The status of the human gene catalogue

Preface

Scientists have been trying to identify every gene in the human genome since the initial draft was published in 2001. Over the intervening years, much progress has been made in identifying protein-coding genes, and the estimated number has shrunk to fewer than 20,000, although the number of distinct protein-coding isoforms has expanded dramatically. The invention of high-throughput RNA sequencing and other technological breakthroughs have led to an explosion in the number of reported non-coding RNA genes, although most of them do not yet have any known function. A combination of recent advances offers a path forward to identifying these functions and towards eventually completing the human gene catalogue. However, much work remains to be done before we have a universal annotation standard that includes all medically significant genes and maintains their relationships with different reference genomes.

Introduction

The Human Genome Project (HGP) was launched in 1990 with two central goals: “analyzing the structure of human DNA” and “determining the location of all human genes”². The recent sequencing and assembly of a complete human genome from telomere to telomere³ accomplished the first of these goals: a complete, gap-free DNA sequence. Achieving the second goal, though, has been far more complicated than originally anticipated, despite a vast increase in our knowledge of the location and function of tens of thousands of human genes. Over time, the task of identifying genes and their functions has been augmented with the goal of identifying their regulatory mechanisms. International efforts have been launched to find all functional elements in the genome^4,5, including genes as well as transcriptional and post-transcriptional regulatory elements.

Early conceptions of the genome treated it as a repository for genes, most of which were thought to encode a single protein-coding transcript^6,7. Today, though, we know that the picture is different, and that human biology can be influenced by thousands of alternative transcripts and transcribed elements that are not translated into proteins^8,9, and by hundreds of thousands of regulatory elements⁴. Further complicating matters, we now know that many transcribed RNA molecules are further processed into smaller RNA fragments that can have functions different from their parent transcripts (Figure 1).

Figure 1:

A major challenge for gene annotation is how to capture the diversity of gene products and functions. For example, although the vast majority of protein-coding genes occur on distinct transcripts, a small number of bi-cistronic transcripts encode two distinct open reading frames on the same transcript. Similarly, introns within protein-coding genes may host noncoding RNAs, including miRNAs, snoRNAs or lncRNAs, which may regulate the transcriptional activity of the locus, or may have catalytic roles unrelated to the main protein product. Alternate splicing of transcripts may give rise to proteins that enhance or inhibit each other. Transcripts that are truncated and cannot produce functional proteins are targeted for nonsense-mediated decay (NMD). These products, together with ubiquitinated proteins (Ub) or unwanted intronic material are rapidly recycled by cellular lysozomes. Other seemingly nonproductive transcripts may be repurposed as functional ncRNAs.

The purpose of this perspective is to revisit the goals of the HGP in light of our increased understanding of the diversity of functional elements in the human genome. While the genome contains many different features, this perspective will focus on genes. In the sections that follow, we will consider how we can finish specific aspects of human gene annotation in the years to come. These include (1) completing the list of protein-coding genes and all of their isoforms; (2) compiling a complete list of RNA genes of all lengths and varieties; and (3) identifying medically important genes and gene variants, and linking them to specific disorders. For each of these discussions, we will review where we are today, and what remains to be done, and then finally (4) we discuss technology needed to complete the annotation of human genes.¹⁰

Protein-coding genes

Protein-coding genes included in major genome annotation databases–e.g., GENCODE, RefSeq, and CHESS–or captured in reference protein annotation databases such as UniProtKB generally have evidence not just for their translation but also, in many cases, for the function of the protein that they encode^11–14. Primary evidence can include the direct biochemical or molecular experiments or inference of function recovered from the scientific literature. The direct observation of function of a gene product or that of a close paralog provides confidence in the assignment of function of the gene and its annotation as protein-coding. In addition, the generation of high-quality genome sequences for a large number of vertebrate species, alongside the development of software (such as PhyloCSF++¹⁵, PhastCons¹⁶, or PhyloP¹⁷) capable of using alignments to identify regions of the genome under purifying selection, as well as indirect evidence of translation from mass spectrometry data, increases our confidence in many protein-coding genes.

Protein-coding gene count

The annotation of protein-coding genes was the primary focus of the Human Genome Project, after capturing the sequence itself, and while this annotation is still incomplete, the scientific community is approaching a consensus on the identities of these genes. From an initial estimate of 50,000–100,000 genes in the 1980s, the estimated number has dropped steadily, falling to 30,000–40,000 with the initial publication of the human genome^18,19, and then further to ~20–25,000^20,21, 22,000¹, and just under 20,000 today³, one recent database release suggests as few as ~19,000 (e.g., 19,370 in GENCODE Release 41).

These refinements came about through a variety of advances, including comprehensive manual review²², improvements in computational annotation methods and analysis, and the generation of ever greater volumes of high-quality experimental transcriptional data. Despite the overall reduction in gene count, novel protein-coding genes continue to be identified, as well as alternative isoforms of known genes.

The Matched Annotation from NCBI and EMBL-EBI (MANE) collaboration²³ recently published a near-complete dataset containing one isoform for each protein-coding gene for which two of the leading annotation projects, RefSeq and GENCODE, agree completely. A secondary goal of this project is to converge on an answer to the question of how many protein-coding genes we have. MANE 1.0 contains 19,062 gene loci, which covers ~98% of the curated protein-coding genes in the major human gene catalogs, bringing us closer than ever to one of the central goals of the HGP. An important caveat is that the MANE annotation is provided on the human reference genome known as GRCh38, which still contains gaps, and not on the finished T2T-CHM13 assembly, which was reported to contain 140 additional protein-coding genes³.

We propose a number of future steps to completing the annotation of protein-coding genes in the human genome:

1. For each protein-coding gene, develop a comprehensive picture of its transcripts and their expression levels in all tissues and cell types available, and determine its conservation in other species.

2. For all proteins that fold into stable structures, determine their 3-dimensional structure and evaluate their stability.

3. Determine all alternative sites of transcription initiation and termination, and record how frequently each site is utilized in normal tissues.

4. Label all reproducible splicing events that lead to non-functional proteins.

5. Catalog and highlight the many exceptional cases where normal rules appear to be violated. These include (a) bicistronic genes, where two distinct protein-coding genes occur on the same transcript; (b) selenoproteins, where UGA can encode selenocysteine rather than functioning as a stop codon; (c) non-standard splice sites with recognition sites deviating from the most common GT-AG, GC-AG, and AT-AC sites²⁴; (d) coding sequences that use a codon other than ATG as the start codon; and (e) extremely short exons, which are often missed or misplaced by current methods.

Although we are nearing consensus on a protein-coding gene set, the precise set of annotated protein isoforms is still in flux^11,25. Determining this number has been challenging for multiple reasons. First, the determination of isoforms today relies primarily on assembly of RNA-seq data, which in turn relies on having a complete sample of all genes in all cell types, including those prevalent during early development. Efforts such as GTEx²⁶ have surveyed a large number of tissues, but still only cover a subset of cell types. Projects such as the Human Cell Atlas aim to identify cell-type-specific RNAs for all human cell types, but much work remains. Second, computational methods do not consistently produce the same splice isoforms from large, complex RNA-seq data sets, in part because short-read RNA-seq sequencing is insufficient to unambiguously determine complete splice structures. Third, even for those isoforms that do appear reproducibly in RNA-seq experiments, many may not encode functional proteins. And fourth, genetic variants in the human population likely introduce splice variants that will only be catalogued as we sequence a greater number of individual humans.

Pseudogenes

Another major challenge, beyond identifying the genes and splice variants themselves, is determining which gene-like elements are pseudogenes. Pseudogenes are sequences that represent defective copies of genes: over 14,000 have been annotated on the human genome. They can be divided into three types: processed (introns removed during retrotransposition), unprocessed (introns retained during duplication), and unitary (pseudogenes without a functioning counterpart in human). Recent evidence using long-read technology suggests that some previously-annotated pseudogenes may in fact be functional^27,28, and other reports indicate that some pseudogenes continue to be translated, although the protein products might not be functional²⁹.

Noncoding RNA genes

Non-coding RNA genes (ncRNAs) include a range of different RNA molecules that are transcribed from DNA, that do not encode proteins, and that provide a function in the cell. A variety of subclasses of ncRNAs have been described, including both long ncRNAs (lncRNAs), defined as RNAs ≥200 nt, and many types of shorter ncRNAs such as microRNAs, small nucleolar RNAs, transfer RNAs, piwi-interacting RNAs, and others. We note that although many non-functional RNA sequences might be transcribed in various cells and conditions, our definition will only call them genes if they have a discernable function at the cellular or organismal level. Similarly to protein-coding genes, the functions of lncRNAs need to be determined by primary experimental evidence that reveals their biochemical or molecular function, which can be obtained, e.g., from perturbation of lncRNAs followed by molecular phenotyping³⁰. In contrast to protein-coding genes, though, it is still unclear if function can be inferred by comparing paralogues, due to our limited understanding of the mechanisms of action for most lncRNAs and to the fact that some lncRNAs contain embedded sequences from retrotransposons. In the near term, most annotation efforts will continue to strive to comprehensively catalogue ncRNA transcripts, regardless of their functional status.

Although annotation strategies that search for conserved protein sequences cannot be used for characterizing ncRNAs, high-throughput RNA-seq experiments have provided an abundant source of evidence for transcription of these genes. Compared to protein-coding RNAs, ncRNAs discovered through RNA-seq appear in relatively low abundance, raising questions about whether they encode functional elements or instead represent transcriptional noise. On the larger question of what ncRNA genes do, many possible functions have been described, including regulating expression of other genes, splicing, chromatin architecture, epigenetic regulation, dysregulation in cancer and other diseases, translation, DNA repair, and more^31–33. And although tens of thousands of ncRNA transcripts are currently annotated in the human genome, their heterogeneity, poorly understood biology, and other characteristics make the comprehensive discovery of all genes in the ncRNA catalogue an unsolved problem.

A summary of lncRNA gene annotation in current catalogues is shown in Table 1. The two most-widely used are RefSeq and GENCODE, both of which employ human annotators along with large-scale cDNA and RNA sequencing resources^12,34,35 to determine which ncRNA genes to include. In parallel, a variety of consortia and individual research laboratories have provided valuable additional resources, including NONCODE, the FANTOM consortium’s CAT resource, LNCipedia, miTranscriptome, CHESS, LncBook, RNAcentral, and others (e.g., see³⁶).

Table 1:

Annotation databases that catalogue long ncRNA genes (figures as of late 2022). Here, “long” refers to loci ≥200 nt.

Resource	LncRNA genes
RefSeq12	17,948
GENCODE11	19,933
NONCODE37	96,411
FANTOM CAT38	27,919
LNCipedia39	56,946
miTranscriptome40	58,648
CHESS13	17,623
LncBook41	95,243

The overlap between these annotation databases is relatively low³⁵, illustrating how far we are from a consensus on the identification of ncRNA genes. This rather fragmented landscape has nonetheless delivered an impressive achievement in charting the enormous variety of noncoding RNA genes.

Other challenges to ncRNA annotation

A variety of evidence suggests that ncRNA catalogues remain incomplete in a number of ways, and the community is still far from agreement on the true number of ncRNA genes and the true number of transcript isoforms. These issues arise from a variety of sources. First, the transcriptomic datasets from which most ncRNAs are derived originate from a non-exhaustive set of tissues/cell types that are over-represented by adult organs, cell lines and tumors. Rare but important cell types (e.g., tissue stem cells) or difficult-to-access developmental timepoints (e.g., embryonic stages) are poorly represented. This leads to incomplete sampling of existing gene loci and transcript isoforms. Second, the majority of transcriptomic data is produced using oligo-dT reverse transcribed RNA, which largely omits less-studied transcripts such as non-polyA and circular RNAs, although different approaches have been used to circumvent these issues (e.g.⁴²). Third, incomplete reverse transcription of cDNA gives rise to transcript models with inaccurate 5’ ends, and RNA degradation (which affects major organs at different rates post mortem) can lead to fragmented annotations and incorrect transcription start site (TSS) annotation.

The unique biology of ncRNAs also contributes to the challenges of annotating them. Current evidence indicates that they tend to be expressed at low levels⁴³, although this might be explained by technical biases in bulk RNA sequencing⁴⁴, or in very specific cell types and tissues, leading to relatively infrequent sampling compared to protein-coding RNAs. Their splicing and post-transcriptional processing tends to be as complex as that of protein-coding genes, leading to an ensemble of transcript isoforms that confuses short-read assemblers and human annotators alike⁴⁵. Note that these same features might also be true of non-functional (noisy) transcripts.

Annotation quality strongly affects our classification of the coding/non-coding biotype of RNAs and interpretation of their biological roles. Non-coding gene annotation efforts are complicated by the fact that some ‘non-coding’ loci in fact encode previously-overlooked protein products. A small but finite fraction of lncRNAs encode ‘micropeptides’ <100 amino acids in length that play diverse and important biological roles²⁵. Their small size confounds conventional open reading frame (ORF) discovery pipelines but may be identified by ribosome profiling or evolutionary signatures of protein conservation using PhyloCSF⁴⁶. Examples of lncRNAs that have been reclassified in this way include the widely-studied TUG1⁴⁷, whose ORF was only discovered after sequencing of a previously overlooked exon, highlighting how incomplete annotation can lead to misclassification of protein-coding status.

While some small ORF-encoded micropeptides display clear cellular phenotypes, the majority remain functionally uncharacterized and it is conceivable, particularly for those lacking significant evolutionary conservation, that they could represent either non-functional ‘translational noise’ or else early steps in the evolutionary birth of novel proteins. Although evidence of translation from some lncRNAs can be detected by mass spectrometry or ribosome profiling, the assignment of function to any such micropeptides will require further biochemical validation. In a converse fashion, protein-coding loci generate substantial numbers of non-protein-coding transcript isoforms, some of which have been shown to be functional; e.g., MYH7b⁴⁸. This blurring of the boundary between ‘coding’ and ‘non-coding’ will present a fascinating challenge to future annotation efforts.

Another challenge arises from the dissonance between standard annotation schemas, involving clearly defined, yet arbitrarily defined genes and transcripts, with the messy biological reality of ncRNA transcriptional units. Conventionally, genes are defined as the union of all overlapping transcripts at a locus, and neighboring genes are separated by a clear gap. These definitions worked well in the past. However, with the advent of deep and comprehensive long-read RNA sequencing, annotations are approaching a point at which read-through transcription events will begin to unite nearly all pairs of neighboring genes. Following classical gene definitions, the result could be a single “super gene” on each chromosome^49,50, which is clearly not a useful abstraction.

Functional annotation

One of the biggest challenges in ncRNA annotation relates to adding functional labels. For protein-coding genes, we have a rich amount of prior functional evidence, in addition to powerful computational methods for predicting gene function based on primary sequence. For example, DNA-binding transcription factors or membrane-bound receptors can often be predicted from translated amino acid sequences. In contrast, we know little about the vast majority of ncRNAs, and have no validated means of predicting function from sequence. Thus, one near-term goal for annotation of ncRNA genes will be describing the different types of evidence supporting them (e.g., tissue-specific expression levels), even though their function might remain unknown.

The majority of ncRNAs have not been properly assayed for function⁵¹: in initial efforts, 10–40% of selected ncRNAs showed some sort of function including effects on cell morphology and proliferation⁵². Scaling these approaches will require coordinated large efforts, including bulk assays in cell models and single cell sequencing from tissue or organoid models. Even so, the weak phenotypes observed when perturbing some ncRNAs, their tissue and cell type specificity, and the growing number of ncRNAs still being discovered, pose huge challenges that will require new technologies to create genome-scale assays. New methods for large-scale screening for interactions with other cellular compartments (e.g., chromatin) will be needed to address these challenges.

To date, many ncRNAs have been assigned names or biotypes that imply some function⁵³; in particular, ncRNAs are often named after a nearby or overlapping protein-coding gene. For example, FAS-AS1 is an anti-sense (AS) transcript whose name reflects its overlap with the protein-coding gene FAS. This may lead to confusion amongst users, because the lncRNAs in question may not have a function related to that of the neighboring protein-coding gene.

Health and medical annotation

A key application of human gene annotation is its use in diagnosing and treating genetic disease. Over five thousand genes and many thousands of variants of those genes have been associated with single gene disorders and disease risk, as catalogued in OMIM⁵⁴. For example, the BRCA Exchange database (https://brcaexchange.org/) currently lists over 34,000 variants in the BRCA1 gene alone, of which 2,228 are labeled as pathogenic⁵⁵.

When assessing variant pathogenicity in a clinical setting, the completeness and accuracy of gene and transcript models is essential. The impacts of variants as determined by programs such as Annovar⁵⁶ and Variant Effect Predictor⁵⁷ depend on the predicted open reading frames of transcripts. Further, designs of oligonucleotide baits and PCR primers used in targeted capture sequencing for clinical diagnostic assays depend on the correct annotations of exon boundaries. Even when whole-genome sequencing (WGS) is used for diagnosis, clinicians do not consider unannotated exons as candidates for interpretation.

Flaws in annotation can lead to serious errors in the clinic. Among many examples that might be cited, one case of a false negative diagnosis was caused by missing exons in a transcript of CDKL5, in a proband with seizures who was ultimately diagnosed by WGS after reannotation detected the missing exons^58,59. Another striking case led to a new diagnosis of Dravet syndrome after reannotation of an isoform of SCN1A revealed that the original annotation was missing a “poison” exon. In that case, the patient had splicing variants leading to expression of the nonfunctional isoform⁵⁹.

The need for a clinical standard

Currently, clinical laboratories commonly use RefSeq transcripts as a reference to report variants in well-known disease-linked genes, typically relying on reports from the literature. When the literature is unclear, laboratories tend to choose a transcript using simple criteria such as length or first appearance in annotation databases. This practice is not ideal as the chosen transcript might not reflect the properties needed for clinical diagnosis and leads to inconsistency among different laboratories. To further compound this issue, clinical laboratories commonly still map variant data onto the previous reference genome GRCh37. To realize the full potential of genomic diagnostics, there is a need for a universal transcript reference for every protein-coding gene.

The MANE collaboration, launched in 2018, addressed this need by generating a set of representative reference transcripts (MANE Select) to be used as universal reporting standards. For a small number of genes, the database provides a second transcript labeled as ‘MANE Plus Clinical’ when one transcript alone is not sufficient to report known clinical variants. The MANE set now covers more than 98% of human protein-coding genes and provides a logical starting point for clinically important gene and transcript annotation. However it would be useful in the future to add clinically important annotations of noncoding RNAs and regulatory elements to MANE, at least for those that have been associated with genomic variants linked to disease risk^60–62. Clinical interpretation and reporting will also benefit if other databases map their contents to the MANE standard. It is worth noting that the MANE set is anchored on GRCh38, so clinical databases that still use GRCh37 need to make the transition to GRCh38 to get the most out of this set. In addition to using a standard transcript for variant reporting, laboratories and databases also need to use standardized descriptions of genetic variants⁶³ to ensure unambiguous mapping to reference genomes.

Consistent annotation across multiple reference genomes

The hg19 (GRCh37) genome was replaced in 2014 by GRCh38. Although both reference genomes are still in use, they differ in many ways: their coordinates are different, some genes are missing from the older version, gene symbols have changed, and many genes have different exon-intron structures. Even for genes that are unchanged across the two releases, there is no standard way to translate coordinates between genomes without creating artifacts.

The advent of a truly complete human genome sequence, T2T-CHM13, promises to provide much more stability in gene coordinates.³ Looking ahead, we are likely to have many reference genomes for different human sub-populations. We already have annotated reference genomes for Ashkenazi⁶⁴, Puerto Rican⁶⁵, and Han Chinese individuals⁶⁶, and many more are likely to be produced. Another approach is to create a single “pan-genome” representing all populations, and the first draft of this approach, using a sequence graph based on 47 diploid assemblies of a genetically diverse individuals, has just appeared⁶⁷. Ultimately, we need a gene-centric alternative to gene symbols for referring to the same gene, and a coherent system to denote the same variants on any human reference genome.

Technology to finish the human gene catalogue

Finishing the human gene catalogue will require innovative new technologies to address the challenges ahead, such as resolving the functional relationships between gene products in a diversity of tissues, cells, and developmental stages. Here we touch on a few technologies that are available now or that may be available soon to solve these problems.

Matched long-read sequencing and proteomic analysis of gene products.

Genome-wide measurements of when and where specific isoforms are expressed are currently needed. Measuring gene expression within tissues and at single-cell resolution has already revealed many coordinated patterns of gene expression in cells and tissues⁶⁸. However, cell-specific splicing estimates from these studies remain problematic⁶⁹, and the number of splicing events is likely underestimated⁷⁰.

RNA-seq analysis at the isoform level currently relies on differential expression of exons within a gene⁶⁹, which is highly dependent on the method of library construction and on sequencing depth⁷¹. Even when expression levels are measured accurately, the relative abundance of a transcript does not correlate perfectly with translation⁷². Ribosomal profiling is a powerful method for measuring the translation of protein-coding isoforms, and it can validate engagement with the translational machinery for many predicted alternate isoforms. Interestingly, transcript analysis from ribosomal scanning and translation complexes at polysome fractions predicts large numbers of unannotated small ORFs^73,74, and these need further exploration to determine if they represent valid functional genes. These predictions might be resolved through full-length sequencing, preferably directly from RNA molecules, coupled with further ribosomal profiling or other methods for detecting translation.

While sequencing with single molecule technologies (e.g., Oxford Nanopore Technologies (ONT) and Pacific Biosciences (PacBio)) is capable of providing full-length direct RNA and cDNA sequences, relatively few experiments to date have used these technologies to survey the RNA landscape from each human cell type. Other confounding issues concern sequencing the poly-A terminating RNAs, using ONT oligo-dT reverse transcription primers and oligo-dT linker ligation⁷⁵. Strategies are being developed to specifically capture total RNA that will use RNA ligases to add a primeable sequence at the true 3’ ends of all RNA transcripts. Another approach uses artificial poly(U) tailing to add a primeable sequence to both capped RNAs and non-capped RNAs^76,77. Information on RNA modification, which can be measured from ONT direct RNA sequencing data, will likely provide a powerful new type of functional annotation.

Validation of protein-coding isoforms will ultimately require protein detection through deeper proteomics sequencing as is currently ongoing⁷⁸ or through other means. Meta-analyses of proteomic data rely heavily on the quality of the transcriptome reference to identify peptides mapping to putative isoforms. However, when coupled with new long-read technologies, dual proteome-transcriptome assemblies are finding evidence of higher isoform diversity than predicted from a representative transcript approach, by resolving peptide fragments that would otherwise fail to map unambiguously to a gene or single isoform. In one recent study, 30% of the gene products identified using a dual PacBio-mass spectrometry approach were distinct isoforms from the same locus, which included thousands of examples where the alternate isoform was not measurable using mass spectrometry alone⁷⁹. We anticipate that soon, progress in long-read technologies will produce more reliable maps of full-length transcript isoforms, quantifiable isoform switching, and isoform dosage at the resolution of individual cells.

Methods to capture low-expressed transcripts.

Capture sequencing has recently been adapted to target specific RNAs, in order to provide higher sequencing coverage for selected regions of the genome using short- and long-read RNA sequencing in a high-throughput manner^34,80. This is particularly useful to enrich for RNA from ultra-low input samples⁸¹ and from genes expressed at very low levels. The use of capture technologies, together with recent increases in the throughput of long-read sequencing platforms, could enormously benefit the study of low-expressed transcripts, particularly lncRNAs, which in turn may be vital for the study of gene regulation in both normal and diseased cells.

Conclusions

Over 20 years after the original publication of the human genome, the number of protein-coding genes is stabilizing around 19,500 (Figure 2), although the number of isoforms of these genes is still a subject of intensive study and discussion. The completion of a human genome sequence itself offers the opportunity to map these genes onto a stable, finished sequence and converge to a final number in the next few years. Greater standardization of gene and isoform annotation will improve our ability to apply this knowledge in a clinical setting.

Figure 2:

predicted and observed human gene counts over time. Counts of protein-coding, pseudogene, and non-coding genes are shown. Time points before 2003 and after 2023 (dashed lines) represent an average of predictions from the literature¹ and extrapolations from this perspective article, respectively. Time points from 2003 to 2023 are based on 20 iterations of the NCBI RefSeq annotation of the human reference genome, including both curated and predicted genes.

In contrast, noncoding RNA genes, particularly lncRNAs, are at an earlier stage of understanding, and are still increasing in number, with current catalogs containing 17,000–20,000 lncRNAs or more. New technologies offer promising avenues to refine this catalog, although a complete functional characterization of lncRNAs is likely many years away. The steady decline in the number of protein-coding genes over the last 20 years makes it only natural to ask if lncRNA numbers may follow a similar trend, as our knowledge of RNA biology and technologies improve.

Finally, we note that even with a complete gene annotation of a finished genome, we will have only one example of the human gene catalogue, one that will not apply to all humans. It is likely that many healthy individuals have more or fewer copies of some genes, and future efforts to survey the diversity of the human population will be an important step towards achieving a more complete view of the gene content of our genome.