Exposure to per- and poly-fluoroalkyl substances in association to later occurrence of type 2 diabetes and metabolic pathway dysregulation in a multiethnic US population

Summary

Background

Growing evidence suggests that exposure to per- and polyfluoroalkyl substances (PFAS) are linked to an increased risk of type 2 diabetes (T2D); however, the effect of PFAS mixtures and underlying mechanisms are not well understood. We examined the associations between exposure to PFAS mixture with later T2D diagnosis and underlying metabolic dysregulations.

Methods

We conducted a nested case–control study within BioMe, an electronic health record-linked biobank of >65,000 patients seeking primary care at Mount Sinai Hospital, New York, since 2007. After excluding prevalent T2D cases at baseline, we selected 180 incident T2D cases (33% African Americans, 33% Hispanics, 33% Whites) and 180 age, sex, and ancestry-matched T2D-free controls. In prediagnostic plasma collected at baseline (∼6 years before diagnosis), we quantified seven PFAS and untargeted metabolomic profiles. We used Weighted Quantile Sum regression to evaluate the PFAS mixture association with the odds for incident T2D. We analysed the associations between ∼650 annotated metabolites and the PFAS mixture or T2D odds using Hierarchical Bayesian Weighted Quantile Sum and logistic regression, respectively, adjusting for matching factors and other confounders. Pathway enrichment analyses were performed using Mummichog.

Findings

Each tertile increase in the PFAS mixture was associated with higher odds of incident T2D (OR [95% CI] = 1.31 [1.01, 1.70]), with Perfluorooctane Sulfonate (PFOS) having the highest contribution to this association. Metabolites associated with both the PFAS mixture and T2D odds were 5-hydroxytryptophan, glucoheptulose, and sulfolithocholylglycine; the associations with sulfolithocholylglycine survived multiple testing corrections. Pathways associated with both the PFAS mixture and T2D were glutamate metabolism, arginine and proline metabolism, and drug metabolism—cytochrome p450.

Interpretation

Exposure to PFAS mixtures may be associated with increased odds for T2D in multiethnic populations via dysregulations in amino acid and drug metabolism. Larger investigations in multiethnic populations are required to elucidate the potential PFAS contribution to metabolic alterations and T2D risk.

Funding

National Institutes of Health (R01ES033688, P30ES023515, R21ES035148, R35ES030435, R01ES032242, R01ES034521, R01ES029944, R01ES030364, U01HG013288, R21ES037112 and P30ES007048).

Research in context.

Evidence before this study

Growing evidence suggests that exposure to per- and poly-fluoroalkyl substances (PFAS), among other endocrine-disrupting chemicals (EDCs), may increase the risk of Type 2 diabetes (T2D). However, studies linking prior PFAS exposure to future T2D diagnosis are limited. Furthermore, the biological mechanisms underlying the potential diabetogenic effects of PFAS exposures are not well understood in humans, particularly in multiethnic populations.

Added value of this study

In this nested case–control study of 180 incident T2D cases and 180 matched controls within the multiethnic BioMe electronic health record-linked Biobank in New York (U.S.), we found that each tertile increase in the PFAS mixture was associated with a 31% increase in the odds of incident T2D, and that these associations could be due to dysregulations in amino acid biosynthesis and drug metabolism.

Implications of all the available evidence

Our findings support that PFAS mixture exposures may increase the risk for T2D through dysregulations in pathways of amino acid biosynthesis and drug metabolism. Findings from this study underscore the importance of preventing PFAS exposures to promote public health and advance knowledge about potential mechanisms underlying the PFAS impacts on human metabolism in multiethnic populations.

Introduction

Type 2 diabetes (T2D) prevalence has risen globally since 1980. In 2019, nearly 34 million adults had T2D, and approximately 88 million adults had prediabetes in the US.^1,2 Compared to non-Hispanic Whites, Blacks and Hispanics in the US are at higher risk for T2D and its long-term complications, such as heart disease, renal failure, neuropathy, macrovascular and microvascular disease, and eye and hearing damage.^3,4 Growing evidence suggests that exposure to per- and poly-fluoroalkyl substances (PFAS), among other endocrine-disrupting chemicals (EDCs), may promote the development of T2D, likely in synergy with other known risk factors, such as poor diet, physical inactivity, and genetics.4, 5, 6, 7 However, epidemiological evidence linking PFAS and T2D has so far been inconclusive.⁸

PFAS are a broad class of persistent chemicals detected in the blood of over 98% of people in the US and elsewhere.^5,9,10 For decades, PFAS have been used as industrial surfactants in textile coatings, fire-fighting foams, and several consumer products.^11,12 Previous studies have shown that PFAS exposures are associated with higher risks for T2D13, 14, 15, 16, 17, 18 and with other adverse metabolic outcomes across the lifespan, such as obesity^19,20 and dysregulations of adipocytokines,²¹ lipids,^17,22 glucose,^17,23 and liver enzymes.^11,24 However, findings linking PFAS with T2D are inconclusive,²⁵ and mostly come from European- and Asian-ancestry populations.^23,26,27 Moreover, biological mechanisms underlying the potential diabetogenic effects of PFAS exposures are not yet well understood in humans. In a recent systematic review, we found that leveraging high-resolution metabolomics in population studies is a promising tool to determine metabolic dysregulations induced by PFAS exposures.²⁸ However, investigations integrating PFAS, T2D, and metabolomics have so far been scarce,^29,30 and there is a lack of data from multiethnic populations that may be more broadly generalisable to the general US population.

Therefore, we conducted an untargeted metabolome-wide investigation within the BioMe biobank cohort in New York City to characterise the associations of exposures to PFAS as a mixture with incident T2D and identify dysregulated metabolic pathways that may underlie the PFAS-T2D associations.

Methods

Study population and design

We conducted a nested case–control study of 180 incident T2D cases and 180 controls within BioMe, an ongoing and broadly consented Electronic Health Record (EHR)-linked biobank, with more than 65,000 adult participants enrolled non-selectively from the Mount Sinai Hospital patient population.³¹ BioMe enrolment was initiated in 2007 and is still ongoing. BioMe participants represent a broad multi-ancestry and socioeconomic diversity, characteristic of the communities served by Mount Sinai Hospital in New York City. BioMe participants were predominantly recruited from outpatient primary care practices and consented at enrolment to be followed throughout their clinical care (past, present, and future clinical visits). At enrolment (study baseline), participants provided a blood sample and information on sociodemographic and lifestyle factors through interview-based questionnaires with a specially trained nurse. The BioMe Biobank is linked to Mount Sinai's system-wide Epic EHR,³² which captures a full spectrum of biomedical phenotypes, including clinical outcomes, covariates, and exposure data from past, present, and future healthcare encounters. BioMe records are updated from Epic weekly using standardised methods. For the present study, we implemented a nested case–control design in the BioMe cohort enrolled as of March 31, 2020 (n = 53,790) (Fig. 1). From all initially enrolled participants, we excluded those with (1) uncertain T2D cases or control status at baseline, i.e., at the time of blood draw (n = 4135), (2) prevalent T2D cases at baseline (n = 9471), and (3) T2D cases diagnosed within one year after the baseline (n = 351) to minimise the likelihood of reverse causation bias. The remaining eligible incident T2D cases diagnosed at least one year after baseline (n = 779) were matched to one eligible T2D-free control by age at diagnosis (control can be of the same age or older), sex, and self-reported ancestry. Our study included cases enrolled prior to 2017 because we limited selection to only incident T2D, and because prevalent cases and cases diagnosed in less than year of follow up were excluded. The average time from baseline to T2D diagnosis within cases was (median [IQR]) 6 [1, 10] years. Out of the 779 eligible matched incident T2D case–control pairs, we selected 180 case–control pairs for the present PFAS and metabolomics analysis based on power calculations (>80% power at alpha level 5%) at the time of study design and considering budget availability, using a stratified (by ancestry random) sampling approach (n = 60 African American pairs, 60 = Hispanic American pairs, and 60 = European American pairs, see Table 1). Our initial sample size calculation with 360 samples yielded over 80% power at alpha level 5% to detect a significant odds ratio with the PFAS mixture index and T2D diagnosis.

Fig. 1.

Flowchart detailing the exclusion and inclusion criteria for the samples in this study. #: one eligible T2D-free control matched to a case of similar age (control can be of the same age or older by six months).

Table 1.

Comparison of main characteristics between T2D cases and controls.

Characteristic	Unit/Category	Incident T2D Cases (n = 179)	Controls (n = 180)	p-value
		Mean (SD) or n (%)	Mean (SD) or n (%)
Age at case diagnosis/matching	In years	60 (12)	60 (12)	0.96
Time interval between T2D diagnosis and baseline enrolment	In years	4.2 (2.5)	n/a
Self-reported ancestry	European American	60 (33.5%)	60 (33.3%)	0.99
	African American	60 (33.5%)	60 (33.3%)
	Hispanic American	59 (33.0%)	60 (33.3%)
Sex	Male	72 (40.2%)	73 (40.6%)	0.95
	Female	107 (59.8%)	107 (59.4%)
Calendar year at blood draw (baseline)		2010 (2)	2011 (2)	8 $\times$ 10−4
Age at baseline	In years	56 (12)	58 (13)	0.001
BMI at baseline	Kg/m2	31.4 (7.6)	28.0 (6.2)	5.5 $\times$ 10−6
Smoking Status at baseline	Yes	40 (22.3%)	23 (12.8%)	0.017
	No	139 (77.7%)	157 (87.2%)
Plasma PFAS concentrations (ng/mL)	% > LOD (=0.01 ng/mL)	GM (95% CI)	GM (95% CI)
PFOS	100.0%	8.90 (7.72, 10.26)	9.41 (8.25, 10.74)	0.71
PFHxS	99.4%	0.77 (0.66, 0.91)	0.83 (0.72, 0.96)	0.82
PFNA	99.7%	1.08 (0.97, 1.20)	1.11 (0.99, 1.24)	0.78
PFOA	100.0%	2.89 (2.51, 3.33)	2.84 (2.51, 3.22)	0.69
PFHpA	41.1%	0.02 (0.02, 0.02)	0.02 (0.01, 0.02)	0.66
PFDA	93.0%	0.27 (0.23, 0.33)	0.29 (0.24, 0.35)	0.21
PFHpS	89.4%	0.07 (0.06, 0.09)	0.08 (0.07, 0.09)	0.91

PFDA: Perfluorodecanoic acid; PFHpA: Perfluoroheptanoic acid; PFHxS: Perfluorohexanesulfonic acid; PFHpS: Perfluoroheptanesulfonic acid; PFOS: Perfluorooctanesulfoninc acid; PFNA: Perfluorononanoic acid; PFOA: Perfluorooctanoic acid; GM: Geometric mean; LOD: Limit of detection; n/a: not applicable.

Ethics

Written and informed consent was obtained from each individual in this study as a part of the BioMe Biobank. This pilot study was granted approval by the Institutional Review Board (STUDY-21-01871) at the Icahn School of Medicine at Mount Sinai in New York.

T2D classification in BioMe

T2D case and control status were defined by algorithms designed as part of the eMERGE Network³³ and customised to and fully implemented and validated in the BioMe biobank.³¹ The algorithm defined a case when a patient was diagnosed with T2D and treated with T2D medication, using the (international classification of diseases) ICD-9 and ICD-10 codes. Patients diagnosed with type 1 diabetes (T1D) and those diagnosed with both T1D and T2D were excluded and classified as uncertain status. Control status was defined as not having a diagnosis for either T2D or T1D, not being treated for T2D, and having glycosylated haemoglobin A1c<6.5% (no prediabetes). One T2D case was excluded from the analysis because PFAS measurements did not pass quality control (QC) criteria due to insufficient blood volume.

Targeted measurements of PFAS

PFAS were measured in plasma samples collected from participants at BioMe entry (baseline) using liquid chromatography with high-resolution mass spectrometry (LC-HRMS) and previously established protocols.³⁴ Briefly, plasma samples were prepared by adding 2.5 μL of containing 24 ¹³C labelled PFAS to obtain a final 5 or 25 ng/mL concentration, depending on PFAS species, to 40 μL of plasma. Proteins were then precipitated by adding 80 μL of cold acetonitrile, vortexing, and centrifuging at 18,000 g for 15 min. The resulting aliquot was diluted with 2 volumes of LC-MS grade water and stored in a refrigerated autosampler until analyses. PFAS analysis was completed by reverse phase chromatography with negative electrospray ionisation, performed with a Thermo Scientific Vanquish Flex UHPLC system with Binary Pump attached to a Thermo Scientific Q Exactive HF-X Orbitrap mass spectrometry system (Thermo Fisher Scientific, Rockford, IL, USA).

After analysis of BioMe samples, PFAS peaks, and their corresponding 13C-labelled internal standards were identified by matching mass m/z with a tolerance of 5 ppm and retention time window of 15 s and integrated using TraceFinder 5.1 (Thermo Fisher Scientific, Rockford, IL, USA). Reference standardisation was used to quantify PFAS concentrations, as described previously.^17,34, 35, 36, 37 To ensure the accuracy of PFAS measurements across batches, National Institute of Standards and Technology 1957 and 1958 were analysed in each analytic batch, and the concentrations that were measured were compared to known PFAS levels. Assay variability was evaluated using (1) replicate analysis of NIST 1957 and 1958 and (2) two separate pooled plasma samples analysed at the beginning, middle, and end of the batch; the coefficient of variation for Perfluorooctane Sulfonate (PFOS), Perfluorooctanoic acid (PFOA), Perfluorohexane Sulfonic acid (PFHxS), Perfluorononanoic acid (PFNA), Perfluorodecanoic acid (PFDA), and Perfluoroheptanesulfonic acid (PFHpS) were all less than 15%. The limit of detection (LOD) for all PFAS was 0.1 ng/mL. We substituted PFAS concentrations below LOD by the value of LOD/√2 for that respective substance.

Untargeted metabolite profiling

Untargeted, high-resolution metabolomic profiling was conducted on plasma samples using LC- Ultra High-Resolution Mass Spectrometry (UHRMS) and previously established methods.³⁸ Samples were prepared by treating plasma with acetonitrile containing ¹³C-labelled internal standards of endogenous and exogenous chemicals. Extracts were analysed using positive and negative electrospray ionisation using hydrophilic interaction liquid chromatography (HILIC) and C18 reversed-phase chromatography. Peaks were detected using a Thermo Scientific Q Exactive HF-X Hybrid Quadrupole-Orbitrap (Thermo Fisher Scientific, Waltham, MA, USA) UHRMS operated at 120,000 resolution over a mass-to-charge ratio (m/z) range of 85 to 1250.^39,40 Two separate pooled reference plasma samples were analysed at the beginning, middle, and end of each analytical batch of 70 samples for quality assurance/quality control (QA/QC) and evaluation of batch effects. Raw data files were extracted using adaptive processing of high-resolution LC/MS data (apLCMS) with modifications by xMSanalyzer, configured to enhance data quality control, and corrected for batch effects using ComBat, with detected signals uniquely identified by their m/z, retention time, and ion intensity.41, 42, 43 We annotated 907 and 748 metabolites (in positive and negative ionisation modes, respectively) by their matching accurate mass m/z, and isotopic patterns against a database of metabolites that includes the most common endogenous metabolites reported on the Metabolomics Workbench using xMSannotator, which leverages a multi-level scoring algorithm to increase metabolite annotation confidence.⁴⁴ Reported metabolites correspond to Metabolomics Standards Initiative (MSI) Level 2 annotations.⁴⁵ In association analyses, however, we included a total of 656 metabolic features (372 and 284 in positive and negative ionisation modes, respectively) that had non-zero values for at least 75% of samples. The distribution of the median (25th and 75th percentiles) coefficient of variation for the metabolic features (in positive and negative ionisation modes) are 0.84 (0.65, 1.14) and 0.82 (0.67, 1.03).

Additional covariates

All statistical models were adjusted for potential confounders that were chosen based on a priori knowledge and directed acyclic graphs (see Fig. 2 where only main associations and potential confounders of the main associations are illustrated). All statistical models were adjusted for matching factors (sex, age at diagnosis, and self-reported ancestry)⁴⁶ and additionally, for measured potential confounders at the study baseline, including body mass index (BMI), calendar year of enrolment (i.e., the date of blood draw), and smoking status. We chose not to adjust for any intermediate factors that fall within the hypothesized causal pathway (mediators) between PFAS exposures and T2D diagnosis.

Fig. 2.

Directed Acyclic Graphs (DAGs) elucidating association between PFAS exposure and Type 2 Diabetes Risk. Only main associations and potential confounders of the main associations are illustrated.

Statistics

All analyses were conducted in R (version 4.2.3). All the PFAS concentrations and signals from metabolites were converted into tertiles to ensure robustness (see PFAS distributions in Table 1). The association analyses were conducted in multiple stages: (1) To evaluate the associations between individual PFAS exposures and T2D, we used conditional logistic (c-logistic) regressions adjusted for covariates. (2) To evaluate the association of PFAS mixture with T2D status, we used the Weighted Quantile Sum (WQS) regression model⁴⁷ under the assumption of dose additivity⁴⁸ (assuming the effect of individual PFAS within the mixture is not strongly synergistic or antagonistic). This model can accommodate correlations among PFAS and summarise overall exposure to the mixture by estimating a single weighted index and considering each component's contribution to the mixture effect using estimated weights. The WQS regression involves estimating the weighted index and then using that created index in a generalised linear model to estimate the overall joint mixture effect. Creating the weighted index uses bootstrapped samples and a subset of randomly chosen exposures (to minimise the bias due to multicollinearity).⁴⁹ Once the mixture index, i.e., the weighted average of the exposures, is estimated, the overall mixture effect is calculated using a generalised linear model given the estimated weights. Because of the importance of this mixture association, we ensured the p-value and the 95% CI were robust. We, therefore, re-estimated the p-value using a permutation-based test, where the PFAS-mixture was randomly permuted 10⁵ times. (3) Next, we evaluated the PFAS associations with plasma metabolites. Since the directionality of PFAS-mixture and metabolite associations could be either positive or negative, we conducted exposure-metabolite wide association studies (ExWAS) using hierarchical Bayesian WQS (hBWQS). In PFAS-metabolites analyses, we used inverse probability weighting using logistic regression to control for potential selection bias due to the nested case–control design based on T2D status.^50,51 The propensity score model included sex, self-reported ancestry, time of blood sample collection, and T2D status. (4) Finally, we used logistic regression with robust heteroscedasticity-consistent standard errors to evaluate the associations between metabolites and odds of T2D diagnosis.^52,53 The Bonferroni procedure was adapted for multiple comparison error correction to handle correlated features. We estimated the effective number of tests for the ExWAS in each ionisation mode and PFAS exposure.54, 55, 56 The nominal significance (0.05) was divided by the effective number of tests (n = 100 and n = 72 for positive and negative modes, respectively) to obtain the adjusted significance cutoffs. The adjusted p-value thresholds were 0.005 and 0.007 for positive and negative ionisation modes. Association estimates (and 95% CIs) from all analyses are expressed per tertile increase in PFAS exposures/metabolites to facilitate comparison of the magnitude of effects across statistical methods.

Mummichog⁵⁷ was used to perform pathway enrichment analysis. The mummichog algorithm takes into account high-throughput metabolomics results obtained from previous analyses. It leverages the organisation of metabolic pathways and networks to predict functional activities and possible biological pathways by circumventing upfront metabolite identification. Across metabolic networks, random permutations of extracted metabolic features were drawn to calculate Fisher's exact statistic on the null randomisation distribution of the test statistic.⁴⁴ The analyses were performed separately for each of the ionisation modes. Although multiple testing error corrections may provide a theoretical guarantee against elevated false discovery rates, given the smaller sample size of this study, it can potentially omit weaker yet biologically relevant features due to their inter-correlations.^36,58,59 Therefore, we included all metabolic features with unadjusted nominal p-values less than 0.05 to achieve sufficient enrichment. Finally, we chose only those enriched pathways with at least three hits of significant metabolites from all the enrichment analyses. The adduct chosen for analysis in the positive ion mode is M + H, whereas it is M−H for the negative ion mode. The boundary on the limit of mass accuracy was set at 10 ppm.

Sensitivity and secondary analyses included: (1) Evaluating linearity between each PFAS concentration, as well as other quantitative covariates (age at diagnosis/matching and BMI at enrolment) with the outcome using B-splines. (2) We incorporated covariate balancing through optimal full matching by dichotomising the weighted PFAS index as high vs low.^60,61 The regression analysis was then repeated on the matched covariate-balanced dataset. (3) We repeated the WQS regression analysis for the PFAS-mixture and T2D association, excluding PFHpA from the mixture, as PFHpA had >50% of concentrations below the LOD. (4) We fitted the ExWAS using linear regression for each PFAS exposure and metabolite to evaluate whether the results for individual PFAS exposure and the PFAS-mixture remained consistent. (5) Lastly, we repeated the WQS regressions stratified and examined effect modification as well as the interaction by sex and self-reported ancestry, as few previous studies have suggested sex- and race-specific effects.^23,40,62,63

Role of funders

The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health, and do not have any role in study design, data collection, data analyses, interpretation, or writing of report.

Results

Study population description

T2D cases compared to controls were more likely to smoke and had higher BMIs on average at baseline (Table 1). PFOS, PFOA, and PFNA were the PFAS detected at the highest concentrations in both cases and controls, with no statistically significant difference in PFAS concentrations noted between the two groups in unadjusted analyses. Correlation between all seven PFAS ranged from 0.02 [PFHpA-PFHxS] to 0.88 [PFHpS-PFOS] and, after excluding PFHpA, correlations ranged from 0.39 (PFHxS-PFDA) to 0.88 (PFHpS-PFOS) (Supplemental Figure S1). The median calendar year for blood draw for cases and controls were 2010 and 2011, respectively.

PFAS associations with the T2D risk

The odds ratio (OR) [95% CI], distribution of weights from WQS, and estimates from c-logistic regressions are shown in Fig. 3 (and Supplemental Table S1). The PFAS-mixture was positively associated with increased risk of incident T2D (OR [95% CI] = 1.31 [1.01, 1.70], p-value is 0.04), with PFOS and PFHxS having the highest contributions in the PFAS-mixture association (with respect to the uniform threshold of equal weights). In individual PFAS analyses, compared to other PFAS, PFOS (1.26 [0.93, 1.71], the p-value is 0.13) and PFHxS (1.17 [0.87, 1.58], the p-value is 0.30) had stronger adverse associations of larger magnitude with T2D. Associations for PFNA and PFOA were in the same adverse direction as for PFOS and PFHxS but more attenuated. We found no evidence of an association between PFDA, PFHpA, or PFHpS and T2D.

Fig. 3.

Associations between PFAS concentrations in prediagnostic plasma (as a PFAS mixture and individual PFAS exposures) with the odds for T2D. (a) OR (95% CI) obtained from WQS regression for the PFAS mixture and conditional logistic regressions for individual PFAS, adjusted for sex, age at diagnosis, self-reported ancestry, BMI at baseline, date of blood draw, and smoking habits. (b) Barplot denotes the weights (contributions) of individual PFAS to the PFAS mixture association with incident T2D. The weights sum up to one. The solid horizontal line denotes the expected weights if all PFAS had contributed equally to the mixture index.

PFAS associations with plasma metabolites at baseline

Results from ExWAS using hBWQS between PFAS mixtures and metabolites are presented in Fig. 4a and c. Multiple metabolites were significantly associated with the PFAS mixture after multiple comparison error corrections in the positive and negative ionisation modes. In positive mode, the top five metabolites (with p-value <0.0001) were N (6)-Methyllysine, lysophosphatidylcholine (LPC) 22:5, ceramide (Cer) 18:1, 3-indole propionic acid, and cysteinyl glycine (Cys–Gly). In the negative mode, the top five metabolites (with p-value <0.0001) were sulfolithocholylglycine, monoacylglycerol (M.G.) 18:2, 2,3-Diphosphoglyceric acid, N-Acetylornithine, and Cortisone. Moreover, the metabolites indole acrylic acid, lysophosphatidylethanolamine (LPE) 20:4, taurine, and tryptophan were significantly associated with the PFAS mixture in both positive and negative ionisation modes. The complete list of metabolites significantly associated with the PFAS mixture, along with the beta coefficients (95% CIs) and p-values, is presented in Supplemental Table S2.

Fig. 4.

Volcano Plot showing beta coefficients and raw p-values for the associations between the PFAS mixture and metabolites in prediagnostic plasma (a and c) and between plasma metabolites and the odds for incident T2D (b and d). The subfigures a and c show associations between the hBWQS PFAS mixture and metabolites in positive and negative ionisation modes, respectively, adjusted for sex, age at diagnosis, self-reported ancestry, BMI at baseline, date of blood draw, and smoking habits. The subfigures b and d show associations between the metabolites and incident T2D in positive and negative ionisation modes, respectively, adjusted for sex, age at diagnosis and self-reported ancestry, BMI at baseline, date of blood draw, and smoking habits. The black horizontal line denotes the p-value threshold at 0.05. The green horizontal line in a and c denotes a p-value threshold of 0.005 (positive ionisation mode), and in b and d, it denotes a p-value threshold of 0.007. The blue and red dots denote significantly negatively and positively associated metabolites, whereas the black dots denote metabolites with non-statistically significant associations.

In individual PFAS analyses, PFDA presented the largest number of associations with metabolites in both modes compared to other PFAS (Fig. 5). Some metabolites (e.g., corticosterone in positive mode—Fig. 5a and sulfolithocholyglycine in negative mode– Fig. 5b) were associated with multiple individual PFAS, as well as with the overall PFAS mixture (Fig. 4). Most detected associations were positive in direction, indicating an increase in metabolite levels per PFAS exposure increase. The complete list of metabolites significantly associated with individual PFAS, along with the beta coefficient and the p-values, is presented in Supplemental Table S3.

Fig. 5.

Figure showing the associations between individual PFAS exposures and metabolites in prediagnostic plasma after adjusting for multiple comparisons and using high-resolution metabolomics in positive (a) and negative (b) ionisation modes. The metabolite classes were represented with different colours on the right side of the figure. Models were adjusted for sex, age at diagnosis and self-reported ancestry, BMI at baseline, date of blood draw, and smoking status.

Metabolites in prediagnostic plasma associated with higher odds for T2D

PC P-34:2/PC O-34:3 and sulfolithocholyglycine metabolites remained significantly associated with incident T2D after correcting for multiple comparison errors in positive and negative modes, respectively (Fig. 4b and d). Furthermore, three metabolites, including 5-hydroxy-tryptophan, sulfolithocholyglycine, and glucoheptulose, were associated with both the PFAS mixture and T2D risk, but only sulfolithocholyglycine passed multiple comparison corrections. The full list of metabolites (beta coefficient and p-value) significantly associated with T2D was presented in Supplemental Table S4.

Pathway enrichment results for the association of metabolic pathways with PFAS exposures and odds for T2D

Enriched metabolic pathways significantly associated with the PFAS mixture were tyrosine metabolism, vitamin C and aldarate metabolism, glutamate metabolism, arginine, and proline metabolism in positive ionisation mode (Fig. 6a), and drug metabolism—cytochrome p450 in negative ionisation mode (Fig. 6b). Enriched pathways associated with the T2D were bile acid biosynthesis, vitamin B3, alanine, and aspartate metabolism (Fig. 6b), as well as fructose and mannose metabolism (Fig. 6b). Glutamate metabolism, arginine and proline metabolism, and drug metabolism—cytochrome p450 were common pathways associated with both the PFAS mixture and T2D.

Fig. 6.

Metabolic pathways associated with the PFAS-mixture (on the left) and incident T2D (on the right), using high-resolution metabolomics in positive (a) and negative (b) ionisation modes. X-axis represents -log (p-value, base = 2). p-Values are based on Fisher's Exact test for the pathway. The size of the point also represents -log₂ (p-value). The red vertical line denotes a p.value equal to 0.05. Associations were adjusted for sex, age at diagnosis, self-reported ancestry, BMI at baseline, date of blood draw, and smoking habits.

Sensitivity analysis

In sensitivity analysis (1) We found no meaningful departures from linearity in the associations between PFAS (or secondary quantitative covariates) with the outcome using B-spline analyses. (2) We present the love plot of the covariate balancing⁶⁴ in Supplemental Figure S2. The repeated regression model on the balanced data yielded almost similar association estimates. (3) After rerunning the WQS regression model with PFAS-mixture and T2D excluding PFHpA, the estimated OR did not meaningfully change (<5% change in estimate), and PFOS remained the top contributor to the PFAS mixture association with T2D as in the main analysis. (4) In the ExWAS using linear regression for each PFAS exposure and metabolite, the overall associations observed for individual PFAS with plasma metabolites were in a similar direction as those observed for the PFAS mixture with plasma metabolites (Supplemental Figure S3). (5) In the sex-stratified analysis, we did not find clear evidence for an interaction by sex in the PFAS mixture-T2D association (OR [95% CI] = 1.40 [0.81, 2.42] in males, and 1.25 [0.81, 1.94] in females, p-sex interaction >0.9). In ancestry-stratified analysis, the odds of T2D associated with the PFAS mixture tended to be higher in European Americans and Hispanics compared to African Americans, but the interaction did not reach the significance level (OR [95% CI] = 1.47 [0.83, 2.63], 1.25 [0.65, 2.40], and 1.19 [0.71, 2.00], respectively, all p-ancestry interactions>0.8). (6) Lastly, missing value was noted only for BMI (n = 1) and imputed using the sample mean under the assumption of missing at random. We did not observe any meaningful change in estimates (i.e., percent change in the OR less than 2%) in complete data analysis without imputing missing values.

Discussion

In this nested case–control investigation within the multiethnic BioMe cohort, we showed that exposure to the PFAS mixture may be associated with higher odds for T2D over a ∼6-year follow-up period. In further support of this association, we found common metabolites, specifically amino acids, steroid sulfates, bile acid conjugates, and monosaccharides carbohydrates, associated with both the PFAS mixture exposure as well as with later odds for T2D. Our findings provide evidence that higher PFAS mixture exposures could increase risk for T2D through dysregulations in amino acid biosynthesis and drug metabolism pathways. This is one of the few investigations on the role of the metabolome in the PFAS-T2D association. Findings require replication in larger multiethnic populations and can help advance the knowledge of PFAS health risks and potential underlying mechanisms in previously underrepresented ethnic minorities.

In line with our findings, a recent systematic review of epidemiology studies integrating high-resolution metabolomics showed consistent results about PFAS associations with alterations in several lipids and amino acid metabolites involved in energy and cell membrane disruption.^28,59,65, 66, 67, 68, 69 PFAS act as ligands for peroxisome proliferator-activated receptors (PPARs) α, γ and β/δ,70, 71, 72, 73, 74 which regulate glucose and lipid metabolism, adipogenesis, fat storage, and inflammation.75, 76, 77 PFAS also stimulates insulin secretion⁷⁸ and may decrease beta-cell viability in the pancreas.^79,80 In rodents, PFAS exposures were demonstrated to have altered levels of lipids, glucose, metabolism-regulating hormones, and cytokines in peripheral blood, which promoted T2D development.81, 82, 83, 84, 85, 86, 87 We found three metabolites, 5-hydroxy-tryptophan, sulfolithocholyglycine, and glucoheptulose, significantly associated with both the PFAS mixture and risk of future T2D diagnosis, out of which sulfolithocholyglycine survived multiple test correction. Sulfolithocholyglycine, a bile acid conjugate, is among the important signalling molecules coordinately regulating lipid, glucose, drug, and energy metabolism,⁸⁸ and altered levels of sulfolithocholyglycine have been observed in patients with T2D,⁸⁹ potentially explaining a link between PFAS exposure and T2D risk. Similar to the result of our study, in a cohort of female firefighters in California, sulfolithocholylglycine was positively associated with PFAS exposures.⁹⁰ Other previous studies also observed that PFAS contributes to lipid and amino acid dysregulation. In previous untargeted high-resolution metabolomics investigations, PFHxS exposure was shown to be associated with dysregulation of glycosphingolipids, linoleic acid, branched-chain amino acids (BCAAs), and aromatic amino acids in US Hispanic adolescents.¹⁷ In a cohort of European children, a mixture of legacy PFAS was shown to be associated with higher serum levels of glycerophospholipids, BCAAs, and aromatic amino acids.²⁴

We found amino acid biosynthesis pathways, namely glutamate metabolism and arginine and proline metabolism, to be significantly associated with both the PFAS mixture and T2D in our study. These findings align with previous literature showing associations between arginine, aspartate, and asparagine in adults with the risk of T2D.^17,91,92 In previous high-resolution mass spectrometry studies of animal models and human studies, PFAS exposures were associated with metabolic alterations in amino acids, including glutamate, tyrosine, tryptophan, arginine, and proline metabolisms.^17,23,84 We found a consistent association between exposures to PFAS mixtures and metabolic dysregulations in amino acid metabolism, consistent with the hypothesis that alterations in amino acid circulating levels may contribute to the risk of developing T2D.⁹³ We also found that drug metabolism—cytochrome P450 was significantly associated with both the PFAS mixture and incident T2D. Previous epidemiological studies have also shown an association between PFAS exposures and alterations in the drug metabolism pathway.^17,36,66 Further, in the mice model, PFAS exposure was shown to induce cytochrome p450 (CYP450) enzymes in mouse liver.⁹⁴ Moreover, in recent studies, PFAS-induced hepatotoxicity was proposed to be caused by interactions with the CYP450.⁹⁵ On the other hand, a recent in vivo study demonstrated that a low chronic inflammatory status associated with T2D modulates CYP450 activities.⁹⁶ Further, patients with T2D have increased activity of inflammatory markers, which, in turn, are associated with the downregulation of several isoforms of CYP450.⁹⁷

Our study has a few limitations, including the restricted sample size, which has limited our study's precision to detect potential interactions by sex or ancestry. Given that this is a novel area of research, understudied in multiethnic populations, we hope that these results will inspire future larger investigations in this field. Furthermore, we only measured seven PFAS, while evidence for other emerging short-chain PFAS is urgent and currently limited in this field of research. The PFAS and untargeted metabolome analysis were conducted at the same time in our study, and therefore, we cannot establish temporality in this association. Furthermore, pathway analysis results should be interpreted cautiously because of their reliance on incomplete pathway databases, the lack of consideration of potentially complex interactions between metabolites, and the potential for an elevated false discovery rate due to multiple comparison errors.^98,99 Further, we acknowledge the limitations of pathway analysis in metabolomics that can still generate misleading results; a deeper biological understanding through wet lab experiments is crucial to understanding the hypothesised dysregulated pathways.¹⁰⁰ We also note that the current sample size is underpowered to identify any ancestry-stratified effects of PFAS exposures on T2D diagnosis. Lastly, given the observational study design and reliance on an EHR-linked biobank, we cannot rule out residual or unmeasured confounding by dietary intake patterns or, in female participants only, parity and breastfeeding history.^18,101

An important strength of our study is the ethnic diversity of the origin BioMe population, as especially African Americans and Hispanics have been underrepresented in prior research.¹⁰² Ethnic minorities in the US are more likely to be exposed to higher levels of PFAS and are also more likely to develop T2D.¹⁰³ Further research in multiethnic larger populations could be critical to elucidate the potential impacts of PFAS exposures in diabetes disparities.¹⁰⁴ Furthermore, state-of-the-art ultra-high-resolution mass spectrometry approaches leveraged in this study offer a more detailed characterisation of endogenous metabolites and can provide a more comprehensive characterisation of early metabolic signatures associated with T2D of potential clinical utility in the future. Our study included two ionisation modes, HILIC and C18, to capture a wider range of metabolites. A HILIC column can be considered complementary to a C18 column since they have opposite retention mechanisms, with HILIC being better for separating polar metabolites, while C18 is better for non-polar metabolites, making them very useful for analysing a wider range of non-targeted compounds when used together.^105,106 In our pathway analysis, because of this complementarity, we observed slight differences in detected pathways; the HILIC positive analysis had more metabolites belonging to protein synthesis pathways, whereas the one in the C18 negative had more metabolites related to carbohydrate synthesis related pathways. The use of cutting-edge biostatistical approaches for analysing complex exposure mixtures and metabolomics associated with T2D is another strength of our study. Findings from this study underscore the importance of preventing PFAS exposures to promote public health and can inform about potential mechanisms underlying PFAS effects on human metabolism in multiethnic populations.

In this nested case–control study of incident T2D within the multiethnic BioMe EHR-linked Biobank, we found that prediagnostic plasma concentrations of PFAS mixtures are associated with increased risk for T2D over ∼6 years of follow-up, and that these associations could be due to dysregulations in amino acid biosynthesis and drug metabolism. Findings may inform public health policy needed to reduce PFAS exposures in the environment, as well as precision-environmental health approaches for T2D prevention and treatment.

Contributors

Vishal Midya: Writing—review & editing, Writing—original draft, Visualisation, Validation, Software, Methodology, Investigation, Formal analysis, Conceptualisation. Meizhen Yao: Writing—review & editing, Writing—original draft, Visualisation, Validation, Software, Methodology, Investigation. Elena Colicino: Writing—review & editing, Investigation, Methodology. Dinesh Barupal: Writing—review & editing, Investigation, Methodology. Xiangping Lin: Data Acquisition, Writing—review & editing. Chris Gennings: Writing—review & editing, Investigation, Methodology. Leda Chatzi: Writing—review & editing, Investigation. Veronica Wendy Setiawan: Writing—review & editing, Investigation. Ruth J. F. Loos: Writing—review & editing, Investigation, Methodology. Ryan W. Walker: Writing—review & editing, Investigation, Methodology. Douglas I. Walker: Writing—review & editing, Investigation, Methodology. Damaskini Valvi: Writing—review & editing, Conceptualisation, Investigation, Methodology, Supervision, Funding acquisition. All authors contributed to editing the manuscript and approved the final manuscript. All authors read and approved the final version of the manuscript.

Data sharing statement

Due to the sensitive nature of the data, it cannot be made available. However, codes and de-identified data can be considered to be made available upon written request (with a proposal of how the data will be used) to the study senior author (Damaskini Valvi: dania.valvi@mssm.edu).

Declaration of interests

Dr. Leda Chatzi has served as an expert consultant for plaintiffs in litigation related to PFAS-contaminated drinking water. Dr. Ruth J. F. Loos received consulting fees and honoraria from Novo Nordisk and Elly Lilli. Dr. Damaskini Valvi received support for attending meetings from the Society of Toxicology. All other authors declare no other competing financial interests or personal relationships that might influence the work reported in this paper.