyses on our data did not find any statistically significant effect of obesity on their levels or interaction between PCOS status and obesity on them (data not shown). In view of our results, the interaction between gut microbiome and metabolomics in women with PCOS warrants further research.

As seen, these biomarkers may have pathophysiological grounds to be associated with PCOS. However, did they demonstrate to be reliable enough to accurately discriminate between women with and without PCOS? Our model combining these biomarkers and AMH yielded a diagnostic accuracy ranging from 83 to 92% in distinguishing between women with PCOS and control women. These figures are equivalent to those of markers widely used in clinical practice such as thyroid-stimulating hormone immunochemiluminimetric assays in discriminating between hyperthyrodism and euthyroidism⁴⁰, or hemoglobin A_1cand fasting plasma glucose in detecting diabetes⁴¹, and considerably better than circulating follicle-stimulating hormone concentrations in discriminating menopausal and premenopausal women⁴². In other words, the conditional probability that a woman does not have PCOS when our model does not predict PCOS (= 1 ─ negative post-test probability)⁴³ ranges from 93.9% to 95.4% according to current figures of PCOS prevalence^43,44. Since the pre-test probability of disease (PCOS status) in a symptomatic woman is very high, it is likely that this combination of biomarkers will not change the clinician’s practice when a high sensitivity threshold is used. However, a high specificity cut-off would yield a low negative likelihood ratio significantly decreasing the probability that the patient has PCOS, avoiding unnecessary misdiagnoses, and thus, prompt allowing the physician to seek another aetiology for the patient’s complaints. In our view, the routine practice in clinics that are not familiar with Reproductive Endocrinology, or conducted by physicians without expertise on this field, would appear to become the main scenario for applicability of these combination of biomarkers.

Regarding the usefulness of introducing circulating AMH into the model, the concordance of the latter and ultrasonography for the diagnosis of PCOM is somehow controversial^7,45,46. Circulating AMH does not appear to be a reliable biomarker for diagnosing PCOS by itself either⁴⁷. In consonance, AMH alone only showed an acceptable ability of discrimination among our study subjects, yet it proved to be a valuable complement to other biomarkers reaching an excellent-outstanding ability of combined discrimination in our regression model. Not surprising, AMH in PCOS is not only a marker of OD, but may be associated with hyperandrogenic features⁷.

Other studies involving circulating biomarkers combinations have been reported. Three miRNAs (miR-223-3p, miR-4488, and miR-151a-5p) were found to be differentially expressed between PCOS women and age-matched controls from South China by sequencing analysis and confirmed by reverse transcription-quantitative PCR⁴⁸. miR-151a-5p and miR-4488 expression levels were significantly upregulated, and miR-223-3p expression was downregulated in the PCOS cohort compared with the control cohort. The AUC_ROC were 0.889, 0.871, and 0.664 for miR-4488, miR-151a-5p, and miR-223-3p, respectively. Of note, combining AMH levels with the three miRNAs resulted in an AUC_ROCof 0.967, and a higher sensitivity and specificity. However, PCOS phenotyping used immunochemiluminiscence assays, which is associated with a significant risk of misdiagnosis⁷. Another case-control study from Spain⁴⁹, including non-obese and obese participants, defined a method for molecular classification of PCOS based on unbiased identification of miRNA biomarkers and decision-tree protocols with a set of four miRNA (miR-93a, miR-28, miR-143, and miR-539). These data were robustly presented, but unfortunately, state-of-the-art methodology for PCOS diagnosis and a performance analysis across diverse PCOS phenotypes were lacking in that study.

At this point, are there any potential hurdles to implementing these diagnostic biomarker combinations in clinical practice? Bearing in mind the challenges for diagnosing PCOS in the clinical setting (such as the integration into routine practice of MS-based methods for sex-steroid measurements or reliable PCOM sonographic explorations), developing a multi-analytical biomarker platform might ease the diagnosis of the existing vast numbers of women with PCOS worldwide. Although AMH testing is widely available, the main drawback for the clinical translation of our findings is the accessibility of LMW metabolites measurements that currently need specialized laboratory techniques. miRNA quantification is relatively affordable through standard techniques already available in routine laboratories such as quantitative PCR, even though several challenges for clinical application exist⁵⁰. The translation of metabolomic assays to real-life application is also far from being solved⁵¹. Nevertheless, some successful examples support further efforts on this field⁵². Recently, a cost-effective nanoenzyme-based reflectometric platform has proven an excellent performance in multiplex detection of LMW metabolites within the mM range⁵³. In the same line, magnetic-microfluidic chemiluminiscence platforms for accurate quantification of miRNA will be likely suitable for point-of-care applications in the very next future⁵⁴. Notwithstanding, any future initiative will need additional multidisciplinary technical actions, as well as rigorous cost-effectiveness analyses to reach routine practice.

Our research may pose, however, several limitations to be solved. Women with PCOS in our series mostly presented with the classic hyperandrogenic phenotype. Although, our discrimination model worked on other phenotypes, the small sample size of these subsets of women prevented us from extrapolating our results to milder non-hyperandrogenic phenotypes of the syndrome. In this regard, the model showed a lower sensitivity and PPV to identify women with non-hyperandrogenic PCOS. Of note, the inclusion into this group of women with an emerging condition such as obesity-related female hypogonadism ‒ otherwise women who fulfill current non-hyperandrogenic PCOS phenotype diagnostic criteria ‒ might contribute to this fact⁵⁵. In the same line, an external validation is of utmost importance to avoid overstating the diagnostic accuracy of any model. Therefore, it is mandatory to validate our findings in a population-based cohort, including a greater number of women with normoandrogenic PCOS prior to generalizing the applicability of our results to all PCOS phenotypes. The possibility exists that this subset of patients need another integrative approach leaned on multi-modal diagnostic inputs by combining ultrasound imaging with metabolic biomarker data as recently proposed⁵⁶. Likewise, the diagnostic accuracy of this combination of biomarkers for PCOS would undoubtedly need to be checked in populations including women with other functional hyperandrogenic disorders – particularly non-classic congenital adrenal hyperplasia – or ovulatory disorders in order to determine if they are specific biomarkers for PCOS. To this concern, miR-142-3p has been consistently associated with obesity and type 2 diabetes^57,58, but to the best of our knowledge, the role of miR-142-3P and LMW metabolites on other hyperandrogenic or ovulatory disorders, excluding PCOS, have not been tested so far. Regarding the biomarkers derived from metabolome, we could not standardize diet before sampling and did not use food diaries to address long-term differences in diet among the subjects. These are factors that could impact on gut microbiota, which is a major contributor to the metabolome, and it also constitutes a limitation of our study²³. However, this scenario may be more appropriate for real-life conditions and, in our experience, the diet composition of women with PCOS does not differ from that of premenopausal non-hyperandrogenic women⁵⁹. Anyhow, future research with basal diet standardization may make sure the metabolite differences are actually caused by PCOS pathophysiology. We did not perform preanalytical protein precipitation and ultrafiltration, and thus, amino acids highly bound to proteins ‒ e.g.: aromatic amino acids ‒ were not able to be accurately quantified, either. Nonetheless, this lack of sample pre-treatment could ease the integration of metabolite profilings into a one-step multianalytical platform. Furthermore, the differences in pre-analytical processing, analysis software, and quantification – semi-quantitative (values expressed in arbitrary units) vs. quantitative (values expressed in µM concentrations) – preclude us from direct comparisons with our earlier reports^14,23. Notwithstanding, the use of an advanced software with a greater performance in selected metabolite quantification compared to prior versions, allowed more robust and reproducible results, minimizing the likelihood of false positives, and enhancing overall data quality. The reliability and validity of our findings are ensured by this refinement in methodology, which provides greater confidence in the reported data. Among other strengths of our study, we may remark the careful phenotyping of the large sample of women studied here, and the inclusion of control women selected to be similar with the cases in terms of BMI.

Conclusions

Patients with PCOS in our series consistently underexpress plasma miR-142-3p. At fasting, they also present with a differential pattern of LMW metabolites with respect to control individuals. These circulating biomarkers, combined with follicular-phase AMH levels, showed an excellent performance in identifying women with hyperandrogenic phenotypes of PCOS, regardless of obesity. However, additional research involving a greater number of normoandrogenic PCOS patients (and possibly population-based cohorts) would be required to confirm the effectiveness of the biomarker panel in those women. Recent advances in multiplexed analyses for the simultaneous identification of diverse biomarkers in one sample^60,61 may ease the diagnosis of these patients. Finally, even though the present study may be an exciting proof-of-concept, more work needs to be done on technology before the practical application of this biomarkers combination.

Abbreviations

AKT

Protein kinase B

AMH

Anti-müllerian hormone

AUC_ROC

Areas under the ROC curve

BCAA

Branched chain amino acids

BMI

Body mass index

CPMG

Carr-Purcell-Meiboom-Gill

ΔCq

Difference in cycle quantification

¹H-NMRS

Proton nuclear magnetic resonance spectroscopy

LC-MS/MS

Liquid chromatography–tandem mass spectrometry

LMW

Low-molecular weight

LR

Likelihood ratio

MS

Mass spectrometry

miRNA/miR

Micro ribonucleic acids

OD

Ovulatory dysfunction

mTORC1

Mammalian target of rapamycin complex 1

PCOM

Polycystic ovarian morphology

PCOS

Polycystic ovary syndrome

PI3K

Phosphoinositol 3 Kinase

ROC

Receiver operating characteristics

S6K1

Ribosomal protein S6 kinase β1

SCFA

Short-chain fatty acid

SD

Standard deviations

VEGF

Vascular endothelial growth factor

Author contributions

M.L.-R. and H.F.E.-M. designed the study, recruited subjects, researched, and analyzed data, wrote the discussion, reviewed and edited the final version. M.Á.M.-G. and M.I. processed samples, performed assays, researched and analyzed data and contributed to the discussion. A.L.-Q. and S.L.V. processed samples, analyzed data and contributed to the discussion. L.N.-C. recruited subjects, analyzed data, and contributed to the discussion. J.Q-S. analyzed data and contributed to the discussion. M.L.-R. and H.F.E.-M. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. All authors have read and agreed to the published version of the manuscript.

Data availability

All data sets generated and/or analyzed during the current study are not publicly available but are available from the corresponding author on reasonable request.

Declarations

Competing interests

Based on the results of this work, M.L.-R., H.F.E.-M., M.A.M.-G., and M.I. have requested for grant of an European patent as inventors (submission number: 300553124; application number: EP25382105.2) which is currently pending. S.L.Q., A.Q.T., L.N.-C., and J.Q.-S. have no conflicts of interest.

Ethics approval and consent to participate

All experiments were performed in accordance with relevant guidelines and regulations. All research involving human research participants was also performed in accordance with the Declaration of Helsinki, and accomplished the law 3/2.018, of December 5, 2008, on the Protection of Personal Data and was accordance with the Regulation (EU) 2016/679 of the European Parliament and of the Council of April 27, 2016 on Data Protection. All patients and controls provided informed consents allowing us to include their data in a database for research purposes including this study. That research database was approved by the local Ethics Committee from Hospital Universitario Ramón y Cajal (Date of approval: 7-March-2002; Reference number: 12/02). The informed consent was revised and again approved by the same local Institutional Review Board on 2022, March 10.