Abstract
The molecular determinants of tissue composition of the human brain remain largely unknown. Recent genome-wide association studies (GWAS) on this topic have had limited success due to methodological constraints. Here, we apply advanced whole-brain analyses on multi-shell diffusion imaging data and multivariate GWAS to two large scale imaging genetic datasets (UK Biobank and the Adolescent Brain Cognitive Development study) to identify and validate genetic association signals. We discovered 503 unique genetic loci that explained more than 50% of the average heritability across imaging features sensitive to tissue compartments. We identified key molecular pathways involved in axonal growth, astrocyte-mediated neuroinflammation, and synaptogenesis during development. Our results provide critical implications for potential targets for pharmacological intervention on neuropsychiatric outcomes.
Main
The human brain develops through complex yet carefully orchestrated neurobiological processes, whereby cortical and subcortical circuitries are integrated for proper functioning1. Neural migration, axonal guidance, and synapse formation are coordinated through spatially distributed molecular gradients spanning across several brain regions 2. Differences in tissue composition are the end result of these developmental processes. We can gain substantial insight into how neural circuitries were formed and supported by investigating the genetic determinants of whole brain patterning with respect to tissue composition.
Recent advances in multi-shell diffusion magnetic resonance imaging (ms-dMRI) and diffusion signal modeling have created an opportunity to evaluate tissue composition in vivo 3–7. Differences in diffusion signals between water molecules of intracellular, extracellular, and unhindered compartments are captured by ms-dMRI, allowing for the estimation of the relative proportions of cell bodies, axonal fibers, and interstitial fluids within a voxel 3–5,8–12. This imaging modality has been used to detect compositional changes driven by neurodegeneration 8,11, development 3, obesity 4, and carcinogenesis 9,10. However, there is currently no genome-wide association study (GWAS) on compositional features derived from ms-dMRI. This omission is critical, as traditional imaging measurements are insensitive to neurite density, short-range fibers, and cellular properties of cortical gray matter and subcortical nuclei7.
Moreover, GWAS of brain imaging measurements usually adopt a univariate approach, performing associations with one brain region at a time 13–18. Patterns encompassing the whole brain have been mostly ignored or controlled away as global effects, potentially biasing associations toward purely regional effects. This risks misattributing the nature of genetic effects on the brain, e.g., cortical surface area is driven by local cortical expansion when it may instead be due to underlying axonal growth. The univariate region-of-interest approach may also be underpowered to detect the full extent of genetic variants associated with canonical neurodevelopmental pathways, especially when effects are spatially distributed 19,20. A multivariate GWAS, focused on detecting loci that have effects across multiple brain regions, has been shown to be highly efficient in discovering many loci 21–23. Nevertheless, to gain biological insight, there is a need to extend these methods to enable meaningful interpretation of results.
Here, we performed a multivariate GWAS on the metrics derived from ms-dMRI to examine the genetic determinants of whole brain patterning of cellular compartments. Using two largest extant imaging genetic studies that have compatible ms-dMRI scans, the UK Biobank24 (UKB) and the Adolescent Brain Cognitive Development□ Study (ABCD Study®) 25,26, we identified 503 unique loci for tissue sensitive diffusion metrics. The discovered loci were enriched for neurogenesis, neuron differentiation, and axonal development. Among the replicated loci, 152 have not been reported previously by GWAS of brain imaging phenotypes. By investigating the spatial distribution of the associated effects, we identified molecular pathways involved in neuroinflammation and axonal growth. Signal overlaps at both the locus level and genome-wide with neuropsychiatric outcomes indicate the functional relevance of our GWAS results, providing grounds for further understanding the biological underpinnings of neuropsychiatric disorders and potential pharmacological targets.
Results
Multivariate GWAS on features of tissue composition across the whole brain
We processed ms-dMRI data from UKB and ABCD with restriction spectrum imaging (RSI) to extract the tissue composition features of the human brain 3–5,8–12,27. RSI decomposes the diffusion-weighted signals as emanating from three separable tissue compartments: intracellular, extracellular, and free water (Figure 1a). Each compartment is characterized by its intrinsic diffusion properties. In this study we consider the intracellular compartment, which is defined by restricted diffusion bounded by cellular membranes, and the free water compartment characterized by the unimpeded diffusion of water molecules. RSI estimates the normalized isotropic restricted signal volume fraction, N0, which captures the relative amount of cell bodies within a voxel, such as the densities of neurons, astrocytes, and oligodendrocytes. The normalized directional restricted signal volume fraction, ND, captures the relative amount of tube-like structures within a voxel, such as axons and dendrites. The free water component, NF, captures the relative amount of free water outside of cell structures.
After the images were harmonized and registered to a common atlas to ensure the alignment of each voxel across subjects (see Method for detailed imaging processing pipelines 25,26), we then performed separate voxel-wise multivariate GWAS on N0, ND, and NF. In the discovery stage (UKB discovery set, imaging acquisition before 2019, n = 23,543), we implemented a combined principal components GWAS (CPCs) to identify associated loci for each feature 22,28 (Figure 1b). Using the UKB discovery set, CPCs calculated the principal components (PCs) from the tissue feature across all voxels. From the whole-brain images in 2 mm resolution per voxel, spanning across 100 by 100 by 130 voxels, the first 5000 PCs were extracted and used in the subsequent analyses, explaining more than 70% of total variance of the imaging data (Extended Data Figure 1). Since all PCs are orthogonal to each other, the statistical inference can be based on combining the associations between genetic variants and each of the derived PCs (Figure 1b). We tuned the hyper-parameters for the combination function to optimize the power for discovery 19,22 by searching through four possible combination sets (see Methods). To account for hyper-parameter tuning and the three tissue features, we set the p-value threshold for genome-wide significance as 5e-8 divided by 12 = 4.2e-9.
After Linkage-disequilibrium pruning (LD R2 > 0.1) and positional clumping (distance < 250K bp), we found 432, 350, and 273 independent genetic loci associated with N0, ND, and NF, respectively (Figure 2a). Combined, there are 503 unique loci across all three tissue features. Of those significant loci, 40% are shared in the same genomic regions across all features whereas 34% are associated exclusively with one tissue feature (Figure 2b). The gene set analyses performed by Functional Mapping and Annotation (FUMA) 29 of the discovered loci shows each tissue feature has distinct pattern of Gene Ontology enrichment. While all tissue features were highly enriched for the Gene Ontology term of neurogenesis, N0 showed stronger enrichment in anatomical morphogenesis, while ND demonstrated more enriched in axon development, neuron projection guidance, and tangential neuronal migration (Figure 2c). This suggests that at the level of the genomic loci, modeling tissue compositions captured differential molecular effects associated with the human brain.
While the SNP heritability for each PC can be as high as 0.32 to 0.36 (Extended Data Figure 2), the multi-dimensional heritabilities 23 indicate the mean signal for N0, ND, and NF are 0.09 (95%CI 0.04 - 0.13), 0.06 (95%CI 0.02 - 0.10), and 0.05 (95%CI 0.01 - 0.09). Hence, the discovered loci explained 59%, 60%, and 58% of the average SNP-heritabilities for N0, ND, and NF.
Loci validated in adults and adolescents
To validate the discovered loci in independent studies, we first calculated imaging scores 30–33 based on eigenvectors and association weights from the discovery set, and then performed the association tests between genetic variants and the derived scores (see Methods). This procedure is similar to confirmatory canonical correlation analysis 23, except with only one variant involved in each regression. We repeated the same confirmatory analysis in the UKB replication set (n = 6,396, scanned after 2019) and ABCD samples (n = 8,189), except for including study-specific covariates and random effects controlling for family relatedness and diverse genetic background in ABCD (see Methods). Among the discovered loci, 335 (79%), 298 (85%), and 222 (81%) were replicated in the independent UKB replication set for N0, ND, and NF, after Bonferroni correction for number of loci discovered. In ABCD, 106 (25%), 153 (43%), and 88 (32%) of the discovered loci were replicated for N0, ND, and NF, despite the large differences in age and other sample characteristics between UKB and ABCD (Figure 2d).
To examine the overlap between our validated loci and previously reported loci in neuroimaging GWAS, we curated the reported loci lists from the NHGRI-EBI Catalog based on keywords in “brain”, “imaging”, “cortical”, “subcortical”, and “white matter”. The final list of reported loci included GWAS on brain connectivity 15, cortical surface measures 13,21,34, derived imaging instruments across all modalities 35, subcortical volumes 14,21,36, brain volumes 16,34,37, white matter hyperintensities 38, and white matter microstructure 18. We queried if any of our validated loci were in linkage-disequilibrium (LD) with or located in 250kb regions of previously reported neuroimaging loci. The results are summarized in Figure 2d. Among the replicated loci, 262 unique loci overlapped with previously published brain imaging GWAS, many of them found to be associated across several different imaging measurements (Figure 2d). This suggests widespread pleiotropy between the development of cortical surfaces and cerebral white matter, with molecular processes working across a spatial gradient rather than focalized in one single anatomical region.
Loci showing differential effects on tissue compositions
Closer inspection of the effect size distributions of the loci provides a unique angle into the molecular processes shaping the human brain. For instance, the 5q14.3 locus at the gene body of VCAN, tagged by a common SNP rs12653308, was found to be strongly associated with N0 (Figure 3a, Extended Data Table 1). It was reported to be associated with various diffusion metrics from white matter fiber tracts18 and cortical surface measurements21 (Figure 2a). Instead of fiber tracts or cortical surface regions, we found that the association strength is particularly strong in the hippocampus bilaterally (Figure 3b, 3c), based on the regional enrichment analysis with 50,000 bootstraps (see Method). VCAN encodes versican, which is a lectican-binding chondroitin sulfate proteoglycan (CSPG) and serves critical roles in astrocyte-mediated neuroinflammation39 (Figure 3d). CSPGs were found to be associated with astrocyte-dependent synaptogenesis within the hippocampus 40. When we examined the associations between genetic variants of genes encoding CSPGs (BCAN, NCAN, and VCAN) and tissue features, we found N0 showed stronger association signals than ND and NF (Figure 3e). Since the effects were replicated in ABCD, our results support the early effects of astrocytic mediated processes on the human hippocampus via CSPGs. Changes in the distribution of CSPGs in the hippocampal formation were observed among patients with Schizophrenia and patients with Bipolar disorders 41,42, linking our findings to neuropsychiatric outcomes. The Drug-Gene Interaction database (DGIdb) shows VCAN as tier one druggable target and was found to be interacting with cyclosporine 43,44, indicating a potential path for pharmacological interventions.
The locus located at 2p23.3, tagged by rs11126784, has strong signals associated with ND (Figure 3f). This locus resides within the gene body of DPYSL5 and has been reported to be associated with cortical surface measures 21. Instead of cortical surface, our whole-brain multivariate GWAS indicates the effect sizes were more diffusely distributed among white matter tracts, especially within cortico-striatal circuitry (Figure 3g, 3h). DPYSL5 belongs to the collapsin response mediator protein (CRMP) family, including DPYSL2, DPYSL3, and DPYSL4, which are essential for axonal growth and neurite morphogenesis 45–47 (Figure 3i). Indeed, all tagged SNPs of the CRMP family proteins show stronger association signals with ND than with N0 and NF (Figure 3j). Our results are concordant with CRMP involvement in neurodevelopment and showing that their effects can be observable among major white matter fiber bundles early on. Our findings are also relevant to neuropsychiatric outcomes, as CRMP has been implicated in schizophrenia and mood disorders, and can be pharmacologically intervened 48.
The 152 novel loci we discovered and replicated in this study are relevant for neuropsychiatric phenotypes and warrant further investigation (Extended Data Table 1-3). An N0-specific novel locus at 5q14.3 is within the gene body of MEF2C, which can influence neural progenitor cell differentiation and regulation of synaptic densities 49,50. This locus overlaps with GWAS findings of educational attainment and intelligence 51. Another locus at 20p12.1, on the gene body of MACROD2, showed consistent signals among adults and adolescents (ND: UKB discovery p = 1e-29, UKB replication p = 1.7e-18, and ABCD replication p = 4.6e-8), and has previously been linked to autism 52 and general cognitive ability 53. The gene MACROD2 was also implicated in educational attainment 51 and risk-taking behaviors 54.
Cell-type enrichment analysis
Although N0, ND, and NF were designed to capture different properties of tissue compartments, the strong overlapping signals across the three features indicates that similar cell processes and populations may shape all three microstructural features. To investigate this, we analyzed the heritability enrichment given cell type annotations using stratified LD score regression (LDSC) 55. A dimensionally-corrected multivariate statistic, such as the scaled χ2, can be used in the context of LDSC for deriving the relative enrichment in the average heritability of the high-dimensional phenotypes 23. Hence, we ran stratified LDSC with tissue-specific chromatin annotations 55 and cell type-specific annotations 56 to obtain cell type-specific enrichment patterns for our RSI phenotypes (Figure 4).
While the overall patterns of the enrichment are similar across three tissue features, ND has the strongest enrichment signals across all activating histone markers (H3K27ac, H3K36me3, H3K4me1, H3K4me3, and H3K9ac) and DNase hypersensitivity sites (Pbon < 0.05). All three features were enriched in the chromatin state of fetal brain and hippocampal tissues whereas ND also shows enrichment in the cingulate cortex and substantia nigra (Figure 4a). With respect to cell populations, cell type-specific chromatin state analysis indicates all three features have significant enrichment in embryonic dopaminergic interneurons and astrocytes (Pbon < 0.05). Moreover, ND shows stronger enrichment signals in oligodendrocytes, as expected for an imaging feature capturing the integrity of tubular structures such as the myelin sheath.
Genetic overlap with neuropsychiatric and immune-related phenotypes
We investigated the proportion of genome-wide signals of the three tissue features which overlap with neuropsychiatric phenotypes 51,54,57–63 and immune disorders 64. Based on a method tailored for unsigned multivariate statistics 23, we evaluated the signal shared between each pair of traits using their summary statistics. The amount of shared signal was the Spearman correlation of the average SNP −log10 p values within each approximately independent LD block. All three tissue features consistently show significant overlap with immune disorders 64, schizophrenia 61, attention deficit hyperactivity disorder 65, bipolar disorder 59, cross-psychiatric-disorders 57, and Alzheimer’s disease 63 (ρ: 0.10 ~ 0.29; all Pbon < 0.05). Educational attainment 51 and risk-related behaviors 54 are also significantly correlated, indicating their relevance to general cognitive features (Extended Data Figure 4). Nevertheless, the patterns of genome-wide signals shared with neuropsychiatric phenotypes were not evidently different across three tissue features, despite the distinct patterns we observed at the locus level and cell-type specific enrichments. While the limited resolutions by the LD block can contribute to this lack of differences, the evident similarities in the genome-wide level can mean that the pleiotropic effects on neurodevelopmental traits are highly polygenic, sharing multiple loci but with different functional outputs.
Discussion
Using imaging features of whole brain tissue compositions, a multivariate GWAS discovered and validated 503 loci, of which 152 had not been reported in previous GWAS of neuroimaging phenotypes. Through in-depth examination of effect size distributions, we demonstrated the specific impact of molecular pathways, including CSPGs and CRMP, on the tissue composition underlying the human brain in vivo. Our findings are highly relevant for neuropsychiatric outcomes, including cognitive functions and psychiatric disorders. By identifying the key protein families and highlighting the susceptible brain regions through enrichment analyses, these results indicate a path to further investigate a potential target for pharmacological interventions.
In addition to these biological insights, our findings also showcase the need for novel analytic approaches in brain imaging genetics. Multivariate GWAS on whole brain phenotypes circumvents the potential “spotlight bias” that region-of-interest approaches are susceptible to. Diffuse effects across brain regions and neurobiological pathways are more easily detected with this approach, as the inference is based on the total sum of the effects. Detecting genetic loci associated with axonal growth, astrocyte-mediated neuroinflammation, and changes in the synaptic density require a more nuanced modeling approach of imaging signals. Moving beyond the metrics of structural volumes or fiber orientation enabled us to detect molecular effects on brain tissue properties, highlighting relevant biological pathways important for human brain development and neuropsychiatric outcomes.
Methods
UK Biobank samples
The inclusion criteria for the UKB sample were as follows: individuals who had valid consent at the time the analyses were performed (Dec 2020), were genetically inferred as having European ancestry, and completed the neuroimaging protocols. Among individuals who were included in the analyses, we further divided samples into two groups based on when the neuroimaging was performed (before or after 2019). We decided to use this naturally occurring cut-point instead of randomized allotment of the groups because of best practice considerations, avoiding potential systematic biases driven by temporally related imaging confounds. Individuals who had valid imaging data before 2019 were assigned as the discovery set (n = 23,543) and those who had valid imaging data, not before, but after 2019 were assigned to the validation set (n = 6,396). The demographic information of the final selected UKB samples can be found in the Extended Data Table 4. Data from UKB is obtained under accession number 27412.
Adolescent Brain Cognitive Development study (ABCD) samples
For replication of results, we selected the full baseline data of the ABCD Study from public data release 3.0 (NDA DOI:10.151.54/1519007). Since ABCD was designed to recruit individuals with diverse ancestral background which reflect the racial/ethnic composition of the United States, we did not exclude individuals based on their genetic ancestries, using linear mixed effects models to control for the family relatedness and heterogeneous ancestral background. We only excluded those who did not have valid imaging and genetic data from release 3.0, resulting in 8,189 individuals in the analyses. The demographic characteristics of the ABCD samples can be found in the Extended Data Table 4.
Imaging data processing
Both UKB and ABCD have diffusion imaging protocols that were compatible for applying RSI models. The MRI scans of UKB were performed at three scanning sites in the United Kingdom, all on identically configured Siemens Skyra 3T scanners, with 32-channel receive head coils. The MRI scans of ABCD were collected by 21 study sites throughout the United States, with scanners from Siemens Prisma, GE 750 and Phillips 3T scanners. To harmonize the imaging data across the two studies, we processed the dMRI data from UKB and ABCD using the ABCD-consistent imaging processing pipeline implemented by the ABCD Data Analysis, Informatics, and Resource Center (ABCD DAIRC). The detailed processing procedures have been published elsewhere25. In short, multi-shell dMRI data acquired with seven b=0 s/mm2 frames and 96 non-collinear gradient directions, with 6 directions at b=500 s/mm2, 15 directions at b=1000 s/mm2, 15 directions at b=2000 s/mm2, and 60 directions at b=3000 s/mm2. Ms-dMRI data were first processed through forward-reverse gradient warping, eddy current correction, and motion correction to reduce the spatial distortion and signal heterogeneities driven by scanner differences. The corrected images were then aligned to a common atlas using rigid body registration 25. The fiber orientation density (FOD) functions were then calculated for each voxel and then the tensor information was fed into multi-channel nonlinear smoothing spline registration, resulting in positional and orientational aligned voxel-wise diffusion data in 2mm resolution.
Restriction spectrum imaging (RSI) models the diffusion signals as mixtures of spherical harmonic basis functions 5,12. Based on the intrinsic diffusion characteristics of separable pools of water in the human brain (i.e. intracellular, extracellular, and unhindered free water), RSI estimates the signal volume fractions of each compartment and their corresponding spherical harmonic coefficients. The measure of restricted isotropic diffusion (N0) is the coefficient of the zeroth order spherical harmonic coefficient, normalized by the Euclidian norm of all spherical harmonics. This feature is most sensitive to isotropically diffusing water in the restricted compartment, within cell bodies. The measure of restricted directional diffusion (ND) is the sum of second and fourth order spherical harmonic coefficients, normalized by the norm of all spherical harmonics. This feature is sensitive to anisotropically diffusing water in the restricted compartment, within oriented structures such as axons and dendrites. The normalized free water diffusion (NF) measure is calculated as the zeroth order spherical harmonic coefficients for the unhindered water compartment. NF is also normalized by the Euclidean norm of all SH coefficients. This normalization makes the RSI features unitless and in the range of 0 to 1. N0, ND, and NF provide greater tissue specificity than the widely-used diffusion tensor metrics and are particularly useful for understanding the variations of cellular organization within the human brain and highly informative for human brain development 3,10,11,27.
Genotype data processing
For UKB, we used the released v3 imputed genotype data. For ABCD, we used the public release 3.0 imputed genotype data. Both datasets were imputed with the HRC reference panel 66. We performed post-imputation quality control to only allow for GWAS on common bi-allelic SNPs. We filtered SNPs which have minor allele frequencies less than 0.5 percent, Hardy-Weinberg disequilibrium (p < 1e-10), and missingness greater than 5 percent. Genetic principal components and ancestral factors were derived using well-called independent SNPs for both datasets and were used for controlling population stratification in our analyses.
Combined principal component GWAS (CPC)
We implemented CPC for our multivariate GWAS on our UKB discovery set. CPC has been shown to be a robust multivariate GWAS method that is well powered to detect loci across different scenarios 19,22,28. The procedures were as follows. First, the principal components and their corresponding eigenvectors were derived given the voxel-wise imaging data. Each SNP was regressed on each of the derived principal component scores, controlled for age, sex, 20 genetic principal components, and genotyping batches. For a given SNP, the Wald statistics for each principal component were combined as a simple linear sum (Figure 1B). Given that PCs are orthonormal, the sum of the Wald statistics follows the χ2 distribution with k degrees of freedom for k PCs combined. Although several different combination functions can be used 19, we found that the global-local combination with Fisher’s method proposed in the original CPC paper has greatest power in detecting genetic loci 22. Therefore, we experimented with four different global-local cut points (50, 100, 500, and 1000 principal components) to see which combinations yield the most discoveries. To reflect this experiment, we lowered the significance threshold to p <4.2e-9 (corrected for 12 multiple comparisons, as 4 thresholds and 3 features were used in the current study).
Replication with confirmatory imaging scoring
To perform the replication test for the discovered loci, we used the confirmatory imaging scoring instead of repeating the GWAS on the independent cohorts. The eigenvectors (νk) and the regression coefficients (βk) obtained from the discovery set were used to calculate the imaging scoring for all subjects in the validation sets.
x stands for the raw imaging data. Given that each PC is independent of the other, it can easily be shown that the SNP regression on the imaging score is equivalent to the comparison of the consistencies of regression coefficients between the discovery set and validation set (Supplemental Materials).
Regional enrichment for spatial distribution across voxels
To provide more interpretability for the multivariate GWAS results, we developed a regional enrichment analysis to show which brain regions have relatively stronger signals. Most previous imaging studies relied on re-doing the voxelwise association tests to show the effect distributions of the discovered loci 14,16–18,36. Given the distributed nature of the effect sizes among imaging measurements, the voxel-wise associations were not an ideal way of localizing effects 20. Instead, we examined the overlap between association patterns and regions of interest in the co-registered anatomical atlas. The enrichment score is the probability-weighted regression coefficients from CPC:
The variance of the enrichment score was estimated by bootstrapping the association patterns from SNPs that did not surpass the significance threshold. We then calculate the corresponding enrichment z-score and the corresponding p-values. In the current study, we obtained 130 probability maps of brain regions defined in the common atlas. We applied the regional enrichment analyses on the loci that showed robust signals across adult and adolescent data.
Loci annotations, overlaps, and gene-set enrichment analyses
To annotate the identified genetic loci, we used FUMA 29 and the GRanges function in R. SNPs with LD of r2<0.1 and within 250kb distance were considered as one single locus. MAGMA 67 was used for calculating the gene-set enrichment.
Calculation of high dimensional heritability
Previous studies on the heritability of high-dimensional phenotypes indicated the average heritability is a valid way of estimating the genetic architecture of human traits 23,68. It is equivalent to the weighted average of heritabilities across each of the principal components. We applied LD score regression for each principal component and then weighted these according to their eigenvalues, deriving the average heritabilities across RSI features.
Stratified LD score regression for cell-type specific heritability enrichment
We rescaled the χ2 according to extended methods 23 and then used stratified LDSC to examine the relative enrichment of heritability for cell type-specific annotations.
Calculation of shared genome-wide signals between two phenotypes
As proposed in other multivariate GWAS efforts23, for a given summary statistics of a phenotype, we first calculated the average magnitudes of associations in each of the approximately independent LD blocks 69, deriving the unsigned polygenic signal profiles of a given trait. Spearman correlations were performed for each pair of the GWAS results, evaluating the level of overlapping in the genome-wide signals.
Author Contributions
CCF, AMD, and OA conceptualized and designed the study. CCF and RL performed the analyses. DJH and OF processed the data. CCF interpreted the results and wrote the manuscript. All other co-authors provide critical inputs for the revision of manuscripts.
Conflict of Interest
Dr. Andreassen has received speaker’s honorarium from Lundbeck, and is a consultant to HealthLytix. Dr. Dale is a Founder of and holds equity in CorTechs Labs, Inc, and serves on its Scientific Advisory Board. He is a member of the Scientific Advisory Board of Human Longevity, Inc. and receives funding through research agreements with General Electric Healthcare and Medtronic, Inc. The terms of these arrangements have been reviewed and approved by UCSD in accordance with its conflict of interest policies. The other authors declare no competing interests.
Acknowledgement
This work was supported by grant R01MH122688 and RF1MH120025 funded by the National Institute for Mental Health (NIMH). Data used in the preparation of this article were obtained from the Adolescent Brain Cognitive Development□ Study (ABCD Study®) (https://abcdstudy.org), held in the NIMH Data Archive (NDA). The ABCD Study is supported by the National Institutes of Health and additional federal partners under award numbers: U01DA041022, U01DA041028, U01DA041048, U01DA041089, U01DA041106, U01DA041117, U01DA041120, U01DA041134, U01DA041148, U01DA041156, U01DA041174, U24DA041123, and U24DA041147.
Footnotes
Refined loci selection and the corresponding relevancy to biological pathways. The focus of the results was shift to make it clear why modeling tissue compositions matters.