Integrated chemical exposome-metabolome profiling of follicular fluid and associations with fertility outcomes during assisted reproduction.

This pilot study leveraged a subset of 84 patients undergoing oocyte retrieval at the Emory Reproductive Center who enrolled in the Emory Prospective Opportunity for Women’s Health Research (EmPOWR) cohort in Atlanta, GA ( Knight et al., 2022 ). The patients were recruited between May 2018 and August 2020 prior to treatment and oocyte retrieval for either IVF (including both intracytoplasmic sperm injection and conventional insemination) or fertility preservation. Participants were eligible if they were 18–50 years of age, could understand written and spoken English, and had no history of chemotherapy. For this analysis, we further excluded two patients with oocyte counts above 50, as explained in the Statistical Analysis section. Our final analytical sample size was 82 patients. All participants provided written informed consent, and the study protocol was approved by the Emory University Institutional Review Board (STUDY00003644). As described in detail previously ( Knight et al., 2022 ), fertility treatment for patients at the Center involved controlled ovarian stimulation through daily injections of gonadotropins to stimulate maturation and growth of follicles. There were four different types of stimulation protocols: only gonadotropins (through either an antagonist, Lupron flare, or agonist/Lupron down-regulation protocol) and gonadotropins combined with oral medication of clomiphene citrate and antagonist. Final oocyte maturation was induced through a trigger injection of either human chorionic gonadotropin (hCG) or gonadotropin-releasing hormone (GnRH) agonist. Oocyte retrieval was performed 35 h later via an ultrasound-guided needle aspiration through the vaginal wall. Mature oocytes were either vitrified (for those undergoing oocyte cryopreservation) or fertilized with sperm the same day as retrieval via conventional insemination or intracytoplasmic sperm injection (ICSI), with subsequent embryo culture for up to six days. Embryos that progressed to the blastocyst stage were graded for quality and then either transferred into the uterus or cryopreserved with or without trophectoderm biopsy. After embryo transfer, a pregnancy test was checked 10–11 days later, and then a clinical pregnancy was determined by ultrasound at six to eight weeks gestation based on the presence of a gestational sac with crown-rump length and cardiac activity. Other information collected included the antral follicle count (AFC), which was the number of antral follicles 2–10 mm on both ovaries counted via ultrasound on the first day of the stimulation protocol, and anti-Müllerian hormone (AMH), which was measured in the blood sample drawn most recently prior to protocol start. For our study, qualified physicians abstracted the medical record data into REDCap on the outcomes of controlled ovarian stimulation, including the number of retrieved, mature, and fertilized oocytes, and for patients subsequently undergoing embryo transfer, the clinical pregnancy outcome. In this study of a pilot cohort, our main outcome of interest was the number of oocytes successfully retrieved ( n = 82), although secondary analyses also evaluated clinical pregnancy ( n = 66) when sample size was sufficient. Other measures included the mature oocyte count ( n = 80), which was the number of total retrieved oocytes that were assessed for maturity and reached metaphase II (MII) stage, as well as the fertilized count ( n = 68), calculated as the number of oocytes successfully fertilized and showing two distinct pronuclei (2PN). There were two patients with completely missing data on mature oocyte count, because they either underwent only conventional fertilization on all oocytes (wherein oocytes are not stripped as they are for ICSI) or vaginal culture (device containing oocytes and sperm inside vagina for five days) and thus did not have any oocytes assessed for maturity. A few patients only had a portion of their oocytes eligible to be counted in the maturity measure (depending on their protocol), which thus could yield an underestimate. We also lacked fertilization data for patients only undergoing fertility preservation. During retrieval, follicular fluid was collected along with the oocytes from each aspirated follicle by a qualified reproductive endocrinologist physician. Following removal of the oocyte(s) from the collection tube by an embryologist, the remaining follicular fluid was retained for research purposes. About 70–100 mL of follicular fluid was pooled from most if not all the aspirated follicle(s), regardless of whether or not the follicle yielded an oocyte. Human tubal fluid flushing media (consisting of Dulbecco phosphate-buffered saline [PBS] at 7.0–7.2 pH) was used to flush the follicles during oocyte retrieval and thus was also present in the pooled follicular fluid. The samples were stored at 4 °C until processed via centrifugation for 15 min at 1200 g , washed with PBS, centrifugated for 8 min at 300 g , and then stored at −80 °C. For LC-HRMS using both C 18 and HILIC columns, the follicular fluid samples were prepared in one batch alongside quality control (QC) samples to ensure data integrity and method reliability. All samples and QCs were prepared together using an Opentrons OT2 automated liquid handler and 96-well plates (Thermo Scientific Abgene 0.8 mL polypropylene microplates). The follicular fluid samples were thawed at 4 °C, and 30 μL of sample was extracted by adding 90 μL acetonitrile (Fisherbrand UHPLC-MS Acetonitrile) containing 13 C labeled internal standards. Treated samples were vortexed for 2 min, equilibrated at 4 °C for 30 min, and then centrifuged for 15 min (3,220× g , 4 °C). Two aliquots of supernatant (30 μL) were then transferred to 96-well plates containing either 60 μL water for C 18 analysis (Fisher Chemical Optima Water) or 60 μL 1:1 acetonitrile/water for HILIC analysis. These were vortexed 2 min and placed in a refrigerated autosampler until analysis. Untargeted analysis was performed using separate systems designated for either C 18 or HILIC chromatography. Each system was comprised of a Vanquish Duo ultra high-performance liquid chromatography system (Thermo Fisher Scientific, Rockford, IL, USA) coupled to an Exploris 120 high-resolution mass spectrometer (Thermo Fisher Scientific, Rockford, IL, USA). Samples were analyzed using dual column chromatography with mobile phases optimized for positive or negative ionization. All samples were analyzed in reverse phase using TARGA C 18 columns (5 μm, 50 × 2.1 mm; Higgins Analytical, Inc, Mountain View, CA, USA) in both positive and negative mode. HILIC separation was accomplished using a SeQuant ZIC-HILIC column (3.5 μm 50 × 4.6 mm; Merck KGaA, Darmstadt, Germany) for positive mode and an XBridge Amide column (3.5 μm 3.0 × 50 mm; Waters Corporation, Milford, MA) for negative mode. Further details on our LC-HRMS methodology have been published elsewhere ( Pan et al., 2025 ) and build upon previous advancements ( Johnson et al., 2010 ). For our validated GC-HRMS methods ( Hu et al., 2021 ), the follicular fluid samples were prepared in a single batch alongside QC samples using automated liquid handling (Opentrons OT2) and glass-coated polypropylene wellplates. Samples were thawed at 4 °C and 100 μL sample was extracted by adding 20 μL of formic acid and 200 μL 4:1 (v/v) hexane/ethyl acetate containing 13 C labeled internal standards. The samples were vortexed at 4 °C for 60 min and then centrifuged at 3220× g for 10 min at 4 °C. The organic layer was collected and transferred to another plate containing Waters Technologies QuEChERS powder of MgSO 4 , primary and secondary amine (PSA), and C 18 sorbent to produce extracts suitable for volatile and semi-volatile compounds analyzed by GC while removing interferences to the column or detector. This mixture was vortexed for 2 min and subsequently centrifuged. The upper organic layer was then used for instrumental analysis. Untargeted analysis was performed with a Trace 1610 GC system coupled to an Exploris GC-HRMS system (Thermo Fisher Scientific, Rockford, IL, USA). Separation was accomplished by injecting 6 uL of extract into a programmable temperature vaporizer (PTV) inlet with a variable temperature program connected to a 15 m, 0.25 mm ID, 0.25 μm Rxi-5Sil column equipped with an integra-guard 10 m column (Restek), 1.0 mL/min flow rate of helium, and the following oven temperature gradient: 50 °C hold for 1.5 min, increase to 320 °C at 25 °C/min and hold for 4 min. The HRMS resolution was set at 60,000 full width at half-maximum (FWHM) at mass-to-charge ratio ( m / z ) of 200. The scan range was set from 70 to 700 m / z ; the automatic gain control (AGC) target and maximum injection time in full-scan MS settings were set to Standard and Auto, respectively. This allowed the instrument to automatically adjust the injection time for an ion count of ~1 × 10 6 for MS. The electron ionization source was maintained at 280 °C and 70 eV during the course of the run. QC samples were analyzed multiple times across the study batch. For LC analysis, this included NIST Standard Reference Material 1950 (Metabolites in Frozen Human Plasma) (x2), method blanks (x3), and repeat analysis of two contrasting plasma pools from BioReclamationIVT that included a mix of races/ethnicities and genders and that were separated for younger individuals aged 21–35 and older individuals aged > 50 (x6 for each pool). QC samples for GC analysis included solvent blanks (x2), instrument blanks (x10), method blanks (x2), process blanks of solvents prepared and extracted alongside samples (x2), alkane mixtures (x2), internal standard solutions (x4), NIST Standard Reference Material 1958 (Organic Contaminants in Fortified Human Serum) (x2), and replicates of the two contrasting plasma reference pools (x3 for each pool). The instrument blanks were analyzed to monitor instrument background and potential for contamination. Internal standards of 13 C labelled chemicals were added to study samples and/or run as QC solutions to assess instrument performance before analysis, to verify consistent response relative to both neat standards and extracted samples, and to flag any issues within a batch related to column deterioration, sample preparation errors, or loss in instrument sensitivity. The labelled chemicals included congeners of major brominated diphenyl ethers, persistent organic pollutants, and polychlorinated biphenyls for GC analysis and amino acids, cortisol, glucose, nicotine metabolites, caffeine, and various PFAS for LC analysis. Process blanks were run to monitor for contamination introduced during sample preparation and extraction. Method water blanks and solvent blanks were used to assess the purity of reagents and detect potential carryover. Pooled plasma samples served as controls to evaluate method reproducibility, retention time stability, and instrument performance across runs. In addition, the certified NIST reference materials were analyzed to verify method accuracy and support traceability to established standards. Leveraging both sets of pooled plasma reference samples, we found strong Pearson correlation coefficients across all features (among those detected in at least half of samples) of 0.99 on average with a range of 0.91 to 1.00. For the first pool, there were 11,144 features with coefficients of variation (CVs) less than 10 % and another 17,434 with CVs less than 20 %. For the second pool, there were 11,175 features with CVs less than 10 % and another 18,142 with CVs less than 20 %. The slight differences reflect potential variability in sample preparation, as these samples were not technical replicates. After analysis, raw instrument files were converted to.mzML files with peak picking and extracted using XCMS ( Smith et al., 2006 ; Tautenhahn et al., 2012 ). Because of ion fragmentation in GC-HRMS analysis, detected features (of those with max intensities five-fold change above method blanks) were deconvoluted into spectral clusters representing mass spectral fragments from the same compound using RamClustR, which grouped detected m / z corresponding to the same compound and cluster-averaged the peak intensities ( Broeckling et al., 2014 ). Retention time indices for GC-HRMS were calculated based on the retention time relative to the two closest eluting alkanes using Equation 1 in the Supplementary Material . The final feature tables for LC and GC data contained information on each feature’s m / z , retention time (LC) or retention time index (GC), and ion intensity. Chemicals were identified by matching detected m / z (±3 ppm for GC and ± 5 ppm for LC) and retention time (± 50 retention time index for GC and ± 15 sec for LC) to an in-house reference library of ~450 GC standards and ~1,200 LC standards of environmental chemicals and endogenous metabolites analyzed using the same method parameters. The standards included classes of phthalates; phenols; pesticides; PFAS; smoking markers; bile acids; sterols; carnitines; short-chain, medium-chain, and long-chain fatty acids; steroid hormones; and organic acids with confirmed m / z , retention time, adduct, and other characteristics. These annotations are considered confirmed identifications due to the comparison to authentic in-house reference standards analyzed using our same method and have the highest annotation confidence of Level 1 under the Schymanski criteria ( Schymanski et al., 2014 ). To quantify concentrations of target PFAS in follicular fluid using our validated method within the same untargeted workflow, an 8-point external calibration curve ranging from 0.05–40 ng/mL was prepared using Accustandard Native PFAS Reference Standard (M-8327) in Gibco newborn calf serum (NBCS, Thermo Fisher Scientific) ( Chang et al., 2025 ). In addition to the calibration curve, single preparations of 1 and 10 ng/mL PFAS standards in NBCS were analyzed at the beginning and end of the study batch to ensure stability. All calibration standards were prepared and analyzed within the same LC batch as the study samples and quality control samples. The PFAS signals were normalized against twelve 13 C isotopically labeled PFAS internal standards. For all PFAS calibration curves, the R 2 values were >0.99. Our laboratory participates in PFAS method accuracy tests three times per year through the CTQ AMAP Ring Test for Persistent Organic Pollutants in Human Serum. The PFAS analyte list is provided in the Supplementary Material . To characterize the presence of novel and unexpected exposures, detected features were annotated using a series of chemical databases to generate Level 4 annotations ( Schymanski et al., 2014 ). Only features detected with max intensities in study samples five-fold or higher above average method blank intensities were considered for annotation to avoid contributions from chemicals related to background or procedural contamination. For annotation of GC features, spectral clusters were matched to two separate databases using the MetaboAnnotation package in R with match results filtered to reverse search score thresholds >0.7. The first database included accurate mass spectra and combined our GC Chemical Library v2.1, the Thermo GC-Orbitrap Contaminants Library, and a published PFAS library ( Casey et al., 2023 ). Spectra were matched using a mass accuracy of 30 ppm, retention index error of 25, and forward and reverse scoring thresholds of 0.5 and 0.7 respectively. The second database included low-resolution, unit-mass spectra and was matched using a mass error of 0.5 Da and retention index error of 25 ( Tsugawa et al., 2020 , Tsugawa et al., 2015 ). For LC annotation, untargeted features were annotated using seven reference databases and xMSannotator. Matching thresholds included a match accuracy of ±3 ppm, ±10 s retention time tolerance between adducts/isotopes, a 0.9 correlation threshold between adducts/isotopes, and a mass defect window of 0.01 for the isotope search ( Uppal et al., 2017 ). Our primary reference database was the Norman Substance Database (Norman SusDat), a collection of over 100,000 unique environmental chemicals, commercial products, their predicted metabolites, and pharmaceuticals from over 70 contributors around the world ( Mohammed Taha et al., 2022 ). For our statistical analyses, we pruned the Norman SusDat annotations to only those from their source lists related to exogenous chemicals, excluding databases of pharmaceuticals and drinking water contaminants that also capture pharmaceuticals and hormone-related compounds (see Supplementary Material for our list of Norman sources, including PFAS-, pesticide-, and plastic-specific lists). Other available annotations for our LC data came from the Reference Set of Metabolite Names (RefMet) ( Fahy and Subramaniam, 2020 ); PubChem Lite version 1.18.0 ( Bolton et al., 2020 ); the DSSTox master list of per- and polyfluoroalkyl substances (PFAS) ( US EPA, 2020 ); curated pesticide-related lists from CompTox; a plastic chemical database compiled from the PlastChem report ( Wagner et al., 2024 ); and the functional use database (FUSEDB), a list of chemicals with functional use information in the Chemicals and Products Database of CompTox ( Dionisio et al., 2018 ). In the case of multiple Level 4 annotation matches to an LC feature, we selected the one with the highest confidence (based on xMSannotator’s confidence schema) and then with the highest number of total PubMed and/or patent reference counts ( Bolton et al., 2020 ). We investigated the associations between chemical exposures, metabolic pathways, and IVF outcomes, namely the number of retrieved oocytes ( n = 82 patients), and in secondary analyses, the odds of clinical pregnancy after embryo transfer ( n = 66 patients). Our statistical approach included: 1) single-chemical regression models of the associations between untargeted chemical feature intensities in follicular fluid and the IVF outcomes, 2) weighted quantile sum (WQS) regression models of the cumulative chemical mixture effects on oocyte count, and 3) pathway enrichment analyses of potential underlying metabolic pathways that are significantly associated with both the chemical exposure mixture effects and the oocyte yield. For statistical analyses, we included untargeted features detected in at least 25 % of follicular fluid samples. We did not apply stricter thresholds so that we could still evaluate potential risk from less common exposures or biological mechanisms, however, we used a manual zero-quantile technique to account for low-detected chemicals in our mixture models (see Section 2.4.3 ). The features’ peak intensities (relative abundances) were log 2 -transformed after substituting zero non-detect values with half the minimum detected value for that feature. Furthermore, we conservatively excluded two patients who exhibited extreme outliers for the oocyte counts (over 50 oocytes retrieved) that could disproportionately distort model results. This yielded a final sample size of 82 patients for our primary analyses. In our preliminary exposome-wide association study (EWAS) analysis, we conducted single-chemical regression models for each untargeted feature and the two outcomes (quasi-Poisson models for oocyte count and logistic models for clinical pregnancy). We adjusted for a parsimonious list of important covariates: age (continuous); current smoking status (yes and no); ovarian stimulation protocol type, categorized into three groups based on study proportions (gonadotropins antagonist protocol; gonadotropins with clomiphene and antagonist; and other [gonadotropins, lupron flare or down-regulation protocol]); and race/ethnicity, categorized into four groups based on data availability and study population proportions (non-Hispanic White, non-Hispanic Black, Asian, and Other races/unknown). Because of the large number of features, we corrected for multiple testing using the Benjamini-Hochberg (BH) procedure and evaluated the significance of the associations under different false discovery rates (FDRs) (5–20 %). We summarized these results in Manhattan plots, along with more in-detail effect sizes for the small subset of features with confirmed environmental chemical identities. These effect sizes for oocyte count were exponentiated to provide rate ratios and scaled to represent a standard-deviation (SD) increase in the chemical intensity, so that the single-chemical effect sizes could be roughly compared to each other as well as the WQS mixture index effect sizes. Even among the confirmed chemicals, we used the same corrected significance threshold as for the entire EWAS. Following our statistical workflow, we leveraged WQS regression with its random subsets (RS) and repeated holdouts (RH) implementations to evaluate the cumulative effect of a chemical mixture index on IVF outcomes and to identify individual chemicals driving the total mixture effects ( Carrico et al., 2015 ). Critically, this advanced mixture method accounts for the fact that untargeted exposure data have 1) high dimensions (number of chemicals ≫ sample size), which are not supported in traditional regression ( Tanner et al., 2019 ), and 2) high correlation patterns, which can introduce bias in traditional regression due to multi-collinearity and co-confounding ( Braun et al., 2016 ; Carrico et al., 2015 ; Weisskopf et al., 2018 ). We used the random subsets implementation of WQS for high-dimensional data to estimate chemical weights in the mixture index over many random subsets of the chemical predictors, instead of over many random bootstrapped samples that use the full set of chemical predictors each time. In this way, each RS leveraged a different combination of chemicals, serving to de-correlate the exposure data, avoid confounding by multi-collinearity, and improve generalization accuracy ( Curtin et al., 2021 ). To improve representativeness and stability of the results, especially given our limited sample size, we used the repeated holdouts option in WQS to repeat the models 100 times over different random partitions of participants into a 40 % training data set and 60 % testing data set ( Tanner et al., 2019 ). Based on the RHs, we determined the 95 % confidence intervals for the median estimates across the distribution of results, thus providing a measure of the uncertainty in chemical contributions. In our WQS analyses, we focused on oocyte count as the outcome (quasi-Poisson), since clinical pregnancy ( n = 66) had a very small sample size especially with WQS partitioning of the data into random training and testing sets in each repeated holdout. We conducted three WQS models for the different mixtures of environmental chemicals detected by GC, LC-HILIC, and LC-C 18 instrument methods in the follicular fluid (positive and negative ion modes within the HILIC and C 18 columns were combined). The chemicals were selected based on: 1) potential biological relevance: negatively associated with oocyte count at p < 0.10 from the single-chemical regression models, and 2) potential environmental exogenous origin: having a confirmed exogenous identity or having an annotation match with an above-zero confidence level in the subset of database sources from Norman SusDat related to environmental chemicals (see Untargeted Feature Annotation section), after excluding confirmed endogenous metabolites. From the annotations, to reduce feature redundancy we also excluded any additional adducts or isotopes (retaining the main M + H or M−H adduct) among those meeting the highest confidence rating of three by xMSannotator. Note that we could not be sure we only selected exogenous environmental chemicals because of uncertainties in annotation matches, although we did remove source lists of pharmaceutical biomarkers and other endogenous substances from the Norman database prior to annotation. Furthermore, as a comparison point, we conducted a smaller WQS model based on the environmental chemicals with confirmed identities across methods (under Level 1 annotation confidence based on ~1,650 available reference standards) with negative associations at p < 0.10; we did remove confirmed features with duplicate identities here, which made a negligible difference in sensitivity analyses. To interpret the effect estimates in multiple ways, we presented them using both the standard WQS units of per-decile increase and our transformed units of per-SD increase (calculated as the median of the effect estimates per SD of the index within each repeated holdout). Before WQS analyses, the chemical intensities were manually quantized into deciles, except that non-detects (zeros) were separated into their own zero-quantile. We performed the WQS models with 100 repeated holdouts, each with 2,000 random subsets (looking at √( # features) at a time by default) in the 40 % training data set. For estimating average weights, we applied a positive constraint to the optimization function and used the exponential of the t-statistic in the signal function, and we only had weights derived from models with positive slopes. We adjusted the models for the same covariates as before: age, race/ethnicity, current smoking status, and ovarian stimulation protocol type. In the results, if the overall mixture effect was significant, the weights in the index could then be interpreted as the relative importance of each chemical component’s contribution to the mixture’s effect on the outcome. We classified the importance based on the percentage of repeated holdouts in which the chemical’s average weight was higher than 1/ n chemicals , the equi-weight threshold if all chemicals contributed equally to the mixture effect ( Bennett et al., 2022 ; Busgang et al., 2022 ). To identify functional metabolic pathways potentially underlying the association between chemical mixture exposure and oocyte count, we conducted a meet-in-the-middle (MITM) analysis leveraging all untargeted LC-HRMS features ( Babin et al., 2023 ). For exposure–metabolite relationships, we used single-chemical quasi-Poisson regression models to test the associations between each WQS chemical mixture index (independent variable) and each untargeted metabolite (dependent variable), adjusted for the usual covariates. For the metabolite–outcome relationships, we used the results from our oocyte count EWAS. We then performed two pathway enrichment analyses by inputting all unadjusted p -values from either the exposure–metabolite or the metabolite–outcome analysis, using the permutation-based weighted hypergeometric test in the metapone R package ( Tian et al., 2022 ). Metapone annotates and predicts functional biological activity of untargeted metabolite features from both positive and negative ion modes simultaneously, while down-weighting the influence of metabolites with multiple annotations and thus less certainty ( Holzhausen et al., 2023 ; Tian et al., 2022 ). It leverages the following databases for pathway annotation: Kyoto Encyclopedia of Genes and Genomes (KEGG) ( Ogata et al., 1998 ), Mummichog ( Li et al., 2013 ), and The Small Molecule Pathway Database (SMPDB) ( Frolkis et al., 2010 ), although we filtered to human-only pathways by using flag==1 from the database at github.com/EMERGE-EXPOSOME/Metapone-pathway . We set the mass tolerance to 5 ppm, the number of permutations to 1,000, the unadjusted p -value threshold for metabolites to 0.05, the fractional count power to 0.5, and the maximum match count to 10, and we filtered to human-only pathways. Annotated adducts included: M, M + H, M−H2O + H, M + Na, M + ACN + H, M + ACN + Na, M + ACN + 2H, 2 M + H, and M + 2H for positive mode, and M−H, 2 M−H, M−2H + Na, M + Hac-H, M−H2O−H, and M + Cl for negative mode. Pathways were deemed significant if the overall BH-adjusted p < 0.05 (corresponding to an FDR < 5 %) and if at least three of its weighted metabolites were significant. We then evaluated pathways that were significant and overlapping for both exposure–metabolite and metabolite–outcome analyses. To help validate the significant metabolic pathways, we subsequently conducted a single-metabolite regression analysis of the associations between the untargeted features with confirmed identities as endogenous metabolites (Level 1 annotation confidence) and the oocyte outcomes (oocyte count, mature oocyte count, fertilized count, and clinical pregnancy). These quasi-Poisson or logistic regression models were adjusted for the same covariates as before and were corrected for multiple testing using an FDR < 20 % threshold. We also gathered the involved metabolic pathways of these confirmed features using their HMDB IDs linked based on InChIKey, PubChem ID, or Chemical Name (when needed).

Table 1 provides characteristics of the 82 patients included in our analysis who underwent oocyte retrieval at the Emory Reproductive Center in Atlanta, GA. They were 36 years old on average, with 62 % identifying as non-Hispanic White, 16 % as Asian, and 15 % as non-Hispanic Black. Male-factor infertility was involved in 39 % of cases (which was not mutually exclusive with female-factor infertility). On average, our patients had 16 oocytes retrieved and 12 oocytes that matured to MII stage. An average of nine oocytes were successfully fertilized (2PN embryos) among those not undergoing oocyte cryopreservation, and an average of four blastocyst-stage embryos were available to use or freeze. Of the 66 patients who had an embryo transfer, 32 (39 %) had a clinical pregnancy, and one (1 %) had an ectopic pregnancy. The number of oocytes retrieved ( n = 82) was highly correlated with other markers of ovarian reserve and outcomes of ovarian stimulation, including the number of mature oocytes (Spearman r = 0.90, p < 0.00001, n = 80), the number of successfully fertilized oocytes ( r = 0.90, p < 0.00001, n = 69), the AMH blood level prior to cycle ( r = 0.71, p < 0.00001, n = 82), and the baseline AFC ( r = 0.71, p < 0.00001, n = 80). Oocyte count was also strongly correlated with the ovarian sensitivity index (OSI), calculated as the oocyte count divided by total gonadotropin dose to account for the effect of stimulation dosage ( Revelli et al., 2020 ) ( r = 0.80, p < 0.00001, n = 82); so, we did not conduct separate analyses for OSI. We detected a total of 104,071 untargeted chemical signals in the follicular fluid samples, including a range of exogenous environmental chemicals, endogenous metabolites, and other biomarkers. After clustering GC-detected signals to account for fragmentation, 69,892 chemical features remained, including 64,662 that were detected in at least a quarter of participants ( Fig. 1 ). Annotation using accurate mass matching yielded 17,727 of those common features with potential identity matches, while the identities of 622 features could be confirmed based on our ~1,650 in-house reference standards (Level 1 confidence). Note that there was potential overlap in detected features across the five instrument methods. Eighty-two features were identified as exogenous environmental chemicals rather than endogenous metabolites (under Level 1 confidence after removing duplicates across instrument modes) ( Table 2 ). Of these, a total of 80 were detected in at least 25 % of samples, 77 were detected in ≥ 50 % of samples, 75 were detected in ≥ 75 % of samples, 64 were detected in ≥ 90 % of samples, and 44 were detected in 100 % of samples. The chemical class with relatively higher variability in detection rates tended to be the pesticides, and nicotine exposure was only detected in five patients (two of which were known current smokers) ( Table 2 ). Twenty-four targeted PFAS were able to be confidently identified and quantified as absolute concentrations in the same workflow as the untargeted data ( Table S1 ). Based on the targeted data, eleven of the PFAS were detected in over 80 % of follicular fluid samples, while the rest were in less than half. Four targeted PFAS were found in every follicular fluid sample: perfluoroheptanoic acid (PFHpA), perfluorohexane sulfonic acid (PFHxS), perfluorononanoic acid (PFNA), and perfluorooctanoic acid (PFOA). Targeted perfluorooctane sulfonic acid (PFOS), perfluoroundecanoic acid (PFUnA), and perfluorodecanoic acid (PFDA) were found in at least 98 % of follicular fluid samples. The summary of concentrations for the quantified PFAS are provided in the Supplementary Material . We identified 11,867 untargeted features in follicular fluid that were significantly associated with the number of oocytes retrieved, after correcting for multiple testing with a false discovery rate of 20 % ( p < 0.04) and adjusting for age, race, smoking, and ovarian stimulation protocol type ( Fig. 2 ). This included 3,081 features associated with oocyte count in the negative (adverse) direction and 8,786 associated in the positive direction. Under a stricter false discovery rate of 5 % ( p < 0.002), there were 2,788 features significantly associated with oocyte count, with 527 in the negative direction and 2,261 in the positive direction. In logistic regression models for clinical pregnancy (among 66 patients who underwent an embryo transfer), no features were significant after correction for multiple testing. However, 221 features in follicular fluid were nominally associated with the odds of achieving a clinical pregnancy at an unadjusted p -value < 0.01 (not correcting for multiple testing), including 101 features in the negative adverse direction (of which 21 were detected by GC) and 120 in the positive direction (of which 26 were detected by GC). Among those features with confirmed exogenous environmental chemical identities, none were significantly associated with oocyte count or clinical pregnancy after correcting for multiple testing under the FDR of 5 % ( Table 2 ). A few chemicals were significantly associated in the positive or negative direction with oocyte count under the less conservative FDR of 20 %. For example, an SD increase in levels of triethyl phosphate (TEP) was significantly associated with a 0.89 times lower expected number of retrieved oocytes (95 % CI: 0.81–0.97; p = 0.01) and an SD increase in benzo[a]pyrene (B[a]P) with a 0.88 times lower expected number of oocytes (95 % CI: 0.80–0.97; p = 0.02). Another three chemicals were associated with oocyte count in the positive direction: thiamethoxam (1.15; 95 % CI: 1.04–1.28; p = 0.01), perfluorooctanoic acid (PFOA) (1.12; 95 % CI: 1.01–1.25; p = 0.03), and perfluoroheptanesulfonic acid (PFHpS) (1.13; 95 % CI: 1.02–1.26; p = 0.03). In WQS models of untargeted environmental chemicals detected in follicular fluid, we found that the chemical mixture indices were negatively associated with the expected number of retrieved oocytes, after adjusting for age, race, current smoking status, and ovarian stimulation protocol type and after averaging across repeated holdouts ( Table 3 ). Specifically, a decile increase in the environmental chemical mixture index was significantly associated with between 0.69 and 0.74 times lower expected number of retrieved oocytes (i.e., 26 % to 31 % lower) on average, in separate models for the mixture of 587 priority chemicals detected by the GC instrument (0.74; 95 % CI: 0.57–0.88), the mixture of 181 priority chemicals detected by LC-HILIC (0.73; 95 % CI: 0.66–0.82), and the mixture of 150 priority chemicals detected by LC-C 18 (0.69; 95 % CI: 0.61–0.81). The number of chemicals contributing to the cumulative mixture effect at an average weight higher than the equi-weight threshold were 199, 63, and 62 for the GC, LC-HILIC, and LC-C 18 chemical mixtures, respectively ( Table 3 ). Of those important chemicals, there was one for GC, 17 for LC-C 18 , 24 for LC-HILIC, and two for Confirmed that had a weight higher than the threshold within at least 50 % of the repeated holdouts in the WQS models ( Fig. 3 ). Although the most important chemical from GC was not annotated, detailed information on the available annotations for all features with average weights above the equi-weight threshold are provided in Table S2 of the Supplementary Material . Among the 125 LC features above threshold, which were selected based on the Norman SusDat database, three were annotated in source lists for PFAS, six for pesticides, and 23 for plastic-related chemicals (not mutually exclusive) at Level 4 annotation confidence. The novel PFAS annotations included 1-perfluoropropylethanol, dodecafluoroheptanol, and bis(nonfluorobutyl)phosphinic acid, which were not among our confirmed standards list. The plastic-related chemical annotations included 4- tert -butylstyrene; diheptyl phthalate (DHPP); triethyl phosphate (TEP); bis- sec -butyl peroxydicarbonate; bis(2-ethylhexyl) decanedioate; butylbenzene; ethylene acrylate; tetralin; 1-dodecyl-2-pyrrolidinone; 2-phenoxyethanol; 2,2′-methylenebis(4-methyl-6- tert -butylphenol); 4-ethenyl-phenol; 4-methoxybenzaldehyde; cyclohexanone, 2-(1-cyclohexen-1-yl)-; glycidyl methacrylate; methylhexahydrophthalic anhydride; N-butyldiethanolamine; N-phenyl-1-naphthylamine; bis[2-(2-butoxyethoxy)ethyl] adipate; tert -butylhydroquinone; and benzenepropanoic acid, 3,5-bis(1,1-dimethylethyl)-4-hydroxy-, among others. Additional annotations of chemicals used in pesticides or pesticide production included 2,4-dichlorophenol, 1-phenylnonane, and dimethyl phosphate. A few annotated features were detected and found significant by both LC-HILIC and LC-C 18 methods, including TEP, diethyl hydrogen phosphate, and 2-phenylquinoline. In addition, 46 of the features were also annotated in the FUSE database. The FUSE database provided harmonized functional use(s), the most common being fragrance (61 %), flavoring/nutrient (48 %), solvent (30 %), softener/conditioner (30 %), deodorizer (26 %), antioxidant (20 %), plasticizer (17 %), and/or surfactant (11 %) ( Table 4 ). However, chemical names should be considered as annotations (Level 4), not confirmations. In a WQS model of the small mixture of seven exogenous environmental chemicals with confirmed identities and preliminary associations, we found that a single decile increase in the confirmed environmental chemical mixture index was associated with a 0.92 times lower expected number of retrieved oocytes (95 % CI: 0.86–0.98) ( Table 3 ). Four of these chemicals had average weights higher than the equi-weight threshold: benzyl butyl phthalate (BBP), propoxur, TEP, and benzo[a]pyrene. The latter two had weights above the threshold in at least 50 % of repeated holdouts, whereas the other two had weights above the threshold in 44–48 % of holdouts. In the larger mixture of any confirmed environmental chemicals regardless of preliminary relevance to oocyte count, the mixture effect of 0.96 did not reach statistical significance (95 % CI: 0.82–1.1) and may have been diluted by the irrelevant subset of the chosen mixture. For further analyses of the confirmed chemicals, we focused on the smaller mixture filtered by preliminary associations. The scatterplots show a clear negative relationship between the WQS cumulative mixture indices for each patient and their retrieved oocyte count, especially for the models of larger annotated mixtures that included hundreds of chemicals ( Fig. 4 ). The negative trends were consistent across the major racial/ethnic groups. As Fig. 4 demonstrates, the cumulative mixture index scores per patient mostly distributed between the third and sixth deciles, since it was unlikely for someone to have consistently low (or consistently high) deciles of exposure across all relevant chemical features aggregated in the mixture index. As such, the cumulative mixture effects may be better interpreted in transformed units of per SD rather than per decile, which were calculated to be 0.79 for GC (95 % CI: 0.70–0.88), 0.73 for LC-HILIC (95 % CI: 0.68–0.80), 0.74 for LC-C 18 (95 % CI: 0.67–0.80), and 0.88 for the confirmed mixture of seven chemicals (95 % CI: 0.81–0.97). These per-SD cumulative mixture effects were larger in magnitude of effect than any individual per-SD effect of the 44 WQS “possible contributors” in the single-chemical regression models, which had effect sizes of 0.90 (GC), 0.79–0.92 (LC-HILIC), 0.85–0.92 (LC-C 18 ), and 0.88–0.91 (confirmed) per SD increase in the single chemical feature’s intensity. We conducted numerous sensitivity analyses to examine implications of the parameters and decisions of our WQS modeling approach, as described in the Supplementary Material . The cumulative mixture effects remained consistent or stronger across nearly all the sensitivity models, even when we loosened the filtering criteria to higher p -value thresholds ( Table S3 ). In our meet-in-the-middle approach, we evaluated overlapping metabolic pathways between 1) WQS chemical mixture indices (i.e., the exposures) associated with untargeted metabolites (i.e., the intermediates) and 2) untargeted metabolites (the intermediates) associated with oocyte count (the outcome). We identified 20 metabolic pathways that were significantly enriched with both a chemical exposure mixture index and the oocyte count ( Fig. 5 ). These included metabolism of: biopterin, tryptophan, xenobiotics (by cytochrome p450), selenocompound, selenoamino acid, nicotinate, nicotinamide, methionine, cysteine, urea cycle/amino group, phenylacetate, glycine, serine, alanine, threonine, and pyrimidine. Other pathways involved the biosynthesis of pantothenate, coenzyme A (CoA), phenylalanine, tyrosine, tryptophan, and aminoacyl-tRNA, as well as ammonia recycling, protein digestion/absorption, and mineral absorption. Another 18 pathways (38 in total) were significantly enriched with a mixture index but not oocyte count, including glycerolipid metabolism, cytochrome p450 drug metabolism, nicotine addiction, beta-alanine metabolism, arginine and proline metabolism, nitric oxide signaling pathway, d-glutamine and d-glutamate metabolism, sulfate/sulfite metabolism, phosphatidylcholine biosynthesis, purine metabolism, prostate cancer, phospholipid biosynthesis, aldosterone-regulated sodium reabsorption, C21-steroid hormone biosynthesis and metabolism, plasmalogen synthesis, porphyrin and chlorophyll metabolism, and carnitine shuttle. Note that significant metabolites in some pathways do overlap with other pathways (for example, those from the prostate cancer pathway are also involved in ovarian steroidogenesis, steroid hormone biosynthesis, and other cancer pathways). There were substantially higher numbers of significantly enriched pathways for the larger untargeted chemical mixture indices (19–24 pathways each) than for the small chemical mixture based on only known, confirmed chemicals (one pathway: carnitine shuttle). To support the assessment of these underlying pathways, we also evaluated the univariate associations between abundances of confirmed endogenous metabolites in follicular fluid and controlled ovarian stimulation outcomes, adjusted for age, race, smoking, and stimulation protocol type ( Table S4 ). Of the metabolites known to play a role in our 20 identified overlapping enriched pathways, 30 unique metabolites were significantly associated with retrieved oocyte count (FDR < 20 %), including valine, uridine, uracil, tryptophan, thymine, thymidine, succinate, serine, phenylalanine, pantothenic acid, pantothenate, N-acetylserotonin, N-acetylglutamate, L-alanine, kynurenine, kynurenic acid, kynurenate, isoleucine, indolelactic acid, indole-3-acetate, histidine, cytosine, creatine, creatinine, citrulline, aspartate, asparagine, 5-hydroxyindoleacetate, 4-acetamidobutanoate, and 1-aminocyclopropanecarboxylate. An additional eight unique metabolites were associated with the number of mature or fertilized oocytes, including tyrosine, phenylpyruvic acid, hippurate, glycerate, glutamate, cysteate, aminoisobutanoate, and nicotinamide ( Table S4 ). Two metabolites were significantly associated with clinical pregnancy: cysteate and 4-acetamidobutanoate. Many other endogenous metabolites not related to the significantly enriched pathways were also associated with ovarian stimulation outcomes but are not listed here ( Table S4 ).

In this integrated exposome–metabolome study of follicular fluid from a diverse human cohort undergoing assisted reproduction, we discovered complex mixtures of known and unidentified environmental chemicals that reach the ovarian microenvironment. Using statistical mixture methods, we found statistically significant cumulative effects of the untargeted chemical mixtures on ovarian response to controlled hormone stimulation, with the mixture effects being larger than any single chemical effect alone. A wide array of potential metabolic pathways was identified as underlying the mixture effects on oocyte count and have mechanistic evidence of their role in ovarian function, demonstrating the power of untargeted discovery-based approaches to uncover novel exposures and biological mechanisms in reproductive biofluid that would have been missed if focusing only on traditional targeted chemicals.

In this pilot study, we assessed the untargeted chemical exposome and metabolome in follicular fluid collected from 82 patients undergoing oocyte retrieval following ovarian stimulation. We identified several complex mixtures of known and unidentified environmental chemicals in follicular fluid that were cumulatively associated with lower numbers of retrieved oocytes as well as several metabolic pathways potentially underlying these associations. Among the chemicals with known identities, 44 were detected in all follicular fluid samples. This indicates universal exposure within the ovarian fluid that houses oocytes by endocrine-disrupting chemicals that can readily cross the blood-follicle barrier. The contaminants included chemical classes previously found in follicular fluid at high detection rates, such as PFAS ( Bellavia et al., 2023 ; Björvang et al., 2022 ; Hallberg et al., 2021 ; McCoy et al., 2017 ; Petro et al., 2014 ), phthalates ( Beck et al., 2024 ; Bellavia et al., 2023 ; Tian et al., 2023 ), polycyclic aromatic hydrocarbons (PAHs) ( Neal et al., 2008 ), organophosphate esters (OPEs) ( Li et al., 2024a ; Yao et al., 2024 ), and certain organochlorine pesticides ( Zhu et al., 2015 ), as well as carbamate, pyrethroid, and other pesticides not yet reported in the literature. The presence of reproductive-toxic chemicals in the oocyte’s immediate microenvironment is concerning given that the follicle is the key functional unit of the ovary and highly sensitive to disruptions with consequences on oocyte competence, early embryo development, and chance of pregnancy ( Da Broi et al., 2018 ; Yu et al., 2022 ). In addition to known chemicals, we detected about 70,000 other signals in the samples using both GC- and LC-HRMS, which is more than has been detected in follicular fluid previously with LC-HRMS alone ( Hallberg et al., 2021 ; Hood et al., 2022 ). Although a portion of this mixture includes endogenous metabolites and adducts or isotopes, there were at least several hundred common chemical features each that could be annotated as other possible PFAS, pesticides, and plastic-related chemicals, suggesting that a wide array of understudied environmental chemicals can infiltrate the ovarian follicle. Many contaminants likely remain unidentified; for example, 13,274 commonly detected LC features and 13,643 commonly detected GC features could not be annotated but may comprise important emerging environmental exposures yet to be known, especially considering GC analysis is optimized for exogenous chemicals ( Zhang et al., 2021 ). In mixture models, we found significant cumulative effects of the environmental chemical mixtures on the total number of oocytes retrieved following controlled ovarian hormone stimulation. In combining the effects of hundreds of chemicals detected in the ovarian follicles, the effect sizes were much larger than that of any single chemical alone. For example, the three main annotated mixtures containing up to 587 chemicals each were associated with a 21–27 % lower expected oocyte yield per SD increase in the index, whereas the small confirmed chemical mixture was associated with 12 % lower oocyte yield per SD, and the important untargeted chemical contributors individually were only associated with 10–21 % lower oocyte yield per SD increase in univariate regression models. Our results highlight the importance of evaluating cumulative risk, as assessing only one chemical at a time may underestimate the true level of risk, because in reality many chemicals affect the same health endpoint ( Kortenkamp, 2014 ). Other studies similarly employed mixture models and, although our effect sizes are not directly comparable, they found that a mixture of 22 EDCs in follicular fluid was significantly associated with lower oocyte yield among 188 women ( Li et al., 2024a ) and that a mixture of 21 EDCs in follicular fluid was significantly associated with higher odds of having a diagnosis of diminished ovarian reserve in a case-control study of 150 women, where the overall mixture effect was larger than the effect of any chemical subgroup ( Tian et al., 2023 ). Oocyte yield is a relevant marker of ovarian reserve that infertility treatment protocols strive to maximize because of the prognosis for successful live birth among those undergoing IVF. For example, a UK-based study of 400,000 fresh IVF cycles, with male-factor infertility as the most common reason (56 %), found a very strong non-linear, positive association between the number of oocytes retrieved and the live birth rate, with the best chances observed at ~15 oocytes ( Sunkara et al., 2011 ). Because the antral phases of follicle development are dependent on gonadotropin hormones, the ovarian response to gonadotropin stimulation during ART treatment reflects a fundamental reproductive function ( Edson et al., 2009 ; McCoy et al., 2017 ). Although our sample size of patients undergoing embryo transfer after egg retrieval was too small to evaluate pregnancy success in mixture models, our regression analyses did suggest that about 100 features representing either endogenous metabolites or environmental chemicals were nominally associated with lower odds of achieving clinical pregnancy, and others with higher odds. In fact, there was suggestive evidence that the confirmed chemicals triphenyl phosphate (TPHP), bisphenol S (BPS), and pesticides deltamethrin, lenacil, and propoxur in follicular fluid were associated with lower odds of pregnancy in our single-chemical models. In addition, of the chemical features with important contributions to the cumulative mixture effects on oocyte yield in our study, we confirmed the identities of four chemicals: butyl benzyl phthalate, triethyl phosphate, benzo[a] pyrene, and propoxur. There is supporting evidence of the ovotoxicity of these chemicals in the literature. Based on a recent scoping review, BPS, a plasticizer alternative to bisphenol A (BPA) used in food contact materials and other plastic products, may impair follicle count, follicle morphology, oocyte maturation, and ovarian reserve, which is of special concern given that BPS has higher oral availability and potential for exposure than BPA ( Peters et al., 2024 ; Zhang et al., 2024a ). BBP is a phthalate plasticizer used in vinyl flooring, other polyvinyl chloride (PVC) materials, food packaging, gloves, personal care products, paints, and adhesives ( Cao, 2010 ; Just et al., 2015 ; Wittassek et al., 2011 ). A previous epidemiologic study found that the primary BBP metabolite (monobenzyl phthalate) in follicular fluid collected from 641 women was significantly associated with lower numbers of retrieved, mature, and fertilized oocytes ( Yao et al., 2023 ). Another ART study ( n = 188) detected BBP’s secondary metabolite (monobutyl phthalate) in follicular fluid and found it to significantly contribute to individual and mixture effects on all outcome stages of retrieved oocytes, mature oocytes, fertilized oocytes, and high-quality embryos ( Li et al., 2024a ). TPHP and TEP fall within the class of OPEs, which are used as flame retardants or plasticizers and have some evidence of disrupting follicular development in animals ( Wang et al., 2021 ) and IVF outcomes in humans, including a TPHP metabolite associated with failed implantation and clinical pregnancy ( Carignan et al., 2017 ). Although reproductive research on TEP is especially limited, a previous study of 319 women undergoing ART found that serum levels of TEP were significantly associated with lower levels of luteinizing hormone (LH) and higher levels of follicle-stimulating hormone (FSH), which play a critical role in the female reproductive system ( Liu et al., 2023 ). B[a]P is a PAH with common exposure from tobacco smoke and air pollution due to incomplete fuel combustion ( Lim et al., 2023 ). A pilot study ( n = 36) found significantly higher levels of B[a]P in follicular fluid from smokers compared to non-smokers and from women who did not achieve pregnancy compared to those who did ( Neal et al., 2008 ). Animal studies have demonstrated its reproductive toxicity, specifically that maternal dosing of B[a]P to mice increases incidence of preterm birth and fetal death ( Zhao et al., 2024 ), decreases follicle formation, increases apoptotic cell death, and damages DNA in germ cells in the fetal ovary ( Stefansdottir et al., 2023 ), and trans -generationally reduces the finite ovarian reserve in daughters and great grand-daughters ( Lim et al., 2023 ). Propoxur is a carbamate insecticide with little known about its female reproductive toxicity. However, studies of male rats have shown that propoxur doses disrupt reproductive hormone levels in blood ( Oyewopo et al., 2022 ) and metabolomic profiles in urine ( Liang et al., 2019 , Liang et al., 2012 ), including the metabolites creatine, lactate, and succinate that we identified as significantly related to oocyte outcomes. Finally, animal studies have shown that deltamethrin exposure impairs the maturation, quality, and survival of oocytes ( Jia et al., 2019 ) and reduces the number of primary and secondary follicles and corpus luteum ( Marettova et al., 2017 ). Lenacil has even less toxicity information, which highlights the utility of our untargeted approach to identify new chemicals of concern. From the mixture model results, we also annotated numerous features as being possible PFAS, pesticides, plastic-related chemicals, and cosmetic chemicals. Many of these annotations are novel emerging chemicals not previously studied in relation to fertility and demonstrate the possible phenomenon of regrettable substitution, in which phased-out toxic chemicals are swapped for less-known chemicals within the same class ( Brase et al., 2021 ; Zimmerman and Anastas, 2015 ; Zota et al., 2014 ). A recent review paper found that common types of plastic chemicals, pesticides, and PFAS are associated with impaired ART outcomes, such as reduced ovarian reserve, oocyte yield, fertilization rate, ovarian sensitivity, or embryo quality ( Shulhai et al., 2024 ). Thus, the literature supports the plausibility of our findings on a wide array of chemical classes influencing ovarian reserve. Because of our integrated analysis of both the exposome and metabolome in follicular fluid, we were able to discover potential metabolic mechanisms of toxic chemicals directly in the ovarian microenvironment of the developing oocyte. This localized approach was important because a previous metabolome study of 135 women found that follicular fluid had a distinct metabolic profile compared to blood, with only ~1,000 features overlapping between the two biofluids and over 5,000 features unique to the follicular fluid ( Hood et al., 2022 ). We found 38 metabolic pathways that were significantly associated with at least one chemical mixture index, of which 20 were also associated with oocyte yield. This result indicates complex biological responses to chemical contamination that may underly impaired ovarian function. Interestingly, the small mixture of confirmed chemicals was only significantly associated with one metabolic pathway, which supports the importance of discovery-based untargeted approaches that reveal other exposures and mechanisms that would otherwise go unnoticed. All of the overlapping pathways between exposure and retrieved oocyte count have some evidence of a role in ovarian function, and numerous have even been identified in human studies of the follicular fluid metabolome as being significantly different in patients with infertility disorders (such as diminished ovarian reserve, PCOS, or endometriosis), including the metabolites phenylalanine, tyrosine, tryptophan, alanine, threonine, and methionine ( Fiscus et al., 2024 ; Kobayashi and Imanaka, 2024 ). Pathways of aminoacyl-tRNA biosynthesis; phenylalanine, tyrosine, and tryptophan biosynthesis; pantothenate and CoA biosynthesis; tryptophan metabolism; and uric acid metabolism have also been shown to differ between patients with diminished versus normal ovarian reserve ( Li et al., 2023 ). The former two pathways have differentiated infertile patients with endometriosis as well ( Wei et al., 2023 ). Tryptophan and its decomposition products, in particular, may be harmful to the quantity and quality of oocytes and the early development of embryos ( Li et al., 2023 ). The mechanism of tryptophan is hypothesized to be related to its role in serotonin and melatonin ( Li et al., 2023 ), however, inflammation can cause tryptophan metabolism to switch from synthesizing serotonin to synthesizing kynurenine ( Myint, 2012 ; Smith et al., 2020 ). Kynurenine has been shown to impair hormone secretion, ovulation, corpus luteum formation, and reproductive function in animals ( Shen et al., 2023 ) and to be associated with PCOS and reproductive hormone levels in humans ( Wang et al., 2022 ). Essential amino acids, including tryptophan, phenylalanine, and methionine, cannot be synthesized in the body, so their changes could reflect disruption in amino acid catabolism and thus energy for the developing oocyte ( Fiscus et al., 2024 ) or dysregulation of pH, osmolality, and nucleotide synthesis during oocyte development ( Seli et al., 2014 ). Biopterin metabolism, an enriched pathway in our study, is important for metabolism of phenylalanine, tyrosine, and tryptophan. Biopterin metabolism, tryptophan metabolism, and tyrosine metabolism were all significant pathways in human blood between urinary phthalate exposure and live birth following IVF ( Hood et al., 2024 ), and in the same cohort, tyrosine metabolism in human follicular fluid was significantly associated with mature oocyte count and biopterin metabolism in serum with peak estradiol levels ( Hood et al., 2023 ). Another pilot study found that essential amino acid metabolite methionine in follicular fluid was significantly correlated with the number of retrieved oocytes, the number of cleaved embryos, and maternal age ( Huang et al., 2022 ). In vivo studies demonstrated that methionine administered to rats can have beneficial effects on follicle growth and ovarian estrogen synthesis at low doses but harmful effects on the number of primordial follicles or the embryonic development at higher doses ( Nazem et al., 2022 ; Yang et al., 2024 ), indicating that this amino acid does play an important role in the ovaries. We also found significantly enriched pathways related to non-essential amino acids, such as glycine, serine, and alanine, which are rich in the reproductive tract. Alanine is important in pH regulation and generates nitric oxide, which is critical in embryo development and establishment of pregnancy ( Seli et al., 2014 ). Levels of serine in blood were significantly different in pregnant versus non-pregnant patients with PCOS ( Zhang et al., 2014 ), and levels of serine, serine-synthesizing enzymes, and glycine levels in follicular cumulus cells were significantly different in young patients with diminished versus normal ovarian reserve ( Lu et al., 2023 ). Serine is converted to glycine for one-carbon metabolism, which can increase anti-inflammatory cytokines while reducing pro-inflammatory cytokines and thus alter steroidogenesis, follicular maturation, and ovulation ( Zhang et al., 2014 ). Follicular cells have developmental stage-dependent expression of amino acid transporters that bring glycine, alanine, and other amino acids into oocytes and support higher oxidative metabolism at specific times of oocyte maturation ( Gao, 2020 ; Gu et al., 2014 ). In animal oocytes, glycine supplementation in in vitro culture medium increased nuclear maturation rate and blastocyst formation rates, decreased apoptosis and reactive oxygen species ( Li et al., 2018 ), and improved mature oocyte quality and embryo development ( Cao et al., 2016 ). During mouse oocyte growth, glycine, serine, and threonine metabolism — as well as nicotinate and nicotinamide metabolism — was significantly enriched comparing P10 oocytes (secondary follicles) to later P15 oocytes (early antral follicles with meiotic competence) ( Zhang et al., 2024b ). Nicotinamide is a form of niacin (vitamin B3, which can also be synthesized from tryptophan) and precursor to NAD+, a coenzyme with potential to treat ovarian aging, PCOS, and other gynecological diseases ( Li et al., 2024b ). In humans, nicotinamide levels were higher in follicular fluid from large follicles than small follicles from the same patient and were significantly correlated with oocyte maturation rate and fertilization rate ( Guo et al., 2022 ), and vitamin B3 metabolism in follicular fluid was significantly associated with peak estradiol levels in a metabolome study ( Hood et al., 2023 ). In animals, it was shown that the reduction in oxidative stress in oocytes by nicotinamide or niacin supplementation played a role in protecting oocyte quality ( Almubarak et al., 2021 ; Guo et al., 2022 ; Min et al., 2021 ). Our identified pathway of another vitamin, pantothenate (B5), is essential for CoA, which plays a role in metabolism of sugar, fat, and protein as well as antioxidation, and uric acid (related to urea cycle metabolism) can serve as an antioxidant; oxidative stress may impair oocyte growth ( Li et al., 2023 ) or embryo development ( Seli et al., 2014 ). Urea cycle/amino group metabolism in human follicular fluid is supported by another study that found significant associations with mature oocyte count ( Hood et al., 2023 ). Pyrimidine is a major energy carrier and a subunit of nucleic acids, which likely facilitates the high amount of DNA and RNA synthesis needed during early embryo development ( Li et al., 2020 ). In cells and fluid from sheep oocytes at different stages, pyrimidine metabolism was upregulated during oocyte maturation, from MII oocyte to 2-cell embryo, and from 8-cell embryo to blastocyst ( Pan et al., 2024a ). Phenylacetate is a metabolite of phenylalanine, and a study found that a urinary pathway of phenylacetate-related metabolism (as well as a tryptophan pathway) was significantly different for humans experiencing pregnancy loss in the first trimester compared to controls with normal gestation at that stage ( Li et al., 2021 ). Genes related to selenocompound metabolism were differentially expressed in human cumulus cells from high-quality oocytes compared to low-quality oocytes ( Liu et al., 2018 ). Selenium and selenoproteins may act as antioxidants and regulators of granulosa cell growth, with some evidence of selenium’s ability to improve IVF outcomes ( Qazi et al., 2018 ). Furthermore, the pathway of xenobiotic metabolism by cytochrome p450 may implicate reactive oxygen species from chemical bioactivation in redox imbalance in the oocyte ( Sobinoff et al., 2010 ). Mineral absorption is also relevant because of the delicate role of minerals like selenium, zinc, copper, iron, calcium, and sodium in reproductive hormone levels, oocyte quality, antral follicle count, ovarian hyperstimulation, embryo development, implantation rate, and live birth rate; mechanisms of oxidative stress and hormonal regulation may contribute to these risks ( Kapper et al., 2024 ). Finally, although not an overlapping pathway with oocyte count, the significant enrichment of C21 steroid hormone biosynthesis with chemical mixture exposure is of particular interest to fertility, as this class of hormones includes pregnenolone and progesterone and comprises the starting materials for other reproductive hormones such as estrogen and testosterone ( Ghayee and Auchus, 2007 ). For many of the overlapping pathways, we were able to leverage identifications from authentic reference standards to validate the relevance of confirmed endogenous metabolites that are involved in those pathways and individually associated with oocyte outcomes, which supports our confidence in these pathways (summarized in Fig. 6 ). Overall, the pathway enrichment results showed that there is an array of different biological perturbations that may underly follicular and oocyte development, and some of these pathways potentially share common mechanisms such as oxidative stress, energy production, minerals, and B vitamin metabolism. Our findings also demonstrate the utility and novelty of follicular fluid as a localized biofluid to both capture environmental chemicals that reach the ovaries and to assay direct metabolomic response profiles during controlled reproduction. While our study provides important insights on novel chemical exposures in follicular fluid and biological pathways related to ovarian response, there are several limitations to note. First, our analytical sample size was only 82 patients, so our results should only be considered preliminary and hypothesis-generating. Second, the generalizability of our results to the general reproductive population is unclear given that our patients were all women undergoing oocyte retrieval for fertility preservation or infertility treatment. Additionally, fertility patients tend to be older and of higher socioeconomic status. However, our cohort was more racially diverse than most previous fertility cohorts, and a large portion of the couples were seeking treatment for male-factor infertility. Third, we relied on medical records and were not able to survey the participants about other environmental, socioeconomic, or psychosocial factors, which limited our ability to adjust for confounders. Sample size also limited the number of covariates we could reasonably adjust for in models, including BMI. However, further adjustment for BMI in sensitivity tests only negligibly affected associations between confirmed chemicals and outcomes (slightly higher p values). Fourth, our study design was cross-sectional with follicular fluid collected at the same time as oocyte retrieval, however, chemical exposure was blinded to the outcome, and the environmental chemicals in follicular fluid were likely transferred from blood which can reflect aggregated exposure over months to years for some chemicals with long half-lives ( Barr et al., 2005 ). Moreover, we prospectively evaluated the outcome of clinical pregnancy at six to eight weeks in secondary analyses. Fifth, number of oocytes retrieved is only a proxy for more clinically meaningful fertility endpoints such as clinical pregnancy and live birth. While our analyses generally assumed that a higher number of oocytes retrieved was a positive monotonic outcome, too many oocytes can indicate reproductive pathologies of polycystic ovary syndrome or ovarian hyperstimulation ( Sunkara et al., 2011 ), and ovarian reserve is limited in predicting underlying fecundity and/or oocyte quality. Our use of generalized linear models for the single-chemical associations also did not allow for the possibility of non-linear effects of the chemicals on oocyte count, which could partly explain the opposing directions of effect seen for some exposures. In addition, the outcome of clinical pregnancy had a smaller sample size (since not all patients underwent embryo transfer) and is impacted by both maternal and paternal factors, which could partly explain why we found fewer associations between exposures and clinical pregnancy than for oocyte count. Sixth, a major challenge with untargeted data is the uncertainties in identification of chemical signals. While we had many annotations for the LC features based on matching to mass spectral databases, these are not confirmations of identities ( Krahl et al., 2019 ; Pourchet et al., 2020 ; Uppal et al., 2017 ). Most GC features remained unannotated because the large external GC–MS databases acquired spectra from unit mass instruments (and are thus prone to false positives) while only small databases are available to match based on accurate mass ( Koelmel et al., 2022 ). However, we were still able to statistically analyze the associations of unidentified chemicals with IVF outcomes and can retroactively identify the features in the future as more chemical reference standards become available. In addition, we confirmed the identities of over 600 detected LC and GC features using standards run on our instruments. Another limitation with the untargeted data is that we only have peak intensities (relative abundances) rather than absolute concentrations of the chemicals, which challenges comparison to other external cohorts. Seventh, the follicular fluid samples can contain variable amounts of residual flushing media, which could dilute chemical abundances for retrievals from patients that required more flushing media than others; in future work, we will measure the volumes of flushing media used and will analyze samples of flushing media used at the time of collection to mon any unintentional chemical contamination. Finally, the limitations with our WQS mixture models include that they did not account for interactive or non-linear effects among chemicals and that we assumed unidirectional effects on oocyte yield. In the models, we also excluded potential unannotated environmental chemicals from the LC data to minimize the unintentional inclusion of endogenous metabolites, although this decision should only underestimate the mixture effects, and our selected GC mixtures were inclusive of unknown chemicals. We further focused on chemicals with nominal univariate associations, which could have missed chemicals with effects that are not observable on their own. Despite these limitations, our study had several key strengths and novelties. We were able to directly measure chemicals and metabolic perturbations within a target reproductive biospecimen that is representative of the oocyte microenvironment. We applied the latest advanced untargeted HRMS technology to comprehensively measure the integrated exposome and metabolome through a systems biology framework. This included combining both GC and LC analysis, which has been greatly underutilized in studies and expands the coverage of detectable environmental chemicals with varied physical–chemical properties ( Manz et al., 2023 ). The untargeted approach also supports hypothesis discovery to identify potentially emerging environmental risk factors that could not be anticipated in advance. Furthermore, we applied the latest statistical mixture methods to high-dimensional untargeted exposome data and extended it into metabolic pathway enrichment analysis for the first time. This statistical workflow offers an innovative way to evaluate cumulative chemical mixture effects at high dimensions while minimizing bias and to explore biological mechanisms underlying reproductive outcomes, which has the potential to inform prevention of infertility and improvement of pregnancy outcomes. Finally, this pilot study was a sub-cohort from the Emory Reproductive Center in Atlanta, which comprised a more racially/ethnically diverse population than most previous ART cohorts ( Chan et al., 2021 ; James-Todd et al., 2016 ). This EMPOWR biorepository supports follow-up research to leverage larger sample sizes, investigate health disparities, and integrate with other multi-omics such as the epigenome.

An estimated 17 % of people experience infertility for a year or more during their lifetime ( Cox et al., 2022 ). Both animal and in vitro studies have implicated endocrine-disrupting chemicals (EDCs) as reproductive toxicants that target the ovary, which is the major female reproductive organ and vulnerable to interference from chemical exposures at key steps during reproduction ( Craig et al., 2011 ). For example, EDCs can impair the development, functionality, and number of ovarian follicles ( Palioura and Diamanti-Kandarakis, 2015 ; Patel et al., 2015 ) as well as the signaling, availability, and production of ovarian and sex hormones ( Craig et al., 2011 ; Ding et al., 2020 ; Mlynarcikova et al., 2014 ; Uzumcu and Zachow, 2007 ). Epidemiologic research has linked human exposure to EDCs with failures in oocyte fertilization, embryo implantation, clinical pregnancy, and live birth ( Carignan et al., 2017 ; Meeker et al., 2011 ; Messerlian et al., 2016 ) and with increased risk of polycystic ovary syndrome (PCOS) and diminished ovarian reserve (DOR) ( Hammarstrand et al., 2021 ; Jin et al., 2019 ; Konieczna et al., 2018 ; Palioura and Diamanti-Kandarakis, 2015 ; Tian et al., 2023 ). The effects on reproductive health are relevant for more than just fertility, as certain ovarian toxicants may also lead to pubertal acceleration or delay ( Kiess et al., 2021 ), irregular menstrual cycles and earlier menopause ( Ding et al., 2022 ), painful gynecological diseases ( Dutta et al., 2023 ), and risk of gynecological cancers ( Liu, 2021 ). Cohorts of patients undergoing assisted reproduction technology (ART) have provided valuable data about the effects of environmental chemicals on reproductive health because the key events of reproduction are precisely controlled, timed, and observable at an early stage ( Hwang et al., 2022 ). ART also offers the opportunity to collect biospecimens from target organs that would otherwise be too invasive to collect in people trying to conceive without medical assistance. Follicular fluid, the liquid that surrounds the oocyte and fills the follicular antrum in an ovarian follicle, is routinely collected during oocyte retrieval as part of in vitro fertilization (IVF) and represents the localized ovarian environment that oocytes are directly exposed to within the reproductive organ ( Björvang et al., 2022 ; Tian et al., 2023 ). The fluid is formed from the transfer of blood plasma components into the follicle and from secretions by theca and granulosa cells within the ovarian follicle ( Pan et al., 2024b ). The rich mixture consists of steroid hormones, gonadotropin hormones, growth factors, cytokines, extracellular vesicles, proteins, and lipid energy sources that are critical for follicle selection, follicle growth, oocyte maturation, ovulation, and fertilization ( Pan et al., 2024b ). The ovarian follicle is a highly fragile microenvironment sensitive to disruption in these processes ( Petro et al., 2012 ), which can have subsequent consequences on embryo survival and pregnancy ( Hallberg et al., 2021 ). Thus, follicular fluid is a more toxicologically relevant target biofluid for the study of reproductive pathologies compared to traditional peripheral blood samples ( Hauser and Mínguez-Alarcón, 2023 ). Of concern, EDCs can cross the blood-follicle barrier and contaminate the follicular fluid surrounding an oocyte. Previous studies have detected a variety of targeted chemicals directly in follicular fluid samples from ART cohorts, including phthalates, parabens, phenols, flame retardants, pesticides, and per- and polyfluoroalkyl substances (PFAS) ( Al-Hussaini et al., 2018 ; Beck et al., 2024 ; Bellavia et al., 2023 ; Björvang et al., 2022 ; Ikezuki et al., 2002 ; Johnson et al., 2012 ; Li et al., 2024a ; Meeker et al., 2009 ; Petro et al., 2012 ; Tian et al., 2023 ). Previous studies have also found associations of certain of these chemicals, including PFAS, with ART outcomes such as lower embryo quality ( Björvang et al., 2022 ; Li et al., 2024a ; Liu et al., 2024 ; Petro et al., 2012 ), diminished ovarian reserve or follicle counts ( McCoy et al., 2017 ; Tian et al., 2023 ), reduced ovarian sensitivity to gonadotropin stimulation ( Bellavia et al., 2023 ; Björvang et al., 2022 ), decreased fertilization rate ( Al-Hussaini et al., 2018 ; Petro et al., 2012 ), or lower odds of pregnancy ( Al-Hussaini et al., 2018 ; Johnson et al., 2012 ) and clinical pregnancy ( Beck et al., 2024 ). However, these studies measured at most 64 chemicals using targeted methods. Only one published study, to our knowledge, has analyzed environmental chemicals in follicular fluid through an untargeted approach. Using high-resolution mass spectrometry (HRMS) with liquid chromatography (LC), the authors found that the follicular chemical environment is distinct from blood and is a location of accumulation for >200 chemical signals ( Hallberg et al., 2021 ). Furthermore, previous metabolomic work from our group found that the follicular fluid has a unique biology compared to blood, with thousands more endogenous metabolites identified in the follicular fluid and only weak-to-moderate correlation between overlapping features in the follicular fluid and serum ( Hood et al., 2022 ). Follicular fluid represents a highly relevant and novel biofluid that is both important to successful reproduction and susceptible to the accumulation of reproductive-toxic chemicals. With recent technological advances in systematically measuring environmental chemicals at an untargeted –omics scale, we can study the reproductive impacts of not only a small number of known chemicals but also discover a larger number of previously unknown toxicants without having to choose analytes in advance ( David et al., 2021 ; Vermeulen et al., 2020 ). This comprises a key component of the “exposome,” the discovery-based study of the comprehensive and cumulative environmental influences on biological systems and health ( Miller, 2024 ; Miller and Jones, 2014 ). For example, untargeted HRMS can now, in a single analytical run, detect over 100,000 signals of both exogenous environmental chemicals (i.e., the exposome) and endogenous metabolites (i.e., the metabolome) in biospecimens ( Balcells et al., 2024 ; Walker et al., 2019 ). However, very few studies have performed untargeted HRMS on follicular fluid to evaluate environmental exposures, and none have coupled it with both LC and gas chromatography (GC) to expand coverage and characterization of low-level environmental chemicals with distinct properties from endogenous metabolites ( Manz et al., 2023 ; Zhang et al., 2021 ). Considering that nearly 70,000 new chemicals were registered for commerce globally in the last decade ( Wang et al., 2020 ) and over 16,000 chemicals are involved in plastic production alone ( Wagner et al., 2024 ), discovery-based chemical analysis provides a critical opportunity to determine the potential reproductive toxicities and health impacts of emerging chemical exposures. With simultaneous measurement of the metabolome, we can also integrate the exposome with the biological mechanisms that may underly the toxicity ( Niedzwiecki et al., 2019 ). In the current study, we leveraged a pilot study of 84 patients undergoing egg retrieval with follicular fluid samples collected from 2018 to 2020 at the Emory Reproductive Center in Atlanta, GA. Our objectives were to: 1) measure untargeted chemicals and endogenous metabolites in follicular fluid using both GC- and LC-HRMS, 2) determine the proximate impacts of the environmental chemical mixtures on IVF outcomes, including oocyte count after controlled ovarian stimulation, and 3) evaluate perturbations in metabolic pathways in the follicular fluid that potentially underly the associations between chemical exposure and oocyte count.

Appendix A. Supplementary data Supplementary data to this article can be found online at https://doi.org/10.1016/j.envint.2025.109787 .