Multiplexed functional genomic assays to decipher the noncoding genome

Abstract

Linkage disequilibrium and the incomplete regulatory annotation of the noncoding genome complicates the identification of functional noncoding genetic variants and their causal association with disease. Current computational methods for variant prioritization have limited predictive value, necessitating the application of highly parallelized experimental assays to efficiently identify functional noncoding variation. Here, we summarize two distinct approaches, massively parallel reporter assays and CRISPR-based pooled screens and describe their flexible implementation to characterize human noncoding genetic variation at unprecedented scale. Each approach provides unique advantages and limitations, highlighting the importance of multimodal methodological integration. These multiplexed assays of variant effects are undoubtedly poised to play a key role in the experimental characterization of noncoding genetic risk, informing our understanding of the underlying mechanisms of disease-associated loci and the development of more robust predictive classification algorithms.

Introduction—Challenges of interpreting noncoding genetic variation

Since the early 2000s, the rapid development and plummeting costs of genotyping arrays and next-generation sequencing (NGS) technologies have revolutionized the fields of genetics and genomics (1). The widespread deployment of genotyping, whole exome and whole genome sequencing studies in the research, clinical and direct-to-consumer spaces have enabled the identification of millions of genetic variants across myriad individuals and ancestral populations as well as the de novo somatic mutational landscape in cancer (2–4). The interpretation of this vast and rapidly expanding catalog of human genetic variation has emerged as a fundamental problem in modern genetics. Understanding the relationship between genotypic and phenotypic variance is critical for elucidating the molecular basis for gene expression, identifying the genetic contribution to diseases, and providing genetic diagnosis and risk stratification in the clinical setting.

The biological interpretation of gene-mapping studies remains a major challenge. In genome-wide association studies (GWAS), linkage disequilibrium (LD) within trait-associated loci obscures the underlying causal variants among the many correlated polymorphisms within the locus, the majority of which are expected to be functionally neutral (5). Additionally, the vast majority of loci and variants identified by GWAS are found within non-protein coding regions, which comprise roughly 98% of the human genome and are under relatively less selective constraint than protein coding regions (6). Careful statistical partitioning of heritability has repeatedly confirmed that the majority of heritability for common disorders is harbored within the noncoding genome, particularly within enhancer domains and intronic splice sites (7,8). Similarly, the bulk of variation identified in NGS studies is found in the noncoding genome, and interpretation of the millions of noncoding rare variants of likely small effects remains notoriously difficult. The contribution of noncoding variation to rare and Mendelian disorders also remains poorly understood, although the existence of mutations that focally disrupt enhancers (9), along with the many mutations known to act via haploinsufficiency, suggests that regulatory mutations affecting gene dosage are more common than currently appreciated. Similarly, examples of single noncoding mutations associated with severe developmental disorders (10) and an expanding array of noncoding driver mutations found in cancer (11,12) demonstrate the broader importance of interpreting noncoding variation in the clinical setting. In contrast to protein altering mutations, noncoding variation maintains unclear functional relationships with putative target genes and may not act within all cell types or tissues (challenges summarized in Table 1). Therefore, high-throughput functional interpretation of noncoding genetic variation is one of the major challenges facing modern genetics (13).

Table 1.

The major challenges in interpreting noncoding loci and variants identified in GWAS or NGS studies, and proposed methodological solutions to address these challenges

Challenge	Exacerbating Variables	Methodological Solutions
Is the variant causal?	- Linkage disequilibrium- Low power/small effect sizes- Genotyping/imputation failure	- Fine-mapping- Deep imputation with improved population reference panels, increased sample sizes
Is the variant functional?	- Same as above- Uncertain function of noncoding genome	- Prioritization algorithms- Functional genomic annotations - Genome editing/in vitro modeling- MPRA
Which gene is regulated?	- Gene dense regions- Strong LD- Complex 3D genomic architecture	- QTL colocalization - 3D interactome assays - Genome editing/in vitro modeling
Which cell-type mediates risk?	- Trait in heterogenous tissue, or unclear causal organ system- Gene dense regions	- Heritability enrichment by cell-type - Transcriptomics/Proteomics- Genome editing/in vitro modeling

Interpretation of noncoding genetic variation using functional genomic maps

Noncoding genetic variation is assumed to primarily function by directly or indirectly influencing gene expression or splicing. Therefore, an efficient and simple prioritization strategy is to overlap variants with functional genomic annotations (14). These annotations are identified through empirical assays leveraging NGS to query DNA regions enriched for specific biochemical modifications (called ‘peaks’) known to correlate with regulatory activity. This includes methods to identify DNA accessibility (DNase-hypersensitivity (15) and ATAC-seq (16)), DNA methylation, histone modifications (Histone ChIP-seq) (17), Transcription Factor (TF) Binding (TF-ChIP, HT-SELEX, PBM (18)), STARR-Seq (19) or direct assessments of transcriptional activity (GRO/PRO-cap (20,21), GRO-seq (22), PRO-seq (23), CAGE (24)). Large-scale consortium initiatives including NIH Roadmap (25), ENCODE (26), IHEC (27), PsychENCODE (28) and FANTOM5 (29) catalog multiple such marks across many cell-lines and tissues and are an invaluable public resource. Other consortia catalog features such as TFBSs (e.g. JASPAR (18), TRANSFAC (30)) or enhancer annotations (e.g. Enhancer Atlas (31)). Moreover, the development of computational tools including ChromHMM (32) and Segway (33) that integrate multiple annotations across tissues provides higher resolution genomic segmentation and more refined descriptions of transcriptional regulatory states. Other tools, including RegulomeDB (34) and HaploReg (35), integrate pre-existing or user defined marks to directly prioritize GWAS variants. Similarly, functional maps are leveraged by a number of computational algorithms and machine learning methods for functional variant prediction (discussed below; tools reviewed in (36)). The mechanistic relevance of functional genomic annotations to gene regulation and variant prioritization is supported by the strong statistical enrichment of GWAS variation within these regions in a tissue and cell-type-specific manner (37–41).

Nevertheless, functional genomic maps are subject to a number of limitations. First, biochemical maps are often discordant and fail to overlap relevant variation. One analysis found little genomic overlap between multiple enhancer annotation methods and found that only a minority of GWAS and eQTL variants (including curated causal variants) overlapped any given enhancer mark (42). Furthermore, experimental characterization of predicted enhancers found that between 30 and 46% are transcriptionally active (43,44), which taken together suggests that a large proportion of enhancers are both miss-specified and undetected. Second, biochemical peaks are broad and lack the nucleotide-level resolution to disambiguate closely spaced variants (45). Third, existing functional maps are static snapshots and may not recapitulate critical demographic, disease or environmental-specific states, nor dynamic stimulus-induced regulatory changes (46,47). Moreover, such maps do not always capture cell-type or stage-specific aspects of transcriptional regulation, although single-cell epigenomic studies (48) as well as cross-tissue imputation strategies (49) attempt to rectify this. Finally, and most critically, overlap of GWAS variation with regulatory annotations alone does not prove causal relationships. For a comprehensive review of advances and challenges in genome wide annotation of regulatory elements, see Xu et al. (this issue) (14).

High-throughput functional assays for the annotation of noncoding variation

The limitations of biochemical annotations have motivated the development of an array of high-throughput experimental methods for genomic characterization (50) (Table 2). These multiplexed assays of variant effects (MAVEs) generally come in two flavors: (1) parallelized, direct genomic manipulation using gene-editing technology, and (2) massively multiplexed synthetic reporter assays using NGS as a readout (51). MAVEs have rapidly proliferated in recent years, leading to the collection and curation of rich functional data across the spectrum of noncoding variation (51,52). In the remainder of this review, we will discuss the impact of both approaches toward characterizing noncoding genetic elements and their respective limitations. These high-throughput experimental methods are especially important because existing computational methods are either not well powered or lack the specificity to define functional variation in a convincing manner (45).

Table 2.

Comparison of different methods used to screen noncoding and regulatory variants

Screening technology	Advantages	Limitations
CRISPR knockout	- Native context - Infer target genes	- Relatively low throughput - Cannot control exact editing result
CRISPRa/i	- Same as above	- Relatively low throughput - Cannot infer exact causal variants
CRISPR base editing	- Same as above - Precise base conversion/direct inference	- Relatively low throughput - Bystander edits - Limited types of conversions - Editing window restriction
CRISPR prime editing	- Same as above - Versatile types of edits	- Relatively low throughput - Low efficiency
MPRA	- Higher throughput- Direct measurement of functional variants	- Artificial expression- Not within native genome context- Cannot infer target genes

Interrogation of noncoding regulatory function through CRISPR/Cas-based genome modification

Genome editing technologies based on CRISPR/Cas systems permit perturbation of genetic elements in their native genomic contexts and subsequent measures of downstream phenotypic response (53). The modular nature of these technologies enables targeted mutagenesis (54), epigenomic modulation (e.g. CRISPRi (55) and CRISPRa (56)) and direct base pair editing (base editing (57) or prime editing (58)), among other applications (Fig. 1). Excitingly, these assays can be scaled by generating pooled libraries of many gRNAs combined with single-cell phenotyping to create high-throughput CRISPR screens (59). However, the bulk of these multiplexed assays have focused on manipulation of the protein-coding genome in pooled knock-out or saturation mutagenesis studies (60).

Figure 1.

Effects of CRISPR-based tools on noncoding variants. (A) CRISPR/Cas9 generates a double-strand break close to the target site, which induces NHEJ, resulting in an unpredictable repair event that often involves knockout of the variant of interest (highlighted ‘X’). (B) Base editing, as exemplified by the cytidine editor, leads to conversion of specific base pairings (highlighted ‘C’ to ‘T’) for all available relevant base instances in an editing window. (C) Prime editing induces an accurate and versatile conversion of a short sequence (‘X’ to ‘Y’ conversion) by utilizing the template for reverse transcription on the prime editing guide RNA (pegRNA); however, the efficiency of prime editing is much lower than the other technologies. (D) Epigenetic modulators, such as CRISPR interference (CRISPRi), use fused transcriptional co-effectors (KRAB-dCas9 in this case) to modify relevant chemical moieties on chromatin to change chromatin states and repress transcription; a range of up to 1 kb of DNA can be influenced.

Nevertheless, pooled CRISPR screens have been used to efficiently survey the noncoding regulatory landscape. The earliest iterations of these approaches paired Cas9-induced mutagenesis of cis-regulatory regions with functional readouts including cellular proliferation or antibody staining paired with FACS (61) (Fig. 2A). Classic examples include mutagenesis and characterization of putative enhancers upstream of the BCL11A gene in HUDEP-2 erythroid cells (62), TP53 enhancers in MCF-7 fibroblasts (63), among others (64,65). Focusing on biochemically predefined enhancers tractably constrained the target search space but likely missed novel regulatory elements (66) (Fig. 2B).

Figure 2.

Common screening strategies for the noncoding regulatory landscape with CRISPR-based tools. (A) In the first strategy (upper panel), a cellular knock-in reporter line is constructed to monitor endogenous expression of the target gene. Following CRISPRi or CRISPR perturbation, cells with low reporter expression are identified by FACS, followed by gRNA quantification to identify enrichment within functional relevant sites. This type of screening is inexpensive and well-powered, but limited to one target gene. In the second strategy (lower panel), single-cell RNA-seq (scRNA-seq) is performed along with single-cell gRNA sequencing (sc-gRNA-seq). Cells expressing each unique gRNA are grouped together and compared with the remaining cells to identify differentially expressed genes; this analytic framework is similar to eQTL analysis and offers genome-wide coverage, but requires much greater sequencing depth and expense. (B) Screening can be performed in either a tiling (upper) or focused (lower) fashion. The tiling array is unbiased but limited to a restricted number of loci, while the focused array is biased toward known epigenomic or enhancer annotations but enables surveillance of many distinct loci simultaneously.

CRISPRi/a screening represents a major technical breakthrough by greatly expanding throughput compared with earlier mutagenesis studies (Fig. 2). For CRISPR interference (CRISPRi), targeted suppression of endogenous transcriptional machinery is achieved through fusion of dCas9 to repressor proteins, typically the Kruppel-associated box domain (KRAB) (55). Other repressor domains include LSD1, DNMTA3, SID4X and HDAC3 (61,67). Conversely, for CRISPR activation (CRISPRa), dCas9 is fused to one or more activating domains including VP64 (e.g. sunTag (68)), p300 (69) or the MS2 RNA hairpin that recruits MCP-P65-HSF1 fusion proteins (e.g. SAM (70)) to drive gene expression. These systems modify transcriptional machinery within ~200–500 bp of the targeting gRNA to enable regional rather than nucleotide level assay resolution (Fig. 1D). Assay multiplexing and incorporation of single-cell transcriptomic measurement allows for the efficient surveillance of the noncoding regulatory landscape at the genome-wide level and includes approaches such CROP-seq (71), Perturb-seq (59), CRISP-seq (72), MOSAIC-seq (73) and CRISPRi-flowFISH (74).

In an early seminal paper, Fulco and colleagues (75) used CRISPRi to characterize a ~1 Mb region surrounding the MYC and GATA1 transcription factors, identifying nine distal enhancers. A more recent study from Gasperini et al. (76) used the CROP-seq method to screen more than 5920 putative enhancers in K562 cells, identifying 664 novel enhancer-gene relationships (Fig. 2A) highlighting the rapid evolution and improving throughput of these approaches. Other examples of CRISPRi and CRISPRa demonstrate a rapidly growing literature characterizing enhancer elements across cell types (reviewed by Shukla and Huangfu (61)). These assays have also been used to characterize other genomic features such as long noncoding RNAs (77), transcription factor binding sites (TFBS) and cistromes (78,79), and even to validate putative regulatory variants identified from eQTL studies, GWAS or cancer-associated somatic mutations identified in clinical cohorts using WGS (80,81).

A major consideration when interpreting CRISPR-based screening methods is the potential for off-target effects (82). This issue is somewhat mitigated by in silico gRNA design tools that use either alignment (83,84) or evidence-based models to predict on-target and off-target efficiency (85–87). While there are several experimental methods to detect off-target effects of single gRNAs in a genome-wide fashion (88–90), such methods are not scalable for CRISPR screening assays. Therefore, using multiple gRNAs to target each individual locus is the typical solution.

Another methodological limitation is the high dropout rate in mRNA detection using single-cell sequencing technologies, which is magnified in pooled CRISPRi screens incorporating eQTL-like analysis pipelines (76). Practically, our experience suggests that differential gene expression (DGE) based on scRNA-seq data is powered to detect changes primarily in the top decile of genes expressed in a given cell population. Additionally, enhancers have smaller effect sizes relative to promoters and may have compensatory or redundant function when perturbed (76). These two issues, combined with a large multiple-testing burden given the scale of the noncoding genome, suggest a significant limitation in detection power for these assays.

Lastly, there is a tradeoff between throughput and base-pair level resolution. CRISPRa/i approaches lead to chromatin modifications at the range of several hundred base pairs, enabling in aggregate extended assessment of the broader genetic landscape without base-pair level resolution. In contrast, base editing approaches allow nucleotide-scale manipulation and have become highly efficient, with desired conversion rates up to 90% (91). However, these nucleotide conversions can be restricted (primarily C-to-T or A-to-G) (92), bystander effects can occur (i.e. unintended edits of similar bases within the editing window) (93) and fixed windows of editing may limit the editability of bases. As opposed to base editing, even the most efficient version of prime editing thus far achieves 10–20% of desired edits (94), which is not pragmatic for high-throughput screening with a gene expression readout. A recent study that performed a saturation prime editing screen on specific coding regions of two genes with 26 pegRNA structures (95) cleverly used functional markers for target enrichment to achieve high editing efficiency (up to 80%). Overall, base pair editing approaches are very promising and with additional technical advances are likely to be useful for high-throughput screening of noncoding variants in the near future.

Massively parallel reporter assays enable direct functional characterization of diverse genomic features

A second set of approaches for the functional characterization of genetic elements are collectively known as massively parallel reporter assays (MPRA), which involve the construction of a synthetic library of genomic elements, each paired with a reporter gene and a unique, genetically encoded barcode. These libraries are delivered to cell lines or tissues of interest, and the functional effects of library elements are assayed through the multiplexed measurement of barcoded reporter transcripts using NGS (96). This method has enabled regulatory characterization of diverse sets of genomic features across a variety of biological contexts (Fig. 3).

Figure 3.

MPRA workflow. (A) Individual cis-regulatory elements (CREs) can either be synthesized on a DNA oligonucleotide array or sampled from genomic DNA using selection or capture for specific target regions. (B) The CREs are cloned into a reporter vector to drive expression of a uniquely barcoded transcript. Each CRE normally corresponds to multiple unique identifying barcodes. In this example, the CRE is upstream of the minimal promoter and barcoded transcript, but other MPRA designs exist, each with specific advantages and drawbacks. (C) The MPRA reporter library is delivered into a cell model, followed by mRNA collection and unique barcode amplification using targeted PCR. (D) MPRA data can be used to infer transcriptional activity of tested CREs or compare the relative transcriptional efficacies of different alleles of the same variant within a given CRE.

The earliest MPRA iterations were developed to test the transcriptional activity of enhancers (43,97–99) by counting uniquely barcoded transcripts deriving from each enhancer element. These synthetic assays have variable library designs and implementations. For example: (1) enhancer elements can be placed upstream of a minimal promoter, or within the 3’ UTR of the reporter gene (STARR-seq (19), suRE (100,101)). (2) Enhancer DNA can be obtained via microarray synthesis, PCR (102) or genomic DNA capture (103,104). (3) Libraries can be delivered using episomal or integrating vectors (105,106). MPRAs have been used to successfully screen for enhancer activity (96), repressor elements (107) or differential activity across a diverse array of prokaryotic and eukaryotic cell types (97–99,108). Packaging libraries within viral delivery platforms, including adeno-associated virus (104,109) and lentivirus (105,110), have broadened the available cellular contexts to include difficult primary cell lines such as neural cells (110–112), and in vivo assays in the mouse retina (113) and brain (104,109). Approaches incorporating saturation mutagenesis (114) or tiling of overlapping elements (115) enable elucidation of TF-binding logic and nucleotide-resolution sequence specificity. Finally, more recent MPRA iterations probe the function of other noncoding genomic features such as 5’ and 3’ UTRs or splice sites and their effects on transcriptional, post-transcriptional or translational regulation. These include MPRA characterization of RNA splicing (116,117), RNA or protein stability (118–120), RNA editing (121), RNA localization (122) and translation efficiency (123).

MPRA can also be used to compare transcriptional regulatory effects between alleles to interrogate noncoding genetic variation, though these assays are performed less frequently due to the challenge of measuring small allelic effect sizes. Kircher et al. (114) performed saturation mutagenesis across 20 disease associated cis-regulatory elements to test for functional effects across ~30 000 SNVs. Others have assessed synthetic configurations of TFBSs to query the sensitivity of DNA-binding elements to disrupting variation (112,124–126). Additionally, MPRA has been used to screen thousands of noncoding variants derived from eQTL (127,128), GWAS and WGS studies. Specifically, MPRA has been used to measure allelic effects of variants associated with vascular traits (129,130), cancer (130–132), obesity (133), diabetes (134), osteoarthritis (135), COPD (136), lupus (137) and neuropsychiatric (138,139) and neurodegenerative disorders (140). Other work has characterized rare autism-associated (141) or other common (142) variants located in human accelerated regions, as well as assessment of human-specific (143) and introgressed (144) regulatory variants.

It has been repeatedly shown that MPRAs are exquisitely sensitive at detecting functional disruption of TF binding sites within gene-regulatory elements (112,145,146). Recently, Cooper et al. (140) used MPRA to screen common variants derived from GWAS for two neurodegenerative disorders, finding that functional common regulatory variants disrupted a network of interacting TFs likely to operate in a cell-type specific manner to modulate disease risk. These results underscore how high-throughput data from functional genetic screens can intersect with network biology to provide broader insights into the etiology of complex traits beyond mechanistic characterization of individual loci.

When compared with gene-editing approaches, MPRAs afford a number of advantages. First, most MPRAs do not require deep single-cell phenotyping and are therefore comparatively cheap. Synthetic library construction is flexible and not limited to sequences or regions, such as those harboring PAM sites, and internal validity and normalization is facilitated through the incorporation of known control sequences. Moreover, assay throughput is an order of magnitude greater, with MPRA libraries typically querying >10 000 elements. Most significantly, MPRAs allow for nucleotide level resolution and testing of a priori defined sets of human polymorphisms.

Technical considerations are important and library design and implementation can have a large impact on results (147). Increasing enhancer size changes regulatory function through the introduction of distal TFBSs with transcriptional modifying effects, which may be more reliable (147). The ‘classic’ placement of library elements in the 5’ UTR increases assay sensitivity to alterations in DNA-binding motifs, whereas elements in the 3′ orientation (STARR-seq) are more influenced by RNA-binding proteins (147,148). Additionally, MPRAs testing allelic effects are dependent on high transfection efficiency, which has necessitated the use of immortalized cell lines in the literature (149), which can impact assay outcomes (150). Therefore, MPRAs are ideally performed in primary cell types or in vivo tissues most relevant to the underlying trait of interest, or using a staged design. Similarly, assays rarely account for state-specific or environmental factors that may interact with genetic elements to influence gene expression. Improvements in library delivery methods and more complex state-specific study designs will begin to rectify this.

Additionally, since MPRAs involve testing DNA elements outside of their native genomic environments they do not account for locus-specific chromatin architectures, DNA–protein and DNA–RNA interactions and adjacent sequence contexts that strongly regulate gene expression. Yet, despite this limitation, they are remarkably reliable. In all cases, interpretation should be cautious, as demonstration of regulatory function does not necessarily imply that a specific variant is ‘causal’ or pathogenic. Similarly, functional variants might regulate multiple downstream genes through one-to-many enhancer relationships. Only experimental follow-up can resolve which functional variants or associated risk genes are mechanistically relevant.

Integration of multiple functional methods overcomes individual technical limitations

Multimodal methodological integration can address many of the limitations of individual approaches. In a recent tour de force, Ray and colleagues (151) characterized ~2776 regulatory SNVs in the multi-disease associated TNFAIP3 locus across three related erythroid cell types using seven approaches: HiChIP, ATAC-seq, TFBS assessment, CRISPRa, CRISPRi and L-MPRA (lenti) and T-MPRA (transfection). In their analysis, they found that CRISPRi and T-MPRA were effective for identifying likely causal variation and particularly favored the integration of these two approaches (151). In another recent study, we also integrated T-MPRA, CRISPRi screens and published biochemical data to functionally prioritize 5706 GWAS variants for two neurodegenerative disorders, Alzheimer’s disease and Progressive Supranuclear Palsy (140). These complementary methods were used to sequentially and iteratively prioritize high-probability causal variants and account for the limitations of individual methods: MPRA provided allele level functional information, CRISPRi verified functionality in native genomic contexts and assignment of downstream genes, and biochemical assays confirmed accessibility and likely functionality within relevant post-mortem tissues (140). Other recent advances include the transMPRA approach, which cleverly combines synthetic MPRA libraries with CRISPRi screens in a combinatorial manner to systematically identify trans regulators of user-defined noncoding genetic elements (152). These results highlight the necessity of a multimodal approach in future studies seeking to meaningfully annotate the noncoding genome.

Data from multiplexed assays can improve predictive classification methods

The development of inferential and predictive methods for identifying functional genomic elements is an active area of ongoing research. Computational methods have the advantage of more throughput and less time and overhead costs compared with experimental approaches. In recent years, an array of statistical and machine learning algorithms leveraging sequence features including biochemical annotations and evolutionary conservation have been developed for functional variant prediction (see Nicholls and colleagues for an overview of machine learning methods (153) and Schipper and Posthuma, this issue, for a broad discussion of analytic methods (155)). Unfortunately, these algorithms still exhibit inconsistent predictive accuracy across all use cases and poor agreement with each other (45,147), limiting utility in the research and clinical setting. As such, the ACMG’s latest guidelines have restrictive rules on the usage of predictive algorithms in clinical reporting (154). To rectify this, it has been proposed that functional data accrued from high-throughput experimental screens can serve as an orthogonal benchmark of algorithm performance (128,156). However, the extent to which computational algorithms recapitulate functional data from MPRA during direct comparison remains highly disparate in the literature (129,140,147,157) and may vary depending on the specific features used to train particular algorithms. Interestingly, recent work demonstrates that functional data from MPRA can be directly incorporated to train computational algorithms and boost predictive performance (158). For example, Movva and colleagues (159) developed MPRA-DragoNN, a convolutional neural network exclusively trained on pre-existing MPRA, and validated this method by fine-mapping an independent GWAS of LDL levels.

These findings are illustrative of a broader point: high-throughput experimental datasets can be leveraged to improve a priori predictive methods across different classes of functional variation. For instance, Fulco and colleagues developed the CRISPRi-FlowFish method to characterize putative enhancers in K562 lymphoblastoid cells and used the resultant functional data to generate an Activity-by-Contact model of enhancer function that outperformed existing predictive methods (74) and effectively prioritized GWAS variation (160). Rosenberg et al. (116) created a functional assay to measure differential splicing effects for more than 2 million synthetic variants, which was used to train a predictive model that vastly outperformed existing algorithms. Other work functionally screened polyribosome loading for more than 280 000 synthetically derived 5’ UTRs to generate an effective predictive model for the translational efficacy of 5’ UTR sequence variants (123). These examples demonstrate that the diverse, rapidly expanding catalog of experimental functional datasets can be utilized to improve predictive algorithms and may even inform the interpretation of human variation in clinical databases.

Conclusions and future perspectives

Recent technological innovations have rapidly advanced our functional understanding of the human regulatory genome (Xu et al., this issue). Tissue-specific, genome-wide annotations of putative regulatory elements are now available across hundreds of tissues and cell lines, as exemplified by work from the ENCODE (148), Roadmap Epigenomics (25) and PsychENCODE (161) consortiums. Cell-type-specific regulatory maps represent the next frontier in this effort. Yet, functional screening is still required to fully delineate the functional effects of noncoding genetic variation.

The complex genetic architectures underlying most common traits require integrative experimental approaches to dissect individual loci and to understand their combined effects. The work on autism spectrum disorders (ASD) is a particularly salient example of genetic studies revolutionizing causal models of diseases. Studies over the past decades have identified causal roles for: (1) common variants, (2) rare, de novo and protein-truncating variants, (3) copy number variants and (4) high-confidence syndromic genes (162). Although the phenotypes of ASD are multifaceted and highly variable, both genetic (163,164) and transcriptomic evidence (165,166) points to disruption of convergent gene networks and pathways during fetal brain development. Comprehensive mapping of such networks advances our understanding of disease etiology. Transcriptional regulation through cis-regulatory elements, where noncoding variants play critical roles, stands as a key piece of the network. Therefore, high-throughput functional assays are required to dissect the variants involved—in order to map their cell-type specificity and downstream target genes. A similar logic can be applied to many other diseases.

CRISPRa/i screens and MPRA are two proven technologies to dissect noncoding variation. Worth mentioning are other high-throughput methods surveying other aspects of noncoding variant function. For example, SNP-SELEX (167) and REEL-Seq (168) are parallelized in vitro assays directly querying DNA/TF interactions. In addition, variants affecting RNA splicing (sQTLs) are also predicted to play a significant role in disease etiology (8,169), but high-throughput functional characterization of sQTLs is in its early stages (e.g. (117)). To improve high-throughput analysis of splicing, a more efficient prime editing assay coupled with an isoform-level transcriptomic readout would be optimal.

One major area of opportunity aside from increased scaling is to couple these technologies with other biological assays. Currently, CRISPR screens rely primarily on transcriptomics or cell proliferation (170) as readouts. However, recent studies have started to incorporate imaging analysis into this repertoire (171,172), which may facilitate the linkage of genetic variants to other biological features specific to disease. Other complex biological traits could be integrated within these assays, such as neuronal activity, metabolomics, etc., which more directly reflect the relevant physiological processes. Ultimately, functional genetic screening will ideally be performed at base-pair resolution and in versatile, native and relevant cellular contexts with physiologically relevant functional readouts. In the near term, this will depend on improved gene editing technologies and cell delivery methods.

Funding

National Institute of Neurological Disorders and Stroke [UG3NS104095 to D.H.G.]; National Institute of Mental Health [1U54NS123746 and 5U01MH116489-05 to D.H.G.]; National Institute of Aging fellowship [1F30AG064832 to Y.A.C.]; UCLA-Caltech MSTP training grant [T32-GM008042 to Y.A.C.].