Functional brain network dynamics mediate the relationship between female reproductive aging and interpersonal adversity

Abstract

Premature reproductive aging is linked to heightened stress sensitivity and psychological maladjustment across the life course. However, the brain dynamics underlying this relationship are poorly understood. Here, to address this issue, we analyzed multimodal data from female participants in the Adolescent Brain and Cognitive Development (longitudinal, N = 441; aged 9–12 years) and Human Connectome-Aging (cross-sectional, N = 130; aged 36–60 years) studies. Age-specific intrinsic functional brain network dynamics mediated the link between reproductive aging and perceptions of greater interpersonal adversity. The adolescent profile overlapped areas of greater glutamatergic and dopaminergic receptor density, and the middle-aged profile was concentrated in visual, attentional and default mode networks. The two profiles showed opposite relationships with patterns of functional neural network variability and cortical atrophy observed in psychosis versus major depressive disorder. Our findings underscore the divergent patterns of brain aging linked to reproductive maturation versus senescence, which may explain developmentally specific vulnerabilities to distinct disorders.

Main

Accelerated reproductive aging is linked to risk for psychopathology across the lifespan^1,2,3,4,5,6. This relationship probably reflects the neuroplastic and stress-modulating effects of gonadal hormones, documented most consistently for biological female organisms^3,7,8,9. Indeed, substantial cross-species evidence suggests that ovarian hormones can counteract hypothalamic–pituitary–adrenal (HPA) axis-evoked neuroplastic processes associated with poorer cognitive performance and psychopathology^10,11, including those foreshadowing dementia risk (for example, hippocampal tau accumulation, atrophy and deficient functional connectivity), leading to associative learning and memory deficits^12,13,14.

Although compelling, the evidence considered above comes primarily from the rodent literature and focuses on short-term fluctuations in gonadal hormones. These findings provide only limited insights into the neuroplastic mechanisms underlying the link between stress exposure/susceptibility and faster progression of reproductive maturation (that is, earlier pubertal timing) or senescence (that is, earlier menopausal onset¹⁵). Moreover, in humans, there is compelling evidence of a bidirectional relationship between reproductive aging and stress, as hormonal fluctuations linked to faster progression towards reproductive aging milestones, such as puberty or menopause, enhance stress susceptibility¹⁰, whereas stress exposure accelerates reproductive aging¹. Although accelerated reproductive aging as a putative marker of greater allostatic load heightens the risk for psychopathology across the life course^1,16,17, its immediate adaptive value and brain correlates are likely to vary with age. For instance, in adolescence, accelerated maturation in response to greater ‘wear and tear’ is regarded as an evolutionarily adaptive mechanism because it maximizes the reproductive opportunities for an organism that is likely set for an early demise^18,19. To our knowledge, there is no comparable argument on the adaptiveness of accelerated reproductive senescence. In earlier life, the link between reproductive aging and psychopathology is likely to reflect the neuroplastic effects of gonadal hormones, with precocious reproductive maturation hindering the fine-tuning of the slower developing association networks that are transdiagnostically involved in psychopathology^17,20. Complementarily, in later adulthood, premature reproductive senescence would deprive individuals of the neuroprotective effects of ovarian hormones, thereby increasing susceptibility to stress and dementia^9,13. Understanding the age-specific neurodevelopmental correlates of the relationship between stress and reproductive aging is thus foundational to the design of personalized detection and intervention paradigms intended to alleviate adverse psychological consequences^21,22.

To probe this question, we analyzed multimodal data from female participants in the Adolescent Brain and Cognitive Development (ABCD) and Human Connectome Project–Aging (HCP-A) studies. We focused on female participants for two reasons. One was the availability of well-validated measures indexing stages of reproductive aging based on quantifiable factors (for example, reproductive cycle characteristics, gonadal hormone levels). A second reason was the greater susceptibility and exposure to adversity, including pain and hostility from others, observed in female relative to male individuals^23,24,25. Use of the ABCD and HCP-A datasets allowed us to compare two life stages associated with higher stress vulnerability among female individuals^10,26, but typified by complementary endocrine contexts (that is, increasing versus decreasing ovarian hormone levels³) and thus, probably, differential capacity to withstand adversity²⁷. In line with prior studies^1,28,29, accelerated reproductive aging was estimated separately within each sample as earlier pubertal timing (ABCD) or more advanced menopausal status (HCP-A) relative to other participants of the same chronological age. To operationalize exposure and sensitivity to stress, we examined participants’ experiential ratings of pain and interpersonal adversity (that is, hostility from others), which cross-species research has shown to robustly engage the HPA axis and associated neuroimmune responses^{12,27,30,31,32,33,34,35}. By focusing on pain and interpersonal adversity as the stressors of interest, we were able to draw direct comparisons between our present study and previous rodent literature in which the inter-relationships among gonadal hormones, stress exposure, neurocognitive functioning, as well as psychopathology-like behaviors, have been studied most extensively.

In previous investigations, accelerated brain development, estimated either longitudinally or cross-sectionally relative to same-chronological-age peers, was independently linked to greater stress exposure/reactivity^36,37 and faster reproductive maturation^2,38,39. However, evidence of its role in mediating the relationship between stress and reproductive aging is lacking, partly due to the fact that most studies have focused on structural markers of neurodevelopment (for example, cortical thickness and gray matter volume (GMV)^2,4; but see ref. ⁴⁰ for an exception), which may not be as sensitive as measures of brain function for capturing individual differences in cognition, affect and psychopathology^41,42.

We therefore examined two intrinsic (that is, resting state) markers of brain function that reportedly play a key role in adjustment to environmental stressors: variability (that is, within-individual standard deviation) in regional time series (TSV) and moment-to-moment variability of inter-regional functional coupling (FCV). These two markers show divergent associations with measures of adaptive functioning. Specifically, greater TSV is thought to reflect a brain’s wider ‘dynamic range’ and thus its capacity to respond in a more differentiated manner to external stimuli and environmental challenges^43,44,45,46. FCV has been linked to maladaptive entrainment with the external milieu⁴⁷. Greater intrinsic FCV, broadly operationalized as more frequent moment-to-moment changes in functional coupling strength between a given region and the rest of the brain, is regarded as a marker of brain network architectural instability over time^48,49,50.

The two variability indices show distinguishable trajectories of lifespan development. From childhood to middle adolescence, intrinsic TSV shows regionally specific maturational trajectories that track intracortical myelination patterns (that is, declines in sensory region and increases in transmodal association areas⁵¹). Intrinsic FCV declines steadily with development, as individualized motifs of functional network segregation and integration emerge^52,53. From middle adulthood onwards, normative declines in TSV predict longitudinal age-related decrements of cognitive performance⁴³, while advancing age has been associated with rising FCV⁵⁴, which, in turn, foreshadows declining cognitive performance^48,50,55.

Our primary objective was to characterize TSV/FCV patterns suggestive of accelerated/decelerated brain aging relative to same-chronological-age peers³⁷ that could explain the previously documented bidirectional relationship between reproductive aging and stress susceptibility^1,10 (see Table 1 for sample demographics). We defined the latter in reference to the participants’ perceptions of recently experienced pain and exposure to interpersonal adversity. For the adolescent sample of the ABCD Study, the longitudinal design allowed us to investigate the temporal stability of the relationship between brain function and reproductive aging (at Time 1/baseline and Time 2/two-year follow-up, respectively) and the potential prospective prediction of Time 2 measures of pain/hostility from the Time 1 brain measures.

Our second objective was to elucidate the broader functional relevance of variations in reproductive and brain aging. We thus sought to shed light on whether our identified TSV/FCV patterns could be relevant to understanding vulnerability to psychiatric disorders typified by premature neural senescence⁵⁶. To do so, we examined the overlap between the spatial topography of functional brain variability tracking reproductive aging versus symptom severity in two psychiatric disorders associated with accelerated brain aging⁵⁶ and with adolescent onset⁵⁷, psychosis and major depressive disorder (MDD)^{58,59,60,61,62} (see Table 2 for psychiatric sample demographics). Extant evidence suggests that enhanced vulnerability to psychopathology during key female hormonal transitions, such as puberty and menopause, may reflect, at least partly, the modulatory effects of gonadal hormones over four neurotransmitter systems (that is, dopaminergic (DA), serotoninergic (5-HT), glutamatergic (GLU) and gamma-aminobutyric acid-related (GABA)) that play a critical role in mental functions relevant to coping with stress (Fig. 1h)^10,27,63,64. Importantly, these four neurotransmitter systems are also critically implicated in both psychosis^{65,66,67,68,69,70,71,72,73} and MDD^{74,75,76,77,78,79,80}. Using a multimodal approach, we sought to uncover the neurochemical correlates of the overlap in the reproductive aging and psychosis/MDD-relevant brain maps with the goal of identifying potential targets for future pharmaceutical interventions targeting neuropsychiatric symptoms linked to accelerated reproductive aging. Our conceptual framework is represented schematically in Fig. 1.

Fig. 1: Schematic representation of our conceptual model.

a–f, Premature reproductive aging in the form of earlier pubertal onset (a) or faster progression towards menopause (b) was hypothesized to be related to patterns of intrinsic variability in time series (TSV) (c) or functional coupling (FCV) (d), suggestive of accelerated maturation/senescence, which would, in turn, predict greater perceptions of hostility (e) and/or pain (f). The dashed red lines in a depict potential trajectories of accelerated reproductive maturation or aging. g, Using the Schaefer brain atlas¹⁶⁵, the spatial profiles of TSV/FCV linking reproductive aging to stress susceptibility were further examined for their similarity with the atrophy, brain variability and chemoarchitectural fingerprint of disorders typified by accelerated brain maturation/senescence. Schaefer networks: TP, temporo-parietal; SAL-VAN, salience/ventral attention; LB, limbic; DMN, default mode; DAN, dorsal attention; SM-A, somatomotor-A; SM-B, somatomotor-B; VIS, visual. ^aPlot in a and b adapted with permission from ref. ²¹¹ under a Creative Commons license CC BY 4.0. Whole-body silhouettes in a, b, e and f are from vecteezy.com.

Results

The results reported in the following were replicated with the Gordon atlas⁸¹ (Supplementary Information) and their specificity was supported by supplemental tests featuring alternate graph measures sensitive to maturation/aging (Supplementary Fig. 4).

Accelerated reproductive aging is linked to greater FCV

ABCD

The full-sample discovery partial least-squares (PLS) analysis linking Time 1 and Time 2 TSV and FCV, respectively, to aging as well as perceived exposure to pain and interpersonal adversity, identified a sole significant brain-behavior latent variable (LV) pair (P = 0.017; shared variance, 37%). Across the ten test folds, the extracted LV pair was found to reflect the Time 2 link between accelerated reproductive aging and greater FCV (r = 0.11, permutation-based P = 0.029, 95% confidence interval (CI) = [0.01; 0.20]) (Fig. 2a). This effect was widely expressed and peaked in the visual (VIS), salience/ventral attention (SAL/VAN), dorsal attention (DAN), default mode (DMN) and control systems (Fig. 2b,c shows the cross-validation and Fig. 2d,e the discovery PLS results).

Fig. 2: The brain LV from the behavioral PLS analysis linking reproductive aging, perceived hostility and pain to TSV and FCV in the ABCD sample.

a, The height of the colored bars indicates the correlation coefficient between the behavioral variables and the predicted brain LV scores (based on a cross-validation procedure) in each condition across all 441 ABCD participants. Error bars, represented as black blocks, constitute the 95% CI values estimated by re-running each correlation on 100,000 bootstrap samples. The individual data points represent the correlation coefficient obtained with each of the 100,000 bootstrap samples. CIs that do not include zero reflect robust correlations between the respective behavioral variable and the predicted brain LV score in a given condition across all participants. The asterisk marks the statistically significant correlation between FCV and accelerated reproductive aging at Time 2 (based on permutation testing as described in the main text). b, The Schaefer ROIs that are (1) robustly correlated (based on cross-validated 99% CIs, as described in the main text) with the brain LV in a and (2) have an absolute value BSR of at least 2 in the discovery PLS analysis (d). These are partial correlations controlling for the confounders listed under ‘Control variables’ in Methods. d, The ROIs that have an absolute BSR value of at least 2 and are thus considered to load reliably (at ~95% CIs, based on 100,000 bootstrap samples) on the brain LV identified with the discovery PLS analysis in the full sample (N = 441). c,e, To facilitate interpretation, the Schaefer network-level representations of the ROI-specific results from b and d, respectively. All bootstrap samples were obtained by sampling with replacement from the original set of 441 ABCD participants. BSR, bootstrap ratio.

HCP-A

A similar pair of brain variability-behavior LVs emerged from the full-sample discovery PLS analysis conducted separately in the HCP-A data (P = 0.046; shared variance, 34%). As in the ABCD dataset, the extracted LV pair reflected the robust link between accelerated reproductive aging and FCV, a relationship that was cross-validated over all ten test folds, (r = 0.20, permutation-based P = 0.024, 95% CI = [0.03; 0.35]; Fig. 3a). Thus, greater FCV related to reproductive aging was widespread and most clearly detected in the VIS, DAN, DMN and temporo-parietal (TP) systems (Fig. 3b,d).

Fig. 3: The brain LV from the behavioral PLS analysis linking reproductive aging, perceived hostility and pain to TSV and FCV in the HCP-A sample.

a, The height of the colored bars indicates the correlation coefficient between the behavioral variables and the predicted brain LV scores (based on a cross-validation procedure) in each condition across all 130 HCP-A participants. Error bars, represented as black blocks, constitute the 95% CIs estimated by re-running each correlation on 100,000 bootstrap samples. The individual data points represent the correlation coefficient obtained with each of the 100,000 bootstrap samples. The asterisk marks the statistically significant correlation between FCV and accelerated reproductive aging (based on permutation testing as described in the text). CIs that do not include zero reflect robust correlations between the respective behavioral variable and the predicted brain LV score in a given condition across all participants. b, The Schaefer ROIs that are (1) robustly correlated (based on cross-validated 99% CIs, as described in the main text) with the brain LV in a and (2) have an absolute value BSR of at least 2 in the discovery PLS analysis (d). These are partial correlations controlling for the confounders listed under ‘Control variables’ in Methods. d, The ROIs that have an absolute BSR value of at least 2 and are thus considered to load reliably (at ~95% CIs, based on 100,000 bootstrap samples) on the brain LV identified with the discovery PLS analysis in the full sample (N = 130). c,e, To facilitate interpretation, the Schaefer network-level representations of the ROI-specific results from b and d, respectively. All bootstrap samples were obtained by sampling with replacement from the original set of 130 HCP-A participants.

FCV mediates the perceived hostility–reproductive aging link

To test whether the sample-specific profiles of FCV mediate the link between accelerated reproductive aging and perceived exposure to pain/interpersonal adversity, we specified the mediational model represented in Fig. 4. Profiles of functional brain variability were estimated separately within each sample as the weighted sum of all region-of-interest (ROI) contributions in the FCV condition. We focused on the FCV condition in the HCP-A data and the Time 2 FCV condition in the ABCD sample, because these were the only conditions in which we observed robust brain–reproductive aging correlations in the ten test folds (Figs. 2a and 3a). The PLS-extracted profiles of FCV mediated the relationship between reproductive aging and sensitivity to hostility (total standardized indirect effect: 0.014, 95% CI = [0.003; 0.034]), but not pain (total standardized indirect effect: 0.004, 95% CI = [−0.005; 0.017]) (Fig. 4). None of the direct or total effects (reproductive aging–perceived hostility/pain) was statistically significant (all P > 0.33), a pattern of results suggestive of an indirect relationship only, which is currently well-accepted in the literature^82,83.

Fig. 4: Mediational model linking reproductive aging to perceived hostility and pain via variability in functional brain organization across the ABCD and HCP-A samples.

The PLS-extracted profiles of FCV mediated the relationship between reproductive aging and sensitivity to hostility, but not pain. ^aP = 0.004. ^bP = 0.008. ^cP = 0.41. ^dP = 0.0001. All P values are two-tailed. Whole-body silhouettes are from vecteezy.com.

Age-related FCV maps, chemoarchitecture and neuropathology

Given the weak association between the confounder-controlled ROI-brain LV correlations estimated in the two samples, separate canonical correlation analyses (CCAs) were conducted for the FCV patterns extracted from the ABCD versus HCP-A data.

In each sample, the discovery CCAs identified a single mode that linked FCV to the chemoarchitecture and disorder-specific neuropathology across all ten test folds (cross-validated r values of 0.50 (ABCD) and 0.66, both P = 1 × 10^−5; Fig. 5a,c). The FCV profile associated with accelerated reproductive maturation in the ABCD sample (Figs. 2b and 5b) showed robust positive correlations with the psychosis-, rather than the MDD-relevant FCV and atrophy maps (Fig. 5b,e,f), as well as with areas of higher 5-HT, DA (D1/D2) and GLU receptor density (Fig. 5g–k). Complementarily, the FCV profile related to faster reproductive senescence in HCP-A (Figs. 3b and 5d) was positively associated with the MDD-, rather than the psychosis-relevant FCV and atrophy maps (Fig. 5e,f), but showed robust negative correlations with the 5-HT, DA (D1/D2) and GLU receptor density maps (Fig. 5g–i,k). Put differently, greater FCV linked to faster reproductive aging was observed in areas of lower 5-HT, DA and mGLU5 receptor density.

Fig. 5: Chemoarchitectural correlates of the reproductive aging–hostility link.

Disorder-specific GMV loss/FCV and receptor density maps linked by CCA to the brain variability profiles observed in the ABCD and HCP-A samples, respectively (cf. Figs. 2 and 3). a,c, Scatter plots describing the linear relationship between predicted values of the brain and the disorder-specific GMV loss/FCV/receptor density CCA variates across the 300 ROIs from the Schaefer atlas for the ABCD (a) and HCP-A (c) samples. b,d, Correlations between the observed disorder-specific/receptor density maps and the predicted value of their corresponding canonical variate across all test CCAs for the ABCD (b) and HCP-A (d) samples. Error bars, represented as black blocks, constitute the 99% CIs estimated by re-running each correlation on 100,000 bootstrap samples. The individual data points represent the correlation coefficient obtained with each of the 100,000 bootstrap samples. Colored columns correspond to variables with robust loadings on their variate in both the Schaefer and Gordon atlas data (based on the bootstrapping-derived 99% CIs). The bootstrap samples were obtained by sampling with replacement from the 300 Schaefer ROIs. e–k, MDD/psychosis difference GMV loss (e) and FCV (f) maps, as well as the 5-HT (g), D1 (h), D2 (i), NMDA (j) and GluR5 (k) receptor density maps. NMDA, N-methyl-d-aspartate.

Fig. 6: STROBE flowchart describing sample selection and outlining the main analytic steps.

The ABCD and HCP-A data were downloaded as part of a package with the data reported in ref. ¹⁶⁶ and ref. ¹⁵⁵, respectively. The rationale for doing so was twofold: (1) allow cross-study comparisons and (2) include participants who provided good quality data across multiple contexts (in particular, for ABCD, given the age of the sample).

Table 1 Demographic information for the ABCD and HCP-A samples

Table 2 Demographic information on the datasets contributing to the FCV and disorder-specific atrophy maps

Control analyses

A series of partial correlation analyses investigated whether the PLS-extracted age group FCV LV (Supplementary Fig. 7a,b,d,e), the ROI-specific intrinsic functional coupling (FC), estimated separately within each sample (Supplementary Fig. 8), and the complementary reproductive aging brain LV (that is, the ABCD brain LV for the HCP-A sample; the HCP-A brain LV for the ABCD sample) would impact the identified CCA modes (Fig. 5a,c and Supplementary Fig. 6a,c) or the structure of the corresponding cross-validated canonical loadings (Fig. 5b,d). With the exception of the neuropathology GMV variable in the ABCD sample, these additional controls did not impact any of the CCA results reported above (partial r values for the cross-validated CCA modes from 0.38 to 0.44, all spin test P = 1 × 10^−5; Supplementary Fig. 9).

Discussion

Although independently associated with stress exposure/reactivity^36,37 and earlier reproductive maturation^2,38,39, the role of neurodevelopment in explaining the link between the two has been challenging to prove. Our investigation (see Fig. 6 for STROBE flowchart) takes a first step towards addressing this issue by demonstrating that the relationship between faster reproductive aging and perceptions of greater interpersonal adversity is mediated by patterns of intrinsic variability in functional brain organization (that is, FCV), spanning slower to faster intrinsic timescale networks^84,85. The identified FCV patterns showed age-group specificity with regard to spatial extent and composition, as well as developmental significance. In adolescence, these FCV patterns were widely expressed and suggestive of delayed brain development relative to same-chronological-age peers^52,53, results that resonate with recent evidence linking a younger estimated brain age to greater psychiatric symptom burden in the transition from childhood to young adulthood⁸⁶. Of note, a robust relationship between accelerated reproductive maturation and greater FCV emerged only at Time 2, plausibly because the corresponding trending association observed at Time 1 was partly obscured by the functional brain network ‘immaturity’ (that is, higher FCV) of the entire ABCD sample at this earlier assessment. In middle adulthood, the FCV patterns linking earlier reproductive senescence and perceptions of greater interpersonal adversity implied steeper brain aging^48,50,54 and were observed primarily in the VIS, DMN and attentional networks, all of which are particularly vulnerable to aging-related functional dedifferentiation processes^50,87. Importantly, we only found evidence of a robust indirect positive relationship between reproductive maturation/senescence and perceived hostility, suggesting that it is the emergence of the age-group-specific FCV profiles that may underpin the bidirectional link between aging and social stress sensitivity^1,10.

The involvement of somatomotor (SM) areas in the maturational and aging brain profiles dovetails with evidence on its role in tracking dynamic whole-brain state transitions⁸⁸ and its contribution to the stability of FC patterns⁵⁴. In line with the documented relevance of neuronal hyper-excitability and intrinsic FC, broadly, as well as SM FC, specifically, to MDD^89,90,91,92 and psychosis^69,72,93,94, we report a robust, yet age-specific, association between the identified FCV patterns and normative profiles of FCV and GMV loss in MDD and psychosis, respectively. In adolescence, the FCV patterns mediating the link between reproductive maturation and perceived interpersonal adversity showed a positive correlation with the FCV map tracking psychosis, rather than MDD, symptom severity, as well as the associated GMV loss map. Thus, slower functional brain development^52,53 may represent one mechanism through which accelerated reproductive maturation and perceived interpersonal adversity increase vulnerability to psychosis, a disorder related to faster structural, but not functional brain aging^58,62. While delayed emergence of cognitive control skills, and thus poorer stress-coping resources^95,96, arise as primary candidates for explaining the psychological implications of the above inter-relationships, more-in-depth studies on the associations between reproductive aging, perceived hostility and neurodevelopment are warranted.

In contrast, the middle-aged FCV profile mediating the relationship between earlier reproductive senescence and perceived hostility overlapped areas of normative MDD-related GMV and regions showing FCV patterns that tracked MDD symptom severity. Dovetailing with evidence on the relevance of SM, VIS and attentional networks to the pathophysiology of MDD^91,92, our results imply that accelerated aging of the aforementioned systems constitutes one mechanism through which earlier reproductive senescence and interpersonal adversity may heighten risk for MDD in later life. This interpretation is consistent with extant theory and evidence linking neurocognitive aging to a declining ability to engage strategically with the external environment in the here and now and thus a likely poorer capacity to cope effectively with novel challenges in the external milieu^97,98.

The more robust molecular correlates emerged for the adolescent FCV profile and spanned several neurotransmitter systems modulated by gonadal hormones and implicated in functional network dynamics, stress reactivity and psychosis-related pathophysiology^{10,69,73,99,100}. Echoing earlier studies on the modulatory effect of estradiol on dopaminergic systems^64,101, the FCV patterns associated with earlier reproductive maturation overlapped with areas rich in D2 and, to a lesser extent, D1 receptor density. These effects are consonant with the well-documented relevance of D1/D2 receptors to stress susceptibility¹⁰², psychosis pathology (D2 (ref. ⁷³)) and their differential involvement in maintaining (D1) versus updating (D2) mental representations in the service of adjustment to environmental challenges^103,104. The overlap between the FCV profile related to faster reproductive maturation and regions of higher GLU receptor density fits nicely with the putative role of excitation/inhibition imbalance in psychosis^68,72 and the involvement of mGLU5 in responses to social stress³⁴. Broadly, the chemoarchitectural findings underscore the potential wide-ranging impact of earlier progression towards a key female hormonal transition. They thus dovetail nicely with epidemiological evidence that adolescence is a key period for the emergence of multiple psychiatric disorders⁶¹. This interpretation fits well with the corresponding neuropathological results, because risk for psychosis relates to multiple neuropsychiatric comorbidities⁵⁷.

More generally, if accelerated reproductive maturation or senescence is experienced as an internal stressor (for example, an inner alarm clock signaling the potential for an earlier demise¹⁹), then our findings could speak to extant stress-related theories¹⁰⁵. The FCV patterns linking faster reproductive aging to greater perceived hostility could reflect an adaptation to chronic stress¹⁰⁶ through which an internal stressor is projected onto the external social environment to alleviate its threat value (that is, the negative affective experience stemming from perceptions of faster reproductive maturation/aging is misattributed to perceptions of greater interpersonal adversity). Alternatively, the observed FCV patterns could reflect a sensitizing mechanism that drives reciprocal interactions between internal and external stressors (that is, perceptions of accelerated reproductive maturation/aging increase vigilance to signs of interpersonal adversity and vice versa). Targeted interventions that address perceptions of biological aging, their associated interpretations (with regards to long-term survival) and perceptions of social support availability may help mitigate these effects.

To disentangle the functional significance of the FCV patterns associated with accelerated reproductive maturation/senescence from the significance of those linked to typical age-group differences, we compared the ABCD and HCP-A samples (Supplementary Fig. 7). These analyses revealed that FCV patterns typical of middle adulthood (rather than adolescence) correlate positively with the FCV and GMV maps tracking psychosis (rather than MDD) severity (Supplementary Fig. 7c,f). Thus, in line with epidemiological evidence of higher MDD prevalence in adolescence and of longitudinal increases in MDD adolescent cases^10,107, our results suggest that, for typically developing female individuals, early teenage years may be linked to greater vulnerability to MDD, rather than psychosis spectrum disorders. Complementarily, our results suggest that typical middle-aged women may be more likely to develop psychosis spectrum disorders (rather than MDD). This effect could stem from age-related vulnerability to attentional problems (Supplementary Fig. 10a,b), which, in turn, increases risk for multiple psychiatric and neurological disorders with which psychosis shares a substantial number of brain pathways¹⁰⁸. This conjecture is compatible with the observed chemoarchitectural map correlated with age-group differences in FCV, which spanned neurotransmitter systems relevant to cognitive control and responses to stress (GABA^109,110), memory persistence and flexibility^103,111, learning (NMDA¹¹²) and broader psychiatric vulnerability and (un)successful aging (5-HT^113,114) (Supplementary Fig. 7c,f).

Limitations and future directions

Our investigation has several limitations that pave the way for future research venues. First, its primary cross-sectional design precludes any conclusions regarding the causality of the reported effects, a caveat that needs to be addressed in carefully controlled longitudinal investigations and/or cross-species experimental manipulations.

Second, due to data availability, our study drew on predominantly Caucasian samples. The neuropathological fingerprint of common psychiatric disorders shows some variability with ancestry, comorbidity with medical conditions and/or country-level socioeconomic conditions¹¹⁵. Consequently, future studies with larger and more diverse samples could elucidate the boundary conditions of the effects reported here.

Third, our study focused on the adverse outcomes of accelerated reproductive senescence^116,117,118. Nonetheless, certain (epigenetic) aging patterns (that is, DNA methylation at certain promoter regions) can have a neuroprotective effect¹¹⁹. Characterizing these aging patterns and their susceptibility to environmental modulation, including perinatal factors (for example, maternal depression¹²⁰) would be worth pursuing in the future.

Fourth, our research probed the relevance of variability in resting-state functional brain dynamics. However, cognitive traits, including those implicated in stress responses, show context-specific associations with brain dynamics¹²¹. Hence, use of a variety of task paradigms (for example, naturalistic movie viewing¹²², acute stress manipulations) and multiple neural dynamics indices^123,124 could enhance the applicability of the present results and shed light on the specific mechanisms linking accelerated senescence to pain versus hostility sensitivity (cf. refs. ^32,125).

Fifth, we failed to find an association between TSV and either accelerated reproductive aging or perceived stress. Associations between resting-state TSV and cognitive performance or stress exposure have been documented in older samples, both from the earlier and later life stages^43,51. Alternatively or additionally, the link between TSV and stress-relevant variables may be most likely to emerge in paradigms assessing blood-oxygen-level-dependent (BOLD) signal dynamics on externally oriented attentional tasks, because higher TSV has been shown to foster finer-grained perceptual processing⁴⁴. Therefore, future investigations including more age-varied samples and assessing TSV during both rest and externally oriented processing tasks could characterize its relevance to accelerated aging and stress susceptibility.

Sixth, our investigation focused on relatively stable patterns of neurobiological aging. However, perceived exposure and susceptibility to stressors is also modulated by reversible brain-aging factors (for example, sleep deprivation^126,127), whose short- versus long-term impact on brain-network organization and dynamics is worth exploring.

Seventh, because biological aging is a multimodal process, use of multiple indicators beyond reproductive senescence¹²⁸ could further elucidate shared mechanisms underpinning stress susceptibility, connectomic maturation/senescence and cognitive functioning at different life phases (cf. ref. ¹²⁹). This research could also pinpoint indicator-specific sequelae, such as those linked to the accumulation of senescent cells and subsequent chronic neuroinflammation, mediated by microglial and astroglial activation, which hinders synaptic plasticity and brain-network functioning¹³⁰. Use of multiple aging indicators, including those indexing neuroinflammation, would further facilitate investigations into the sex specificity of the associations among biological senescence, dynamic connectomics and stress vulnerability. Such studies would build on extensive evidence of sex differences in brain–immune system interactions^131,132, brain network organization¹³³, genetically modulated risk for stress-related disorders^134,135 and responses to pain/hostility^{7,136,137,138}.

Eighth, although we focused on two key female hormonal transitions, mounting evidence attests to the long-term neurobiological impact of an additional pivotal hormonal event, specifically, pregnancy and motherhood¹³⁹. For instance, pregnancy is linked to changes in intrinsic functional network organization, reduced brain inflammation and increased neurogenesis^140,141, which, in the long term, may foreshadow a younger brain age and foster higher cognitive reserve^139,142. Most relevant to the present study is the finding of post-pregnancy changes in the relationship between cycling hormones and psychological stress responses, which probably reflects alterations of GABA-ergic signaling and HPA-axis output¹⁴³. Hence, pregnancy may impact the inter-relationships among reproductive aging, perceived stress and intrinsic FCV, a possibility worth exploring in future research.

Finally, FCV did not mediate the reproductive aging–pain relationship. However, pain is multifaceted and best viewed through a biopsychosocial lens¹⁴⁴, as its physical features (for example, location—single site versus multi-region) interact with psychological and sociocultural factors to shape experience¹⁴⁵. Given the high prevalence of chronic pain in both adolescent¹⁴⁶ and middle-aged²⁴ samples, more fine-grained assessments of pain, including location and duration of each episode, could advance understanding of its relevance to neurobiological aging and stress susceptibility¹⁴⁷.

Conclusions

Using cross-sectional data primarily, we have documented age-specific profiles of variability in functional brain organization that mediate the relationship between accelerated reproductive aging and perceptions of interpersonal adversity in adolescence and middle age. The two profiles indicated divergent changes in aging-related trajectories (that is, delayed in adolescence, but accelerated in middle adulthood) and showed opposite relationships with patterns of functional neural network variability and cortical atrophy observed in psychosis versus MDD. Our findings speak to the importance of brain dynamics in mediating the relationship between reproductive senescence and perceived stress, while also underscoring its age-specific role in alleviating versus enhancing risk for psychological maladjustment.

Methods

Participants

Both samples contributed complete data on all the analyzed variables (one time point for the HCP-A sample; baseline/Time 1 and two-year follow-up/Time 2 data for the ABCD sample). Inclusion of the imaging data was guided by recommendations of the ABCD and HCP-A teams, respectively. Because pubertal timing, including its association with mental versus cardiometabolic health and, to a lesser extent, menopausal onset, vary with racial background^{148,149,150,151}, and there were insufficient participants to allow well-powered analyses of racial differences, the samples were exclusively (ABCD) white—as reported by the children’s parents—or predominantly (62%, HCP-A) white, as self-reported by the participants. This choice is justified, as one of the comparison clinical groups (that is, MDD) with relevant data was also exclusively white. Participants in both samples were biologically unrelated to one another (that is, they did not share a family ID). Table 1 summarizes the demographic information for both the ABCD and HCP-A samples. The demographic information for the MDD and psychosis samples is included in Table 2. Figure 6 contains the STROBE flow diagram detailing participant selection for our present study.

ABCD

The present research uses baseline and two-year follow-up data downloaded in November 2021 as part of the ABCD Study Curated Annual Release 4.0 (https://data-archive.nimh.nih.gov/abcd). The non-imaging variables were screened for completeness. Inclusion of the imaging data was guided by recommendations of the ABCD team. Specifically, global information about the data recommended for inclusion by the ABCD team was recorded as a binary decision (0 (‘no’)/1 (‘yes’)) in the ‘abcd_imgincl01’ file, which was part of Annual Release 4.0. Run-specific information was recorded as ‘fail/pass/unspecified/no value’ under ‘qc_outcome’ in the file ‘fmriresults01’. We only used data from participants who had passed the qc for six resting-state runs (three runs at baseline and two-year follow-up, respectively).

All participating parents provided informed consent, and the participating children provided assent in accordance with the guidelines of the local research ethics review boards across the 21 participating sites (that is, Children’s Hospital Los Angeles, Florida International University, Laureate Institute for Brain Research, Medical University of South Carolina, Oregon Health & Science University, SRI International, University of California San Diego, University of California Los Angeles, University of Colorado Boulder, University of Florida, University of Maryland at Baltimore, University of Michigan, University of Minnesota, University of Pittsburgh, University of Rochester, University of Utah, University of Vermont, University of Wisconsin-Milwaukee, Virginia Commonwealth University, Washington University in St Louis and Yale University). Participants were compensated for time and travel expenses. More specific compensation details have not been publicly released.

HCP-A

Participants were aged 36 to 60 years, an interval that captures the transition from reproductive to late post-menopausal status. Exclusion criteria included sensory (hearing/vision) deficits, uncontrolled high blood pressure, major organ failure and a Montreal Cognitive Assessment (MoCA) score of 19 or lower¹⁵².

All participants provided informed consent in accordance with the guidelines of the local research ethics review boards across the four participating sites (that is, University of California Los Angeles Medical School, University of Massachusetts Medical School, University of Minnesota Medical School and Washington University at St Louis). Participants were compensated for time and travel expenses up to US$400. More specific compensation details have not been publicly released.

Reproductive aging

ABCD

Reproductive aging was assessed with the five-item Pubertal Development Scale (PDS), which is significantly correlated with other indices of pubertal development, including physician ratings¹⁵³. The questionnaire comprises three gender-general items (growth spurt, changes in skin and hair growth) and two gender-specific items (breast development and menarche (girls)). The instrument, which uses a four-point Likert-type response format, ranging from 1 (no development) to 4 (development already completed), was completed separately by the youth and their guardian/parent. Because youth ratings showed poorer reliability at baseline (Cronbach’s alpha of 0.54), we opted to use parental ratings at both baseline and two-year follow-up (Cronbach’s alpha of 0.65 and 0.78 at baseline and two-year follow-up, respectively). For participants with available hormonal data, we verified that parental PDS ratings showed the expected positive correlation with pubertal hormones, both at baseline—r(423) = 0.27, P < 0.001 (testosterone), r(414) = 0.28, P < 0.001 (dehydroepiandrosterone (DHEA)), r(409) = 0.09, P = 0.06 (estradiol)—and at the two-year follow-up—r(317) = 0.25, P < 0.001 (testosterone), r(313) = 0.29, P < 0.001 (DHEA), r(299) = 0.17, P = 0.003 (estradiol).

An index of accelerated maturation relative to same-chronological-age peers was computed at each time point by regressing from the PDS score the youth’s chronological age, such that a positive residual score indicated accelerated reproductive aging^1,29 (Supplementary Fig. 1a presents the distribution of accelerated maturation scores across the ABCD sample). This residual score showed robust positive associations with contemporaneous levels of DHEA and testosterone across both time points, (r values from 0.22 to 0.28, all P < 0.001). However, a significant relationship between accelerated reproductive maturation and estradiol levels emerged only at the two-year follow-up (r(299) = 0.17, P = 0.003).

HCP-A

Responses on the self-report survey of the Stages of Reproductive Aging Workshop (STRAW-10¹⁵⁴) were used to estimate menopausal status. We used the same strategy as in ref. ¹⁵⁵ to convert STRAW codes to menopausal status. In the 110 participants with available data, more advanced menopausal status correlated with higher levels of follicle stimulating hormone (r(107) = 0.81, P < 0.001) and lower levels of estradiol (r(107) = −0.41, P < 0.001) after controlling for whether blood was collected after fasting¹⁴⁹.

Reproductive aging was estimated by regressing out chronological age from the menopausal status index described above. Positive residual scores reflected accelerated reproductive aging, and negative scores decelerated reproductive aging (Supplementary Fig. 1b presents the distribution of accelerated aging scores across the HCP-A sample). As expected, this residual score correlated positively with follicle stimulating hormone levels (r(107) = 0.32, P < 0.001), but not estradiol levels (r(107) = −0.05, P > 0.60), potentially because the latter is more strongly modulated by chronological age-related processes relevant to menopause. Finally, we verified that age at menarche, a significant predictor of age at menopause, was not significantly associated with this index of reproductive aging (r(128) = −0.08, P = 0.34).

Perceived stress

None of the perceived stress measures described below had a cutoff for severe or clinically significant pain experience or perceived hostility because they were all primarily intended to assess individual differences in healthy individuals. Two-independent-samples Mann–Whitney U-tests confirmed that the distribution of neither experienced pain (standardized test statistic 0.27, P = 0.79) or perceived hostility (standardized test statistic 0.46, P = 0.64) is statistically different in the ABCD versus HCP-A samples (details on the measures are presented in the following).

Pain experience

ABCD

The two-year follow-up was the earliest time point at which participants’ exposure to pain in the month preceding their study participation was assessed with four items from an adjusted version of the Seattle Child And Adolescent Pain questionnaire¹⁵⁶, which gauged presence/absence of pain (0/1), highest pain intensity (10-point Likert-type scale from ‘1 not at all’ to ‘10 the worst’), typical duration of a painful episode (four-point Likert-type scale from ‘1 less than an hour’ to ‘10 all day’) and the extent to which pain hindered engagement in typical activities (10-point Likert-type scale from ‘1 not at all’ to ‘10 stopped me from doing anything’). This revised four-item scale showed good reliability (Cronbach’s alpha of 0.78). Based on the observed scores (Supplementary Fig. 1), ~50 participants reported having experienced some pain during the month preceding their study participation, with most pain episodes being rated as moderate to severe (Supplementary Fig. 1c), although relatively short (Supplementary Fig. 1d) and with reduced impact on daily activities (Supplementary Fig. 1e).

HCP-A

Participants’ exposure to pain during the preceding week was quantified through their responses to two National Institutes of Health (NIH) Toolbox measures: the one-item Pain Intensity Survey (‘1 no pain’ to ‘10 worst pain imaginable’) and one item from the Pain Interference Survey, which assesses the extent to which pain interfered with daily activities (‘1 not at all’ to ‘5 very much’). Data unavailability prohibited use of other, domain-specific items from the Pain Interference Survey. However, we think that for the sake of comparability with the ABCD measures, use of the above two items is best advised. Score on the two pain items were strongly correlated (r of 0.74), which is why they were first standardized across participants and then averaged to create a single index of pain exposure/sensitivity. Based on the observed scores, approximately half the sample reported some pain in the week preceding their participation (Supplementary Fig. 1f), which had some impact on their daily activities (Supplementary Fig. 1g).

Perceived hostility

ABCD

At the two-year follow-up, participants completed the 18-item Revised Peer Experiences Questionnaire¹⁵⁷. This measure uses a five-point Likert-type scale (‘1 Never’ to ‘5 A few times a week’) to gauge the extent to which the respondent had been victimized in the past (for example, ‘Another kid gossiped about me so others would not like me’). The nine victimization-relevant items showed good reliability (Cronbach’s alpha of 0.80) and were averaged to create an index of perceived hostility/interpersonal adversity. Although the perceived hostility scores were relatively low (Supplementary Fig. 1h), ~75% reported having experienced some form of victimization in the past.

HCP-A

The eight-item Perceived Hostility Survey from the NIH Toolbox gauged participants’ perceptions of having been treated poorly by people in their lives in the month preceding their study participation. The measure uses a five-point Likert-type scale (‘1 Never’ to ‘5 Always’) to assess the frequency with which others have criticized, yelled, argued with the participant. The scale showed good reliability (Cronbach’s alpha of 0.91) and the responses to the eight items were averaged to create an index of perceived hostility. Inspection of Supplementary Fig. 1i revealed that 88% of the participants had experienced some interpersonal adversity in the month preceding their study visit.

Functional brain variability

These analyses focused on regional TSV and cross-regional time-series variability (that is, FCV), because these measures are widely used to capture distinct aspects of functional MRI (fMRI) dynamics¹⁵⁸. TSV captures variability in the BOLD signal within an individual brain region, which is thought to reflect that region’s capacity to adjust its activity in response to changes in the external milieu and, thus, represent environmental dynamics more accurately¹⁵⁹. FCV captures variability in pairwise regional BOLD signal correlations and, as such, it is regarded as a marker of (in)stability in functional brain network organization⁴⁸. We selected dynamic, rather than static, functional indices because of their relevance to environmental adjustment, including psychopathology, and sensitivity to broader biological aging processes^{47,48,159,160}. Specifically, past research has shown that, in adulthood, greater TSV predicts superior cognitive performance and treatment outcome in individuals with social anxiety disorder, whereas greater FCV has been linked to higher anxiety and accelerated metabolic aging. Our focus on regional rather than whole-brain measures provides the spatial specificity required for investigating associations with receptor and neuropathological maps.

Resting-state scans

Both samples contributed four resting-state fMRI (rfMRI) scans (eyes open with passive crosshair viewing) for a total of 20 (ABCD) or 26 (HCP-A) minutes at each sampled time point (ABCD, baseline and two-year follow-up; HCP, one time point). For the ABCD sample, due to data availability, we used only three scans from each time point (six resting-state scans in total). In the HCP-A sample, two rfMRI scans were acquired with an anterior-to-posterior (AP) phase encoding sequence and the other two with a posterior-to-anterior (PA) phase encoding sequence. Because comparisons with the ABCD data were based on a single time point, we used only two of the four rfMRI scans so as to make the scan duration comparable across samples. To facilitate potential future comparisons with the HCP-A task fMRI data on inhibitory control, which were collected with a PA phase encoding sequence, we only used the two resting scans acquired with a PA phase encoding sequence.

MRI data acquisition

The ABCD participants were scanned across 21 US sites, with a protocol harmonized for Siemens Prisma, Philips and GE 3T scanners, whereas scanning for the HCP-A sample was performed across four US sites on Siemens Prisma 3T scanners (32-channel coil). For both samples, to account for magnet and sociodemographic differences among sites, we introduced site ID as a covariate in all analyses. A similar multiband acquisition sequence was used for both samples (see refs. ^161,162).

fMRI data preprocessing

ABCD

Our analyses used minimally preprocessed rfMRI data (described above), which was available as part of the ABCD Study Curated Annual Release 4.0. All 441 participants contributed three complete resting-state runs at baseline/Time 1 and the two-year follow-up/Time 2, for a total of six resting-state scans across the two time points (see the ‘Participants’ section for the quality-control checks used in this study).

The main processing steps applied to these data by the ABCD Study team are outlined as follows (for further details on specific steps, see ref. ¹⁶²). The minimal preprocessing pipeline applied to the rfMRI data involved correction for head motion, spatial and gradient distortions, bias field removal and co-registration of the functional images to the participant’s T1-weighted structural image. We further applied the following steps: (1) elimination of initial volumes (eight volumes (Siemens, Philips), five volumes (GE DV25) and 16 volumes (GE DV26)) to allow the MR signal to reach steady-state equilibrium, (2) linear regression-based removal of the mean time courses of cerebral white matter (WM), gray matter (GM) and cerebrospinal fluid (CSF), as well as the quadratic trends and 24 motion terms (that is, the six motion parameters, their first derivatives, and squares from the time course of each Schaefer parcel). Before being regressed, the motion terms were filtered to eliminate signals within the respiratory effect range (that is, 0.31–0.43 Hz, cf. ref. ¹⁶³).

HCP-A

The rfMRI data were processed by applying the Generic fMRI Volume and Surface Processing Pipelines, multi-run independent component analysis (ICA) FIX denoising and multimodal surface-matching registration, as detailed in ref. ¹⁶⁴.

ROI definition

Our main fMRI analyses featured the Schaefer 300 parcel-functional atlas¹⁶⁵, downloaded from https://github.com/ThomasYeoLab/CBIG. We used the 17 functional network version of the atlas, which is available in the Montreal Neurological Institute (MNI) standard space.

ABCD

As in refs. ^166,167, to align the atlas with the participants’ native space for each resting-state run, the following steps were implemented in Functional Magnetic Resonance Imaging of the Brain Software Library (FSL): (1) the middle image in each resting-state run was converted to MNI space (via the MNI-152 brain template available in FSL), and the inverse transformation (MNI-to-participant native space) was estimated; (2) the inverse transformation was used to align the Schaefer atlas to each participant’s functional images, separately for each resting-state run, based on nearest-neighbor interpolation (to keep the integer values of the ROI IDs).

HCP-A

Voxel-wise time series were extracted and averaged within each of the Schaefer ROIs using the ‘cifti-parcellate’ command with the MEAN method in the Connectome Workbench (https://www.humanconnectome.org/software/get-connectome-workbench).

TSV

Within each element of fully preprocessed resting-state run data, ROI-level TSV was computed as the within-individual standard deviation (s.d.) of BOLD signal fluctuations over time⁴⁴. This is the simplest measure of variance-based neural variability¹⁵⁹, which has been robustly implicated in lifespan development⁴³ (for a review of relevant findings, see ref. ¹⁵⁹) and psychopathology^160,168. Importantly, this s.d.-based measure of neural variability is strongly correlated with an alternative deviance-based index of TSV based on mean-squared differences between successive time series^169,170, particularly in datasets (such as the present ones) where preprocessing included the removal of linear and quadratic trends in the BOLD signal.

ROI-to-ROI correlations

Each resting-state run was broken down into 24-s-long (that is, 30 volumes) non-overlapping windows for a total of 33 windows at each time point in the ABCD sample and 30 windows in the HCP-A sample. The window length and type (that is, non-overlapping) were selected with the goal of augmenting sensitivity to patterns of dynamic functional brain reconfiguration and individual differences in such processes. Sliding windows that were ~30 s long met both criteria^171,172,173.

Pairwise Pearson correlations between regional time series extracted from the Schaefer-atlas ROIs were computed separately in MATLAB and expressed as Fisher’s z-transformed scores. As per previous work¹⁷⁴, we retained the positive Fisher’s z-scores and set negative values to zero.

Network-level analyses

The Network Community Toolbox (NCT¹⁷⁵) was used to estimate all network-level metrics.

To characterize patterns of ROI-based clustering in functional community organization during awake rest, we used a multilayer generalized Louvain-like community detection algorithm¹⁷⁶ and implemented it in the NCT^155,166. As in earlier works^{155,166,174,177}, the two free parameters in the modularity quality optimization were set to the default value of 1 (the spatial resolution parameter) and 0.5 (the cross-layer window-to-window connection strength parameter), the latter to to allow heterogeneity in window-to-window mental states¹⁷⁴.

ROI-level variability in functional organization was estimated using the ‘flexibility’ function in the NCT as the number of times a given ROI changed functional communities between two consecutive windows. To account for the near degeneracy of the modularity landscape¹⁷⁸, the multilayer community detection algorithm was initiated 100 times and the ROI-level variability indices were averaged across all iterations.

Specificity analyses

To probe the specificity of any observed relationships between TSV/FCV and reproductive aging, we included four additional functional connectivity indices that fluctuate with age^{179,180,181,182}, but that we hypothesized to be more weakly associated with adjustment to environmental challenges. First, using the NCT, we estimated two ROI-specific dynamic indices: functional network segregation (called ‘recruitment’ in the NCT), operationalized as the number of times a given ROI was assigned to the same community as the other ROIs in its native functional system, versus functional network integration, operationalized as the number of times a given ROI was assigned to the same community as ROIs from different functional systems. In both instances, the native functional system was the one defined in the Schaefer functional atlas. Second, we applied the ‘participation_coef_sign.m’ function from the Brain Connectivity Toolbox¹⁸³ to the participant-specific ROI-to-ROI static correlation matrices. The participation coefficients thus estimated reflect the diversity of an ROI’s connections (that is, the likelihood that an ROI has positive versus negative connections equally distributed across ROIs from all the functional systems defined in the Schaefer atlas). Participation coefficients of higher value indicate a less differentiated pattern of positive and/or negative functional connectivity.

Disorder-specific FCV maps

An overview of the contributing samples is presented in Table 2. Details on participant compensation have not been publicly released for these samples. The FCV and GMV loss maps for psychosis were based on data from the Human Connectome Project–Early Psychosis (HCP-EP), Data Release 1.1. Informed consent was obtained from the participants or their legal authorized representative/guardian in accordance with the guidelines of the local research ethics review boards across the four participating sites (Indiana University, Beth Israel Deaconess Medical Center–Massachusetts Mental Health Center, McLean Hospital and Massachusetts General Hospital).

The FCV data for MDD were taken from the Perturbation of the Depression Connectome (PDC) Project because, to our knowledge, this is the largest publicly available database of depressed patients with high-quality rfMRI data. However, due to a lack of an appropriate healthy control group (the control group for the PDC overlaps with the HCP-A sample), we estimated MDD-specific atrophy maps with data from the Kansas Musical Depression Study^184,185 and the Russia fMRI Depression Study^186,187. Informed consent was obtained from all participants in the MDD studies in accordance with the guidelines of the local research ethics review boards (University of California Los Angeles (PDC), University of Kansas Medical Center (Kansas Musical Depression Study) and the multi-profile clinic Pretor/International Institute of Psychology and Psychotherapy (Novosibirsk, Russia) (Russia fMRI Depression Study)).

FCV maps

The rfMRI data for both the PDC (two resting-state runs, each lasting ~6 min and 30 s) and the HCP-EP (four resting-state runs, each lasting ~5 min and 30 s) were collected with the same parameters as for the HCP-A sample. Half of the runs for each sample were acquired with an AP-phase encoding sequence and the other half with a PA-phase encoding sequence. Our analyses featured both PDC runs and combined data from the HCP-EP to amount to the same duration as in the PDC and with an equal number of scans acquired with an AP- versus PA-phase encoding sequence. The data from both samples were preprocessed with the same pipeline described for the HCP-A sample and run through the same analytic steps described for the ABCD/HCP-A data above, to obtain ROI-specific FCV indices. We focused only on FCV because, of the two brain variability indices scrutinized, it was the only one to show a robust association with reproductive maturation/senescence.

In the PDC, before any treatment intervention, depressive symptom severity was assessed with the Hamilton Depression Rating Scale¹⁸⁸ (HDRS), which is one of the most widely used clinician-administered rating scales for primary depression. In the HCP-EP, psychosis severity was measured with the Positive and Negative Syndrome Scale¹⁸⁹ (PANSS). Disorder-specific FCV maps were estimated by computing the Spearman’s partial correlation between the total HDRS score at baseline (MDD) or the total PANSS score (psychosis) and ROI-specific FCV indices across all participants, controlling for age, average relative scan-to-scan displacement, gender (both samples) and, additionally, in the HCP-EP only, also site, race (white versus non-white) and antipsychotic medication dosage. The two disorder-specific maps were standardized, then a difference map was created by subtracting the psychosis–FCV map from the MDD–FCV map. Thus, in the MDD/psychosis difference map, higher values typify ROIs that are relatively more sensitive to MDD symptoms, and lower values typify ROIs relatively more sensitive to psychotic symptoms.

Relevance to subclinical depression scores

To test whether the MDD/psychosis difference map could differentiate between subclinical variations in depressive versus non-depressive symptoms, we focused on the 440 ABCD and 130 HCP-A participants included in our main analyses who had available data on the DSM-oriented depression, anxiety, attention deficit hyperactivity disorder (ADHD) and antisocial personality and somatic disorder from the Child Behavior Checklist¹⁹⁰ and Adult Self-Report¹⁹⁰. We thus took the following steps: (1) separately, within each sample, using the confounders listed under ‘Control variables’, we computed partial Spearman’s correlations between each ROI’s FCV and DSM-oriented scores on depression, anxiety, ADHD, antisocial personality disorder and somatic disorder; (2) the resulting five DSM disorder–FCV partial correlation maps for each sample were standardized and then averaged for the corresponding disorders across the two samples; (3) a CCA with tenfold cross-validation probed the association between the MDD-psychosis–FCV difference map and the five DSM disorder–FCV partial correlation maps averaged across the ABCD and HCP-A samples.

The results of the above-mentioned CCAs, which are depicted in Supplementary Fig. 10, suggest that FCV differences linked to subclinical depression differentially map onto the MDD–FCV, rather than the psychosis–FCV, map. Complementarily, subclinical variations in anxiety and ADHD symptoms mapped onto the psychosis, rather than the MDD, FCV map. Although the MDD/psychosis difference map seems capable of differentiating between depressive and non-depressive symptoms, we note that the goal of the neuropathology CCA analyses presented in the main text was to investigate whether the brain differences linked to faster reproductive aging and perceived hostility/pain would resemble the brain profiles observed in MDD and psychosis, suggestive of likely future risk for these disorders, but not necessarily linked to the current presence of symptoms relevant to either disorder.

GMV loss maps

Disorder-specific atrophy maps were created using the exact pipelines described in refs. ^191,192. Each clinical disorder group (MDD, psychosis) was compared against demographically matched healthy controls using a two-sample t-test. Following extant practices in the literature¹⁹³, the group t-maps for each disorder were averaged and parcellated in the MNI-152 space based on the Schaefer/Gordon atlas using the ‘Parcellater’ function from neuromaps¹⁹⁴ (https://github.com/netneurolab/neuromaps) implemented in Python 3. Similar to the FCV maps, the two disorder-specific GMV maps were standardized, then a difference map was created by subtracting the psychosis GMV map from the MDD GMV map. Thus, in the MDD/psychosis difference map, higher values typify ROIs that are relatively more sensitive to MDD symptoms, whereas lower values typify ROIs that are relatively more sensitive to psychotic symptoms.

Receptor density maps relevant to reproductive aging

We analyzed receptor densities for DA, 5-HT, GLU and GABA, because these neurotransmitter systems are under substantial gonadal hormone modulation. We estimated receptor density with the normative atlas established by ref. ¹⁹³, which is based on positron emission tomography (PET) data from over 1,200 healthy individuals. For each DA, GABA, GLU and 5-HT tracer map we downloaded group-averaged preprocessed PET images from https://github.com/netneurolab/hansen_receptors/tree/main/data/PET_nifti_images. The ‘Parcellater’ function from neuromaps¹⁹⁴ segmented each tracer map in the MNI-152 space based on the Schaefer atlas. An adapted version of the ‘make_receptor_matrix.py’ script (https://github.com/netneurolab/hansen_receptors/tree/main/code) from ref. ¹⁹³ was used to estimate the weighted average of the receptor density maps corresponding to DA, GABA, GLU and 5-HT in Python 3. We did this separately for each tracer, taking into account the number of participants in the respective study, as implemented in ref. ¹⁹³.

As in ref. ¹⁹³, group-level estimates of receptor density were obtained for each of the Schaefer ROIs then standardized across all the parcels in the Schaefer atlas. To avoid biasing the canonical correlation analyses results towards the 5-HT density maps, we averaged the standardized scores corresponding to the four moderately correlated 5-HT receptor maps (HT1a, HT2, HT4 and HT6; r values from 0.28 to 0.57), but kept the HT1b density index separate.

Statistical analysis

CCA and PLS^195,196 were selected to probe the relationships among our variables of interest. Given its superior performance in datasets containing highly correlated within-set variables^197,198, PLS was used to test the relationship of reproductive aging and stress susceptibility with ROI-level functional variability (TSV, FCV). In contrast, because of its greater sensitivity to cross-set relationships¹⁹⁸, CCA was selected to characterize the relevance of functional brain variability to both the chemoarchitecture and disorder-specific neuropathology. All CCA and PLS models were cross-validated using a tenfold procedure, then 99% CIs for each correlation coefficient (CCA) and loading/salience (PLS) were obtained by using the ‘bootci’ function in MATLAB (with default settings and 100,000 bootstrap samples).

PLS analysis

The MATLAB PLS toolbox was downloaded from https://github.com/McIntosh-Lab/PLS/. PLS was run separately in the ABCD and HCP-A samples, respectively. In these analyses, the ‘behavioral’ set included the reproductive maturation (ABCD)/aging (HCP-A) measures, as well as self-reported pain and perceived hostility from others (both samples). The brain matrix contained the ROI-specific TSV and FCV indices, aggregated across all available resting-state scans at each scrutinized time point (ABCD: baseline/T1, two-year follow-up/T2; HCP-A: one time point). The discovery PLS analyses in the ABCD sample contained four separate conditions (Time-1 TSV, Time-2 TSV, Time-1 FCV and Time-2 FCV). These conditions were intended to elucidate the temporal stability of the brain–reproductive aging relationship (at Time 1 and Time 2, respectively) and the potential longitudinal relationship between the Time-1 TSV/FCV measures and the Time-2 measures of pain and/or hostility. Due to data availability, the discovery PLS analyses in the HCP-A sample contained only two conditions (TSV and FCV). In addition to the two aforementioned ‘behavioral PLS’ analyses, we ran a group (task) PLS analysis to identify age-group-related differences in brain variability patterns and test whether they could account for the observed relationships between the accelerated reproductive-aging FCV and chemoarchitectural/neuropathological maps. The group PLS analysis only featured the ROI-specific FCV data for which an association with reproductive aging had been identified (cf. Figs. 2 and 3 and Supplementary Figs. 2 and 3).

Significance and generalizability testing

The significance of the extracted brain–behavioral LV pairs was estimated through permutation testing (100,000 permutations, 200 times larger than the guidelines provided by ref. ¹⁹⁹) in the discovery PLS analysis, which was conducted separately in the ABCD and HCP-A samples. The generalizability of our PLS models was tested using a tenfold cross-validation procedure. Specifically, the PLS analysis was run on nine folds of data, and the weights associated with the identified brain LVs were used to compute the predicted value of the brain LV in the test fold. This procedure was repeated until each of the ten folds served as ‘test data’ once. The significance of the correlation between the predicted brain LV score and the behavioral variables with which it was found to be reliably associated (see below) was estimated in the test fold using permutation-based testing with 100,000 samples.

Reliability testing

In the discovery PLS analyses, we used 100,000 bootstrap samples (that is, 1,000 times larger than the guidelines provided by ref. ¹⁹⁹) to (1) identify ROIs making a reliable contribution to the identified LV (that is, BSRs, ROI weight/standard deviation, greater than 2 in absolute value^200,201, which approximate a 95% CI) and (2) estimate 95% CIs for the correlation between the ‘behavioral’ variables and the extracted brain LV¹⁹⁹. In the cross-validation procedure, we used the ‘bootci’ function in MATLAB (with default settings and 100,000 bootstraps) to characterize the (1) 99% CIs of the correlation coefficient corresponding to each ROI-level TSV/FCV and the scores of the associated LV (estimated in the test fold) and (2) 95% CIs for the correlations between the predicted value of the brain LV and the observed value of the variables in the ‘behavioral’ set¹⁹⁹.

Mediation analysis

To test whether patterns of functional brain variability account for the shared variance between accelerated biological aging and stress susceptibility, we combined the data from the two samples and conducted one mediation analysis using generalized maximum likelihood estimation in MPLUS 8.10 (ref. ²⁰²). Accelerated reproductive aging and functional brain variability (that is, FCV, based on the output of the PLS cross-validation analyses; ‘Results’) constituted the predictor and mediator variables, respectively. Susceptibility to pain and hostility, respectively, represented the two parallel outcomes. The mediation model was tested with 95% CIs estimated with 100,000 bootstrapping samples. As recommended in the literature^82,203 bootstrapping-based 95% CIs for the indirect effects were used as effect size estimates. Although we report bias-corrected bootstrapping-based results, we verified that identical indirect effect estimates would be obtained with percentile bootstrap CIs.

CCA

To probe the relationship between functional brain variability and disorder-specific FCV (1 map (MDD–psychosis)) and atrophy (1 map (MDD–psychosis)), as well as receptor density maps (seven receptor density maps), we conducted a series of CCAs with cross-validation¹⁹⁵. CCA was implemented in MATLAB using the ‘canoncorr’ module. As recommended in ref. ¹⁹⁵, we ran all CCAs on sample sizes more than ten times greater than the number of contributing variables. To evaluate the performance of our CCA-derived models, we employed a tenfold cross-validation procedure. For all sets of CCAs, discovery analyses were conducted on nine folds of data and the resulting CCA weights were employed to derive predicted values of the brain variability and disorder-specific atrophy/receptor density variate in the left-out (‘test’) fold. This procedure was repeated until each of the ten folds served as ‘test data’ once. The 10 test folds were concatenated and the correlation between the predicted variates across the full sample was evaluated using a permutation test with 100,000 samples.

We described the relationships between measured variables and their corresponding CCA-derived variate through correlations between the observed value of a given variable and the predicted value of its corresponding variate across all test folds. Standardized canonical coefficients were not included because the unique association between a given variable and its corresponding variate was of limited value in the case of our present analyses.

The 99% CIs for each correlation coefficient were obtained by using the ‘bootci’ function in MATLAB (with default settings and 100,000 bootstrap samples). The CIs were selected to match those used in the PLS analyses, for which clearer guidelines exist (‘PLS analysis’). The aforementioned correlation coefficients are a more conservative estimate of the traditional canonical loadings as they are estimated in the test, rather than discovery, folds.

To account for correlated disorder-specific atrophy/receptor density patterns based on anatomical proximity²⁰⁴, we used Vasa’s ‘rotate_parcellation’ function in MATLAB (https://github.com/frantisekvasa/rotate_parcellation/commit/bb8b0ef10980f162793cc180cef371e83655c505) to generate 100,000 spatially constrained permutations of the Schaefer brain LVs, as identified in the behavioral PLS analyses for the ABCD and HCP-A samples, respectively. These spatially constrained permuted functional-brain LVs were used to assess the significance of the correlation between the Schaefer brain LV and the disorder-specific atrophy/receptor density variates.

Control variables

To maximize the interpretability of the model estimates, the discovery PLS analyses were conducted on data that had not been residualized for the confounders listed below. To demonstrate the robustness of our results, the PLS cross-validation and the mediation analyses controlled for the following variables that were of no interest in the present study: (1) chronological age, to ensure that the reported inter-relationships among accelerated reproductive aging, functional brain variability, pain and hostility hold irrespective of the participants’ chronological age; (2) handedness (coded as ‘0’ for right-handedness and ‘1’ for non-right-handedness), to control for potential differences in the lateralization of the observed effects (for example, refs. ^205,206); (3) serious medical problems based on the ABCD Parent Medical History Questionnaire (ABCD) or self-reported habitual use of medications, lifetime history of traumatic brain injury and body mass index as a proxy for cardiometabolic health (HCP-A), given their potential to influence BOLD signal (variability)^207,208; (4) socioeconomic status based on (family) income-to-needs (dependents) in both samples and years of education (HCP-A only), because it can impact reproductive aging, particularly pubertal timing¹; (5) scanner site, to account for scanner-related differences, as well as broad differences in family education and socioeconomic status across sites²⁰⁹; (6) race (HCP-A only), because it can impact reproductive senescence¹⁴⁹; (7) average modality-specific (that is, resting state) motion per participant, because it can adversely impact functional connectivity metrics, in our case FCV²¹⁰. Thus, the models derived from the ABCD and HCP-A data through the discovery PLS analyses were cross-validated via partial correlation analyses (brain LV–behavior (Figs. 2a and 3a) and brain LV–individual ROIs (Figs. 2b and 3b)), which controlled for the above confounders. Similarly, before the mediation analysis, the variables of interest were residualized for the above confounders separately within each sample.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this Article.