The Human Pangenome Project: a global resource to map genomic diversity

Preface

The human reference genome is the most widely-used resource in human genetics and is due for a major update. Its current structure is a linear composite of merged haplotypes from more than 20 people, with a single individual comprising most of the sequence. It contains biases and errors within a framework that does not represent global human genomic variation. A high-quality reference with global representation of common variants, including single nucleotide variants (SNVs), structural variants (SVs), and functional elements is needed. The Human Pangenome Reference Consortium (HPRC) aims to create a more sophisticated and complete human reference genome with a graph-based, telomere-to-telomere representation of global genomic diversity. We will leverage innovations in technology, study design, and global partnerships to construct the highest-possible quality human pangenome reference. Our effort will improve data representation and streamline analyses to enable routine assembly of complete diploid genomes. With attention to ethical frameworks, the human pangenome reference will contain a more accurate and diverse representation of global genomic variation, improve gene-disease association studies across populations, expand the scope of genomics research to the most repetitive and polymorphic regions of the genome, and serve as the ultimate genetic resource for future biomedical research and precision medicine.

Need for a Complete Human Pangenome Reference

The human reference genome is the foundational open-access resource of modern human genetics and genomics, providing a centralized coordinate system for reporting and comparing results across studies^1–4. Its release set the bar for genomic data sharing, essential for nearly all human genomics applications, including alignments, variant detection and interpretation, functional annotations, population genetics, and epigenomic analyses. The current human reference (GRCh38.p13) is a mosaic of genomic data assembled from >20 individuals, with ~65% of the sequence contributed by a single individual^5,6. Dependence on a single, mosaic assembly (which does not represent any one person’s sequence) creates reference biases, adversely affecting variant discovery, gene-disease association studies, and the accuracy of genetic analyses^7,8. More than two decades after the first human genome reference sequences were released, the current reference genome still contains errors, rare structural configurations that do not exist in most human genomes, and gaps in regions that have been difficult to assemble^9,10 because of their repetitive and highly polymorphic nature. The human reference genome, like most technology-driven resources, is overdue for an upgrade¹¹.

For years the Genome Reference Consortium (GRC) has updated the linear reference by fixing errors, filling in gaps, and adding newly discovered variants^1,4,9,12. When enough changes accumulate, new builds are generated and released. While this process has served the community well, shortcomings have been identified along the way. Segments of genome sequences sampled from individuals may differ considerably from the reference genome, leading to errors in read mapping to the reference and reducing the accuracy of variant calls^13,14. Identification of structural variants (SVs; >50bp deletions, insertions, tandem duplications, inversions, and translocations) relies on detecting patterns of discordant read pairs or split read alignments, which in turn depend on the accuracy of read mapping^15,16. Assembling and detecting these SVs are challenging when the reads are too short to cover long, repetitive regions of the genome⁷. This is because short reads (50–300bp) may be identical and/or overlapping with one another such that it is impossible to determine where they should map. Both the limitations of short reads and reference biases mean that we may have missed >70% of structural variants in traditional whole genome sequencing studies^17,18.

Advances in sequencing technologies and a greater appreciation for the importance of genetic diversity make improving the human reference sequence both timely and practical. First, the development of long-read (>10kb) sequencing technologies has enabled the assembly of large, repeat-rich regions, facilitated phasing and assembly of maternal and paternal haplotypes, and improved representation of GC-rich regions of the genome that are often missing in short-read assemblies^19–22. Second, growing recognition of the importance of diversity and inclusion in human genomics²³ has led to widespread calls to improve representation and methods for detecting and presenting global variation.

In this Perspective, we outline the goals, strategies, challenges, and opportunities for the Human Pangenome Reference Consortium (HPRC). We will engage scientists and bioethicists in creating a human pangenome reference and resource that represents genomic diversity across human populations while improving technology for assembly and developing an ecosystem of tools for analyses of graph-based genome sequences. This new reference will maintain essential ties to the original reference for continuity, even as we strive to develop complete and error-free, telomere-to-telomere (T2T) assemblies of all chromosomes of individual human genomes, referred to here as ‘haplotypes’.

Goals and Strategies of the HPRC

A “pangenome” is the collective whole of genomic information for a given species. Originally popularized in the context of highly dynamic bacterial genomes²⁴, the concept has been adapted to the field of human genomics, in which the full extent of human genomic variation is expected to be much broader than has thus far been revealed. The pangenome data infrastructure depends on the high-throughput production of high-quality, phased haplotypes (segments of a chromosome identified as being maternally or paternally inherited) that improve upon the current human reference genome. Highly accurate and complete haplotype-phased genome assemblies will be organized into a graph-based data structure for the pangenome reference that compresses and indexes information^25–27. This data structure will contain a coordinate system with a simple, intuitive framework for referring to genomic variants while preserving backward compatibility with GRCh38 and prior linear reference builds. Managing and interpreting these data requires trans-disciplinary collaboration and innovation, focused on the development of novel conceptual frameworks and analytic methods to construct the pangenome infrastructure and tools for downstream analyses and visualization. The HPRC’s goals are laid out in Box 1.

Box 1: Goals of the HPRC.

• Identify individuals representing diverse genomic and biogeographic backgrounds to include in the pangenome reference, with at least 350 reference quality haplotype-phased human diploid genomes (700 haplotypes total).

• Integrate ethical, legal, and social implications (ELSI) scholarship in the development of recommended policies and protocols for inclusion, data acquisition, and stewardship from study recruitment to publication of findings.

• Prioritize the use of long-read and long-range technologies and assemblies with haplotype-aware algorithms to generate the highest quality phased genomes possible.

• Create methods to finish diploid genomes from telomere-to-telomere across complex regions, closing gaps and ensuring hard-to-measure variants are identified.

• Foster an ecosystem of pangenome reference tools to facilitate the annotation of genes and other genomic features.

• Implement an iterative design-development-engagement process to understand and respond to user community needs.

• Develop communication strategies that will assure understanding of the pangenome reference resource, including the community’s ability to fix and report errors.

• Enable appropriately controlled access to data through genomics platforms such as the International Nucleotide Sequence Database Collaboration (INSDC)^28,29, the National Center for Biotechnology Information (NCBI), the UCSC Genome Browser^30,31, Ensembl^32,33, the WashU Epigenome Browser^34,35 and NHGRI’s cloud-based analysis platform, AnVIL^36,37.

• Foster an international human pangenome reference alliance that actively engages the diverse populations it seeks to represent.

The HPRC functions through multi-disciplinary collaborations, convening cross-institutional and multi-national working groups dedicated to sample collection and consent, population genetic diversity, technology and production, phasing and assembly, approaches to construction of a pangenome reference, resource improvement and maintenance, and resource sharing and outreach (Figure 1). The HPRC has begun the process of engaging international partnerships with the Australian National Centre for Indigenous Genomics (NCIG)³⁸, the FDA-recognized Clinical Genome Resource (ClinGen)³⁹, the NIH-funded Human Heredity and Health in Africa (H3Africa) Consortium⁴⁰, the Personal Genome Project (PGP)⁴¹, the Vertebrate Genomes Project⁷ and the Global Alliance for Genomics and Health (GA4GH)⁴². The HPRC will integrate perspectives from the international scientific community through these collaborators and others yet to be identified to inform the development of HPRC references, methods, and standards.

Figure 1: The Human Pangenome Reference Consortium.

This diagram provides an overview of the HPRC’s several components. Collect: I,000 Genomes samples jump-start the project and will be followed by additional samples collected through community engagement and recruitment. Sample selection efforts will ensure the graph-based reference captures global human genomic diversity. Sequence: Long-read and long-range technologies are used to generate genome graphs and bridge gaps in difficult-to-assemble genomic regions. Assemble: Telomere-to-telomere finished diploid genomes will foster variant discovery, especially in complex, difficult to assemble genomic regions. Construct: Scalable bioinformatics approaches assemble, QC, call variants, and benchmark graph assembly accuracy. The graph is annotated with gene descriptions and transcriptome data, making it more accessible and interpretable. Utilize: Collaboration across scientific and stakeholder communities will create a new ecosystem of analysis tools. Clinical applications and research use will involve analysis, validation, interpretation, and publication of results. Outreach: Members of the HPRC Outreach community engage and educate the user community and broadly share all genomic products and informatics platforms. ELSI: ELSI scholars will develop selection processes and policy frameworks that meet investigator needs while respecting research partner autonomy and cultural norms.

Inclusion Criteria

In the first phase of the project, HPRC investigators selected individual genomes for high-quality sequencing among existing cell lines established by the 1000 Genomes Project (1KG), which offers a deep catalog of human variation from 26 populations⁴³. These cells were originally collected from volunteer donors using consent procedures designed for unrestricted data use, and the cell lines are available in the NHGRI Biorepository at Coriell⁴⁴. The selected cell lines were prioritized based on a combination of criteria, ranging from genetic and geographic diversity of the donors, to availability of relevant parental data (for haplotype-phasing), and limited time in cell culture (to minimize the accumulation of de novo mutations).

Individuals were initially identified using clustering/visualization techniques (UMAP clusters generated from 1KG data) and observed allelic diversity (heterozygosity) then selected for inclusion in the first phase of HPRC (N=100). Our inclusion criteria and recruitment strategies are evolving with the project, and we recognize that there are inherent limitations to clustering algorithms and using only the 1KG dataset.

While useful for the first phase of the HPRC, genomes selected from the 1KG data represent a limited scope of geographic and genomic diversity. One reason is that the resource was developed by sampling in 26 geographic locations across the globe, and the discrete number of individuals included from each location limits the amount of genomic variation representing those regions, especially regarding rare variants that are less likely to be observed in small sample sizes. The genomes of individuals sampled from each 1KG location cannot be assumed to harbor sufficient variation to be comprehensive of the genomic diversity in the natural population of the region, let alone to represent an entire continent. Furthermore, 1KG populations were often selected by asking potential study participants questions about their racial, ethnic, or ancestral identities, assigning ancestry based on geographic location, or some combination, which would not necessarily produce a representative sampling of any natural population. Since population descriptors can be inconsistent on clinical forms⁴⁵ and are fluid across cultural contexts^46–48, there are many unknown layers of diversity within each geographic sampling cohort of 1KG data.

Because 1KG data are insufficient to support the ambitious sampling and genetic diversity goals of the HPRC, the consortium will include additional genomes from participants identified through the BioMe Biobank at Mount Sinai and a cohort of African Americans recruited by Washington University. Those participants will be invited to contribute to the HPRC, with informed consent that comports with de-identified open data and creation of cells lines at Coriell. In later phases, the HPRC will foster additional domestic and international partnerships to explore additional avenues to broaden diversity and enhance inclusion (Box 2).

Box 2: Commitment to Diversity and Inclusion.

There are many aspects of diversity to consider for broad inclusion, and the first step is to assess current gaps in diversity. Researchers have demonstrated a lack of diversity in genomics research using biogeographic ancestry groupings at the continental level, as well as socio-cultural categories such as racial and ethnic identities⁴⁹. It is important to distinguish biological and sociocultural diversity, since sociocultural labels are not derived from genotype data, and vice versa. Due to gaps in genomic sampling worldwide, the distribution of allelic variants appears to be correlated with continental-level biogeographic ‘ancestry’. However, variants are rarely unique to a single biogeographic ‘population’, and factors such as effective population size, founder events, and genetic drift are responsible for differences in allele frequencies between human groups.

Capturing the full range of human genomic diversity is a daunting task: some gaps are understood and predictable, but we also face unknown unknowns. The initial HPRC dataset cannot be comprehensive of global genomic variation, but it can set a foundation to build upon. The HPRC will produce high-quality genome data for 350 individuals (700 haploid genomes) selected to maximize global representation within the logistical constraints of the initial HPRC efforts. Strategic partnerships with organizations such as the GA4GH and H3Africa are underway, which we anticipate will help facilitate international engagement and broaden our understanding of the cultural, ethical, legal, social, and political considerations of the HPRC. However, further partnerships will be needed to include populations that are under-represented or entirely missing from current data resources. HPRC actively welcomes additional partners and collaborators to join us in rising to this challenge.

Embedded ELSI Scholarship

Most human genomics has been based on individuals of European ancestry, and the datasets available for analyses are thus biased. At present, precision medicine is based on genomic variation found in populations with primarily European ancestry. Much of the global genetic diversity that contributes to clinical phenotypes is missing from clinical genetic tests. Many ethical, legal, and social challenges arise in efforts to include hitherto excluded populations, communities, or groups.

The HPRC has formed an ‘embedded’ team of scholars to address Ethical Legal and Social Implications (ELSI) of its work, with expertise at the intersections of genomics with biomedical ethics, law, social sciences, demography, community engagement, and population genetics. The main objective of the HPRC-ELSI team is to identify, investigate, and ultimately offer consortium investigators advice about the issues they face, which must be addressed if the HPRC is to meet its goals. In the embedded model, with ELSI scholars participating in key meetings where decisions are made, investigators can engage these colleagues in discussions that deepen their understanding and appreciation of what is at stake as we seek to improve the human reference genome.

Large-scale human population genetics projects aimed at broadening the diversity in genomic datasets and analyses have often missed the mark in demonstrating respect for individuals and communities. The Human Genome Diversity Project (HGDP) encountered strong opposition three decades ago⁵⁰, facing objections that its approach was extractive and its goals benefited scientists and rich country institutions, but did not match the priorities of Indigenous Peoples or people in resource-poor regions who were asked to donate their samples and data. The Havasupai Tribe of northern Arizona sued the Arizona Board of Regents in 2002, when they learned that samples donated for diabetes research were shared with other researchers and re-used for studies of schizophrenia and population origins, to which tribal members did not agree. That case was settled in 2010^51,52, but the impact it had on relations between tribal communities and genomics research has persisted.

More recently, the Wellcome Sanger Institute was criticized for licensing access to data arising from southern African samples, while institutions in Africa asserted terms of informed consent did not permit commercial uses. The NIH was also criticized for inadequate tribal engagement and consultation in the All of Us program^53,54. With a keen awareness of this history, the HPRC is initiating a process to consult with, engage, and genuinely include groups currently not well represented in genomic database. Indigenous scholars have spearheaded a movement for Indigenous data sovereignty⁵⁵, for example, including the development of the CARE principles for Indigenous Data Governance⁵⁶ to layer onto the FAIR principles that support the open-science approach that the HPRC and similar projects take⁵⁷. The HPRC is reaching out to Indigenous geneticists, leaders, and community members to engage and collaboratively develop a truly global and inclusive reference resource, taking into account FAIR and CARE principles. Furthermore, similar efforts will be made for other diverse populations that the HPRC will work with.

Some groups we seek to engage may develop sampling and sequencing efforts parallel to, rather than directly participating in the HPRC. In developing a state-of-the-art pangenome reference sequence, the HPRC will continue to disseminate standards for accuracy and completeness of sequencing, while emphasizing the importance of ELSI considerations. It is a priority for the HPRC to actively communicate with parallel genome-wide sequencing efforts to ensure compatibility between efforts, thus enabling integration into a global pangenome reference resource. HPRC is committed to assessing local policies and promoting broad sharing of resources developed through interdisciplinary engagement, scholarship, and innovative technical solutions. The HPRC will establish procedures for navigating potential tensions among its technical, research, and resource-generating objectives with local customs, laws and data sharing policies for the groups within the HPRC as well as for those in the parallel projects.

Initial Data Generation and Release

Technological advances in genomics enable sequencing long repeats, physical mapping to chromosomes, and phasing maternally and paternally inherited haplotypes (Box 3).

Box 3: Sequencing and Assembly.

Notable improvements in long-read technologies have resulted in complete chromosome assemblies^20–22 and have demonstrated the ability to broaden variant analysis to span large, complex human SVs⁵⁸. Use of highly accurate consensus reads (99.9%, or Q30) of moderate length (e.g. 10–20 kb), such as high fidelity (HiFi) reads from Pacific Biosciences (PacBio), routinely resolves long tandem repeats, or satellite arrays, and large segmental duplications^20,21,59. In parallel, the nanopore-based sequencing platform (Oxford Nanopore Technologies) offers long-read data that routinely generate substantial coverage of reads that are 100s of kb in length (or “ultra-long” data, UL) with an increasing number of reported reads greater than one million bases. Like the HiFi data, UL data is used to close large and persistent assembly gaps, including in human centromeres, subtelomeric regions, and large segmental duplications^22,59. Further, chromosome conformation capture methods produce long-range data for both short-read (Hi-C⁶⁰) and long-read sequencing (Pore-C⁶¹). Such chromatin crosslinking protocols generate chimeric DNA fragments from interacting chromosomal regions that are covalently linked together. These ligated DNA molecules are sequenced to help determine phasing and spatial organization at the level of an entire chromosome. With continuing gains in both HiFi read length and nanopore single read base-level quality, and improved methods for the use of chromosome conformation capture methods to guide phased haplotype assembly, we are entering into a new era of routine complete chromosome-level assemblies^20–22,37. In collaboration with the T2T Consortium, which aims to use long-read sequencing and cutting-edge algorithmic approaches to close the hundreds of gaps persisting in the human reference genome²², the HPRC will generate accurate assemblies of entire chromosomes. These assemblies will empower us to characterize variations in large, repeat-rich regions that have historically been out of reach for standard genetic analysis and interpretation.

For the initial phase of the project, we sequenced a single individual, HG002, whose genomic sequence has been thoroughly characterized by the Genome in a Bottle (GIAB) consortium⁶². We evaluated multiple sequencing technologies and assembly algorithms to identify the optimal combination of platforms and develop an automated pipeline that generated the most complete and accurate genome representation (Figure 2)⁶³. We began with the now well-established assumption that long reads (>10kb) yield more complete genome assemblies than short reads alone⁷. The technologies tested included PacBio and/or ONT long reads for generating contigs, 10X Genomics linked reads, Hi-C paired reads, Strand-Seq long reads, and/or Bionano optical maps for scaffolding contigs into chromosomes. This pilot benchmark study created the standards for sequencing technologies and computational methodologies critical to HPRC’s success.

Figure 2: Standards were developed through a pilot benchmark study of one individual.

Multiple long-read and long-range technologies and computational methods were evaluated to develop the combination of platforms and an automated pipeline that provides the most complete and accurate genome graph.

We found that the trio approaches using parental short-read sequence data to sort haplotypes of the offspring’s long-read data gave the most complete assemblies of each haplotype with the fewest structural errors⁶³. Further, all methods attempting to separate haplotype sequences performed much better in generating highly contiguous assemblies than those that merged the consensus between haplotypes into one assembly. The algorithm that gave the highest haplotype separation accuracy for contigs was HiFiasm⁶⁴, which incorporates separation of reads of each haplotype into the assembly graph⁶⁵. Generation of contigs were more structurally accurate than scaffolds, where the HPRC identified areas of improvements necessary to prevent contig miss-joins, missed-joins, collapsed repeats, and other structural assembly errors. Based on these findings, an initial set of 47 1KG genomes from parent-offspring trios was assembled with Hifiasm⁶⁴, creating high quality diploid contig-only genome assemblies. Going forward, we will further optimize sequencing, assembly, and analysis methods with the goal of creating fully-phased T2T diploid genomes, including repetitive and structurally variable regions such as centromeres, telomeres, and segmental duplications. We anticipate that the high-quality assemblies created in the project will drive tool creation and improvement for diploid genome assembly and QC where new and recently created existing tools (from the T2T assembly of CHM13²²) are applied to diploid genome assembly.

The first HPRC data release is comprised of the sequencing data from 47 participants – mostly from the 1KG Project (listed and described in Supplemental Table 1). All sequencing data are publicly available and can be downloaded without egress fees from the Amazon Web Services (AWS) Public Datasets program and can be analyzed with the AWS cloud. Data is also available for analysis within the AnVIL cloud platform, organized as a public workspace⁶⁶. AnVIL is the NHGRI’s genomic data science Analysis, Visualization, and Informatics Lab-space that provides a cloud environment for analysis of large genomic datasets, and supports multiple globally used analysis tools including Terra, Bioconductor, Jupyter, and Galaxy³⁶.

Pangenome Reference

We are building a pangenome reference with three complementary parts: 1) the haplotypes: the sequences within the input assemblies; 2) the pangenome alignment: a sequence graph and an efficient embedding of each of the input haplotypes as paths within this graph; and 3) the coordinate system: a backward-compatible coordinate system and set of sequences that make it possible to refer to all variations encoded within the reference equally (Figure 3). The haplotypes provide hundreds of individual representations of the genome, spanning global diversity. Each haplotype assembly will be useful individually as a reference for studying genomic sequences that are divergent from the current human reference assembly. The pangenome alignment represents the homology relationships among the individual assemblies. This canonical alignment will support coordinates translation (liftOver) between the haplotypes and defines the allelic relationships. It will form the substrate for many emerging pangenome tools and pipelines that will improve important genomic workflows, for example, by making genotyping accuracy less dependent on ancestry. The coordinate system provides a global, unambiguous means to refer to all the variations within the pangenome. It makes all the variations within the haplotypes first class objects that can be referred to equally. Ultimately, it will provide a more complete means to refer to variations not contained within the existing linear reference, proving useful for databases and tooling that will build upon the pangenome reference.

Figure 3: The Human Pangenome Reference.

Graph-aware mappers can be used to genotype samples by directly mapping against the graph. This simplified example shows how to create a pangenome graph for four people and calculate the allele frequency of three variants. Iterating through each individual produces the graph structure, which improves as new genomes are added. Genomic data is arranged into a sequence variation map based on edges. Alternative haplotypes are depicted as alternate pathways across the graph, with the edges being the primary data-bearing elements. The pangenome reference catalogs genomic variation and allows for population- scale analysis thanks to its graph structure. Tracing a path through the network and connecting sequences at access edges yields haplotypes for individuals. For clinical interpretation, allele frequencies are reported.

Supporting these parts is a new proposed set of file standards⁶⁷, notably the rGFA format for representing a pangenome and the GAF format for representing read mappings to a pangenome. We hope these will have an impact on the field similar to how SAM/BAM⁶⁸ and VCF⁶⁹ formats generated a broad range of interoperable tools that became widely used and accessible. To kick-start this process, we have developed the vg toolkit⁷⁰ and minigraph⁶⁷, which incorporate downstream tools for graph construction and long- and short-read mapping and genotyping.

We anticipate releasing an alpha pangenome reference based on existing variant calls and assembled contig genomes. Using the proposed incremental coordinate system, we will subsequently release updated graphs incorporating the growing numbers of assemblies.

Variant Detection

A central aim of this research is to document the genetic similarities and differences among the human genomes included in the pangenome reference. Comprehensive variant detection, however, is still a challenge even when high-quality genome assemblies are available. No single data type or bioinformatic approach yet achieves high performance across all variant classes and genomic regions^58,71. Therefore, we are pursuing multiple complementary approaches to variant detection using a combination of whole-genome multiple assembly alignment, pairwise assembly-assembly alignment, and traditional reference-based read alignment.

Ideally, we will accomplish variant detection in a single step designed to build pangenome graphs directly from whole-genome, multiple assembly alignments. Genetic variants will be represented naturally as features in the resulting graph because any variant would be captured by the assembly process. This offers a significant advantage, enabling optimal breakpoint reconstruction via joint analysis of all input genomes. Accurate multiple-alignment and graph construction of entire human genomes is extremely challenging, but recent improvements to tools such as minimap2⁷², minigraph⁶⁷, cactus⁷³ and pggb⁷⁴ make this feasible. However, variant calling errors can still arise from assembly and sequence alignment errors, especially in repetitive regions of the genome. Given this, and the fact that pangenome graph construction tools have not been thoroughly evaluated at scale with real-world data, we are also pursuing the complementary approaches described below.

An alternative approach to multiple-alignments is to map variants from pairwise assembly-assembly alignments. Towards this end, we are using minimap2 and Winnowmap to align each draft assembly to the GRCh38 and T2T-CHM13 references to perform variant detection of SNVs, indels and other SVs. This approach is more straightforward than whole-genome multiple-alignment; however, complications can arise from reference genome effects and the need to merge results across many pairwise comparisons. The exact coordinates of complex and repetitive variants may differ due to alignment ambiguity. To alleviate reference effects, we are mapping variants via pairwise alignment of the two haploid genomes of each individual, enabling detection of natural heterozygous variants within sequences that are missing or poorly represented in GRCh38 and CHM13. Pairwise alignment assembly methods help control for potential errors from the multiple-alignment and graph construction process outlined above; however, they still fail to detect variants that are not captured in the underlying assemblies. We are also running a host of traditional variant callers that rely on alignment of raw reads to the GRCh38 and CHM13 references to control for potential assembly errors. Though limited by reference genome quality and alignment accuracy, these traditional tools are able to capture a subset of variants that are not accurately assembled, and they will serve as a cross-check on newer and less mature assembly-based tools.

In sum, we expect the above methods to capture most genetic variants in genomic regions accessible to current assembly and alignment methods. We will compare our variant calls to published call sets from the 1KG Project⁷⁵, HGSVC⁷⁶, and GIAB⁶² using samples from these projects that are also included in the HPRC references to assess quality. We will evaluate and validate variant calls using independent data types generated by the HPRC but not used for contig assembly – such as ONT, Hi-C, Strand-Seq, and BioNano data – and assess the read-level support for each variant call based on alignment of raw data to assemblies and pangenome graphs.

Achieving comprehensive T2T variant detection across the entire genome will require improved methods for genome assembly, multiple-alignments and graph construction. The development and application of these methods in subsequent years is a major goal of the HPRC, and will help extend the impact of pangenomics to the full spectrum of variant classes.

Pangenome Annotation

Annotation of the current GRCh38 reference includes genes and genomic features, such as repeats, CpG islands, regulatory regions, and ChIP-seq peaks among others. The pangenome reference will have these same utilities and more, including the following:

Genes:

The two primarily used gene sets in genomic analysis are NCBI’s RefSeq⁷⁷, which exists as independent mRNA definitions, and EMBL-EBI’s Ensembl/GENCODE⁷⁸, which is built on the GRCh38 reference. The pangenome reference will support both the RefSeq and Ensembl/GENCODE gene sets. We will map both annotations to each haplotype. Specifically, we will evaluate the mapping of the core reference set of human transcriptome data to each haplotype and incorporate putative new genes that are not represented in either RefSeq or Ensembl/GENCODE. Mapping these gene sets in conjunction with other transcriptomic data sets will annotate the pangenome graph. Other tools will support spliced alignments and transcript reconstruction on a mature graphical data structure. We will integrate the results of these approaches into the annotation released for each haplotype, accompanied by a description of whether transcripts are identical by both methods or whether changes were identified, including transcripts that are disabled, duplicated, or missing on a given haplotype. We will also annotate all transcript haplotypes for their global frequency using Haplosaurus tools⁷⁹. We will initially annotate haplotype-by-haplotype, and will explore methods for direct annotation of the pangenome, such as those currently being developed in the GENCODE Consortium. Direct annotation methods simultaneously cover all relevant haplotypes and result in both an annotated genome graph and haplotype-specific annotations. One of the critical use cases of direct annotation of the pangenome will be large transcriptomic datasets aligned directly to the graphical structure that natively annotate it.

Functional Elements and Other Genome Features:

A central goal in biology is to understand how sequence variants affect genome function to influence phenotypes. Genome function includes regulatory regions that influence gene expression, enhancers that modulate expression levels, and the three-dimensional interactions that control chromosome structural organization within a cell. We will use the pangenome reference to annotate such functional information using existing RNA-Seq, MethylC-Seq, and ATAC-Seq datasets from Roadmap Epigenomics, ENCODE, 4D Nucleome (4DN), Genotype-Tissue Expression (GTEx), and Center for Common Disease Genomics (CCDG), among others. This will enhance the functional human genetic variation catalog.

Integrating functional data with the pangenome reference will facilitate the development of toolkits and analysis pipelines that evaluate the impact of genetic variants on complex traits and variation in phenotypes. The HPRC will work with developers to define rules and mechanisms to engage with multimodal ‘Big Bio-Data’ for both data providers and consumers. We will co-create user-friendly informatics platforms to manage, integrate, visualize, and compare highly heterogeneous datasets in the context of the genetic diversity represented in the pangenome. Box 4 lists available resources for working with pangenome graphs. We will also make all haplotype-by-haplotype annotation methods available in AnVIL so that others can run them to create custom annotation tracks on all or a selected subset of assemblies. These platforms will serve as a foundation for significant clinical datasets and global biobank initiatives that will ultimately improve precision medicine and medical breakthroughs. For example, the NHGRI is establishing the Impact of Genomic Variation on Function (IGVF) Consortium, which aims to develop a framework for systematically understanding the effects of genomic variation on genome function. Data generated by the IGVF will include high-resolution identification and annotation of functional elements and cell type-specific perturbation studies to assess the impact of genomic variants on function. The pangenome will be an important foundation for predicting functional outcomes in these studies.

Box 4: Pangenome Graph Tools.

Graph Building

• minigraph⁶⁷

• PGGB⁷⁴

Graph Aligners

• deBGA⁸⁰

• BGREAT⁸¹

• BrownieAligner⁸²

• GenomeMapper⁸³

• HISAT2⁸⁴

• VG⁸⁵

• GraphAligner⁸⁶

• GRAFࡊ⁸⁷

• PaSGAL⁸⁸

• SPAligner⁸⁹

Graph Indexing

• CHOP⁹⁰

• PSI⁹¹

Graph Visualization

• Bandage⁹²

• GfaViz⁹³

• SGTK⁹⁴

• AGB⁹⁵

• Sequence Tube Map⁹⁶

• MoMI-G⁹⁷

• VG view⁸⁵

• VG vis⁸⁵

• ODGI viz⁹⁸

Gene Prediction

• Path Racer⁹⁹

Variant Detection

• PanGenie¹⁰⁰

• Cortex with Bubbleparse¹⁰¹

• BayesTyper¹⁰²

• Paragraph¹⁰³

• GraphTyper2¹⁰⁴

• VG⁸⁵

Data Sharing

To enhance community access and sharing, we will submit sequence data (HiFi, ONT, Hi-C, and others), assemblies, and pangenomes produced by the consortium to AnVIL³⁶ and INSDC²⁸. Data will also be stored and made publicly available on both S3 and Google Cloud Storage. This general model supports future efforts to use cloud-based strategies for biological data analysis spanning multiple centers. Users of various clouds worldwide will know that they are using the same datasets. Data coordination within the Consortium will leverage the established methods in use and constant development since the inception of 1KG more than a decade ago^105,106. These processes will ensure that we rapidly release data in an organized manner, with proper accessioning of archival datasets, and future traceability of analysis objects and primary data items. Data stored in INSDCs will use BioProjects and BioProject umbrella structures similar to the 1KG and Vertebrate Genome Project⁷ to ensure that data are appropriately organized and easily identifiable in the public archives. This approach ensures that sample identifiers are effectively managed via the BioSamples database¹⁰⁷, including metadata provisions, and makes any data generated from the same samples readily tractable. INSDC will archive all reads, assembly data, and other relevant archives will be used, as appropriate for a specific data type. Each haplotype assembly will receive a Genome Collections accession number (GCA_*), which we will version as we make assembly updates. We will address additional data sharing considerations as they arise through our expanded recruitment and sampling efforts to broaden the diverse representation of global variation.

Adoption and Outreach

Achieving widespread international adoption of a pangenome reference will be a challenge¹¹. HPRC will design a pragmatic model and transition plan that is simple and compelling enough to gain traction among researchers and clinical laboratories. Working across scientific and other stakeholder communities, we will foster a new ecosystem of analysis tools. We will maintain and improve the reference, establish scalable bioinformatics methods for resolving errors, improve resolution in difficult-to-resolve genomic regions, and respond to user feedback. Importantly, we envision an integrated pangenome transition plan that involves broad community engagement via outreach and education, from tool developers to end-users. These efforts will create a software ecosystem and expert user base to support the next generation of human genetics. The pangenome reference will provide improved genomic research standards, data sharing, and reproducible cloud-based workflows. Understanding the barriers to adoption will lead to effective outreach and training, ensuring that the pangenome reference resource is widely adopted.

Adoption will ultimately be driven by the creation of a data resource that sustains continued improvement in its accuracy and completeness, enables a range of uses, and improves genomic analyses. We will actively publicize the benefits of using the pangenome. As a starting point for our outreach efforts, we have created a website (humanpangenome.org) to publicize the Consortium. We have also created human pangenome social media accounts that directly connect our Consortium with the end-user community (e.g. @HumanPangenome on Twitter).

To facilitate adoption, we will explore who the user community will be, their needs, and, most importantly, the technical and non-technical barriers they may encounter. Addressing potential obstacles is essential, since we know that adopting an updated version of the linear reference resulted in significant bottlenecks for many laboratories. The cost of switching can be significant, and HPRC is aware that many clinical labs worldwide still use the GRCh37 build from February 2009 for this reason. HPRC will examine how to reduce switching costs and expedite transition. User data will be collected in self-reporting surveys, including user characteristics, location, specific applications, and barriers to adopting a pangenome reference framework.

Creating a coordinate system that builds on GRCh38 and includes both GRCh37 and GRCh38 assemblies is central to user adoption. HPRC will develop training materials that explain the additional sequences included in the pangenome reference coordinates and how these sequences relate to GRCh37/GRCh38. Existing linear-reference tools will continue to work with the expanded pangenome reference coordinate system, and pangenome-based results will be translatable to these existing coordinate systems with improved genotype accuracy.

We will develop liftOver tools that make it easy to go backward from the pangenome reference to GRCh37/GRCh38 when necessary. We already have algorithms for this purpose and demonstration of functionality to predict read mappings from a prototype pangenome to GRCh37/GRCh38. We will precompute all mappings between the previous assemblies and the pangenome and provide these coordinate translation functions with the pangenome reference release. This information should ease the transition of other databases and resources that rely on these coordinates and provide an annotation directly onto the GRCh37/GRCh38 assemblies in areas where mappings and interpretation on the pangenome are more reliable than current, linear sequence representations.

We will augment the human genome browser’s displays to transition to any haplotype assembly in the pangenome reference and display the haplotype alignments. Visualizations will include relevant genetic backgrounds for specific tracks, for example picking the right HLA haplotype for a read mapping track. To ensure we use these tools effectively, we will add detailed information that explains these novel views to our existing training materials and make this information part of our respective workshops.

We have adopted the GA4GH principles and will develop exchange formats analogous to SAM/BAM and utility libraries analogous to htslib/samtools, facilitating the development of transition tools and workflows for the pangenome reference. We will deposit these tools and their guides in the HPRC resource repository. We have also developed a prototype transcript archive that facilitate annotation discovery in GRCh37, GRCh38, CHM13 and the pangenome, and visualize the differences between two transcripts (for example, on two different genomes).

We aim to engage pilot users to obtain feedback about these resources. The HPRC program and related tool developers connected with the community of users will develop new tools that gain additional value from using the pangenome reference rather than linear-reference genome assemblies. We will report on our discoveries in publications and talks, through the blog, webinars, and on the HPRC website and provide educational tools and forums on using and switching to a pangenome reference.

Relevance to Disease Research

We expect that the resources and methods we are developing will profoundly impact studies of the genetic basis of human disease and precision medicine. While we recognize that adoption by the clinical research community will take time, there are three important benefits to using a pangenome reference. First, a more complete reference that incorporates and displays human genetic diversity will produce fewer ambiguous mappings and more accurate copy-number variation analyses throughout the genome when patient samples are sequenced and analyzed^59,108. This will improve genetic diagnosis and the functional annotation of variants. Second, the resource will enable the discovery of disease risk alleles and previously unobserved rare variants, especially in regions that are inaccessible to standard, short-read sequencing technologies. Studies of unsolved Mendelian genetic disease, for example, have shown that ~25% of “missing” disease variants can be recovered when longer reads are applied and more complex repetitive regions are characterized¹⁰⁹. Important genetic risk loci such as SMN1/2 (spinal muscular atrophy), LPA (lipoprotein A and coronary heart disease), CYP2D6 (pharmacogenomics), as well as numerous triplet repeat expansion loci are now being sequenced and assembled in large human cohort studies. These studies are revealing the standing pattern of natural genetic variation for loci typically excluded from previous analyses^58,59. Resolution of these loci by long-read sequencing in even a limited number of human haplotypes improves our ability to genotype them in other patient-derived short read datasets, allowing for the discovery of new genetic associations, through both GWAS and eQTL methods⁵⁸. Finally, the pangenome approach represents a fundamental change in how human genetic variation is discovered. Instead of simply mapping sequence reads to a reference, we are constructing phased genome assemblies and aligning them to the graph, which in turn will pinpoint all genetic differences both large and small at the base-pair level^26,110. As long-read sequencing costs fall and pangenome methods evolve²⁶, we predict patient samples will likely be sequenced using long-read technology to increase sensitivity and accuracy.

Outlook

As we write this Perspective, the world is reeling from the COVID-19 pandemic and the spread of new SARS-Cov-2 variants. Scientists can trace the virus’s epidemiology, determine why humans are susceptible^111,112, and determine why some individuals are more susceptible than others^113,114. The current GRCh38 human reference is one of many resources that made this possible, but we know that it can be improved. Through years of strategic investments in the public and private sectors, we find ourselves with the technologies and methods to build additional references that better represent global human genomic diversity.

The human pangenome reference will collect accurate, haplotype-phased genome assemblies generated by efficient algorithmic innovations, which we anticipate will be widely used by the scientific community. The collection of individual genomes, comprised of sequence information, genomic coordinates, and annotations, will be a critical resource with more accurate representation of human genomic diversity. The original Human Genome Project enabled major advances in human health and genomic medicine^1–4; it is time to build a more inclusive resource with better representation of human genomic diversity to better serve humanity.