Mitochondrial and psychosocial stress-related regulation of FGF21 in humans

Abstract

Fibroblast growth factor 21 (FGF21) is a metabolic hormone induced by fasting, metabolic stress and mitochondrial oxidative phosphorylation (OxPhos) defects that cause mitochondrial diseases (MitoD). Here we report that acute psychosocial stress alone (without physical exertion) decreases serum FGF21 by an average of 20% (P < 0.0001) in healthy controls, but increases FGF21 by 32% (P < 0.0001) in people with MitoD, pointing to a functional FGF21 interaction between the stress response and OxPhos capacity. We further define co-activation patterns between FGF21 and stress-related neuroendocrine hormones and report associations between FGF21 and psychosocial factors related to stress and wellbeing. Overall, these results highlight a potential role for FGF21 as a stress hormone involved in meeting the energetic needs of psychosocial stress.

The hormone FGF21 regulates systemic energy metabolism¹. It is released by the liver and other organs in response to metabolic challenges (for example, fasting) and acts centrally and peripherally to regulate nutrient intake and metabolic processes^2–5. Extended Data Fig. 1 and Supplementary Table 2 show RNA transcript levels for FGF21 (highest in liver, muscle and immune cells) and its receptor complex consisting of FGFR1 (all isoforms) and β-klotho (encoded by KLB; highest in arteries, pancreas, small intestine and adipose tissue) across 50 human tissues, establishing their expression patterns across several organs.

FGF21 is elevated in most clinical cases and animal models of genetic OxPhos defects^3,4,6,7, making it a relatively sensitive and specific biomarker of MitoD. Naturally, OxPhos capacity and energy metabolism are closely tied to mental and psychosocial states⁸, as perceived threats trigger endocrine responses (for example, cortisol, norepinephrine) that drive the metabolic shifts required to fuel ‘fight-or-flight’ responses, and mounting a stress response costs energy^9,10. Elevated FGF21 levels are also observed in animal models of chronic distress¹¹ and in human psychiatric conditions^12,13, suggesting that mitochondrial defects and psychosocial factors may jointly influence FGF21 biology.

To experimentally probe this question, two groups of participants were enrolled in the Mitochondrial Stress, Brain Imaging and Epigenetics (MiSBIE) study¹⁴: individuals who served as controls with normal mitochondrial biology (n = 70) and patients with MitoD (n = 40), including 15 people with a single, large-scale mtDNA deletion and 25 people with a pathogenic mitochondrial DNA variant (m.3243 A > G mutation, five of which had mitochondrial encephalomyopathy, lactic acidosis, stroke-like episodes [MELAS], which is the most severe form of MitoD). See Supplementary Table 1 for a breakdown of symptoms by group. Detailed information about the study procedures and measures has been previously described¹⁴ and is detailed in the Methods.

We first confirmed two hallmarks of FGF21 biology: its induction by OxPhos defects and its link to fasting and metabolism. Participants arrived at 09:00 h and completed anthropomorphic measurements and a fasting blood draw at ~10:00 h, after which they received a standar dized breakfast, completed questionnaires and remained relatively sedentary before consuming a standardized lunch. Blood was collected after lunch at ~13:00 h to assess FGF21 levels in the fed state. Compared to controls, participants with MitoD exhibited markedly elevated serum FGF21 levels (180–420% higher, P < 0.0001; Fig. 1a), which were highest in the deletion subgroup (Extended Data Fig. 2a,b). As expected, morning (AM)-fasted FGF21 concentrations were higher than afternoon (PM)-fed concentrations across groups (Ps < 0.001; Extended Data Fig. 2a,c). The morning-to-afternoon decline was more robust and consistent in controls (Δ −73.3%, g = 0.5, P < 0.0001) than MitoD (Δ −14.7%, g = 0.1, P < 0.0005; Extended Data Fig. 2c). Receiver operating characteristic curve analyses also confirmed FGF21 as a relatively sensitive and specific biomarker of OxPhos deficiency, with characteristics comparable to prior studies (Extended Data Fig. 3a–d). FGF21 measured in the fed state was most predictive of MitoD (Extended Data Fig. 3b), and stratifying the MitoD cohort based on the presence or absence of major clinical symptoms revealed that individuals with documented symptom(s) showed significantly higher PM–fed FGF21 levels (Extended Data Fig. 4).

Fig. 1 |. Elevated serum FGF21 in MitoD is exacerbated in response to acute mental stress and correlated with disease severity.

a, Distribution of the study cohorts comparing FGF21 levels in controls and MitoD when fasted (AM) (controls: M = 235.0 pg ml⁻¹, n = 65 vs MitoD: M = 657.8 pg ml⁻¹, n = 34, P < 0.0001, two-sided Mann–Whitney U-test) or fed (PM) (controls: M = 109.0 pg ml⁻¹, n = 65 vs MitoD: 567.8 pg ml⁻¹, n = 33, P < 0.0001). Box plots depict means for each group; error bars, s.e.m. b, Schematic of the modified TSST paradigm used to assess FGF21 and hormone responses to acute stress. Pictogram of physiological stress reactivity is for illustrative purposes only; no data are represented. c, Changes in blood FGF21 concentrations in response to the TSST in controls (at each timepoint: −5 min, n = 65; 5 min, n = 63; 10 min, n = 60; 20 min, n = 61; 30 min, n = 61; 60 min, n = 61; 90 min, n = 60; 120 min, n = 62) and MitoD (−5 min, n = 33; 5 min, n = 33; 10 min, n = 33; 20 min, n = 32; 30 min, n = 31; 60 min, n = 30; 90 min, n = 28; 120 min, n = 30). Values are mean absolute change (pg ml⁻¹) relative to baseline measurements at −5 min (absolute change = value − baseline). Inset, relative change in FGF21 (pg ml⁻¹) in controls (relative change = value / baseline). Error bars, s.e.m. d, Spearman’s rank correlations between disease severity scores and FGF21 levels at 90 min post stress in the MitoD cohort; fitted trend line with 95% CIs. Illustrations for a, b and d created with BioRender.com. MELAS, mitochondrial encephalomyopathy, lactic acidosis, stroke-like episodes; NMDAS, Newcastle Mitochondrial Disease Adult Scale; COMPASS, Composite Autonomic Symptom Score; CNS, Columbia Neurological Score.

Consistent with prior research^15,16, fasting FGF21 in controls was comparable between females and males and was positively correlated with age (r_s = 0.37, P = 0.002), per cent body fat (r_s = 0.25, P = 0.046), total fat mass (r_s = 0.31, P = 0.015), fasting blood glucose (r_s = 0.49, P < 0.0001), insulin (r_s = 0.027, P = 0.026) and total cholesterol (r_s = 0.30, P = 0.016); and negatively correlated with high-density lipoprotein (r_s = −0.35, P = 0.0043) (Extended Data Fig. 5a–d). Apart from the correlation between FGF21 and percent body fat (r_s = 0.42, P = 0.013), these relationships were not observed in the MitoD group, suggesting that OxPhos defects disproportionally elevate FGF21 and obscure the relationships observed in healthy controls.

To isolate the effects of psychosocial stress without the confounding effects of exertion or hunger¹⁷, participants completed a modified Trier Social Stress Test (TSST) around 13:00 h. The TSST, which involved 2 min of preparation and a 3 min speech in front of an evaluator, is a validated lab stressor that reliably elicits psychophysiological reactivity¹⁴ (see Methods). Serum FGF21 was measured in blood drawn immediately before the stressor (−5 min) and at +5, 10, 20, 30, 60, 90 and 120 min post stress, enabling the first assessment of the acute response of FGF21 to psychosocial stress in both controls and patients with MitoD (Fig. 1b).

In the control group, FGF21 dropped by 20% (Δ −13.10 pg ml⁻¹, g = 0.75, P < 0.0001) 20 min after the stressor onset and returned to baseline by 90 min (Fig. 1c and Extended Data Fig. 6b). By contrast, the MitoD group showed a 32% FGF21 increase (Δ +180.8 pg ml⁻¹, g = 1.17, P < 0.0001), peaking at 90 min (Fig. 1c), with similar patterns across all MitoD subgroups (Extended Data Fig. 6c). A mixed-effects model showed that mitochondrial OxPhos defects moderated the FGF21 stress response, with controls and MitoD participants displaying distinct trajectories (Extended Data Fig. 6a). When considered alongside cortisol and GDF15 trajectories (Extended Data Fig. 7)^18,19, these results suggest that healthy individuals may temporarily suppress FGF21 during acute stress to prioritize rapid energy use, whereas patients with MitoD may upregulate FGF21 to compensate for impaired OxPhos-derived energy transformation.

To assess the clinical relevance of FGF21, we examined how serum levels—measured while fasted, after lunch and during stress—related to mitochondrial disease severity, functional capacity and fatigue. Patients with MitoD with higher peak FGF21 (90 min post stress) had more severe disease manifestations (|r_s| = 0.66–0.71; Fig. 1d), and these associations were stronger than those observed with fasting levels (Extended Data Fig. 8a,b). These results suggest that stress-induced FGF21 changes reflect clinically meaningful links between acute mental stress and mitochondrial dysfunction.

To explore why acute stress affects FGF21 differently in controls and patients with MitoD, we examined its interactions with hormones of the hypothalamic–pituitary–adrenal (HPA), hypothalamic–pituitary–gonadal (HPG) and sympathetic nervous system (SNS) axes, measured in parallel across the stress time course.

Compared to controls, patients with MitoD showed distinct correlation patterns between FGF21 and stress-related steroid hormones of the HPA (cortisol, cortisone) and HPG (DHEA, testosterone) axes across the stress–response timeline (Fig. 2a). Controls had weak negative or non-significant correlations between FGF21 and cortisol or cortisone, whereas patients with MitoD showed moderate to strong positive correlations. A similar pattern emerged with DHEA, with mostly negative correlations in controls but positive correlations in MitoD. For testosterone, the pattern was reversed: controls showed mixed weak correlations, whereas patients with MitoD had consistently strong negative correlations. These findings document altered co-regulation of FGF21 with endocrine axes in individuals with mitochondrial OxPhos defects.

Fig. 2 |. Differential patterns of co-regulation between FGF21 and hormones of the HPA, HPG and SNS axes and correlations between FGF21 and psychosocial factors in Controls and MitoD.

a,b, Top: Spearman’s correlation matrices depicting differential patterns of co-regulation between FGF21 and corresponding measurements of hormones of the HPA and HPG (a) and SNS axes (b) across the TSST time course in controls and MitoD. All measurements are log₁₀ transformed. Scatter plots with Spearman’s correlations for selected time points featuring moderate to high correlations in MitoD compared to controls. Linear estimates with 95% CIs are provided for illustrative purposes. For scatter plots: cortisol: controls, n = 57 vs MitoD, n = 31; cortisone: controls, n = 57 vs MitoD, n = 31; testosterone: controls, n = 45 vs MitoD, n = 23; DHEA: controls, n = 27 vs MitoD, n = 8; norepinephrine: controls, n = 61 vs MitoD, n = 30; epinephrine: controls, n = 36 vs MitoD, n = 17; dopamine: controls, n = 62 vs MitoD, n = 33; serotonin: controls, n = 61 vs MitoD, n = 31. Note that the results for DHEA and epinephrine should be interpreted with caution given the number of undetectable samples, which we interpret as a technical error but may also reflect physiologically low levels. Illustrations in a and b created with BioRender.com. c,d, Associations between FGF21 levels and psychosocial self-report measures categorized as positive (c) and negative factors (d), organized from largest (bottom) to smallest (top) for the MitoD group. Group-specific associations were computed using linear regression models adjusting for age and group-specific per cent body fat. Circles represent the effect size of the adjusted bivariate association between FGF21 and the self-report measures, computed as Cohen’s r with 95% CIs. Diamonds indicate overall Cohen’s r for each group across each category, with s.e.m. Scatter plots show the three strongest relationships in each category. The y-axis scale for FGF21 in the Couples Satisfaction plot differs from the others because Couples Satisfaction scores only exist for participants in romantic relationships; MitoD participants with the highest FGF21 levels were not in romantic relationships, so they did not complete this scale. Two-tailed significance is indicated as *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001.

We also examined associations between FGF21 and catecholamines downstream of the SNS axis (that is, norepinephrine, epinephrine, dopamine, serotonin; Fig. 2b). In controls, correlations with FGF21 were generally weak or non-significant. By contrast, patients with MitoD who had higher FGF21 levels generally exhibited lower serotonin and dopamine levels across timepoints. Although the roles of these circulating neurotransmitters are not fully understood, these patterns support the notion that mitochondrial biology influences stress neuroendocrinology²⁰ and the FGF21 response to mental stress in humans.

To interrogate the generalizability of these group-level findings, we leveraged the naturally occurring inter-individual variation in FGF21 responses across our 64 control participants. Stratifying by stress-induced FGF21 reactivity, we found that 30% of controls showed mild increases in the same direction as patients with MitoD. These individuals with net positive FGF21 reactivity had distinct hormonal and metabolic profiles compared to those with net negative reactivity, including higher cortisol and cortisone responses, lower baseline testosterone and higher body fat and insulin levels (Extended Data Fig. 9). Given that cortisol and cortisone are catabolic hormones and testosterone is anabolic, these results could indicate that mental stress influences interactions between FGF21 and the anabolic–catabolic axes linked to metabolic health.

To more fully examine the psychobiology of FGF21, we next explored relationships between FGF21 (AM–fasting and PM–fed, separately) and trait-like psychosocial factors traditionally linked to disease risk or wellbeing. These measures include a range of positive (for example, environmental mastery, couple’s satisfaction, social support) and negative factors (for example, loneliness, childhood neglect, work stress) measured with a comprehensive battery of validated questionnaires in the MiSBIE study (see Methods). Consistent with the response of FGF21 to acute stress, chronic stressors and a lack of positive protective factors were more strongly associated with FGF21 in patients with MitoD than in controls, especially in the fed state (Fig. 2c,d and Extended Data Fig. 10). In patients with MitoD, higher environmental mastery (r_s = −0.44, P < 0.0001) and couple’s satisfaction (r_s = −0.43, P < 0.01; Fig. 2c) were associated with lower FGF21, whereas greater loneliness correlated with higher FGF21 (r_s = 0.40, P < 0.0002). These results suggest that patients with MitoD may be more metabolically sensitive to psychosocial stress, further pointing to a psychobiological link between OxPhos defects and FGF21 regulation.

Finally, we validated the associations between FGF21 biology and psychosocial factors from the small MiSBIE study in 51,818 UK Biobank participants of various health statuses²¹ (https://proteome-phenome-atlas.com). Consistent with our couple’s satisfaction result, after correcting for demographic variables including age, sex and body mass index (BMI), UK Biobank participants who experienced a marital separation or divorce had elevated levels of plasma FGF21 (P = 4.3 × 10⁻⁷, β = 0.066, 95% confidence interval (CI) = [0.041, 0.091], n = 17,281). Financial difficulties were also associated with elevated FGF21 (P = 6.5 × 10⁻¹¹, β = 0.074, 95% CI = [0.051, 0.096], n = 23,320). On the other hand, engaging in non-exercise-related and non-alcohol-related leisure and social activities was associated with lower FGF21 levels (P = 2.7 × 10⁻¹³, β = −0.065, 95% CI = [−0.082, −0.048], n = 35,860), substantiating a potential ‘social’ dimension to FGF21 regulation.

Together, these findings highlight three key points: serum FGF21 is rapidly regulated by mental stress in a manner dependent on mitochondrial health; FGF21 shows dynamic associations with anabolic–catabolic endocrine axes; and FGF21 relates to modifiable psychosocial factors—both positive and negative experiences—linked to disease risk and ageing²². This study thus extends our understanding of FGF21 from a metabolically responsive hepatokine to a psychometabolic signal, making FGF21 a mental stress hormone, sensitive to systemic mitochondrial OxPhos capacity.

The diverging FGF21 trajectories observed during acute mental stress—that is, suppression in controls versus marked elevation in patients with MitoD—support an ‘energetic’ framework for the stress-to-disease cascade⁸, wherein a robust OxPhos capacity may buffer the metabolic costs of threat, whereas compromised mitochondria may require compensatory FGF21 upregulation. These dynamics position mitochondria as a central nexus through which psychological stimuli may be transduced into endocrine signals, offering an energetic pathway for how mental stress becomes biologically embedded.

Although FGF21 is widely regarded as a relatively static biomarker of mitochondrial disease, chiefly driven by energy insufficiency and fasting–fed status, our findings reveal that mental stress alone can double circulating FGF21 in some m.3243 A > G carriers (average group-level change, +30% at 90 min). One potential avenue for clinical translation and implementation involves tracking treatment responses across mitochondrial and stress-related conditions; however, further research is needed to clarify how OxPhos capacity and stress interact in different individuals to regulate FGF21 over time. Our findings suggest that a single FGF21 measurement taken during a stressful clinic visit or research assessment could overestimate disease burden and misrepresent a patient’s underlying metabolic state. This possibility underscores the need for repeated measurements and the consideration of situational context when interpreting FGF21.

Finally, the strong correlations between FGF21 and chronic psychosocial factors in patients with MitoD suggest that adverse mental states may chronically elevate FGF21, whereas positive factors may reduce its circulating level. Given that many patients report that negative mood exacerbates their most debilitating symptoms²³, longitudinal monitoring of FGF21 (and probably GDF15) could help clinicians gauge the dual efficacy of therapies that target both metabolic and psychological domains. Integrating FGF21 into psychobiological models of mitochondrial disease may thus support a more comprehensive and personalized approach to disease management.

Taken together, our findings position circulating FGF21 as a psychosocial stress hormone in individuals with impaired mitochondrial OxPhos capacity. However, the observational nature and limited scope of the present study preclude any inference about causality. Whether FGF21 actively modulates stress-related behaviours or simply mirrors the underlying metabolic-stress axis remains an open question calling for longitudinal and mechanistic studies.

Methods

Ethics statement

Participants were enrolled in the MiSBIE study¹⁴ in accordance with the guidelines set by the New York State Psychiatric Institute (Institutional Review Board protocol no. 7424) and the Columbia University Medical Center (Institutional Review Board protocol no. AAAU9470). The study was registered under ClinicalTrials.gov registration: NCT04831424.

Participants

Participants were recruited locally from our clinic at the Columbia University Irving Medical Center and across the USA and Canada. Given that the TSST was a primary component of our study design, we used a screening process whereby we excluded individuals with cognitive impairment who were incapable of providing informed consent. We used the Telephone Interview for Cognitive Status, a widely used, brief test of mental status and excluded anyone who received a score of ≤30. All participants provided written informed consent to participate in the study and authorized the potential dissemination of results. Recruitment took place between June 2018 and May 2023. A total of 110 participants aged 18–60 years were enrolled, including 70 healthy controls (n = 48 female, 22 male) and 40 individuals who had genetically defined MitoD (n = 28 female, 12 male). The participants’ sex was determined by self-report. Eligibility criteria have been previously reported¹⁴.

Procedure

The fasting assessments and psychosocial stress paradigm described here were part of an extended 2-day MiSBIE study protocol that has been previously described in detail¹⁴. Day 1 included a fasting blood draw in the morning, followed by a clinical assessment, completion of self-report questionnaires and an extensive psychophysiological protocol that included a modified TSST, which is a well-established method for inducing and measuring psychosocial stress responses²⁴.

The modified TSST speech task administered in this study consisted of 120 s of speech preparation followed by 180 s of speech delivery under evaluation. To induce stress, the study coordinator entered the room carrying a clipboard and a video camera mounted on a tripod. The coordinator silently set up the video camera to give the impression that it was recording and placed a full-length mirror in front of the participants so they could see themselves throughout the task. Participants were instructed to prepare a speech in response to an uncomfortable scenario: being accused by a security guard of stealing merchandise from a store. After the instructions and scenario were presented, participants completed a subset of items from the Primary Appraisal Secondary Appraisal threat questionnaire. They were given 2 min to prepare their speech and were then instructed to deliver it to an evaluator for 3 min. The ‘evaluator’, also wearing a white coat and carrying a clipboard, entered the room silently and took notes as the participants delivered their speeches. After the 3-min period, the evaluator left, and the camera and mirror were removed simultaneously. Blood and saliva samples, body temperature and affect ratings were collected 5 min before the initiation of the speech task, and again 5, 10, 20, 30, 60, 90 and 120 min post TSST. This observational study included no intervention or random assignment; assays were run under blinded conditions on de-identified samples, with group labels revealed only after data collection. No data points were excluded from the analyses.

Preliminary results for the study’s predefined primary outcome measures can be accessed at https://clinicaltrials.gov/study/NCT04831424.

Sample collection and processing

Blood.

Blood was collected by an intravenous catheter in the antecubital vein at nine different timepoints. The fasting sample was collected in the morning of day 1 (two 8.0 ml serum clot activator tubes (BD-367820)), and eight afternoon samples were collected (eight 6.0 ml serum clot activator tubes (BD-367815)) during the stress reactivity protocol of day 1. Samples were inverted 10–12 times and allowed to rest for at least 30 min before serum isolation. Serum was isolated by centrifugation and aliquoted to be stored at −80 °C until used. Minor deviations between the handling of the morning and afternoon samples can be found in the detailed blood processing protocol¹⁴.

Saliva.

Saliva was collected with a cotton salivette (Sarstedt, 51.1534.500) at nine timepoints, including the day 1 morning sample and the eight afternoon samples collected during the stress reactivity protocol. Collected samples were placed on ice and transported to the lab immediately (morning) or were stored on ice until the end of the protocol (afternoon) and then transferred to the laboratory, where the salivettes were centrifuged at 1,000g for 5 min at 4 °C. The supernatant was carefully aliquoted into 2 ml cryogenic storage tubes and immediately stored in a −80 freezer. Minor deviations between the handling of the morning and afternoon samples can be found in the detailed blood processing protocol¹⁴.

Measures

FGF21.

FGF21 levels were measured in serum using a high-sensitivity ELISA kit (R&D Systems, DF2100) according to the manufacturer’s instructions. Similar lot numbers were used throughout all participant samples, and the average coefficient of variation between lot numbers was determined using reference samples for quality control. Serum samples were diluted with the calibrator diluent provided in the kit at a 1:4 ratio. Absorbance was measured at two wavelengths: 450 nm and 540 nm. Optical imperfections in the plate were corrected by subtracting readings at 540 nm from the readings at 450 nm. Samples were analysed on duplicate plates, with concentrations calculated as the average of the duplicates. A standard curve (five samples per plate) and serum reference samples (three samples per plate, the same sample per batch) were run with each assay, and the inter-assay coefficient of variation was monitored. All standard curves and references were overlaid to detect failed runs. Samples at all concentration readings were included in the final analysis, as the control cohort had low detection rates. For samples with concentrations below the standard curve, adjustments were made by calibrating the values to half of the minimum standard curve concentration. Despite this calibration of low concentrations, the significance of the various analyses performed was not affected. After confirming that adjusting the values for low detection had no impact on the results, all original concentration values were used in the analysis. A potential limitation of this ELISA assay is that it has not been explicitly validated for measuring active and inactivated (carboxy-terminal-cleaved) forms of FGF21, and there were no FAP inhibitors added to the blood to prevent cleavage. As such, our measures may reflect both active and inactive forms of FGF21.

Across the entire study cohort, we had a total of 131 missing samples owing to failed blood draws (for example, veins too small) or technical issues in processing samples. There were 11 missing samples for the fasting timepoint and a total of 120 across the stress time course. For controls, there were five missing samples at both fasting and the baseline stress (−5 min) timepoints. For the MitoD group, there were six missing samples at fasting and seven missing samples for baseline.

Catecholamine hormones.

Norepinephrine, epinephrine, dopamine and serotonin were measured in serum using liquid chromatography–tandem mass spectrometry (UPLC–MS/MS) at the Biomarker Core Laboratory at Columbia University’s Irving Institute for Clinical and Translational Research. The metabolites were extracted from samples spiked with deuterated internal standards by protein precipitation, followed by derivatization with dansyl chloride. The supernatant was then subjected to liquid–liquid extraction using ethyl acetate, and the extracted metabolites were resuspended in acetonitrile for LC–MS analysis. Chromatographic separation was achieved on a Waters ACQUITY UPLC HHS C18 column (2.1 × 100 mm, 1.8 μm) maintained at 40 °C, with gradient elution using water and acetonitrile containing 0.1% formic acid as mobile phases at a flow rate of 300 μl min⁻¹. LC–MS/MS analysis was performed in positive electrospray ionization mode with multiple reaction monitoring (transitions: norepinephrine, 869.2 > 170.3; epinephrine, 883.3 > 170.2; dopamine, 853.30 > 170.37; serotonin, 643.29 > 146.05) on a Waters Xevo TQS MS integrated with the ACQUITY UPLC system (Waters).

Steroid hormones.

Cortisol, cortisone, DHEA and testosterone levels were measured in saliva samples by a high-performance LC–MS/MS using an API 5000 QTrap mass spectrometer (AB Sciex, 6669), following previously established methods^25,26. Previously aliquoted saliva samples were thawed, and 100 μl of saliva was added to a tube containing 50 μl of internal standard and 100 μl of a methanol and water solution with zinc sulfate. The mixture was vortexed and centrifuged, and the supernatant was injected into the LC–MS/MS system for analysis. Cortisol, cortisone and testosterone were detected within a range of 0.001–10 ng ml⁻¹, and DHEA was detected within a range of 0.01–20 ng ml⁻¹, with intra-assay and inter-assay coefficients of variation between 4.3% and 10.8%.

Salivary steroid measurements are widely accepted as a reliable technique because they reflect the bioactive (unbound) hormone fraction, are noninvasive and can be taken repeatedly with minimal participant burden or stress. Early foundational work²⁷ demonstrated that salivary cortisol measurements correlate closely with serum levels, supporting their validity in psychobiological research. This foundational work, and subsequent research, underscored salivary sampling as an equal—if not superior—alternative to blood-based methods, especially in acute or stress-oriented endocrinological studies.

Across the entire study cohort, out of 880 samples for catecholamine and steroid measurements, the dataset included a total of 72 missing samples for cortisol, 73 for cortisone, 267 for testosterone, 547 for DHEA, 445 for epinephrine, 116 for norepinephrine, 116 for dopamine and 116 for serotonin. Data were missing either because saliva or blood could not be collected or because the analyte concentration was lower than the detectable limit.

Metabolic markers.

Glucose, insulin, total cholesterol, high-density lipoprotein, low-density lipoprotein and triglyceride levels were measured from fasting samples by a CLIA-certified diagnostic laboratory at Columbia Presbyterian Hospital according to their standard protocols.

Body composition markers.

The height and weight of participants were measured on the morning of day 1, and BMI was calculated as BMI = weight (kg) / height (m)². The per cent fat mass and fat-free mass were estimated using a four-point bioelectrical impedance device (InnerScan PRO multi-frequency, segmental, wireless body composition monitor; Tanita).

Disease severity markers.

Disease severity and symptom profiles were evaluated using five different methods: (1) the Newcastle Mitochondrial Disease Adult Scale, a validated tool to assess symptoms and functional limitations across various life domains, and (2) the Columbia Neurological Score, used to measure neurological symptoms focused on the central nervous system (both of which are administered by a qualified clinician); (3) the Composite Autonomic Symptom Score, which evaluates autonomic dysregulation across six domains; (4) the 30-s sit-to-stand test to measure functional capacity; and (5) the Modified Fatigue Impact Scale, which is used to measure perceived fatigue and its effects on physical, cognitive and psychosocial functioning²⁸.

Psychosocial self-report measures.

Aspects of participants’ psychosocial lives were measured using a battery of standardized self-report measures, which included the domains outlined below.

Perceived stress and stress exposure.

The dimensions of perceived stress and stress exposure that were assessed included perceived stress, measured with the Perceived Stress Scale²⁹; chronic stress, measured with the Trier Inventory for the Assessment of Chronic Stress³⁰; daily hassles, measured with the Daily Hassles Scale³¹; childhood trauma, measured with the Childhood Trauma Questionnaire³²; life events, measured with the Life Events Questionnaire³³; and lifetime stressor exposure, measured with the Stress and Adversity Inventory for Adults³⁴.

Mental health and wellbeing.

The dimensions of health and wellbeing that were assessed included anxiety, measured with the State and Trait Anxiety Inventory³⁵; depressive symptoms, measured with the Beck Depression Inventory³⁶; burnout, measured with the Maslach Burnout Inventory³⁷; post-traumatic stress disorder symptoms, measured with the Post-traumatic Stress Disorder Checklist for the DSM-5 (PCL-C)³⁸; personal wellbeing, measured with Ryff’s Psychological Well-Being Scale³⁹; aspects of an individual’s sense of coherence, measured with Antonovky’s Sense of Coherence Scale⁴⁰; and daily emotions, measured with the Modified Differential Emotions Scale⁴¹.

Social life.

The dimensions of social life that were assessed included perceived social status, measured with the MacArthur Ladder⁴²; social support, measured with the Social Support Questionnaire⁴³ and the Multidimensional Scale of Perceived Social Support⁴⁴; couple’s satisfaction, measured with the Couples Satisfaction Index⁴⁵; and loneliness, measured with the UCLA Loneliness Scale Version III⁴⁶.

Statistical analysis

The study was powered based on our primary study outcomes (not on FGF21), and final sample sizes were constrained by recruitment constraints of the MitoD groups. All hypothesis tests were two-sided and non-parametric. Normality and equal variances were not formally tested; individual data points are shown to display distributions. Spearman’s rank correlations were computed on raw values with exceptions for catecholamines and steroid hormones, and metabolic biomarkers, which were log transformed. Group differences between controls and MitoD cohorts, stratified by sex, and individual group differences between the control, mutation and deletion groups were evaluated using the non-parametric Mann–Whitney U-test and Kruskal–Wallis test, followed by Dunn’s multiple-comparison test for post hoc analysis when appropriate. Differences between two timepoints in the same individuals (for example, morning to afternoon, day 1 fasting to fed) were analysed using the non-parametric Wilcoxon signed-rank test for paired data.

Stress reactivity was defined as either a relative or an absolute change from baseline at each timepoint. To determine whether acute psychosocial stress affected serum FGF21 levels over time, and if these results differed between groups (MitoD vs controls), we used a linear mixed-effects model to test the effects of group, time and the group × time interaction. Given that the assumption of sphericity was violated, the Geisser and Greenhouse correction was applied. Post hoc Dunnett’s tests were performed for multiple within-group comparisons to test FGF21 stress reactivity within each group. Hedges’ g was used as a measure of effect size and was computed as the standardized mean difference between baseline and the peak-response timepoint, with baseline values normalized to mean = 0 and s.d. = 0, and the standard deviation of the individual peak change scores used as the denominator and group-level average change and standard deviation across participants used as the numerator. Correlations, Wilcoxon tests, the mixed-effects model and Dunnett’s tests were performed using GraphPad 10.0 (GraphPad Software).

Associations between FGF21 levels and psychosocial self-report measures were computed using linear regression models in RStudio (v.2023.12.1 + 402). An interaction term of the respective psychosocial self-report measures with group (MitoD vs controls) served to compute the disease group-specific associations between FGF21 and the psychosocial self-report measures. All models were adjusted for the covariates age and per cent fat because of their relevant influence on FGF21 (Extended Data Fig. 9a,b). The adjusted bivariate association between FGF21 and the self-report measures was computed and tested for significance per disease group using simple slope analyses (R package emmeans), reporting Cohen’s r as a standardized effect size measure.

FGF21 and psychosocial associations from the UK Biobank plasma proteome atlas

To compare our MiSBIE results with FGF21 findings from the atlas of plasma proteins (https://proteome-phenome-atlas.com) generated from the UK Biobank, we examined 2,920 plasma proteins in relation to 986 health-related traits in 53,026 individuals. Regression models were constructed using the ‘lm’ and ‘glm’ functions from the ‘stats’ package in R (v.4.2.0) and the ‘polr’ function from the R package ‘MASS’. Associations between plasma proteins and binary health-related traits (for example, exposures) were assessed with logistic regression, and associations with ordered categorical traits were assessed using proportional odds logistic regression models. All regressions were performed with the adjustment of participants’ baseline information of age, sex, ethnicity, Townsend deprivation index, BMI, smoking status, fasting time, season of blood collection (summer–autumn (June to November) vs winter–spring (December to May)) and blood age (date of blood collection to date of protein examination). Significant associations were defined with a threshold of P_adj < 1.71 × 10⁻⁸ (P < 0.05 / (2,920 × ~1,000 traits)) to account for multiple testing. Adjusted P values, β, 95% CIs and sample sizes are reported in the text.

Extended Data

Extended Data Fig. 1 |. FGF21 and receptor-complex expression across the human body.

Mean transcript levels (normalized Transcript Per Million, nTPM) for (a) FGF21 together with its (b) high affinity receptor FGFR1 (all isoforms) and its obligate co-receptor (c) β-Klotho (KLB) expression across 48 human tissues, from the GTEx v8 RNAseq dataset. Datapoints represent the mean from 21 to 803 individuals per tissue. Tissues are grouped and color-coded by their systemic functions and location. The five tissues with the highest and lowest expression are annotated. See Supplemental Table 2 for corresponding values. Two-sided Wilcoxon signed rank test showed that liver FGF21 expression is 104.2-fold higher (p < 0.0001) than the second highest tissue, skeletal muscle, and 123.7-fold higher than the average of all human tissues. Note: The relatively uniform expression of FGFR1 (all isoforms) and KLB across vascular, digestive, and glandular tissues, along with two brain regions is consistent with circulating FGF21 having a broad signaling in human physiology.

Extended Data Fig. 2 |. Morning to afternoon dynamics of FGF21 across groups.

(a) Two-sided Mann-Whitney’s U-test comparing fasting and fed FGF21 levels (pg/ml) between Controls (n = 65) and MitoD subgroups (3243A>G, n = 18; MELAS, n = 3–4; Deletion, n = 12). In the AM–Fasting state, the 3243A>G group was 0.8-fold higher (M = 427.6 pg/ml, P = 0.004) than Controls (M = 235 pg/ml), whereas the Deletion group was 3.6-fold higher (M = 1088 pg/ml, P < 0.0001). In the PM–Fed state, all MitoD groups were significantly higher than Controls (M =109 pg/ml): 3243A>G was 1.8-fold higher (M = 304 pg/ml, P = 0.0005), MELAS was 1.6-fold higher (M = 284 pg/ml, P = 0.006), and Deletion was 8.5-fold higher (M = 1035 pg/ml, P < 0.0001). Error bars represent standard error of the mean (SEM) (b) Hedge’s g comparing MitoD subgroups to Controls at each time point. (c) Fasting (AM) to fed (PM) change in individual participants’ FGF21 (pg/ml) levels with the group difference expressed as percent difference between time points. Two-sided paired Wilcoxon tests showed significant decreases from fasting-AM to fed-PM levels in Controls (n = 61, P < 0.0001, Median: AM = 173.8 pg/ml, PM = 79.08 pg/ml) and the MitoD group combined (n = 30, P = 0.0005, Median: AM = 535.4 pg/ml, PM = 309.1 pg/ml). When split by subgroup, there was a significant decrease in the m.3243A>G point mutation group only (n = 18, P = 0.004, Median: AM = 307 pg/ml, PM = 191.6 pg/ml).

Extended Data Fig. 3 |. ROC Analyses examining FGF21 as a Diagnostic Biomarker for Mitochondrial Disease.

Receiver operating characteristic (ROC) curve analyses for the prediction of mitochondrial disease based on the serum levels of FGF21 measured by enzyme-linked immunosorbent assay. Plots depict percentage sensitivity plotted against percentage specificity to assess diagnostic performance of serum FGF21 in distinguishing MitoD from healthy participants. (a) Two curves comparing the diagnostic performance of fasting (AM) versus fed (PM) measures of serum FGF21, showing that fed (PM) FGF21 shows enhanced performance with AUC = 0.86. (b) Two plots comparing the diagnostic performance of fed and fasting FGF21 for Deletion and Mutation (m.3243A>G + MELAS) groups separately, alongside a groups-combined curve. Curves demonstrate highest discriminative ability for the Deletion group in the fed state (AUC = 0.98). (c) Comparison of FGF21 and GDF15 diagnostic performance in the fed state and (d) the performance of FGF21 and GDF15 combined. The combined curve was generated using individual FGF21 and GDF15 measures and calculating predicted probabilities by logistic regression for each participant measurement at PM-fed time point.

Extended Data Fig. 4 |. FGF21 dynamics stratified by major clinical symptoms in participants with mitochondrial disease.

Two-sided Mann-Whitney’s U-tests comparing (a) baseline (PM-fed, pre-stress) serum FGF21 concentrations in MitoD participants who either do (Symptom +) or do not (Symptom –) carry a given clinical symptom and (b) FGF21 reactivity, expressed as the net percent change from baseline. Bar graphs represent mean FGF21 values. Error bars represent standard error of the mean (SEM).

Extended Data Fig. 5 |. Demographic and metabolic variations in serum FGF21 levels by group: Analysis by sex, age, body composition, and metabolic markers.

(a) Violin plot showing FGF21 (pg/ml) between females and males for each group. Solid line marks median and dashed lines mark first and third quartiles (IQR). A two-sided Kruskal-Wallis test with Dunn’s multiple comparisons revealed no difference in FGF21 levels between females (n = 64) and males (n = 34). (b) Two-sided Spearman’s correlations between age and fasting FGF21 (pg/ml) by group. Error bands represent 95% confidence intervals. (c) Two-sided Spearman’s correlations between fasting FGF21 levels (pg/ml) and different body composition factors in Controls and MitoD participants. Percent body fat was significantly correlated with FGF21 in both groups (Controls, n = 63, P = 0.046, and MitoD, n = 34, P = 0.013), whereas fat mass was only significantly correlated in Controls (n = 63, P = 0.015). Error bands represent 95% confidence intervals. (d) Forrest plot depicting two-sided Spearman’s correlations coefficients between fasting FGF21 and metabolic biomarkers in Controls and MitoD participants; all measurements are Log10 transformed. Circles represent Spearman’s coefficients and error bars represent 95% confidence intervals. In Controls (n = 65), FGF21 was positively correlated with blood glucose (P < 0.0001), total cholesterol (P = 0.016), and insulin (P = 0.026), and negatively correlated with HDL (P = 0.0043). The MitoD group showed no significant correlation with any metabolic biomarkers. P < 0.05 (*), P < 0.01 (**), P < 0.001 (***), P < 0.0001 (****).

Extended Data Fig. 6 |. Mixed-effects models comparing MitoD to Controls over time with individual and group FGF21 trajectories in response to the speech task.

(a) A mixed effects model with Restricted Maximum Likelihood (REML) estimation showed significant effects of Group, F(1,100) = 44.34, P < 0.0001, Time, F(1.2,109.9) = 13.05, P = 0.0002, and a significant Group X Time interaction, F(7, 640) = 12, P < 0.0001, n = 102, indicating that the trajectory of serum FGF21 over time varied significantly in the MitoD group compared to Controls. Using Dunnett’s tests to compare each subsequent time point to baseline (−5min), we found that Controls showed significant differences at 5mins (P = 0.0021, n = 63), 10mins (P = 0.0003, n = 60), 20mins (P = 0.0007, n = 61), and 30mins (P = 0.0014, n = 61), whereas the MitoD group showed significant differences at 10mins (P = 0.0091, n = 33), 20mins (P = 0.006, n = 32), 30mins (P = 0.0004, n = 31), 60mins (P = 0.0013, n = 30), and 90mins (P = 0.0052, n = 28). Each timepoint represents mean FGF21 (pg/ml). (b-c) Group and individual trajectories for each subgroup are show relative to baseline values (Value/Baseline). Each data point represents a unique biological replicate that is derived as an average from two technical ELISA replicates from each participant (unit of study = individual participant). Control = 65 participants, m.3243A>G = 17 participants, Deletion = 12, MELAS = 3. Relative change between time points of the same group was calculated as the value at a given time point divided by the baseline value (−5min). For each group, paired Wilcoxon tests were used to compare the FGF21 peak to its own baseline measurements. Error bars represent standard error of the mean (SEM). (b) Average percent change in FGF21 (pg/ml) in Controls as a group, as well as individual trajectories for each participant. Individual timepoint represent mean FGF21 (pg/ml). Error bars represent standard error of mean (SEM). (c) Average percent change for each MitoD subgroup relative to their baseline, as well as individual trajectories for each participant. Individual time point represent mean FGF21 (pg/ml). Error bars represent standard error of mean (SEM). P < 0.05 (*), P < 0.01 (**), P < 0.001 (***), P < 0.0001 (****).

Extended Data Fig. 7 |. Stress-induced FGF21, GDF15, and cortisol trajectories in the MiSBIE Study.

Schematic depicting distinct average trajectories of FGF21, GDF15, and cortisol as percent change from baseline after the Tier Social Stress Test (TSST) in Controls (Healthy Mitochondria) vs MitoD (Mitochondrial Diseases). Preliminary data is available at ClinicalTrials.gov #NCT04831424.

Extended Data Fig. 8 |. Psychosocial stress enhances correlations between FGF21 and disease severity indices in patients with MitoD.

(a) Scatter plots illustrating two-sided Spearman’s correlations between FGF21 and composite scores of various disease severity indices at fasting (AM) and post-stress (+90min). Error bands represent 95% confidence intervals (b) Spearman’s correlation matrix showing two-sided correlations between serum FGF21 levels across time points and subscales of disease severity indices. NMDAS = The Newcastle Mitochondrial Disease Adult Scale.

Extended Data Fig. 9 |. Differences in Controls based on FGF21 stress-reactivity.

(a) Controls classified as having a net positive or net negative FGF21 stress-reaction based on percent Area Under the Curve (AUC). AUC calculations were derived from baseline corrected values expressed as percent change relative to baseline FGF21, with both positive and negative peaks included in this calculation, resulting in a net percent AUC for each participant. Scatter plots show two-sided Spearman’s correlations between FGF21 reactivity and (b) cortisol reactivity (P = 0.011, n = 64), (c) cortisone reactivity (P = 0.0049, n = 55), (d) baseline testosterone levels (P = 0.019, n = 48), (e) percent body fat (P = 0.0005, n = 61), and (f) fasting insulin levels (P = 0.0069, n = 64). Reactivity for cortisol and cortisone was calculated as Percent Reactivity = ((Max_Reactivity_Value - Baseline_Value) / Baseline_Value) * 100; Baseline testosterone = measurements (pg/ml) at −5min of the TSST time course; Fasting insulin = measurements (U/ml) at the fasting (AM) time point. Bar graphs for two-sided unpaired Mann-Whitney’s U tests with error bars representing standard error of the mean (SEM) show that Controls with positive reactivity have (b) 80% higher mean cortisol reactivity (P = 0.026, n = 44 vs 17), (c) 30% higher mean cortisone reactivity (P = 0.0023, n = 37 vs 18), and (e) 6.6% higher mean percent body fat (P = 0.0056, n = 42 vs 19). (d) Inversely, Controls with higher FGF21 reactivity had 17.61 pg/ml lower mean testosterone at baseline (P = 0.0066, n = 34 vs 14). (f) There was no statistically significant difference between high and low FGF21 reactors on baseline insulin (P = 0.16, n = 45 vs 19).

Extended Data Fig. 10 |. Forrest plots of fasting and fed FGF21 levels with psychosocial self-report measures.

Forrest plots illustrating associations between psychosocial self-report measures categorized as (a) positive or (b) negative and fasting or fed FGF21 levels adjusted for age and percent fat. The group-specific associations were computed in linear regression models adjusting for age and group-specific percent fat. Circles represent the size of the adjusted bivariate association between FGF21 and the self-report measures as Cohen’s r value with 95% CI error bars. Fasting is presented on the left and fed on the right, with the comparison showing that the differences between Controls and MitoD are more apparent in the fed state. They are organized from smallest (bottom) to largest (top) association for the MitoD group in the fed state. Diamonds represent the overall Cohen’s r for each group represented as mean and standard error of the mean (SEM). P < 0.05 (*), P < 0.01 (**), P < 0.001 (***), P < 0.0001 (****).

Supplementary Material

Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s42255-025-01388-6.

Data availability

All source data for the analyses in this paper are available for download. Source data for steroids and catecholamines is available at http://osf.io/pc9f7. MiSBIE study information, detailed protocols and procedures are available at www.picardlab.org/MiSBIE. Source data are provided with this paper.

Code availability

Code used for the analyses is available at https://github.com/mitopsychobio/FGF21_MitoD_Kurade.git.