ANALYSIS

Cis eQTL* with Signif. eQTL eQTL (Genes) not Present in GTEx**

not found in Cohort

(typically 100–150). Recently, eforts to understand gene expression changes in neuropsychiatric23–26 and neurodegenerative diseases27–34 have generated brain RNA-seq from disease and normal tissue, as well as genome-wide genotypes. These analyses have found little evidence for widespread disease-specific eQTL, as well ashigh cross-cohort overlap24,35, meaning that most eQTL detected are disease-condition independent. Vis implies that meta-analysis across disease-based cohorts will capture eQTL which are unconfounded by disease state despite diferences in disease ascertainment of the samples,and leverages thousands of available samples to produce a well-powered brain eQTL resource for use in downstream research.

NSNPs 4905 4099 5437 4906 4886 4432 4956 1055 3600 3725 3294 5253 4767 5364 5651 3844 4527

GWAS 2.72E-09 1.28E-09 1.66E-06 2.58E-08 2.58E-08 7.44E-07 1.02E-08 1.16E-19 3.37E-08 2.03E-08 1.92E-06 3.68E-12 2.30E-12 1.15E-08 2.84E-06 1.57E-06 4.03E-08 eQTL 2.28E-21 4.42E-10 8.85E-45 1.10E-13 2.21E-39 7.85E-07 1.91E-09 5.52E-07 1.63E-07 5.63E-31 6.90E-62 1.50E-09 1.29E-20 1.16E-06 5.38E-10

Ve replication of fetal eQTL, while signifcant,siosmewhat lower than the replication of adult eQTeLpresented in GTEx. Vis may be due to multiple factors. Ve fetal eQTL analysis was generated frominbrhaomog-e nate, rather than dissected brain regions, thoughhte lower replication likely also refects broad trsacnriptional diferences between developing and mature brain58. Vese transcriptional diferences may also explainwhy we fnd substantially more eQTL than a recently published, similarly sized eQTL analysis which uses samples from across developmental and adult timepoin2t5s, and why this meta-analysis shows higher replicatoin of GTEx eQTL.

Previous studies report a lack of widespread diseea-s pecific eQTL observed in schizophrenia (CMC24) and Alzheimer9s (ROSMAP)35. In accordance, we fnd a strong overlap among eQTacLross disparate disease samples, particularly those with neuropsychiatric nad neurodegenerative disorders, as well as normanldi ividuals from these and other cohorts such as GT24E,3x5. Vis suggests that disease-specifc eQTL, if they xeist, are likely few in number and/or small in efect size, rleative to condition-independent eQTL in general. tIhfey do exist, disease-specifc eQTL discovery may be succes ful in more targeted analyses or with larger samepslizes or meta-analyses, but was not explored for the puorspe of this general resource. Vus, the heterogenseosaumples derived from diferent disease-based cohortsncabe meta-analyzed to create a general-purpose branieQTL resource representing adult gene regulation, despeitcomprising samples with diferent disease backgronuds, along with normal controls. Verefore, these eQTlLl bwei useful both within and outside these specifdcisease contexts. For example, since these eQTL are not deiasse specifc they may be used to understand healthygene expression regulation in the brain, as well as tonfi er colocalization of eQTL signatures with diseraiske for any disease whose tissue etiology is from the brain, since these signatures are refective of normal brainartieogunl. It should be stated that while many eQTL are not disesae specifc, i.e. they are identifed under various central nervous system (CNS) disease diagnoses and in control brains, they may still contribute to common CNeaSsedsisas previously demonstrate2d4,32–34,45,46. While we have demonstrated a proof-of-concept colocalization analysis with a previously published schizophrenia GWAS, these eQTL are a broadly useful resource for studying nepusryochiatric and neurodegenerative disorders, as well as for understanding the landscape of gene raetgiounl in brain. Methods RNA-seq Re-alignment. For the CMC studies (MSSM-Penn-Pitt, HBCC), RNA-seq reads were aligned to GRCh37 with STAR v2.4.0g159 from the original FASTQ fles. Uniquely mapping rdesa overlapping genes were counted with featureCounts v1.56.02using annotations from ENSEMBL v75.

For the AMP-AD studies (ROSMAP, Mayo RNAseq), Picadr v2.2.4 (https://broadinstitute.github.io/picard)/ was used to generate FASTQ fles from the availablBeAM fles, using the Picard SamToFastq function. Paircd SortSam was frst applied to ensure that R1 and R2erads were correctly ordered in the intermediate SAMfle before converting to FASTQ. Ve converted FASTQs ewaerligned to the GENCODE24 (GRCh38) reference genome using STAR v2.5.1b, with twopassMode set Basasic. Gene counts were computed for each sample by STAR by setting quantMode as GeneCounts.

RNA-seq normalization. To account for diferences between samples, studieesx,perimental batch efects and unwanted RNA-seq-specifc technical variations,we performed library normalization and covariatejuasdtments for each study separately using fxed/mixed efects modeling. A mixed efect model was required to jointly normalize both tissues from the Mayo cohort. All other cohorts contained only one tissueed,seofeactfmxodel was used. Ve workfow consisted of the following steps: 1. Gene fltering: Out of ~56 K aligned and quantifed genes, only genes showing at least modest expression were used in this analysis. Genes that were expressed more than 1 CPM (read Counts Per Million total reads) in at least 50% of samples in each tissue and diagnosis category were retained for analysis. Addi-tion ally, genes with available gene length and percentage GC content from BioMart December 2016 archive were subselected from the above list. Vis resulted in approximately 14 K to 16 K genes in each study. 2. Calculation of normalized expression values:Sequencing reads were then normalized in two steps. First, conditional quantile normalization (CQ6N1 )was applied to account for variations in gene length and GC content. In the second step, the confdence of sampling abundance was estimated using a weighted linear model using the voom-limma package in biocondu6c2t,6o3.rVe normalized observed read counts, along with the corresponding weights, were used in the following steps. 3. Outlier detection: Based on normalized log2(CPM) of expression values, outlier samples were detected using principal component analysis (PCA64),65 and hierarchical clustering. Samples identifed as outliers using both the above methods were removed from further analysis. 4. Covariate imputation: Before identifying associated covariates, important missing covariates were i-mput ed. Principally, post-mortem interval (PMI), or the latency between death and tissue collection, whisch i frequently an important covariate for the analysis of gene expression from post-mortemsburea,inwatiss imputed for a portion of samples in Mayo RNAseq data for which true values were unavailable. Genomic predictors of PMI were estimated using ROSMAP and MSSM (an additional RNA-seq study available through AMP-AD) samples and were used to impute missing values as necessary. 5. Covariate identifcation: Normalized log2(CPM) counts were then explored to determine which known covariates (both biological and technical) should be adjusted. Except for the HBCC study, we usetedpa- s wise (weighted) fxed/mixed efect regression modeling approach to select the relevant covariatesinhgava signifcant association with gene expression. Here, covariates were sequentially added to the model iyf the were signifcantly associated with any of the top principal components, explaining more than 1% oi- f var ance of expression residuals. For HBCC, we used a model selection based on Bayesian informationacriteri (BIC) to identify the covariates that improve the model in more than 50% of genes. 6. SVA adjustments: Ager identifying the relevant known confounders, hidden-confounders were identifed using the Surrogate Variable Analysis (SV4A1.)We used a similar approach as previously defne2d4 to fnd the number of surrogate variables (SVs), which is more conservative than the default method provided by the SVA package in R66. Ve basic idea of this approach is that for an eigenvector decomposition of-permut ed residuals each eigenvalue should explain an equal amount of the variation. By the nature of eig-enval ues, however, there will always be at least one that exceeds the expected value. Vus, from a series of 100 permutations of residuals (white noise) we identifed the number of covariates as shown in Supplemryenta Table 1. We applied the <irw= (iterative re-weighting) version of SVA to the normalized gene expression matrix, along with the covariate model described above to obtain residual gene expression for eQTL analysis. 7. Covariate adjustments: We performed a variant of fxed/mixed efect linear regression, choosing mixed-efect models when multiple tissues or samples, were available per individual, as shown here: gen expression ~ Diagnosi+s Sex + covariates+ (1|Donor), where each gene was linearly regressed ind-e pendently. Here Donor (individual) was modeled as a random efect when multiple tissues from the same individual were present. Observation weights (if any) were calculated using the voom-lim6m2,6a3 pipeline, which has a net efect of up-weighting observations with inferred higher precision in the linear model ftting process to adjust for the mean-variance relationship in RNA-seq data. Ve diagnosis conmenptowas then added back to the residuals to generate covariate-adjusted expression for eQTL analysis.

Vis workfow was applied separately for each study. For the AMP-AD studies, gene locations were liged over to GRCh37 for comparison with the genotype imputation panel (described below). For HBCC, samipthleasgwe <18 were excluded prior to analysis. Supplementary Table 1 shows the covariates and surrogate varleiasbidentifed in each study.

AD diagnosis harmonization. Prior to RNA-seq normalization, we harmonized thOe ALD defnition across AMP-AD studies. AD controls were defned asaptients with a low burden of plaques and tangless, awell as lack of evidence of cognitive impairment. For the ROSMAP study, we defned AD cases to be individuals with a Braak67 greater than or equal to 4, CERAD sco6r8eless than or equal to 2, and a cognitive diagnosis of probable AD with no other causes (cogd=x4), and controls to be individuals with Braak lestshan or equal to 3, CERAD score greater than or equal to 3, and cognitivegdniaosis of 8no cognitive impairment9 (c=og1)d.xFor the Mayo Clinic study, we defned disease status based on neuropathology, where individuals with Braak sceogreater than or equal to 4 were defned to be AD cases, and indiidvuals with Braak less than or equal to 3 were defend to be controls. Individuals not meeting A<D case= or c<ontrol= criteria were retained for anyaslis, and were categorized as o<ther= for the purposes of RNA-seq adjustment.

Genotype QC and imputation. Genotype QC was performed using PLINK v16.99. Markers with zero alternate alleles, genotyping call rate≤ 0.98, Hardy-Weinberg p-value < 5e−5 were removed, as well as individuals with genotyping call rate< 0.90. Samples were then imputed to HRC (Version r1.1 201462)as follows: marker positions were liged-over to GRCh37, if necessaMrya.rkers were then aligned to the HRC loci using HR-C 1000G-check-bim-v4.2 (http://www.well.ox.ac.uk/~wrayner/tool)s,/which checks the strand, alleles, position, reference/alternate allele assignments and frequenices of the markers, removing A/T & G/C single nucoletide polymorphisms (SNPs) with minor allele frequency (AMF) > 0.4, SNPs with difering alleles, SNPs with > 0.2 allele frequency diference between the genotyped samples and the HRC samples, and SNPs not in the reference panel. Imputation was performed via the Michigan Ipmutation Server70 using the Eagle v2.371 phasing algorithm. Imputation was done separately by cohort and by chip within cohort, and markers wi2th≥R0.7 and minor allele frequency (MAF)≥ 0.01 (within cohort) were retained for analysis.

Genetic ancestry inference. GEMTOOLs72 was used to infer ancestry and compute ancestry components separately by cohort. Ve GEMTOOLs algorithm usessaignifcance test to estimate the number of eigenvteocrs (ancestry components) necessary to represent the variability in the da7t3a. For each cohort, we used the top c o-m ponents as estimated by the GEMTOOLs algorithm, whchi resulted in some variation in the number of coom-p nents selected. For MSSM-Penn-Pitt and HBCC, whichare multi-ethnic cohorts, only Caucasian samples wre retained for eQTL analysis. eQTL analysis. eQTL were generated separately in each cohort and tissue using MatrixEQT7L4adjusting for harmonized Diagnosis and inferred Ancestry compontseunsing c<is= gene-marker comparisons: Expression~ Genotype+ Diagnosis+ PC1 + … + PCn,, where PCk is the kth ancestry component, using Expression variables which were previously covariate adjusted as descriebd above. Here we defne c<is= as ± 1 MB around the gene, and GRCh37 gene locations were used for consistenwciyth the marker imputation panel. Meta-analysis was performed via fxed-efect mod7e5lusing an adaptation of the metareg function in the gap package in R. In order to assess potential infation of Type 1 error, wrefpoermed 5 permutations of the gene expression vsa,lureelative to genotype and ancestry components, within diagnisofsor each cohort, and repeated the regression laynsaes as described above. For each of the 5 iterations orfmpuetation, a meta-analysis was then performed acrostshe 4 cohorts. We found that Type 1 error was well conlletrdo(Fig. 1a). Given that multiple tissues were present, we also evaluated a random-efect model, but found substantially defated p-values (less signifcant) in the permautitons, relative to the expected distribution, suggesting that this model is over-conserva.tive Comparison with GTEx and fetal eQTL. Full summary statistics for the GTEx v271 eQTL for all available brain regions were obtained from the GTEx Porthatltp(s://gtexportal.or g),/and fetal eQTL were obtained from Figshare76. For each replication comparison (e.g. meta-anailsyvss GTEx or meta-analysis vs. fetal eQTL), only markers and genes present in both the external eQTL and our analysis were retained for comparni.sAos this was done separately for GTEx and for the fetal eQTL resource, the list of genes and SNPs varies slightly for ea-ch com parison. Ve replication rate was estimated as theπ1 statistic using the qvalue package77 in R as follows: we extracted the meta-analysis p-values for all SNP-gnee pairs, which were signifcant in GTEx at FDR ≤ 0.05. We then applied the 8qvalue9 command to the meta-anasilys p-values to generateπ1 = 1 − π0, which corresponds to estimated proportion of non-null p-va7l7u.eVse 8smoother9 option was used to estimπ0ataes a function of the tuning parameter λ as it approaches 1. Ve variance around this estimtae is relatively small (see Supplementary Figures 1 and 2 for example) and does not materially afect the observations in this manuscript.

Conversely, we estimated the replication rate ofgsniifcant meta-analysis eQTL SNP-gene pairs in GTEx. Analogous methods were used to estimate all otheprlrication rates. For the purposes of reportingttohetal number of eQTL not present in GTEx, and genes withuoteQTL in GTEx (Table 1), we have included genes and SNP-gene pairs not present in GTEx in the count, hwoever this accounts for a relatively small propoornti of the diference (472,995 eQTL and 1481 genes).

GTEx eQTL generation. In order to verify that the observed power increasned replication imbalances were not due to methodological diferences betwehenistmanuscript and those performed by GTEx, we obi ntaed access to the GTEx v7 data, and generated eQTL focrortex, anterior cingulate cortex, and frontal ceoxrutsing our approach. We used gene expression and imputedengotypes as provided, as well as the provided coviatres, which included 3 ancestry covariates, 14-15 surrogtae variable covariates, sex and platform. We thenepreated the comparisons with the meta-analysis described inthe previous section, using a MAF cutof of 0.03h,iwch best appeared to control Type 1 error, as observed by permutation between genotype and geneesesxiporn, while maximizing the number of signifcant eQTL in the true data. Results did not change materially. Pathway analysis of cerebellar eQTL genes. In order to identify whether genes showing cerebellar-specifc eQTL patterns showed any biolocgail coherence, we performed a pathway analysis asof-l lows. For genes with at least 5 signifcant cerebealrl eQTL, we computed the Spearman correlation ofeecft-size between cerebellum eQTL and cortical eQTL for theolci that were signifcant in cerebellum. We then seelcted genes for which the efect-sizes did not show positive correlation (Spearmρa<n90s.1) between the two tissues as showing diferent eQTL association patterns across the gene and performed a pathway analysihs wGiOt biolo-g ical processes Fisher9s exact test. Note that dueot the (power-mediated) greater detection of eQTLcoinrtex, we did not perform the reverse comparison. Ve results were relatively robust to the choice of umminni mumber of signifcant eQTL, correlation cutof, and choice of correlation statistic (Spearman vsoPena)r.s Coloc analysis. We applied Approximate Bayes Factor colocalizati o(cnoloc.abf7 )from the coloc R package to the summary statistics from the PGC2 Schizophrenia GWA36Sdownloaded from the PGC websithet(tp://pgc. unc.edu), and the summary statistics from the eQTL meta-analysis. Each gene present in the meta-analysis was compared to the GWAS in turn, and suggestive andgsinifcant GWAS peaks withp-value < 5e-6 were considered for analysis.

Data availability Data for the ROSMAP37 and Mayo cohor4t0sare available through the AMP-AD Knowledge Por3t1a.lData for the MSSM-Penn-Pitt and HBCC cohorts are available through the CommonMind Knowledge P2o3.rtal eQTL results for the ROSMAP78, Mayo TCX79, Mayo CER80 and cortical meta-analysis81 are available through the AMP-AD Knowledge Portal. Vese results includeNSP (location, rsid, alleles, and allele frequencya)nd gene (location, gene symbol, strand and biotype)foinrmation, as well as estimated efect size (beta)s,tatistic (z), p-value, FDR, and expression-increasing allele.

Code availability An R package with all code for the gene expressionnormalization is available athttps://github.com/SageBionetworks/ampad-difex.pAll other analyses were generated using packagespublicly available from their respective authors.

Acknowledgements For the ROSMAP and Mayo RNAseq studies, the resultspublished here are in whole or in part based on dtaa obtained from the AMP-AD Knowledge Portal (doi:3103.7/syn2580853). ROSMAP study data were provided by the Rush Alzheimer9s Disease Center, Rush University Medical Center, Chicago. Data collection wauspsported through funding by NIA grants P30AG10161, R01AG15891, R01AG17917, R01AG30146, R01AG36836, U01AG32984, U01AG46152, the Illinois Department oPfublic Health, and the Translational Genomics Reserach Institute. Mayo RNA-seq study data were provided bythe following sources: Ve Mayo Clinic AlzheimerseDasie Genetic Studies, led by Dr. Nilufer Ertekin-Taner and Dr. Steven G. Younkin, Mayo Clinic, Jacksonville, FL using samples from the Mayo Clinic Study of Aging, theyMoaClinic Alzheimer9s Disease Research Center, andthe Mayo Clinic Brain Bank. Data collection was suppoedrtthrough funding by NIA grants P50 AG016574, R01 AG032990, U01 AG046139, R01 AG018023, U01 AG006576, U01 AG006786, R01 AG025711, R01 AG017216, R01 AG003949, NINDS grant R01 NS080820, CurePSP Foundation, and support from Mayo Foundation. Study data includes samples collected through the Sun Health Research Institute Brain and Body Dnoation Program of Sun City, Arizona. Ve Brain and Body Donation Program is supported by the National Intestoituf Neurological Disorders and Stroke (U24 NS072026 National Brainnad Tissue Resource for Parkinsons Disease and Reeladt Disorders), the National Institute on Aging (P30 A1G9610 Arizona Alzheimers Disease Core Center), theArizona Department of Health Services (contract 211002, Azroina Alzheimers Research Center), the Arizona Biomdeical Research Commission (contracts 4001, 0011, 05-901nad 1001 to the Arizona Parkinson9s Disease Consoumrt)i and the Michael J. Fox Foundation for ParkinsonseRaerch. Vis study was in part supported by NIH RF1 AG051504 and R01 AG061796 (NET). For CommonMind, dtaa were generated as part of the CommonMind Consortium supported by funding from Takeda Pharemuaticcals Company Limited, F. Hofmann-La Roche Ltd and NIH grants R01MH085542, R01MH093725, P50MH066392, P50MH080405, R01MH097276, RO1-MH-075916, P50M096891, P50MH084053S1, R37MH057881, AG02219, AG05138, MH06692, R01MH110921, R01MH109677, R01MH109897, U01MH103392, and contract HHSN271201300031C through IRP NIMH. Brain tissue for the study was obtainedrof m the following brain bank collections: the MoSiunnati NIH Brain and Tissue Repository, the University oPfennsylvania Alzheimer9s Disease Core Center, the University of Pittsburgh NeuroBioBank and Brain and Tissue Roepsitories, and the NIMH Human Brain Collection C o.re CMC Leadership: Panos Roussos, Joseph Buxbaum, Anedwr Chess, Schahram Akbarian, Vahram Haroutunian (Icahn School of Medicine at Mount Sinai), BernievDlin, David Lewis (University of Pittsburgh), Raqeul Gur, Chang-Gyu Hahn (University of Pennsylvania), EnricoDomenici (University of Trento), Mette A. PeterSso,lveig Sieberts (Sage Bionetworks), Vomas Lehner, Geetha Senthil,eSftano Marenco, Barbara K. Lipska (NIMH). SKS, TP, KKD, JE, AKG, LO, BAL, and LMM were additionalyl supported by NIA grants U24 AG61340, U01 AG46170, U01 AG 46161, R01 AG46171, R01 AG 46174. All data used in this manuscript have been previously released through their respective consortia and have beenvrieewed by IRBs at their institution of origin. Inrmfoed consent has been obtained from all individuals.

Competing interests Ve authors declare no competing interests.

Additional information Supplementary information is available for this paper ahtttps://doi.org/10.1038/s41597-020-00642-.8 Correspondence and requests for materials should be addressed to S.K.S. or L.M.M.

Reprints and permissions information is available at www.nature.com/reprints.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional aoliations.

Open Access This article is licensed under a Creative CommonsAttribution 4.0 International License, which permits use, sharing, adaptation, dsitribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to- the Cre ative Commons license, and indicate if changes werme ade. Ve images or other third party material inhtis article are included in the article9s Creative Commons license, unless indicated otherwise in a crediltine to the material. If material is not included in the article9s Creative Commons license and your intended use is no-t per mitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, vhitstipt://creativecommons.org/licenses/by/4.0/ © Ve Author(s) 2020 The CommonMind Consortium (CMC) Schahram Akbarian11, Jaroslav Bendl12, Michael S. Breen13, Kristen Brennand12, Leanne Brown11, Andrew Browne12,14, Joseph D. Buxbaum12,13,14, Alexander Charney12, Andrew Chess15,16,17, Lizette Couto11, Greg Crawford18,19, Olivia Devillers11, Bernie Devlin20, Amanda Dobbyn12,21, Enrico Domenici 22, Michele Filosi 22, Elie Flatow 11, Nancy Francoeur 21, John Fullard12,21, Sergio Espeso Gil 11, Kiran Girdhar 12, Attila Gulyás-Kovács 16, Raquel Gur23, ChangGyu Hahn24, Vahram Haroutunian12,17, Mads Engel Hauberg12,25, Laura Huckins12, Rivky Jacobov11,Yan Jiang11,JessicaS.Johnson 12,BibiKassim 11,YungilKim 12,LambertusKlei 20, Robin Kramer26, Mario Lauria27, Thomas Lehner28, David A. Lewis20, Barbara K. Lipska26, Kelsey Montgomery1, Royce Park11, Chaggai Rosenbluh16, Panagiotis Roussos11,14,15,29, Douglas M. Ruderfer30, Geetha Senthil28, Hardik R. Shah 14, Laura Sloofman 12, Lingyun Song18, Eli Stahl12, Patrick Sullivan31, Roberto Visintainer22, Jiebiao Wang32,Ying-Chih Wang15, Jennifer Wiseman11, Eva Xia 21, Wen Zhang12 & Elizabeth Zharovsky 11 11Division of Psychiatric Epigenomics, Department of Psychiatry, Icahn School of Medicine at Mount Sinai, New York, New York, USA1.2Division of Psychiatric Genomics, Department of Psychiatry, Icahn School of Medicine at Mount Sinai, New York, New York, USA13.Seaver Autism for Research and Treatment, Department of Psychiatry, Icahn School of Medicine at Mount Sinai, New York, New York, U1S4FAr.iedman Brain Institute, Icahn School of Medicine at Mount Sinai, NewYork, NewYork, US A15.Institute for Genomics and Multiscale Biology, Department of Genetics and Genomic Sciences, Icahn School of Medicine at Mount Sinai, NewYork, NewYork, U1S6DAe. partment of Cell, Developmental and Regenerative Biology, Icahn School of Medicine at Mount Sinai, New York, New York, USA. 17Department of Neuroscience, Icahn School of Medicine at Mount Sinai, New YNoerwkY,ork, USA.18Center for Genomic & Computational Biology, Duke University, Durham, North Carolina, 1U9DSAiv.ision of Medical Genetics, Department of Pediatrics, Duke University, Durham, North Carolina, US2A0D. epartment of Psychiatry, University of Pittsburgh School of Medicine, Pittsburgh, Pennsylvania, US21AD.epartment of Genetics and Genomic Sciences, Icahn School of Medicine at Mount Sinai, NewYork, NewYork, USA22.Laboratory of Neurogenomic Biomarkers, Centre for Integrative Biology (CIBIO), University of Trento, Trento,23INtaelyu.ropsychiatry Section, Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Pheilpahdia, Pennsylvania, USA. 24Neuropsychiatric Signaling Program, Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania, USA. 25Department of Biomedicine, Aarhus University, Aarhus, Denmark. 26Human Brain Collection Core, National Institutes of Health, NIMH, Bethesda, Maryland, US2A7D.epartment of Mathematics, University of Trento, Trento, Ita2l8yN. ational Institute of Mental Health, Bethesda, Maryland, USA. 29Psychiatry, JJ Peters VA Medical Center, Bronx, New York, USA3.0Department of Medicine, Psychiatry and Biomedical Informatics, Vanderbilt Genetics Institute, Vanderbilt Univertsyi Medical Center, Nashville, Tennessee,