Massively Parallel Reporter Assays: Defining Functional Psychiatric Genetic Variants across Biological Contexts.

Abstract

Neuropsychiatric phenotypes have long been known to be influenced by heritable risk factors, directly confirmed by the past decade of genetic studies which have revealed specific genetic variants enriched in disease cohorts. However, the initial hope that a small set of genes would be responsible for a given disorder proved false. The more complex reality is that a given disorder may be influenced by myriad small-effect noncoding variants, and/or by rare but severe coding variants, many de novo. Noncoding genomic sequences—for which molecular functions cannot usually be inferred—harbor a large portion of these variants, creating a substantial barrier to understanding higher-order molecular and biological systems of disease. Fortunately, novel genetic technologies—scalable oligonucleotide synthesis, RNA sequencing, and CRISPR—have opened novel avenues to experimentally identify biologically significant variants en masse. An especially versatile technique resulting from such innovations are Massively Parallel Reporter Assays (MPRAs), powerful molecular genetic tools that can be used to screen ≥thousands of untranscribed or untranslated sequences and their variants for functional effects in a single experiment. This approach, though underutilized in psychiatric genetics, has several useful features for the field. Here, we review methods for assaying putatively functional genetic variants and regions, emphasizing MPRAs and the opportunities they hold for dissection of psychiatric polygenicity. We discuss literature applying functional assays in neurogenetics, highlighting strengths, caveats, and design considerations—especially regarding disease-relevant variables (cell type, neurodevelopment, and sex), and ultimately propose applications of MPRA to both computational and experimental neurogenetics of polygenic disease risk.

Introduction

Psychiatric diseases are genetically influenced by both heritable variation (common and rare) and non-inherited, de novo mutations. Estimated common variant (frequency ≥ 1%) influence on disease liability ranges from 10–33% for major depressive disorder (MDD) (1–3) and schizophrenia (SCZ) (1,4,5) to over 50% in autism spectrum disorders (ASD) (6,7). The remaining familial heritability of psychiatric—especially neurodevelopmental and psychotic (8,9)—diseases is largely conferred by rare variants (7). Two major hurdles have prevented variant data from illuminating disease mechanisms: the volume of variant discoveries/associations to test for functionality and causality, and imperfect methods of predicting variant consequences.

Variant-disease associations arise from correlational methodologies. Genome-wide association studies (GWAS) identify overrepresented single nucleotide polymorphisms (SNPs), tagging hundreds of mostly untranscribed, linked SNPs (7). Similarly, family studies identify thousands of proband-specific (de novo) or -enriched (rare, inherited) variants, though 0–2 per patient may be causal. However, discovery-oriented approaches alone are incapable of specifying which variants have biological function.

Predicting whether and how noncoding variants are functional is a nontrivial enterprise. The majority of GWAS loci bear indirect indication(s) of transcriptional regulatory function, including expression quantitative trait locus (eQTL) associations, chromatin accessibility, or histone marks (10–12). As others have noted, these data alone cannot define functional regulatory elements/variants (13,14). However, even within one cell type, such data are often mutually discordant: an emerging (i.e., preprinted) study examining six epigenomic datasets from K562 cells showed 49% of functional regulators did not overlap any epigenomic annotations; another 40% only overlapped one of the six (15). Similarly, MPRA of chromatin-based K562 enhancer predictions found only 30% regulated transcription (16). Unsurprisingly, these discrepancies apparently extend to disease variant interpretation: only a minority of GWAS variants (except for blood traits) overlap tissue-specific regulatory predictions (17) from histone marks (18). Such findings collectively suggest that heritable, disease tissue-specific regulatory phenomenon are both missed and mislabeled when relying solely on chromatin states.

Despite the clear excess of de novo variation in coding sequences in ASD and other neurodevelopmental disorders, and though coding variant consequences can often be predicted (e.g., nonsense mutations), this constitutes the minority of heritable risk for several psychiatric diagnoses (19). The remaining burden falls within putative transcriptional (19) and translational regulatory elements (e.g., promoters, UTRs) (20,21). ASDs provide a representative case: among 1,902 subjects, over 255,106 de novo variants were recently identified, with thousands each in upstream/promoter sequences and untranslated regions (UTRs) (22). UTRs regulate transcript stability and miRNA interactions (23); emerging work further implicates UTRs in nuclear transcript trafficking in the brain (24). The occurrence of most disease-linked variation in the least-well understood features of the genome/transcriptome thus obstructs understanding of disease biology. Collectively, these two problems necessitate high-throughput assays with functional readout for putative regulatory elements and variants. Such assays enable identification of functional variants and the biological contexts in which they act. This knowledge can shape hypotheses regarding shared mechanisms by which disparate genetic factors converge on shared phenotypic endpoints.

Here, we will primarily discuss MPRAs for high-throughput parcellation of genetic discoveries. MPRA technology pairs genomic features (e.g., each allele of a genomic sequence) to a reporter gene bearing unique, transcribed barcodes, allowing multiplexed RNA-level readout of element activity (25,26). Critically, there is substantial potential for MPRAs to identify functional variants from neuropsychiatric-associated loci. In part one, we discuss uses of MPRAs in functional identification of gene regulatory elements and variants, design/interpretation considerations for MPRA, and methods to complement/follow-up findings. In part two, we discuss potential applications of MPRA to identify mechanistic convergence across polygenic risk space.

Part 1: MPRAs for Identification of Functional Regulatory Elements and Variants

MPRAs offer a flexible framework to study elements regulating transcription (e.g., enhancers, promoters), splicing, protein translation, and post-transcriptional phenomenon. Though too numerous to review deeply here, we point readers to published and emerging applications of MPRA to splicing (27–29), RNA editing (30), and protein translation (31). MPRAs have been most broadly applied to explore and computationally model transcriptional “regulatory grammar”—how sequence features such as binding motifs, their abundance, and arrangement affect regulatory capacity (16,32–38). More recently, these approaches have been applied to characterize UTR functions in RNA stability and translation (39–43), and to identify SNPs and rare variants influencing transcription (44–50).

As shown in Figure S1A–B, a canonical ‘enhancer’ MPRA utilizes a promoter with candidate elements either upstream or in a 3’UTR (STARRseq) (51). Each element is paired with unique barcodes in the transcribed UTRs, which are sequenced as quantitative readouts. Expression—representing transcription or RNA stability—is measured as the RNA barcode counts per DNA barcode count (Figure S1E). This measure can be leveraged to define active or differentially active enhancer elements. Functional elements have been defined by either comparing to minimal-promoter only barcodes (16,34,37,38,52,53), or individual sequences against their shuffled counterparts (32,38); MPRAs have also successfully compared activity between alleles (44,45,47–50,54).

MPRAs also enable study of post-transcriptional regulatory elements. As shown in Figure 1E–G, the RNA/DNA expression metric assess UTR effects on transcript stability. Published and emerging UTR MPRAs have not yet considered human variants directly, but have distinguished functions of ASD/ID-implicated CELF proteins and related RBPs (42,55,56), and defined 3’UTR (39–41) and 5’UTR (43,57) features influencing transcript stability and translation. Both UTR and enhancer MPRAs enable assessment of disease-associated variant function across model systems (Figure 1A–C; Figure S2).

Figure 1 |. Example Allele-Differential Phenomenon in Common MPRA Approaches, and Analysis of MPRA Data.

A) In a transcriptional-regulatory assay, a putative regulatory SNP may create, ablate, strengthen, or weaken a TF binding site. As a result, one allele drives more transcription (detected via its 3’UTR barcodes) per encoding DNA than the other allele. B) In a 5’UTR assay, a functional SNP may sequence features controlling translation initiation. For example, a variant allele may introduce an upstream start codon out of frame with the reporter gene, resulting in nonsense mediated decay, and thus, decreased detection of the barcodes paired to that UTR allele. C) In a 3’UTR assay, a variant may alter an RBP binding site; in this example, an RBP site specific to one allele increases the stability of the reporter transcript, and thus of the barcode paired to it. D) After transfection/transduction, RNA is collected from specimens and prepared along with DNA (often the delivered DNA, though sometimes this is recovered from the specimens as well) to generate sequencing libraries to quantify expression of the delivered elements in the RNA, compared to starting abundance in the DNA. E) Example read counts, presented visually, for the DNA and RNA barcode counts of one barcode paired to each allele. F) MPRA analysis centers on taking the ratio of RNA/DNA counts (or counts per million), represented by the sequence fragments at top left, as a measure of expression—i.e., approximating the number of transcripts generated per encoding DNA. These can be compared relative to the expression of all elements to find the strongest features (e.g., strongest enhancers and repressors, or most stabilizing and destabilizing UTR elements), or G) compared on a variant-wise basis to determine significant allelic regulatory effects.

MPRAs Identify Functional Elements in Specific Cellular Contexts

Perhaps the most exciting—if underappreciated—property of MPRAs is the ability to assay elements using disease-relevant cells and conditions. Functional elements are defined by each cell type’s unique milieu of expressed TFs, chromatin modifiers, miRNAs, and RBPs, which mediate regulatory element activity. The breadth of published and emerging tissue/cell type differences in gene expression (58,59), chromatin marks (18,60) and chromatin interactions (61–63) all illustrate the magnitude of these regulatory differences. The importance of cell type was illustrated by an MPRA of the same elements in U87 glioblastoma and neural progenitor cells (NPCs): the most active enhancers in each cell type contained entirely different motifs and sequence features (37). Recent (64,65) and emerging (66) approaches have identified highly cell type-specific brain enhancers using adeno-associated viral (AAV) vectors alongside traditional (e.g., immunofluorescent) readouts. Moreover, a novel, AAV-based MPRA (i.e., using barcodes) identified novel functional enhancers for somatostatin interneurons (67). Aside from these examples, MPRAs in neural cells have been limited to date. Several early MPRAs utilized explanted retina to explore influences of TF binding sites and their arrangements on activity (38,48,53,68). One novel study, relevant to functional contexts (discussed below), assayed mouse neuron enhancers for activity changes following KCl depolarization (69). Other studies include an MPRA characterizing temporal patterns of cis-regulatory element activity across seven timepoints in human NPC differentiation into neurons (70). This delineation of regulatory element function illustrates the power of regulatory assays to reveal timepoints and cellular states wherein gene regulation—especially for neurodevelopmental disorders—may exert its causal effects.

In vivo regulatory assays—including in the brain—have more recently been demonstrated, generally at smaller scales than in vitro MPRAs. Osterwalder, et. al (71) singly or multiply knocked out putative limb development enhancers in mice, illustrating enhancer redundancy—that is, limb development disruption only with perturbation of multiple elements. McClymont et. al (72) identified 2,000 candidate embryonic mouse enhancers in purified midbrain dopamine neurons, and validated the developmental and regional specificity of a subset using transgenic reporter mice. The scale of these assays has been expanded by groundbreaking implementation of MPRA in the brain in vivo (48,67) to query functional effects in native cell contexts.

This transition to in vivo MPRA is beneficial because, while cell type overwhelmingly influences regulatory assays, additional conditions may equally alter outcomes (Figure 2A–C). Age, sex, pharmacology, and environment (e.g., stress)—all can shape gene expression. For example, MPRAs have identified elements responsive to hormonal contexts such as steroid-responsive glucocorticoid receptor binding (73). Altogether, MPRAs enable identification of functional regulatory elements across varied internal and external environments.

Figure 2 |. Regulatory assays are influenced by a range of conditions, from environment to sequence context.

The range of conditions that influence regulatory assays (from top to bottom) starts when considering A) the environment, e.g., sex, time, and pharmacology. These parameters have the potential to affect various –omic profiles in a given system. B) The next level of consideration is the organism, which can include human-derived tissue or one of the many model organisms. Human genomic context is ideal for studying the biology of human disease – though a comparatively limited scope of techniques for human-derived tissues exists. C) Next, one should consider the selected cell type(s) and whether to assay in vitro or in vivo. Each of these provides a unique set of benefits, and one approach can be used to validate findings from the other (45,74). In the case of modeling the brain and psychiatric genetic variants, cell type-specific/enriched MPRAs in vivo would constitute the highest-fidelity model of variant effects by accounting for regulatory effects of endogenously interacting cell types. D) Lastly, the sequence context will be influenced by the delivery method, which results in transcription from extragenomic or intragenomic MPRA DNAs. In either case, only limited length of sequence surrounding a feature of interest is preserved (e.g., in ~120bp tiles of genomic sequence in custom oligonucleotide cloning, or ≤ 1kb in clone-and-capture methods), preventing assessment of any interactive effects from elements further away. (A recent study suggests that size of a tile negatively correlates with reproducibility of expression driven compared to that driven by ~120bp tiles, emphasizing the importance of this consideration (93)). While AAV-transduced episomes gain histones (105) and chromosome-like nucleosome spacing (106), it is unknown whether gene-regulatory histone marks on these episomes mirror those of endogenous regulatory chromatin. For these reasons of both local sequence context and chromatin context, we suggest corroboration of MPRA findings in native genomic settings, by, for example, introducing the variant to the genome of a cell line using CRISPR methods.

MPRAs, assay context, and functional variants

MPRAs can be designed not only to identify functional elements, but to assay and compare genetic variants in contexts known—or predicted—to mediate disease. As transcriptomic and epigenomic studies highlight an enormous role for cell type, it is unsurprising that this influence extends to regulatory variants. For example, variants exert cell type-dependent effects on chromatin structure even within a neurodevelopmental lineage: an emerging study discovered chromatin accessibility QTLs in human NPCs and neurons, with ~80% of QTL SNPs being specific to one of the cell types (74). Cell type roles in putatively functional variation are also implicated by GWAS SNPs enrichment in tissue-specific eQTLs (59,75), neural cell type-specific chromatin interactions (62), and eQTLs that evade detection in bulk brain (i.e., multi-cell type) tissue but are evident in purified populations like dentate granule neurons (76). MPRA has likewise demonstrated the essentiality of cell type in defining functional variants: the Critical Assessment of Genome Interpretation 5 (CAGI5) consortium performed an MPRA on saturation-mutagenized human enhancers and disease-associated promoters in numerous cell lines, challenging analysts to computationally predict functionality and effect size for held-out variants. The most predictive annotations for a given cell line were often from the same cells across several top-performing analyses (77). Thus, experimental study of putative disease-associated variants requires firm hypotheses on where (tissue/cell type), when (development/differentiation), and how elements are expressed/active and biologically relevant. Careful consideration needs to be given to the appropriate cellular context when designing assays for psychiatric genetics: key variant-interacting TFs and RBPs expressed in neurons may not be present in convenient cell lines (e.g., K562), potentially rendering functional neural elements/variants apparently silent.

Despite their potential, few MPRAs have examined disease-associated variants while considering both cell type and –omic predictions. Tewhey, et. al (49) screened 30,000 eQTL SNPs from human lymphoblastoid cell lines (LCLs) using MPRA in LCLs, maintaining the discovery context in their assay. Over 3,400 active regulatory sequences were identified, including 850 activity-modulating variants (24%), consistent with functional (expression-modulating) SNP associations tagging linked, non-functional SNPs, akin to GWAS. Illustrating MPRA’s sensitivity, significant allelic differences in activity were detectable at effect sizes <2-fold. Emerging work by Choi, et. al (78) prioritized over 800 SNPs—guided by fine-mapping and epigenomics—from 16 melanoma GWAS loci, to assay for transcriptional-regulatory activity in cultured melanocytes. Candidate variants with concordant eQTL signal in independent melanocyte data were further investigated, ultimately enabling experimental demonstration of biophysical (TF binding), molecular (target gene expression), cellular (growth rate), and in vivo (melanoma rate in transgenic zebrafish) variant mechanisms. Finally, a recent MPRA of autoimmune GWAS loci yielded replicable findings across 12 donor lines of CD4+ T-cells, which were discordant with the more easily accessible—but leukemic—Jurkat cell line (79). These experiments exquisitely illustrate MPRA’s capacity for sensitivity, context specificity, and high discovery rates, especially when integrating both association data and multi-omic annotations.

As with functional element assays, functional variant assays have recently moved in vivo, again including the brain. Kvon, et. al (80) utilized a novel knock-in system to generate transgenic mice expressing a LacZ reporter expressed under putatively regulatory elements containing rare, polydactyly-associated variants; subsequent LacZ staining clarified which variants were functional based on alterations of limb bud LacZ patterns. Excitingly, a small-scale MPRA has recently emerged using in vivo tissue: after prioritizing a single SNP from a bipolar disorder GWAS locus using epigenomic annotations, the two alleles of this sequence region were paired to 20 barcodes each and electroporated into embryonic mouse brains to confirm variant function (48).

Limitations and Design Considerations

With the powerful opportunities of MPRA come limitations. A major caveat lies in gathering candidate variants to assay. For example, a prominent and functionally characterized schizophrenia GWAS locus in the major histocompatibility complex (MHC) region (4)—containing hundreds of linked SNPs—turns out to mark heritable copy number variations in the complement C4a gene (81); assaying only SNPs from this locus would not reveal the primary causal variant. Likewise, an MDD-associated SNP tags the absence of a transposon with regulatory effects on a noncoding RNA (82). In other words, MPRA’s usefulness is contingent on investigation (and size—see below) of sequences to be assayed.

Further considerations include appropriateness of biological ‘contexts’ (Figure 2). At the level of ‘sequence context,’ MPRAs generally use multiplex oligonucleotide synthesis to custom-design sequences and variations by the thousands. However, such approaches are size-limited to ~300bp, which precludes assay of large or spaced regulatory sequences. Oligonucleotide synthesis also is error-prone; tagging each element with multiple barcodes safeguards against error-driven false-positives. Bulk capture-and-clone strategies circumvent these issues by utilizing larger, genomic DNA fragments directly (47,83–85) at the expense of precision assay design. Finally, element positioning can substantially influence results and replicability. While STARR-seq is favorable for one-step cloning (putative enhancers doubling as 3’UTR barcodes), emerging works illustrate that enhancer-like sequence placement in 3’UTRs yields results which cluster separately from other MPRA designs testing the same sequences (86), and that such sequence placement can exert RNA stability effects that, without correction, may confound interpretation (87).

Reporter gene features are also important in regulatory assays. Previous enhancer MPRAs have demonstrated replicability by testing the same elements with a second promoter, with element activities highly correlated between the two (45,88,89). However, these cross-promoter correlations (Pearson r 0.7–0.8) have been weaker than often reported for biological replicates in MPRA (r>0.9). Promoter choice thus can influence assay results, via, for example, absence—or species differences in—promoter elements a cis-regulator requires. Likewise, UTR regulatory elements may be sensitive to the stoichiometry of transcripts and RBPs or miRNAs in the cell; excess transcript production by a strong promoter could potentially render effects of interacting regulators undetectable. In brief, rigorous MPRAs or follow-up assays should use both a minimal promoter and either a strong exogenous (e.g., CMV) or a genomic promoter from the pertinent cell/tissue type (e.g., a constitutively expressed, neuron-specific gene).

Importantly, the ability to test candidate sequences in their endogenous locus is not a feature of MPRAs. Thus, ‘genomic context’—that is, episomal (AAV, plasmid) vs. genome-integrated (lentiviral) approaches—require consideration. Emerging comparisons find these approaches correlate well (86), though certain applications may require a specific approach (e.g. MPRA of chromatin conformation (90)). The comparative throughput for a fixed number of cells is greater for plasmid transfection—thousands of plasmids per neural cell in vitro (91) compared to viral transduction (≤ tens of sequences/cell). These limitations and alternative methods are further considered in Table 1.

Table 1 |. Strengths and Limitations of Functional-Regulatory Assays in Terms of Throughput and Sequence and Cellular Contexts.

Method family: An umbrella term covering multiple adaptations of an assay. Technique: The particular adaptation of the family’s assay. “CRISPR editing” signifies precision replacement of an endogenous genomic sequence with a desired sequence (as opposed to CRISPR mutagenesis). Largest target/sequence assayable: Largest sequence currently supported by viral vectors (AAV assays), pooled oligonucleotide synthesis (MPRA), or guide RNAs (CRISPR), respectively. Simultaneous throughput for variants/perturbations per sample: Based on literature cited in the text and in this table, the largest range of elements that have been simultaneously targeted or tested within experimental replicates for the technique. Can assay cellular phenotypes? For CRISPR assays, the phenotype is generally cell division or death, measured by enrichment/depletion of genome-integrated sgRNAs in a cell culture relative to the pool of sgRNAs originally transfected. Genome-integrated? Does/can the assay utilize genome-integrated sequences (i.e., lentiviral delivery)? Assays at the endogenous genomic sequence? Whether the assay tests the putative regulatory sequence in its standard location in the genome (rather than testing a short piece of genomic sequence elsewhere in the genome or outside of the genome entirely). Can identify target gene of endogenous transcriptional regulator? Whether the assay allows for pairing of a functional element to its target gene(s). This is effectively contingent on testing the element in its endogenous genomic position.

Method Family	Method	Can assay variants (e.g. SNPs) for function?	Can assay elements (e.g., TFBS, enhancers) for function?	Largest sequence / target	Simultaneous throughput for variants / perturbations per sample	Can assay cellular phenotype?	Genome-integrated?	Assays at the endogenous genomic sequence?	Demonstrated in model organism CNS in vivo?	Demonstrated in human primary or iPSC-derived neural stem cells, NPCs or neurons?	Can identify target gene of endogenous cis-regulator?
MPRA	Multiplex (Barcoded) AAV Transcriptional Regulator Assays	Yes	Yes	3–5 kb	100s-1000s	No	No	No	Yes (67)	No	No
	Enhancer MPRA and STARR-seq			~150–200 (custom oligos); ~700 (capture-and-clone)	10,000–10^6		Often not (exception: lentivirus random or targeted integration)		Yes (48)	Yes (48,95,97)	No
	3’UTR MPRA/PTRE-seq								Not demonstrated but see above	No	N/A
	5’UTR MPRA								“	No	N/A
	RNA Splicing MPRA								“	No	N/A
	Protein Translation MPRA								“	No	N/A
CRISPR Regulatory Disruption Assays	CRISPRi	No	Yes	~50 bp	•Max demonstrated in CNS in vivo: 5 targets (2 sgRNAs each) •Max demonstrated in neural cell types in vitro: ~12,000	Yes	Yes	Yes	Yes (136)	Yes (107,137,138)	Yes
	CRISPRa				•Max demonstrated in CNS in vivo: 10 targets (5 sgRNAs each) •Max demonstrated in neural cell types in vitro: 3	Yes	Yes		Yes (139–141)	Yes (113,138,142,143)	Yes
	CRISPR Mutagenesis of regulatory elements				•Max demonstrated in neural cell types in vitro: 26,000 targets (2 sgRNAs per target)	Yes	Yes		No	Yes (104)	For an a priori defined gene (144,145)
	CRISPRi-FlowFISH				~900	Not demonstrated	Yes		No	No	Yes
Low/single throughput	CRISPR Editing	Yes	Yes	Several kb	1–2	Yes	Yes	Yes	Yes	Yes	Yes
	AAV Transcriptional Regulator Assays with traditional readouts (fluorescence, LacZ, etc.)			3–5 kb	1	No	No	No	Yes (esp. tacitly via cell-type targeted optogenetic, chemogenetic, and circuit-labeling techniques, as in (99))	Yes (99)	No
	Luciferase Reporter Assay			3–5 kb	1	No	No	No	Yes (146)	Yes	No

Other considerations include determining an appropriate ‘cellular’ and ‘organismal’ context (Figure 2B–C). Common strategies for choosing cellular contexts include using pathology (e.g., substantia nigra in Parkinson’s disease), expression patterns of disease-associated genes (e.g., cortical excitatory neurons in SCZ (92)), or GWAS-eQTL overlaps (e.g., neurogenic niches of mid-fetal brain in ASD and SCZ (93)). Cell type prioritization is further covered elsewhere in this Special Issue (94).

A notable opportunity is utilization of MPRAs in human iPSC-derived neural cell types, which offer the ability to conduct cell type-specific assays in a human genetic context. Very recent (70) and emerging (48,95) MPRAs are proof-of-principle for this approach, supporting advancement to assaying variants in the setting of iPSC derivates. Moreover, while cell type-specific MPRAs have been restricted to in vitro settings, where reproducing tissue physiology (e.g., inter-cell type interactions, hormones, stress) is difficult, barcoded multiplex AAV regulatory assays (67) indicate in vivo, cell type-specific MRPA is possible. Nonetheless, negative MPRA results should be interpreted cautiously; absence of function in one context may not extend to all contexts.

Statistical considerations in MPRA include appropriate library size (number of elements and paired barcodes) for the cell type to be tested. Generally, library size should be downsized for rarer, hard-to-maintain, or hard-to-transfect/transduce cell types to ensure robust barcode recovery and measurement. MPRAs have tested ~10⁷ sequences simultaneously in easily transfected cancer cell lines (84,85), though in physiological cell types, like NPCs, this capacity is 10⁴-10⁵ (37,69,96), with emerging work approaching 10⁶ (95,97). Library size is further constrained by element-per-cell (i.e., lentiviral) approaches. In other words, the fidelity of the model system and the MPRA library size it can support are generally anticorrelated. A consensus on the depth of barcodes-per-element is, to date, absent, ranging from 1 (STARR-seq (51)) to several hundred in previous (70) and emerging (97,98) work, with highly correlated replicate measurements across this range. Tewhey, et. al estimated that statistical benefits for small-effect transcriptional-regulatory variants accrue by 5 barcodes, and asymptote around 25–50 (49); another finds > 10 barcodes consistently yield inter-replicate r>0.97 in several cell types (86). Whether these guidelines apply to UTR assays remains unclear. Overall, MPRA power guidelines would benefit substantially from deep assessment by modelers and statisticians.

Finally, given a finite number of elements that can be simultaneously assayed, one can choose whether to prioritize candidate variants using epigenomic data, or simply include all linked SNPs (Figure S2). An assay’s ‘hit rate’ may be improved by prioritizing variants with indirect evidence of function, with the caveats of relying on epigenomic data discussed previously. However, foregoing such prioritization enables analysis of how well such features actually predict measured expression. Thus, the decision of prioritization must balance the value of ‘hits’ vs. identifying functionally predictive indirect measures (epigenetics) for the target cell type or disease.

Complementary methods in high-throughput study of DNA and RNA regulatory elements

There are a variety of other approaches that complement MPRAs (Table 1). Of course, lower throughput enhancer assays allow screening of the same elements or variants across a variety of contexts, even in vivo. Whether conducted using AAV (e.g., (99)), or transgenesis (e.g. (100)), these should remain gold standard approaches for validation and deep characterization of small numbers of elements and variants, including those identified by MPRAs.

A primary limitation of MPRA is the inability to test regulators in their endogenous genomic position and sequence context. Sequence-specific targeting using CRISPR/Cas9 has enabled several additional techniques for probing molecular and cellular effects of regulatory variation, with the caveat that, unlike MPRA, these techniques do not currently allow for the multiplexed study of post-transcriptional/translational regulatory variants. Nonetheless, these techniques enable study of putative disease gene roles in gene expression networks and cellular phenotypes. Perturb-seq (101) combines genewise perturbation by CRISPR with single-cell RNA-seq to identify gene sets dysregulated by loss of function of each candidate gene. These have, for example, been used to discover co-transcribed gene networks involved in neuronal remodeling (102) and for in vivo assessment of genes harboring de novo loss of function mutations in ASDs (103). Likewise, CRISPR screens can be used to define functional elements influencing selectable traits (e.g., proliferation), as in an emerging study perturbing both genes and cis-regulatory elements to define their roles in human neural stem cell proliferation (104). Finally, CRISPR editing has been used in vitro to assess single-transcript noncoding variant effects by comparing allelic RNA and genomic DNA abundances in edited cultures (105), a potential means of single-variant validation/follow-up of UTR MPRA findings. To our knowledge, such assays have not been conducted at a genome-wide scale in psychiatric disease, but have been used to identify genes that alter expression of the Parkinson’s-associated PARKIN (106).

Cis-regulatory MPRAs cannot identify the endogenous target gene(s) of functional elements. Fortunately, CRISPR-derived methods using a mutagenically-‘dead’ Cas9 (dCas9) conjugated to a transcriptional activator or repressor allow targeted potentiation or repression of endogenous genomic regulatory elements (CRISPRa and CRISPRi, respectively) to assess altered gene expression and other outcomes. These technologies are already online in state-of-the-art human neuroscience models: a recent CRISPRi study knocked down over 2000 genes by targeting their promoters in iPSC-derived excitatory neurons, defining their context-specific roles in their survival, differentiation, and proliferation—including gene effects altered by co-culture with astrocytes (107). Emerging work has further leveraged CRISPRi’s cell type specificity to study ASD-associated gene knockdown effects in an etiologically relevant cellular context (NPCs) (108). A recently introduced extension of CRISPRi (‘CRISPRi-FlowFISH’) targets intergenic regulators, identifying their target gene by concurrent fluorescent in situ hybridization against genes from the same chromodomain. Fluorescence-intensity sorting into bins and subsequent RNA-seq can then pair regulators (via guide RNA sequence) and target genes (altered FISH signal in a guide RNA’s presence) (109). While this assay was performed in K562 cells, it is not hard to envision its extension to neural cell types in vitro or in vivo. Altogether, CRISPR-based follow-up of MPRA candidates to define target genes and verify of genomic activity of regulators/variants will be key to developing insights in psychiatric genomics.

Part 2: MPRAs as an avenue to dissect multiallelic and polygenic mechanisms of neuropsychiatric traits

While MPRAs cannot intrinsically scale up to functional demonstration of cell-, tissue-, or behavior-level phenotypes, they have the potential to provide key information to guide molecular hypotheses for how these higher-order phenotypes emerge from large sets of regulators and/or their target genes. We focus here on examination of variants in trans space—that is, defining shared and recurrent features among MPRA-nominated functional variants across the genome that may collectively underlie large portions of polygenic disease risk. A brief examination of functional SNP interactions within linkage blocks can be found in the Supplemental Text.

The utility of MPRAs in identifying commonalities from variants across the genome

The most vexing question that remains after individual functional variant mechanisms are elucidated is how variants collectively contribute to phenotypic risk. MPRAs provide several ways to begin addressing this question: 1) identifying regulatory features shared by across several functional risk variants; 2) identifying functional modules enriched for variant-impacted genes; 3) providing functional annotations to variants for computational genomic approaches; and finally, 4) by enabling study of variant-by-environment interactions by performing MPRA across conditions.

Firstly, MPRA experiments running the gamut from basic regulatory genomics to human traits and variation have defined ‘regulatory grammars’ of assayed contexts. Identification of functional variants in the MPRA setting enables similar establishment of the ‘regulatory grammar’ of a trait or disease. Functional variants identified by MPRA across several UTRs may feature a specific RBP’s binding site, for example, or could be used to deliberately define functional activity of a disease-associated miRNA, like miR-137 (4). Likewise, variants associated with a trait could be more likely to fall in particular TF binding sites or be enriched in cell type-specific marks of genomic regulation. Evidence of this convergence is seen in de novo variants associated with ASD: several distinct variants disrupt binding sites for a single TF, NFIX (110). Similarly, putative gene targets of schizophrenia-associated variants are also putative—biases aside (111)—Fmrp targets (112). MPRA has also identified such regulatory convergence by, for example, intersecting identified functional SNPs with TF ChIP-seq datasets in pertinent cell types to discover recurrently disrupted TF binding sites (98). Assays of downstream consequences of variation also confirm biological convergence across association loci. A four-element-target CRISPRi/a assay revealed that schizophrenia risk genes act synergistically via shared influence on synaptic activity, and concurrent alteration of expression of all four genes results in a cellular transcriptome more accurately reflective of postmortem schizophrenia brain tissue (113). For both rare and common variants, identifying common regulators among risk genes provides information which can refine predictions of disease-related cell types based on TF, RBP, or epigenomic mark expression.

Secondly, genes and gene networks affected by statistically associated variation are often predicted using MAGMA (114), which in essence scores genes based on proximity to an associated variant and its linkage partners. Resulting gene sets are subjected to analyses such as Gene Ontology enrichment or are examined for enrichment in WGCNA (coexpression) networks from candidate tissue types to identify pathways and mechanisms on which these genes converge. While its use is ubiquitous in genomic studies, standard MAGMA gene association statistics for psychiatric disorders only modestly correlate to those from a tissue-specific, chromatin configuration-aware modification of MAGMA (115), suggesting that biological hypotheses from MAGMA gene sets may miss disease-associated genes in brain. Being able to refine implicated genes by functional validation using—or in follow-up to—MPRA will help to benchmark such approaches and refine prediction convergence with ‘truly’ dysregulated candidate genes.

Thirdly, epigenomic data alone is not comprehensively predictive of active regulators. However, well-informed analyses of human genetic findings rely heavily on such annotations to convert associations into biological hypotheses. Critically, these epigenomic data—unlike MPRA data—can be collected from postmortem human tissue. MPRAs focused on neuropsychiatric disorder associated variation stand to benefit from high-information datasets by aiding variant prioritization for assay inclusion. Several recent datasets on synthetic UTRs (39,43), RNA binding proteins (116,117), and postmortem human brain multi-omics (60,118–127) are worth noting for readers investigating disease-associated variation. Integrative computational analyses have brought these datasets together predict functional variation in SCZ, bipolar disorder, and ASDs (128,129). These constitute high-priority candidates for experimental validation by MPRA. Furthermore, emerging work reveals a symbiotic relationship developing between epigenomics and functional assays: functional element/variant information from MPRA has been used alongside epigenomic annotations to improve machine learning predictions of functional variants (130). Predictions from these refined algorithms are another low-hanging fruit for candidates to assay by MPRA; those results could then constitute new training data. Such refinement of epigenomic data interpretation coupled with functionally-demonstrated regulatory variation would mutually benefit one another and myriad downstream analyses, such as variant enrichment in genomic features and disease gene identification. For example, TWAS (131) and Predixcan (132) intersect gene expression QTLs (eQTL) with trait-associated variants to predict expression differences between cases and controls, thus identifying dysregulated gene sets. MPRA data can disentangle which eQTL SNPs are truly functional from those associated only due to LD, which could thus refine variant-gene pairings used in these analyses. Altogether, MPRA can serve to refine both epigenomic and genic definitions of truly causal disease features.

Finally, the context-specificity of MPRA represents a newfound ability to assess variant effects on gene regulation en masse under different biological and environmental contexts, including with in vivo models. While issues of convergent disease effects across genes and regulators are indeed complex, environmental effects—perhaps most canonically, stress—on these regulators are questions at the forefront of understanding polygenic risk in neuropsychiatric disorders. Pharmacologic variables have been successfully tested in MPRA, namely in the identification of glucocorticoid-responsive (73) and p53-responsive (133) regulatory elements. MPRAs could further be layered with concurrent gene perturbations (e.g., knockdown of a putative regulator), or cell culture conditions for in vitro identification of variant-environment interactions, exemplified by MPRA identification of neuronal activity-dependent enhancers (69). As mouse and human brain cell types and their gene expression patterns are largely (though not entirely) conserved both in development (134) and adulthood (135), the extension of MPRAs to the in vivo context will enable study of broader endogenous and exogenous disease-associated factors, such as sex or stress. Identifying variants with environment-dependent functions would be a start toward identifying convergent molecular mechanisms behind conditional disease risk in disorders such as MDD.

Conclusion

MPRA presents unique opportunities to dissect polygenicity of psychiatric disorders via simultaneous identification of functional variants across identified risk space. Beyond the primary benefits of identifying ‘true positive’ functional variants in specific biological and environmental contexts, MPRAs stand to rapidly broaden, deepen, and refine hypotheses and mechanisms of both noncoding disease risk and of gene-regulatory architecture itself.