Reproducible Sex Differences in Personalized Functional Network Topography in Youth

TL;DR

This study finds reproducible sex differences in brain network organization in youth, which may help explain why mental health disorders are more common in females.

Contribution

The study identifies robust sex differences in personalized functional brain network topography during the transition to adolescence.

Findings

Sex differences in PFN topography were most prominent in association networks like fronto-parietal and default mode networks.

Machine learning models accurately classified sex based on PFN topography patterns.

These differences are reproducible across large youth samples and may relate to female-biased mental health risks.

Abstract

A key step towards understanding psychiatric disorders that disproportionately impact female mental health is delineating the emergence of sex-specific patterns of brain organization at the critical transition from childhood to adolescence. Prior work suggests that individual differences in the spatial organization of functional brain networks across the cortex are associated with psychopathology and differ systematically by sex.

We aimed to evaluate the impact of sex on the spatial organization of person-specific functional brain networks.

We leveraged person-specific atlases of functional brain networks defined using non-negative matrix factorization in a sample of n = 6437 youths from the Adolescent Brain Cognitive Development Study. Across independent discovery and replication samples, we used generalized additive models to uncover associations between sex and the spatial layout (“topography”) of personalized functional networks (PFNs). We also trained support vector machines to classify participants’ sex from multivariate patterns of PFN topography.

Sex differences in PFN topography were greatest in association networks including the fronto-parietal, ventral attention, and default mode networks. Machine learning models trained on participants’ PFNs were able to classify participant sex with high accuracy.

Sex differences in PFN topography are robust, replicate across large-scale samples of youth. These results suggest a potential contributor to the female-biased risk in depressive and anxiety disorders that emerge at the transition from childhood to adolescence.

Full text

Introduction

Many psychiatric disorders show sex differences in prevalence, presentation and trajectory. For example, the lifetime prevalence of internalizing disorders such as depression and anxiety is nearly twice as high in females^1^, and developmental disorders such as attention-deficit hyperactivity disorder often present differently in males and females leading to disparities in diagnosis and treatment. These sex differences tend to emerge during the transition from childhood to adolescence, a time when functional brain networks implicated in these disorders are refined^2,3^. Previous research has begun to link sex differences in internalizing disorders with sex differences in multimodal neuroimaging measures, including in studies of youth^4–6^. Therefore, understanding and treating mental health conditions for all individuals, including those that are more prevalent in and differentially impact females, requires a clear understanding of sex differences in neurodevelopment.

Prior neuroimaging studies have revealed significant sex differences in functional networks supporting cognitive and emotional processes, including the fronto-parietal^7,8^ and default mode^9^ networks. Dysfunction within these networks has been linked with psychiatric disorders, including anxiety and depression^10–14^. Critically, these functional networks are highly person-specific in their spatial organization across the cortex (“functional topography”). Substantial individual differences in the size, shape, and spatial location of brain regions comprising these networks emerge gradually during neurodevelopment with evidence of sex-specific patterning^3,15^ associated with X-linked gene expression patterns^15^. Innovations in precision brain mapping approaches have begun to chart the person-specific functional topography of personalized functional brain networks (PFNs)^16–18^ and have uncovered novel associations with internalizing psychopathology^14,19,20^ and cognition^3,21^.

In a recent study of individuals across a broad age range (*n=*693, 8–22 years old)^15^, we presented the first report of sex differences in PFN functional topography. Given the ongoing “reproducibility crisis” in psychology and neuroscience wherein a large proportion of research findings fail to replicate in new datasets^22^, it is important to determine whether sex differences in functional topography are replicable across demographically diverse samples with a wider variety of MRI scanning locations and procedures. Moreover, it remains unclear whether these sex differences are consistently observed at the critical transition from childhood to adolescence when many psychiatric disorders first emerge, and whether these differences are associated with pubertal hormone levels. Here, we examine sex differences in PFN topography in youth using non-linear modeling and machine learning in data from the Adolescent Brain Cognitive Development (ABCD) Study^®23^ (n=6,437, ages 9–10). We hypothesized that sex differences would be greatest in association networks, as the functional topography of these networks showed the strongest associations with sex assigned at birth and cortical X-linked gene expression patterns in our previous work^15^. Of note, the novel participant sample used in the present study differs from that used in previous work^15^ across a number of dimensions, including sample size, age range, pubertal stage, scanner types and protocols, data collection sites, functional MRI tasks, racial/ethnic diversity, and socioeconomic status, allowing us to rigorously test the reproducibility and generalizability of our findings.

Method

Participants

Participants from the ABCD Study^®23^ baseline assessment were drawn from the ABCD BIDS Community Collection (ABCC, ABCD-3165^24^). These data were collected across 22 sites in the United States, with Institutional Review Board (IRB) approval from the University of California, San Diego, as well as each of the respective study sites. Written informed consent (parents or guardians) and assent (children) were obtained. Criteria for participation in the ABCD Study^®^ are described in detail in previous work^25^. From the full baseline sample (n=11,878, 9–10 years old), we excluded participants with incomplete data or excessive head motion during fMRI scanning (Figure S1), yielding a final sample of n=7,459. Analyses were conducted in matched discovery (n=3,240, 50.46% female) and replication (n=3,197, 49.13% female) samples drawn from the ABCD Reproducible Matched Samples (ARMS^24,26^), with siblings excluded separately in the discovery and replication samples to avoid leakage across subsamples during model cross-validation (Figure S1). Importantly for the present study, we note that participant “sex” was assessed using a binary caregiver-reported question regarding the assignment of sex at birth on the original birth certificate. Hereafter, we use the term “sex” to refer to sex assigned at birth, the term “female” to refer to individuals assigned female at birth, and the term “male” to refer to individuals assigned male at birth. Demographic information for the participants included in the present study are presented in Table S1.

Functional Magnetic Resonance Imaging (fMRI) Processing:

As in our prior work^21,28^, we leveraged data from the ABCD Community Collection (ABCC) 3165 processed with the ABCD-BIDS pipeline which included distortion correction and alignment, Advanced Normalization Tools (ANTS^29^) denoising, FreeSurfer^30^ segmentation, and surface and volume registration with rigid-body transformation^31,32^. Following this, further processing was done using the DCAN BOLD Processing (DBP) pipeline which includes de-meaning and de-trending of fMRI data with respect to time, denoising using a general linear model with regressors for signal and movement, bandpass filtering between 0.008 and 0.09 Hz using a 2nd order Butterworth filter, applying the DBP respiratory motion filter (18.582–25.726 breaths per minute), and applying DBP motion censoring (frames exceeding an FD threshold of 0.2 mm or failing to pass outlier detection at +/− 3 standard deviations were discarded). We then concatenated cleaned time series data for resting-state and task-based scans as in previous work^21,28^ to maximize the data available for analysis. We excluded participants who had fewer than 600 remaining TRs after motion censoring, as well as those who failed ABCD quality control for their T1 or resting-state fMRI scan.

Definition of personalized functional networks (PFNs)

Detailed information about the neuroimaging acquisition for the ABCD Study^®^, including scanner manufacturers and MRI scanning protocols, have been described previously^27^. Following the same fMRI preprocessing steps (Supplemental Information) as in our prior work in this dataset^21,28^, we maximized the available high-quality data for our analyses by concatenating fMRI timeseries from three task-based scans (Emotional N-Back Task, Stop-Signal Task, and Monetary Incentive Delay Task) and two resting-state scans, and retained only the individuals passing strict motion correction (a minimum of 600 remaining TRs in total after motion censoring). Functional brain regions comprising large-scale networks have been shown to vary substantially in their size, shape, and spatial location across individuals^16,17^. We therefore employed a precision brain mapping approach as in previous work^3,15,19,21,28^ that leverages spatially-regularized non-negative matrix factorization (NMF)^33^ to define individual-specific atlases of functional brain network organization (Figure 1a)^3,34^. This approach has been implemented in previous studies using this dataset^21,28^ to identify seventeen PFNs, revealing substantial inter-individual differences in the spatial layout of functional brain regions, with greatest heterogeneity in association networks (Figure 1b).

Statistical analyses

We aimed to (1) evaluate whether individual-specific patterns of PFN topography were associated with sex and (2) assess the extent to which sex can be accurately classified from patterns of PFN topography in new individuals. To this end, we first conducted a mass univariate analysis relating vertex-wise PFN topography to sex, then trained multivariate classification models using rigorous cross-validation, as described in detail below.

Mass univariate analysis:

To determine whether sex is associated with distinct patterns of PFN topography, we first evaluated vertex-wise associations, as in our previous work^15^ using generalized additive models (GAMs) with penalized splines. These GAMs were fit at each vertex and included a linear covariate for in-scanner head motion (mean fractional displacement), a nonlinear term for age, and a random effect covariate for data collection site. We accounted for multiple comparisons within each PFN by controlling the false discovery rate (FDR; Q<0.05). Spatial maps of GAM loadings were compared across discovery and replication samples using conservative spin-based permutation testing to account for spatial autocorrelation^35^. To determine the role of pubertal development and hormone levels in shaping potential sex differences in PFN topography, we also conducted mass univariate analyses using data from the Pubertal Development Scale (PDS)^36^ and salivary hormone levels for DHEA, testosterone, and estradiol^37^. DHEA and testosterone were collected for both sexes; estradiol was collected for females only.

Multivariate classification:

To leverage the high-dimensional data from individual-specific patterns of PFN topography across the whole cortex simultaneously, we next trained a linear support vector machine (SVM) to categorize participant sex based on their multivariate PFN loadings matrix. SVM is a common form of classifier that is well suited to make use of high dimensional data for binary classification and has been shown to perform well in previous work^15^. Replicating the procedure in our prior work^15^, we applied nested two-fold cross-validation (2F-CV), with the inner loop used to determine the optimal tuning parameter C to balance model bias and variance, and the outer loop used to estimate model accuracy in held-out data. Classifier performance was evaluated using accuracy, sensitivity, specificity, and the area under the receiver operating characteristic (ROC) curve. We also evaluated classifier performance relative to a set of 1000 null models, where participant sex was permuted relative to PFN topography on each iteration.

Prior to model training and testing, we eliminated siblings to avoid leakage of family structure across subsamples, yielding a total sample of n = 6,437 (discovery: n = 3240, 50.46% female; replication: n = 3197, 49.13% female) for all multivariate classification analyses. Before beginning our 2F-CV procedure, we first split the data between the matched discovery and replication samples according to the previously defined ABCD Reproducible Matched Samples^24,26^. Then, separately within the discovery and replication samples, we performed 2F-CV as follows (Figure S2). For the outer 2F-CV loop, we trained and tested the SVM model using split-half subsets separately within either the discovery or replication sample. After training the model in one half of the data and testing its performance in the other held-out half of the data, we then repeated this procedure in reverse. Prior to model training, covariates for age, site, and in-scanner head motion (mean fractional displacement) were regressed from each feature, separately in the training and testing sets to avoid leakage. To determine whether classification accuracy was driven by the choice of split, we repeated this analysis using 100 permuted splits of the data, each time randomly dividing the discovery and replication samples into independent training and testing sets.

Inner 2F-CV loops were used to determine the optimal tuning parameter C by further randomly dividing the training set of the outer 2F-CV loop into two subsamples. The first split-half subsample was used to train the SVM model with each of 15 possible C parameter values: [2^−5^, 2^−4^, … , 2^8^, 2^9^]. These models were each tested in the second held-out subsample as in our previous work^15^. We then repeated this procedure using the second subsample for training and the first subsample for testing, calculating the average held-out classification accuracy across the two subsamples for each value of the parameter C. The optimal C parameter value was selected as the C with the highest average held-out classification accuracy, and this optimal C parameter was used to train the models within the outer 2F-CV loop. It is worth noting that even the smallest subdivisions of the data in our nested 2F-CV procedure still contained over one thousand participants each at a minimum, yielding sufficient statistical power to train and test our machine learning models using the most conservative possible (fewest folds) cross-validation approach.

To evaluate the relative importance of each feature within the SVM model, we first extracted feature weights for each network loading at each vertex and averaged these weights across the 100 randomly permuted splits of the data. Then, to avoid challenges with interpretation due to the covariance structure among feature weights, we applied the Haufe transformation^38^ to invert the models prior to feature weight interpretation. Next, we averaged the Haufe-transformed weight maps across the training and testing sets from the outer loop of the matched samples 2F-CV procedure. As in our univariate analysis, spatial maps of SVM weights were compared across samples using spin-based permutation testing^35^.

Results

Association between sex and person-specific functional topography

To characterize sex differences in functional brain network topography just prior to the transition from childhood to adolescence, we leveraged previously defined maps of personalized functional networks (PFNs; Figure 1) for each individual in the ABCD Study^®^ dataset^21^ (n=6,437, 9–10 years old, 49.8% female). These maps reflect each individual’s unique functional topography of seventeen canonical large-scale networks. To determine whether a participant’s sex is reflected in their person-specific patterns of functional brain network organization, we first conducted mass univariate analyses using generalized additive models (GAMs) to relate vertex-wise PFN loadings to sex.

We found spatially heterogeneous associations between sex and PFN topography in both discovery and replication samples. Sex differences in functional topography were greatest in association networks (Figure 2A–C; Figure S3–4), with some PFNs exhibiting greater loadings in females (e.g., fronto-parietal and dorsal attention networks) and others exhibiting greater loadings in males (e.g., default mode and ventral attention networks). We evaluated the total effect of sex at each vertex by summing the absolute value of the Z-statistic across all 17 PFNs. This analysis revealed that associations between sex and PFN topography are greatest in association cortices such as the inferior parietal lobule, ventrolateral prefrontal cortex, and orbitofrontal cortex (Figure 2D; Figure S5). We observed highly consistent spatial distributions of GAM loadings across discovery and replication samples (r=0.90, pspin<0.001; Figure 2E) and with our prior work in an independent dataset^15^ (r=0.59, pspin<0.001; Figure 2F), using conservative spin-based spatial randomization testing to account for spatial autocorrelation^35^. These results were also found to be consistent in sensitivity analyses that included pubertal stage, pubertal timing, and salivary hormone levels as covariates (Figures S6–S9), and we observed no significant associations between PFN topography and pubertal measures, including pubertal stage, pubertal timing, and salivary hormone levels (Figures S10 and S11).

Next, we sought to confirm these vertex-wise mass univariate results by using multivariate classification to leverage the full pattern of PFN topography across the cortex. To evaluate how multidimensional patterns of PFN topography relate to sex, we trained linear support vector machine (SVM) classifiers to categorize participants’ sex from PFN topography patterns using conservative cross-validation. These models were able to correctly identify held-out participants’ sex as male or female from PFN topography patterns with high accuracy averaged across the 100 SVM iterations within each subsample (discovery: 87.4%, replication 87.2%; Figure 3A, Figure S12A), successfully replicating our prior work^15^. Model sensitivity and specificity were 0.876 and 0.872, respectively in the discovery sample (replication: 0.870 and 0.870), with a large area under the ROC curve (discovery: 0.966; replication: 0.965), indicating excellent model performance on held-out data that exceeded chance-level accuracy from randomly permuted null models (mean: 0.50, p<0.001; Figure 3A, Figure S12A, inset histograms).

Model performance was robust to the choice of split in participants between the training and testing sets, as evidenced by repeated random cross-validation (discovery: mean accuracy = 87.4% ; 95% CIs = [0.873, 0.875]; replication: mean accuracy = 87.2% ; 95% CIs = [0.871, 0.873]). To identify which brain regions contributed most to the correct classification of participant sex from functional topography, we examined the SVM feature weights after applying the Haufe transformation^38^ to invert the models for interpretability. Replicating prior results^15^, we found that association networks contributed most to the classification of participant sex, primarily those within the fronto-parietal, ventral attention, and default mode networks (Figure 3B–C; Figure S12B–C). Vertex-wise patterns of feature weights also provided convergent results with mass univariate analyses (discovery: r=0.86, pspin<0.001; replication: r=0.83, pspin<0.001; Figure 3D; Figure S12D). The spatial pattern of feature weights was also highly consistent across samples (r=0.93, pspin<0.001; Figure S13). We also found convergent results when SVM models were trained separately on vertex-wise loadings from each PFN independently, with ventral attention, default mode, and fronto-parietal networks showing the best model performance across discovery and replication samples (Figure S14).

Discussion

Our results demonstrate robust and replicable associations between sex and the spatial patterning of functional brain networks in youth. Across analytic approaches and independent samples, we consistently find that the spatial patterning of person-specific functional brain networks significantly differs based on sex as a biological variable. While no single brain region or network is systematically larger or smaller in its spatial extent across all males or females, we find that the greatest sex differences in functional topography tend to be disproportionately found in association areas like the fronto-parietal, default mode, and ventral attention networks, with weaker effects found in sensory and motor cortices. These results represent a successful replication of prior findings^15^ in a large sample of participants and suggest that sex might be one of many factors that shape the development of functional networks in youth at the precipice of the critical transition to adolescence.

Sex differences in personalized functional brain network topography in youth

Extending prior work describing sex differences in neuroimaging features in young adults^39,40^, our results suggest that sex differences in functional topography are consistently observed in children just prior to the transition to adolescence. This critical transition period that often coincides with pubertal changes is marked by the emergence of many common psychiatric disorders, including depression and anxiety, which disproportionately affect females^1^. This time period also coincides with the maturation of functional brain networks, including the protracted development of association networks like the fronto-parietal and default mode networks^41,42^ which have been shown to have distinct profiles of functional development between males and females^43^. These association networks also exhibit the most person-specific patterns of functional topography among all large-scale brain networks^3^ and are associated with symptoms of psychopathology^19,20^. Our observation that these networks also reflect an individual’s sex aligns with previous findings^44^, including recent findings in adults^39,40^, and suggests that sex differences in functional brain networks may play a role in the emergence and exacerbation of sex differences in psychiatric disorders during the transition to adolescence. Thus, future studies may seek to further investigate potential behavioral consequences of sex differences in association network topography in youth as well as the potential role of functional brain network development as an early biomarker for sex-specific psychiatric symptom emergence in youth.

Sex differences in functional topography consistently replicate across independent datasets

Replication studies often fail^45^, and even successful replication studies most often yield results with smaller effect sizes than initial discoveries^46^. The present study not only successfully replicates findings observed in our prior work, it also uncovered effect sizes that were approximately the same or even larger than in the previous study^15^. Specifically, the present study confirmed the presence of sex differences in PFN topography and replicated the observation that these sex differences are primarily found in association networks. This successful replication is especially notable in light of the many differences between the datasets in each study, including sample size, age range, scanner types and protocols, data collection sites, functional MRI tasks, racial/ethnic diversity, and socioeconomic status. Thus, the present study represents a strong counterexample to the ongoing reproducibility crisis in psychology and neuroscience^22^.

Several important distinctions between the present study and this previous work provide context for interpreting these results. First, the previous study^15^ used data from the Philadelphia Neurodevelopmental Cohort (PNC; n=693). Here, we applied the same analytical approach to a dataset that is an order of magnitude larger (ABCD Study^®^; n=6,437). This considerable increase in sample size may explain the improvement in model performance on held-out data between studies (from 82.9% to 87.1% accuracy), as models trained in larger datasets with rigorous cross-validation are less likely to be overfit^47,48^. Second, the previous study^15^ assessed individuals aged 8–23 years old, while the present study leveraged data from the baseline assessment of the ABCD Study^®^ when participants were 9–10 years old. The more restricted age range in the present study may also help to explain the improved model performance, since functional brain network topography changes throughout development^2,3^. Though age was included as a model covariate in both studies, it is possible that the smaller age range in the present study still yielded some advantage in classifying sex from patterns of functional topography at a more restricted time period of brain development.

Limitations

There are several limitations of this study worth noting. First, sex was assessed using a binary parent-reported question regarding the assignment of sex at birth on the original birth certificate, and we lacked a sufficiently large sample size to examine functional topography of intersex youth. Importantly, existing data suggest that binary classifications of sex do not align well with the complex mosaics of male and female characteristics observed in individual brains^49^. Thus, further research is warranted to more comprehensively characterize person-specific patterns of male, female, and intersex characteristics in functional brain network topography. Second, prior work has shown that functional brain network connectivity is associated with both sex and gender in youth^50^. As the present study aimed to understand sex differences in functional topography, future work is also needed to investigate potential effects of continuous gender dimensions such as gender identity and expression. Given that only 0.5% (n=58) of baseline ABCD Study^®^ participants reported being or possibly being transgender^51^, and given that gender continues to develop throughout early adolescence, future studies in longitudinal timepoints will be key in investigating potential individual or interactive effects of sex and gender in shaping neurodevelopment.

Third, the present study leveraged a cross-sectional sample at a single time point from within an ongoing longitudinal study of youth. As youth from the ABCD Study^®^ continue to participate in follow-up study sessions from childhood to adulthood, it will become increasingly possible to investigate changes in sex-specific functional brain network topography with critical developmental changes such as puberty across longer timescales than investigated in the present study. Moreover, because puberty was already underway in a substantial portion of females in the ABCD Study, future studies of younger individuals will be required to investigate the activational role of pubertal hormones, which begin before physical changes become observable, on sex differences in functional topography. Future longitudinal studies considering the complex interplay of biopsychosocial factors related to sex and gender development may also reveal mechanistic links between sex-specific patterns of functional brain network topography and sex differences in psychiatric illness manifestation (e.g., internalizing symptoms). Fourth, the present study focused on sex differences in functional rather than structural differences in brain organization, though sex differences in gross structural anatomy (e.g. head size) are well documented^52^. However, recent work has demonstrated that sex differences in functional brain organization do not appear to be systematically associated with structural imaging measures such as surface area or microstructural organization^44^.

Future directions: using precision brain mapping to inform female mental health

In addition to the future directions noted above, our observation that person-specific patterns of functional brain network topography show sex differences, particularly in association networks related to psychiatric symptoms^19,20^, also lays important groundwork for future studies of sex differences in mental health, including mental health conditions that disproportionately impact females. First, future work should further examine how PFN topography develops across the female reproductive lifespan, with a particular focus on changes across critical hormonal transition periods such as puberty, pregnancy, and menopause. These hormonal transition periods are known to have substantial impact on neurodevelopment and often align with the timing of psychiatric illness onset^53^, yet have been historically underfunded and understudied^54^. Extending the study of PFNs across the lifespan therefore has potential to improve our understanding of how neuroplasticity during hormonal shifts impacts functional topography and trajectories of psychiatric illness. Second, longitudinal studies examining how sex differences in PFN topography emerge during development may inform early preventions or personalized treatments for psychiatric illnesses such as personalized neuromodulation via transcranial magnetic stimulation (TMS), filling critical gaps in existing treatment options.

Finally, it is worth noting that sex-specific individual differences in the topography of association networks may also reflect childhood environments and socio-economic status^28^, which have also been shown to explain a large portion of inter-individual variance in psychopathology symptoms^55^. Taken together with evidence of sex differences in stress responses across the lifespan^56^, our findings motivate future research into whether sex differences in the effects of environmental stressors are associated with sex differences in association network topography and psychiatric illness. Additionally, given that environmental stressors have been shown to confer vulnerability to psychiatric symptoms during future reproductive timepoints characterized by significant hormonal fluctuations such as pregnancy^57^ and menopause^58,59^. Future work may therefore seek to parse the independent and interactive effects of hormonal, genetic, and environmental factors that may together shape individual-specific spatial patterning of functional networks across the female reproductive lifespan.

Conclusions

In this study, we demonstrate reproducible sex differences in person-specific patterns of functional brain network organization in youth. The ability to successfully classify sex from the spatial configuration of PFNs is primarily driven by sex differences in the functional topography of association networks. By characterizing sex differences in functional topography in youth, this study provides a key stepping-stone toward addressing sex differences in susceptibility to psychiatric symptoms that emerge during the transition to adolescence.

Supplementary Material

1