CSF estrogens' relationships to neuroinflammatory markers and brain networks in middle-aged and older black and white women.

This study was approved by the Emory University and Rutgers University Institutional Review Boards (IRB). All work was planned and conducted in accordance with the WMA’s Declaration of Helsinki as revised in 2024. Signed and dated informed consents were obtained from all participants. This parent study is longitudinal and on-going, and a cross-sectional design was used to characterize how estrogen levels correlate across biofluids (plasma, CSF), with CSF proteomics, and brain MRI volumes at baseline. Participants were recruited through health fairs, churches, Greek organizations, community presentations, Research Match, mailed research postcards, and from previous cohorts (PI Wharton) who agreed to future contact using Community Engaged Participatory Research (CBPR) principles. All research procedures take place at Emory University. Participants receive $300 total compensation. Participants completed questionnaires, cognitive testing, fasting phlebotomy and lumbar puncture (LP), MRI analysis, vascular ultrasound, and overnight sleep monitoring using a single-lead EEG (Sleep Profiler). Participants were eligible if between the ages of 45–85; no/minimal memory complaints or diagnosis of MCI; Montreal Cognitive Assessment (MoCA) ≥ 20 ( n  = 81 for ≥ 26, n  = 17 between 23 and 25, n  = 11 for < 23) [ 56 ]. Participants were excluded if there were contradiction for LP or MRI; residence in a skilled nursing facility; any significant systemic illness or unstable medical condition which could affect cognition, cause difficulty complying with the protocol, or consenting for study procedures; neurological disease which would influence cognition or CSF profiles (e.g., Parkinson’s disease, amyotrophic lateral sclerosis, multiple sclerosis, large territory stroke, traumatic brain injury); recent history of untreated major depression, substance use disorder; pregnant, nursing, or planning to get pregnant. Each participants completed questionnaires on medical and medication history, sleep (Epworth Sleepiness Scale), Patient Health Questionnaire 4, and Center for Epidemiologic Studies Depression Scale. In addition, they also completed a comprehensive questionnaire designed to characterize the sexual, reproductive and hormone-related history of middle-aged and older women, including current menopausal status and symptoms (if any), current or past hormone use (type, route), current use of aromatase inhibitors, and history of surgical hysterectomy or oophorectomy. Hormone replacement therapies (HRT) were classified as containing estrogen (E +), progestin (P +), or both. Neuropsychological analysis was performed by experienced technicians to assess domain-specific cognitive functions, including verbal memory (Buschke Selective Reminding Test [SRT]), attention/executive (Digit Span, Trail making test [TMT] A&B, Symbol Digit Substitution Test [SDST]), language (letter-guided fluency [FAS], Category Fluency-Animals), visuospatial (Judgement of Line Orientation [JOLO]) and overall (MoCA) functions. Up to twenty mL of CSF was collected via research LP using a 23.5 G non-traumatic Sprotte needle or a 22 G Quincke or Whitacre needle by a board-certified neurologist after overnight fasting. CSF was immediately (centrifuged at 4 °C for 10 min), aliquoted, labeled, and frozen until analysis. Approximately 45 mL of blood was collected on the same day in K 2 -EDTA tubes (for plasma) and SST tubes (for serum). For plasma, blood was centrifuged at 1000 g and 4 °C for 15 min before aliquoting, labeling, and freezing at −80°C until analysis. Plasma and CSF levels of estrone (E1), estradiol (E2), progestin, and sex hormone-binding globulin (SHBG) were measured using mass spectrometry (MS) at Mayo Clinic in a manner consistent with CLIA requirements. For estrogen, E2 and E1 were extracted from 0.5 mL of biofluid (plasma or CSF) using methylene chloride, and Deuterated 17β-estradiol-d5 and estrone-d4 were added to each sample before the liquid extraction as internal standards. After derivatization with dansyl chloride, sample extracts underwent high-pressure liquid chromatography (HPLC; Agilent Technologies 1290 Infinity UPLC) and then LC–MS/MS (Agilent Technologies 6490 Triple Quadrupole; ESI interface, operated in the multiple-reaction monitoring positive mode). Concentrations were derived on a nine-point standard curve (range of 0–100 pg/mL). For progesterone, levels were measured using a competitive electrochemiluminescence assay (Cobas e411, Roche Diagnostics) following manufacturer’s protocol. For SHBG, a microparticle-based immunoassay (Cobas e411, Roche Diagnostics) was used following manufacturer’s protocol. Concentrations for both progesterone and SHBG were determined via a instrument-generated calibration curve. The lowest level of quantitation (LLOQ) was 1.00 pg/mL for E1, 0.30 pg/mL for E2, 0.20 pg/mL for progesterone, and 5 pg/mL for SHBG. Concentrations below these levels were substituted with LLOQ/2 for subsequent analysis. A custom CSF targeted proteomic array (1,500 assays, with 33 targeting duplicate proteins to identify differences between assay versions; Table S2) was developed using proteins associated with inflammation, transcriptional regulation, ubiquitination, and organellar functions (e.g., lysosome) based on targets’ known biological functions as well as our firsthand experience measuring similar proteins from the greater SomaLogic CSF panel ( n  = 7,596 as of 2023). Frozen CSF samples from 96 participants having sufficient CSF were shipped from Emory University over-night on dry ice to Rutgers, randomized, and then sent to SomaLogic. Because CSF protein levels were influenced by freeze-thawing cycles during internal pilot experiments conducted by Rutgers and SomaLogic, samples were only thawed at time of analysis to minimize this effect. In addition to participant samples, five pairs of adjacent CSF aliquots were sent to undergo the same SomaScan. For each analyte, quality control (QC) steps included assessment for quantitation and intermediate precision. Limit of detection (LOD) was determined using a conservative threshold based on 10 blank samples (mean + 2 SD). Out of 1,467 analytes, 365 were below LOD and excluded from subsequent analysis, while 27 had coefficients of variation (CV) ≥ 20%. The final SomaScan panel for analysis included 734 proteins with CV < 10% and 341 with 10 ≤ CV < 20%. CSF AD biomarker (Aβ42, Aβ40, t-Tau) levels were measured using the fully automated Lumipulse G1200 using commercial kits (Fujirebio Europe, Ghent, Belgium). Aβ42/Aβ40 ratio is used as an indicator for AD neuropathologic changes (ADNC) as approved by the US FDA. A subset of participants (64 women and 24 men) had structural MRI of adequate quality performed at the Emory Center for Systems Imaging for detailed analysis. Those with MRI were similar to those without MRI (Supplementary Information). Among women, there was no difference in proportions reporting B/AA race/ethnicity ( p  = 0.600), menopause ( p  = 0.467), history of hysterectomy ( p  = 1.000), history of oophorectomy ( p  = 1.000), current use of E + HRT ( p  = 0.613) or P + HRT ( p  = 0.588). MRIs were conducted mostly at the same visit as blood and CSF were collected, but no more than three months after. High-resolution anatomical images were acquired using a T1-weighted magnetization-prepared rapid gradient-echo (MPRAGE) sequence. Imaging parameters included a repetition time (TR) of 2300 ms, an echo time (TE) of 2.95 ms, and an inversion time (TI) of 900 ms. The flip angle was set to 9 degrees. Images were obtained with a field of view (FOV) of 256 × 256 mm and a matrix size of 256 × 256, yielding a voxel size of 1.0 × 1.0 × 1.2 mm3. The slice thickness was 1.2 mm, with a total of 160 slices acquired. The acquisition time for this sequence was approximately 5 min. This sequence provided high-resolution anatomical images optimized for gray and white matter differentiation, suitable for volumetric and structural analyses. Cortical reconstruction and volumetric segmentation were performed with the FreeSurfer image analysis suite (version 7.4.1), which is documented and freely available for download online ( http://surfer.nmr.mgh.harvard.edu/ ). Volumetric estimates for regions of interest were derived based on the Automated Anatomical Labeling (AAL) atlas [ 57 ].

Between November 2020 and September 2024, 81 women (median age 64, range 45–78) and 28 men (median age 67.5, range 51–83) completed structured clinical, neuropsychological, brain MRI, plasma, and CSF analysis (Table  1 ). Approximately three-quarters of the women (28/38 B/AA and 32/39 NHW, p  = 0.377) reported to be post-menopausal, and 12 (14.8%) had a history of hysterectomy only ( n  = 7), oophorectomy only ( n  = 1), or both ( n  = 4). 10 women (12%) and no men were on HRT containing estrogen only ( n  = 4), progestin only ( n  = 3), or both ( n  = 3). Most participants had BMI ≥ 25 kg/m 2 (72% of women and 74% of men), and – consistent with our and others’ prior observations – B/AA participants had lower CSF t-Tau levels (standardized B = 0.249, 95% CI: 0.087, 0.411, p  = 0.003) and CSF t-Tau/Aβ42 ratio (standardized B = 0.387, 95% CI: 0.174,0.600, p  < 0.001; Table  1 ). 29 (27%) had CSF Aβ42/Aβ40 consistent with (≤ 0.055, n  = 18) or indeterminate for (between 0.055 and 0.075, n  = 11) ADNC. Using a robust demographically-adjusted normative data involving 3,366 NHW and 1,463 B/AA older (> 50) adults with no cognitive impairment enrolled in a prevention study for mild cognitive impairment [ 66 ], two NHW women and one NHW man had MoCA scores < 7th percentile, and four more participants (two NHW women, one B/AA woman, and one B/AA man) had score between the 7th and 16th percentile. None had abnormal scores in SRT, and there was no difference in proportion of participants having abnormal or intermediate CSF Aβ42/Aβ40 ratio between those with MoCA above or below the 16th percentile (equivalent to Z-score of-1.00, p  = 0.492). More detailed assessment using ten neuropsychological tests showed below-expectation attention (forward digit span, Z-score < −1.50) was the most common reason for having at least one abnormal test score (Supplementary Information). Compared to men, post-menopausal women had lower plasma E1 (log 10 -transformed mean 1.21 vs. 1.45, p  < 0.001, Fig.  1 A) and E2 (log 10 -transformed mean 0.54 vs. 1.33, p  < 0.001, Fig.  1 B) but higher plasma SHBG levels (log 10 -transformed mean 1.59 vs. 1.40, p  < 0.001). Among post-menopausal women not on E + HRT, B/AA participants had greater plasma E1 and E2 levels than NHW participants. We did not detect this difference according to race/ethnicity in the smaller group of men, consistent with an external study of over 12,000 men [ 67 ]. In both women (Pearson’s correlation coefficient ρ  = 0.846, p  < 0.001; consistent with prior finding [ 68 ]) and men ( ρ  = 0.712, p  < 0.001), plasma E1 and E2 levels were highly correlated. Fig. 1 Plasma E1 and E2 levels according to sex, race/ethnicity, and BMI. B/AA and NHW participants were examined according to biological sex, self-reported race/ethnicity, and self-reported menopausal states in women (Post-M: post-menopausal; M: menopausal; Peri-M: peri-menopausal; Pre-M: pre-menopausal; Unk: unknown menopausal status; A & B or BMI ( C & D ) for plasma E1 and E2 levels. For this analysis, women reporting other race/ethnicity ( n  = 4) were excluded. Hormone replacement therapy (HRT) is shown to contain estrogen (E +), progestin (P +), or both (E +/P +). * Greater in men than all women combined, p  < 0.01; †Greater in B/AA than NHW women not on E + HRT, p  < 0.05 Plasma E1 and E2 levels according to sex, race/ethnicity, and BMI. B/AA and NHW participants were examined according to biological sex, self-reported race/ethnicity, and self-reported menopausal states in women (Post-M: post-menopausal; M: menopausal; Peri-M: peri-menopausal; Pre-M: pre-menopausal; Unk: unknown menopausal status; A & B or BMI ( C & D ) for plasma E1 and E2 levels. For this analysis, women reporting other race/ethnicity ( n  = 4) were excluded. Hormone replacement therapy (HRT) is shown to contain estrogen (E +), progestin (P +), or both (E +/P +). * Greater in men than all women combined, p  < 0.01; †Greater in B/AA than NHW women not on E + HRT, p  < 0.05 Seven women (9%) on E + HRT had higher plasma E1 ( p  = 0.002), E2 ( p  = 0.046), and SHBG ( p  = 0.019) levels than those not on E + HRT. However, they also tended to be younger (58 vs. 63.5 yr, p  = 0.098) and have lower BMI (24.8 vs. 28.4 kg/m 2 , p  = 0.091). Linear regression analysis showed that higher plasma E1 levels was associated with lower BMI in NHW women (Fig. 1 C, 95% confidence interval of slope: −3.76, −0.64 for NHW vs. −0.061, + 3.02 for B/AA women, p  < 0.005), and this difference persisted after adjusting for E + HRT and post-menopausal status (Table  2 ). A similar trend was observed for plasma E2 (Fig.  1 D, 95% confidence interval of slope: −4.271, + 0.382 for NHW vs. + 1.033, + 6.169 for B/AA women, p  = 0.006), also persisting after adjusting for E + HRT and age (Table  2 ). Because only three (out of seven) women on E + HRT had plasma E1 levels higher than menopausal/post-menopausal women, and only one had plasma E2 levels higher than menopausal/post-menopausal women, none of the women on E + HRT were excluded from subsequent analyses. No effect of race/ethnicity was seen on plasma E1 or E2 levels in men. Table 2 Factors associated with plasma E2 and E1 levels from linear regression Factor influencing log 10 (plasma E1) B (95% CI) P log 10 (BMI) 0.723 (−0.832, 2.278) 0.357 Post-menopausal state −0.314 (−0.497, −0.130) 0.001 Estrogen-containing HRT 0.389 (0.117, 0.660) 0.006 Race/ethnicity  B/AA Reference  NHW 3.071 (0.135, 6.007) 0.041  Other −20.75 (−36.99, −4.528) 0.013 Race/ethnicity X BMI 0.003  B/AA X scaled log 10 (BMI) Reference  NHW X scaled log 10 (BMI) −2.215 (−4.420, −0.191) 0.032  Other X scaled log 10 (BMI) 15.15 (3.312, 26.99) 0.013  Factor influencing log 10 (plasma E2) B (95% CI) P  log 10 (BMI) 2.295 (0.032, 4.557) 0.047  Age (every 10 years after 40) −0.358 (−0.493, −0.222)  < 0.001  Estrogen-containing HRT 0.297 (−0.110, 0.704) 0.150 Race/ethnicity  B/AA Reference  NHW 5.141 (0.877, 9.404) 0.019  Other −17.069 (−41.26, 7.121) 0.164 Race/ethnicity X BMI 0.015  B/AA X scaled log 10 (BMI) Reference  NHW X scaled log 10 (BMI) −3.672 (−6.609, −0.735) 0.015  Other X scaled log 10 (BMI) 12.741 (−4.924, 30.41) 0.155 Factors associated with plasma E2 and E1 levels from linear regression There was a moderate correlation between CSF and plasma levels of both E1 ( ρ  = 0.600, p  < 0.001) and E2 ( ρ  = 0.554, p  < 0.001), but these may have been inflated by the number of participants with undetectable CSF estrogen levels (see Methods). High plasma E1 levels were associated with higher CSF E1 categories (O.R. = 7.48, 95% CI: 1.02–54.63, p  = 0.047 for detectable/low, and O.R. = 58.33, 95% CI: 5.07, 671, p  = 0.001 for detectable/high). Similarly, high plasma E2 levels were associated with higher CSF E2 categories (P.O. = 7.24, 95% CI: 2.16–24.29, p  = 0.001) . BMI correlated specifically with CSF E1 and age correlated specifically with CSF E2 (Table  3 ). After accounting for BMI and age, race/ethnicity did not appear to influence the relationship between plasma and CSF estrogen levels in this cohort (Table  3 ). Progesterone was not detectable in any of the CSF samples. Table 3 Association between CSF tertiles and plasma levels according to multi-nominal logistic regression analysis (for CSF E1, parallel assumption not met for ordinal regression) or ordinal regression analysis (for CSF E2). N.S. = not significant (factors with p ≥ 0.150 at the model level are excluded) Odds ratio for detectable/higher CSF E1 p Odds ratio for detectable/lower CSF E1 p Proportional odds for higher CSF E2 p log 10 (plasma E1) 58.33 (5.07, 671) 0.001 7.48 (1.02, 54.63) 0.047 N.S log 10 (BMI) 2.1 × 10 4 (5.0 × 10 –1 , 9.1 × 10 8 ) 0.067 3.0 × 103 (3.5 × 10 –1 , 2.6 × 10 7 ) 0.083 N.S log 10 (plasma E2) N.S N.S 7.24 (2.16, 24.29) 0.001 Age (every 10 years after 40) N.S N.S 2.02 (0.89, 4.57) 0.091 E(+)HRT N.S N.S 3.84 (0.73, 20.23) 0.112 Association between CSF tertiles and plasma levels according to multi-nominal logistic regression analysis (for CSF E1, parallel assumption not met for ordinal regression) or ordinal regression analysis (for CSF E2). N.S. = not significant (factors with p ≥ 0.150 at the model level are excluded) The association of BMI and older age with CSF estrogens in women is consistent with prior reports of non-gonadal estrogen sources including the brain. To identify brain processes potentially influenced by CSF estrogen levels, we first analyzed their relationships to CSF ADNC biomarker Aβ42/Aβ40 in all participants adjusting for age, sex, and BMI. Plasma E1 and E2 were associated with CSF Aβ42/Aβ40 only in men ( p  = 0.002 for E1, p  < 0.001 for E2) but not women ( p  = 0.570 for E1, p  = 0.197 for E2). Neither relationship was influenced by adjusting for age or BMI. On the other hand, CSF Aβ42/Aβ40 was associated with CSF E1 tertiles for women ( p  = 0.024) and potentially men ( p  = 0.065), but not CSF E2 tertiles. ANCOVA adjusting for age, sex, race/ethnicity, and BMI showed CSF Aβ42/Aβ40 to differ between CSF E1 categories (F(2,108) = 4.104, p  = 0.019) but not CSF E2 categories (F(2,108) = 0.030, p  = 0.971), with undetectable CSF E1 associated with more pathologic CSF Aβ42/Aβ40 (Fig.  2 B). Thus, neither plasma estrogens nor CSF E2 – the more biologically active estrogen – was associated with ADNC in women. Fig. 2 Relationship between plasma sex hormones, CSF sex hormones, and proteins related to age at menopause. CSF levels and E1 and E2 were detectable in fewer than half of the women ( A ), and were therefore subsequently analyzed as ordinal variables (0 = undetectable, L = detectable/lower, H = detectable/higher). CSF Aβ42/Aβ40 differed according to CSF E1 ( B ) but not E2 levels (age- and race/ethnicity-adjusted values shown, with dashed line corresponding to pathologic Aβ42/Aβ40 threshold for 65-year-old NHW participant). Hypothesis-driven analysis of gene products previously linked to age at menopause showed higher CSF E1 levels to associate with higher levels of nudix hydrolase 12 (NUDT12) and acyl-CoA dehydrogenase, long chain (ACADL); higher CSF E2 levels to associate with higher levels of capping actin protein, gelsolin-like (Capg) and interferon-induced protein 10 (IP10/CXCL10). * p  < 0.05 by ANCOVA; † p  < 0.05 and †† p  = 0.052 by multivariate multiple regression Relationship between plasma sex hormones, CSF sex hormones, and proteins related to age at menopause. CSF levels and E1 and E2 were detectable in fewer than half of the women ( A ), and were therefore subsequently analyzed as ordinal variables (0 = undetectable, L = detectable/lower, H = detectable/higher). CSF Aβ42/Aβ40 differed according to CSF E1 ( B ) but not E2 levels (age- and race/ethnicity-adjusted values shown, with dashed line corresponding to pathologic Aβ42/Aβ40 threshold for 65-year-old NHW participant). Hypothesis-driven analysis of gene products previously linked to age at menopause showed higher CSF E1 levels to associate with higher levels of nudix hydrolase 12 (NUDT12) and acyl-CoA dehydrogenase, long chain (ACADL); higher CSF E2 levels to associate with higher levels of capping actin protein, gelsolin-like (Capg) and interferon-induced protein 10 (IP10/CXCL10). * p  < 0.05 by ANCOVA; † p  < 0.05 and †† p  = 0.052 by multivariate multiple regression We then analyzed the levels of 1,075 CSF proteins in 96 of 109 participants using aptamer-based SomaScan (SomaLogic, Boulder, CO; see Methods) for proteomic correlates of CSF estrogen levels using two complementary approaches: (1) a hypothesis-driven analysis based on menopause-related genes from ROS/MAP, and (2) a data-driven approach using principal component analysis (PCA) of all measured proteins. The hypothesis- driven analysis based on 60 genes previously identified from post-mortem tissue in the ROS/MAP study [ 31 ] to associate with age at menopause (Table S1.1). MMR analysis accounting for age, race/ethnicity, CSF E1 or E2 tertiles, HRT, and multiple hypothesis testing showed 21 of 61 CSF analytes to correlate with at least one of these factors at the model level. Models including CSF E2 tertiles showed greater influence from demographic factors than models including CSF E1 tertiles (Tables S1.2-S1.8). Four proteins were directly associated with CSF E1 and E2 levels (Fig.  2 C): NUDT12 (nudix hydrolase 12) and ACADL (acyl-CoA dehydrogenase, long chain) for E1, and CAPG (capping actin protein, gelsolin-like) and IP10/CXCL10 levels for CSF E2 (Tables S1.9-S1.12). Including an interaction term between race/ethnicity and CSF E1 levels identified two additional proteins with race/ethnicity-specific associations: CFB (complement factor B) and Aβ/APP (a SomaScan measure targeting but only modestly and inversely correlated with AD-related Aβ42). While the above hypothesis-based approach isolated five identifiable CSF proteins which associated with CSF estrogen levels, it is difficult to extrapolate biological significance from these alone due to the limited targets examined [ 69 ] We thus leveraged a systems biology approach to identify CSF proteomic features associated with CSF estrogen tertiles, age, race/ethnicity, BMI, HRT, MoCA, and CSF Aβ42/Aβ40. Scaled levels of 1,075 CSF proteins were analyzed through PCA to generate 49 PCs (Table S2.1, Fig S2) representing protein groups or clusters which show co-variance across participants [ 61 ]. All PCs were then entered into an MMR to test for the effects of age, race/ethnicity, BMI, HRT, CSF estrogen tertile, MoCA, CSF Aβ42/Aβ40, and the interaction term between race/ethnicity and CSF estrogen tertile. For CSF E2, MMR showed five PC scores to have model-level significance at FDR < 0.05 (Table S2.2), with one associated with CSF E2 tertiles and race/ethnicity (PC43 = 5.468—3.927 × log 10 [BMI] + 0.446[if B/AA]—0.699[if detectable/high CSF E2] + 0.437[if detectable/low CSF E2], p  = 0.006; Fig.  3 A & D, Fig S1A & B; Table S2.3) and two with the interaction between race/ethnicity and CSF E2 tertiles (PC19 = 7.327–5.025 × log 10 [BMI]—0.839[if detectable/low CSF E2] + 1.299 [if B/AA and detectable/high CSF E2] + 1.105 [if B/AA and detectable/low CSF E2], Fig.  3 B & E; PC 42 = 1.745 [if B/AA and detectable/high CSF E2], Fig.  3 C & F; Tables S2.4 & S2.5). On the other hand, PC41 (Fig.  4 A & S1G-H; Table S2.6) and PC29 (Fig.  4 B & S1I-J; Table S2.7) were associated with the interaction between race/ethnicity and CSF E1 tertile, and PC29 was additionally associated with CSF E1 tertile (Table S2.7). We observed that, for each PC, PCA consolidated the collective effect sizes of age, BMI, race/ethnicity, and CSF estrogen levels on top loading proteins into the PC (Figs. 3 D & 4 C). It is notable here that two PCs associated with CSF E2 had top loading from neurodegenerative disease risk gene products (positive loading of APOE and negatively loading of BST1 for PC43, positive loading of CD33 for PC42) [ 70 – 72 ], and CSF Aβ42/Aβ40 was associated with PC41 but did not diminish its relationship with CSF E1. Repeating the analysis using a smaller number of PCs [ 33 ] resulted in consolidation of some PCs but similar correlations (Fig S3 & S4). Fig. 3 Relationship between CSF protein groups/modules and CSF E2 levels. Principal components (PCs) derived from targeted CSF proteomics identified three ( A - C ) to associate with CSF E2 tertiles (0 = undetectable, L = detectable/lower, H = detectable/higher; closed circles represent B/AA participants, and open circles represent NHW participants). Representative proteins positively or negatively loading onto each PC are shown (scores for the strongest loading protein in parentheses). Gene Ontology biological processes associated with each PC were identified according to number of proteins included in each process and enrichment signal, with stringent thresholds for enrichment signal (> = 1.0) and false discovery rate (FDR) given the smaller gene sets from each PC entered into STRING-DB. Top generic process such as “inflammatory response” or “immune response” for each PC was included as reference for signal, protein count, and FDR. Effect sizes from representative proteins loading onto each PC are shown according to effects from independent variables, with negative loading proteins’ signs reversed (*) Fig. 4 Relationship between CSF protein groups/modules and CSF E1 levels. Principal components (PCs) derived from targeted CSF proteomics identified two ( A , B ) to associate with CSF E1 tertiles (0 = undetectable, L = detectable/lower, H = detectable/higher; closed circles represent B/AA participants, and open circles represent NHW participants). Representative proteins positively or negatively loading onto each PC are shown (scores for the strongest loading protein in parentheses). Gene Ontology biological processes associated with each PC were identified according to number of proteins included in each process and enrichment signal, with stringent thresholds for enrichment signal (> = 1.0) and false discovery rate (FDR) given the smaller gene sets from each PC entered into STRING-DB. Top generic process such as “inflammatory response” or “immune response” for each PC was included as reference for signal, protein count, and FDR. Effect sizes from representative proteins loading onto each PC are shown according to effects from independent variables, with negative loading proteins’ signs reversed (*) Relationship between CSF protein groups/modules and CSF E2 levels. Principal components (PCs) derived from targeted CSF proteomics identified three ( A - C ) to associate with CSF E2 tertiles (0 = undetectable, L = detectable/lower, H = detectable/higher; closed circles represent B/AA participants, and open circles represent NHW participants). Representative proteins positively or negatively loading onto each PC are shown (scores for the strongest loading protein in parentheses). Gene Ontology biological processes associated with each PC were identified according to number of proteins included in each process and enrichment signal, with stringent thresholds for enrichment signal (> = 1.0) and false discovery rate (FDR) given the smaller gene sets from each PC entered into STRING-DB. Top generic process such as “inflammatory response” or “immune response” for each PC was included as reference for signal, protein count, and FDR. Effect sizes from representative proteins loading onto each PC are shown according to effects from independent variables, with negative loading proteins’ signs reversed (*) Relationship between CSF protein groups/modules and CSF E1 levels. Principal components (PCs) derived from targeted CSF proteomics identified two ( A , B ) to associate with CSF E1 tertiles (0 = undetectable, L = detectable/lower, H = detectable/higher; closed circles represent B/AA participants, and open circles represent NHW participants). Representative proteins positively or negatively loading onto each PC are shown (scores for the strongest loading protein in parentheses). Gene Ontology biological processes associated with each PC were identified according to number of proteins included in each process and enrichment signal, with stringent thresholds for enrichment signal (> = 1.0) and false discovery rate (FDR) given the smaller gene sets from each PC entered into STRING-DB. Top generic process such as “inflammatory response” or “immune response” for each PC was included as reference for signal, protein count, and FDR. Effect sizes from representative proteins loading onto each PC are shown according to effects from independent variables, with negative loading proteins’ signs reversed (*) Because of the disproportionate number of participants with undetectable CSF estrogen levels, we performed post-hoc analysis of relationships between plasma estrogen levels and these CSF PCs. When plasma E2 levels and E + HRT (two factors influencing CSF E2 levels) were included instead of CSF E2 tertiles, we also found PC43 to associate with plasma E2 ( p  = 0.027) as well as E + HRT ( p  = 0.039), PC19 with only E + HRT ( p  = 0.019), and PC42 with only race/ethnicity ( p  = 0.045). When plasma E1 levels were included in the analysis in place of CSF E1 tertiles, we found PC29 to associate with plasma E1 ( p  = 0.077) but PC41 to associate with race/ethnicity ( p  = 0.024) and age ( p  = 0.003). To better understand the biological functions represented by protein groups associated with CSF E2 and E1, we conducted pathway enrichment analysis using STRING-DB, a widely used resource for mapping protein interactions and enriched biological processes [ 65 ]. Given the intentional overrepresentation of inflammatory proteins in our SomaScan panel, we prioritized identifying more specific biological processes in Gene Ontology (GO) beyond general immune/inflammatory functions, which were only included for comparison with more specific immune-related pathways. More GO biological processes were identified by STRING-DB to associate with CSF E2 than CSF E1 tertiles, but leukocyte proliferation/chemotaxis was represented in four out of the five PCs. At the same time, a closer examination of CSF analytes linked to these GO terms showed functional and/or structural specificity within each PC which informs potential mechanisms underlying these CSF protein groups/modules: analytes associated with leukocyte proliferation (e.g., bone marrow stromal cell antigen 1 [BST1], leptin [LEP], migration inhibitory factor [MIF], CD81) negatively loaded onto PC43; CC chemokines – C6-CC chemokines in particular, including CCL15 positively and CCL21/CCL23 negatively – loaded onto PC42; [ 73 ] CXC chemokines sharing receptor CXCR2 (CXCL1, CXCL2, CXCL3, associated with neutrophil chemotaxis [ 74 – 76 ]) positively and CXC chemokines sharing receptor CXCR3 (CXCL9 CXCL10, CXCL11, associated with monocyte chemotaxis) negatively loaded onto PC41 [ 77 ]. Other unique processes associated with CSF E2 included positive loading of ERK1/2 cascade on PC43 (apoE, ACE2, Notch2; Fig.  3 A), positive loading of glia-related proteins on PC19 (RAGE/AGER, GRN), negative loading of complement regulation on PC19 (C2, C3, C5, C9, CFB, CFI, Fig.  3 B), and positive loading of markers for classical monocytes on PC42 (siglec-3/CD33, IL-13Ra1) [ 78 ]. We interpret these findings to suggest that, among women with undetectable CSF estrogen, B/AA participants show greater neuroinflammatory markers associated with neutrophil chemotaxis than NHW participants which diminish with higher CSF estrogen levels. Conversely, higher CSF estrogen levels are accompanied by markers associated with complement activation in B/AA and monocyte/microglia activation in NHW participants. Finally, because estrogen can be secreted by or stored in cerebral structures, we examined whether CSF E2 levels associated with measurable differences in regional brain volumes among a subset of women who underwent MRI analysis (45, 11, and 8 women with undetectable, lower, and higher CSF E2 levels, respectively). Cortical volumes measured using FreeSurfer were first analyzed by PCA to identify eight PCs (Fig S7), with five PCs having significant loading from mostly symmetric bilateral structures (Table S3). Six PCs correspond to previously described brain networks including somatosensory (PC1), default mode (DMN; PC2), limbic (PC3), visual (PC4), frontal pole auto-association [ 79 ], and posterior-medial (PC7) [ 80 , 81 ] networks. The remaining two PCs correspond to the salience network (PC5, relatively more loading from left than right) and a network previously linked to movement initiation (PC8) [ 82 ]. MMR of the eight cortical volume PC scores co-varying for total intracranial volume, age, race/ethnicity, BMI, HRT, and MoCA showed PC1 and PC7 scores to associate with CSF E2 tertile ( p  = 0.015 for Roy’s largest root, p  = 0.094 for Wilks’ lambda). Compared to participants with undetectable CSF E2 levels (Table S3), participants with detectable/lower CSF E2 levels had smaller PC1 (somatosensory network, −0.700 [−1.360, −0.040], p  = 0.038, Fig.  5 A) and participants with the detectable/higher CSF E2 levels had smaller PC7 (posterior medial, −0.963 [−1.718, −0.208], p  = 0.013, Fig.  5 B) volumes involving the entorhinal and parahippocampal regions, even after controlling for total intracranial volume and borderline effect from race/ethnicity. These are consistent with differences seen between men and women in the UK Biobank [ 83 ]. For CSF E1, participants with detectable but lower levels had smaller PC2/DMN volumes (−0.899 [−1.551, −0.247], p  = 0.04, Fig.  5 C). These relationships were not replicated when plasma estrogens were used in place of CSF estrogen. A closer examination showed that whereas CSF estrogens’ had direct effects on PC1, PC2, and PC7 volumes after adjusting for other factors, plasma estrogens only modified the effects of another variable (race for PC1) or had no effects at all (Table S3.2-S3.4). Fig. 5 MRI correlates of CSF E2 and E1 levels. MRI principal components (PCs) corresponding to somatosensory (PC1, A ) and posterior medial (PC7, B ) networks differed according to CSF E2 tertiles (* p  < 0.05 in post-hoc analysis; 0 = undetectable, L = detectable/lower, H = detectable/higher). MRI PC corresponding to default mode network ( C ) also differed according to CSF E1 tertiles (†), but this was mediated by CSF Aβ42/Aβ40 level MRI correlates of CSF E2 and E1 levels. MRI principal components (PCs) corresponding to somatosensory (PC1, A ) and posterior medial (PC7, B ) networks differed according to CSF E2 tertiles (* p  < 0.05 in post-hoc analysis; 0 = undetectable, L = detectable/lower, H = detectable/higher). MRI PC corresponding to default mode network ( C ) also differed according to CSF E1 tertiles (†), but this was mediated by CSF Aβ42/Aβ40 level

The role of sex hormones on menopausal and post-menopausal women’s brain health remains controversial despite negative outcomes from the Women’s Health Initiative (WHI) and the Kronos Estrogen Prevention Study (KEEPS) [ 17 – 20 , 84 , 85 ]. Here we prospectively measured compartment-specific E1 and E2 levels in a group of mostly post-menopausal B/AA and NHW women, and found only modest correlation between plasma and CSF levels. Using targeted CSF proteomics and volumetric MRI analysis, we also found higher CSF estrogen levels to correlate with neuroinflammatory markers in keeping with several AD-related changes according to race/ethnicity, and volumes differences in two networks previously linked to differ between men and women independent of race/ethnicity. We discuss these findings below. Technical challenges in measuring estrogen levels are not new, and plasma measurements have undergone greater assay standardization than CSF measurements. Radioimmunoassay-based studies previously showed moderate correlation between CSF and plasma E2 across ages in women [ 86 – 88 ], but a LC–MS/MS study showed stronger CSF-blood correlation for E1 than E2 in a small sample size [ 89 ]. Because we observed race/ethnicity, age, and BMI to influence plasma and CSF estrogen levels, study participant characteristics very likely could lead to different effect sizes across studies. Notwithstanding these caveats, our data support positive – albeit modest – relationships between plasma and CSF estrogen levels [ 90 ]. While low CSF estrogen levels limit in-depth investigation into their source (brain vs. systemic circulation) and impact, targeted CSF proteomics here – favored over LC/MS–MS to investigate proteins of low abundance such as chemokines and cytokines – begin to identify protein groups/modules as correlates of each CSF estrogen state. Several proteins linked to CSF estrogens here were previously implicated in models of menopause. Experimental knock-down of estrogen receptor α resulted in increased gene expression of apoE- and RAGE-related pathways [ 91 ]. This is most in keeping with increased PC43 (top loading protein of apoE, ATG5, MIF, BST1) and PC19 (top loading protein of sRAGE) scores in B/AA women with undetectable CSF E2. Also in support of this relationship are previous findings that ATG5 has a tissue- and condition-specific bidirectional relationship with estrogen [ 92 ], and estrogen potently downregulates MIF expression [ 93 ]. For proteins loading onto PC19, RAGE is a cell surface receptor whose activation on immune cells leads to cytokine and chemokine releases [ 94 ], but it also mediates the influx of AD-related Aβ peptides across the blood brain barrier on the brain endothelial cell surface [ 95 ] and intraneuronal Aβ uptake on the neuronal surface [ 96 ]. RAGE gene expression was increased in the brain of ovariectomized rats [ 97 , 98 ], and alternative splicing or ectodomain shedding creates sRAGE as a soluble decoy receptor for cell surface RAGE. Despite these findings, it is not clear why no association (PC43) or even the opposite association (PC19) with CSF E2 was observed in NHW women. Chemokines, cytokines, and complement proteins associated with PC19 and PC42 may provide additional insight into our observations. In keeping with our pathway enrichment analysis, negative loading of CRP, C9, Factor B, and C3d onto PC19 suggests the greater score to favor the alternative (reduced CRP with depletion of C9, Factor B, C3d) over classical complement pathway. Previously observed negative effects of E2 and ERα-selective receptor agonists on membrane attack complexes [ 99 ] again agree with observation in B/AA women but not NHW women. Related to this, frailty-associated levels of CD5 and CXCL6 in extracellular vesicles were recently found to be higher in NHW than B/AA women [ 100 ]. These race/ethnicity-based differences in CSF inflammatory proteins are consistent with our prior reports [ 39 , 55 ], and we now also expand the list to include CXCL1, CXCL2, and CXCL3 (PC41, Figs. 3 D, 4 C, S5) as well as EDA whose receptor is a tissue-independent hallmark of aging [ 101 ]. Also associated with CSF estrogen levels are multiple Siglecs including CD33/Siglec-3 which is a well-known genetic driver for AD risks [ 72 ]. Siglec proteins are leukocyte-expressed cell surface receptors of sialic acid which can up or downregulate immune responses. The protective CD33/Siglec isoform has been shown in a mouse AD model to compacts extracellular plaques as well as enhance microglia-mediated Aβ phagocytosis [ 102 , 103 ]. Adding to this was a recent spatial proteomics analysis in human brain associating AD tissue with enrichment of microglia-expressed Siglec-3/CD33 coupled with reduced apoE expression [ 104 ]. Interestingly, we observed this pattern in our cohort (PC43, PC42) to associate with high CSF E2 levels in B/AA women, a group with high AD risks. Among other Siglecs, Siglec-5 was the CSF analyte most strongly associated with CSF E1 levels (Fig.  4 C & S5). Cell surface Siglec-5 [ 105 ] and extracellular Aβ peptides [ 106 ] both interact with extracellular Hsp70, suggesting that soluble Siglec-5 might compete for Hsp70 – as it does for PSGL1 [ 107 ] to antagonize its leukocyte recruitment – to create more permissive conditions for higher order Aβ species. Even so, the impact of estrogen levels on these AD-related observations remains unclear. At least one Siglec (CD22/Siglec-2) downregulates B cells upon E2 stimulation [ 108 – 110 ]. Because CSF levels were only available for four Siglecs (−3, −5, −7 [which loaded negatively onto PC43], −9), our work here could not determine the if some of these relationships generalized to other Siglecs. Understanding of soluble Siglecs’ biology also continues to evolve [ 111 ], with majority of recent association studies between estrogen and Siglecs occurring in breast cancer models [ 112 – 115 ]. Similarly, little of C11orf49/CSTPP1 – associated with E1 in post-hoc analysis (Fig S5) – biology is well established [ 116 ]. C11orf49/CSTPP1 has a promoter region which is hyperactive in mast cells [ 117 ], which are activated by estrogen [ 118 , 119 ] and implicated in endometriosis as well as sexual dimorphism of allergic diseases [ 120 ]. Siglec-5 and IL-9 (loading onto E1-related PC29 and PC41) are also expressed by mast cell [ 121 , 122 ], but a microglial source for all remains possible as we previously detected C11orf49 transcripts in CSF macrophages and monocytes [ 52 ]. CSF estrogen’s relationships with multiple Siglecs here should stimulate more investigation into the latter’s role in aging and menopause, especially since Siglec-15 is already being targeted in osteoporosis treatment [ 123 ]. We were not able to perform meaningful pathway analysis using menopause-related genes from ROS/MAP due to the small number of CSF estrogen correlates. Three of the four proteins identified through this process were most different between women with extreme CSF estrogen levels: NUDT12, a hydrolase for nicotinamide adenine dinucleotide -capped messenger RNAs [ 124 ]; ACADL, a mitochondrial enzyme which may share epitope with several brain-specific isoforms [ 125 ]; and CXCL10/IP-10 which we [ 39 , 126 ] and others [ 127 ] previously linked to inflammaging as well as AD. The fourth protein, Capg, was the exception. Capg blocks the ends of actin to regulate motility of non-muscle cells, and an increase in its CSF level has been linked to microglia activation [ 128 ]. The low CSF Capg levels in women with intermediate CSF E2 levels independent of race/ethnicity is thus most reminiscent of MRI PCs’ associations with CSF estrogens. Because plasma E2 levels did not differ between those having undetectable and detectable but low CSF E2 levels, a dissociation between CSF and plasma E2 – not seen for E1 – may reflect feedback regulation of activated microglia by E2 [ 129 ]. Together, the above proteome-derived findings from PCA and ROS/MAP generate new hypotheses relating to racial AD disparities, and should be thoroughly verified using cell-based techniques such as CSF single cell RNA sequencing [ 52 ]. Aside from the possibility that estrogen mechanistically biases towards adaptive immunity [ 130 ], it is similarly possible that estrogen alterations result from neuronal and glial alterations. Neurons can synthesize estrogen [ 131 ], and brain injury is accompanied by increased aromatase expression as well as estrogen release by astrocytes [ 132 – 134 ]. Our brain volumetric analysis seems to support this explanation, showing smaller volumes in the somatosensory and posterior-medial networks for those with intermediate CSF E2 or E1 tertiles. While a comparison with those having low levels appears consistent with the finding from WHI that women on conjugated estrogen HRT experienced greater brain atrophy [ 135 ], estrogen-associated PCs loaded by neurodegenerative marker NfL (PC42) and A2 astrocyte marker S100A10 (PC19) showed changes in the opposite directions of neuronal injury in our study. As discussed above, intermediate CSF E2 levels – without increase in plasma E2 compared to those with low CSF E2 levels – may be associated with microglial activation rather than neurodegeneration. What’s more, inclusion of MoCA to account for cognitive decline or neurodegeneration did not show a mediator effect between CSF estrogen tertiles and brain volume, and AD has been previously characterized by lower – not higher – brain [ 136 ] as well as circulating E2 levels [ 137 ]. Thus, the dissociation between CSF and plasma E2 in those with detectable but low CSF E2 may represent an intermediate state preceding neurodegeneration. Potentially consistent with our observation are the prior findings that 1) estrogen can enhance the connectivity between the parahippocampal gyrus (part of the posterior-medial network enriched with estrogen receptors [ 138 ]) and other brain regions [ 139 ], and 2) increased connectivity can associate with lower regional gray matter volume via higher intracortical myelin content [ 140 ]. This alternate explanation should be tested in the future by examining white matter content and functional connectivity within the somatosensory and posterior-medial networks. If so, investigators should practice caution in interpreting gray matter volume-only findings especially in healthy older adults. Our study generated multiple testable hypotheses for larger replication and model-based follow-up studies as well as new protein markers to improve the interpretation of mixed HRT studies [ 141 – 145 ], but has a number of limitations. The number of participants in each group and the proportion on HRT are small compared to prior epidemiological or intervention studies related to estrogen, but the in-depth CSF proteomic analysis conducted here is not feasible in large studies. For the same reason, internal cross-validation was not practical. We included small numbers of women with unclear and pre-menopausal states as well as those on HRT, which increased the cohort’s heterogeneity and complexity in interpretation even if these factors were included as independent variables. Inclusion of women whose MoCA scores placed them in the MCI range also risks additional confounding effects, despite our inclusion of CSF Aβ42/Aβ40 as a marker for Alzheimer’s Disease Neuropathologic Changes. The technical sensitivity for CSF E2 and E1 is low compared to normative levels in this population, limiting our ability to fully model the effect of CSF estrogen on other CSF proteins on a continuous (rather than ordinal) scale. CSF progesterone was also not detectable. However, plasma E1 and E2 levels did not differ between women with undetectable and low CSF E2 levels, while they did between women with undetectable and low CSF E1 levels. This floor effect of plasma E2 – much more commonly measured than CSF E2 – may suggest a relatively greater sensitivity for CSF E2. While we relied on bioinformatics to speculate on biological processes underling CSF protein changes, we did not provide cellular or other corollary to test these hypotheses. Even though we demonstrated an association between CSF estrogen and brain volumes, we could not identify the same relationship when plasma estrogen was used in place of CSF estrogen. Finally, we did not include genetic ancestry nor social determinants of health to further explore whether the race/ethnicity-associated modifications resulted from genetic or environmental differences, although the often strong collinearity between these factors calls for a well-balanced cohort in the future to address causal factors. In conclusion, we used proteomic, MRI, and statistical procedures to identify race/ethnicity-based differences on plasma estrogen levels and the associations between CSF estrogen level and immune processes, but did not find race/ethnicity to influence network brain volumes. These results provide insight into estrogen-associated biological processes, reinforce the need to recruit diverse participants, and encourage caution when relying on plasma estrogen levels alone to examine sex hormones’ impact on the brain. Investigators should continue to explore high sensitivity assays to better detect CSF estrogen forms, and prospectively test the pathways hypothesized here in a replication cohort as well as other models towards future therapies to reduce AD-related disparities.

Women and Black/African Americans (B/AA) represent two groups facing disproportionate risk for Alzheimer’s disease (AD) dementia, and the causes are complex [ 1 ]. For women, their rates of brain atrophy due to aging or AD differ from men [ 2 – 5 ]. Women in the middle-to-older age transition (45–65 years) are also more likely to have positive imaging biomarkers for brain amyloid and neuro-degeneration [ 6 , 7 ]. The most frequently examined factor to account for these AD or dementia disparities between women and men has been exposure to sex hormones across the life span [ 8 , 9 ], as brain gray matter morphology, white matter hyperintensity, and functional connectivity all undergo detectable changes during the menstrual cycle or menopause, and exposure to exogenous sex hormones [ 10 – 15 ]. This is consistent with the post-mortem finding that earlier surgical menopause was associated with greater AD neuropathology [ 16 ]. While early observational studies linked hormone replacement therapies (HRT) with reduced dementia risks [ 17 , 18 ], two large randomized controlled trials failed to demonstrate reduced dementia risks with HRT during early [ 19 ] or late [ 20 ] menopause – including doubling of dementia risks for women on estrogen-plus-progesterone HRT in Women’s Health Initiative Memory Study [ 20 ]. However, a recent meta-analysis using data from over 40,000 HRT- or placebo-treated participants linked mid-life use of estrogen-only HRT to reduced AD or dementia risks, and late-life use of estrogen-plus-progestogen therapy to increased AD or dementia risks [ 21 ]. The life course timing of estrogen deprivation or exposure relative to emergence of AD-related changes thus remains key to elucidation of biological events underlying the relationship between systemic and brain-specific processes. In the brain, estrogen receptors are expressed in neurons as well as glia [ 22 , 23 ]. Estrogen receptor binding leads to nuclear localization and transcription or membrane signaling independent of transcription [ 24 , 25 ]. Ovariectomy in rats and non-human primates result in transcriptional changes involving neurotransmitter receptors, neuropeptides, growth factors, and especially inflammatory mediators [ 26 – 28 ]. In humans, the Religious Order Study and Memory and Aging Project (ROSMAP; mostly non-Hispanic White participants) combines one large longitudinal US national study and one large Illinois-based study to provide rich neuropathologic data in over 2,500 participants, primarily involving dorsolateral prefrontal cortical tissue [ 29 , 30 ]. Post-mortem transcriptomics in ROSMAP linked age at menopause (natural or surgical) to altered expression of 2,685 genes, with an overall genetic correlation that earlier menopause is associated with poorer cognition [ 31 ]. Compared to post-mortem work, limited biofluid markers of metabolism and inflammation have been examined in relation to menopause or related symptoms [ 32 – 34 ]. However, there are key knowledge gaps between human blood-based and post-mortem neuropathologic studies to elucidate biochemical changes in the brain [ 35 – 37 ]. Proteins in the cerebrospinal fluid (CSF) better reflect brain-associated biological processes [ 38 , 39 ], yet we are not aware of any study having paired CSF sex hormone with CSF proteomic analysis relevant to AD biomarkers and neuroinflammatory markers. It is also incompletely understood why older B/AA adults have 64% greater lifetime dementia risks than non-Hispanic White (NHW) adults [ 40 ]. Differences in early life education quality may account for disparate performance on standardized cognitive measures [ 41 , 42 ] and cognitive reserve, but social determinants of health inadequately account for race/ethnicity-based AD health disparities in the United States [ 43 ]. Compared to NHW women, B/AA women experience younger age of menarche [ 44 ], higher menstrual E2 levels [ 45 ], and earlier entry into menopause [ 46 ]. Over the past 20 years, B/AA women have also experienced greater increase in breast cancer incidence [ 47 ], higher rates of surgical menopause [ 48 ], nearly four more years of menopausal vasomotor symptoms [ 49 ], and slower rates of decline in post-menopausal hormone use [ 50 ] than NHW women. The mid-to-late life transition – strongly associated with one major wave of accelerated molecular aging [ 51 ] – thus not only provides an important baseline to study divergent trajectories of brain health into the future but also an opportunity to investigate biological correlates of potentially disparate aging-related changes before diverging AD outcomes. To accomplish both of these goals, we designed and conducted a study to examine the relationships among sex assigned at birth (hereafter referred to as sex), age, CSF sex hormones, and 1,075 CSF proteins related to inflammation, ubiquitination, cellular signaling, RNA processing, and other biological processes [ 52 ]. Leveraging our success in enrolling older B/AA and NHW participants (median age of 70–75) into modern research to discover biological correlates of AD health disparities [ 39 , 53 – 55 ], we recruited approximately equal numbers of B/AA and NHW women with no cognitive impairment. We previously used this balanced design to identify race/ethnicity-related differences in markers for tau pathology [ 39 , 53 ], AD-related inflammation [ 39 ], and blood–brain barrier integrity [ 55 ]. By introducing plasma and CSF measures of estradiol (E2) and estrone (E1), we sought to examine CSF protein clusters or groups which vary according to estrogen levels, race/ethnicity, or both.

Participants with missing plasma or CSF E1 and E2 data were excluded from the baseline and subsequent analyses. Thirteen participants had sufficient CSF aliquots for AD biomarkers and sex hormone measurements, but did not have sufficient CSF aliquots available for detailed proteomic analysis. They were excluded for analysis involving the proteomic panel. Compared to those with proteomic measures, these 13 had younger age (Mann–Whitney U of 866, p  = 0.023), but were otherwise similar in MoCA (U of 669, p  = 0.668), CSF Aβ42/Aβ40 ratio (U of 477, p  = 0.169), plasma E1 levels (U of 608, p  = 0.973), plasma E2 levels (U of 409, p  = 0.060), sex (X 2  = 1.262, p  = 0.261), race/ethnicity, CSF E1 tertiles, and CSF E2 tertiles (Fisher’s exact probability of 0.465, 0.108, and 0.540). All tests are two-tailed, and coefficient values are shown with 95% confidence intervals (95% CI). We used p  < 0.05 as the nominal threshold for statistical significance, but present a wider range of p-values for transparency [ 58 ]. All demographic and clinical data were analyzed using IBM SPSS 28 (Aramonk, NY). For baseline comparisons, Chi-squared tests or Fisher’s probability tests were used for categorical variables. The one-sample Kolmogorov–Smirnov (KS) test determined that age, Buschke SRT delayed recall, letter fluency (FAS), category fluency (animals), and SDST met the normality assumption. Age was thus examined using analysis of co-variance (ANCOVA) for differences according to sex and race/ethnicity (Table  1 ). For cognitive scores, adjusted scores were used to determine differences according to sex and race/ethnicity. Generalized linear models were used to examine adjusted cognitive Z-scores or scaled scores as well as CSF AD biomarker levels regardless of if they met the assumption of normal distributions for consistency. Table 1 Demographic and sex hormone levels of participants Women Men B/AA ( n  = 38) NHW ( n  = 39) Other ( n  = 4) B/AA ( n  = 13) NHW ( n  = 15) Age 62.5 (57, 70) 65 (61, 71) 57 (48, 64) 68 (63, 73.5) 67 (57, 71) BMI 28.6 (26.1, 31.9) 26.6 (23,0, 30.4) 22.3 (20.9, 25) 26.7 (24.8, 29.4) 26.0 (24.4, 29.3) MoCA 27 (24, 28) 28 (26.25, 29) 30 (30, 30) 26 (23.3, 26.8) 27 (26, 29) SRT Total 46.5 (40, 50.3) 48 (45, 55) 49 (47.3, 53) 39 (29, 43) 42 (29, 51) SRT Delayed Recall 6 (4, 8) 7 (5, 9) 10.5(8.5, 11.8) 3 (2.5, 5.5) 7 (4, 8) Forward Digit Span 7 (5.8, 8) 8 (7, 8) 5.5 (5, 7.5) 7 (6, 7) 7 (6, 7) Reverse Digit Span 5 (4, 5) 5 (4, 7) 6 (4.5, 6.8) 4 (3.5, 5) 5 (4, 6) Letter fluency(FAS)*† 43 (37.8, 52.3) 44 (33, 57) 55 (46, 61.8) 36 (28, 53) 41 (30, 47) Animal fluency 21.5 (17, 23) 23 (19, 26) 27 (18.8, 37.5) 19 (16, 22) 23 (18, 26) Trail Making Test A 29.5 (23, 36.5) 26 (21, 32) 22 (16.8, 27.8) 37 (31, 45) 26 (23, 31) Trail Making Test B* 72 (63.8, 95) 66 (56, 83) 46 (37.8, 53.5) 90 (69, 151) 71 (58, 108) JOLO 18 (13.8, 23) 25 (18, 27) 26 (21, 28.8) 19 (15.5, 27) 25 (24, 27) Menopausal State N/A N/A  Post-menopausal 28 (74%) 32 (82%) 2 (50%)  Menopausal 3 (8%) 3 (8%) 1 (25%)  Peri-menopausal 5 (13%) 2 (5%) 0  Pre-menopausal 2 (5%) 1 (3%) 1 (25%)  Unknown 0 1 (3%) 0  Hysterectomy 5 (13%) 6 (15%) 0 N/A N/A  Oophorectomy 2 (5%) 3 (8%) 0 HRT containing  Estrogen 1 (3%) 6 (15%) 0 0 0  Progestin 2 (5%) 4 (10%) 0 0 0 Plasma hormone  E1 20.5 (14.3, 20.5) 14 (9.9, 22) 12 28 (21, 49) 25 (15, 35)  E2 4.45 (3.02, 9.23) 2.6 (1.8, 5.8) 5.0 24 (17, 29) 21 (15, 27)  Progesterone ND ND ND ND ND  SHBG 34.2 (23.9, 47.5) 44.0 (29.3, 60.8) 64 (31, 125) 25.9 (19.6, 37.2) 26.4 (21.7, 35.0) CSF AD biomarkers  Aβ42 (pg/mL) 604 (507, 761) 676 (445, 848) 418 (408, 477) 573 (514, 943) 509 (437, 723)  Aβ40 (ng/mL) 7.0 (5.5, 9.1) 9.2 (7.7, 10.4) 5.8 (5.0, 7.8) 6.3 (5.5, 11.7) 7.3 (5.9, 10.1)  t-Tau (pg/mL) †† 230 (177, 285) 277 (218, 390) 143 (131, 230) 271 (227, 341) 277 (218, 562)  Aβ42/Aβ40 0.09 (0.08, 0.10) 0.08 (0.05, 0.10) 0.09 (0.07, 0.10) 0.09 (0.08, 0.10) 0.08 (0.05, 0.10)  t-Tau/Aβ42 †† 0.35 (0.30, 0.41) 0.41 (0.28, 0.78) 0.28 (0.27, 0.54) 0.41 (0.30, 0.52) 0.42 (0.34, 1.30)  MRI analysis 28 (74%) 32 (82%) 100% 9 (69%) 15 (80%) Median and interquartile ranges are shown for continuousvariables. Raw scores for neuropsychological tests areshown, but group-level differences were analyzed using age, sex, andrace/ethnicity-adjusted scores in generalized linear models MoCA Montreal Cognitive Assessment, SRT Buschke Selective Reminding Test, JOLO Judgement of Line Orientation, HRT hormone replacement therapy, E1 estrone, E2 estradiol, SHBG sex hormone bindingglobulin, ND not detected except in 3 women’s samples, N/A  not applicable tomen * Greater in B/AA than NHW, p <0.001, †greater in women, p =0.003 (see also Supplementary Information), †† greaterin NHW than B/AA, p <0.005 Demographic and sex hormone levels of participants Median and interquartile ranges are shown for continuousvariables. Raw scores for neuropsychological tests areshown, but group-level differences were analyzed using age, sex, andrace/ethnicity-adjusted scores in generalized linear models MoCA Montreal Cognitive Assessment, SRT Buschke Selective Reminding Test, JOLO Judgement of Line Orientation, HRT hormone replacement therapy, E1 estrone, E2 estradiol, SHBG sex hormone bindingglobulin, ND not detected except in 3 women’s samples, N/A  not applicable tomen * Greater in B/AA than NHW, p <0.001, †greater in women, p =0.003 (see also Supplementary Information), †† greaterin NHW than B/AA, p <0.005 Plasma E1 and E2 levels were log 10 -transformed before analysis due to non-normal distribution. Analysis of CSF E1 and E2 was limited by the high number of samples having undetectable levels or floor effects (13 of 28 men and 56 of 81 women, including 3 of 7 women on E + HRT). Therefore, CSF E1 and E2 levels were categorized as undetectable, detectable/lower, detectable/higher. For CSF E1, the three groups’ concentrations were < 1.0 pg/mL ( n  = 70), 1.0–1.4 pg/mL ( n  = 20), and 1.5–9.5 pg/mL ( n  = 20). For E2, the three groups’ concentrations were < 0.30 pg/mL ( n  = 65), 0.3–0.4 pg/mL ( n  = 21), and 0.5–1.9 pg/mL ( n  = 24). To determine relationship between CSF and plasma estrogen levels, we used multinomial logistic regression to derive odds ratio (O.R.) for higher CSF E1 categories (due to failed parallel line assumption) or ordinal regression to derive proportional odds (P.O.) for higher CSF E2 categories, adjusting for age, log 10 (BMI), and race/ethnicity. CSF Aβ42/Aβ40 showed binomial distribution, likely reflecting one cohort without and one cohort with underlying ADNC. We thus calculated Z-scores based on those with normal values (> 0.065) [ 59 , 60 ] whose ratios were normally distributed (0.0927 ± 0.0099). In this setting, a Z-score lower than −2.91 was considered positive for ADNC. Relationship between CSF Aβ42/Aβ40 and estrogen levels (E1 or E2) were examined using analysis of covariance (ANCOVA) adjusting for age and race/ethnicity. After QC steps above, 1,075 CSF analyte levels underwent standardization and outlier detection across both sexes. Following our own prior workflow as well as consensus recommendations [ 52 , 61 ], dimension reduction was achieved via Principal Component (PC) Analysis which was conducted in female participants. Briefly, each individual CSF protein was first examined to assess for normal distribution across the sample cohort using Kolmogorov–Smirnov Test, and CSF proteins which did not have normal distribution were log 10 -transformed before principal component analysis (PCA). While a normal distribution is not necessary before standardizing data for PCA, we have found the PCA solutions for biofluid proteins – consistent with other scenarios [ 62 ] – to be more reproducible across cohorts and platforms when normalization is attempted (e.g., log 10 -transformation). After log 10 -transformation, all CSF analytes were standardized, and outliers (≥ 4.0 or ≤ −4.0) were imputed in SPSS using K-Nearest Neighbors ( n  = 5). PCA was conducted using co-variance matrix due to all proteins having standardized arbitrary concentration units [ 63 ], and the optimal number of PCs was selected based on 90% variance as an eigenvalue threshold gave too many PCs and there was no clear “elbow”. Varimax rotation was performed for ease of interpretation and pathway analysis. Among the 33 proteins with two assay versions, only the later version was used in analysis. Volumes of 34 cortical regions on each side were standardized before PCA, and co-variance matrix was used due to all brain regions having standardized arbitrary volume units [ 63 ]. Elbow rule and total variance accounted by PCs were used on conjunction to determine eight PCs as the optimal solution. While left and right brain structures were allowed to independently undergo PCA, all PCs had loading from bilateral brain structures. Varimax rotation was performed for ease of interpretation and pathway analysis. Multivariate multiple regression (MMR) analysis was performed with CSF or structural MRI PCs in women only. For each, CSF E2 level (undetectable, lower, higher), race/ethnicity, menopausal status, HRT, age, MoCA, and CSF Aβ42/Aβ40 were entered as independent variables. The MMR was selected to test the likelihood of null hypothesis that none of the PCs were associated with the variables of choice. Omnibus parameters (Wilks lambda, Roy’s Largest Root) were first examined for the independent variables. Based on the intermediate number of comparisons, Hochberg’s correction of Bonferroni procedure was used to determine threshold of significance from multiple tests [ 64 ]. This reduced the number of proteomic PCs associated with CSF E2 and E1 from 5 and 6 (nominal) to 3 and 4 (Table S2.2). For CSF proteomic PCs, a race/ethnicity x CSF E2 tertile interaction term was entered as this improved the overall models. Top positive and negative loading proteins were entered into STRING-DB using GO Biological Processes to identify pathways associated with changes in PC scores [ 65 ]. A similarity score of 0.7 was used to group GO processes sharing significant gene/protein products, and the top generic process such as “inflammatory response” or “immune response” for each PC was included as reference for signal, protein count, and FDR. The same was then repeated for CSF E1 tertiles, log 10 (plasma E2), and log 10 (plasma E1). For volumetric MRI PCs, total intracranial volume was also entered as an independent variable. The interaction term of age X CSF Aβ42/Aβ40 achieved omnibus significance and was included in the CSF estrogen models, and the interaction term between race/ethnicity and plasma estrogen levels achieved omnibus significance. Hochberg’s correction was also applied for MRI PCs to adjust for multiple comparisons, which reduced the number of MRI PCs associated with CSF E2 and E1 from 4 and 3 (nominal) to 3 and 2. Datasets used and/or analyzed during the current study are available from the corresponding authors on reasonable request according to NIH-approved Resource Sharing Plan and approval by Emory University.

Supplementary Material 1. Supplementary Material 2. Supplementary Material 1. Supplementary Material 2.