Over 50 million (one in five) Americans are affected by psychiatric disorders, which include diverse conditions that vary in symptoms, severity, outcomes, and treatment responsiveness. These polygenic disorders arise from a combination of variants that contribute to risk with differing frequencies and penetrance. The extreme clinical heterogeneity among patients with the same diagnosis, coupled with the frequent overlap in clinical presentation across disorders, can delay accurate diagnosis, which remains highly subjective (
1). Despite a substantial investment in genetic studies that have now cataloged hundreds of disease-associated variants, findings have not translated into clinical insights. Why is this important? A better understanding of the molecular underpinning of disease will make possible the prediction of clinical outcomes years before the onset of symptoms and facilitate the discovery of drugs that might delay, mitigate, or reverse disease pathophysiology.
Precision Medicine: Preventing Disease Before It Arises
Large genetic studies can identify rare “genetic heroes,” carriers of Mendelian mutations without clinical manifestation of disease (
2). For example, and with relevance to the brain, case reports include one such genetic hero for Alzheimer’s disease: an autosomal dominant
PSEN1 mutation carrier, which results in early disease onset (at 30–50 years of age), who also happens to be homozygous for the protective
APOE3 Christchurch mutation (R136S) and has shown no evidence of early-onset Alzheimer’s disease for three decades (
3). More commonly, although
APOE4/4 carriers have a 15-fold increased risk for Alzheimer’s disease (
4), 1 in 10 “resilient” individuals do not develop symptoms. In either case, the variable penetrance of
PSEN1 and
APOE4/4 variants can be reframed as evidence of naturalistic “cures” that can prevent disease in individuals with otherwise high genetic predispositions. In the widely accepted additive risk model for genetic disease, does there exist a discrete “tipping point” between health and disease? Can ameliorative early interventions “untip” genetic disease risk?
Even for highly penetrant conditions with no current treatment strategies, a spectrum of phenotypes are possible. Individuals with Down syndrome have IQs ranging from <40 to 120 (
5), and symptoms of Huntington’s disease can onset from 2 to 79 years of age (
6,
7). This vast heterogeneity reflects complex interactions between genes, circuits, and environments. Understanding the contexts that buffer genetic risk will allow individuals to achieve their greatest phenotypic potential and improve life outcomes.
Stem Cell Models: Resolving Cell-Type-Specific and Context-Dependent Effects of Genetic Risk
Many psychiatric disorders, notably schizophrenia, include a long prodromal period prior to the onset of symptoms. In order to study the molecular mechanisms underlying disease initiation prior to diagnosis, physicians and scientists have historically tracked high-risk individuals for decades, waiting for symptom onset in a small subset of their cohort. Instead, using stem cell models, it is possible to recapitulate neurodevelopment in vitro, essentially observing risk in a laboratory dish. We and others have repeatedly observed that neurons and glia derived from human induced pluripotent stem cells (hiPSCs) most resemble their prenatal in vivo counterparts (
8–
13), and so are particularly well suited to test the neurodevelopmental impact of the many psychiatric risk variants predicted to influence fetal cortical development (
14). Put simply, these risk-in-a-dish models make it possible to observe cellular phenotypes related to disease, uncover molecular mechanisms driving these perturbations, and screen for drug and/or gene therapies with which to prevent or reverse them.
The first human embryonic stem cell lines were grown in 1998, with the theoretical capability to differentiate to every cell type in the human body, and so they represented a nearly limitless source of material for studies of human disease (
15). Research into the patterning of these cells, toward understanding the molecular networks regulating pluripotency—their unique ability to differentiate into all cell lineages—uncovered four genes that, when overexpressed, were sufficient to reprogram adult somatic cells into hiPSCs (
16). This transformative discovery made it possible to generate stem cells from any individual, and therefore to make the precise cell types affected by any disease, in order to observe disease-associated processes in real time. Unfortunately, the results have been anything but clear-cut. For example, in schizophrenia hiPSC-based studies, phenotypes have been reported in each of glutamatergic neurons (
17), GABAergic neurons (
18), dopaminergic neurons (
19), neural progenitor cells (
8), astrocytes (
20), oligodendrocytes (
21), and microglia (
22). Moreover, as the case-control hiPSC experiments of schizophrenia grew larger, it became clear that patients did not all differ from control subjects in the same way (
23), and also that there was substantial variation between individual control subjects (
13). In hindsight, the field’s earliest experiments proved naive. But why?
Genetics and Genomics: The (Increasingly) Complex Risk Architecture of Psychiatric Disorders
Over the past decade, our understanding of the complex genetics of psychiatric disorders has exploded, although admittedly, there is much more yet to learn. Ten years ago, the largest-ever genome-wide association study (GWAS) of schizophrenia identified seven loci associated with disease risk (
24). Today, with over 10-fold more individuals analyzed, there are 270 loci (
25), with no clear estimate as to how many variants will ultimately be implicated in disease risk (
26,
27). There are now hundreds of significant loci associated with psychiatric (
28–
31) and neurodegenerative disease risk (
32,
33). Risk variants linked to schizophrenia (
28,
34–
36), autism spectrum disorder (ASD) (
37–
39), and more broadly across the neuropsychiatric disorder spectrum (
14,
40–
43) are enriched for genes involved in synaptic biology and gene regulation (
39,
44–
48) and for those expressed during fetal cortical development (
14,
44,
49,
50). Although highly penetrant rare mutations underlie a fraction of cases (
51), psychiatric disorders more frequently arises from risk variants that each confer only a tiny increase of risk and are common in the population at large (
25). Thus, individually small risk effects combine to yield much larger impacts in aggregate, although the precise mechanisms involved remain unclear (for a review, see reference
52).
CRISPR: Engineering Risk for Psychiatric Disorders
Fortuitously, in parallel with these genetic advances were substantial developments in genome engineering (
53). CRISPR-associated proteins (Cas) use short and easily synthesized “clustered regularly interspaced short palindromic repeats” (CRISPR) sequences as guide RNAs to recognize specific complementary strands of DNA. Together, they form the basis of a technology known as CRISPR-Cas, which can be used to efficiently edit DNA to introduce variants or rescue mutations (
54), and also to alter the epigenome—for example, manipulating histone modifications (
55,
56), DNA methylation (
57), and chromatin interactions (
58), directly activating or repressing expression of target genes (
59), and cleaving RNA (
60). Engineering can be multiplexed (
61), parallelized (
62), and conducted at near genome-wide scale (
63). Thus, it is now theoretically possible to introduce, or reverse, the myriad risk loci associated with psychiatric disorders, alone or in combination, in specific neuronal cell types.
Many hiPSC-based studies of psychiatric risk genes recruit patients with highly penetrant rare mutations (e.g.,
MECP2 [
64],
FMR1 [
65],
CACNA1C [
66],
SHANK3 [
67],
NRXN1 [
68],
NLGN4 [
69], 22q11.2 [
70], and
DISC1 [
71]), observing phenotypes such as altered neuronal development and/or synaptic function. Likewise, engineered deletions of
FMR1 (
65),
SHANK3 (
72),
NRXN1 (
73), and many more genes (
74) revealed similar impacts. Although the common risk variants identified through GWAS are predicted to result in relatively more subtle phenotypic effects, several such risk loci have successfully been explored within case-control cohort designs (e.g.,
C4 [
22],
CACNA1C [
75]). To improve our ability to resolve the small effect sizes associated with common variants (termed single-nucleotide polymorphisms or SNPs), CRISPR engineering can be used to make comparisons across a shared genetic (“isogenic”) background. This strategy has demonstrated the causal role of schizophrenia-associated GWAS SNPs in gene regulation through enhancer-promoter looping (
CACNA1C [
76]), 3D-genome folding (
PCDHα [
77]), miRNA abundance (miR-137 [
78]), and mRNA levels (
FURIN [
79]). In fact, CRISPR-based isogenic comparisons of one such noncoding SNP (rs4702) produced genotype-dependent transcriptomic and neuronal activity differences that were not just cell-type-specific but also microRNA-dependent (
79). Likewise, for two predicted target genes of schizophrenia GWAS (
SNAP91,
TSNARE1), CRISPR-based approaches to activate and repress endogenous gene expression resolved distinct pre- and postsynaptic perturbations (
79).
While CRISPR-based approaches can manipulate a handful of variants or genes, higher-throughput methods are necessary to comprehensively survey all of the loci associated with psychiatric disorders (for a review, see reference
80). Toward this, massively parallel reporter assays (MPRAs) can evaluate the regulatory activity of the thousands of loci associated with human cortical development (
81,
82) and psychiatric disorder risk (
83–
85). Additionally, CRISPR expression quantitative trait loci (eQTL) mapping can identify SNPs that differentially regulate proximal (e.g., crisprQTL [
86]) and distal (e.g., CRISPRi-FlowFISH [
87]) target gene expression. Finally, population-scale “village-in-a-dish” experiments, whereby single-cell RNAseq is applied to pooled populations of hiPSC-derived neurons or glia from dozens of genotyped donors (e.g., Census-seq [
88–
91]) are extremely well powered to link SNPs to gene expression and phenotypic effects.
Traditional functional genomic approaches generally involved one-gene-at-a-time “genotype-to-phenotype” studies, applying prior knowledge in a hypothesis-driven manner to test the causal role of specific genes. By contrast, large-scale screens are “phenotype-to-genotype” approaches, broadly manipulating many variants or genes at once in order to characterize those that result in specific changes. Although traditionally conducted in an arrayed format that requires a substantial time and resource investment, state-of-the-art CRISPR or village-in-a-dish screens incorporate a pooled design. Coupling CRISPR-based perturbations to single-cell RNA sequencing for analytical readouts (e.g., Perturb-seq [
92] and ECCITE-seq [
93]) further expands the breadth of biological questions that can be tested in each pooled screen. Altogether, this makes possible the comprehensive interrogation of candidate disease genes (
94) or unbiased gene lists (
95–
97) in a single-tissue culture well, dramatically reducing the time and financial costs involved.
Fundamental Questions Lacking Answers
A major challenge in the field is connecting disease-associated risk variants (the vast majority of which fall in noncoding sequences) to their respective target gene(s), pathway(s), and causal cellular phenotype(s) (
52), and in doing so, translating genetic findings into disease-relevant cell biology. Given that a spectrum of developmental and neuropsychiatric phenotypes share overlapping genetic architectures (
52), there is great value in exploring their shared heritability.
Rare and common variants underlying risk for psychiatric disorders show evidence of cumulative effects (
98,
99) and are predicted to converge at the pathway level (
100). Toward this, we recently perturbed a dozen schizophrenia risk genes in isogenic glutamatergic neurons, observing convergence on a subset of genes and subnetworks involved in synaptic function (
101). Recent studies likewise queried an overlapping set of ASD genes in vitro in human neural progenitor cells (27 genes) (
94) and human brain organoids (three genes) (
102), and in vivo in fetal mouse brains (35 genes) (
103) and
Xenopus tropicalis (10 genes) (
104). Using hypothesis-free pooled (manipulating many genes at once in the same “chimeric” cell culture experiment or animal) and/or parallelized (manipulating many genes one at a time in “arrayed” format across cell culture wells or animals) approaches, the studies reporting convergence of ASD genes impacting neurogenesis (
94,
102–
104), Wnt signaling (an important neurodevelopmental pathway that controls cell fate decisions and tissue patterning) (
94) and gene expression (
102,
103). Moreover, in one report, two subgroups of ASD genes were resolved, which correlated with abnormal Wnt signaling: one that inhibited and one that enhanced spontaneous cortical neurogenesis (
94). Do these points of convergence represent novel points of therapeutic intervention?
It follows that we must also resolve how the effects of risk variants combine to yield much larger impacts in aggregate. Within individual patients, it is possible that genetic risk factors sum in different patterns depending on whether their target genes are expressed in the same cell types and/or converge within the same biological pathways. Individual risk variants may sum linearly (
105,
106) or be amplified (
107) or buffered (
108) by epistatic effects. We previously uncovered an unexpected combinatorial effect between risk genes that was not predicted from single gene perturbations, one that concentrated on synaptic function and linked the rare and common variant genes implicated in psychiatric disorders (
79). Likewise, a single GWAS risk locus that regulates expression of more than one target gene can result in highly significant gene×gene interaction terms, suggesting that they synergistically contribute to the molecular and synaptic phenotypes observed (
109). Does understanding the combinatorial interactions between risk variants improve genetic diagnosis and/or indicate new drug targets?
The task is complicated in that many of the genes regulated by disease-associated variants are most likely expressed and regulated in cell-type-specific and context-dependent manners. To date, efforts to explain psychiatric disorder risk through annotation of GWAS signals with functional genomics have not accounted for the impact of environment or stressors. This failure to incorporate gene-by-environment interactions may be especially critical in studies of those psychiatric disorders that require or include specific stressors (e.g., posttraumatic stress disorder) or exposures (e.g., substance use disorders, anorexia nervosa) among the diagnostic criteria, or that count exposure to illicit substances or extreme stress as among the critical risk factors for disease (e.g., schizophrenia). Glucocorticoid receptor signaling is highly associated with trauma response (
110); therefore, understanding the genetic basis for dysregulated stress response may yield novel therapeutic targets. Studies of in vitro brain organoids generated from healthy control subjects suggest that excessive glucocorticoid exposure could interfere with neuronal maturation (
111). Furthermore, we observed robust glucocorticoid hypersensitivity in hiPSC neurons derived from individuals with posttraumatic stress disorder (
112). How do risk variants interact with the environment across the diverse cell types that comprise the human brain?
Whereas every current antipsychotic drug targets dopaminergic signaling, genetic risk for schizophrenia is enriched for genes expressed by glutamatergic and subsets of GABAergic neurons (
113), suggesting that disease etiology reflects cell types distinct from observed pharmacology. Consistent with this, studies of neuronal convergence of ASD genes in brain organoids reveal asynchronous development of inhibitory GABAergic neurons and deep-layer glutamatergic projection neurons through distinct molecular pathways (
102). In fact, the mechanisms underlying psychiatric disorders are thought to drive symptoms through iterative pathological changes in circuit function (
114). Thus, the challenge is to assemble in vitro models that capture neuronal circuitry (
115,
116), glial support (
117–
119), vasculature (
120,
121), and blood-brain-barrier functions (
122) while retaining the capacity to deconvolve cell-type-specific effects. For example, mutations in
CACNA1C associated with Timothy syndrome cause deficits in calcium signaling in both glutamatergic (
66,
115) and GABAergic (
115) neurons, which specifically lead to perturbations in interneuron migration into cortical brain organoids (
123). How do risk variants impact cellular function within circuits?
Finally, the extent to which clinical drug responsiveness is heritable and/or stable throughout the lifetime, across the spectrum of neuropsychiatric disease, needs further investigation, but promising examples such as lithium-responsive bipolar disorder (
124) have been identified. We helped to show that hyperexcitability in hippocampal neurons derived from lithium-responsive, but not nonresponsive, patients with bipolar disorder is ameliorated after lithium treatment (
125). In an independent follow-up, a similar phenotypic analysis was capable of predicting with 92% accuracy whether hippocampal neurons were derived from a patient with or without clinical responsiveness to lithium (
126). Is clinical treatment response predictable? Future genomic approaches will provide the means to stratify patients with overlapping risk combinations into “genetically defined” cohorts. An improved drug screening strategy would better recapitulate disease pathophysiology and integrate advances in psychiatric genetics. Toward this, we described a proof-of-concept application of transcriptomic drug screening to hiPSC-based models, demonstrating that drug-induced differences (including many known antipsychotics) in patient-derived neural progenitor cells were capable of reversing postmortem schizophrenia transcriptional signatures and were enriched for genes related to schizophrenia biology (
127). Altogether, these studies reveal major advantages of incorporating cell type and patient-specific platforms in drug discovery.
Conclusions
Each person’s distinct genetic, epigenetic, and environmental risk profile predisposes them to some brain disorders and confers resilience to others. There will be no one-size-fits-all cure for psychiatric disorders. Each patient represents a unique aggregation of risk factors, which may converge on different pathways, interact in dissimilar cell types, and impact distinct developmental windows. Pursuing a functional genomics approach that integrates stem cell models and genome engineering will help to resolve the impact of patient-specific variants across cell types, genetic backgrounds, and environmental conditions. Striving to translate risk “variants to genes,” “genes to pathways,” and “pathways to circuits” will reveal the convergent, additive, and synergistic relationships between risk factors within and between the cell types of the brain. These insights could identify therapeutics tailored to an individual’s specific risk profile, and so springboard the development of novel, personalized approaches to treat disease.