ABSTRACT
Genome graphs have gained prominence and are becoming increasingly pertinent in the genomic research landscape. Despite their innate advantages, there is a shortage of techniques to comprehensively analyse the structural properties of genome graphs and systematically unearth the underlying genomic complexity of the population or species they represent. In this study, we formulated a novel framework to represent and capture the intricate structural complexities inherent in genome graphs. This approach opens up the opportunity to visualise the entire human genome at once and enables the prioritisation of sites of interest that are valuable for in-depth research. We applied the formulated technique to visualise and compare the structural properties of two human pan-genome graphs: one that augments only the variants commonly present in different human populations and the other that augments all the variants, including the rare ones. We also developed and benchmarked various genome-graph-based variant calling workflows and analysed human whole genomes with them. We compared the variant-calling performance of the two constructed graphs with each other and with the linear reference genome. We identified that genome graphs are better reference structures than their linear counterparts, and the proposed structural analysis framework can effectively analyse, visualise and compare the complexities embedded in them.
INTRODUCTION
The reference genome is a cardinal element in genome analysis. Despite its broad impact, the first draft of the reference genome (1), unveiling the 3 billion base pairs of DNA, missed much of the variation unique to heterogeneous human populations (2). Since the release of the first draft, the reference genome has been improved regularly to redress its inadequacies. The current widely used human reference genome GRCh38, released in 2013 (3) and recently updated in 2019 (4), is a haploid DNA sequence majorly derived from merged haplotypes from individuals of Caucasian and African ancestries, with a single individual comprising most of the sequence (5, 6). Various studies (7–10) elucidate the genetic diversity and heterogeneity of human populations. However, recent advances in genomics have shown that owing to its haploid linear structure, the current reference cannot effectively accommodate the known variants in different subpopulations. Thus, it cannot sufficiently capture the genetic diversity in human populations (11). As only one allele of a variant can be present in a haploid genome, using such a reference structure leads to reference allele bias (12). These drawbacks of a linear reference genome highlight the need for a better reference structure to efficiently represent the sequence information in the DNA of heterogeneous and diverse human populations.
Human genome graphs have the potential to represent highly heterogeneous populations while overcoming the drawbacks that usually accompany using traditional linear structure-based references (13). Genome graphs encode the genetic variants within a population and embed within them paths that can represent possible sequences from a population for which the reference is built (14, 15). They methodically segment sequences into discrete entities termed nodes. The interconnection of these nodes through edges establishes a network, facilitating seamless traversal within the graph from one node to another. Unlike a traditional linear reference genome that can be biased towards a specific individual or a population, genome graphs can incorporate information from multiple individuals from varied populations. This makes them more versatile for studying diverse genomic landscapes as they provide a flexible and inclusive framework for representing known genetic variation, structural variations, and population-specific differences in human populations (16). Previous studies (17) broadly classify genome graphs as population-specific, capturing the genetic variations like single nucleotide polymorphisms (SNPs), insertions, deletions (INDELs), and structural variants (SVs) that are prevalent within a particular population and pan-genomes representing the genetic diversity across multiple populations.
Genome graphs can compactly represent sequence information, including the genetic variants across populations. The structure of genome graphs is dynamic as the nodes and edges of the network depend on the set of variants augmented during its construction. So, by altering the variant set used for construction, genome graphs can be easily modified and updated. Such an effortless correction of the reference structure is impossible with linear reference genomes. As genome graphs can represent alternate alleles at specific genomic loci, they overcome reference allele bias (12). That is, when a genome graph incorporates variants known to be present in a population, it enables alignment of reads of newly sequenced genomes from that particular population with fewer mismatches, enabling superior variant calling and more accurate downstream analysis (18). Genome graphs also make it feasible to focus on hypervariable regions of the human genome at a population and personalised level.
Despite the inherent advantage of genome graphs to focalise hypervariable regions in the genome, there is no method yet to comprehensively study the structural complexities underlying them (19). There is a need to quantify the complexity of genome graphs that would strategically shed light on the regions of the genome with salient structural features. Such a method would enable researchers to locate previously unexplored parts of the genome with potential functional significance. Population-level studies could benefit from such a method as genome graphs encompass the heterogeneity of the subpopulation they represent. By comparing the structural complexities of different population-specific genome graphs, it is possible to highlight the similarities and differences in the polymorphism patterns of different populations.
Past studies that establish methods to work with genome graphs and demonstrate the advantages of genome graphs over the traditional linear reference genome have constructed their reference genome graphs in myriad ways. Each study has augmented a different set of variants onto the linear reference to construct their graph reference without a clear rationale behind the variant set selection (20). Some studies have augmented the complete set of variants from a prominent species-specific database (21–23). Other studies have filtered out rare variants from the variant set before adding them to the genome graph reference (24–26). The effects of adding rare variants onto the genome graphs is not fully understood yet, and a systematic analysis is required to resolve this issue at the scale of human genomes.
In our study, we formulated novel methods to quantify and capture the structural complexity of genome graphs. We applied these techniques to visualise and compare the structural properties of two human pan-genome graphs. The first augments only 8,496,706 common variants from the diverse cohort of 2504 samples from the 1000 Genomes Project with an allele frequency of at least 5%. The second one augments all the 85,123,169 variants, including the rare ones. We recorded the variant-calling performances of these two constructed genome graphs with each other. To this end, we developed and benchmarked various genome-graph-based variant calling workflows and identified the optimal computational pipeline. Using the selected pipeline, we compared the variant-calling performance of the two constructed human genome graphs with each other and with the linear reference genome applied with the established linear-genome-based variant-calling workflow.
METHODS
Construction of human pan-genome graphs
Recent developments using genome graphs as a reference structure for genome analysis have spawned numerous tools for their construction and allow researchers to analyse newly sequenced genomes with genome graphs (27). Even though most of the existing tools operate well for smaller genomes, they rarely perform optimally when the size of the genome scales up. The software that runs smoothly at the scale of human genomes is few and far between. vg toolkit (14) is one of the widely used, openly available tools that can efficiently handle the construction of human genome graphs and analyse whole genome sequences with the same. It is an actively maintained software with an extensive set of functionalities. We have used the vg toolkit in our study to construct human genome graphs and parts of downstream analyses.
Genome graph construction with the vg toolkit involves incorporating variants over a linear reference structure to generate the edges and nodes that comprise the graph-based reference structure. We have used the GRCH38 (hg38) reference genome as the linear reference structure. The 1000 Genomes Project (7), built on the Human Genome Project, sequenced hundreds of individuals from diverse populations and revealed millions of genetic variations, paving the way for a deeper understanding of human diversity. The variants from this project (1KGP variants) elucidated the genetic heterogeneity of the species well and can be used to construct the human pan-genome graph. As the 1KGP variant set called with hg38 reference genome was incomplete for chromosome Y, we lifted over the Y chromosome variants called from hg37 to hg38 and have used it for the construction and analyses of genome graphs.
Human pan-genome graphs can be constructed by augmenting the 1KGP variants onto the hg38 reference. However, most of the variants from the 1KGP project were rare (Table 1), and the added advantage of augmenting such variants onto the genome graph was uncertain. To overcome this irresolution, we constructed and compared two types of genome graphs: one with the entire 1KGP variants augmented onto hg38, called the 1KGP Complete Genome Graph, and the other excluding rare variants with alternate allele frequency less than 5%, called the 1KGP Common Genome Graph (Figure 1).
1KGP Complete Genome Graph augments all the variants from the 1000 genomes project. 1KGP Common Genome Graph augments only the variants from the 1000 genomes project, with an alternate allele frequency of 5%.
Structural analysis of genome graphs
Genome graphs encased within them a reference path that corresponds to the linear reference genome on which the variants were augmented. The human pan-genome graphs constructed in our study encompassed a path that retraced the hg38 reference genome. Nodes in the reference path could be backtracked on the linear coordinate system, and this property of the genome graph was a bridge between the two reference structures. Every variant augmented onto the genome graph created a path that diverged from the reference path at a particular reference node. Each variant path that diverged from a reference node would increase the out-degree of that corresponding reference node by one. This implied that the out-degree of the reference node directly correlated with the complexity of the human genome at that position. This phenomenon was used to credibly quantify the complexity of the entire human genome through the lens of genome graphs. To systematise the structural analysis of genome graphs, we propose the definitions in Box 1.
Definitions of terms used in structural analysis of genome graphs.
Reference path: A path in the genome graph corresponding to the linear reference genome used in its construction.
Reference Node: A node present in the reference path.
Variant path: Paths in the genome graph that arise due to the variants augmented during its construction.
Variant node: A node present in a variant path.
Out-Degree: Number of outgoing edges that emanate from a node.
Variable node: A node with a minimum out-degree of 2.
Variability: The presence of variable nodes in a region.
Hypervariable node: A node with a minimum out-degree of 5. An out-degree of 5 is chosen to distinguish the variable nodes that are more complex than an SNP.
Hypervariability: The presence of hypervariable nodes in a region.
Invariable region: Parts of reference path that are of a size more than 1000 nodes but devoid of any variability.
The genome graphs created in the vg toolkit were extracted in the Graphical Fragment Assembly format (GFA) and were incorporated into a custom Python program built on top of NetworkX (28). NetworkX enabled the effortless application of graph algorithms at the scale of human genomes. We subdivided the reference path in the constructed human pan-genome graphs into bins of 10 Mbp. Out-degrees of all the nodes were calculated for the binned subgraphs from both the genome graphs. Variable and hypervariable nodes from each bin were enumerated. A complete panoramic view of the human pan-genome graphs was obtained by collating the count of such nodes throughout the human genome. Circos plots (29) were used to get a bird’s-eye view of the genome graphs. These views were then used to compare them and identify the key structural similarities and differences in the reference genome graphs. The functional significance of crucial findings from the structural analysis of human pan-genome graphs was studied in-depth.
Variant calling with genome graphs
Genome graphs have been established as a superior method to capture novel variants (17). A major drawback of using genome graphs as a reference structure in variant calling was that they were not easily scalable to the human genome size, and they required compute resources and runtimes much higher than a linear reference genome. Even though the vg toolkit scaled seamlessly to the human genome size, the default variant caller included in it was compute-intensive and time-consuming. As the mapping algorithm gave output in GAM format, employing existing runtime-efficient variant callers that could work only with the well-established BAM format was not straightforward. However, previous works (30) have established a workaround by converting the GAM file format to the BAM file format by projecting the reads mapped to the variant paths of the genome graph to their originating reference nodes. This opened an avenue to use existing established and runtime-efficient variant calling algorithms with the vg toolkit.
In our study, we designed multiple genome-graph-based variants calling workflows by incorporating existing linear genome-based variant callers, namely GATK Haplotypecaller (31), Bcftools mpileup (32), and FreeBayes (33). We compared them with the default variant caller vg call that is bundled with the vg toolkit. We mapped the raw sequence reads to the 1KGP Common Genome Graph using the runtime-efficient vg giraffe algorithm from the vg toolkit suite. We converted the output GAM files to BAM format using vg surject. The obtained BAM files were processed according to the requirements of the variant caller used. We then compared the performances of these developed workflows by benchmarking them with the acknowledged Genome In a Bottle (34) dataset (GIAB). HG002 was used for the comparison study, and the variants called in each workflow were compared with the high-confidence variant set of HG002 provided by the GIAB consortium. All the developed workflows were run on a machine with 64 cores and 512GB memory to maintain uniformity. The performances of all the developed genome-graph-based variant calling workflows are presented in Table 4. The better pipeline among them was selected based on the runtime of the workflow and the F1 score, which is the harmonic mean of precision and recall from the GIAB benchmarking.
The parallel implementation of the Freebayes-based variant calling workflow got a balance of runtime efficiency and good benchmarking scores. Figure 2 describes the optimal genome-graph-based variant-calling pipeline step-by-step. We wrapped this computational pipeline in the Snakemake workflow management system for improved reproducibility and better ease of use in the future (35). We estimated the benefits of adding rare variants to the human genome graph with the finalised genome-graph-based variant calling pipeline. We processed GIAB samples with the 1KGP Common and Complete Genome Graphs. The variant calling capabilities of these human pan-genome graphs were also compared with that of the hg38 reference. To this end, a computational pipeline was constructed to call variants with the linear reference genome. BWA-MEM was used as the aligner, and the GATK suite was used to call and process variants with hg38. The variants called in the seven GIAB samples by these three workflows were compared correspondingly with their high-confidence variant calls provided by the GIAB consortium. hap.py (36) was used to benchmark all three pipelines for both SNPs and INDELs.
A step-by-step depiction of the complete genome-graph-based variant calling pipeline that balances performance with runtime. This pipeline was used to compare the advantages of variant-calling performance when adding rare variants onto the genome graphs.
RESULTS
The prevalence of rare variants in the 1000 Genomes project with an alternate allele frequency of less than 5% was studied for different sample sizes and is summarised in Table 1. The individuals in each sample size were picked in a stratified random order for population and gender. As the sample size increased, the total number of variants increased steadily, with rare variants constituting most of them. With a growing sample size, more variants specific to one or few individuals were called, causing the count of common variants to be stable.
We constructed and compared two genome graphs to understand the benefits of adding the rare variants that were in the majority to the human pan-genome graph. All the 85,123,169 variants identified across 2504 individuals were augmented onto hg38 to create the 1KGP Complete Genome Graph. After applying a minimum 5% alternate allele frequency filter on these variants, 8,496,706 common variants from these diverse individuals were extracted and augmented onto hg38 to create the 1KGP Common Genome Graph.
Structural analysis of genome graphs reveals fundamental biological insights
The reference path in the genome graphs was subdivided into bins of window size 10MB. The prevalence of variable and hypervariable regions in each bin in each chromosome was used to get panoramic visualisations of the entire genome graphs. Figure 3 portrays the bird’s-eye view of the human pan-genome graphs using Circos plots.
Each sector in the Circos represents a chromosome. Each tangential bar spans a 10MB genomic window. The number of hypervariable nodes in each bin is represented in the outer track. The number of variable nodes in each bin is represented in the inner tracks. The maximum value for the inner and the outer tracks are present at the plots’ central points.
Interesting structural properties from the human pan-genome graphs can be inferred from Figure 3. The inner track, representing the prevalence of variable nodes, was capped at 524K for the 1KGP Complete Genome Graph, whereas it was 61K for the 1KGP Common Genome Graph. This result was anticipated as the rare variants were removed in constructing the 1KGP Common Genome Graph. Hence, the variability of the common graph would be lesser when compared to the 1KGP Complete Genome Graph. However, the contours of the bars capturing the distribution of variable nodes in the inner track were very similar for both the genome graphs, implying that the polymorphism in the human genome that arose due to common variants alone retained certain genome topography from the polymorphism that arose from the entire set of genetic variants.
However, the contours of the bars in the outer track capturing the occurrence of hypervariable nodes were starkly different for both the genome graphs. Hypervariable nodes were present throughout the 1KGP Complete Genome Graph, the majority of them were located in chromosome X. Whereas, the 1KGP common Genome Graph was depleted of hypervariable nodes all over the genome, even in chromosome X. Rare variants that were removed from the 1KGP Common Genome Graph, were mostly driving the hypervariability in the genome. This could mean that the hypervariability observed in the human genome, particularly in chromosome X, originated from different populations independently, as the common genetic variants present across populations did not engender the structural complexity to the same level.
The location of the bins with a high frequency of variable nodes for both the genome graphs is recorded in Table 2. A 10MB bin in chromosome 6 spanning from 30-40 Mbp showed the highest variability only in the 1KGP Common genome graph. This bin encompassed a majority of the Major Histocompatibility Complex known as the human leukocyte antigen (HLA) of the human genome. Previous studies have established that the HLA contains one of the most polymorphic gene clusters of the entire human genome. The presence of multiple HLA alleles in the population ensures that at least some individuals within a population will be able to recognise protein antigens produced by any microbe, thus reducing the likelihood that a single pathogen can evade host defences in all individuals in a given species (37). Table 3 shows that the bin containing HLA had a comparatively lower frequency of variable nodes in the 1KGP Complete genome graph. It might imply that a notable portion of the variants driving polymorphism in HLA were not rare and were shared by individuals in different human subpopulations.
In both the human pan-genome graphs, other genomic regions with high variability were observed in chromosomes 8 and 16. The first 10MB bin from chr8 with the highest variability in 1KGP Complete graph encompassed parts of the β-defensin gene clusters called DEFB. DEFB was previously known to be highly variable and was associated with evolutionary importance and disease risk (38). Chromosome 16 enclosed a couple of highly variable bins in both human pan-genome graphs as it hosted several large polymorphisms, often associated with segmental duplications (39). As these bins were highly variable in both the genome graphs, it implied that the polymorphism in these regions originates from common and rare variants.
Structural analysis enables the quantification of the complexity of genome graphs
Genome graphs capture the complexity and diversity of genomic structures in a species, and nodes with high out-degree can be crucial for understanding specific aspects of the genome. The maximum possible out-degree of a node representing an SNP in the genome graphs was four (one reference and three possible alternate alleles). Intending to capture the highly complex parts of the genome, we defined hypervariable nodes as more complex nodes than an SNP node. Hypervariable nodes, representing genomic regions with high levels of polymorphism, can be essential for studying population-specific genetic diversity and capturing regions of the genome that substantially impact various biological processes and functional relationships. These hypervariable nodes can potentially be used as markers that can distinguish different genome graphs built for the same species, as other graph-derived metrics like the number of nodes and edges cannot achieve this.
The degree-based distribution of hypervariable nodes in the constructed human pangenomes is depicted in Figure 4a. As expected, the 1KGP Common Genome Graph had restricted distribution of hypervariability not only for genomic position but also with degree. Nearly 150 hypervariable nodes were observed in the 1KGP Common Genome Graph, and all had an out-degree of either five or six. On the other hand, the hypervariability is well distributed in terms of both degree and genomic position in the 1KGP Complete Genome Graph. Nearly 7500 hypervariable nodes were observed, and two nodes were seen with the maximum out-degree spanning twelve. When the exact location of the most hypervariable nodes was pinpointed, it was found that both these nodes were adjacent to each other and were present in chromosome 1. Figure 4b depicts the path-level representation of the adjoining 12-degree nodes in the 1KGP Complete genome graph obtained using the default visualisation command bundled with the vg toolkit.
(a) Degree distribution of the hypervariable nodes in the 1KGP Complete Genome Graph (Left) and the 1KGP Common Genome Graph (Right). (b) Path-level representation of the most complex nodes in the 1KGP Complete Genome Graph. Markers are placed on the edge that trace the reference path. The 12-degree nodes are present in chromosome 1, and the node IDs are 19298453 and 19298461.
The human genome is comprised of large zones of invariance
Invariant regions are essential components of the genome, reflecting the conservation of genomic segments across different individuals, populations, or species. While invariant zones across individuals suggest a critical role in maintaining essential biological functions, invariance across the genomes of species helps infer evolutionary relationships and divergence patterns among different organisms. Identifying and studying invariant zones in the genome can contribute to various aspects of genomics, including evolutionary biology, functional genomics, and biomedical research.
Genome graphs, apart from emphasising hypervariable nodes, also offer a convenient way to pin down the invariant zones in the genome. In our study, we defined invariant zones as the regions in the genome graph composed of at least 1000 nodes devoid of any variant paths. This essentially translated to a stretch of the genome at least 32 Kbp in size with no variants found in the samples included in the study being defined as an invariant zone. As the 1000 genome project was a very diverse cohort, invariable regions identified in the human pan-genome graphs were conserved for individuals across diverse populations.
The 1KGP Common Genome Graph was created by depleting the rare variants from the variable regions of the 1KGP Complete Genome Graph. Removal of the rare variants reduced the number of variable nodes from the genome graph and, depending on the removal area, could have three possible effects: i) when variable nodes in a region were depleted enough, new invariant zones were created; ii) when variable nodes present at the edges of existing invariant zones were cleared, the length of the corresponding invariant zone increased; iii) when variable nodes were not reduced enough in a region, no change was observed for the invariant zones.
The distribution of invariant regions in both the human pan-genome graphs is presented in Figure 5. It can be inferred from Figure 5a, b that all the three effects mentioned above have taken place when rare variants were removed from the genome graph.
(a) The total number of invariant zones present in each chromosome for the two human pangenomes constructed. Chromosomes 3, 4, 6, 11, 12, 13, 14, 17, 18, 20, 21, 22 have the same number of invariant regions for both the genome graphs. (b) The median length of the invariant zones in each chromosome. The length of the zones in BP is calculated as (number of nodes in the zone) times 32. (c) The volume of chromosomes invariant in the human pan-genome graphs. Invariability is calculated as (Total length of invariable zones in a chromosome in BP)/(Total length of the chromosome in BP).
The number of invariant zones was unchanged between the two genome graphs for twelve chromosomes, but the aggregate invariability differed for all the chromosomes. This implied that only the lengths of existing invariant regions have increased in these chromosomes due to the rare variant depletion. In these chromosomes, all the variable node removal at the edges of existing invariant zones increased their lengths, and removing other variable nodes was not enough to create new invariant zones. Meanwhile, the reduction of the median length of invariant zones in the 1KGP Common Genome Graph, for example, in chromosomes 2, 7, 16 and 19, implied the creation of newer invariant zones that were of smaller lengths than the ones observed in the 1KGP Complete Genome Graph. Sizable invariant zones with more than 1 MBbp lengths were observed at various regions of both genome graphs (Figure S1).
Existing genome-graph-based variant calling workflows can be optimised
To overcome the extensive computational requirements of the genome-graph-based variant calling pipelines, we developed multiple workflows that employed linear-genome-based variant callers such as Bcftools mpileup, GATK Haplotypecaller and Freebayes-parallel, in conjunction with genome-graph-mapped files. To this end, we mapped the GIAB sample HG002 to the 1KGP Common Genome Graph using the vg giraffe algorithm and converted the obtained GAM file to the BAM format. We also called variants using the variant caller vg call bundled with the vg toolkit and compared the results with the other developed workflows. The variants generated from each workflow were compared with the high-confidence variant calls published by the GIAB consortium. hap.py was used to benchmark these developed genome-graph-based workflows.
Table 4 summarises the metrics quantifying the performance of developed variant-calling workflows. The runtimes of each post-graph-alignment step for the developed workflows are presented in detail in Table S1a-d. It can be seen that using Bcftools mpileup and Freebayes significantly reduced the computing time of the genome-graph-based variant calling. The sharp increase in the runtime of the GATK HaplotypeCaller pipeline was due to the extensive preprocessing steps. As reflected in the F1 scores, particularly in the case of INDELs, GATK HaplotypeCaller was the best-performing workflow. However, the parallel implementation of Freebayes scores was very similar to F1 scores and was the fastest pipeline. Considering the balance of performance and runtime, we used the Freebayes-parallel workflow for our downstream analyses.
Genome graphs are better reference structures than linear genomes for variant identification
GIAB samples were processed with both the 1KGP Common Genome Graph and the 1KGP Complete Genome Graph to better understand the variant calling benefits of adding rare variants to the human pan-genome. The Freebayes-based variant calling workflow that benchmarked well with HG002 was used to process all the GIAB samples with both the constructed genome graphs. To compare different reference structures, we also processed the samples with the linear hg38 with the established BWA-GATK workflow. The metrics extracted from the variants called in all three workflows are presented in Figure 6.
(a) The total number of variants captured in all the seven GIAB samples by the three reference structures. BWA-GATK was used to process samples with GRCH38, and the finalised vg-giraffe-Freebayes-parallel was used to process samples with the genome graphs. (b) Distribution for the ratio of the number of transitions to the number of transversions for the variants called in GIAB samples by the different reference structures.
The genome graph referenced workflows captured, on average, 393K more variants than the linear reference workflow. Examining the variants closer, we observed that in the case of SNPs, the genome graph workflows always outperformed the linear workflow, whereas in the case of INDELs, in two samples, the linear pipeline called more of that variant type (Figure S2). The 1KGP Complete Genome Graph (p-value = 0.0012, one-tailed paired t-test) and the 1KGP Common Genome Graph (p-value = 0.0012, one-tailed paired t-test) reference structures captured more variants than the linear hg38. Regarding the quality control metric, all three reference structures achieved a Ts/Tv ratio of nearly 2.0, an expected threshold for the human whole genome variants (40).
However, adding rare variants to the genome graph did not give any significant improvements in terms of the number of variants called. The 1KGP Complete Genome Graph that augmented the rare variants from diverse individuals could not capture more variants (p-value = 0.08) than the 1KGP Common Genome Graph. The average increase in the variant count obtained by the more complex genome graph was only 111. In HG005 and HG006, the less complex 1KGP Common Genome Graph was observed to capture slightly more variants than the 1KGP Complete Genome Graph, which augmented all variants.
To verify if the new variants captured by the genome graphs were not false positives, we benchmarked the variants called by all the reference structures with the high-confidence variants set. Except for HG002, the F1 scores for the SNPs and INDELs were higher for the genome-graph-based workflows than for the linear genome-referenced workflow (Figure 7, Table S2). We observed that the 1KGP Common Genome Graph (p-value = 0.006, one-tailed paired t-test) and the 1KGP Complete Genome Graph (p-value = 0.006, one-tailed paired t-test) had higher F1 scores for SNPs (Figure 7, Table S2). Meanwhile, in INDELs, the difference between the graph and the linear referenced pipelines was not significant (p-value = 0.02 for both, Figure 7, Table S2).
(a) Distribution of F1 scores of SNPs called by the three reference structures for the seven GIAB samples. (b) Distribution of F1 scores of INDELs. The F1 scores are calculated as the harmonic mean of the precision and recall.
Even though it was unclear from the visualisation, the F1 score for SNPs was observed to be higher for the 1KGP Common Genome Graph by an average of 1.2 x e-05 than the 1KGP Complete Genome Graph (p-value = 0.001). The difference between the genome graphs was not significant (p-value = 0.14) for the F1 scores of INDELs. Considering the number of variants and the benchmarking results, the linear reference genome-based workflow underperformed the genome-graph-based workflows. As the performance of the 1KGP Common Genome Graph was at par with that of the 1KGP Complete Genome Graph, no benefits in terms of variant calling were observed while adding rare variants to the human genome graph.
DISCUSSION
Genome graphs are gaining prominence and becoming increasingly important in genomic research. They aid in enhancing our understanding of genetic diversity and its implications for health, evolution, and biodiversity. Their ability to better capture the genomic complexity than the traditional linear reference genome makes them immensely useful in population-level studies. However, there is a lack of methods to thoroughly analyse the structural properties of genome graphs and systematically uncover the underlying genomic complexity of the populations or species they represent. Understanding the structural implications of genome graphs holds significance in fundamental biological and evolutionary contexts and in shaping strategies for preventing and treating genetic diseases. Disease susceptibilities are known to be divergent for individuals from different populations, and polymorphic regions of the genome may influence drug metabolism and response. Understanding the variability in drug-metabolising enzymes and drug targets is crucial for personalised medicine. Therefore, identifying the polymorphic sites of the genome distinct to certain populations has significant implications for personalised medicine.
Previous works have used only simple metrics, such as the number of nodes, edges and connected components, as indicators of the complexity of the graph (19). Existing visualisation techniques are limited to small genome graphs and have been infeasible for graphs with hundreds of thousands of nodes (41–45). In this work, we designed a novel framework to represent and capture the intricate structural complexities inherent in genome graphs. The proposed structural analysis method can be effectively scaled to the whole human genome size and opens up the opportunity to visualise the entire human genome graph at once and get a panoramic view of the complexities of the human genome. Such a view of the human genome graph can be a foundation for visualising, navigating, unearthing, and prioritising sites of interest from the human genome, which is valuable for in-depth research.
We segregated the complex regions of the genome graphs into three broad categories: variable, hyper-variable and invariable. While the in-depth study of variable and hypervariable regions is essential for understanding the genetic makeup of populations and exploring the functional implications of genetic diversity, invariant zones of the genome help identify conserved regions with high evolutionary significance. The structural analysis also yielded a novel visualisation technique that encapsulates the entire genome graph in a single figure and presents a panoramic view. We applied these techniques to analyse, visualise and compare the structural properties of two human pan-genome graphs constructed in this study. Variable, hypervariable, and invariable regions were identified from both the human genome graphs and used to compare and contrast them. Through the proposed structural analysis framework, we can efficiently capture the complexity of genomes while enabling the effective analysis, visualisation, interpretation and comparison of genome graphs.
The variant calling performance of both the human genome graphs was also examined in depth. Multiple genome-graph-based variant calling workflows were built and benchmarked to identify the optimal computational pipeline. Using the finalised workflow, we attempted to understand the added benefits in the variant calling performance when augmenting the rare variants onto the human pan-genome graph. We observed that genome-graph-based workflows captured nearly 393K variants more than the linear-genome-based workflow while achieving higher F1 scores in GIAB benchmarking (p-value = 0.006). This implies that genome graphs are better reference structures than the linear reference genome. However, no immediate advantages were observed regarding variant calling when incorporating rare variants into the human genome graph, as the performance of the 1KGP Common Genome Graph was comparable to that of the 1KGP Complete Genome Graph. Only 111 more variants were captured by the latter, and the former had better performance when it came to F1 scores of SNPs. The underperformance of high complexity genome graphs when compared to graphs with simpler structures, aligns with previous studies (20) that evaluated the read-mapping performance of genome graphs with varying complexities. Through our study, we extended this comparison to the variant calling performance of genome graphs and validated it on sequenced data instead of synthetic data.
When analysing the variants captured in the 1000 genomes project for different stratified sample sets, we identified that with increasing sample size, the percentage of rare variants increased while the number of common variants remained almost constant. So, the construction of an extensive human genome graph with only the common variants from diverse individuals can be achieved with a smaller sample size. In contrast, constructing a well-representative human genome graph that augments the rare variants would require a large sample size. Genome graphs can be useful for analysing WGS from divergent populations from those that make up hg38 as it overcomes reference allele bias. As population-level studies are done in phases (17), and fewer samples are sequenced at the beginning of the study, analysis of the new genomes can benefit from the construction of a population-specific genome graph that augments only common variants that can stay stable with increasing sample size when compared to all the variants that expand constantly. Nevertheless, a detailed variant count analysis specific to the population is required before such a population-specific genome-graph-based analysis begins.
The framework for the structural analysis of genome graphs elucidated in this paper is coded for and tested on the genome graphs constructed with the vg toolkit. However, the proposed idea can be extended to genome graphs built with other software, as the structural analysis framework is pivoted around the traversal of the reference path, and this only requirement is ubiquitous to most genome graphs. In this study, we have used hg38 instead of the latest T2T (46) as the reference genome on which the variant paths were added to create the genome graphs. It can be noted that the formulated structural analysis framework would function well, irrespective of the reference genome. Moreover, this proposed idea is not just limited to human genome graphs but can extend smoothly to genome graphs representing the genetic complexity in other species.
A further development that annotates the genome graphs with meta-information can complement the proposed structural analysis well and open the avenue for applying more sophisticated graph algorithms on genome graphs. When such methods are applied to population-specific human genome graphs, researchers can comprehensively assess genetic diversity, population structure, disease associations, and evolutionary dynamics within and across populations. This information can enhance our understanding of the diversity in human genetics and how it impacts health, diseases, and the historical dynamics of populations.
SOFTWARE AND CODE AVAILABILITY
The codes are available at https://github.com/IBSE-IITM/HumanGenomeGraphs. The BWA-GATK workflow is containerised as a docker pipeline and is available at https://hub.docker.com/r/ibse/genome-india.
FUNDING INFORMATION
The work was supported by the Department of Biotechnology, Govt. of India (BT/GenomeIndia/2018) to MN, KR, HS and the Centre for Integrative Biology and Systems Medicine, IIT Madras (BIO/18-19/304/ALUM/KARH) to KR, HS. MN was supported by the Wellcome Trust/DBT grant (IA/I/17/2/503323).
COMPETING INTERESTS
The authors have declared no competing interest.
ACKNOWLEDGEMENTS
We thank Philge Philip, Sai Sruthi Amirtha Ganesh, Keerthika Moorthy, Harshita Agarwal, and members of the Centre for Integrative Biology and Systems Medicine (IBSE) for their valuable discussions and comments.
REFERENCES
