Article Figures & Data
Figures
- Figure 1.
Path diagram of genomic SEM model. The summary statistics of two RSNs that have shown to significantly correlate with global connectivity will be used as input together with summary statistics of the global connectivity GWAS. In this way, rg between the unique components (u) of the two RSNs can be estimated while taking global connectivity into account. The example in this diagram shows the global and unique genetic effects on functional connectivity (FC) for the ventral attention network (VAN) and default mode network (DMN), but a similar model was used for other RSN pairs and for measures of structural connectivity (SC). This method is embedded in a flowchart that describes the sample (see Extended Data Figs. 1-1 and 1-2 for sample characteristics and exclusion criteria) and all methods used in this manuscript (Extended Data Fig. 1-3).
- Figure 2.
Multitrait Manhattan plots of SNP-based GWAS for (a) within-RSN functional connectivity strength (RSN-FC) and (b) within-RSN structural connectivity property (RSN-SC). The light gray dashed horizontal line indicates traditional genome-wide significance (p < 5 × 10−8), whereas the red dashed horizontal line indicates genome-wide significance after additional correction for the number of traits tested (p < 3.13 × 10−9). SNPs with p > 0.01 are omitted for visualization purposes. Manhattan plots per RSN are provided as Extended Data Figures 2-1 (FC) and 2–2 (SC), replication efforts of these results (Extended Data Figs. 2-3, 2-4) are plotted in Extended Data Figure 2-5, and characteristics of all loci, lead, and candidate SNPs are available in Extended Data Figures 2-6 and 2-7. Heritability estimates based on these GWAS results are provided in Extended Data Figures 2-8 and 2-9.
- Figure 3.
eQTL and Hi-C gene mapping of structural connectivity (SC) and functional connectivity (FC) network measures. a, Within-visual network-SC SNPs were mapped to METTL10, FAM53B, and METTL10-FAM53B readthrough (RP11-12J10.3) through chromatin interaction mapping (orange). METTL10 was additionally mapped by 46 SNPs because of their eQTL associations in cerebral cortex tissue. b, FUMA gene mapping, based on established eQTL associations (green) in human temporal cortex, link eight within-limbic network-FC SNPs on chromosome 10 to CYP2C8. All FUMA gene-mapping results are displayed in Extended Data Figure 3-1, with fine-mapping results in Extended Data Figure 3-2.
- Figure 4.
Multitrait Manhattan plots of gene-based GWAS for (a) FC and (b) SC within resting-state networks (RSNs). The light gray dashed horizontal line indicates significance after correcting for the number of genes tested per trait (p < 2.65 × 10–6), whereas the red dashed horizontal line indicates significance after an additional correction for the number of traits tested (p < 1.66 × 10–7). See Extended Data Figure 4-1 for the association p-values of all genome-wide significant genes, Extended Data Figure 4-2 for local rg summary statistics between Alzheimer’s disease and default mode network-FC (plotted in Extended Data Fig. 4-3), and Extended Data Figure 4-4 for gene-set analysis statistics.
- Figure 5.
Global rg (±SE) between functional connectivity (FC) and structural connectivity (SC) within the same RSN. Genetic correlations as performed in LDSC do not show estimates significantly different from 0 (Extended Data Fig. 5-1). Additional estimation of local rg did not yield significant overlapping loci between SC and FC within each RSN either. The colors correspond to the RSN colors in Figure 2 and 4.
- Figure 6.
Genome-wide rg across RSN measures of (a) functional connectivity (FC) and (b) structural connectivity (SC). If one of the two RSNs showing significant LDSC rg showed additional significant rg with global FC/SC, we instead report the residual rg (rg between the two RSNs while taking global FC/SC into account in Genomic SEM; see Materials and Methods and Fig. 1). The significant rg that survived correction for multiple testing (p < 1.19 × 10−3) is indicated with an asterisk (*).
Tables
- Table 1
Local genetic correlations across RSNs within the functional (FC) and structural (SC) domains
Chr Start Stop RSN 1 RSN 2 ρ 95% CI p-value 1 2,215,496 2,983,519 FC SMN FC VN 0.77 0.47 1.00 3.39 × 10−5 1 18,427,821 19,238,924 FC DMN FC FPN 0.72 0.45 1.00 9.48 × 10−6 1 211,082,893 212,347,582 FC VAN FC SMN 1.00 0.74 1.00 1.75 × 10−7 2 113,930,669 115,203,835 FC FPN FC SMN 0.88 0.64 1.00 3.42 × 10−7 2 207,726,595 208,674,588 FC FPN FC VN 0.97 0.72 1.00 1.02 × 10−6 5 4,636,543 5,828,694 FC DMN FC DAN 0.73 0.47 1.00 2.72 × 10−5 5 68,006,994 71,468,651 FC VAN FC SMN 0.79 0.53 1.00 2.05 × 10−5 5 75,959,516 77,290,255 FC DMN FC DAN 0.91 0.65 1.00 3.42 × 10−6 6 10,416,551 11,790,671 FC VAN FC SMN 0.83 0.55 1.00 2.49 × 10−5 7 50,894,509 51,951,647 FC LN FC VN 0.88 0.57 1.00 5.12 × 10−6 8 64,215,359 66,018,204 FC DMN FC VN 0.86 0.59 1.00 1.09 × 10−5 9 93,441,051 94,175,374 FC FPN FC SMN 0.90 0.61 1.00 1.73 × 10−5 9 93,441,051 94,175,374 FC FPN FC VN 0.87 0.62 1.00 4.58 × 10−6 10 89,971,629 91,021,321 FC VAN FC VN 0.96 0.67 1.00 1.23 × 10−6 15 39,238,841 40,604,780 SC DAN SC FPN 0.85 0.53 1.00 9.51 × 10−7 17 13,648,447 14,508,610 FC DMN FC LN 0.89 0.69 1.00 3.50 × 10−9 18 2,839,843 3,722,828 FC DMN FC DAN 0.70 0.45 1.00 2.66 × 10−5 19 17,045,964 17,750,518 FC LN FC DAN 0.73 0.47 1.00 2.34 × 10−6 19 17,045,964 17,750,518 FC DMN FC DAN 0.79 0.53 1.00 1.43 × 10−5 Loci with Bonferroni-corrected significant [p < (0.05/774=) 6.46 × 10−5] rg (ρ with lower and upper limit of 95% confidence interval) between RSN-FC or RSN-SC as performed in LAVA. Within these loci, global FC or SC did not show significant univariate h2 or rg with either of the two RSNs. See Extended Data Table 1-1 for all statistics. SMN = somatomotor network, VN = visual network, DMN = default mode network, FPN = frontoparietal network, VAN = ventral attention network, DAN = dorsal attention network, LN = limbic network.
Extended Data Figure 1-1
Mean age (years) and female percentage in discovery and holdout samples for GWAS on RSN-FC/RSN-SC. Download Figure 1-1, XLS file.
Extended Data Figure 1-2
Sample size and exclusion criteria for discovery and holdout sample in GWAS for RSN-FC/RSN-SC. From all subjects in the latest neuroimaging release, we randomly assigned 5000 subjects to a holdout set for validation. Genotype quality control exclusion criteria included UKB-provided relatedness, discordant sex or sex aneuploidy. Outliers include “ununsable” subjects as well as outliers based on methods advised by UKB Neuroimaging Documentation. Download Figure 1-2, XLS file.
Extended Data Figure 1-3
Flowchart of methods involved in the current study. Functional and structural connectivity (FC/SC) within resting-state networks (RSN) as defined by Yeo et al. (2011) were obtained similarly as previously described (Wei et al., 2019). SNP-based and gene-based GWAS and in silico follow-up were performed on a discovery sample and were validated in a replication sample. rsfMRI = resting-state functional magnetic resonance imaging, DWI = diffusion weighted imaging, SNP = single nucleotide polymorphism, FC = functional connectivity, SC = structural connectivity, GWAS = genome-wide association study, LDSC = linkage disequilibrium score regression, LAVA = local analysis of [co]variant annotation, FUMA = functional mapping and annotation, MAGMA = multimarker analysis of genomic annotation, MsigDB = molecular signatures database, PGS = polygenic score. Download Figure 1-3, TIF file.
Extended Data Table 1-1
Bivariate and partial local genetic correlations from LAVA between summary statistics of RSN-FC/RSN-SC. Locus = locus number. chr = chromosome of locus. start = basepair start position of the locus. stop = basepair end position of the locus. n.snps = number of snps within the locus. n.pcs = number of PCs within the locus. phen1 = one of the two phenotypes involved in the bivariate rg. phen2 = one of the two phenotypes involved in the bivariate rg. z = the conditional phenotype in partial rg, rho = local genetic correlation. rho.lower and rho.upper = 95% confidence interval of rho. r2.phen1 z and r2.phen2 z = the proportion of genetic variance explained for phen1/phen2 by z. pcor = partial correlation coefficient. ci.lower and ci.lower = 95% confidence interval for partial correlation coefficient; P = p-value for local rg. Download Table 1-1, XLS file.
Extended Data Figure 2-1
Manhattan plots of SNP-based (left) and gene-based (right) GWAS for RSN-FC. The light grey dashed horizontal line indicates (left) traditional GWS (p < 5 × 10−8) or (right) significance after correcting for the number of genes tested per trait (p < 2.65 × 10−6). The red dashed horizontal line indicates significance after an additional correction for the number of traits tested (left; p < 3.13 × 10−9) or (right; p < 1.66 × 10−7). Download Figure 2-1, EPS file.
Extended Data Figure 2-2
Manhattan plots of (left) SNP-based and (right) gene-based GWAS for RSN-SC. The light grey dashed horizontal line indicates (left) traditional GWS (p < 5 × 10−8) or (right) significance after correcting for the number of genes tested per trait (p < 2.65 × 10−6). The red dashed horizontal line indicates significance after an additional correction for the number of traits tested (left; p < 3.13 × 10−9) or (right; p < 1.66 × 10−7). Download Figure 2-2, EPS file.
Extended Data Figure 2-3
Best p-value threshold and PGS prediction (with PRSice) for FC/SC within RSNs. Phase 1 = GWAS of FC/SC within RSNs in discovery sample. Phase 2 = best p-value threshold determination in half replication sample. Phase 3 = validation of best p-value threshold PRS in other half replication sample. Threshold = best p-value threshold. PRS R2 = variance explained by the PRS. Full R2 = variance explained by the full model (including the covariates). Null R2 = variance explained by the covariates. Coefficient = regression coefficient of the model. SE = SE of regression coefficient of the model. P = p-value of the model fit. N SNPs = number of SNPs included in the model. Download Figure 2-3, XLS file.
Extended Data Figure 2-4
Association statistics for discovery lead SNPs in the holdout set for replication. Phenotype = phenotype of the discovery lead SNP. lead SNP = original UKB variant name. CHR = chromosome. POS = position. REF = reference allele. ALT = alternative allele. A1 = effect allele. OBS_CT = per SNP sample size. BETA = effect size. SE = SE of the effect size. L95 = the lower endpoint of the confidence interval. U95 = the upper endpoint of the confidence interval. TSTAT = t statistic. P = p-value. ALT_FREQS = frequency of alternative allele. RSID_UKB = rs identifier of the SNP. MAF_UKB = MAF as provided by UKB (full file). INFO_UKB = INFO as provided by UKB (full file). Download Figure 2-4, XLS file.
Extended Data Figure 2-5
Probability curves of discovery lead SNPs being significant in a replication sample of increasing size [log10(N) scale]. The current replication sample size of this study is represented by the vertical grey line. FC, functional connectivity; SC, structural connectivity; SMN, somatomotor network; LN, limbic network; FPN, frontoparietal network; VN, visual network. Download Figure 2-5, EPS file.
Extended Data Figure 2-6
Genomic loci and lead variants in the GWAS of RSN-FC/RSN-SC as identified by FUMA. uniqID = ID of the top lead variant within the locus formatted as 'chromosome:base pair position:alleles in alphabetic order'. Chr = chromosome. Pos = base pair position lead SNP. P = p-value. Start = locus BP start position of the locus. End = stop position of the locus. nSNPs = number of unique candidate SNPs in the genomic locus, including non-GWAS-tagged SNPs (which are available in the use selected reference panel). nGWASSNPs = number of the GWAS-tagged candidate SNPs within the genomic locus. This is a subset of “nSNPs.” nIndSigSNPs = number of the independent significant SNPs in the genomic locus. IndSigSNPs = rsID of independent significant SNPs in the genomic locus. nLeadSNPs = the number of lead SNPs in the genomic locus. LeadSNPs = rsID of lead SNPs in the genomic locus. Download Figure 2-6, XLS file.
Extended Data Figure 2-7
FUMA output for all candidate SNPs. uniqID = unique ID of SNPs consists of chr:position:allele1:allele2 where alleles are alphabetically ordered. rsID = rsID of SNPs as provided in the input GWAS, otherwise extracted from the specified reference panel. chr = chromosome. pos = position on hg19. MAF = minor allele frequency computed based on the reference panel. P = p-value from cerebellar volume GWAS [non-GWAS tagged SNPs (extracted from the reference panel) are “NA”]. Beta = beta from cerebellar volume GWAS [non-GWAS tagged SNPs (extracted from the reference panel) are “NA”]. SE = SE of beta from cerebellar volume GWAS [non-GWAS tagged SNPs (extracted from the reference panel) are ““NA””). r2 = the maximum r2 of the SNP with one of the independent significant SNP. IndSigSNP = rsID of an independent significant SNP which has the maximum r2 of the SNP. Nearest gene = the nearest gene of the SNP based on ANNOVAR annotations. Dist = distance to the nearest gene. SNPs which are locating in the gene body or 1 kb upstream or downstream of TSS or TES have 0. Func = functional consequence of the SNP on the gene obtained from ANNOVAR. CADD = combined annotation dependent depletion score. RDB = RegulomeDB score. posMapFilt = whether the SNP was used for positional mapping or not, 1 is used, otherwise 0. eqtlMapFilt = whether the SNP was used for eQTL mapping or not, 1 is used, otherwise 0. ciMapFilt = whether the SNP was used for chromatin interaction mapping or not, 1 is used, otherwise 0. Download Figure 2-7, XLS file.
Extended Data Figure 2-8
SNP-based heritability of RSN-FC/RSN-SC. Heritability estimates were calculated using LD score regression (LDSC). h2 = heritability estimate. h2 SE = SE of heritability estimate. Intercept = the LDSC intercept (should be close to 1). Lambda = median(chi2)/0.4549. Mean chi2 = mean chi-square statistic. Ratio = (intercept-1)/(mean(chi2)-1), measures the proportion of the inflation in the mean chi^2 that the intercept ascribes to causes other than polygenic heritability. Ratio SE = SE of ratio. Download Figure 2-8, XLS file.
Extended Data Figure 2-9
Partitioned heritability of RSN-FC/RSN-SC in genomic categories. Partitioned heritability estimates were calculated using stratified LDSC regression. Prop. SNPs = proportion of total SNPs that belong to a given category (categories are not mutually exclusive; i.e., a SNP may be assigned to more than one category). Prop. h2 = proportion of heritability ascribed to SNPs in the category. Enrichment = proportion of heritability divided by proportion SNPs. Enrichment SE = SE of the enrichment estimate. Enrichment p-value = p-value of the enrichment estimate. Download Figure 2-9, XLS file.
Extended Data Figure 3-1
FUMA gene mapping based on positional location, eQTL association and/or chromatin interactions in cerebral cortex tissue. “Ensg = ENSG ID. Symbol = gene symbol. Chr = chromosome. Start = starting basepair position of the gene. End = ending basepair position of the gene. Strand = strand of the gene. Type = gene biotype from Ensembl. entrezID = entrez ID (if available). HUGO = HUGO (HGNC) gene symbol. pLI = pLI score from ExAC database. The probability of being loss-of-function intolerant; ncRVIS = noncoding residual variation intolerance score. The higher the score is, the more intolerant to noncoding variation the gene is. posMapSNPs (posMap) = number of SNPs mapped to gene based on positional mapping. posMapMaxCADD (posMap) = maximum CADD score of mapped SNPs by positional mapping. eqtlMapSNPs (eqtlMap) = number of SNPs mapped to the gene based on eQTL mapping. eqtlMapminP (eqtlMap) = minimum eQTL p-value of mapped SNPs. eqtlMapminQ (eqtlMap) = minimum eQTL FDR of mapped SNPs. eqtlMapts (eqtlMap) = tissue types of mapped eQTL SNPs. eqtlDirection (eqtlMap) = consequential direction of mapped eQTL SNPs after aligning risk increasing alleles in GWAS and tested alleles in eQTL data source. ciMap (ciMap) = “Yes” if the gene is mapped by chromatin interaction mapping, “No” otherwise. ciMapts (ciMap) = tissue/cell types of mapped chromatin interactions. minGwasP = minimum p-value of mapped SNPs. IndSigSNPs = rsID of the independent significant SNPs that are in LD with the SNPs that are mapped to the gene. GenomicLocus = index of genomic loci where mapped SNPs are from. Download Figure 3-1, XLS file.
Extended Data Figure 3-2
95% credible sets as identified by FINEMAP. PIP = posterior inclusion probability (SNPs with PIP > 0.1 in bold), is the posterior probability that this SNP is causal. log10bf = log10 Bayes factors. The Bayes factor quantifies the evidence that the SNP is causal with log10 Bayes factors greater than 2 reporting considerable evidence. mean = marginalized shrinkage estimates of the posterior effect size mean. sd = marginalized shrinkage estimates of the posterior effect size SD. mean_incl = conditional estimates of the posterior effect size mean. sd_incl = conditional estimates of the posterior effect size SD. Download Figure 3-2, XLS file.
Extended Data Figure 4-1
Association p-values of all genome-wide significant (GWS) genes for RSN-FC/RSN-SC. Displayed are genes significant after Bonferroni correction for the number of genes and the number of traits tested. Entrez ID = Entrez ID of gene. CHR = chromosome. Start = start position of the gene in base pairs. Stop = end position of the gene in base pairs. N SNPs = number of SNPs in gene. N Param = number of relevant parameters used in the model. N = per gene sample size; Z = Z-value for the gene, based on its p-value; P = SNPwise mean p-value (model that uses sum of squared SNP Z-statistics as test statistic). Download Figure 4-1, XLS file.
Extended Data Figure 4-2
Bivariate local genetic correlations from LAVA between summary statistics of Alzheimer’s disease and default mode network-FC. Locus = locus number. chr = chromosome of locus. start = basepair start position of the locus. stop = basepair end position of the locus. n.snps = number of snps within the locus. n.pcs = number of PCs within the locus. phen1 = one of the two phenotypes involved in the bivariate rg. phen2 = one of the two phenotypes involved in the bivariate rg. rho = local genetic correlation. rho.lower and rho.upper = 95% confidence interval of rho. r2 = proportion of genetic signal for phen1 that is explained by phen2. r2 lower and upper = 95% confidence interval of r2. ci.lower and ci.lower = 95% confidence interval for partial correlation coefficient. P = p-value for local rg. Download Figure 4-2, XLS file.
Extended Data Figure 4-3
Local rg between default mode network-FC and Alzheimer’s disease as performed in LAVA. Only loci that passed the univariate h2 threshold (p < 1 × 10−4) were tested for bivariate rg, resulting in the Bonferroni-corrected significance threshold represented by the red line. Significant loci are visualized with their rg estimate. Within these loci, global functional connectivity strength (FC) did not show significant univariate h2 and could therefore not bias these results. Download Figure 4-3, EPS file.
Extended Data Figure 4-4
Results of MAGMA’s gene-set analysis for FC/SC within RSNs in all curated and gene-ontology gene-sets from MSigDB (sets C2 and C5). Gene set = set tested to be associated with cerebellar volume gene-based GWAS sumstats. N genes = number of genes used in the analysis. Beta = effect size. Beta SD = standardized effect size. SE = standard error. P = p-value. Download Figure 4-4, XLS file.
Extended Data Figure 5-1
Global genetic correlations between summary statistics of RSN-FC/RSN-SC in LDSC and genomic SEM. rg = global genetic correlation. SE = SE of rg; Z = z-score of rg. P = p-value for rg. h2 obs = observed scale h2 for trait 2. h2 obs SE = SE of observed scale h2 for trait 2. h2 int = single-trait LD Score regression intercept for trait 2. h2 int SE = SE of single-trait LD score regression intercept for trait 2. gcov int = cross-trait LD score regression intercept. gcov int SE = SE of cross-trait LD score regression intercept. Download Figure 5-1, XLS file.
In this issue
