Functional reorganization of brain networks across the human menstrual cycle

The brain is an endocrine organ, sensitive to the rhythmic changes in sex hormone production that occurs in most mammalian species. In rodents and nonhuman primates, estrogen and progesterone’s impact on the brain is evident across a range of spatiotemporal scales. Yet, the influence of sex hormones on the functional architecture of the human brain is largely unknown. In this dense-sampling, deep phenotyping study, we examine the extent to which endogenous fluctuations in sex hormones alter intrinsic brain networks at rest in a woman who underwent brain imaging and venipuncture for 30 consecutive days. Standardized regression analyses illustrate estrogen and progesterone’s widespread associations with functional connectivity. Time-lagged analyses examined the temporal directionality of these relationships and suggest that cortical network dynamics (particularly in the Default Mode and Dorsal Attention Networks, whose hubs are densely populated with estrogen receptors) are preceded—and perhaps driven—by hormonal fluctuations. A similar pattern of associations was observed in a follow-up study one year later. Together, these results reveal the rhythmic nature in which brain networks reorganize across the human menstrual cycle. Neuroimaging studies that densely sample the individual connectome have begun to transform our understanding of the brain’s functional organization. As these results indicate, taking endocrine factors into account is critical for fully understanding the intrinsic dynamics of the human brain. Highlights Intrinsic fluctuations in sex hormones shape the brain’s functional architecture. Estradiol facilitates tighter coherence within whole-brain functional networks. Progesterone has the opposite, reductive effect. Ovulation (via estradiol) modulates variation in topological network states. Effects are pronounced in network hubs densely populated with estrogen receptors.

1 Introduction Figure 1. Timeline of data collection for the 30 experiment sessions. Endocrine and MRI assessments were collected at the same time each day to minimize time-of-day effects. 98 To monitor state-dependent mood and lifestyle measures over the cycle, the following 99 scales (adapted to reflect the past 24 hours) were administered each morning: Perceived 100 Stress Scale (PSS) (Cohen et al., 1983), Pittsburgh Sleep Quality Index (PSQI) (Buysse et al.,101 1989), State-Trait Anxiety Inventory for Adults (STAI) (Spielberger and Vagg, 1984), and testosterone) and the pituitary gonadotropins luteinizing hormone (LH) and follicle 118 stimulating hormone (FSH). One 10cc mL blood sample was collected in a vacutainer SST 119 (BD Diagnostic Systems) each session. The sample clotted at room temperature for 45 min 120 until centrifugation (2,000 ×g for 10 minutes) and were then aliquoted into three 1 mL 121 microtubes. Serum samples were stored at -20 • C until assayed. Serum concentrations 122 were determined via liquid chromatography-mass spectrometry (for all steroid hormones) 123 and immunoassay (for all gonadotropins) at the Brigham and Women's Hospital Research 124 Assay Core. Assay sensitivities, dynamic range, and intra-assay coefficients of variation 125 (respectively) were as follows: estradiol, 1 pg/mL, 1-500 pg/mL, < 5% relative standard 126 (https://caseforge.co/) (days 8-30 of Study 1, days 1-30 of Study 2). Overall 148 motion (mean framewise displacement) was negligible (Figure S1), with fewer than 130 149 microns of motion on average each day. Importantly, mean framewise displacement was 150 also not correlated with estradiol concentrations (Study 1: Spearman r = −0.06, p = .758; 151 Study 2: Spearman r = −0.33, p = .071). Volterra expansion of translational/rotational motion parameters, accounting for 168 autoregressive and nonlinear effects of head motion on the BOLD signal (Friston et al., 169 1996). All nuisance regressors were detrended to match the BOLD timeseries. Our use of 170 coherence allows for the estimation of frequency-specific covariances in spectral 171 components below the range contaminated by physiological noise. Nevertheless, to 172 ensure the robustness of our results, we re-analyzed the data with global signal regression  Functional network nodes were defined based on a 400-region cortical 177 parcellation (Schaefer et al., 2018) and 15 regions from the Harvard-Oxford subcortical 178 atlas (http://www.fmrib.ox.ac.uk/fsl/). For each day, a summary timecourse 179 was extracted per node by taking the first eigenvariate across functional volumes (Friston 180 et al., 2006). These regional timeseries were then decomposed into several frequency 181 bands using a maximal overlap discrete wavelet transform (Daubechies extremal phase 182 filter, length = 8). Low-frequency fluctuations in wavelets 3-6 (∼0.01-0.17 Hz) were 183 selected for subsequent connectivity analyses (Patel and Bullmore, 2016). We estimated 184 the spectral association between regional timeseries using magnitude-squared coherence: 185 this yielded a 415 × 415 functional association matrix each day, whose elements indicated 186 the strength of functional connectivity between all pairs of nodes (FDR-thresholded at 187 q < .05). Coherence offers several advantages over alternative methods for assessing 188 connectivity: 1) estimation of frequency-specific covariances, 2) simple interpretability (values 189 are normalized to the [0, 1] interval), and 3) robustness to temporal variability in 190 hemodynamics between brain regions, which can otherwise introduce time-lag confounds 191 to connectivity estimates via Pearson correlation. association between connectivity and hormones, the coherence data at each edge were 200 randomly permuted, models were fit, and two-tailed p-values were obtained as the 201 proportion of models in which the absolute value of the permuted test statistics equaled or 202 exceeded the absolute value of the 'true' test statistics. We report edges surviving a 203 threshold of p < .001. We did not apply additional corrections in an effort to maximize 204 power in our small sample size; Study 2 instead offers an independent validation of the 205 observed whole-brain effects.

206
Next, we sought to capture linear dependencies between hormones and network 207 connectivity directed in time using vector autoregressive (VAR) models. Here we chose to 208 focus exclusively on estradiol for two reasons: 1) the highly-bimodal time-course of 209 progesterone over a natural cycle confers a considerably longer autocorrelative structure, 210 requiring many more free parameters (i.e. higher-order models, ultimately affording 211 fewer degrees of freedom); and 2) progesterone lacks an appreciable pattern of periodicity 212 in its autocovariance with network timeseries, suggesting less relevance for time-lagged 213 analysis over a single cycle. In contrast, estradiol has a much smoother time-course that is 214 well-suited for temporal evolution models such as VAR.

215
In short, VAR solves a simultaneous system of equations that fits current states of the 216 brain and estradiol from the previous states of each. For consistency, we considered only 217 second-order VAR models, given a fairly reliable first zero-crossing of brain/hormone 218 autocovariance functions at lag two (this was based on common criteria noted in other 219 instances of time-delayed models; Boker et al. (2014)). Fit parameters for each VAR 220 therefore reflect the following general form: where error terms, i,t , are assumed to be uncorrelated and normally-distributed. Given 222 that the design matrix is identical for each outcome measure, they can be combined in 223 matrix form, and a least-squares solution to the system of equations can be obtained via 224 maximum likelihood.

225
With respect to brain states, we modeled both edgewise coherence and factors related given network). These were derived using the relevant functions for weighted graphs in 231 the Brain Connectivity toolbox (Rubinov and Sporns, 2010 Ultimately, regardless of brain measure, each VAR was estimated similarly to the 239 time-synchronous analyses described above: data were Z-scored, models were fit, and 240 model-level stats (test-statistics, R 2 , and RM SE) were empirically-thresholded against 241 10,000 iterations of nonparametric permutation testing. Here, however, both brain and 242 hormonal data were permuted under the null hypothesis of temporal stochasticity (i.e. no 243 autoregressive trends and no time-lagged dependencies between variables). As before, we 244 did not apply additional corrections and offer Study 2 as an independent validation set.   Table 1). Progesterone concentrations surpassed 5 ng/mL in the luteal phase,

276
signaling an ovulatory cycle (Leiva et al., 2015). In Study 2, the participant was placed on which to test the reliability of estradiol's influence on intrinsic brain networks.  did not impact the results ( Figure S4).

290
To further explore cycle-dependent variability, we tested the hypothesis that from Study 1) greatly reduced whole-brain associations ( Figure S6B). Thus, whole-brain 329 functional connectivity appears highly-sensitive to estradiol regardless of reproductive 330 status.

Time-lagged associations between estradiol and whole-brain
332 functional connectivity 333 We then employed time-lagged methods from dynamical systems analysis to further 334 elucidate the degree to which intrinsic functional connectivity is sensitive to fluctuations 335 in estradiol: specifically, vector autoregression (VAR), which supports more directed 336 temporal inference than standard regression models. As described previously, we report 337 results from second-order VAR models: thus, in order to assess connectivity or hormonal 338 states on a given day of the experiment, we consider their values on both the previous day 339 (hereafter referred to as 'lag 1') and two days prior (hereafter referred to as 'lag 2').

340
Ultimately, if brain variance over time is attributable to previous states of estradiol, this 341 suggests that temporal dynamics in connectivity may be driven (in part) by fluctuations in 342 this hormone.

343
When assessing edgewise connectivity states, a powerful disparity emerged between 344 the brain's autoregressive effects and the effects of estradiol in Study 1. We observed vast, 345 whole-brain associations with prior hormonal states, both at lag 1 and lag 2 ( Figure 5A).

346
Perhaps most immediately striking, the sign of these brain-hormone associations inverts 347 between lags, such that it is predominantly positive at lag 1 and predominantly negative 348 at lag 2-this holds for all networks when considering their mean nodal association 349 strengths ( Figure 5B). We interpret this as a potential regulatory dance between brain   Note. p-values empirically-derived via 10,000 iterations of nonparametric permutation testing. 388 As expected, estradiol demonstrated significant autoregressive trends across all models in 389 Study 1. However, between-network integration was only tenuously associated with 390 previous states of estradiol: in several intrinsic networks, overall model fit (variance 391 accounted for, R 2 , and root mean-squared error, RM SE) was at best marginal compared 392 to empirical null distributions of these statistics, and we therefore urge caution in associated with cross-network participation, but model fit as a whole was low (see Table   401 S1). Attention Network (A, left). Timeseries for each network statistic are depicted above in (B,C) and estradiol for each VAR is plotted below. Data are in standardized units and begin at experiment day three, given the second-order VAR (lag of two days).

Figure 6. Dorsal Attention Network topology is driven by previous states of estradiol (Study 1). Observed data (solid lines) vs. VAR model fits (dotted lines) for betweennetwork participation (B, middle) and within-network efficiency (C, right) in the Dorsal
Importantly, we failed to replicate these effects in Study 2 under hormonal 403 suppression ( Table S2). The autoregressive trends in estradiol were generally blunted,

Estradiol and global efficiency 413
In contrast to between-network integration, estradiol was more strongly associated with   Note. p-values empirically-derived via 10,000 iterations of nonparametric permutation testing.
We observed a similar pattern of results in the Dorsal Attention Network (R 2 = 0.37, 426 p = .022; RM SE = 0.77, p = .023; Figure 6C; Table 3). Estradiol again demonstrated 427 significant autoregressive trends at lag 1 (b = 1.17, SE = 0.19, t = 6.30, p < .0001) and lag  (Table S3). 440 In contrast to between-network participation, within-network efficiency yielded Networks; however, aside from the SomatoMotor Network (R 2 = 0.34, p = .039; 453 RM SE = 0.76, p = .018), overall fit in these models was nonsignificant (Table S4) mirrors the cyclic pattern of estradiol release typical of the macaque menstrual cycle (Hao 502 et al., 2006;Ohm et al., 2012). Pairing estradiol with cyclic administration of progesterone 503 blunts this increase in spine density (Ohm et al., 2012). In the hippocampus, progesterone 504 has a similar inhibitory effect on dendritic spines, blocking the proliferative effects of 505 estradiol 6 hours after administration (Woolley and McEwen, 1993). Together, the 506 preclinical literature suggests that progesterone antagonizes the largely proliferative 507 effects of estradiol (for review, see Brinton et al. (2008)). We observed a similar 508 relationship, albeit at a different spatiotemporal resolution, with estradiol demonstrating 509 positive associations with coherence across numerous cortical networks and progesterone 510 having an opposite, negative association on average. In sum, animal studies have 511 identified estradiol's influence on regional brain organization at the microscopic scale.

512
Here, we show that estradiol and progesterone may have analogous effects evident at the 513 mesoscopic scale of whole-brain connectivity, measured by spectral coherence, and 514 macroscopic features of network topology.

528
Inconsistencies between studies could be due to a number of factors such as differences in 529 cycle staging methods, divergent functional connectivity estimation methods, or 530 unaccounted for intra/inter-individual variability (Beltz and Moser, 2019). Our results 531 suggest that failure to properly capture the complete ovulatory window, when estradiol 532 levels rapidly rise, could lead to skewed estimates of stability within functional brain 533 networks across the menstrual cycle (Hjelmervik et al., 2014). As such, dense-sampling 534 studies provide a novel solution to capturing pivotal moments experienced across a 535 complete human menstrual cycle. Arélin et al. (2015) sampled an individual every 2-3 536 days across four cycles and found that progesterone was associated with increased 537 connectivity between the hippocampus, dorsolateral PFC and sensorimotor cortex, 538 providing compelling evidence that inter-regional connectivity varies over the cycle. This 539 particular dense-sampling approach allowed the authors to examine brain-hormone 540 relationships while accounting for intra-individual cycle variation. 541 genomic-effects on the central nervous system. For example, spine density on 543 hippocampal neurons varies by ∼30% over the rodent estrous cycle. In-vivo MRI evidence 544 in mice demonstrates that these hormone-mediated changes can occur rapidly, with 545 differences detectable within a 24-hour period. To capture time-synchronous (rapid) and  The following considerations could enhance the interpretation of these data. First, our 559 investigation deeply sampled a single woman, limiting our ability to generalize these 560 findings to other individuals. To enrich our understanding of the relationship between sex 561 hormones and brain function, this dense-sampling approach should be extended to a 562 diverse sample of women. Doing so will allow us to examine the consistency of our results 563 with respect to inter-individual differences in network organization over the menstrual proportionally-coupled nonlinear oscillators (Boker et al., 2014). Within-person cycle 574 variability is almost as large as between-person cycle variability, which hints that there are 575 highly-complex hormonal interactions within this regulatory system (Boker et al., 2014;576 Fehring et al., 2006). The VAR models we have explored reveal linear dependencies 577 between brain states and hormones, but other dynamical systems methods (e.g. coupled 578 latent differential equations) may offer more biophysical validity (Boker et al., 2014). 579 However, the current sample size precludes robust estimation of such a model.

580
Third, while permutation tests have been used as empirical null models for 581 VAR (Hyvärinen et al., 2010) and its statistical relatives (e.g. Granger causality; Barnett 582 and Seth (2014)), the practice of temporally-scrambling a timeseries will drastically alter 583 its autocorrelative structure and potentially skew observed dependencies over time.

584
Phase-shifting, surrogate data tests such as the amplitude adjusted Fourier transform 585 (AAFT) may offer more robust null distributions. However, AAFT also makes strong 586 distributional assumptions about the original timeseries (Gaussian normality) that, 587 unfortunately, are not met by these data. Additionally, the small sample size over a single 588 cycle precludes the ability to derive robust surrogate realizations of the timeseries. While 589 AAFT is arguably an ideal procedure for analyses such as those reported here, these data 590 simply cannot meet the assumptions required for valid surrogate testing and thus is a 591 major limitation within the current study. Future investigations involving larger samples 592 of women over several cycles that allow implementation of such models will be critical.

593
Fourth, while coherence is theoretically robust to timing differences in the 594 hemodynamic response function, hormones can affect the vascular system (Krause et al., 595 2006). Therefore, changes in coherence may be due to vascular artifacts that affect the 596 hemodynamic response in fMRI, rather than being neurally-relevant. Future investigations 597 exploring the assumptions of hemodynamics in relation to sex steroid hormone 598 concentrations will add clarity as to how the vascular system's response to hormones 599 might influence large-scale brain function.

600
Fifth, these findings contribute to an emerging body of work on estradiol's ability to 601 enhance the efficiency of PFC-based cortical circuits. In cycling women performing a 602 working memory task, PFC activity is exaggerated under low estradiol conditions and 603 reduced under high estradiol conditions (Jacobs and D'Esposito, 2011). The same pattern 604 is observed decades later in life: as estradiol production decreases over the menopausal transition, working memory-related PFC activity becomes more exaggerated, despite no 606 difference in working memory performance (Jacobs et al., 2016a). Here, we show that 607 day-to-day changes in estradiol enhance the global efficiency of functional networks, with 608 pronounced effects in networks (DMN and FCN) that encompass major regions of the 609 PFC (Schaefer et al., 2018;Yeo et al., 2011). Together, these findings suggest that estradiol 610 generates a neurally efficient PFC response at rest and while engaging in a cognitive task.

617
In turn, estradiol enhances dopamine release and modifies the basal firing rate of 618 dopaminergic neurons (Becker, 1990;Pasqualini et al., 1995;Thompson and Moss, 1994), a 619 possible neurobiological mechanism by which alterations in estradiol could impact 620 cortical efficiency. Future multimodal neuroimaging studies in humans can clarify the link 621 between estradiol's ability to stimulate dopamine release and the hormone's ability to 622 drive cortical efficiency within PFC circuits.

623
Sixth, we observed surprisingly few autoregressive effects in brain measures across 624 our time-lagged models. This was despite relatively strong day-to-day similarity in 625 whole-brain patterns of connectivity (Figure S3), and clear evidence for autocorrelation 626 when assessing the brain data in an independent, univariate fashion. Thus, the 627 incorporation of sex hormones into a time-lagged modeling framework attributed more 628 temporal variability in the brain to fluctuations in hormone concentrations. Nevertheless, 629 an ongoing debate within the network neuroscience community surrounds test-retest 630 reliability in resting-state functional connectivity analyses. Some studies state that large 631 amounts of data (> 20 minutes) are necessary for test-retest reliability (Gratton et al.,632 2018a; Noble et al., 2017), while others argue that reliability can be derived from shorter 633 (5-15 minutes) scans (Birn et al., 2013;Van Dijk et al., 2010). We are limited in our ability to 634 assess whether the ostensibly-weak autoregressive trends suggested by our time-lagged 635 models would be replicated under longer scanning durations and hope future work 636 addresses this issue.

637
Finally, we chose to apply a well-established group-based atlas (Schaefer et al., 2018) 638 to improve generalizability to other individuals, as a key goal of our investigation was to 639 demonstrate how sex steroid hormones explain variability in intrinsic network topologies 640 based on regional definitions shown to be reliable across thousands of individuals 641 (Schaefer et al., 2018;Yeo et al., 2011). Yet, group-based atlases can lead to potential loss in 642 individual-level specificity, and recent work has demonstrated that fixed atlases may not 643 capture underlying reconfigurations in the parcellations themselves within an 644 individual (Bijsterbosch et al., 2019;Salehi et al., 2020a,b). Therefore, future work using 645 individual-derived functional networks will be necessary to determine whether spatial 646 reconfigurations in parcellations emerge as a function of the menstrual cycle, over and identifying the large-scale network disturbances that underlie, or predict, the emergence 710 of disease symptomology by incorporating sex-dependent variables (such as endocrine 711 status) into clinical models. This may be particularly true during periods of profound 712 neuroendocrine change (e.g. puberty, pregnancy, menopause, and use of hormone-based 713 medications, reviewed by Taylor et al. (2019)) given that these hormonal transitions are 714 associated with a heightened risk for mood disorders.