The power of multiplexed functional analysis of genetic variants

Abstract

New technologies have recently enabled saturation mutagenesis and functional analysis of nearly all possible variants of regulatory elements or proteins of interest in single experiments. Here we discuss the past, present, and futureof such multiplexed (functional) assays for variant effects (MAVEs). MAVEs provide detailed insight into sequence-function relationships, and they may prove critical for the prospective clinical interpretation of genetic variants.

Genome sequencing is now routine, and tools for annotating gene structures and regulatory elements are becoming increasingly mature. Yet in this time of genomic plenty, researchers remain poor at predicting genotype–phenotype relationships, that isthe consequences of genetic variation. Which single-nucleotide changes will affect gene regulation? Which amino acid changes will affect protein function? Under what circumstances do the resulting biochemical phenotypes give rise to organismal phenotypes? For regulatory, protein-coding, and organismal phenotypes, whatis the distribution of effect sizes within the space of all possible sequence variants? What risk does each confer for disease, and to what degree do they affect characteristics such as age of onset and severity of disease? Although methods for answering these questions by computational prediction have proliferated, their effectiveness is limited, and the conventional approach to confirm that an individual variant has a meaningful biochemical or organism-level effect is still to assay it in an in vitro systemor model organism¹. This one-by-one, post hoc approach doesnot scale to the vast numbers of genetic variants that are being discovered each day by clinical exome and genome sequencing.

Within the past decade, various innovations have enabled the assignment of functional effects to hundreds to thousands of sequence variants in a single, highly multiplexed experiment; here we term these ‘multiplexed assays for variant effects’, or MAVEs. MAVE experiments have mapped sequence–function relationships with base-pair resolution for both proteins and regulatory elements in the form of deep mutational scans² and massively parallel reporter assays³, respectively. Deep mutational scans build on low-throughput predecessors such as alanine scanning, in which each variant is cloned and assayed individually⁴, and surveys of variant libraries by protein-display methods^5–7, which suffer from reliance on capillary sequencing that allows function to be assigned to only a small number of winning variants. Massively parallel reporter assays build on methods such as saturation mutagenesis⁸ that are similarly limited with respect to scalability.

From a technical perspective, underlying the development of MAVEs are advances in massively parallel DNA synthesis⁹ and DNA sequencing¹⁰ that, respectively, enable the multiplex construction of genetic variant libraries and the multiplex quantification of functional consequences. Using oligonucleotide libraries generated by massively parallel DNA synthesis, wecan now—within a single experiment—program every single-nucleotide change in a regulatory region or every possible amino acid change in a protein. Using next-generation DNA sequencing, it is now possible to track and quantify the functional effects of all of these variants within a single experiment. Even more recently, breakthroughs in genome engineering^11–13 have enabled MAVE approaches for assaying the functional consequences of variants in their native genomic context¹⁴. With MAVE methods, the number of variants that can be functionally tested increases to hundreds or thousands per experiment.

Already, MAVE approaches have shown their utility for sequence–function analysis of diverse classes of sequence, including enhancers, promoters, mRNA untranslated regions, splice sites, and numerous kinds of proteins (Fig. 1). Though analogously multiplexed assays have been developed that validate thousands of putative regulatory elements at once¹⁵ or scan endogenous genomic space to dissect novel regulatory elements¹⁶, we focus here primarily on the dense dissection of sequences of interest, that is, measuring the effects of all possible nucleotide substitutions in a regulatory element or amino acid substitutions in a protein.

Figure 1 |.

Multiplexed assays for variant effects (MAVEs) throughout the central dogma. In the past decade, MAVEs have functionally tested at least tens to hundreds of thousands of variants across diverse sequences corresponding to all parts of the central dogma. Gene regulatory elements such as enhancers, promoters, and untranslated regions (UTRs) have been dissected, as have regions that affect splicing and RNA stability. All possible amino acid changes across proteins of diverse function have been characterized, yielding insights into biochemistry and disease.

Although there is considerable variety in the details of each implementation, MAVE experiments share a basic framework, with key steps represented in Figure 2: (1) construction of a variant library (i.e., allelic series) of the sequence of interest, (2) delivery of this variant library to an in vitro or in vivo system, (3) the functional assay (i.e., the stratification of variants by function), (4) sequencing to quantify each variant’s representation in the context of the assay, and (5) calculation and calibration of functional scores for each variant. Ideally, this workflow resultsin sequence–function maps that capture the effect size of every possible variant at every position in the sequence of interest, with respect to the function assayed and potentially its correlates as well (e.g., clinical phenotype) (step 6 in Fig. 2). Below, we discuss each of these steps in greater detail.

Figure 2 |.

The key steps of MAVE. Though diverse, all MAVE experiments rely on the same steps: (1) Construction of a variant library or allelic series of the sequence of interest. These variants might include all possible amino acid changes in a protein or all single-nucleotide changes in a regulatory element. (2) Delivery of this variant library to a model system. Variant libraries can be delivered episomally or via genomic integration by genome editing, by random insertion, or at a safe harbor locus. (3) A functional assay to stratify variants by function. Effects on RNA expression from variant regulatory sequences are measured by using sequencing to count transcripts under the influence of each variant. In protein MAVEs, assays are used that separate coding sequences for functional versus nonfunctional variants. (4) Sequencing to quantify each variant’s representation in the context of the assay. For regulatory sequences, DNA and RNA that tag each variant can be sequenced to quantify effects on transcriptional output. Protein-coding variants (or variant-associated tags/barcodes) are sequenced before and after functional selection. (5) Calculation and calibration of functional scores for each variant. Sequencing read counts must be converted into a score for each variant. These scores range over a distribution of all possible effect sizes, and this distribution can be benchmarked by variants of known effect. (6) The genotype–phenotype relationship at every position in the interrogated sequence is represented in sequence–function maps.

Construction of a variant library

It is challenging to generate a variant library (also referred to as an allelic series) that uniformly represents all possible nucleotide or amino acid substitutions in an efficient and cost-effective manner. Classic methods such as error-prone PCR suffer from polymerase bias¹⁷. Doped oligonucleotides are limited in length (~200 bp), and although they can produce uniform libraries, methods that rely on individually synthesized oligonucleotides (i.e., one or two primers for each programmed mutation of a single base or codon) are costly and labor-intensive^18–20. However, despite these limitations, these remain viable options for constructing variant libraries of both regulatory and protein-coding sequences.

In 2004, Cleary et al.⁹ demonstrated that the products of microarray-based DNA synthesis can be used in a preparative rather than analytical fashion—that is, they can be released from the array surface and used as high-complexity oligonucleotide libraries. Cost-effective, array-based DNA synthesis technologies are still improving in terms of the length and quality of their products. Despite this, the ability to program complex variant libraries from arrays has facilitated much of the MAVE work of the past decade. There are many ways by which array-derived libraries can be used to program variant libraries for MAVE experiments. For example, all possible alternative codons can be programmed on array-derived primers and incorporated into a coding sequence by sequential primer extensions from a wild-type sequence²¹ or by Gibson assembly²².

Delivery of the variant library

Variant libraries can be introduced to populations of cells via episomes or by insertion into the genome. If one is measuring regulatory activity via transcribed barcodes (Fig. 2, left), a high multiplicity of delivery per cell (i.e., multiple episomal or lentiviral reporters in a single cell) is permissible, if not desirable. However, in assays where the effect of a variant is assessed on the basis of the phenotype of its host cell (Fig. 2, right), the method must limit the number of alleles delivered per cell to one to avoid confounding the impact of any single variant. For bacteria- and yeast-based assays, alleles can be delivered by plasmid becausethe contribution of cotransformed cells is negligible. Delivery of a single allele to a mammalian cell is more challenging. Alleles can be randomly inserted in the genome by viral transduction^23,24 at a multiplicity of infection of less than one, or by targeting only one locus for integration via integrase- or recombinase-mediated insertion^25,26. We have also used CRISPR/Cas9 genome editing to introduce libraries of variants to their endogenous locus, such that ploidy limits the number of copies introduced per cell¹⁴. Although currently somewhat inefficient, this approach has the major advantage of enabling variants to be assayed in their native genomic context.

Stratification of variants by function

The quality of MAVE measurements hinges on the ability of a functional assay to accurately stratify variants by their impact on the biochemical or cellular activity of interest. The design and validation of a well-performing functional assay is perhaps the most challenging aspect of implementing MAVEs. Broadly speaking, the design considerations are different for regulatory elements than for protein-coding sequences.

The impact of regulatory-sequence variation is most often stratified by changes in the transcriptional output of a reporter gene. To assess the effect of programmed regulatory variants on gene regulation, targeted RNA-seq can be used to count cis-linked reporter transcripts that contain a barcode uniquely paired witha specific variant²⁷. For example, for enhancer reporter assays, regulatory variants might increase or decrease transcriptional activation of associated barcodes, relative to the wild-type enhancer^22,27,28. Similar approaches have also been used to measure variant effects on splicing, wherein the regulatory variants might also serve as the barcode^14,29,30. Alternatively, a fluorescent reporter protein can be used as a proxy for RNA expression, with cells separated into brightness bins by fluorescence-activatedcell sorting^31,32. In this case, the variants present in each bin canbe quantified from the sorted DNA and inferred to have either increased or decreased expression of the reporter gene.

Designing suitable assays for protein MAVEs is more challenging. Protein MAVEs generally require delivery of one construct per cell, and protein function needs to be tied to either cellular growth or reporters that can be sorted by flow cytometry. Though each assay is highly specific to the protein of interest, recurring themes in the functional assays that are used for protein MAVE include protein display and capture^33–35, antibiotic resistance^22,36,37, cellular growth^38–40, viral infectivity^41–43, and protein- or antigen-binding affinity^44,45.

Looking to the future, we predict the development of new functional assays in two complementary directions. First, to enable the effective scaling of MAVEs to larger swaths of the proteome, we predict that MAVEs that measure generic protein properties such as stability or localization will be of great utility. Second, to accurately measure the effects of variants on genes associated with disease risk, there will be a strong incentive to develop multiplexing-compatible assays that specifically model the activities of a protein that are thought to be most relevant to its role in disease⁴⁶.

Sequencing to count variant frequency

After stratification in a functional assay, MAVEs rely on massively parallel DNA sequencing to provide a digital ‘count’ for each variant. Variants (or associated barcodes) are amplified from the functionally stratified DNA or RNA and sequenced to determine the frequency of each in each post-assay sample.

Provided that the assay results in changes in the representation of variants, the mutagenized region can be sequenced directly³³. However, if the variant is not part of the assay ‘output’, or if the mutagenized region is long and cannot be easily covered by cost-effective sequencing platforms, then each mutant can be tagged with a short barcode for the purposes of readout/quantification²⁷. Such barcodes can be linked to each variant in the mutagenized regulatory element or protein via subassembly^28,34,47 or long-read sequencing (L.M. Starita, M. Kircher, J. Underwood & J. Shendure, unpublished data). After barcodes have been linked to variants, only the barcodes need be sequenced to track or quantify the variants.

Calculating and calibrating functional scores for each variant

In order for MAVE experiments to be readily interpreted, the number of sequencing reads for each variant must be converted to a meaningful functional score. Multiple statistical models exist to convert read counts into scores^48–51, including simple but effective ratios of variant frequencies in RNA/DNA or selected/unselected populations.

A powerful and unique aspect of MAVE experiments, relativeto conventional one-by-one functional assays, is that they resultin a distribution of effect sizes for a large number of potential variants of a sequence of interest, all generated within a single experiment. To interpret this distribution, ideally one should compare a variant against benchmarks of known or expected effect. In protein MAVEs, stop codons represent the worst outcome, whereas synonymous changes are expected to have neutral effects. These can be used to validate and calibrate the distribution of observations, as well as to quantify uncertainty in measurements. For regulatory MAVEs, well-characterized regulatory motifscan be used as positive controls in noncoding MAVEs, whereas scrambled sequences with no expected regulatory effect can be used as negative controls⁵². The remaining variants of unknown effect are then compared to such controls. For disease-relevant sequences, previously observed pathogenic and benign substitutions can be used to calibrate MAVE scores for use in a clinical context^45,53. We predict that the increase in the number of available human genotype–phenotype data sets, such as that generated by the Exome Aggregation Consortium for estimating allele frequencies⁵⁴, and the expansion of ClinVar as a source for benign and pathogenic substitutions⁵⁵ will allow scientists to better calibrate and interpret the dense sets of experimental variant effects resulting from MAVEs of disease-relevant proteins and regulatory elements.

The future of MAVE

As the progress of human genetics is increasingly limitedby the interpretation of genetic variants rather than by their ascertainment, we predict that the adoption and application of MAVE experiments will accelerate in the coming years. Ideally, by the 20th anniversary of Nature Protocols, sequence–function maps for thousands of proteins and regulatory elements^56–58 will have been generated at single-residue resolution.

Such high-resolution sequence–function maps may provide the substrate for training more accurate computational models for directly predicting the impact of genetic variation on phenotypes. We are already observing this potential. For example, a model trained on the few hundred thousand splice sites in the human transcriptome⁵⁹ is less accurate than one trained on muchlarger data sets created via MAVE-like experiments designed to learn the rules of splice-site selection from millions of synthetic exons³⁰. By revealing relationships between regulatory sequence composition and gene regulation, MAVEs will also shed lighton the fine-scale ‘functional anatomy’ of the transcription factor binding motifs that comprise enhancers and promoters, and thereby allow scientists to build better algorithms to predictthe effects of single-nucleotide variants on gene expression. For proteins, models trained on large numbers of MAVE experiments could also provide information about the relationship between primary sequence and protein properties such as stability, folding, and mutational tolerance.

Beyond advancing genomics in general, we also expect that MAVEs of disease-associated genes and regulatory elements will deliver experimental data needed to guide variant interpretation in the clinic. Genetic testing is identifying vast numbers of variants of unknown significance (VUSs). VUSs are often missense substitutions with unknown effects on the biochemistry of a disease-associated protein (though, importantly, splicingand other regulatory mutations can also lead to disease). To patients and physicians alike, ‘VUS’ is a confusing and unhelpful categorization. These variants are found at a high rate in panelsof genes associated with cancer risk, where VUSs can outnumber interpretable ones by 95 to 1 (ref. 60). As a solution to the problem of VUS identification, the results of MAVEs have the potentialto give rise to clinically calibrated scores for the impact of every possible amino acid change on the biochemistry of the tested protein-coding gene. Generating these variant interpretations prospectively will position those in the field to incorporate them into the routine practice of clinical genetics.

These high rates of VUSs come from surveying only coding sequences associated with disease, which represent about 1%of the genome. Scientists have just begun to understand how noncoding genetic variation in the other 99% of the genome influences disease. More than 24,000 loci, most of which occur in noncoding regions, have been associated with human phenotypes through genome-wide association studies⁵⁸. MAVEs can not only help determine which of these implicated sequences contributeto gene expression, but also test noncoding variants associated with disease^61–63. In recent regulatory MAVE studies, CRISPR-facilitated allelic replacement was used to functionally validate individual prioritized variants, as studying these hits in their native genomic context conferred biological advantages over episomal assays; in the genome, contextual transcription factors, polymerases and 3D architecture are maintained for promoters and distal regulatory elements⁶², and native splicing machinery is available^14,63. Saturation genome editing has already been applied to survey thousands of variants in the genome¹⁴, and we predict that as genome engineering protocols rapidly improve, the clear advantages of testing variants in their endogenous context will motivate the use of genome editing in the next decade’s MAVEs.

Additionally, we predict that genome editing may enable MAVEs of a given allelic series in the context of more thanone cell line (i.e., reflecting the different genetic backgroundsof patients, which are known to influence the penetranceor variable expressivity of many genetic diseases even whenthe causal mutations are identical). One can also envision quantifying epistatic effects between variants or allelic series that are concurrently introduced to multiple endogenous locations scattered across the genome. Such experiments have the potential to shed light on how epistatic effects influence phenotypic traits, which at present remains poorly understood.

Massively parallel functional dissection of proteins and regulatory elements generates empirical measurements of the consequences of thousands of variants per experiment. The distributions of effect sizes and sequence–function maps inform biology and may have a critical role in the clinical interpretation of genetic variation. We anticipate that as methods pioneered over the past 10 years are scaled up over the next decade, MAVEs may enable measurement of the functional consequences of millions to billions of genetic variants.