Characterizing the follicular fluid metabolome: quantifying the correlation across follicles and differences with the serum metabolome.

Participants were part of the Environmental and Reproductive Health (EARTH) study ( 25 ). Briefly, the EARTH study was a prospective cohort study that examined how dietary and environmental factors influenced fertility. Men and women attending the Massachusetts General Hospital (MGH) Fertility Center (2005–2019) for infertility evaluation or treatment were invited to participate. At enrollment, participants completed a questionnaire about their diet, environmental exposures, and reproductive health. Participants also had their height and weight taken by trained study staff to calculate body mass index (BMI; kg/m 2 ). The Institutional Review Boards at MGH and the Harvard T.H. Chan School of Public Health approved this study. All participants provided written informed consent. For the serum metabolomics study, 200 women undergoing ART cycle where the intention was to transfer fresh embryo(s) between 2005 and 2016 were randomly selected, as previously described ( 8 ). One hundred and twenty-five women had both a blood sample and follicular fluid sample taken from their first dominant follicle, which is the final analytical sample size used to compare the serum and follicular fluid metabolome. In addition, to compare follicular fluid samples across follicles during the same ovarian stimulation cycle, 26 women had both a first and second sample, 23 women had a first and third oocyte sample, and 18 women had a second and third oocyte sample. Women provided a non-fasting blood sample (~6 ml) during a routine morning appointment (7–10 AM) during controlled ovarian stimulation. Most commonly (78%) the blood sample was collected on day 3 of the ART cycle (the first day of ovarian stimulation) although some samples were collected during either the monitoring phase or on the day of oocyte retrieval (22%). Blood samples were centrifuged and stored at −20 °C at MGH until being transferred to Harvard to be stored at −80 °C. During oocyte retrieval, a follicular fluid sample was taken from women’s first three follicles with a 16 G needle. Each sample was collected in a separate tube prepared with 1 ml of flushing media. Once the oocytes were removed, the follicular fluid from each separate follicle was centrifuged to separate the supernatant and pellet. The resulting aliquots were then stored at −80°C. The blood serum and follicular fluid supernatant samples were sent in separate batches to Emory University for metabolomics analysis. In both instances, samples were randomized prior to analysis to minimize batch effects. Since some women had multiple follicular fluid samples, these samples were always included in the same batch but their order was randomly assigned. We applied a comprehensive workflow that has previously been successful in analyzing non-fasting samples and is described elsewhere ( 26 , 27 ). Briefly, samples were treated with an acetonitrile mixture. Both the serum and follicular fluid samples were analyzed using liquid chromatography with high resolution mass spectrometry (Dionex Ultimate 3000 RSLCnano; Thermo Orbitrap Fusion). To facilitate greater feature separation and detection, we used two chromatography columns, the C18 hydrophobic chromatography column with negative electrospray ionization (ESI) and the hydrophilic interaction chromatography (HILIC) column with positive ESI. The sheath and auxiliary gases were set at 30 and 5 for the negative ESI, respectively. The spray voltage was set at −30 kV for the negative ESI. For the positive ESI, the sheath and auxiliary gases were set at 45 and 25, respectively. The spray voltage was set at 3.5 kV. The mass-to-charge ratio (m/z) range was 85 to 2000, with a resolution of 60000 and 5.00 × 10 5 ions collected. With each batch, we used two quality control samples, NIST 1950 ( 28 ) and pooled human plasma (Equitech Bio). We used ProteoWizard to covert raw data files to .mzML files using apLCMS and xMSanalyzer ( 29 – 31 ). Unique features were characterized based on m/z, retention time, and ion intensity. We restricted our analysis to metabolic features that were detected in at least 20% of samples in the serum and follicular fluid because we were interested in metabolites that appeared in a majority of samples. Additionally, we restricted our comparison of the serum and follicular fluid metabolome to the women with a follicular fluid sample from their first aspirated follicle (n=125). We identified overlapping metabolic features between the serum and follicular fluid matrices based on a 10 ppm threshold for m/z using the find.overlapping.mzs function in the apLCMS package in R ( 29 ). From the initial list of overlapping features, we excluded features that had >40 s difference in retention time. We allowed for a more lenient matching criteria for the retention time because the two matrices were run independently. Finally, we screened the overlapping features to ensure a 1 to 1 match between the serum and the follicular fluid. If a serum feature matched with more than one follicular fluid feature (or vice versa) we only included the feature that had the closest retention time. We examined three types of Pearson’s correlations between the serum and follicular fluid: 1) the overall correlation, 2) the feature-specific correlation, and 3) the subject-specific correlation ( Supplemental Figure 1 ). In the sub-set of women with multiple follicular fluid samples, we examined these three correlations between features measured in a woman’s first and second follicle (n=26), first and third follicle (n=23), and second and the third follicle (n=18). The overall correlation provides a single correlation value that describes how correlated the average intensity of the metabolomic features are across all samples. The feature-specific correlation provides a correlation coefficient for each overlapping feature using linear regression; in this method, for each feature, there is a data point for each individual subject. We examined both raw and adjusted feature-specific correlations accounting for women’s age, BMI, initial infertility diagnosis, ovarian stimulation protocol, and year of sample collection. Variables for the adjusted feature-specific correlation were selected because they were suspected to be confounders based on a priori knowledge. Lastly, the subject-specific correlation provides a correlation coefficient for each subject across all the overlapped features between the two sample matrices. For all analyses, we log transformed the intensity of each feature to reduce the skewness of the distribution. We described each of these correlation coefficients using both descriptive statistics (medians and ranges) and plots. We confirmed the overlapping metabolic features between the serum and follicular fluids and across the follicles with level-1 evidence via comparison of the m/z to authentic standards (n=738) analyzed in our lab with the same methods ( 32 ). We further screened annotated features by their retention times and excluded those with >40 s difference. If features matched to multiple metabolites based on m/z and retention time, we selected those with the closest retention times. If there were still multiple metabolites identified, we selected metabolites that have been previously identified in both the serum and the follicular fluid. We used the Human Metabolome Database ( 33 , 34 ) to identify the super-class of each matched metabolite and violin plots to describe the distribution of adjusted feature specific correlation for each metabolite super-class. We investigated several demographic, reproductive, and study design characteristics that we hypothesized a priori might affect the subject-specific correlation between serum and follicular fluid metabolites. These included time between serum and follicular fluid sample collection (≤7 days, >7 days), initial infertility diagnosis (male, female, and unexplained), age (≤37 years, >37 years), and BMI (<25 kg/m 2 , ≥25 kg/m 2 ). In all instances we investigated whether there was a significant association between the variable and the subject-specific serum/follicular fluid metabolite correlation using one-way analysis of variance. Additionally, we restricted our analysis to only women with day 3 blood samples to further account for changes in the metabolome due to hormonal stimulation. We also conducted a sensitivity analyses adjusting for two dietary patterns, Western and Prudent, collected from a validated food frequency questionnaire ( 35 ), to examine the influence of diet on our results. We included both in our adjustment set for the feature-specific correlation. Dietary pattern diet was only available for a subset of women (n=120; 96.0%) because some women enrolled prior to the introduction of the food frequency questionnaire.

On average, the women in our study were 34.8 years (standard deviation [SD]: 3.9) and had a BMI of 23.9 kg/m 2 (SD: 4.6) ( Supplemental Table 1 ). The most common initial infertility diagnosis was unexplained (43.0%). The majority of women contributed a follicular fluid sample from their first follicle (n=125; 92.6%) while fewer had a follicular fluid sample included from a second (n=35; 25.9%) or third follicle (n=28; 20.7%). In the serum metabolome, 10,803 and 12,968 unique features were identified in the C18 negative and HILIC positive chromatography columns, respectively ( Figure 1 ). In the follicular fluid from women’s first follicle, 14,397 and 17,168 unique features were identified in the C18 negative and HILIC positive chromatography columns, respectively. After excluding features that appeared in less than 20% of samples, in the C18 negative column, 7,830 serum features and 10,790 follicular fluid features remained and between them 3,790 overlapping features were identified based on m/z. After screening the retention time and checking for 1:1 matching, 1,928 features overlapped. In the HILIC positive column, after applying the same exclusion criteria, 9,074 serum features and 5,542 follicular fluid features remained and 2,352 overlapping features were found between these two matrices. After checking retention times and 1:1 matching, 1,149 features overlapped between the serum and the follicular fluid. The overall correlation between the overlapped features in the serum and follicular fluid was moderate in the C18 negative column (r=0.45) and slightly stronger in the HILIC positive column (r=0.63) ( Table 1 ). The median feature-specific correlation was initially negligible but increased to a weak correlation after adjustment for covariates in both columns (C18: raw r=0.07, adjusted r=0.35; HILIC: raw r=0.10, adjusted r=0.37). Upon further investigation, adjustment for year of sample collection had the largest impact on the adjusted correlation coefficients followed by protocol, infertility diagnosis, age, and BMI ( Supplemental Table 2 ). Similar to the overall correlations, the median subject-specific correlations were also moderate, and were lower in the C18 negative column (r=0.42) compared to the HILIC positive column (r=0.59). We confirmed the identity of 27 overlapping metabolites in the C18 negative column ( Supplemental Table 3 ) and 36 overlapping metabolites in the HILIC positive column ( Supplemental Table 4 ) with level-1 evidence. Additionally, for features that had more than one match after screening, we reported all potential matches in Supplemental Table 5 . In the C18 negative column, the most common super-classes identified were 1) organic acids and derivatives (n=7; 25.9%), 2) organic oxygen compounds (n=6; 22.2%), and 3) lipids and lipid-like molecules (n=5; 18.5%) and benzenoids (n=5; 18.5%). In the HILIC positive column, the most common super-classes of the metabolites identified with level-1 evidence were 1) organic acids and derivatives (n=10; 27.8%), 2) organoheterocyclic compounds (n=9; 25.0%), and 3) lipids and lipid-like molecules (n=8; 22.2%). Across both columns, we did not observe any clear patterns in the adjusted feature-specific correlation by super-class ( Supplemental Figure 2 ; Supplemental Figure 3 ). In the C18 negative column, glutamine had the highest correlation (r=0.578) between the serum and follicular fluid while catechol had the lowest correlation (r=0.29). In the HILIC positive column, phosphatidylcholine had the highest correlation (r=0.59) between the serum and follicular fluid while isoleucine had the lowest correlation (r=0.26). In the C18 negative column, when comparing samples from women with a first and second follicle (n=26) and excluding features that appeared in less than 20% of samples, 10,776 overlapping features were identified ( Supplemental Figure 4 ). Similarly, when comparing the first and third (n=23) and the second and third (n=18) follicles, 10,387 and 11,195 overlapping features were observed, respectively. In the HILIC positive column, when comparing samples from women with a first and second follicle (n=26) and excluding features that appeared in less than 20% of samples, 5,929 overlapping features were identified ( Supplemental Figure 5 ). Similarly, when comparing the first and the third follicles (n=23) and the second and third follicles (n=18), 6,336 and 6,634 overlapping features were observed, respectively. The C18 negative column consistently had more overlapping features than the HILIC positive column. This was mainly due to the higher number of features that were present in >20% of the samples. The overall correlation between metabolites from different follicles within a woman was close to 1 and much higher across the three follicles than the serum-follicular fluid analysis (C18 & HILIC first-second: 0.74 & 0.79; C18 & HILIC first-third: 0.76 & 0.78; C18 & HILIC second-third: 0.81 & 0.85) ( Table 2 ; Table 3 ). The median feature-specific correlation was also high across all three follicles (C18 & HILIC first-second: 0.89 & 0.91; C18 & HILIC first-third: 0.88 & 0.90; C18 & HILIC second-third: 0.89 & 0.90). In total, we identified 23 metabolites across the follicles in the C18 negative column and 42 metabolites across the follicles in the HILIC positive column using level-1 evidence. In the C18 negative column, the most common super-classes of these metabolites were: 1) organic acids and derivatives (n=7; 30.4%), 2) organic oxygen compounds (n=6; 26.1%), and 3) benzenoids (n=4; 17.4%) ( Supplemental Figure 6 ). In the HILIC positive column, the most common super-classes were: 1) organoheterocyclic compounds (n=14; 33.3%), 2) organic acids and derivatives (n=12; 28.6%), and 3) lipid and lipid-like molecules (n=7; 16.7%) ( Supplemental Figure 7 ). In both the C18 negative and HILIC positive columns, time between serum and follicular fluid sample collection was not significantly associated with the subject-specific correlation (C18 p-value: 0.59; HILIC p-value: 0.64) ( Supplemental Table 6 ). Similarly, there was no evidence of effect modification by age or BMI. When examining infertility diagnosis, the subject-specific correlation did not vary in the C18 negative column (p-value: 0.77). However, in the HILIC positive column, individuals who had male infertility diagnosis (median r=0.601) had a slightly higher subject-specific correlation than those diagnosed with female (median r=0.586) or unexplained (median r=0.578) infertility (p-value<0.005). After adjustment by diet, median feature-specific correlation was similar to the unadjusted median feature-specific correlation ( Supplemental Table 7 ). Finally, when we restricted our sample to only women with blood samples collected before hormonal stimulation began (on cycle day 3), our results remained unchanged ( Supplemental Table 8 ).

Among women undergoing ART, we observed limited overlap and weak to moderate correlation between metabolomics features in the serum and follicular fluid. However, across follicles within a woman, we observed a high overlap and a high correlation between metabolites detected in follicular fluid samples from the first three dominant follicles. In summation, the follicular fluid represents a novel matrix, which may be a rich source of biochemical predictors of female fertility and reproductive outcomes than the serum. Based on our results, a single follicular fluid sample may be sufficient for estimating many subject-specific biomarker concentrations.

In our prospective cohort of women undergoing a fresh autologous IVF cycle, we observed limited overlap between metabolomics features identified in serum and follicular fluid. Furthermore, the overall, feature-specific, and subject-specific correlation coefficients demonstrated only a weak to moderate association between the intensity of overlapping features in these two matrices. In contrast, we observed a high degree of overlap with strong correlation coefficients between metabolomic features identified in the follicular fluid of a woman’s first three dominate follicles. Lastly, we detected more features in the follicular fluid metabolome than the serum metabolome which may indicate that there are metabolic features unique to the follicular fluid that may be relevant to reproductive health but additional studies are needed to further investigate this idea. While we are unaware of any previous untargeted metabolomics studies addressing the same research question as ours, there have been previous studies that have compared the concentrations of targeted metabolites in the serum and follicular fluid. In a targeted metabolomics study of 30 postpartum dairy cows, a weak-to-moderate positive correlation was observed for glucose (r=0.29), total cholesterol (r=0.24), and calcium (r=0.30) concentrations in the serum and follicular fluid ( 36 ). In a study of 14 women undergoing IVF, researchers observed lower concentrations of several sugars including glucose, glycerol, galactose, and galactosamine in the follicular fluid as compared to the serum ( 37 ). Additionally, they observed significant correlations for galactose (r=0.6540) and glycerol (r=0.6598) between the two matrices ( 37 ). A couple small, pilot studies among women undergoing IVF treatment showed that trace elements were detected in the majority of follicular fluid samples but at much lower levels than in the blood and urine ( 38 , 39 ). and only weak to moderate positive correlations were observed (range: 0 for Mn in urine/follicular fluid to 0.64 for Hg in blood/follicular fluid). In a study of 70 women with PCOS and healthy controls, three metabolites were identified in both the serum and the follicular fluid that were associated with PCOS ( 40 ). However, the follicular fluid metabolome demonstrated greater amino acid disturbances as compared with the serum metabolome which indicates that certain biological processes may be differentially altered depending on the biofluid ( 40 ). In another study of 15 women undergoing IVF treatment, Schweigert et al, observed that overall the protein content of the follicular fluid was different than the serum and they hypothesized that selective transport, rather than simple filtration, was occurring at the blood-follicle barrier ( 41 ). Taken together, these studies support the finding that metabolites in the serum and follicular fluid are positively, albeit weakly correlated, with certain metabolites, like those involved in energy and lipid metabolism possibly having a stronger concordance. The reason for the distinct metabolome is likely due to several factors. First, there are differences in the underlying purpose of each biological fluid. Serum allows for the transportation of nutrients, oxygen, and other necessary metabolites across for the body. In contrast, the follicular fluid is the medium that allows communication between the oocyte, the granulosa cells, and the theca cells ( 10 , 11 ). Second, the blood-follicle barrier, which exists between these two biological fluids, is an important structure that limits the access of foreign compounds and/or harmful substances (e.g., toxicants, drugs) to the developing follicles ( 42 ). It is an integral part of follicle development as it regulates the composition of the follicular fluid during folliculogenesis, so that different components (e.g., proteins, peptides, electrolytes, ions, sugars, and others) can be precisely and rapidly recruited to the follicular fluid in response to the needs of developing follicles ( 42 ). The microbiome is another source of variability that could have given rise to distinct metabolomes. In fact, several of the overlapping features that we identified with level-1 evidence are microbiota-dependent metabolites. For example, we might expect that metabolites linked to the gastrointestinal microbiome would be present in serum but less so in the follicular fluid due to the upper reproductive tract having a distinct bacterial community ( 43 ). Many of the overlapping metabolites we identified have been previously linked to several negative health outcomes. For example, leucine and isoleucine (both branched-chain amino acids) have been implicated as biomarkers for Type 2 diabetes ( 44 ) and linoleate, oleate, gamma-linolenic acid, and urate (all free fatty acids) have been linked to higher risk of cardiovascular disease ( 45 , 46 ). We also identified many metabolomic markers of inflammatory diseases including tyrosine and tryptophan. This indicates that although the follicular fluid metabolome is distinct from serum, there are many well-characterized metabolites that overlap. The increasing evidence linking reproductive history, and more specifically sub-fertility, as a marker of women’s long-term health ( 47 , 48 ), further suggests that reproductive-specific biospecimens, such as follicular fluid, may provide biological insight above and beyond the current biomarkers although further research is required. Among the first three follicles collected for each woman, we observed a high overlap of metabolomics features detected across the follicular fluid metabolomes. Additionally, all three correlation coefficients were much higher than those reported between the serum and follicular fluid. The high correlations were a bit unexpected given the previous literature. In a study of 26 women undergoing IVF treatment, Yang et al, observed that the size of the follicle influenced the metabolome and that smaller follicles had, for example, significantly higher levels of dehydroepiandrosterone (DHEA), a precursor for several steroid hormones, when compared to larger follicles ( 24 ). Because we only focused on a woman’s three most dominant follicles, which all tended to be large, our choice of study design could have limited the influence of follicle size on the follicular fluid metabolome and reduced this source of variability. This restriction also likely limited the influence of oocyte maturation on the follicular fluid metabolome. Another class of commonly investigated metabolites in the follicular fluid is lipids, specifically high-density lipoproteins (HDL) and related compounds ( 49 , 50 ). Two studies offer slightly different pictures. Kim et al investigated six women undergoing IVF and observed little variability in the concentration of 19 HDL particles across the women’s two ipsilateral follicles. In contrast, Bloom et al, examined two contralateral follicles from 180 women undergoing IVF and found substantial differences in HDL concentrations ( 50 ). In our study, we did not directly investigate lipid molecules but among the identified metabolites of the lipid and lipid-like super-class, we did observe a high (albeit wide-ranging) adjusted feature-specific correlation. While we did not directly collect information on whether the repeated follicular fluid samples in our study were retrieved from ipsilateral or contralateral follicles, the majority were probably ipsilateral, which could have resulted in a higher feature-specific correlation coefficients. While our study has several novel findings, it also has several limitations. First, although we used the same lab protocols and instruments for measuring the untargeted metabolomics, the serum and follicular fluid samples were run independently, which likely reduced the number of overlapping features since we were unable to fully match on retention time. Therefore, caution should be taken when interpreting the results of our serum-follicular fluid analysis. Second, the follicular fluid samples were treated with 1 mL of flushing solution, which could have resulted in variable dilution. In spite of this, we observed a high correlation between the follicles, but future studies should consider how flushing medium could alter results. Third, the metabolic standard we used to identify metabolites were based on serum samples which could have introduced bias into our investigation of the follicular fluid metabolome. However, metabolites would not be expected to have vastly different m/z and retention time across biological fluids so the bias introduced by using serum standards is likely low. Fourth, we used non-fasting serum samples, which could have impacted results. To minimize this potential impact, we applied a comprehensive metabolomics workflow using pooled standards and internal references. We also conducted a sensitivity analysis adjusting for dietary patterns and observed little difference. Fifth, while we adjusted for many key demographic and reproductive characteristics, residual confounding from other unmeasured confounders is still possible. Sixth, we were only able to identify, with level-1 evidence, a fraction of the metabolites that overlapped between the serum and the follicular fluid. This is a common problem in untargeted metabolomics where a vast number of metabolites go unidentified; it is likely that many of these unidentified metabolites play an important role in female reproductive function and future studies should investigate these further. Seventh, similar to the profile of most infertility clinic patients, the women in our study were mostly white and highly educated. While this may have helped increase internal validity by limiting noise due to these factors, we did observe some evidence of effect modification by infertility diagnosis with participants who had male infertility having a higher subject-specific correlation compared to participants with unexplained or female infertility in the HILIC column. Additionally, women in our sample were undergoing ovarian stimulation, which could have influenced the metabolomes. However, we conducted sensitivity analysis restricting to individuals with blood samples collected prior to stimulation and observed limited evidence to suggest this influenced our results. Still, caution must be taken in extending these results to the general population. Our study does have several strengths. Chiefly, we utilized established methods and state of the art technology to compare the metabolome of serum, an often utilized and well-characterized matrix, with follicular fluid, which is a more novel and less explored biofluid. By leveraging the extensive lifestyle and clinical data that was collected as part of the EARTH Study, we were also able to limit bias from a variety of sources. Finally, using a list over 738 chemical standards, we were able to utilize level-1 evidence to annotate metabolites, which facilitated a more detailed investigation.

Metabolomics is a relatively new subfield in the broader “omics” field that examines the presence and concentration of low molecular weight metabolites ( 1 ). In recent years, metabolomics has been applied to describe underlying biological pathways between exposure and disease as well as identify potential biomarkers of both exposure and the disease of interest ( 2 – 8 ). Metabolomics has utilized several different biological matrices including blood, urine, salvia, and to a lesser extent organ-specific tissues and biofluids, which are often more challenging to obtain. One of the underutilized biological matrices is follicular fluid ( 9 , 10 ). The follicular fluid is the immediate microenvironment surrounding the oocyte and is routinely obtained in women undergoing assisted reproductive technology (ART) during oocyte retrieval. This fluid is primarily composed of granulosa and theca cells, which by bidirectional communication with the oocytes, ensures the oocyte remains in meiotic arrest and aids in ovulation and fertilization ( 9 – 11 ). The follicular fluid provides the necessary components for oocyte development and assists in several metabolomic pathways including lipid and energy metabolism ( 9 , 11 ). To date, several studies have used metabolomic methods to characterize differences in the follicular fluid obtained from patients with various reproductive diseases, such as endometriosis ( 12 , 13 ), polycystic ovary syndrome (PCOS) ( 14 – 17 ), and diminished ovarian reserve ( 18 ); or to identify potential biomarkers of oocyte quality in the follicular fluid as a means to improve ART outcomes and success ( 19 – 23 ). While the follicular fluid emerges as an exciting biofluid for biomarker development, key methodological questions still remain. Several previous investigators have begun characterizing the biological variability of proteins, lipids, hormones, and specific metabolites in human follicular fluid, very little work has focused on describing the variability of the follicular fluid metabolome. In particular, it is not entirely clear if the follicular fluid metabolome is similar to the blood metabolome. Understanding the correlation patterns between these two biomatrices may provide evidence for the utility of serum biomarkers, which can be easily obtained, in place of follicular fluid biomarkers. While oocytes do not come into direct contact with the blood, the blood and follicular fluid do exchange metabolites between the blood-follicular fluid barrier. Therefore, we sought to compare the metabolomic features between the serum and follicular fluid of women undergoing ART. Additionally, given that some previous studies that have shown certain targeted metabolites may differ between follicles ( 24 ), we also sought to characterize the variability in metabolomic features across up to three follicles within a woman. We hypothesized that we would observe a lesser overlap in the serum and the follicular fluid metabolomes compared to the overlap in the metabolomes across the follicles.

Supplemental Figure 7. Violin plot demonstrating the adjusted feature-specific correlations for each super-class of metabolites identified between the follicles in the HILIC positive column using level-1 evidence Supplemental Figure 1. Graphical representation of three types of correlation Supplemental Figure 2. Violin plot demonstrating the adjusted feature-specific correlations for each super-class of metabolites identified in the serum and follicular fluid metabolomes in the C18 negative column using level-1 evidence (n=27) Supplemental Figure 3. Violin plot demonstrating the adjusted feature-specific correlations for each super-class of metabolites identified in the serum and follicular fluid metabolomes in the HILIC positive column using level-1 evidence (n=36) Supplemental Figure 4. Flowchart of follicular fluid features A in the C18 negative column among participants with multiple follicular fluid samples who were enrolled in the metabolomics sub-study of the Environmental and Reproductive Health cohort. A Number of features meeting the 20% rule may differ with follicles because the number of subjects vary for each comparison (i.e. the subjects with follicular fluid samples from their first and second follicles may not be the same subjects who have follicular fluid samples from their first and third follicles) Supplemental Figure 6. Violin plot demonstrating the adjusted feature-specific correlations for each super-class of metabolites identified between the follicles in the C18 negative column using level-1 evidence Supplemental Figure 5. Flowchart of follicular fluid features A in the HILIC positive column among participants with multiple follicles collected who were enrolled in the metabolomics sub-study of the Environmental and Reproductive Health cohort A Number of features meeting the 20% rule may differ with follicles because the number of subjects vary for each comparison (i.e. the subjects with follicular fluid samples from their first and second follicles may not be the same subjects who have follicular fluid samples from their first and third follicles)