Mapping causal pathways from genetics to neuropsychiatric disorders using genome-wide imaging genetics: current status and future directions

Abstract

Imaging genetics aims to identify genetic variants associated with the structure and function of the human brain. Recently, collaborative consortia have been successful in this goal, identifying and replicating common genetic variants influencing gross human brain structure, as measured through MRI. In this review, we contextualize imaging genetic associations as one important link in understanding the causal chain from genetic variant to increased risk for neuropsychiatric disorders. We provide examples in other fields of how identifying genetic variant associations to disease and multiple phenotypes along the causal chain has revealed a mechanistic understanding of disease risk, with implications for how imaging genetics can be similarly applied. We discuss current findings in the imaging genetics research domain, including that common genetic variants can have a slightly larger effect on brain structure than on risk for disorders like schizophrenia, indicating a somewhat simpler genetic architecture. Also, gross brain structure measurements share a genetic basis with some, but not all, neuropsychiatric disorders, invalidating the previously held belief that they are broad endophenotypes, yet pinpointing brain regions likely involved in pathology of specific disorders. Finally, we suggest that in order to build a more detailed mechanistic understanding of the effects of genetic variants on the brain, future directions in imaging genetics research will require observations of cellular and synaptic structure in specific brain regions beyond the resolution of MRI. We expect that integrating genetic associations at biological levels from synapse to sulcus will reveal specific causal pathways impacting risk for neuropsychiatric disorders.

Introduction:

The human genome is composed of approximately three billion base pairs. Between any two humans, the sequence of the genome is different at approximately four million locations¹. Those genetic variations, in aggregate, have a strong impact on inter-individual variability for almost all well measured traits, including brain structure, personality, and risk for neuropsychiatric disorders². Most individual variants have an infinitesimally small effect on any of these traits. Based on recent technological advancements reducing the cost of microarray technology and DNA sequencing³, as well as the collaborative efforts of multiple large consortia^4,5, genome-wide association studies (GWAS), whole exome sequencing, and whole genome sequencing studies have identified many loci in the genome that, when modified, impact measurable changes in brain or behavioral traits^6–10. Neuropsychiatric genetics have reached the point where the wheat (risk alleles) can be separated from the chaff (alleles without a detectable association).

The identification of these genetic loci offers considerable promise because it represents the first knowledge of the causal basis of behaviorally defined neuropsychiatric disorders for which we do not understand pathophysiology¹¹. Because environmental insults like medication or stressful life events do not modify the sequence of the genome at specific loci, and genetic variation is identical in every cell in the body from conception (though small exceptions exist¹²), we make the claim that these loci are causally involved in disorder risk. Risk-associated alleles have a unidirectional, although not a direct, effect on risk for these disorders. The causal basis of these genetic findings is in stark contrast to years of case-control studies in neuropsychiatry measuring differences across almost every measurable trait^13–16 that could be either causal or reactive¹⁷.

While exciting, such genetic associations are only a launching pad to reveal biological mechanisms underlying complex disorders. After all, genetic variation itself does not directly cause the altered behavior observed in patients with neuropsychiatric disorders. Instead, genetic variation impacts multiple levels of biology, at certain time points during development, within certain cell-types in the brain, changing development and structure, affecting brain function, and resulting in the behavioral manifestations of the disorder. Thorough characterization of these causal chains is a critical step for developing rational therapeutics. Already, genetics-inspired therapeutic design is underway for many disorders^18–20, and beginning for neuropsychiatric disorders²¹.

Maps of genetic variation impacting multiple levels of biology, termed “quantitative trait loci” or QTLs, allow inference of causal chains leading to risk for neuropsychiatric disorders. The specific phenotypes that are impacted along a causal mechanistic pathway bridging genetic variation to behavioral outcomes are often referred to in psychiatric literature as endophenotypes, and in genetics literature as links in a causal chain, or pathways²². At the beginning of a causal pathway, genetic risk loci may affect the regulation of gene expression within a given cell type at a given developmental time period. A single nucleotide polymorphism (SNP) may lead to a new transcription factor binding site and allelic differences in chromatin accessibility, called chromatin accessibility QTLs (caQTLs)²³. When genetic variation impacts gene expression, sometimes through alterations in chromatin accessibility, these loci are referred to as expression QTLs (eQTLs)²⁴. Large scale maps of genetic variation impacting molecular measures, like caQTLs and eQTLs, have been and are continuing to be developed to allow the identification of regulatory elements or genes impacted by genetic variants associated with neuropsychiatric disorders^25–28 (Figure 1). Similarly, maps of loci associated with gross brain structure and function derived from MRI have been used for almost two decades to suggest brain regions or functions within the causal chains leading to a variety of neuropsychiatric disorders²⁹.

Figure 1:

From genetic association to mechanism. Studies seeking to define causal pathways between genetic risk and the manifestation of neuropsychiatric disorders may employ a generalized three-step approach. In step 1, high-powered genetic association studies are used to identify variants associated with risk for a disorder. In step 2, genetic association with endophenotypes (chromatin accessibility, gene expression, and brain structure) are used to infer causal pathways leading to risk for a disorder. In step 3, experimental manipulations in human or animal model systems are used to validate mechanistic hypotheses.

The use of gross brain structure and function as links in the mechanistic chain is fueled by two main assumptions - that genetic variation has a stronger impact on brain traits than on heterogeneous and behaviorally defined disease categories, and that genetic variation associated with brain changes will allow a greater understanding of the mechanism leading to risk for behaviorally defined disorders³⁰. How well are these assumptions met by MRI measures of brain structure and function? How have genetic associations to brain structure and function informed our understanding of the causal chain leading to risk for neuropsychiatric disorders? How can we leverage new imaging methods for a deeper understanding of genetic risk factors for neuropsychiatric disorders?

In this review, we seek to 1) unify the concepts of endophenotypes from psychiatric literature and causal chains/pathways from genetics literature to understand causal mechanistic pathways leading from genetic variation to risk for neuropsychiatric disorders, 2) present detailed molecular pathways impacting risk for disorders discovered from association studies largely outside the realm of neuropsychiatry to inform the study of neuropsychiatric disorders and identify where imaging associations can be leveraged, 3) discuss for which disorders there is evidence that gross brain structure, measured through MRI, lies along a causal genetic pathway, and 4) discuss the application of higher resolution imaging of post-mortem brain tissue to reveal cellular and synaptic phenotypes and refine causal hypotheses derived from genetic associations to MRI measures.

The history and use of endophenotypes in psychiatric literature

Gottesman and Shields adapted the term “endophenotype” to describe internal phenotypes for schizophrenia that were detectable through biochemical tests or microscopy and were ostensibly caused by genetic variation^31,32. They hoped that studying schizophrenia endophenotypes would side-step the challenging phenotypic heterogeneity found at the clinical level to instead focus on molecular pathways that mediated aspects of the disorder. Brain structure and function as measured through MRI were particularly thought of as excellent endophenotypes²⁹. Following this prescient idea, there are currently two predominant uses of endophenotypes in neuropsychiatric literature:

• (1)To identify genetic variation associated with the endophenotype, rather than a heterogeneous, clinically defined disorder, because the endophenotype is “closer to the underlying biology”, increasing the power of genetic search while still being informative about disorder risk.
• (2)To provide mechanistic connections linking genetic variation to behavioral manifestations of neuropsychiatric disorders.

To achieve these goals, there are several assumptions that must be satisfied. First, to achieve higher powered search, genetic variants must have a stronger effect (a higher effect size of the association) on the endophenotype than the disorder. Phenotypes that follow this assumption are often referred to as “closer to the underlying biology” or “closer to genetics” and are detectable within smaller sample sizes^22,33–35. Second, to provide mechanistic insight, genetic variants impacting the endophenotype must lead to risk for the disorder by way of a causal chain²². Such causation is captured by a mediational model²², whereby neuropsychiatric disorder risk alleles impact brain structure and function and in turn drive the development of the disorder. In contrast, a liability index model does not provide a mechanistic understanding because a risk allele independently impacts both brain structure and risk for a neuropsychiatric disorder via pleiotropic effects²² (Figure 2). We note that “endophenotype” and “intermediate phenotype” terms are often used without precise definition³⁶, often failing to differentiate mediational models from liability index models of genetic risk or incorrectly implying that there is only one link on a causal chain between genetic variant and disorder^22,35,37. The genetics community instead uses the term “causal pathway” to describe the impact of genetic variation along multiple phenotypes leading to disease risk. We prefer the specificity of this language and continue to use it here.

Figure 2:

Different models that could explain an observed genetic correlation. In the mediational model, genetic variation causes a change in brain structure which then leads to risk for a disorder. If this model is true, experimental manipulation of brain structure will alter risk for the disorder. (Imagine wiggling brain structure and observing that disorder is also wiggling). In the pleiotropy, or liability-index model, genetic variation causes changes in both brain structure and neuropsychiatric disorders, but they are not causally linked. If this model is true, experimental manipulation of a brain structure phenotype does not impact risk for a disorder. (Imagine wiggling brain structure and observing that disorder is staying still). Note that we have simplified this graph to only include one phenotype between genetic variation and a disorder, although we expect that many phenotypes will be found in the true causal pathway.

Despite the behavioral definitions of neuropsychiatric disorders and their phenotypic heterogeneity, genetic search has already identified many risk loci by compiling data from tens of thousands of individuals^38–40. This somewhat obviates the proposed use of endophenotypes as a tool for identifying genetic risk loci for neuropsychiatric illness with higher power (use 1, above). Rather, the most pressing need is to unravel the complex etiology of neuropsychiatric disorders by mapping the underlying mechanisms of risk alleles (use 2 above). Below, we discuss the current evidence about how well gross brain structure satisfies both of these criteria.

How genetic variation has led to identification of causal pathways impacting risk for complex disorders

To illustrate the promise of elucidating causal mechanistic pathways that underlie complex disorders using loci identified from GWASs, we present three examples where considerable causal mechanistic understanding has been achieved in order to motivate similar studies in the context of neuropsychiatric disorders. While two of these examples come from outside the field of psychiatry, their general framework could be readily applied to understand disorders impacting neural tissues. Generally, these studies begin with a common genetic variant of small effect, generally in non-coding and poorly annotated regions of the genome, which is used to identify phenotypes at the molecular, cellular, and systems levels that mediate risk for complex disorders including obesity, type 2 diabetes, and schizophrenia.

A causal pathway by which genetic variation impacts risk for obesity

One of the first large GWASs for body-mass index identified a locus within an intron of the FTO gene impacting risk for obesity⁴¹. This locus overlapped a non-coding region so it is unlikely to directly affect protein structure, as is true for most GWAS-identified loci⁴². Instead, given additional functional information describing the activity of the locus in different tissue types⁴³, the locus likely functions as a regulatory element, serving to alter the expression of a proximal gene⁴⁴. But what gene(s) does this regulatory element alter? The non-coding variant of interest was mapped to a gene that it regulates using an eQTL approach, where genotypes in a resource of genetically diverse samples were related to gene expression derived from the tissue of interest. Risk alleles at this locus were associated with increased expression of two genes, IRX3 and IRX5, relatively far (~0.5 and 1.1 Mb, respectively) from the BMI-associated locus in human adipose progenitors⁴⁴. Notably, proximity on the genome was insufficient to determine the gene of action, as is often the case⁴⁵. Does the risk allele impact the morphology of adipocytes, or fat cells? Again, using a genetically diverse sample derived from multiple individuals with human adipocytes it was found that individuals carrying the risk allele had larger adipocytes than those with the non-risk allele⁴⁴. Finally, does modification of IRX3 or IRX5 expression lead to changes in metabolism and body weight? Modification of expression levels of these two genes in both human adipose cells and genetically manipulated mice was found to impact metabolism and body weight⁴⁴. These findings demonstrate validation of a largely complete causal chain that mapped a non-coding obesity-associated locus to specific biological pathways influencing human obesity.

A causal pathway by which genetic variation impacts risk for type 2 diabetes

In another example, large GWASs have identified many loci associated with risk for type 2 diabetes⁴⁶. One locus of interest again overlapped a non-coding region near KLF14 with evidence of regulatory potential in adipose cells⁴³. Which gene(s) might this regulatory element regulate? Again, when genotypes, methylation of the DNA, and gene expression were acquired from a large sample of genetically diverse human tissue samples, the risk allele for type 2 diabetes was associated with increased methylation at the enhancer region and decreased expression of the closest gene, KLF14, specifically in adipose tissue⁴⁷. Do carriers of the risk allele exhibit altered morphology of adipocytes? Microscopy of these cells in a genetically diverse sample showed that carriers of risk alleles at this variant have an increased adipocyte volume and area, again demonstrating a morphological consequence to genetic variation. Do carriers of the risk allele also have increased risk of other phenotypes associated with type 2 diabetes? Through integration with other GWASs, risk alleles at this locus were also found to be associated with increased fasting insulin, decreased high-density lipoprotein cholesterol, and increased triglycerides⁴⁷. Finally, does modulation of the expression of KLF14 impact risk for type 2 diabetes? Mice harboring a conditional knockout of Klf14 in adipose tissue recapitulated insulin resistance, decreased high-density lipoprotein cholesterol, and increased triglyceride phenotypes, experimentally validating the observed genetic associations⁴⁷. This again demonstrates a causal chain describing the mechanism for a common variant locus impacting risk for type 2 diabetes.

A causal pathway by which genetic variation impacts risk for schizophrenia

Within the realm of neuropsychiatric disorders, there has been great success in identifying common variants impacting risk for multiple disorders^5,48, yet very few examples explicitly connect genetic loci to causal pathways. This process is significantly more difficult in brain tissues where the specific cell-types, developmental time periods, or brain regions leading to risk for altered behavior are often not known⁴⁹. Despite this, progress has been made in several disorders, and notably, over 100 loci have been detected that are associated with risk for schizophrenia^10,39. Genome-wide association for schizophrenia risk detected the locus of strongest association in a large, multigenic region comprising the major histocompatibility complex (MHC) locus¹⁰. Fine-mapping of the MHC locus revealed that the schizophrenia-associated single base pair risk alleles were tagging (or correlated with) a structural variant that increased the copy number of genes C4A and C4B⁵⁰. Do these extra copies of C4A and C4B present in the genome influence how much that gene is expressed? An eQTL study in post-mortem human brain tissue demonstrated that increasing copy number of C4A and C4B was associated with increased expression of each of these genes⁵⁰. Interestingly, C4 acts within the complement cascade involved in synaptic pruning, a process thought to be aberrant in schizophrenia^51,52. Such synaptic pruning deficits may explain neuroimaging findings showing an accelerated loss of frontal grey matter in schizophrenia patients⁵³. Does modulation of C4 expression impact synaptic pruning? Mice lacking copies of C4 showed fewer synaptic inputs in a visual circuit that normally undergoes synaptic pruning⁵⁰, experimentally validating the functional impact of these associations. Adding further causal evidence, a recent study replicated this pruning effect in human cells using patient-derived human neural cultures, and even showed that this mechanism of pathology can be targeted therapeutically⁵⁴. These studies^50,54 explain a small, yet important part of the causal pathophysiology, including cell-types and biological processes that lead to risk for schizophrenia⁵⁵, and are particularly exciting as until this point there had been no causal pathophysiology identified for this complex disorder.

The role human neuroimaging of gross brain structure can play in explaining causal pathways

The studies outlined above, and other similar studies not discussed^56–58, describe how genetic loci have led to a new understanding of the etiology of complex traits and share commonalities in design (Figure 1). First, genetic variation has served as an important causal anchor to begin understanding the mechanism leading to complex phenotypes like obesity, type 2 diabetes, or schizophrenia. Second, maps of how genetic variation relates to multiple phenotypes, in multiple tissues, and at multiple developmental time periods allow an inference of the causal chain leading to risk for a disorder. And third, experimental manipulation of genes within model systems, via gene editing in both human cell culture and mice test the causal predictions generated from the integration of genetic association maps.

An effective starting point for conducting these investigations is measuring the impacts of genetic variation associated with a neuropsychiatric disorder on gene regulation. While caQTLs and eQTLs provide valuable mechanistic evidence proximal to the source of genetic risk, a list of genes impacted by a risk locus is insufficient to describe a complete causal chain that results in behavioral abnormalities. Colocalizing genetic risk loci with QTLs at higher biological levels, like cell morphology and brain structure, allows description of more complete causal pathways. In the example described above concerning a genetic risk locus for obesity, identification of eQTLs was supplemented with observations that the risk locus induced aberrant adipocyte size⁴⁴. Together, these lines of evidence explained a mechanism for developing obesity where changes in gene expression shifted a key cell differentiation pathway to fundamentally alter lipid metabolism. We envision that genetic risk loci associated with brain traits can be similarly leveraged to characterize causal pathways in neuropsychiatric disorders at higher biological levels, and that causal hypotheses can be strengthened when this information is integrated with maps of gene regulation QTLs.

Identification of genetic risk loci for neuropsychiatric disorders is accelerating based in large part on the collaborative efforts of psychiatric genetics consortia^5,59. Mapping how genetic variation impacts multiple levels of biology requires measuring multiple relevant phenotypes (cell or tissue specific chromatin accessibility, cell or tissue specific gene expression, and brain structure) in large, genetically diverse populations. Such maps are being created for chromatin accessibility and gene expression in large samples of post-mortem and stem-cell derived brain cells^25–27. Similar maps of genetic influences on neuroimaging measures serve as another crucial tool for identifying and characterizing links on causal pathways. They allow interpretation of whether genetic variants associated with neuropsychiatric illness are also associated with the structure or function of the human brain within specific regions. Finally, directed stem cell differentiation protocols producing specific cell types from multiple brain regions and modern genetic engineering techniques allow the experimental validation of predicted causal pathways in a controlled human system^60–64. This review seeks to place maps of genetic variation associated with gross brain structure into context and explain their utility for describing causal pathways impacting risk for neuropsychiatric disorders.

Genetic associations to neuroimaging traits

We note that in our description of the literature going forward, we avoid discussing “candidate gene” studies, despite the fact that they represent the majority of the neuroimaging genetics literature. In a candidate gene study, a researcher selects a limited number of genes of interest and variants near or within these genes, then measures genetic association with phenotypes of interest (brain structure or function measured via MRI) collected in generally small sample sizes (<1000 subjects) without performing strict multiple comparisons correction across all independent variants present in the genome. The premise of these studies is that using our knowledge of genes involved in brain structure, function, and development, we are able to identify individual variants that are likely to impact the brain or risk for neuropsychiatric disorders. To summarize over ten years of work in this field, only an exceedingly small number of those “candidate gene” associations are replicated in much larger sample sizes^65–67. For example, of 32 candidate gene associations to brain structure or function, zero survived genome-wide significant association levels (P<5×10⁻⁸) to any subcortical structure or intracranial volume in sample sizes often 100 times larger than the candidate gene association sample size⁸. We, as scientists, are remarkably poor at guessing which genetic variants impact any trait, including brain structure. In response, consortia have assembled in order to accumulate large numbers of research participants to gain enough statistical power to identify genetic variants impacting many well measurable traits without pre-selection and with strict statistical thresholds^5,68,69. The Enhancing Neuroimaging Genetics through Meta-Analysis (ENIGMA) consortium⁴, the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) consortium⁷⁰, and the UK Biobank⁷¹ represent the largest published association studies for neuroimaging traits to date. Strict genome-wide significance thresholds and large sample sizes have led to highly replicable associations not just in brain traits but across all genetic search literature⁷² (see Table 1 for good practices). In imaging genetics, pushback to the ideas of meta-analysis, universally applied strict statistical thresholds, and requirement for replication persists⁷³. This resistance stems from an assumption that heterogeneity induced from meta-analysis can dilute or alter effects. Arguing against this point, a recent genetic correlation study measured whether genetic effects are shared between two genome-wide association analyses: the meta-analytic combination of many sites via the ENIGMA consortium and one UK Biobank site for global surface area and thickness⁷⁴. The genetic correlations approach one, indicating an almost complete shared genetic basis between meta-analysis and one-site analysis⁷⁴. A lack of strict statistical thresholds or requirements for replication have led to many publications without reproducible associations^75–77. We make the assertion that brain imaging traits are not special; hence, genetic associations to brain imaging traits are subject to the same statistical thresholds and replication criteria as all other traits. With this in mind, here we focus on imaging genetic findings from large sample sizes reported by consortia. With high confidence consortium-based associations, we discuss how well neuroimaging traits satisfy the previously described criteria of endophenotypes.

Table 1:

Good practices for imaging genetics studies of gross brain structure.

Good practices for imaging genetics studies of gross brain structure:
Assess reliability of phenotype measurements prior to conducting association studies to understand noise levels
Conduct unbiased genome-wide association rather than biased selection of candidate genes
Appropriately power an association test based on effect sizes from similar phenotypes (current gross brain structure GWAS suggests ~10,000 samples are needed to detect significant common variant associations)
Apply rigorous statistical thresholds for association across the whole genome (P<5×10−8 for common variant associations with more stringent thresholds if multiple phenotypes are tested)
When conducting a genome-wide association meta-analysis to achieve sufficient sample sizes, extensive quality checking of individual site level data is necessary prior to meta-analysis
Attempt replication of significant associations in independent datasets

Effect sizes of genetic variants on brain structure

Rare variants can have very large effects on the structure of the human brain. Early genetic associations identified rare, mostly genic, alleles of strong effect that lead to disorders that strongly change the structure of the brain including microcephaly, macrocephaly, polymicrogyria, and cobblestone lissencephaly^78–80. While rare variants can have large effects, they do not explain the majority of variability in brain structure within a population. To address this, consortia have recently identified hundreds of genome-wide significant common variant loci, each generally composed of multiple correlated SNPs, that impact human brain structure including intracranial volume (7 loci)^7,8,81–83, subcortical volumes (38 loci)^{6,8,70,81,84,85}, ventricular volumes (7 loci)⁸⁶, cortical surface area and thickness (150 loci)^74,87, and even white matter anatomy (225 loci)⁷¹. No variants have yet been identified impacting brain function as measured through fMRI⁷¹, though a handful of variants reach genome-wide significance for oscillatory brain activity as measured with electroencepholograms (2 loci)⁸⁸.

Before these large consortia identified replicable genetic associations, it was hypothesized that structural and functional neuroimaging traits were “closer to the underlying biology”. Therefore, genetic associations would have a much stronger effect (higher effect size) on quantitative measures of brain structure or function than on heterogeneous and clinically defined disorders^22,33–35. This assumption fueled the low sample size of “candidate gene” association tests, described above. Given recent replicated genetic associations to many different traits, we can now empirically test the higher effect size assumption. SNPs can indeed have a higher effect on some traits as compared to others. For example, individual SNPs can explain well over 50% of the variability of the expression of a gene⁸⁹ and therefore are detectable in sample sizes of ~100. No individual SNP has been detected with that level of effect on brain structure or function. However, individual SNPs with slightly higher impact on brain structure (maximum of ~0.8% of phenotypic variance explained for a SNP associated with putamen volume) than for schizophrenia risk (maximum of ~0.2% of phenotypic variance explained for a SNP associated with schizophrenia; see Figure 3) have been detected⁹⁰. So, there are instances where common genetic variation shows a stronger impact on brain structure than on risk for neuropsychiatric disorders. This gives support to the hypothesis that structural neuroimaging traits are indeed closer to the underlying biology, just not to the degree previously hypothesized in the candidate gene studies. We note that winner’s curse, an upward bias in effect sizes in association studies that first discover significant relationships⁹¹, could inflate effect sizes discovered in these imaging genetics studies, so the true differences will require further replication samples. Nevertheless, comparing effect sizes in replication samples alone also demonstrates examples of higher effect sizes for brain structure as compared to disease⁹⁰. Replicable influences of common genetic variation on brain function have been more difficult to detect, indicating that either effect sizes are lower for genetic associations to functional brain traits or that higher noise limits power⁹².

Figure 3:

Effect size relationships across traits. Each bar corresponds to a common genetic variant, a single-nucleotide polymorphism (SNP), with an association to cortical structure⁷⁴, subcortical structure or intracranial volume (ICV)⁸, height⁶⁹, schizophrenia¹⁰, and educational attainment¹⁵⁶. The highest effect size SNPs for each trait are shown. Effect sizes were measured in percent variance explained, the fraction of total trait variance that is accounted for by the genetic variant, (for quantitative traits) or percent variance explained on the liability scale (for disease categories). Error bars depict 95% confidence intervals. Statistical significance was calculated between the highest effect size SNPs associated with brain structure and the highest effect size SNP associated with schizophrenia. Significance was calculated by transforming effect size to Fisher’s Z and calculating the significance of the difference in Z-values. SNPs associated with brain structure traits can account for a greater portion of the total phenotypic variance compared to SNPs correlated with risk for schizophrenia disorder. This figure is extended from previous work⁹⁰.

Measuring the effect size differences at an individual locus for two different traits determines whether the maximum effect of a variant is higher in one trait than another. Recent studies^93–95 have asked a broader question: if effect size distributions of common variants across the genome are in general stronger for neuroimaging traits than they are for neuropsychiatric disorders. Effect size distributions can be estimated by clustering the effect sizes of LD-independent SNPs across the genomes into one of two general categories: (1) susceptibility SNPs which show some level of non-zero effect sizes (although do not necessarily survive genome-wide significance), and (2) SNPs which have signals so low they cannot be distinguished from no effect. The clustering is applied through a mixture model approach. Effect size distributions are inversely related to polygenicity: the greater the number of susceptibility SNPs, the smaller the effect of each of those SNPs on the trait. Genetic variants impacting psychiatric disorders are amongst the most polygenic studied with an estimated 10,000-50,000 susceptibility SNPs⁹³, indicating that many genetic variants each of exceedingly small effect size impact risk for these disorders^93,94. In comparison, ulcerative colitis or asthma are estimated to have only 1,000-2,000 susceptibility SNPs⁹³. Within the realm of imaging genetics, the degree of polygenicity of the putamen is over 30 fold less than that of schizophrenia, again indicating a higher effect size distribution of a brain structure trait as compared to a disorder⁹⁵. Effect size distributions have not been comprehensively evaluated for all associations to neuroimaging traits and then compared with those from neuropsychiatric disorder risk, although this is an interesting research direction.

While the effect size of genetic variants on brain imaging traits is still modest and generally requires sample sizes close to or over 10,000 subjects to reliably detect any genome-wide significant association^70,81, current evidence suggests that common variants do have a slightly larger effect on at least some brain structure traits than on risk for neuropsychiatric disorders. Such variants imply that at least some brain structure traits (like putamen volume) are affected by relatively less polygenicity by residing closer to underlying biological processes, thus satisfying the first criteria of an “endophenotype”. Associations with these variants are promising starting points for experimental manipulations because they are both detectable with slightly smaller sample sizes and can highlight brain regions and circuitry shouldering the consequences of genetic risk.

Assessing the global shared genetic basis between gross brain structure and risk for neuropsychiatric disorders

How well have neuroimaging traits fulfilled the second usage of ‘endophenotype’, thereby promoting mechanistic understanding of the effects of genetic risk loci? An initial approach to addressing this question is to ask whether the same genetic variants impact both disorder risk and the gross structure or function of the brain. Using modern statistical genetics techniques, termed genetic correlations, to analyze summary statistics from GWASs⁹⁶, it is possible to determine if common risk alleles across the genome are shared between ancestries for the same disorder or between disorders. Strongly positive genetic correlations are observed between schizophrenia in individuals of European descent and of African descent (r_g=0.61), indicating a largely shared common genetic architecture across ancestries^48,97,98. Perhaps unsurprisingly, risk alleles are shared between schizophrenia, major depressive disorder and ADHD with positive genetic correlations (r_g>0.2), indicating a shared genetic basis and blurring etiological distinctions between them⁹⁹. No significant sharing of risk alleles was detected, however, between Alzheimer’s disease and any psychiatric disorder, indicating that they are affected by largely independent genetic variants and have distinct etiologies⁹⁹.

Significant genetic correlations observed between alleles influencing brain structure in the population and alleles impacting risk for a neuropsychiatric disorder could help localize brain regions critical for disease pathology. In this spirit, genetic correlations have been performed between hippocampal volume and Alzheimer’s disease risk. These demonstrated a significant negative genetic correlation (r_g=−0.15), indicating that alleles associated with decreased hippocampal volume in the general population are also, in part, associated with increased risk for Alzheimer’s⁶. Giving further credence to this finding, structural deficits in the hippocampus are well known to be associated with Alzheimer’s disease¹⁰⁰.

For psychiatric disorders where causally implicated brain regions are not known, genetic correlations may help to identify critical regions involved in the pathology. For example, negative genetic correlations have been observed between the volume of the caudate and nucleus accumbens and bipolar disorder (r_g=−0.17 and r_g=−0.28, respectively)⁸⁴. Other negative genetic correlations have been observed between global cortical surface area and both major depressive disorder (r_g=−0.13) and ADHD (r_g=−0.17)⁷⁴. Further supporting the negative genetic correlation between cortical surface area and ADHD, intracranial volume, highly correlated with cortical surface area, is also negatively genetically correlated with ADHD risk (r_g=−0.23)¹⁰¹. Positive genetic correlations have been observed between global cortical surface area and both Parkinson’s disease (r_g=0.22) and intelligence (r_g=0.22)⁷⁴. These studies are notable because they implicate both the brain regions and the direction of phenotypic effects associated with neuropsychiatric pathologies, thus linking genetics to both disorder and brain structure.

There are some important limitations to the genetic correlation method. It cannot distinguish between causal mediation and a liability index model (Figure 2). As described above, in a mediational model, the neuropsychiatric disorder risk alleles impact brain structure which then impacts risk for a neuropsychiatric disorder through a causal pathway. In a liability index model, a risk allele independently impacts both brain structure and risk for a neuropsychiatric disorder²², indicating pleiotropy instead of causality. To determine which model best fits the data, one needs to experimentally alter brain structure in certain randomly chosen individuals and determine if risk for schizophrenia is altered compared to individuals without alteration in brain structure. If so, this supports the causal mediation model. If not, this supports the liability index model. Of course, experimental manipulation of brain structures is not ethically possible in humans, but our innate genetic differences provide a useful approximation. Assuming that alleles at brain structure associated variants are like treatments randomly assigned in a randomized clinical trial, we can make some causal inferences (given several assumptions detailed in the following references^102,103). Using natural genetic diversity in human populations, we can select alleles that are associated with brain structure as a proxy for perturbing brain structure in humans and determine if they also impact risk for neuropsychiatric disorders through a regression framework. If so, this provides support for a mediational model that describes the causal influence of brain structure on risk for neuropsychiatric disorders using a so-called Mendelian randomization (MR) approach. To our knowledge, this method has not yet been applied to neuropsychiatric disorders and brain structure, but has shown that alleles associated with increases in intracranial volume also associated with increased intelligence, putatively showing not just a genetic correlation but also a causal relationship⁶⁸. Applying this method will be increasingly effective as more loci that impact brain structure are identified, since the power of MR inferences increases with the number of associated SNPs for the phenotype.

Assessing the local shared genetic basis between gross brain structure and risk for neuropsychiatric disorders

The above methods assess globally, across the genome, whether common variants are associated with both changes in brain structure and risk for neuropsychiatric disorders. In order to determine if the same causal SNP(s) impact both traits at an individual locus, we search for SNPs significantly associated with both phenotypes. However, due to linkage disequilibrium (LD), the correlation between genetic variants, an individual SNP significantly associated with two traits could be the result of two separate causal variants in close proximity on the genome, each independently influencing just one of these two different traits. In this case, the observed significance of a SNP in both traits would be an artifact from decaying significance with LD and could be misinterpreted as the same causal SNP jointly contributing to both traits. To infer if the same causal variant(s) influences two traits, one can determine whether the association statistics follow expected LD patterns driven by a single causal SNP or set of causal SNPs^104–108. Colocalization tools have been recently applied to genome-wide association data from schizophrenia and eQTL data from post-mortem brain to identify genes impacting risk for this disorder²⁸. These analyses have identified 40 colocalized signals which provide strong evidence for specific genes whose altered expression is associated with schizophrenia. Although colocalization tools have not yet been formally applied to identify individual loci that impact brain structure and risk for neuropsychiatric disorder, this analysis could identify shared causal variants impacting both brain structure and risk for neuropsychiatric disorders. As such, colocalization analyses will be essential for understanding the causal chains leading to risk for neuropsychiatric disorders.

Importantly, colocalization of signals does not guarantee a causal relationship. For example, if a colocalized SNP affects both brain structure and risk for ADHD, it is not known whether the SNP creates risk for ADHD through changes in brain structure or if the SNP influences ADHD independent of brain structure. This ambiguity makes mechanistic interpretation and subsequent experimental design difficult, as it is not clear whether modification of brain structure will have an effect on ADHD risk. If all data are acquired in the same subjects, it is possible to infer causality at the single locus level using mediation analysis^27,109,110. To our knowledge, however, no tools yet exist to prioritize causality on the individual SNP level using summary statistics alone.

The puzzling lack of shared genetic basis between some neuropsychiatric disorders and brain structure

Several well-powered genetic studies of neuropsychiatric disorders, most notably for schizophrenia^10,39, have so far not demonstrated a significant genetic correlation with any structural or functional brain imaging trait^74,90. This indicates that the same common genetic variants associated with many different brain traits are not demonstrably associated with risk for schizophrenia. (Note that we cannot accept the null hypothesis of no genetic correlation, but we do not observe a genetic correlation significantly different than zero in current sample sizes.) Given the observed brain structural differences between individuals with schizophrenia and neurotypical controls^13,111, as well as the close relationship between structure and function at multiple levels of the brain^112,113, why do common genetic variants impacting risk for schizophrenia leave the gross structure of the human brain undetectably changed?

There could be many reasons to explain this puzzling finding. (1) We are underpowered to detect genetic correlations. However, current efforts would set an upper bound on which genetic correlations would be detected so genetic correlations identified in larger samples must be smaller than those currently observed. (2) Primary causal environmental influences in the absence of genetic effects, for example infection, are driving the observed brain structural differences. However, though environmental influences have been shown to have effects on brain structure and function^114,115, a primary causal role for these influences in schizophrenia is still difficult to establish^116,117. (3) Environmental influences, for example medication, taken in response to schizophrenia diagnosis causes could be driving the observed brain structural differences between cases and controls¹¹⁸. In this case, the observed structural differences are not causing the disorder but are a result of the disorder. (4) Rare variation contributing to schizophrenia risk that is unmeasured in these common variant association studies could be driving the observed brain structural differences. This is unlikely given that common variation, when considered in aggregate, is the greatest contributor to risk for schizophrenia in the population and even individuals harboring a rare mutation also have a polygenic common variant burden^48,99,119. (5) We are measuring genetic influences on brain structure within developmental time periods not critical to disease pathology. Developmental fetal and infant imaging perhaps would detect genetic correlations unobserved in adults^120,121. (6) Brain function, manifesting as the altered behaviors of individuals, could be changed in the absence of structural changes. Given the intimate relationship between brain structure and function¹²², this seems unlikely. Or, (7) brain changes that predispose to schizophrenia risk happen at the cellular or subcellular level and do not manifest at the gross level at which brain images are taken with MRI. Though many of the above are possibilities, we find the last most compelling and detail further ways in which we can study genetic effects on cellular or subcellular level brain traits.

3D Brain Imaging beyond MRI

Genetic association studies continue to identify both common and rare variants that impact brain structures at the gross anatomical level^123–125. MRI is suited to measure gross structural traits of intracranial volume and regional phenotypes like hippocampal volume or cortical surface area and thickness^{6–8,74,81–83}. These phenotypes can be measured in large numbers of living humans at a relatively affordable cost. While such macroscale imaging is valuable for identifying some affected brain regions and subregions, a typical MRI voxel comprises a brain volume of roughly 1 cubic millimeter: a space that may contain tens of thousands of neurons and millions of synapses^126–128. So, genetic associations to gross brain structure lack the resolution to explain the cellular or subcellular basis of observed allelic differences. Understanding the cellular and molecular impact of genetic variation demands imaging on microscopic and ultrascopic scales (Figure 4).

Figure 4:

Structural brain imaging modalities for the future of imaging genetics. Each of three neuroimaging techniques provides distinct advantages and disadvantages at different biological scales. Neuroimaging of gross brain structures through MRI (leftmost diagram) reveals macro-scale anatomy of specific brain regions at millimeter resolution. The center depicts “micro-scale” imaging achieved by tissue-clearing followed by light sheet microscopy at micrometer resolution. Ultra-scale imaging is possible with serial block-face scanning electron microscopy (SBFSEM) at nanometer resolution. Combined observations at all three scales using genetically informative experimental designs will facilitate the mapping of causal pathways underlying genetic risk for neuropsychiatric disorders.

For example, if an allele at a particular variant is strongly associated with reduced cortical thickness as measured by MRI in an adult, there are many cellular and circuit changes that could lead to this macroscale alteration. Is the total number of cortical cells reduced, or are they more densely packed? Do neurons in this region have fewer dendritic arborizations? Are relative contributions of specific cell-types altered? Is synaptic density or structure disrupted? Is the spatial architecture and pattern of connectivity within the structure impacted? None of these questions can be adequately addressed with MRI. Measurements of cell numbers, cell types, synapses, circuit connectivity and arrangements of all of these components in 3D space will require both access to post-mortem tissue and image resolution sufficient to capture cellular and molecular features.

We expect that data illuminating these features will be critical to further map causal pathways from genetic risk loci to neuropsychiatric dysfunction, and also may explain why genetic correlations are not observed at gross brain structural levels with some neuropsychiatric disorders. Measuring phenotypes manifested at the cellular and molecular levels, closer to the effects of a genetic risk factor on the causal chain, will likely be more mechanistically informative and have higher effect sizes - though until association studies are conducted, we will not know the effect sizes. Below, we discuss two emerging imaging technologies that provide cellular resolution or subcellular 3D images of post-mortem human neural tissues, and how they could be applied in a genetically informative design. We describe how modern fluorescence-aided imaging of tissue-cleared samples and high-magnification serial electron microscopy can help to more completely explain the causal mechanisms underlying the genetic risk for disorders of the brain.

Imaging brain structure at cellular resolution with tissue clearing

Our understanding of nervous system function is critically dependent on visualizing the three-dimensional structure of the brain. Cellular resolution volumetric imaging enables interrogations of cell number, density, morphology, and spatial architecture impossible with macroscale imaging, and increases the accuracy and detail of these measurements compared to 2-D image sections. The ability to label and quantify specific cell-types using fluorescence immunohistochemistry in 3-D could help explain the cellular basis of allelic differences in specific brain regions implicated in macroscale imaging genetic studies. For instance, 3-D microscale imaging may reveal cell-type specific spatial disruptions, aberrant cellular morphology, imbalanced numbers of excitatory versus inhibitory neurons, or disorganized cortical lamination.

Cellular resolution optical imaging of post-mortem brain tissue is often restricted to thin 2-D sections due to light’s inability to penetrate deeply into the sample. Imaging many serial 2-D sections to reconstruct 3-D images is possible¹²⁹, but inefficient and may distort features within the images. Several tissue-clearing techniques have recently been developed to address this problem^130–134. These techniques remove light-scattering lipids while preserving cellular spatial arrangements of proteins to yield intact, transparent tissue amenable to light microscopy. Most techniques involve a series of immersions in chemical solvents that dehydrate the sample, dissolve away lipids and induce chemical modifications to create a uniform refractive index, thus minimizing destructive interference¹³⁵. In combination with light sheet microscopy for quick image acquisition¹³⁶, these methods enable cellular and circuit measurements at micrometer resolution, and are compatible with the use of modern histochemical labeling methods in human tissue. Tissue-clearing approaches are also being scaled and adapted specifically for use on post-mortem human brain tissues, providing a valuable tool for obtaining high resolution volumetric observations of neuropsychiatric pathologies¹³⁷. As an example, tissue cleared brains from Alzheimer’s disease patients showed remarkably detailed 3D measurements of amyloid plaque volume, morphology, and arrangement which varies distinctly across subjects¹³⁸.

These techniques make it possible to design an imaging genetics study of post-mortem brain structure at micrometer resolution. Such a study may be able to directly answer some of the questions posed above about the cellular basis for imaging genetics findings identified with MRI, as well as to identify new genetic influences on cellular level brain structure that were not observable in current association studies given the poor resolution of MRI. In order to design such a study, one needs access to a large set of post-mortem brain samples from genetically diverse and genotyped individuals. Brain banks are therefore a key resource in order to accomplish this goal¹³⁹. In addition, genotyping can be completed on existing brain-banked samples and is at this point quite affordable (<$100 USD for arrays which allow genome-wide genotype imputation).

There are of course some barriers to completing such a study. First, post-mortem tissue can be stored either through flash freezing or fixation, and though tissue clearing has been performed on both types, there are limitations: for archival frozen tissue, large chunks may be difficult to fix fully, and for archival fixed tissue, over-fixing may mask the epitopes. Second, it is not known how post-mortem interval affects the degradation of proteins or the ability to label and image them. Third, collecting large sample sizes will likely involve the concerted efforts of multiple different brain banks working together, which can introduce technical variation. Quantifying the degree of technical variation across samples processed at different facilities, minimizing that technical variation via standard operating procedures, and statistical adjustment for uncontrolled for technical variation has been accomplished for many fields including gene expression via RNA-seq^140,141, induced pluripotent stem cells¹⁴², and other post-mortem tissue biobanks¹⁴³, so will likely be possible in this field as well. Fourth, tissue clearing allows labeling most molecules (assuming an antibody or probe exists and is able to diffuse within the tissue) with specific expression in a cell-type. It is often not clear which specific cell types are driving an imaging genetic association, so it may be difficult to design the experiment labeling proteins of interest. When trying to explain the specific cell types underlying an observed gross brain structural hit, the proximity on the genome or functional evidence linking to a specific gene may guide cell types to probe⁴⁵. If however, there is no prior hypothesis, studying the most prevalent cell types may be a good place to start, for example labeling upper and lower layer excitatory neurons in the cortex as well as all nuclei. This can lead to future hypotheses about cell fates, spatial architecture and density. Fifth, tissue clearing and light sheet microscopy are not currently possible within an entire intact human brain. Current working distances of light-sheet microscopes allow imaging of chunks of tissue near the size of a mouse brain (10×10×5.6 mm). Given that cortical thickness in humans is on average 2.5 mm¹⁴⁴, it is possible to image the entire cortical depth of a tissue-cleared sample at a particular location, but covering the entire cortical wall across the ~2500 cm² surface area of the cortex would require imaging over 2000 chunks of tissue. Finally, few tools exist for image segmentation, analysis, and storage in large tissue cleared images^145–147. Image segmentation currently works best on features that are sparse and with simple morphology. For example, using a nuclear label for sparse inhibitory neurons in the cortex will be much easier and more accurate to segment than all nuclei. Additionally, computational storage space for images of voxel size ~1×1×4 μm/pixel is around 0.5TB for each channel for an image the size of a mouse brain (or a chunk of the human cortical wall). Raw data for a sample size of 100 donors for the cortical wall with 4 different types of cells labeled would be 200 TB alone, not including copies of data made in downstream processing. Clearly, large computational resources would be needed to tackle this problem.

By observing tissue-cleared cell populations in brain areas where MRI has identified structural aberrations associated with genetic risk, we can begin to build causal pathways that more explicitly connect genetic risk to brain disorders. When important phenotypes occur in small subcellular structures, such as synapses, electron microscopy techniques that can resolve ultrastructures may be more appropriate, and are discussed below.

Imaging brain structure at subcellular resolution with scanning electron microscopy

Genetic variation may impact synaptic morphology and density, fine-wiring of neuronal processes, axonal myelination and diameter, or other subcellular features. Such ultrastructural phenotypes will be difficult to capture with gross anatomical imaging or optical microscopy. A long standing tool to observe ultrastructures is electron microscopy (EM) which has the power to resolve subcellular organelles, neuronal processes, and the detailed structure of synaptic machinery.

Scanning electron microscopy (SEM) focuses and sweeps a beam of electrons across a fixed and dehydrated biological sample. As the electron beam interacts with molecules within the sample, emitted electrons are detected to create a micrograph image with lateral pixel resolution down to 3.7nm¹⁴⁸. While SEM micrographs provide an impressive depth of field, traditionally a given image is limited to a single surface of a tissue section. However, modern SEM approaches can reconstruct a 3D volume from a series of two-dimensional images made on thin sections of tissue prepared using a microtome or cryostat^149,150. One way this process has been streamlined is via serial block-face scanning electron microscopy (SBFSEM), which combines high-throughput tissue sectioning and subsequent image acquisition¹⁵¹. In this technique, the surface of a tissue block is imaged with repeated passes of Scanning EM as the top layer of the tissue is shaved away by a microtome to access successively deeper sections. Initial 3D reconstructions using SBFSEM achieved a per-section depth of 50nm with lateral resolution less than 10nm, sufficient to resolve synaptic nanostructures and map circuitry by following axonal paths¹⁵¹. Kasthuri and others employed a similar technique when they captured SEM micrographs of brain volumes serialized using an automatic and tape-collecting microtome to visualize a small volume of mouse cortex with astonishing detail, such that nearly all subcellular elements could be identified with a voxel size of 3×3×30nm¹⁵². Their 3D reconstructions clearly resolved axons, dendrites, interactions between glia and neurons, and many subcellular ultrastructures, including synapses, synaptic vesicles, spines, postsynaptic densities, and mitochondria - all features that could play major mechanistic roles in disorders of the brain.

Volumetric ultrascale imaging techniques could be employed to understand the underlying subcellular influences of genetic variation associated with gross brain structure, and may identify novel genetic associations to structural phenotypes that are undetectable using tools that are limited to measuring gross brain structure. As an example, let us revisit the study of Sekar and others, who connected a specific genetic locus associated with schizophrenia to the complement immune response and its role in pruning synaptic connections⁵⁰. No gross brain structural associations have been detected at this locus to date. However, detailed ultrastructural imaging of synaptic number, morphology, and density in post-mortem samples carrying the C4 risk alleles or protective alleles would provide direct and valuable evidence to support a mechanism of regionally specific synapse loss in schizophrenia. Similarly, SBFSEM could probe cellular hypotheses to explain genetic effects on the volumes of particular brain areas. The volume of a brain structure could be affected by changes in cell size, the density of neuronal processes, rearrangement of spatial architecture, or a mixture of these effects. All of these possibilities could be investigated with ultrascale electron microscopy on samples from the brain region of interest.

There are again limitations to performing such a study, some of which we detail here. Kasthuri et al’s impressive saturated reconstructions at 3nm/pixel resolution were performed on just a 40×40×50 μm subset of the tissue block to provide an extremely detailed inventory of one apical dendrite’s surroundings. This size represents ~2% of the thickness of the human cortical wall and less than a millionth of total human cortex volume¹⁵³. In addition, even with expected improvements in acquisition speeds, a 1 mm³ section of brain can take ~2 months to acquire¹⁵⁴. Because of this, specific regional hypotheses and/or scaling up imaging acquisitions are needed to apply this technique for imaging genetics studies. With such detailed ultrastructural images, data storage and downstream processing present technical bottlenecks¹⁵⁴. A single SBFSEM volumetric reconstruction may create an image stack comprising thousands of micrographs¹⁵², which can occupy hundreds of gigabytes, and this would represent only a fraction of the data needed to describe the reach of one pyramidal neuron¹⁵⁰. For large-scale ultrastructural studies, automated analysis tools and data storage solutions are needed to facilitate the processing of large imaging data sets collected from many individuals across broad swaths of the brain¹⁵⁵. While obstacles in access to tissue samples, image collection time, and data storage are nontrivial, we highlight these imaging techniques because of their unique potential for providing direct evidence of structural disruptions producing pathology. Further, we hope that appraisal of both the promise and limitations associated with these approaches can accelerate their development and eventual application.

The combination of well-powered genetic studies and gross anatomical imaging of human brain tissues with MRI has provided valuable hints towards the brain regions mediating genetic risk for neuropsychiatric disorders. Microscale and ultrascale imaging genetics will likely identify genetic influences on cellular density, number, arrangement, and synaptic connections that may be able to both explain the cellular basis of gross imaging genetics associations as well as identifying novel associations to brain structure.

Conclusions:

Many loci in the genome have a replicable association with risk for neuropsychiatric disorders. To understand how variation at these loci leads to alterations in cognition and behavior, we need to understand the cell-types, developmental time periods, brain regions, and biological processes impacted by those variants. To do this, we can map webs of QTLs between genetic variation and multiple endophenotypes leading to disorder symptoms. We provide examples of successful integration of multiple lines of genetic association data to explain the basis of genetic risk for other complex traits like obesity and diabetes. We assert that similar approaches augmented by increasingly high-powered and high-resolution genetic associations to brain structure and function will help us understand the causal basis for disorders of the brain. Indeed, MRI measurements have demonstrated significant genetic correlations between certain brain structures and ADHD, major depressive disorder, bipolar disorder and Alzheimer’s disease, though the causality of these effects remains to be confirmed. Subsequent modulation of endophenotypes along a causal chain with experiments in model systems can validate the downstream effects of those genetic variants. Layering multiple levels of genetic association with imaging data and experimental validation will generate important mechanistic connections that can illuminate previously dimly-lit causal pathways creating risk for neuropsychiatric illness.