Estimation of genetically regulated gene expression across brain regions at biobank scale

We used Joint-Tissue Imputation [43], a recently developed state-of-the-art method from computational genomics, to estimate the genetically regulated expression of 18,647 genes across 10 cortical and subcortical brain regions for 45,549 people from the UK Biobank (64 ± 7.7 years old, 52% female) and 657 people in the HCP (29 ± 3.6 years old, 52% female).

Joint-Tissue Imputation models estimate genetically regulated gene expression (gr-expression) as a weighted linear combination of SNPs that are close to the gene of interest along the linear genome. These models learn weights for each tissue–gene pair by training on genetic sequences and directly measured gene expression from postmortem samples (Fig 1A). Joint-Tissue Imputation leverages shared patterns of genetic regulation across brain regions to improve the estimation of gr-expression in individual regions. In this way, this method extends and generalizes PrediXcan, a pioneering estimation method that models gr-expression by training models only on expression data from the brain region of interest [55].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Estimation of genetically regulated gene expression from genetic data.

(A) Pipeline for estimation of gr-expression with Joint-Tissue Imputation. Left: Joint-Tissue Imputation models are trained on genetic sequences and directly assayed gene expression from postmortem brain samples in the GTEx and PsychEncode projects. Center: The models are trained to estimate gr-expression as a weighted sum of SNPs that are close to the gene of interest along the linear genome. The estimation includes elastic-net regularization because the number of these SNPs typically exceeds the number of samples in the training data. Right: The trained models were used to estimate gr-expression from genetic sequences of neuroimaging-genomic samples in the UK Biobank and the HCP. (B) An illustration of the 10 cortical and subcortical regions with available models of gr-expression. Numbers in parentheses refer to all models that passed baseline performance thresholds for the prediction of observed gene expression on held-out data (r² > 0.01 and p_FDR < 0.05). (C, D) Predictive performance of gr-expression models on held-out data from the GTEx data set. (C) Histograms of r², the variance of directly assayed gene expression explained by estimated gr-expression. (D) Histograms of p-values (−log₁₀ p_FDR) on these r² values. Regions are colored as in panel B. FDR, false discovery rate; GTEx, Genotype-Tissue Expression Project; HCP, Human Connectome Project; SNP, single-nucleotide polymorphism.

https://doi.org/10.1371/journal.pbio.3002782.g001

In our study, we used Joint-Tissue Imputation models that were previously trained on whole-genome sequences and gene-expression data from 838 brain samples in the Genotype-Tissue Expression Project (GTEx) [56]. The samples comprise 10 cortical and subcortical regions (Fig 1B). To test the replicability of our analyses, we additionally used the same models trained on sequencing and expression data from 415 independent samples of the dorsolateral prefrontal cortex (DLPFC) in the PsychENCODE Project [57]. Collectively, we considered 94,345 Joint-Tissue Imputation models, or all performant brain-regional models currently available in the literature.

Joint-Tissue Imputation models have been extensively validated in previous work [44–48]. This validation included quantifying the relationship of gr-expression to directly assayed expression. In this study, we adopted all models of gr-expression that passed baseline performance thresholds for the prediction of observed gene expression on held-out data (r² > 0.01 and p_FDR < 0.05). In practice, the predictive performance of gr-expression models spanned a wide range (Fig 1C and 1D). Low predictive performance does not necessarily mean that the models are inaccurate because the genetic regulation of gene expression—the upper bound on predictive performance—varies considerably for individual genes. Moreover, relatively low associations between gr-expression and assayed expression are more than offset by gains in statistical power of transcriptome-wide association analyses, as we describe below.

Genetically regulated gene expression recapitulates the organization of directly assayed gene expression

We began by testing the extent to which gr-expression recapitulated existing knowledge of genetic-ancestry relationships, brain-regional identities, as well as inter-regional correlations of directly assayed gene expression.

First, we tested if gr-expression patterns reflected known genetic-ancestry relationships from the ethnically diverse sample of the UK-Biobank cohort (Methods, S1 Table). Genetic ancestry denotes genetic commonalities within groups of people but does not necessarily reflect genealogical ancestry (family lines) or self-reported ethnicity. We followed standard practice to estimate genetic ancestry using principal component analysis of gene data. We specifically used principal component analysis to generate low-dimensional embeddings of brain-wide gr-expression from each person (using the people × [brain-wide gr-expression] matrix). As expected, this analysis partitioned people into clusters of African, Asian, and European populations with gradients between these clusters reflecting known patterns of genetic admixture (Fig 2A). This embedding reflects patterns of genetic ancestry that are known and were previously described in analyses of genetic-sequence data [58].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Genetic ancestry, regional identity, and inter-regional organization of estimated gr-expression.

(A, B) Principal component embeddings of estimated gr-expression from the ethnically diverse sample of the UK-Biobank cohort (S1 Table). (A) An embedding of brain-wide gr-expression: scatter plots of principal components of the people × [brain-wide gr-expression] matrix, where people denote people from the UK-Biobank sample and brain-wide gr-expression denotes brain-wide estimates of gr-expression for all genes that had Joint-Tissue Imputation models for each of the 10 regions. (B) An embedding of regional gr-expression: scatter plots of principal components of the regions × [regional gr-expression] matrix where regions denote the 10 regions of people from the UK-Biobank sample and regional gr-expression denotes regional estimates of gr-expression for all genes that had Joint-Tissue Imputation models for each of these regions. (C–K) A 3 × 3 matrix of plots of inter-regional coexpression: correlations between directly assayed expression and estimated gr-expression. The first row and column show results on directly assayed gene expression data from the Allen Human Brain Atlas. The second row and column show results on directly assayed gene expression data from the GTEx project. The third row and column show results on estimated gr-expression from the ethnically diverse sample of the UK-Biobank sample. (C, G, K) Associations between inter-regional coexpression and Euclidean distance in each data set. (D, E, H) Associations between inter-regional coexpression across data sets. P-values denote the probability of obtaining coexpression of at least equal magnitude in data with preserved correlation coefficients between coexpression and Euclidean distance (estimated from 10,000 random samples). (F, I, J) Heatmaps of inter-regional coexpression, averaged across people in each data set (regional numbers follow numbers in panel B). DLPFC, dorsolateral prefrontal cortex; GTEx, Genotype-Tissue Expression Project.

https://doi.org/10.1371/journal.pbio.3002782.g002

Second, we tested if gr-expression patterns reflected regional brain identities across people in the same sample. For this analysis, we generated principal component embeddings of individual region-specific gr-expression (using the regions × [regional gr-expression] matrix). This analysis partitioned gr-expression into well-delineated regional clusters and revealed anatomically interpretable groups of cortical, limbic, and basal ganglionic clusters (Fig 2B). Collectively, these results show that gr-expression simultaneously reflects genetic-ancestry identities across people and brain-regional identities within people. They imply, specifically, that associations of gr-expression, or TWAS, can capture variation across people, similarly to GWAS, as well as variation across regions, similarly to regional transcriptomic studies.

Third, we compared inter-regional correlations of estimated gr-expression to inter-regional correlations of directly assayed expression data from the Allen Human Brain Atlas and the GTEx Project. Recent studies have shown that inter-regional coexpression exponentially decays as a function of inter-regional distance [59,60]. We reproduced these relationships by showing strong inverse nonlinear relationships between inter-regional coexpression in the Allen and GTEx data and Euclidean distance: Allen versus distance r_spearman = −0.711 and GTEx versus distance r_spearman = −0.721 (Fig 2C and 2G). We found a similar, albeit weaker, relationship in the estimated gr-coexpression data: UK Biobank versus distance r_spearman = −0.480 (Fig 2K). More directly, we found strong linear relationships between the inter-regional coexpression in the Allen and GTEx data: Allen versus GTEx r_pearson = 0.683 (Fig 2D). We found similar relationships between estimated and directly assayed inter-regional coexpression: UKB versus Allen r_pearson = 0.613 and UKB versus GTEx r_pearson = 0.861 (Fig 2E and 2H). Heatmaps of all coexpression patterns reflected associations between cortical, basal ganglionic, and other subcortical systems (Fig 2F, 2I and 2J). Finally, we showed that the relationship of coexpression with distance was not sufficient to explain these similarities of coexpression (p ≤ 0.005 for all tests).

Collectively, these results provide multifaceted support for the biological validity, anatomical interpretability, and practical utility of estimated gr-expression. In this way, they establish a foundation for the use of gr-expression in neuroimaging TWAS.

TWAS link genetically regulated gene expression with regional gray-matter volumes

We hypothesized that the integration of multiple SNPs into models of regional gr-expression would allow us to detect novel and neurobiologically meaningful associations. To test this hypothesis, we performed TWAS to identify associations between individual variation of regional gr-expression and gray-matter volumes (Fig 3A). Gray-matter volumes are heritable phenotypes that have been linked to many genetic variants in previous GWAS [16,17,31]. We focused our association studies on 8 regions with available FreeSurfer [61] segmentations and therefore excluded substantia nigra and hypothalamus from subsequent analyses (see Methods for regional definitions).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Variation of regional gr-expression and regional volumes across people.

(A) A pipeline for transcriptome-wide association studies, or TWAS, of neuroimaging phenotypes. The inputs to TWAS comprise values of regional gr-expression (left) and regional phenotypes (right), estimated in the same people. The outputs are associations between the individual variation of regionally specific gr-expression and neuroimaging phenotypes across people (center). (B) Within-regional associations of gr-expression and gray-matter volumes for 2 representative regions. Each point denotes an association between the individual variation of gr-expression and volume in the same region. The horizontal axis shows the chromosome location of individual genes. The vertical axis shows the p-values (–log₁₀ p) of associations. Solid-color points represent associations that pass the thresholds of p_FDR = 0.05 or p_Bonferroni = 0.05 (horizontal lines). Source data can be found in S2 Table. (C) Associations between SNP-based GWAS and gene-based TWAS for 2 representative regions. Left: Scatter plots of p-values (–log₁₀ p) for associations of all genes and SNPs. These plots preserve all genes and SNPs but lack the one-to-one relationship between genes and SNPs. Right: Corresponding scatter plots for the best-performing genes and SNPs. Each gene in TWAS matches with its best-performing SNP in GWAS. Similarly, each SNP in GWAS matches with its best-performing gene in TWAS. These plots show one-to-one relationships but exclude many genes and SNPs. (D) Numbers of associations (p_FDR < 0.05 or p_Bonferroni < 0.05) detected with TWAS and GWAS. Solid colors denote numbers of associations detected with TWAS alone. Beige colors denote number of genes detected with GWAS alone. Stripe patterns denote numbers of genes detected with both TWAS and GWAS. The top bar for each region adopts an FDR correction for TWAS associations (p_FDR < 0.05), while the bottom bar adopts a stricter Bonferroni correction (p_Bonferroni < 0.05). (E, F) Enrichment analyses of TWAS for biological annotations in the NHGRI-EBI GWAS Catalog. (E) Enrichment for biological annotations of genes whose gr-expression predicted regional volumes (p_FDR < 0.05). Each point represents a biological annotation associated with at least 1 gene. The horizontal axis shows the p-values (–log₁₀ p_FDR) of individual annotations. Source data can be found in S3 Table. (F) Relationship between p-values and brain-relatedness of biological annotations. The horizontal axis shows bins of p-values (–log₁₀ p_FDR). The vertical axis shows the fraction of brain-related annotations within each bin. The p-value on the correlation coefficient was computed by permuting the annotations (estimated from 10,000 random samples). (G, H) Heatmaps of inter-regional TWAS between gr-expression and regional volumes. (G) Absolute numbers of associations. Numbers of genes whose gr-expression in 1 region (columns) predicted (p_FDR < 0.05) the volume of another region (rows). Source data can be found in S4 Table. (H) Overlap coefficients. Number of genes that were common to both intra-regional and inter-regional associations in G, normalized by the size of the smaller of the intra- and inter-regional gene sets. FDR, false discovery rate; GWAS, genome-wide association studies; SNP, single-nucleotide polymorphism; TWAS, transcriptome-wide association studies.

https://doi.org/10.1371/journal.pbio.3002782.g003

Our first TWAS inferred associations between gray-matter volumes and gr-expression of the same regions. To minimize the confounders of genetic ancestry, we restricted our analyses to the “White British” sample of the UK-Biobank cohort (S1 Table) [37]. We therefore performed TWAS on 39,565 people (52.2% female, 64.3 ± 7.7 years old), with covariates of genetic ancestry, sex, and age (Methods).

We identified 1,065 associations (of 778 unique genes) between gr-expression and the volumes of 8 brain regions (p_FDR < 0.05, Fig 3B and S1 and S2 Tables). The number of regional associations varied from 68 genes in the amygdala to 205 genes in the cerebellar hemisphere. Many genes that were found in this analysis, including CRHR1, ARL17A, NSF, and OGFOD2, have been implicated in previous GWAS of regional brain volumes, and have also been linked to brain disorders, including epilepsy, schizophrenia, and brain cancer [62–64].

TWAS reinforce GWAS associations and discover novel associations

To directly show the methodological advantages of gene-based TWAS, we directly compared these studies to SNP-based GWAS. We made this comparison in 3 complementary ways.

Direct relationship to GWAS. First, we performed a GWAS on the same sample and compared our TWAS associations for individual genes to GWAS results for the SNPs that formed part of corresponding models of gr-expression. These comparisons were dominated by many-to-many relationships between genes and SNPs, because several SNPs typically predict the gr-expression of a single gene, and similarly, a single SNP can help predict the gr-expression of several genes. The correlations between GWAS and TWAS p-values were moderate but statistically significant (0.275 ≤ r_spearman ≤ 0.373, p < 0.001 for all regions, Fig 3C left, S1 Fig). To focus on the strongest TWAS and GWAS signals, we filtered these data in a way that retained the lowest p-value SNP for each gene and, simultaneously, the lowest p-value gene for each SNP. This process resulted in much stronger and strictly one-to-one relationships (0.479 ≤ r_spearman ≤ 0.583, p < 0.001 for all regions, Fig 3C right and S1 Fig). Collectively, these results show that gene-based TWAS associations are related to, but also distinct from, SNP-based GWAS associations.

Statistical power. Second, we investigated the nature of these differences by contrasting the number of associations detected by TWAS and GWAS. The high multiple-testing burden of GWAS typically requires strict genome-wide Bonferroni corrections. By contrast, the relatively smaller number of statistical tests in TWAS results in a lower multiple testing burden, and the expected polygenic associations of many phenotypes make it common to adopt less strict false discovery rate (FDR) corrections as an alternative to Bonferroni [48]. In our analyses, TWAS under both corrections identified many more genes than the corresponding GWAS (Fig 3D). Specifically, under FDR correction, TWAS detected associations of 673 unique genes (p_FDR < 0.05) that lacked GWAS associations of corresponding SNPs (p_Bonferroni < 0.05). Many of these genes have been previously linked to brain-related disorders, including Alzheimer’s disease (WDR12, AGFG2, and CDK5RAP3), schizophrenia (SRA1, WDR55, CORO7, DDAH2, PCDHA8), autism spectrum disorder (MAPK3, PCDHA13), and major depressive disorder (ZMAT2 and ITIH4) [65–74]. Separately, under Bonferroni correction, TWAS detected associations of 110 unique genes (p_Bonferroni < 0.05) that lacked GWAS associations of corresponding SNPs (p_Bonferroni < 0.05). These results show that TWAS discovers associations of many genes that are undetected with GWAS.

Neurobiological interpretability. Third, to interpret the function of discovered genes more systematically, we tested the enrichment of our TWAS results using the NHGRI-EBI GWAS Catalog, a catalog of gene annotations curated from all human GWAS in the current literature [75]. We discovered 276 enriched biological annotations at p_FDR < 0.05 (Fig 3E and S3 Table) and found that brain-related annotations were much more likely to be enriched than other annotations in the catalog (p < 0.001). Moreover, in addition to the overall enrichment for brain-related annotations, we found a strong positive correlation between the p-values of the enrichment and the fraction of discovered brain-related annotations (r_spearman = 0.964, p < 0.001, Fig 3F). In other words, we found that the most enriched gene annotations were primarily brain related. S2 Fig shows that these enrichments were replicable with a Bonferroni correction on TWAS associations. Collectively, these results show the neurobiological relevance of our discoveries.

TWAS discover associations of genetically regulated gene expression in one brain region with gray-matter volumes of other regions

Separately, we built on our region-specific TWAS findings to test for associations between gr-expression in one brain region and gray-matter volumes of other regions. Such associations are undefined for SNPs (because all cells share the same genome), but are interpretable for gr-expression (because of known inter-regional similarities in gene expression and organization [15,20,23,25]). In practical terms, these analyses also help to discover associations of regional volumes with genes for which these regions currently lack models of gr-expression (Fig 1B).

Inter-regional TWAS discovered between 73 and 209 (median 133) associations (p_FDR < 0.05) of gr-expression in one region with the volume of another region (Fig 3G and S4 Table). gr-Expression in the amygdala and anterior cingulate had the largest number of such associations (Fig 3G, columns) relative to the total available number of gr-expression models in each region (Fig 1B). For example, the gr-expression of FOXO3 in the anterior cingulate predicted the volumes of all 8 regions. This gene has been strongly linked to healthy aging in diverse human populations [76–78]. By contrast, the volume of putamen was predicted by the largest number of genes from other regions (Fig 3G, rows). Several of these genes—including MYLK2, KTN1, DCC, BCL2L1, TPX2, and HELZ—were associated with putamen volume in previous studies [16,79–84]. In particular, in our study, the gr-expression of MYLK2 and KTN1 predicted putamen volume in all regions that had gr-expression models of these genes (in 8 and 4 regions, respectively). In other cases, gr-expression of some genes in many regions predicted volumes of many other regions. For example, the gr-expression of LRRC37A2 in all 8 regions predicted volumes of all regions except putamen and caudate. Similarly, gr-expression of MAPT in the cerebellar hemisphere predicted all volumes except putamen and caudate. Both LRRC37A2 and MAPT have been linked to Parkinson’s disease, and MAPT encodes for tau and has been well studied in the Alzheimer’s disease literature [50,85,86].

We finally quantified the overlap between intra-regional and inter-regional associations. A heatmap of overlap coefficients of these associations formed 3 anatomically distinct groupings of cortical, basal ganglionic, and limbic regions (Fig 3H). These groupings show that the volumes of anatomically similar regions are more likely to share gene associations or, alternatively, that genes from one region are associated with volumes of anatomically similar regions. S2 Fig shows that these groupings were replicable with a Bonferroni correction on TWAS associations.

Collectively, these results suggest a strong relationship between gr-expression profiles of anatomically similar brain regions and, more generally, show the utility of inter-regional TWAS of neuroimaging phenotypes.

Genetically regulated gene expression links regional volumes with clinical phenotypes

We next moved beyond literature-based annotations to test whether gr-expression associations can link regional volumes with clinical phenotypes. To achieve this, we integrated our results with a separate TWAS on a sample of 70,439 people in BioVU, a biobank that contains DNA samples and de-identified electronic health records for patients at Vanderbilt University Medical Center [53,87,88]. Clinical phenotypes derived from electronic health records in BioVU were represented by phenotype codes extracted from International Classification of Diseases (ICD-9) billing codes. The BioVU TWAS used the same Joint-Tissue Imputation models to estimate gr-expression and to discover clinical associations (Fig 4A). In what follows, we filtered this clinical TWAS to focus on 156 brain-related clinical phenotypes. We then compared associations of regional gr-expression with these phenotypes to associations in our inter-regional neuroimaging TWAS.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Association of gr-expression with both neuroimaging and clinical phenotypes.

(A) Pipeline for BioVU TWAS: transcriptome-wide association studies of regional gr-expression and clinical phenotypes from the BioVU Biobank. Top left: Inputs to TWAS comprise electronic health records and DNA samples of the same people. Top center: Clinical phenotypes are extracted from ICD-9 codes present in electronic health records. Bottom left and center: Regional gr-expression is estimated from DNA samples of the same people. Right: Clinical phenotypes and regional gr-expression are combined in the BioVU TWAS. (B) Heatmap showing the number of times by which genes (rows) with regional gr-expression (columns) were linked to both regional volumes and clinical phenotypes. Each count denotes a regional gr-expression that was associated (p_FDR < 0.05) with both a regional volume in the UK Biobank TWAS and with a brain-related clinical phenotype in the BioVU TWAS. (C) Heatmap showing the number of genes with regional gr-expression that linked regional volumes (columns) with clinical phenotypes (rows). Each count denotes a regional gr-expression that was associated (p_FDR < 0.05) with both a regional volume in the UK Biobank TWAS and with a brain-related clinical phenotype in the BioVU TWAS. (D) Enrichment of clinical phenotypes for genes whose gr-expression predicted (p_FDR < 0.05) regional volumes (rows) in the UK Biobank TWAS. Each point represents a brain-related clinical phenotype associated with at least 1 gene. The horizontal axis shows the p-values (–log₁₀ p_FDR) of individual phenotypes. Source data can be found in S5 Table. FDR, false discovery rate; TWAS, transcriptome-wide association studies.

https://doi.org/10.1371/journal.pbio.3002782.g004

We identified 98 genes whose gr-expression in a specific region associated (p_FDR < 0.05) with both volumes in the UK Biobank TWAS and with brain-related clinical phenotypes in the BioVU TWAS (Fig 4B). There were 22 genes in this set whose gr-expression in 4 or more regions linked volumes and clinical phenotypes. In previous GWAS and clinical studies, these genes have been associated with neurogenesis (WNT3) [89,90], neurodevelopmental delays (QRICH1) [91,92], addiction (HCG27) [93], depression (CCDC71, CYP21A2) [94,95], and other brain-related disorders [96,97].

BioVU clinical phenotypes that shared associations of gr-expression with regional volumes included a variety of nervous system symptoms and disorders including, most prominently, demyelinating diseases, motor-related symptoms, and dementia (Fig 4C). Several HLA genes that play a major role in the immune response (including HLA-B/C, HLA-DRB1, and HLA-DRB5) were associated with 2 or more regional volumes and simultaneously with demyelinating diseases, including multiple sclerosis, a prominent immune-mediated disorder [98]. In addition, genes in the HLA-DR and HLA-DQ families were associated with volumes of the cerebellar hemisphere and hippocampus in the UK Biobank and simultaneously with the abnormal movement phenotype in the BioVU TWAS. These associations represent candidate causal mechanisms for linking these genes with Parkinson’s disease and other movement disorders [99–102]. Genes C4B, MST1, and LRRC37A showed similar patterns of associations, in this way supporting and expanding previous links to motor disorders [86,103–105].

Separately, we identified 9 brain-related clinical phenotypes that were enriched (p_FDR < 0.05) for genes whose gr-expression predicted regional volumes (Fig 4D and S5 Table). Most of these phenotypes were enriched for genes that predicted multiple regional volumes. For example, myoclonus was enriched for genes that predicted volumes of 6 regions, while multiple sclerosis and lack of coordination were enriched for genes that predicted volumes of 4 regions. Further, senile dementia was enriched for genes that predicted hippocampal and cerebellar volumes, while speech disturbances was enriched for genes that predicted anterior cingulate volume. The majority of motor-related clinical phenotypes were enriched for genes that predicted volumes of the cerebellum, a well-known center of motor control. S3 Fig shows that our association and enrichment analyses were replicable with a Bonferroni correction on TWAS associations.

Overall, these results show that associations of gr-expression with phenotypes at different biological scales can be combined to reveal genes that link regional volumes and clinical phenotypes. Despite differences in samples and phenotype modalities, we identified a large overlap in the 2 TWAS between associations with regional gr-expression. Furthermore, we found evidence in related literature that supports associations between regional volumes and an array of brain-related disorders. Collectively, these findings highlight the integrated relationships between gene expression and brain phenotypes and the implications of these relationships for the study of brain-related disorders.

Polygenic models of genetically regulated gene expression detect associations in a small neuroimaging data set

Recent studies have shown the potential of combining the gr-expression of multiple genes into polygenic models to improve the prediction of phenotypes (Fig 5A) [106–108]. Such polygenic models may be particularly relevant for highly polygenic phenotypes of brain anatomy and activity. They can also capture the polygenic nature of structural and functional MRI phenotypes and further reduce the number of statistical tests.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Associations of polygenic gr-expression with neuroimaging phenotypes.

(A) A framework for polygenic modeling of regional phenotypes. Polygenic gr-expression was defined as the mean normalized gr-expression of genes that were nominally associated with phenotypes at p < 0.001, uncorrected. To make the mean well defined, the signs of gr-expressions with negative associations were reversed. Model performance was evaluated using permutation testing. (B) Representative scatter plots of neuroimaging phenotypes and polygenic gr-expression. Points represent individuals, and colors denote regions as labeled in C. (C) Pearson correlation coefficients between neuroimaging phenotypes and polygenic gr-expression (square points), polygenic gr-expression from permutation tests (box plots; n = 10,000), and best single-gene gr-expression from TWAS (round points). Stars represent p-values of polygenic associations with permutation testing (* p < 0.05, ** p < 0.005, *** p < 0.0005). Source data can be found in S1 Data. (D) Comparison of p-values (−log₁₀ p_FDR) from polygenic gr-expression associations and best single-gene TWAS. Colors denote regions, while lines denote p = 0.05. DLPFC, dorsolateral prefrontal cortex; FDR, false discovery rate; TWAS, transcriptome-wide association studies.

https://doi.org/10.1371/journal.pbio.3002782.g005

Here, we tested the power of such analyses using the HCP [54], a small but prominent neuroimaging genomic data set with high-quality functional MRI data. To minimize the confounders of genetic ancestry, we restricted our analysis of this data set to a sample of 657 non-twins of European genetic ancestry. Our analyses considered regional volume phenotypes, as well as a representative set of functional MRI phenotypes. The functional MRI phenotypes track properties of regional activity (amplitude of low-frequency fluctuations [109]), within-regional correlation (regional homogeneity [110]), and average inter-regional correlated activity (mean coactivity [111]). Specifically, amplitude reflects the power of low-frequency activity, homogeneity reflects the extent of intra-regional correlated activity, while coactivity complementarily reflects the extent of inter-regional correlated activity (Methods). These phenotypes provide insights into the organization of brain activity and have been extensively studied in neuroimaging genomics [26,27,112–114].

We first performed a single-gene TWAS on these phenotypes. The relatively small size of our sample, however, necessarily resulted in few associations that survived corrections for multiple comparisons. For example, and in contrast to the UK Biobank TWAS, most regional phenotypes in this analysis showed no associations at p_FDR < 0.05. Moreover, as expected, the strongest associations in this sample had much higher p-values (best p_FDR = 0.017) than the strongest associations in the UK Biobank TWAS (best p_FDR = 1.34 × 10⁻²¹).

We then estimated polygenic gr-expression as the mean normalized gr-expression of genes that had nominal TWAS associations with phenotypes (p < 0.001, uncorrected). We tested associations of polygenic gr-expression against null associations of equivalently estimated polygenic gr-expression on data with randomized (permuted) assignment of phenotypes to subjects.

Associations of polygenic gr-expression with phenotypes had mean ± standard deviation r = 0.434 ± 0.113 (Figs 5B and S4). For regional volume and homogeneity phenotypes, these associations tended to be higher than null associations (p < 0.05) and have lower p-values than single-gene associations (Fig 5C and 5D and S1 Data). By contrast, for amplitude and coactivity phenotypes, these associations did not tend to be higher than null associations and had similar p-values as single-gene associations (Fig 5C and 5D and S1 Data). Note also that polygenic gr-expression estimated from more selected genes tended to have higher associations in absolute terms and relative to the null associations (S5 Fig). Collectively, these analyses show that polygenic modeling can further improve the ability of TWAS to infer associations of groups of genes with complex phenotypes.

Replicability of estimated genetically regulated gene expression and TWAS

We finally tested the replicability of our analyses in 3 complementary ways.

First, we tested the replicability of gr-expression models by comparing the estimated gr-expression of the DLPFC using models trained on 2 distinct postmortem samples: our main sample from GTEx and an independent replication sample from PsychEncode [43]. We found that models trained on the 2 samples had highly similar patterns of gr-expression (r_pearson of gr-expression: median 0.799, Q1–Q3 0.559–0.917, Fig 6A). Likewise, we found similar TWAS of these models with DLPFC volumes (r_spearman = 0.540, p < 0.001, Figs 6B and S6). These results suggest that our framework for estimating gr-expression is robust to the training data, at least for sufficiently large samples.

Second, we tested the replicability of association p-values and magnitudes in the UK Biobank using the independent HCP TWAS. As we saw above, the small HCP sample produced almost no associations at p_FDR < 0.05. Correspondingly, we found that a small percentage of associations with p_FDR < 0.05 in the UK Biobank were also present at the nominal threshold of p < 0.05 in the HCP TWAS (median 7.00%, Q1–Q3 4.53%–7.75%, Fig 6C). By contrast, the magnitudes of individual associations are strongly correlated with p-values (r_spearman between magnitudes and −log₁₀p: median 0.783, Q1–Q3 0.773–0.786, S6 Fig) but, unlike p-values, are relatively independent of the sample size [115]. Correspondingly, we found consistently strong correlations between magnitudes of associations that passed p_FDR < 0.05 in the UK Biobank TWAS (r_spearman: median 0.518, Q1–Q3 0.486–0.622, all p < 0.005, Figs 6D, 6E and S6).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Replicability of estimated genetically regulated gene expression and TWAS.

(A, B) Replication of estimated gr-expression trained on independent PsychEncode data. (A) Histogram of correlations between gr-expression of the DLPFC estimated with models trained on GTEx data and independent PsychEncode data. (B) Scatter plot of TWAS associations based on gr-expression of the DLPFC estimated with models trained on GTEx data and independent PsychEncode data. Each point denotes p-values of associations between estimated gr-expression and DLPFC gray-matter volumes in the white-British sample of the UK-Biobank cohort. (C–E) Replication of genes that passed p_FDR = 0.05 in discovery TWAS of gray-matter volumes. (C) Percentages of genes that were replicated at nominal p < 0.05 in replication TWAS. Source data can be found in S2 Data. (D) Correlations between effect magnitudes of genes in the replication and discovery TWAS. Dots denote analyses on the full UK Biobank (discovery) and HCP (replication) samples. Box plots denote analyses of discovery-replication splits of the white-British UK-Biobank sample, ordered from small to large replication samples. Each box plot was estimated from 300 random splits. (E) Scatter plots of effect magnitudes in the UK Biobank and HCP TWAS. Each point denotes effect magnitudes for a gene that showed p_FDR < 0.05 in the UK Biobank TWAS. DLPFC, dorsolateral prefrontal cortex; FDR, false discovery rate; GTEx, Genotype-Tissue Expression Project; HCP, Human Connectome Project; TWAS, transcriptome-wide association studies.

https://doi.org/10.1371/journal.pbio.3002782.g006

Third, we repeated these analyses on TWAS of discovery and replication subsets generated from 1,200 random splits of the white-British UK-Biobank sample (S2 Data). These additional analyses showed that replication samples of the same size as our HCP sample (657 people) had similarly small percentages of replicable associations (median 6.60%, Q1–Q3 6.20%−6.82%) and that larger samples showed much higher percentages (Fig 6C). Likewise, these analyses showed that replication samples of the same size as our HCP sample had strong correlations between magnitudes of effects (r_spearman: median 0.575, Q1–Q3 0.568–0.579; all p < 0.001) and that larger samples showed modestly increased correlations between magnitudes (Figs 6D and S6).

Collectively, these analyses suggest that the estimated gr-expression and magnitudes of TWAS associations were generally replicable, while the p-values of TWAS associations were replicable in large replication samples.

Interactive application to facilitate adoption of TWAS in neuroimaging genomics

To increase the accessibility of our results, we created a browser-based application to explore our SNP-based and gene-based associations (https://github.com/nhunghoang/twas-webapp). The application allows users to compare neuroimaging GWAS with neuroimaging TWAS and with clinical TWAS and, in this way, links analyses of SNPs, genes, neuroimaging phenotypes, and clinical phenotypes. It also allows users to interactively explore associations and provides more direct gene-based interpretations of SNP-based results.

Summary

We adopted state-of-the-art methods from computational genomics to estimate genetically regulated gene expression, or gr-expression, across 10 cortical and subcortical brain regions in more than 40,000 people. First, we showed that estimates of gr-expression across people and brain regions recapitulate the neurobiological organization of directly assayed gene expression. Second, we showed that TWAS based on estimated gr-expression aligned with, and extended, associations from corresponding GWAS. Third, we integrated these results with a set of independent associations between regional gr-expression and brain-related clinical phenotypes extracted from electronic health records. Fourth, we showed that polygenic models of gr-expression can further increase the statistical power of our approach. Finally, we showed that estimated gr-expression levels and the magnitudes of TWAS associations were generally replicable while the p-values of TWAS associations were replicable in large samples.

Advances

Our study shows that gene-based association analyses can bridge gaps in existing neuroimaging genomic studies. Specifically, the approach begins to fill the mechanistic gap in traditional GWAS by extending these studies to identify associations of phenotypes with regionally specific gr-expression, rather than with regionally agnostic SNPs. Moreover, the approach reduces the multiple-testing burden of GWAS by orders of magnitude. Separately, the approach complements regional transcriptomic studies by extending these studies to thousands of people with available genomes.

We demonstrated the unique combination of these advantages with 3 complementary analyses. First, we showed that, like GWAS, our method separates people by genetic ancestry (Fig 2A). Second, we showed that, like regional transcriptomic analyses, our method separates brain regions by patterns of gr-expression (Fig 2B). Third, we showed strong similarities between the inter-regional correlation of gr-expression from the UK Biobank and directly assayed expression from the Allen Human Brain Atlas and the GTEx Project (Fig 2C–2K). In this way, we showed that our approach produces neurobiologically interpretable estimates of regionally specific gr-expression in large human populations.

Our method allowed us to directly benchmark gene-based TWAS against SNP-based GWAS. First, we showed that TWAS associations are broadly related to, but also distinct from, GWAS associations (Fig 3C). Second, we showed that TWAS can discover associations of many genes that are otherwise undetected with GWAS (Fig 3D) and that these discoveries strongly favor brain-related annotations (Fig 3E). Third, we showed that inter-regional associations are interpretable and further increase the utility of TWAS (Fig 3F). Collectively, these results directly demonstrate the conceptual and practical strengths of TWAS.

Separately, our study built on these results in 3 additional ways. First, it integrated estimates of regionally specific gr-expression with gray-matter volumes and brain-related clinical phenotypes (Fig 4). Second, it extended the single-gene TWAS to build polygenic models of neuroimaging phenotypes (Fig 5). Third, it showed that the magnitudes of TWAS associations are generally replicable but that the p-values of TWAS associations are highly sensitive to sample sizes (Fig 6). These results outline a path towards the replicable integration of polygenic gr-expression with complex neuroimaging and clinical phenotypes.

Separately, the study of gr-expression provides unique advantages over the study of directly assayed expression because it allows to focus on the stable, genetically regulated aspects of gene expression without the need to control for the potential confounders of environmental factors and acquisition biases, including batch effects [116]. Similarly, an important advantage of this study relative to group-averaged transcriptomic studies is the lack of evident bias attributable to distance effects [60]. This lack of bias arises because the associations are computed over people, rather than over brain regions. For example, while we found that the observed inter-regional correlation between estimated gr-expression and directly assayed expression could not be explained solely by distance dependence, a distance-based explanation would not invalidate our results because it would reflect biological, rather than artifactual, effects.

Limitations

Our approach has many benefits, but it also has limitations. First, our analyses still require large samples to enable replicable associations (Fig 6C–6E). Nonetheless, the lower multiple-testing burden of TWAS makes this problem less acute than for tests of millions of SNPs in GWAS. Similarly, much like correlations of adjacent SNPs in GWAS, correlations of gr-expression in TWAS, while generally smaller and less common, can make it difficult to fine-map causal genes. Future studies could adopt mendelian randomization to enable causal inference, although this approach comes with its own limitations, including the difficulty of accounting for horizontal pleiotropy (the effect of one gene on unrelated phenotypes) [117–119]. Finally, our focus on genes necessarily misses the effects of variants that operate through means other than the regulation of gene expression.

Second, relative to spatially specific transcriptomic studies, our approach is restricted to a small number of brain regions. In future studies, we propose to overcome this limitation by modeling known relationships between regional and network organization [120]. We also propose that the adoption of similar models will allow researchers to integrate regional associations with high-resolution single-cell atlases of gene expression and link these associations to specific cell types [121].

Third, our results integrate genomic biobanks with available data on genetics, gene expression, neuroimaging, and clinical phenotypes. Such integration necessarily comes with the challenges of demographic diversity and matching. Our analyses were primarily based on European populations and may not necessarily generalize to other populations. As genomic, transcriptomic, neuroimaging, and clinical data continue to increase in size and scope, it will be important to extend these results to analyses of other populations.

Joint-Tissue Imputation models

We used Joint-Tissue Imputation models of gr-expression that were previously trained on postmortem gene expression data from GTEx. In this section, we describe the main aspects of quality control and training of these models. We refer the readers to the original studies of Joint Tissue Imputation [43] and the GTEx v8 data set [56] for a more detailed discussion of these approaches.

Joint-Tissue Imputation models estimate gr-expression as the linear combination of SNPs that are close to the gene of interest along the linear genome. The training of these models, therefore, required data on tissue-specific gene expression and whole-genome sequencing from the same people. The GTEx v8 data set included these data for brain regions of 838 donors that passed internal GTEx biospecimen quality controls [122]. The donors had the following demographics: Age, 21 to 70 years (mean 53); sex, 34% female; ancestry, 85.3% European American/12.3% African American/1.4% Asian American. The data set contained RNA-seq from ten brain regions. Table 1 summarizes the names of these regions and the number of samples used to train models in each region. For completeness, it also summarizes our definitions of the corresponding regions in volumetric Allen Human Brain Atlas data, and in surface-based UK-Biobank and Human-Connectome Project data.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. GTEx regions (tissues) and corresponding definitions of volume-based and surface-based imaging parcellations.

https://doi.org/10.1371/journal.pbio.3002782.t001

The following steps were taken to maximize the predictive accuracy of estimation and to minimize confounders. First, the assayed gene expression levels were controlled for sex, sequencing platform, the top 5 principal components, as well as probabilistic estimation of expression residuals, a Bayesian model of hidden confounders [123]. Second, the models were trained only on biallelic SNPs that had a minor allele frequency of at least 0.05 and that were in Hardy–Weinberg equilibrium (p > 0.05), i.e., only on SNPs that had both sufficient and stable variation. Third, to reduce the effects of linkage disequilibrium, highly correlated (r² ≥ 0.9) SNPs were pruned and the models were trained only on SNPs near the gene of interest. The optimal threshold for proximity was determined separately for each gene by cross-validation. Finally, to additionally control for overfitting, the models incorporated elastic-net regularization and the training was based on 5-fold cross-validation.

As part of the replication analysis (S6 Fig), we also considered Joint-Tissue Imputation models that were trained on 415 samples of sequencing and expression data from the DLPFC in the PsychENCODE project [57]. These data were processed and trained in the same way as the original study of Joint-Tissue Imputation. All pretrained models are available online at https://doi.org/10.5281/zenodo.3842289 (GTEx-trained models) and https://doi.org/10.5281/zenodo.3859065 (PsychEncode-trained models).

Genotype-Tissue Expression Project (GTEx) and Allen Human Brain Atlas Data

Our analyses of inter-regional correlations (Fig 2) compared the estimated gr-expression data described in the previous section to directly assayed expression data from GTEx and the Allen Human Brain Atlas. This section describes our preprocessing of these latter data sets.

We downloaded the most recent release (v8) of the GTEx gene-expression data from https://gtexportal.org/home/downloads/adult-gtex. These data were acquired from 340 donors (an average of 199 donors per region). Gene expression levels were quantified and normalized by GTEx, and genes were selected based on expression thresholds as previously described [56].

We downloaded the Allen Human Brain Atlas microarray gene-expression data from https://human.brain-map.org/static/download. The data were acquired from 6 donors (42 ± 12 years old, 1 female). Brain-wide gene expression levels were quantified and normalized by the Allen Institute, as previously described [130].

Our preprocessing of these data followed current best practices [131]. All imputed and directly assayed expression data were normalized to have zero mean and unit variance across regions. In addition, data from the Allen Human Brain Atlas were filtered to exclude genes whose expression level did not exceed the background signal (as specified by the file PACall.csv). These data were also nonlinearly registered to reference coordinate space [132], assigned to regions with a 2 mm distance threshold, and averaged across all available probes and the left and right hemispheres.

UK Biobank genomic and neuroimaging data

We analyzed data from 45,549 people, or all available people from the UK Biobank with genome-wide genotyping and neuroimaging volumes. Our sample had the following demographics: Age, 64 ± 7.7 years old; sex, 52% female; self-reported ethnicity, 96.7% white/0.6% black/1.1% South Asian/0.3% Chinese/0.5% Mixed/0.8% Other (S1 Table). In this section, we describe the main aspects of quality control and processing of these data by the UK Biobank. We refer the readers to the original publications [17,37,133] for a more detailed discussion of these and other questions.

Genome-wide genotype imputation was performed using data from the Haplotype Reference Consortium [134] as the main imputation reference panel, as well as merged UK10K and 1000 Genomes Phase 3 data sets as the secondary imputation reference panel [135]. The data passed an automated quality-control pipeline [37]. This pipeline comprised marker-based and sample-based quality control. Marker-based control included tests for batch effects, plate effects, deviation from Hardy–Weinberg equilibrium, sex effects, array effects, and sequencing replicability. Separately, sample-based control included tests for unusually high fractions of heterozygous or missing loci, as well as for mismatch between self-reported sex and the intensity of sex-chromosome markers.

Our TWAS (Figs 3 and 4) sought to minimize the confounders of genetic ancestry by focusing on the “White British” sample of the UK-Biobank cohort (39,565 people). We followed UK Biobank analyses to select people who self-reported as “White British” and who had similar genetic ancestry based on UK-Biobank principal component analysis on 147,604 genotype markers (pruned to minimized linkage disequilibrium) over 407,219 unrelated people [37]. By contrast, our analyses of genetic ancestral relationships, regional identities, and inter-regional correlations (Fig 2) focused on the remaining ethnically diverse sample of the UK-Biobank cohort (5,984 people).

All UK Biobank neuroimaging data were processed by the UK Biobank automated brain-imaging pipeline [133]. The pipeline flagged missing and distorted data, registered images to common reference space, and computed imaging-derived phenotypes. We used phenotypes computed on MRI scans from the first imaging visit. Our analyses included only cortical and subcortical brain regions that had available estimates of gray-matter volume and gr-expression and that passed UK Biobank quality control exclusion criteria (see the original reference [133] for a detailed discussion). Volumes of these regions were computed by the UK Biobank using FreeSurfer cortical and subcortical segmentations [61] (Table 1) and were averaged across the left and right hemispheres.

Human Connectome Project genomic and neuroimaging data

The full HCP contains 1,142 people with brain-wide genotyping sequences and neuroimaging data. This cohort had the following demographics: Age 29 ± 3.7 years old, 54% females, 149 pairs of monozygotic twins. In this section, we describe our curation of these data to generate a sample of 657 people. We also describe the main aspects of quality control and processing of these data by the HCP, and our estimation of phenotypes from these data. We refer the readers to the original HCP publications [136–139] for an additional extensive discussion of quality control and data processing.

Genotyping of all people was derived from blood or saliva samples. The genotype data comprised probabilities of single-nucleotide variants, estimated using the Illumina Multi-Ethnic Global Array. Quality-control procedures of these data included verification of self-reported common ancestry for siblings, as well as zygosity for twins.

As in our analyses of the UK Biobank, we sought to minimize the confounders of genetic ancestry by focusing our association analyses on a sample of 657 non-twins of European genetic ancestry. We estimated genetic ancestry using the principal components of genotyping data from this cohort, computed with EIGENSTRAT [140]. We defined people to be of European ancestry when they self-reported as European and when they had similar genetic ancestry based on principal-component structure [37]. Finally, we randomly removed a single person from all pairs of monozygotic twins.

We analyzed structural and resting-state functional MRI phenotypes from the HCP. All data were processed using the HCP minimal preprocessing pipeline [137] and were passed through a standardized quality control pipeline [138]. Structural MRI acquisitions were initially reviewed for image blurriness, motion, and other artifacts. Volumes of these regions were then estimated using FreeSurfer cortical and subcortical segmentations [61]. FreeSurfer-based reconstructions were inspected for obvious errors. Separately, resting-state functional MRI acquisitions were scored for 9 quality control measures that centered on the temporal signal-to-noise ratio, image smoothness, as well as the extent of absolute and relative head motion. The data were registered with MSM-All [141] and denoised with ICA-FIX [139].

In our study, we computed 3 functional MRI phenotypes on these data. First, we computed the amplitude of low-frequency fluctuations as the total power of spontaneous intra-regional activity within the 0.01 to 0.08 Hz range. Second, we computed regional homogeneity as the mean Pearson correlation between all pairs of intra-regional voxels. Third, we computed mean coactivity as the mean Pearson correlation between the activity of the region and all other regions of interest.

Individual variation in homogeneity and coactivity strongly correlated with individual variation of the global signal, the mean activity of all brain voxels. This variation can reflect artifact but also aspects of vigilance and non-neuronal physiology [142]. To focus on correlation structure unaffected by such properties, we computed these phenotypes after regressing out the global signal from voxel time series (for homogeneity) or regional time series (for coactivity).

We computed each phenotype separately for each scan of each person and then averaged the phenotypes across the 4 available scans and the left and right hemispheres.

Analyses of genetic ancestry, regional identity, and inter-regional correlation structure

We created principal component embeddings of ancestral and regional gr-expression for the ethnically diverse sample of the UK-Biobank cohort. We first constructed a 3D array of 5,984 people × 1,892 genes × 10 regions, where people comprised the UK-Biobank sample (S1 Table), and genes comprised all genes with available estimates of gr-expression in all the 10 GTEx regions. We then analyzed reshaped versions of this array. First, to extract ancestral structure, we computed principal components of the 5,984 × 18,920 matrix of brain-wide gr-expression across people. Second, to extract regional structure, we computed principal components of the 59,840 × 1,892 matrix of region-specific gr-expression across people.

We likewise compared inter-regional correlations of expression on subsets of genes that simultaneously had available expression in the UK-Biobank, GTEx, and Allen Human Brain Atlas data. These subsets ranged from 2,837 to 4,220 genes (median 3,642) and differed for each region of interest because each region had a distinct set of available gr-expression models. For all pairs of regions, we computed inter-regional coexpression using the subsets of genes that had expression data in both regions. Finally, we averaged the inter-regional coexpression matrices across all people in each data set.

To test for distance effects, we computed Spearman correlation coefficients between inter-regional coexpression and Euclidean distance between centroids of regions in the volume-based parcellation (Table 1). To test the effects of distance on the similarity of inter-regional coexpression, we generated 10,000 coexpression matrices with permuted ranks and empirical Spearman correlations with Euclidean distance.

Association of genetically regulated gene expression with neuroimaging phenotypes

Gene-set enrichment analyses

We performed gene-set enrichment analysis for biological annotations from the NHGRI-EBI GWAS catalog [75] (Fig 3E). Each phenotype in this catalog is situated within the Experimental Factor Ontology, a general ontology that includes terms from multiple more specialized ontologies and describes a wide range of measurements, including healthy and diseased phenotypes. All annotations reflect findings from curated GWAS analyses.

We used a semi-automated pipeline to detect brain-related terms in this ontology in 2 steps. In the first step, we flagged each term as brain-related if words in its ontology tree included at least one of the following word segments: nerv, neur, cogn, psyc, ment, brai. Second, 2 authors (NH and MR) manually and independently checked these candidate terms to confirm or exclude their brain-related nature. S3 Table lists the phenotypes that were enriched for TWAS genes at p_FDR < 0.05 and also lists their brain-relatedness indicator.

We performed gene-set enrichment analyses for clinical phenotypes in the BioVU TWAS, a database of associations between genetically regulated gene expression and clinical phenotypes derived from the Vanderbilt Biobank. For these analyses, we considered 70,439 people of European ancestry. Phenotypes were represented as phenotype codes based on ICD-9 codes [144]. We restricted our analyses to mental disorders and neurological phenotype code categories and included all gene-phenotype pairs that showed associations in the BioVU TWAS at p_FDR < 0.05.

We performed gene-set enrichment analyses using WebGestalt [145,146]. In both cases, we used a hypergeometric null model to test the enrichment of genes that had TWAS associations of p_FDR < 0.05 against a reference set of all genes in the TWAS (in other words, against all genes with relevant gr-expression models that passed baseline performance thresholds).

Interactive application

We developed a browser-based application for interactive analysis of our association results. This application is available on GitHub at https://github.com/nhunghoang/twas-webapp.