Main

Exposure to early-life adversity (ELA) is a risk factor for behavioral and emotional problems in childhood as well as long-term health consequences1. Neuroimaging studies provide a wealth of evidence for the association of perinatal adversity with neurodevelopmental outcomes2,3. These findings include alterations in structure and connectivity in brain regions that are implicated in common mental disorders. The association of ELA with structural and functional development of the brain4 is now thought to represent adversity-related adaptations rather than stress-induced damage. ELA serves as a signal of the prevailing environmental conditions that influence the pace of brain development as an adaptive response to match the demands of unfavorable developmental conditions, at the potential cost to adult well-being5. In a compromised environment, accelerating development to achieve independence may be prioritized over extended neuroplasticity that benefits the development of higher brain function6. The ‘stress acceleration’ hypothesis proposes that exposure to ELA accelerates development, especially in fear/stress-related domains and emotion circuits7. Rodents raised in stressful environments show accelerated development of fear learning and memory retention8,9. In humans, children exposed to maternal distress and deprivation show adult-like limbic brain features (for example, larger amygdala volumes, functional connectivity patterns typically observed in adults)10,11,12. Moreover, children exposed to early-life stress show accelerated biological age measured via either telomere length or DNA methylation-derived epigenetic age13,14.

The current literature on ELA and neurodevelopmental trajectories is limited by a lack of longitudinal neuroimaging data during childhood that are required for a within-subject assessment of the developmental acceleration hypothesis. Most adversity-related studies with large neuroimaging datasets are retrospective cross-sectional studies where the neuroimaging data are collected in adults reporting on adverse childhood experiences (ACE). Therefore, developmental trajectories are not assessed directly but are inferred based on adult data. In addition, existing neurodevelopmental cohorts (for example, the Adolescent Brain Cognitive Development (ABCD), Human Connectome Project Development (HCP-D) and Pediatric Imaging, Neurocognition, and Genetics (PING) studies) cover large age ranges with few subjects below age 7 years, possibly due to the challenges of collecting high-quality neuroimaging data in preschool children. Thus, there is a critical gap in the literature on neurodevelopmental trajectories from early to late childhood, making it difficult (1) to assess the effect of high ELA exposure on brain development and (2) to identify sensitive time windows for intervention during childhood.

Recent studies on neurodevelopmental trajectories have focused on the correlation between structural and functional connectivity, that is, structure–function coupling (SC–FC), as a measure to capture changes in brain organization and maturation. Between the ages of 8 and 22 years, SC–FC changes in a functional-network-specific manner, with decreases in highly conserved motor regions and increases in transmodal cortices15. In adults, SC–FC is highest in unimodal cortical regions and lowest in transmodal cortices16. Whereas high SC–FC implies functional communication supported directly by local white-matter pathways, low SC–FC suggests that functional communication relies on polysynaptic indirect pathways (circuit-level modulation) and a greater capacity for plasticity15. The extent of SC–FC is linked to behavioral outcomes, such as executive function and higher cognitive abilities17,18,19. Notably, significant SC–FC in the reward network is associated with poor performance on executive function tasks in later childhood19. Therefore, SC–FC captures information on age and the current state of plasticity (highly conserved unimodal regions versus transmodal regions with increased potential for plasticity)16.

Current SC–FC studies focus on adolescents and adults, such that SC–FC trajectories have not been studied in children. Studies of ELA will predict outcomes apparent in early childhood, consistent with early biological embedding of experience. To examine this hypothesis, we leveraged data from the deeply phenotyped Growing Up in Singapore Towards healthy Outcomes (GUSTO) birth cohort (Table 1 and Extended Data Fig. 1) to investigate whether or not neurodevelopment, estimated by SC–FC averaged across the whole cortex, is accelerated in children exposed to high ELA. We focused on measures of prenatal adversity (Supplementary Section 1.1) based on previous neuroimaging studies showing associations with brain structure at birth20,21,22 as well as analyses revealing the fetal neurodevelopment period as a time of peak expression of genes associated with a range of neuropsychiatric disorders23. Despite the evidence of prenatal adversity as a risk factor for psychopathology, it is still unclear how prenatal adversity influences the pace of neurodevelopment in childhood. We first modeled SC–FC changes over three time points (ages of 4.5, 6 and 7.5 years) in our full neuroimaging cohort, which is a mixture of distinct and repeated samples (Fig. 1 and Supplementary Section 1.2). Our three time points correspond to two standard childhood phases—preschool and mid-childhood24. We then stratified our cohort via groups with exposure to different levels of adversity (no adversity, low adversity and high adversity) and compared the SC–FC trajectories between these groups. Next, we investigated whether or not the trajectory differences were specific to transmodal association regions with protracted development. We specifically examined the frontoparietal network (FPN) and the visual network (VIS), functional networks that represent transmodal regions and unimodal regions, respectively25. We hypothesized that accelerated development would be observed in the high-adversity group, which would be specific to transmodal association regions. Finally, we performed exploratory analyses studying whether or not SC–FC at our earliest time point (age 4.5 years) altered a child’s susceptibility to developing psychopathologies, that is, it modulated the effects of adversity scores on behavioral outcomes in childhood.

Table 1 Summary of demographics for the neuroimaging dataset and the dataset with complete neuroimaging and adversity scores data
Full size table
Fig. 1: Study design and aims.
figure 1

Regional SC–FC values were calculated for a 114-region cortical parcellation through the Spearman correlation of structural connectivity (SC) and functional connectivity (FC) values of each region (colors are arbitrary and for qualitative purposes). Whole-cortex SC–FC was calculated by averaging values across all 114 regions. Network-specific coupling was calculated by averaging values across regions assigned to a network. Cumulative adversity scores were calculated based on seven components. This enabled us to model non-linear SC–FC trajectories (solid black line ± 95% confidence interval) over three time points (age 4.5, 6 and 7.5 years; age range, 4.4–8 years), as well as stratified by exposure to adversity (‘No’ (no adversity), score 0, N = 121, total of 206 scans; ‘Low’ (low adversity), score 1–2, N = 199, total of 353 scans; ‘High’ (high adversity), score >3, N = 34, total of 60 scans). The Child Behavior Checklist (CBCL) was administered at age 7 years to detect internalizing and externalizing problems. Brain network images were made using the ggsegYeo2011 package73 (https://doi.org/10.5281/zenodo.4896734).

Full size image

Results

SC–FC trajectories over childhood and stratified by ELA

Resting-state functional MRI (rs-fMRI) and diffusion tensor imaging data from 549 participants (Table 1) were included in the current study, for a total of 917 scans over the three time points (age range, 4.4–8 years). For each participant, structural connectivity (indexed by streamline densities) and functional connectivity matrices were estimated for each pair of a 114-region cortical parcellation. Whole-cortex SC–FC was calculated by deriving the Spearman correlation between non-zero structural and functional connectivity values for each region, and taking the mean of all 114 regions. Whole-cortex SC–FC trajectories, modeled via generalized additive models (GAMs), decreased linearly from age 4.5 to 7.5 years (Fig. 2a; effective degrees of freedom (e.d.f.) = 1.006, F = 114.9, P < 0.001).

Fig. 2: Modeling SC–FC trajectories with GAMs.
figure 2

a,In the left panel, SC–FC (N = 549, total of 917 scans) averaged across the whole cortex, modeled over age 4.5 to 7.5 years using GAMs, is shown as the mean estimate (purple line) ±95% confidence intervals (purple shading). Individual trajectories based on actual data across the three time points are plotted in black. The right panel displays the GAM-estimated additive effect of age (solid line ± shaded 95% confidence interval). Black markings on the x-axis show individual age data points. b, SC–FC trajectories stratified by ELA scores (N = 354, total of 619 scans; categorized as no adversity, low adversity and high adversity) displayed on the same scale. c,d, Trajectories estimating the effect of the time point displayed separately for each adversity group (c) and the difference curves graphically showing the differences between each pair of trajectories (d). Trajectories are considered to be significantly different if the confidence interval does not include zero (dashed red line). GAM results for b–d are displayed as the mean estimates (solid lines) ±95% confidence intervals (shaded areas).

Full size image

Prospective data collection of pregnancy and early-life measures were used to calculate the ELA exposure in a subset of participants (N = 354, total of 619 scans). When the cohort was stratified by ELA exposure (Fig. 2b,c), we observed a linear decrease in SC–FC averaged across the whole cortex for the no-adversity and low-adversity groups (e.d.f. = 1, F > 30, P < 0.001). A curvilinear decrease was observed only for the high-adversity group (e.d.f. = 1.8, F = 9.6, P < 0.001; Supplementary Section 2), suggesting that SC–FC decreased at a different rate between 4.5 and 6 and between 6 and 7.5 years of age.

To compare the trajectories, we plotted the difference curves for each pair of trajectories (Fig. 2d). The trajectories for the no-adversity and low-adversity groups were similar. The difference curves that compare the no- or low-adversity groups with the high-adversity group have similar shapes, with a negative value (a greater decrease for high adversity relative to low adversity) between ages 4.5 and 6 years, and a positive value (a lesser decrease for high adversity relative to low adversity) between ages 6 and 7.5 years. The greatest difference was observed between the low- and high-adversity trajectories.

Modeling using GAMS suggested that there was a greater decrease in SC–FC in the high-adversity group relative to the low-adversity group during early childhood. We tested this statistically using linear mixed effects (LME) models (ages 4.5 to 6; N = 251, total of 343 scans) and found that AdversityLow:Age was significantly higher relative to AdversityHigh:Age (Fig. 3a; estimate = 0.024, s.e.m. = 0.012, t = 2.05, P (one-tailed) = 0.022; Supplementary Section 3a) on SC–FC averaged across the whole cortex.

Fig. 3: Model estimates from LME models.
figure 3

a,b, An accelerated decrease in SC–FC between ages 4.5 and 6 years in the high-adversity group (N = 25, total of 34 scans) relative to the low-adversity group (N = 143, total of 197 scans) was observed for the whole cortex (P = 0.022) (a) and for the FPN (P = 0.007) (b) but not for the VIS (P = 0.438) (b). Results were estimated using LME models, and one-tailed P values reported are for the AdversityLow:Age interaction term. In a, the bar charts show model estimates of the mean ± s.e.m. In b, the boxplots are displayed as the Tukey’s five number summary (the bold horizontal line denotes the median; the lower and upper hinges denote the first and third quartiles, respectively; the whiskers extend to the furthest data point within 1.5 × the interquartile range; and the dots exceeding the whiskers denote outliers). c, DNA methylation-based epigenetic age acceleration (N = 241) at age 6 years is significantly higher in the high-adversity group (N = 25) relative to the low-adversity group (N = 132; P = 0.01). The P values (one-tailed) reported were estimated using a linear regression model. Data in the bar charts show the mean ± s.e.m. of the actual data. No multiple comparison corrections were performed. For all plots, *P < 0.05, **P < 0.01. The brain network images in b were made using the ggsegYeo2011 package73 (https://doi.org/10.5281/zenodo.4896734).

Full size image

We next explored whether or not this finding was different in specific functional networks (Fig. 3b). LME models were performed specifically for the VIS and the FPN, which represent a unimodal sensory network and a transmodal network that matures slowly, respectively. We found that AdversityLow:Age was significantly higher relative to AdversityHigh:Age for the FPN (estimate = 0.033, s.e.m. = 0.013, t = 2.51, P (one-tailed) = 0.007) but not for the VIS (estimate = 0.004, s.e.m. = 0.023, t = 0.16, P (one-tailed) = 0.44), suggesting a higher susceptibility of transmodal brain networks. To confirm this finding, we performed LME models (1) without covariates to obtain unadjusted estimates and (2) with potential confounders (Supplementary Section 3b). We also performed LME models for the remaining networks, and observed a similar accelerated decrease for the default mode network (DMN) and salience network, both of which are transmodal networks (Supplementary Section 3c and Extended Data Fig. 2).

Validation

Epigenetic age

Given that SC–FC decreased over our time period of interest, the accelerated decrease between ages 4.5 and 6 years observed for the high-adversity group suggests accelerated neurodevelopment. To validate our finding, we assessed age acceleration at age 6 years based on DNA-derived methylation epigenetic clocks (N = 241). We found that age acceleration was significantly higher in the high-adversity group relative to the low-adversity group (Fig. 3c; estimate = 0.28, s.e.m. = 0.12, t = 2.36, P (one-tailed) = 0.0095). We observed a similar pattern to SC–FC where the greatest difference was found between the high- and low-adversity groups, whereas the low- and no-adversity groups were similar to each other.

Sensitivity analysis

We replicated our LME findings in a small independent dataset (30 scans) (Supplementary Section 4a). Given the confounding effects of motion on neuroimaging data, especially in young children, we performed sensitivity analyses on an extended dataset excluding subsets of scans based on different motion criteria. The main findings remained unchanged (Supplementary Section 4b). SC–FC estimates were also assessed to be reliable for our quantity of available data (Supplementary Section 4c), and with strict motion parameters (Supplementary Section 4d).

Association with behavioral outcomes

The ELA scores were positively and significantly correlated with child behavioral problems at age 7 years (N = 427). High adversity scores were associated with an increased risk of both internalizing (Pearson’s r = 0.22, P < 0.001) and externalizing behaviors (Pearson’s r = 0.20, P < 0.001).

We then examined whether or not SC–FC would moderate the association of ELA scores with the CBCL at age 7 years (N = 117). We used SC–FC averaged across the whole cortex at age 4.5 years as the moderator variable. The interaction between SC–FC and adversity scores had a significant association with both internalizing (estimate = 0.23, s.e.m. = 0.086, t = 2.66, P = 0.0089) and externalizing behaviors (estimate = 0.28, s.e.m. = 0.086, t = 3.31, P = 0.0013). Simple slope analyses revealed that a significant positive correlation between adversity scores and externalizing/internalizing behavior was only present when SC–FC was high at age 4.5 years (Fig. 4a,b).

Fig. 4: Adversity scores, whole-cortex SC–FC and the CBCL.
figure 4

a,b, A significant interaction between adversity scores and whole-cortex SC–FC (N = 117) was observed for internalizing behaviors (a) and externalizing behaviors (b) assessed by the CBCL. Model estimates were obtained from linear regression models. c, Summary of the coefficients of variables selected by the elastic net regression models. The direction of modulation, positive or negative, is displayed as green or red, respectively. VAN, ventral attention network; SOM, somatomotor network; LIM, limbic network.

Full size image

To explore network-specific modulation, we examined the interaction between SC–FC for seven functional networks and adversity scores. The dorsal attention network (DAN) interaction term is highlighted as having the highest estimated coefficient, suggesting a strong positive modulatory effect, especially on internalizing behavior (Fig. 4c).

Similarly, we found network-specific modulatory effects on relational aggression and frustration levels (Supplementary Section 5).

Discussion

Our study modeled SC–FC changes from preschool to mid-childhood. Thus, we address a critical gap in the neurodevelopmental literature, which previously lacked longitudinal neuroimaging data starting from the preschool age. The relevance of these data can be best appreciated when considering the importance of this period of neurodevelopment for a range of socioemotional and cognitive functions26. Our finding of an accelerated decrease in SC–FC for the high-adversity group between ages 4.5 and 6 years suggests this time period as a potential window for intervention to reduce the effects of ELA on later outcomes. In addition, our findings portray the heterogeneity in neurodevelopment over different stages of childhood and highlight the limitations of collapsing children into a single cohort.

We observed that SC–FC changes as a function of age during childhood—specifically, there is a positive association between structural connectivity and functional connectivity that weakens from age 4.5 to 7.5 years. This reflects the diverse processes underlying neurodevelopment that follow different timelines. From the neurodevelopmental literature, the brain undergoes pronounced yearly change during the preschool years26. White-matter tracts show increases in fiber density and bundle size, suggesting a gradual increase in structural connectivity27. By contrast, functional connectivity tends to follow a pattern of overconnectivity, followed by pruning and restructuring to achieve functional network segregation28. In addition, brain SC–FC is thought to be a measure of plasticity where high coupling implies high regional specialization/low plasticity15. During early childhood, low specialization and high plasticity are expected due to the ongoing neurodevelopmental processes. Our understanding of neurodevelopmental processes during early childhood matches the observed decrease in SC–FC. In sum, SC–FC is a promising summary measure for capturing changes in brain connectivity and organization during childhood that can be used to identify abnormalities in developmental trajectories.

We showed that exposure to prenatal adversity is associated with a steeper decrease in SC–FC between ages 4.5 and 6 years, suggesting accelerated development. However, we acknowledge that the steeper slope between 4.5 and 6 years observed in the high-adversity group may be due to the higher SC–FC at age 4.5 years, and could potentially represent ‘catch up’ development to the norm at age 7.5 years. Nevertheless, it is evident that the developmental trajectory of brain maturation in young children is altered as a result of ELA exposure. For instance, exposure to low socioeconomic status (SES) has been linked to changes in cortical thickness and functional segregation5. Cortical thickness peaks at around age 2 years, and exposure to low SES is associated with cortical thinning occurring earlier in life29. These findings are suggested to reflect an earlier curtailment of synaptic proliferation and a decreased window for synaptic pruning. Other studies suggest that higher SES is linked to more protracted functional network development30,31. Tooley et al. suggest a theoretical model where the extended period of structural development associated with high SES is reflected in functional segregation, that is, a slower trajectory of functional network segregation5. Our findings suggest that whereas accelerated development is sufficiently widespread to be detected by averaging over the whole cortex, regions with a longer maturation window are particularly susceptible. Indistinct cortical boundaries of transmodal association networks32 support the hypothesis that these brain networks undergo protracted development and are more susceptible to factors influencing development. Lin et al. postulate that protracted development enables adaptive developmental plasticity in behaviors such as learning33. This understanding suggests that accelerated development occurs in networks such as the FPN and DMN because these regions are important for experience-dependent learning. Our study adds to this literature by showing that accelerated development in the form of a reduced window for neural plasticity—as captured by a steeper SC–FC decrease—in transmodal regions with protracted development may occur during early childhood as a result of exposure to prenatal adversity. This is not surprising as substantial brain development occurs in utero. Moreover, exposure to prenatal adversity is postulated to increase susceptibility to postnatal influences2, highlighting the distinct complementary effects of prenatal and postnatal adversity on neurodevelopment.

The literature on prenatal stress suggests several mechanisms through which the in utero environment could affect downstream development. These include alterations in the neuroendocrine and immune system that have a sustained impact on later life development34. Glucocorticoids and pro-inflammatory cytokines have widespread effects in the brain and have been considered as candidate mechanisms for the effects of prenatal stress on child outcomes35,36,37. These chemical messengers act as signals that influence cell-signaling pathways (for example, regulating enzyme activity), which in turn influence neurotransmission and the proper formation of neural circuits and maintenance.

Contrary to previous literature using the ACE framework, we did not observe a graded response between our three groups. One possibility is that our adversity score was computed with population-based measures (for example, SES or birth weight) as opposed to the typical ACE questionnaire which focuses on childhood maltreatment38. Our adversity score did not include any categories of abuse. Thus, the literature on neglect/poverty and developmental support, which are also highly predictive of child developmental outcomes39, may be more relevant to our study. In addition, Keding et al. show that physical neglect is linked to widespread accelerated maturation, although exposure to abuse in girls is associated with delayed maturation in emotion circuitry40. Although we observed the greatest differences between the low- and high-adversity groups, we did not observe any differences between the no- and low-adversity groups. When we combined the no-adversity and low-adversity groups, we obtained the same results—accelerated decrease between ages 4.5 and 6 years in the high-adversity group relative to the low-adversity group (Supplementary Section 6).

Results from our exploratory analyses showed that a positive association between prenatal adversity scores and both internalizing and externalizing behaviors was only observed for high SC–FC at age 4.5 years. Follow-up analysis suggests that the DAN plays a role in modulating the effect of ELA on externalizing and internalizing behaviors. Herzberg et al. also show circuit-specific adaptations following early-life stress—within-DAN functional connectivity differed between previously institutionalized youths and controls, and was positively associated with both internalizing and externalizing symptoms41. The DAN may also be important for recovery processes after stress exposure, a potential mechanism through which later life behavior is affected42. We also found positive whole-cortex and network-specific modulatory effects with a child-report measure of relational aggression and a task-based measure of emotion regulation (Supplementary Section 5). It may be worthwhile to explore network-specific SC–FC in relation to different dimensions of psychopathologies in future studies.

Some limitations should be considered when interpreting the study findings. First, as our study population is largely a typically developing cohort, the sample size for the high-adversity group was small relative to the low/no-adversity groups. This feature may increase the risk of overfitting curves and limits our ability to stratify further the population by sex, even though sex was a significant covariate in our analyses (Supplementary Section 3a). Our findings also suggest that the associations between adversity and behavioral problems are different at different degrees of SC–FC, but may be under-powered to detect a mediation effect. Second, motion is always a possible confounder for neuroimaging studies, and collecting high-quality neuroimaging data is especially challenging in pediatric populations under age 7 years. We corrected for motion during pre-processing, and also conducted a sensitivity analysis based on motion criteria, and the main findings remain unchanged. Moreover, a relatively short rs-fMRI sequence (∼5.32 min) was used to reduce the burden on young participants. However, the acquisition time of ∼5½ min has been shown to produce stable functional connectivity estimates in children43, as well as stable SC–FC estimates in our study (Supplementary Section 4c). Third, our study lacks imaging data before age 4.5 years; thus we were unable to establish the SC–FC developmental trajectory before age 4.5 years and can only interpret SC–FC changes between the ages of 4.5 and 7.5 years. As our dataset is semi-longitudinal, our trajectory estimates are at the group, not individual, level. Fourth, we focused on prenatal exposures of adversity, but results could potentially be correlated with postnatal adversity. Therefore, our findings do not necessarily inform on the timing of effects. Finally, our study data represent a Singaporean population, and replication in other cohorts is needed to assess the generalizability of our findings.

We present evidence that accelerated neurodevelopment occurs in regions with protracted development after exposure to ELA. This observation is probably an adaptive response to adjust to a suboptimal environment that results in a shortened window for plasticity-related learning. Our findings suggest that the period before the age of 6 years is critical, and they highlight the importance of early detection and intervention to ameliorate the effects of ELA on later life outcomes.

Methods

Methods are described in detail in Supplementary Section 1.

Subjects

Participants were part of the GUSTO study44,45, a longitudinal, Singaporean community-based birth cohort. Neuroimaging data from 549 participants were included, with a total of 917 scans at ages of 4.5, 6 and 7.5 years. A subset of participants (N = 354) had sufficient data for adversity score computation (Table 1). The GUSTO study was approved by the National Healthcare Group Domain Specific Review Board (D/2009/021 and B/2014/00411) and the SingHealth Centralized Institutional Review Board (D/2018/2767 and A/2019/2406). All investigations were conducted according to the principles expressed in the Declaration of Helsinki. Written informed consent was obtained from all guardians on behalf of the children enrolled in this study. Participants received 150 SGD for each MRI session and an additional 120 SGD for questionnaires and laboratory-based tasks. The study followed the STROBE (strengthening the reporting of observational studies in epidemiology) reporting guidelines for cohort studies46.

Adversity score calculation

Adversity scores focused on prenatal exposures and were calculated as described previously47. Adversity scores were based on seven components, and participants scored 1 point for each component if the respective criteria were met, resulting in a maximum score of 7 (Supplementary Section 1.1). On the basis of the ACE literature on cumulative adversity exposure, scores were re-categorized into three groups: no adversity (score 0, 33%), low adversity (score 1–2, 57%) and high adversity (>3, 10%) (Supplementary Section 1.1d).

MRI acquisition and pre-processing

Neuroimaging data were acquired at two sites using 3 T MRI scanners (Siemens). For each subject, diffusion-weighted, rs-fMRI and T1-weighted images were collected.

Diffusion data were processed in FMRIB’s Software Library (FSL v.6.0.4)48. A brain mask was created from the b0 image, underwent motion correction using the eddy tool (with an outlier threshold of 3 s.d.) and was de-noised using a local principal component analysis method49. The rs-fMRI data were processed with the default pre-processing and de-noising pipelines using the functional connectivity toolbox Conn (v.20b)50 as described previously19 (Supplementary Section 1.2).

Regions of interest and connectivity matrices

The regions of interest (ROIs) were 114 cortical regions51 that were assigned to the seven functional networks identified by Yeo and co-workers52. For each scan, functional connectivity matrices were computed by measuring the bivariate correlation coefficients of the BOLD (blood oxygenation level-dependent) time series between each seed and target ROIs through a hemodynamic response factor-weighted general linear model. Structural connectivity matrices were computed by calculating streamline densities between each seed and target ROIs derived from probabilistic tractography using FSL’s BEDPOSTX and PROBTRACKX tools53,54.

Deriving structure–function coupling

The SC–FC was calculated as described in Baum et al.15, where the Spearman correlation between non-zero structural and functional connectivity values was obtained for each region. Whole-cortex SC–FC was calculated by averaging values across all 114 regions. Network-specific coupling was calculated by averaging values across regions assigned to a network. As the first two time points (that is, ages 4.5 and 6 years) were collected at site 1, while the age 7.5 years data were collected at site 2, each region was harmonized across site using longitudinal ComBat (v.0.0.0.90)55 (Supplementary Section 7).

Epigenetic clock

Epigenetic age was used as a validation measure of development. Blood was collected from participants at age 6 years, and was processed to obtain the buffy coat layer for DNA extraction. DNA methylation was profiled using an Infinium MethylationEPIC BeadChip array following the standard protocol14. Wu’s epigenetic clock56,57,58 and age acceleration (adjusted for cell counts, ageAcc3) were computed using the DNAmAge function from the methylclock package (v.0.99.25)59. The Wu clock was chosen as it was trained on both pediatric data and blood samples.

Child behavior outcomes

The Child Behavior Checklist (CBCL)60 is used to detect behavioral and emotional problems in children. The CBCL was administered at age 7 years, and is a maternal-reported questionnaire that categorizes 118 items into internalizing and externalizing behaviors61. The CBCL subscales showed very good reliability in our cohort (Cronbach’s α = 0.855 for the internalizing subscale; Cronbach’s α = 0.888 for the externalizing subscale). Additional dimensions of psychopathology in relation to ELA are described in Supplementary Section 5.

Statistical analysis

All statistical analyses were performed in R (v.4.04)62. Sex was included as a covariate for all models. The alpha level was set at P < 0.05 (two-tailed unless stated otherwise).

Trajectory analysis

GAMs were used to model whole-cortex SC–FC (outcome) over the three time points (age range, 4.4–8.0 years) using the mgcv package (v.1.8-39)63. To model possible non-linear trajectories, a smoothing function ƒ was applied to the age (predictor), where the e.d.f. statistic reflects the degree of non-linearity of a curve64. To account for longitudinal data, we also included a smoothing function f, equivalent to adding a random effect, for each subject65:

$${{\mathrm{SC}}}{\mathrm{-}}{{\mathrm{FC}}}=f\left({{\mathrm{age}}}\right)+f\left({{\mathrm{subject}}}\right)+{{\mathrm{sex}}}+\in.$$

To model the SC–FC trajectories stratified by adversity groups, a second GAM was run with individual trajectories estimated for each adversity group:

$${{\mathrm{SC}}}{\mathrm{-}}{{\mathrm{FC}}}={{\mathrm{adversity}}}+f\left({{\mathrm{age}}},{{\mathrm{by}}}={{\mathrm{adversity}}}\right)+f\left({{\mathrm{subject}}}\right)+{{\mathrm{sex}}}+\in.$$

Comparing adversity trajectories

Adversity trajectories were compared visually with difference curves and statistically with LME models. Difference curves were computed using the gratia package (v.0.7.2)66 to assess differences between each pair of adversity trajectories (for example, no adversity minus low adversity). Trajectories are significantly different if the confidence interval of the difference curve does not include zero.

LME models were performed with the nlme package (v.3.1-160) and were used to assess whether or not the change in SC–FC between ages 4.5 and 6 years differed significantly between adversity groups (reference: high adversity), that is, a significant interaction between the adversity group and the time point67. Given our hypothesis of accelerated development, one-tailed P values were reported for interaction terms.

Validation of accelerated development

Regression analysis was used to compare whether or not epigenetic age acceleration was significantly different between the adversity groups, with high adversity as the reference group. The first three principle components of the genotyped autosomal single nucleotide polymorphisms were used as co-variates to adjust for population stratification as typically performed in analyses involving genetic datasets68.

Associations with behavioral outcomes

Regression models were used to explore whether or not whole-cortex SC–FC modulated the effects of adversity scores on behavioral outcomes. Separate models were run for each measure. Simple slope analyses were conducted using the interactions package (v.1.1.5)69 to analyse the post-hoc differences if the interaction was significant. Elastic net regression was performed, using the caret (v6.0-92) and glmnet (v4.1-2) packages70,71, to explore network-specific modulators of behavioral outcomes. Predictors included adversity scores, seven functional-network-specific SC–FC measures and seven interaction terms (adversity scores × network-specific SC–FC).

Power analysis

The current study is part of an ongoing birth cohort; study objectives are not a primary outcome of the original cohort recruitment. Furthermore, given the longitudinal follow-up involved, the data in our study (especially the first time point of age 4.5 years) were collected several years before our study conception. Therefore, a power analysis was not conducted for this study as it has been suggested that post-hoc power analyses may not be very informative72.

Reporting summary

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