Aggregate variant calling using short reads enables population and disease studies for paralogous genes

Summary

Variant calling in paralogous genes using short-read sequencing is problematic due to mapping ambiguity between highly similar sequences. Aggregate variant calling, which treats paralogous loci as a single locus by realigning reads to a masked reference genome, can enable variant detection in paralogous genes. We used our informatics tool Parascopy to assess the accuracy of aggregate variant calling in paralogous genes using short-read data. Parascopy achieved significantly higher recall compared to standard variant calling without sacrificing precision. We identified 158 paralogous genes with over 25 percentage points improvement in recall using simulated data and 118 genes with at least 10 percentage points improvement in recall across Genome in a Bottle (GIAB) reference samples. Across 1000 Genomes samples, aggregate genotypes in paralogous genes were highly concordant between whole-genome and whole-exome data (r² = 0.9998). Using sequence data from population and disease cohorts, we utilized aggregate variant calls to perform population-genetic analysis, case-control association analysis, and haplotype phasing in specific disease-relevant paralogous genes that are inaccessible to standard diploid variant calling. Specifically, in the SMN1 gene, we identified tag-SNPs for two-copy SMN1 haplotypes and showed that specific low-frequency missense variants in African populations occurred exclusively on such haplotypes. For the CFC1 gene—previously implicated in heterotaxy syndrome—association analysis using exome data indicated that loss-of-function variants are unlikely to be associated with heterotaxy in individuals with congenital heart disease. Finally, we utilized aggregate genotypes to perform haplotype phasing for the X-linked gene IKBKG, achieving 99.2% concordance between trio-based and statistical phasing approaches.

Paralogous genes with high sequence similarity are difficult to analyze using standard short-read sequencing. Prodanov et al. show that aggregate variant calling, which pools reads across paralogous copies, recovers variants missed by conventional callers and enables population-genetic analysis and disease association studies in clinically relevant human genes.

Introduction

Advances in DNA sequencing technologies have transformed the ability to sequence human genomes and identify disease-causing variants. The human genome is highly repetitive, and approximately 7% of the genome consists of long segments of repetitive DNA sequences—known as segmental duplications or low-copy repeats (LCRs).^1,2 Hundreds of genes are known to overlap these segmental duplications, and many are associated with rare and complex human diseases.³ Sequence variants can be detected with high accuracy in unique regions of the genome, where short-read mappings are typically unambiguous. However, variant calling in low-mappability regions is challenging. Mandelker et al. identified 193 medically relevant genes that are challenging for variant calling using short-read sequencing due to high sequence homology with other genomic regions. Similarly, Ebbert et al.⁴ performed an extensive analysis of mappability across the human genome and identified 748 “dark” genes in the human genome that are problematic for short-read sequencing. The genomic analysis of genes overlapping segmental duplication is further complicated by the presence of frequent copy-number variation⁵ and gene conversion between proximal repeat copies.^2,6 Variants resulting from gene conversion are difficult to identify using standard variant calling, as paralogous sequence variants (PSVs) can no longer be used to distinguish the repeat copies.

In paralogous genes such as SMN1 (MIM: 600354)/SMN2 (MIM: 601627), a significant proportion of short reads cannot be mapped uniquely. As a result, standard variant calling tools that rely on mapping quality exhibit very low recall. However, since the number of repeat copies is small (two for SMN1/SMN2), if all except one of the repeated sequences are masked in the reference genome, reads coming from all repeat copies will align uniquely to the masked reference genome. Hence, variant calling can be performed using standard variant calling tools with higher ploidy (4 for 2 repeat copies). In other words, masking repeat copies reduces the problem of variant calling in paralogous genes in diploid genomes to a “localized polyploid” variant-calling problem. Nevertheless, variants identified using this approach have an ambiguous position: instead of a single genomic position, each variant is associated with two or more (depending on the number of repeat copies) paralogous positions across the repeat copies. For each variant, an aggregate genotype (sum of the genotypes for each paralogous position) can be estimated, but information about the true location of a rare variant is typically unknown.

This approach—which we refer to as aggregate or pooled variant calling—has been successfully used in several studies to identify variants in disease-associated genes (summarized in Table S1). Cummings et al.⁷ developed a custom pipeline to identify variants in the triplicated repeat regions of the NEB (MIM: 161650) and TTN (MIM: 188840) genes in a cohort of individuals with neuromuscular disease. They identified two candidate pathogenic variants in these genes that were confirmed by follow-up genotyping. Gould et al.⁸ developed a workflow to detect variants in the duplicated exons of PMS2 (MIM: 600259) using a similar approach. Similarly, Boisson et al.⁹ were able to identify a pathogenic intronic splice variant in the IKBKG (MIM: 300248) gene—a known Mendelian disease gene for incontinentia pigmenti (MIM: 308300). Analysis of whole-genome sequencing (WGS) data from the 1000 Genomes Project (1KGP) using this approach enabled estimation of variant allele frequencies and confirmed that the pathogenic variant was absent in control subjects. Fleming et al. also used a genome masking strategy to enable the detection of variants in the HYDIN (MIM: 610812) gene¹⁰ and identified bi-allelic HYDIN variants in 15 families with the disease phenotype.

Ebbert et al.⁴ formalized this approach to detect and analyze variants in so-called dark genes using short-read sequencing and identified a rare frameshift deletion in the CR1 (MIM: 120620) gene in individuals with Alzheimer disease (MIM: 104300). Since then, several computational methods have expanded on this approach to enable more accurate variant calling in paralogous genes. Prodanov and Bansal¹¹ proposed a two-step framework for variant calling in paralogous genes: (1) aggregate variant calling using ploidy estimated from aggregate copy-number analysis¹² and (2) using PSVs to resolve aggregate genotypes into locus-specific genotypes. This method was shown to have a higher precision and recall than standard diploid variant-calling methods such as Freebayes and GATK,¹¹ particularly in genes with high sequence similarity. However, the paper focused only on benchmarking diploid variant-calling accuracy. More recently, Gilissen and colleagues described a method called Chameleolyser¹³ that identifies short sequence variants (SNVs and short insertions or deletions [indels]) and gene conversion events in duplicated genes using whole-exome sequencing (WES) data. The approach was validated using long-read sequencing and enabled the detection of disease-causing variants in 25 individuals with rare diseases. The authors refer to variants identified from aggregate variant calling as variants with ambiguous positions (VAPs). In spite of the positional ambiguity, VAPs can still be useful when searching for pathogenic variants in disease-associated genes. If a VAP is predicted to impact gene function (e.g., premature stop codon), the true location of the variant can be determined using targeted approaches. Additionally, for genes with intra-genic segmental duplications such as NEB¹⁴—any variant identified from aggregate variant calling lies within the gene itself and is likely to be functional.

Aggregate or polyploid variant calling is expected to be less accurate than diploid variant calling due to the higher ploidy. Cooke et al.¹⁵ performed benchmarking of single-sample polyploid variant calling using synthetic human datasets (created by merging multiple human samples). They found that the accuracy of polyploid genotyping at the same level of sequencing depth decreased with increasing ploidy. However, 94% of incorrect genotypes were due to incorrectly estimating variant allele copy number by ±1. This is not surprising since using 30×–40× sequencing depth, distinguishing between genotypes that differ by one allele copy (e.g., AAaa and Aaaa), can be difficult due to sampling noise. To establish the usefulness of aggregate genotypes for disease association studies, it is essential to evaluate genotyping accuracy at both the individual and population levels. Several important questions remain unaddressed. First, how accurate are aggregate genotypes across paralogous genes with different numbers of copies (e.g., two vs. three copies)? Second, can aggregate variant calls and genotypes enable rare variant association analyses in disease-associated paralogous genes when diploid variant calling has low recall? Third, can population-genetic analysis be done using aggregate genotypes? Finally, is aggregate variant calling beneficial across all paralogous genes or primarily for those with high sequence similarity?

In this paper, we utilized a targeted aggregate variant calling method that we recently developed¹¹ to address some of these questions. We performed a head-to-head comparison of aggregate variant calling with diploid variant calling in paralogous genes using both simulated and real human datasets. We benchmarked the precision and recall of aggregate variant calling in paralogous genes overlapping segmental duplications using data from the Genome in a Bottle (GIAB) project^16,17,18 and identified genes that show significantly higher recall using aggregate variant calling compared with standard variant calling. For common variants, we also assessed the accuracy of aggregate genotypes using population-level data. In addition, we performed orthogonal long-read-based validation of aggregate variant calls by comparison against haplotype-resolved de novo assemblies from the Human Pangenome Reference Consortium (HPRC). We applied aggregate variant calling to multiple disease-associated paralogous genes using sequence data from the 1KGP as well as disease-specific exome datasets. We used four disease-associated paralogous genes to showcase the utility of aggregate variant calls for population-genetic analysis (SMN1), genotyping of known pathogenic variants (NCF1), rare-variant association analysis (CFC1), and haplotype phasing (IKBKG). These analyses are either infeasible or incomplete using standard diploid variant calling due to the low recall and precision of variant calling in these genes.

Material and methods

aggregate variant calling

Consider a segmental duplication with n repeat copies such that we have a multiple sequence alignment (MSA) of the reference sequence for the repeat copies. Then, a variant is represented by an n-tuple of genomic positions (one for each repeat copy), where n is the number of repeat copies. This corresponds to the definition of VAP proposed in a recent paper.¹³ A PSV corresponds to positions in the MSA where the reference sequence differs between the different repeat copies.

For each segmental duplication, we use one of the repeat copies to perform aggregate variant calling and define the reference sequence. For a bi-allelic variant, we use A and a to denote the two alleles, where A (a) is the reference (variant) allele. For each individual with ploidy K, there are K + 1 possible aggregate genotypes that correspond to 0,1,…K copies of the A allele.¹⁹ K is determined by the aggregate copy number (total number of copies of the gene and its homologs) of a given individual. We denote the aggregate genotype for an individual with ploidy K and k reference allele copies as k//K. Another genotype representation involves writing down A and a alleles in a single string. In this representation, a genotype with k = 3 and K = 5 will be represented as the string AAAaa.

The method for aggregate variant calling implemented in Parascopy has been described previously.¹¹ We provide a brief overview for reference. First, we re-map reads from all repeat copies to a single copy and use polyploid variant calling with appropriate ploidy (illustrated in Figure 1A for a two-copy duplication). For WGS data, ploidy is set to the aggregate copy number estimated from the sequence data. For WES data, we set the ploidy to be 2 times the number of paralogous gene copies. Note that re-mapping is done locally and does not involve global re-mapping of reads. This enables the use of all reads mapped to paralogous regions, regardless of mapping quality, for sensitive variant detection. Variant detection was performed using a widely used variant caller Freebayes.²⁰ Parascopy uses standard genotype likelihood calculations to infer the most likely polyploid genotype for each individual. aggregate variant calling is available in the call module of Parascopy.¹²

Figure 1.

Overview of aggregate variant calling in paralogous genes and applications

(A) Reads originating from two paralogous copies of a gene are re-mapped to a single reference copy.

(B) Polyploid variant calling is performed on the re-mapped reads to identify variants. The aggregate variant calls for a variant with alleles A (reference) and a (variant) can be represented in two ways: “A … a” or “number of reference alleles//total number of copies.” Reference sequences of the two paralogous genes differ at paralogous sequence variants (PSV); corresponding PSV genotypes are either “trivial” when the genotype matches the reference or “non-trivial” otherwise.

(C) Genotype frequency comparison across different sample groups (case-control, population studies).

(D) Pairwise correlation between aggregate genotypes, revealing clusters of correlated variants (illustrated for SMN1).

(E) Variant annotation and rare variant association analysis: candidate loss-of-function and missense variants can be identified even when the genomic position is ambiguous.

(F) Statistical phasing of aggregate genotypes on chromosome X, where collapsed haplotypes for male individuals can be used to deconvolute tetraploid aggregate genotypes in females.

Defining target regions

To define target regions for performing aggregate variant calling, we first obtained a set of duplicated exons by intersecting the Parascopy homology table v.1.7¹² against the union of GENCODE v.38 protein-coding exons²¹ and Agilent v.6 exon regions. We added a padding of 1,000 bp to the selected regions and merged regions that were within 1 kb of each other. We then separated the resulting regions by copy number (Parascopy examine) and filtered out subregions that are (1) non-duplicated, (2) of high copy number (>8), (3) within highly complex duplication (marked as tangled within the homology table), (4) within dissimilar duplications (sequence similarity < 97%), (5) too short (<3 kb), or (6) covered multiple times (if accounting for all repeat copies). Finally, we added 100 bp padding to the filtered regions and merged regions within 1 kb of each other. All long regions (>40 kb) were split into shorter ones (≤25 kb). The resulting set of 1,876 target regions covers 12.2 Mb across chromosomes 1–X and overlaps 919 protein-coding and 772 non-coding genes or pseudogenes.

Simulated data

Following Prodanov and Bansal,¹¹ we first simulated a baseline diploid call set containing, on average, one artificial variant per 1 kb, with 80% substitutions, 10% insertions, and 10% deletions. For the Sim-Poly dataset, we additionally transferred PSV alleles from one copy to another according to allele frequencies, calculated by Parascopy¹² across 503 European individuals from the 1KGP. Simulated PSVs were introduced across 1,876 previously selected target regions; all other artificial variants overlapping PSVs were removed. Next, we used ART Illumina v.2016-06-05²² to generate a 30× dataset simulating HiSeq 2500 paired reads with 150 bp length and 500 bp fragment size.

Constructing aggregate variant calls

Variant call sets, generated by standard variant calling tools such as GATK HaplotypeCaller, as well as the GIAB high-confidence variant call set, contain diploid variant calls and genotypes in the VCF format. To enable comparisons between such diploid variants and genotypes with aggregate variants produced by Parascopy, we combined the diploid variants from various repeat copies into aggregate variants and genotypes. While it is easy to obtain aggregate genotypes for bi-allelic SNVs by concatenating corresponding diploid genotypes (a variant with diploid genotypes Aa and AA in SMN1 and SMN2, respectively, would receive aggregate genotype AAAa), in regions with multiple variants close to each other or regions where PSVs are adjacent to variants, it can be challenging to construct the aggregate genotypes. Instead, it is easier to construct aggregate genotypes using haplotype information. To do that, we use the Parascopy pool-vcf module to first construct haplotypes from input diploid variants and genotypes, pool haplotypes together, then split them into aggregate genotypes. Since haplotype information is not always available, we marked parts of haplotypes with nearby unphased variants (≤10 bp apart) as unavailable for variant aggregation. Another challenge is that for some paralogous genes, the GIAB benchmarking calls do not include all paralogous gene copies. Subregions where at least one of the pooled haplotypes was unavailable (absent in the GIAB call set or has close unphased variants) were excluded from comparison. Similarly to our previous study,¹¹ we also limited the comparison to regions where Parascopy predicted no deviations from the reference aggregate copy number.

Evaluating aggregate variant calling accuracy

We utilized high-confidence variant call sets for seven human genomes (HG001–HG007) generated by the GIAB Consortium.^16,18 The latest version (v.4.2.1) of these variant calls leverages long-read sequence datasets and covers many paralogous genes. For each individual, Illumina WGS data (30×–40× coverage) provided by the GIAB was used for variant calling. We used the RTG Tools²³ vcfeval module v.3.12.1 to compare aggregate Parascopy¹² 1.17, Freebayes²⁰ v.1.3.5, DeepVariant²⁴ v.1.0, and GATK HaplotypeCaller²⁵ v.4.2.2 calls against the baseline GIAB calls. Based on the GIAB call set, we separated PSVs into trivial (reference allele at each repeat copy) and non-trivial. Trivial PSVs were excluded from benchmarking since they contain a reference allele on every repeat copy. We then intersected true positive (TP), false positive (FP), and false negative (FN) variant call sets produced by RTG tools with various BED files containing boundaries of the protein-coding genes, highly similar segmental duplications (≥99.5% similarity), or trivial/non-trivial PSVs. Next, we calculated the total number of TP/FP/FN variants within each region and each combination of regions, counting variant overlaps partially, if only part of the reference allele was covered by the corresponding feature. This allowed us to calculate precision/recall for each individual gene, as well as to separately evaluate accuracy for non-trivial PSVs and non-PSVs. Additionally, for the per-gene accuracy analysis (Figure 2), we excluded genes with complex duplication structure (overlapping 6 or more separate segmental duplications), where transferring variant calls across paralog copies makes per-gene benchmarking unreliable.

Figure 2.

Improvement in variant-calling accuracy (precision, recall, and F₁ score) for aggregate variant calling relative to diploid variant calling across paralogous genes

The left-hand images show the accuracy on the simulated data (488 protein-coding genes with at least 10 variants), while the right-hand images show average accuracy across 7 GIAB samples (270 genes). Genes overlapping complex duplication structure (>6 individual segmental duplications) were removed beforehand due to difficulties with variant-calling evaluation. Standard variant-calling results were obtained using the GATK HaplotypeCaller tool, and aggregate variant calls were obtained using Parascopy. Each circle represents multiple genes with identical accuracy values (rounded to the nearest hundredth); the circle size represents the number of genes in each set. For simulated data, the circle at (1,1) (corresponding to 89–133 genes) is not drawn to scale. Accuracy values are highlighted for four disease-associated paralogous genes—NEB, NCF1, SMN1, and CFC1; among these, only NEB had benchmarking variants available for the GIAB samples.

Variant-calling concordance for the targeted PMS2 dataset

Gould et al.⁸ used long-range PCR (LR-PCR) to perform targeted sequencing of the paralogous PMS2 and the PMS2CL regions in 707 samples, generating two separate sets of reads per sample: one from the PMS2 locus and one from the PMS2CL locus. Of these, 141 samples overlapped with samples with WGS data from the 1KGP.²⁶ We downloaded FASTQ files for the PMS2-locus reads from the 141 samples (NCBI SRA Project PRJEB27948) and mapped the reads for each sample (using BWA-MEM) to a modified GRCh38 reference genome where the PMS2CL sequence was masked (replaced with “N” characters) so that the reads map uniquely to the PMS2 gene region. Next, we performed diploid variant calling (Freebayes, ploidy = 2) on the mapped reads for the 141 samples and obtained the allele frequencies from the VCF output. We performed the same analysis on the PMS2CL-locus reads for the 141 samples to obtain diploid variant calls. The diploid variant calls for the two paralogous regions (PMS2 and PMS2CL) were combined to create aggregate variant calls and genotypes for comparison with WGS data. We performed aggregate variant calling using Parascopy on WGS data (aligned CRAM files from the 1KGP) for the 141 samples using ploidy equal to 4 for the PMS2 region. Variant-calling comparison was limited to the paralogous region of PMS2 that was also targeted in the LR-PCR assays (chr7:5973212–5988480 [GRCh38]). Variants and allele frequencies between the two sets of calls were compared using custom Python scripts, and concordance between allele frequencies was calculated using the Pearson correlation coefficient.

aggregate variant calling on 1KGP samples

In order to evaluate the concordance of Parascopy calls between WES and WGS, we performed aggregate variant calling for paralogous genes with sequence similarity greater than 0.99 for 502 European-ancestry samples from the 1KGP using Illumina WGS²⁶ and WES²⁷ datasets. Parascopy variant calling for WGS data takes into account predicted sample copy numbers across the target loci. For aggregate variant calling using WES data, we also used WGS-based copy-number predictions for ploidy. Aggregate genotypes from WGS and WES data for individual variants and samples across 15 paralogous genes (listed in Table S3) were compared using a custom Python script. In order to limit the comparison to high-quality predictions, we discarded genotypes with low quality (<10) and low read depth (less than 5× ploidy). Finally, to simplify the comparison, we removed a small fraction of multi-allelic variants and variants with non-matching predicted alleles in the two datasets. aggregate variant calling was performed using Parascopy on the same set of 15 paralogous genes using WGS data for samples in the 1KGP from four other superpopulations (EAS [East Asian], SAS [South Asian], AFR [African], and AMR [admixed American]). Allele frequencies for aggregate variants were calculated by summing allele counts across aggregate genotypes.

Validation of aggregate variant calls using HPRC assemblies

We used haplotype-resolved genome assemblies for 30 European-ancestry individuals from the HPRC²⁸ for benchmarking aggregate variant calling and genotyping accuracy. For each of the 15 paralogous gene loci, we extracted the GRCh38 reference sequence of the primary gene region and aligned it to each haplotype-resolved assembly for individual samples using minimap2,²⁹ allowing for secondary and partial alignments. The resulting hits were converted to alignments in BAM format relative to the GRCh38 reference genome. We then identified variants and called genotypes jointly across all 30 individuals (60 haplotype assemblies) using Freebayes. For each individual, we aggregated genotypes across the two haplotype assemblies (maternal and paternal) to obtain a per-individual aggregate genotype for each variant site. We excluded large indels (>50 bp) from the HPRC-based variant calls prior to comparison. We compared the HPRC-based aggregate variant calls to the corresponding Parascopy aggregate variant calls from short-read WGS data (described above) for the same 30 individuals across all 15 genes. Variant sites were classified into three groups: sites shared between both call sets, sites unique to the HPRC calls (FNs in Parascopy), and sites unique to the Parascopy calls (FPs relative to HPRC). For shared variant sites, we calculated genotype concordance by comparing the HPRC-based and Parascopy-based genotype for each sample; Parascopy genotypes with quality below 10 were excluded from the concordance calculation. Since aggregate genotypes at different ploidies are not directly comparable, individuals that had a different ploidy (between the HPRC-based and Parascopy-based copy-number estimates) across more than 10% of the genotypes for any gene were excluded from the concordance calculation for that gene. Such discordances in aggregate copy number are a result of errors in the HPRC assemblies or the short-read-based copy-number estimates.³⁰ The NAIP gene was excluded from the comparison, as the aggregate copy number for this locus differed between WGS and HPRC assemblies across all samples. For the remaining 14 genes, 21 sample-gene pairs (∼1.5 per gene on average) were filtered. Variants overlapping long homopolymers and tandem repeats (GIAB stratification BED file; web resources) were also excluded from the comparison. For calculating mean (and median) statistics, we down-weighted the HYDIN variant calls (weight factor of 0.1), as it had almost 20 times the number of variants compared to other genes.

Variant annotation

For identifying coding variants in paralogous genes, we annotated aggregate variant calls using the SnpEff tool.³¹ Variants were annotated using RefSeq transcripts GenBank: NM_000344.4 (SMN1), NM_000265.5 (NCF1), and NM_032545.4 (CFC1).

SMN1 copy-number and variant analysis

Paralog-specific copy numbers for SMN1 and SMN2 were estimated jointly for all samples in each superpopulation using Parascopy.¹² We estimated haploid copy-number frequencies for each population using the observed SMN1 (and SMN2) diploid copy-number counts and assuming Hardy-Weinberg equilibrium. Estimation was done using a standard expectation-maximization (EM) algorithm that models each individual’s diploid genotype as a pair of unordered haplotypes. To analyze the rare missense variants in the AFR superpopulation, the same algorithm was applied to haplotypes defined jointly by the SMN1 copy number and the variant allele (reference or non-reference).

To identify tag-SNPs for two-copy SMN1 haplotypes, we compared aggregate genotype frequencies between individuals with an SMN1/SMN2 copy number of (4,0) and those with (2,2). Individuals with a (4,0) profile carry four copies of SMN1 and no copies of SMN2; based on the estimated haploid copy-number frequencies (Table S6), such individuals are overwhelmingly expected to carry a pair of two-copy SMN1 haplotypes (SMN1-SMN1). In contrast, individuals with a (2,2) profile are unlikely to harbor an SMN1-SMN1 haplotype, as the frequency of two-copy SMN2 haplotypes is very low.³² SNVs with non-reference alleles enriched in the (4,0) group were considered tag-SNPs for the SMN1-SMN1 haplotype.

Estimating pairwise correlation from aggregate genotypes

To calculate correlation between pairs of bi-allelic variants using aggregate genotypes, we used a multinomial EM algorithm for polyploid genotypes (equation 9 in Gerard³³), one of several estimators proposed in that work. The pairwise correlation (r²) calculated by this method is a polyploid extension of the standard linkage disequilibrium (LD) r² measure used for analyzing diploid genotypes. Genotypes with quality < 10 were set to missing prior to estimation. We used the r² estimates for pairwise correlation analysis and visualization.

NCF1 gene and exome data for complex immune phenotype cohort

Similuk et al.³⁴ reported results from clinical exome sequencing of more than 1,000 families with complex immune phenotypes. WES data (BAM files aligned to the GRCh37 reference genome) for 1,880 individuals (1,545 affected individuals and 335 control subjects) were downloaded from the Database of Genotypes and Phenotypes (dbGaP: phs001899). aggregate variant calling on the 1,880 samples using WES data for the NCF1 gene was performed using Parascopy (ploidy = 6). GATK HaplotypeCaller v.4.2 was run on the BAM files using default parameters. Association analysis for the aggregate genotypes was performed using Fisher’s exact test.

CFC1 variants and association with heterotaxy

For identifying coding variants in the CFC1 gene, we used aggregate variant calls for 2,504 1KGP samples. The Pediatric Cardiac Genetics Consortium (PCGC) is a cohort study of individuals with congenital heart defects (CHDs) that contains DNA sequence data from more than 15,000 individuals. We identified a total of 331 probands presenting with heterotaxy (among other phenotypes) using dbGaP phenotype data. Paired-end sequencing reads (FASTQ files) from exome-targeted sequencing were downloaded from dbGaP: phs000571. Reads were aligned to the GRCh38 human reference genome using BWA-MEM³⁵ (v.0.7.17). PCR duplicates were marked using Picard, and multiple sequencing runs for each sample were merged to produce a single BAM file per individual. For comparison, 438 control samples (without heterotaxy and preferentially without any CHD phenotype) were selected from the same cohort, prioritizing individuals with good sequence coverage (>5 billion total aligned bases). All control samples were processed using the same alignment and post-processing pipeline as described above. aggregate variant calling using WES data for the CFC1 gene (processed region: chr2:130592165–130599575) was performed using Parascopy with ploidy set to 4 to account for the two-copy segmental duplication. Variants with low genotype quality (<20) were filtered out. All identified variants were functionally annotated using SnpEff (v.5.2) based on Ensembl gene models (GRCh38.76). For comparing the frequency of the carriers of rare missense variants between affected individuals and control subjects, only variants that were annotated as missense for the canonical transcript (GenBank: NM_032545.4) were used.

IKBKG phasing analysis

For IKBKG phasing, Parascopy copy-number calling and variant calling were run in a 57 kb target region (chrX:154536200–154593327 [GRCh38]) encompassing IKBKG and the corresponding segmental duplication. Variant calling and subsequent analysis were done in 3,202 1KGP WGS samples²⁶ (including 602 trios), grouped by superpopulation. For phasing analysis, we identified 11 variants, present in every superpopulation with a minor-allele frequency (MAF) of 5% or greater. Trio and statistical phasing was restricted to individuals with no copy-number variation. For male individuals with a single copy of IKBKG and IKBKGP1, the genotype formed by combining the haplotypes of the two loci is referred to as a “collapsed haplotype” or diploid genotype.

Trio phasing

In trios, a male child possesses only maternal alleles, which can therefore be subtracted from the maternal tetraploid genotype to split it into two diploid genotypes. For a female child, we phased all variants, called in all trio members by first separating paternal and maternal alleles by using the father’s diploid genotype. Then, we split the mother’s genotypes by subtracting maternal alleles from either male or female children. This method works for aggregate genotypes of any copy number when all repeat copies are located on the X chromosome irrespective of the distance between repeat copies.

Statistical phasing

Since the two repeat copies at the IKBKG locus lie close to each other (separated by 57 kb of unique sequence), the haplotypes of the gene and its pseudogene are not completely independent. For statistical phasing, we enumerated all diploid genotypes across 11 common variants for all male individuals without copy-number variants (CNVs) and tabulated their frequencies. The aggregate genotype for a female individual is a vector of length 11 with ploidy 4. For each such genotype, we identified all possible combinations of diploid genotypes (from males) that sum to the tetraploid genotype. The likelihood of each phasing was defined as the product of the diploid genotype frequencies, and the phasing with the highest likelihood was selected for each female individual. A subset of females could not be phased using this simple strategy and were not used for comparison to trio-based phasing. To compare the two phasing methods, we selected all females without CNVs that were successfully phased with both methods. Then, we matched two pairs of diploid genotypes to each other to minimize the mismatch and calculated the number of variants that were phased differently.

Ethics declaration

This study is a secondary analysis of previously published and publicly available datasets. No new human subjects data were collected. WGS and WES data from the 1KGP and the GIAB Consortium are publicly available. Access to controlled-access data from the PCGC and the immune disease cohort was obtained through dbGaP (project #28953). The immune disease sequence data were approved for general research use (no institutional review board [IRB] approval required). The PCGC exome data were approved for health/medical/biomedical research and disease-specific research (congenital heart disease). In this study, the PCGC data were used only for analyzing the association of variants in the CFC1 gene with heterotaxy (congenital heart disease).

Results

Overview of aggregate variant calling

The aggregate variant calling approach implemented in Parascopy implicitly constructs a masked reference genome (all repeat copies except one are masked) for a paralogous genomic region and realigns reads to the masked genome (Figure 1A). Variant calling is performed using Freebayes,²⁰ a widely used short-read variant-calling tool that can perform variant calling for one or more samples with different ploidies. Sample ploidy corresponds to the aggregate copy number of the region, estimated using the Parascopy cn module¹² from WGS data. If copy-number information is not available (e.g., WES data), ploidy can be assumed to be constant (twice the number of repeat copies). Subsequently, the Freebayes variant calls are analyzed within the Parascopy call module to estimate aggregate genotypes across multiple samples and filter out low-confidence variants. We denote the reference and alternate alleles—relative to the main repeat copy—as “A” and “a,” respectively. Following the notation of Gerard et al.,¹⁹ we also use a shortened notation R//K to represent aggregate genotypes or allele dosage, where R is the number of reference alleles “A” and K is the total ploidy. Correspondingly, a tetraploid genotype with three reference alleles and one alternate is written as “AAAa” or 3//4 (Figure 1B).

Benchmarking aggregate variant calling accuracy on individual genomes

Prodanov and Bansal¹¹ focused on benchmarking the improvement in variant-calling accuracy in paralogous genes using a combination of aggregate variant calling and locus-specific genotyping. However, when repeat copies have very high sequence similarity or when PSVs cannot be used to distinguish paralogous copies due to gene conversion, paralog-specific genotypes cannot be determined for many variants. In such cases, aggregate variants and genotypes represent the best information obtainable from short-read sequencing. To identify paralogous genes for which aggregate variant calling can provide significantly better recall than standard variant calling, we utilized simulations (see material and methods). These simulations incorporated information about the reliability of PSVs estimated by the Parascopy copy-number estimation method.¹² We compared the recall of Parascopy (aggregate variant calling [agVC]) with the recall of GATK HaplotypeCaller²⁵ and Freebayes²⁰ (standard variant calling [stVC]). Among a set of 488 duplicated protein-coding genes with closely related paralogous sequences and at least 10 benchmarking variants, the agVC recall was higher than GATK recall in 347 cases (Freebayes: 350) and exceeded GATK recall by 25 percentage points (p.p.) for 158 genes (Freebayes: 165). For the disease-associated gene CFC1 (MIM: 605194), aggregate variant calling showed perfect precision and recall (both = 1.0) compared to Freebayes precision and recall of 0.250 and 0.220, respectively (GATK: 0.211 and 0.176). Similarly, for the SMN1 gene, which is associated with spinal muscular atrophy (SMA [MIM: 253300]), precision and recall were 0.984 and 0.924, respectively, while both stVC tools showed precision of 0.368 and recall of 0.124 (Figure 2). Notably, genes overlapping a duplication with 99% sequence similarity and at least 3 kb overlap were significantly enriched among those genes for which Parascopy showed >25 p.p higher recall compared to GATK (Fisher’s exact test p < 10⁻³⁹, odds ratio > 15). Parascopy showed recall at least 10 p.p. lower than GATK for 15 of the 488 simulated genes (Figure 2; per-gene values in Table S2); 10 of these 15 genes overlapped duplications with moderate sequence similarity (<98%), where standard variant calling has sufficient sequence divergence for accurate variant calling.

Next, to assess the significance of these results on real data, we analyzed WGS data for seven human genomes from the GIAB Consortium for which high-confidence variant call sets covering a subset of paralogous genes have previously been published.^16,17,18,36 To enable a direct comparison of the precision and recall of aggregate variant calling with standard variant calling, variants called using stVC methods were converted to aggregate variant calls (see material and methods). In addition to GATK and Freebayes, we included comparison with DeepVariant²⁴ using previously generated variant calls.³⁷ Similar to the results on simulated data, Parascopy showed higher recall than GATK on 221 protein-coding out of 270 genes (Freebayes and DeepVariant: 222 and 224), improving recall over stVC by at least 10 p.p. for 118 (Freebayes: 121, DeepVariant: 125) genes (Figure 2), including many disease-associated genes such as NEB, which showed an improvement in recall of 54 p.p. (0.329 compared to 0.867). Notably, the GIAB benchmark variant calls completely or partially exclude several disease-associated genes such as CFC1 and SMN1, which showed significant improvement in recall on simulated data. A small subset of genes (9/270) had at least 2 p.p. better precision or recall using stVC (GATK) compared to aggregate variant calling (Table S2). 8 of these 9 genes overlapped duplications with moderate sequence similarity (similarity <98%), most likely explaining the higher recall using diploid variant calling.

Among the seven human genomes with high-confidence variant calls, HG007 has the largest GIAB benchmarking call set across the target regions, containing 4,663 PSVs and 7,440 other variants. As shown in Figure S1, Parascopy achieved high accuracy for both variant classes (precision ≥ 0.99; recall = 0.980 and 0.965 for PSVs and non-PSVs), while standard variant callers showed lower accuracy at non-PSVs (GATK precision and recall = 0.909 and 0.877) and especially at PSVs (0.819 and 0.856), where reference sequence of one repeat copies appears on the other copy, often causing read mapping errors. Not surprisingly, the difference becomes larger when only considering 447 variants, 500 bp deep into almost indistinguishable repeats (similarity ≥ 99.5%). While Parascopy precision and recall drop to 0.956|0.945 and 0.896|0.874 (non-PSVs|PSVs), stVC methods perform markedly worse, with GATK precision reaching 0.765|0.253 and recall of 0.549|0.323. Compared to GATK and Freebayes, DeepVariant more accurately genotyped non-PSVs within highly similar regions (precision = 0.878) but detected fewer variants (recall = 0.422), resulting in an F₁ score of 0.57 compared to 0.64 for GATK and 0.63 for Freebayes.

For all variant callers, accuracy on the simulated data across highly similar duplications was similar to that of the HG007 dataset. Nevertheless, because the simulated dataset includes more challenging segmental duplications that are excluded from the GIAB benchmarks (7,036 of 38,090 simulated variants located deep within highly similar duplications), the accuracy of standard variant calling across all duplications decreased by over 15 p.p. (Figure S1B).

All variant callers perform well on trivial PSVs (F₁ score ≥ 0.99 on HG007), where all repeat copies carry their respective reference allele. Parascopy’s F₁ score drops slightly within highly similar duplications (0.961 on HG007), whereas standard callers retain near-perfect accuracy since they analyze the repeat copies independently (no variant call is a correct call). This apparent advantage for stVC tools comes at the cost of substantially lower recall for non-trivial PSVs.

Accuracy of aggregate variant calling and genotypes for population data

Benchmarking of aggregate variant callingusing Parascopy on GIAB data demonstrated high precision and recall for individual genomes. For aggregate variant calling to be useful for disease association analysis, genotyping accuracy must remain high for both common and rare variants. To assess population-level accuracy of aggregate variant calling, we utilized a targeted sequence dataset for the PMS2 gene.⁸ This dataset was generated by LR-PCR-based amplification of PMS2 (and its paralog PMS2CL) followed by Illumina HiSeq sequencing to a very high depth of coverage (∼1,000×). Variant calling for the targeted sequence data was performed using Freebayes for PMS2 and PMS2CL separately (see material and methods). 141 of the 707 individuals sequenced in this study had WGS data available from the 1KGP. We performed aggregate variant calling in the 17 kb duplicated region of PMS2 using Parascopy on these 141 individuals. Comparison of allele frequencies across 469 variants shared between the two sets of variant calls showed a very high correlation (r² = 0.9944). Of the 107 variants with an allele count of 1 in the targeted sequence data (singleton variants), 95.3% (102) were also detected in the WGS data using Parascopy and with an identical allele count (across all samples). Similarly, of the 40 variants with an allele count of 2 in the 141 individuals, 38 were detected in the WGS data (37/38 had a matching allele count). These results showed that aggregate variant calling had high sensitivity (95.2% for singletons and doubletons) for rare variants and high concordance for allele frequencies.

Next, we performed concordance analysis on aggregate variant calls and genotypes from WGS and WES data for 502 individuals of European descent across 15 paralogous genes with high sequence similarity (Table S3). The set of genes included 12 paralogous genes with two copies and 3 genes with three copies. We stratified variants based on the predicted aggregate copy number (4 and 6 for two-copy and three-copy genes, respectively) and MAFs (rare: <1% and common). Since dense clusters of variants can result in reduced target capture efficiency and lower sequencing depth in exome sequencing,³⁸ we removed dense clusters of variants (at least 10 variants within 150 bp) from the comparison. Across 210 common variants with aggregate copy number (CN) equal to 4 (6.2 × 10⁴ genotypes), the genotype concordance rate was 99.5%. For 237 common variants with aggregate CN = 6 (3.4 × 10⁴ genotypes), concordance was lower at 98.1% (Figure 3A). This is not surprising since it becomes more difficult to distinguish genotypes that differ by a single allele (such as 2//6 and 3//6) as ploidy increases.¹⁵ Concordance was very high (>99.96%) for 545 rare variants (2.1 × 10⁵ genotypes) since most samples did not have an alternate allele and it is easier to separate dosage 5//6 from 6//6 than 2//6 from 3//6. When considering all variants together irrespective of rarity and genotype copy number, allele frequencies in the WGS and WES datasets were highly concordant (Pearson’s r² > 0.9998; Figure 3B).

Figure 3.

Population-level benchmarking of aggregate variant calls and genotypes across disease-associated paralogous genes

(A and B) Concordance of aggregate variant genotypes between WGS and WES data. Genotypes for 502 individuals of European ancestry (from 1KGP) across 832 bi-allelic variants in 15 disease-associated paralogous genes with high sequence similarity were used for comparison.

(A) Total number of sample-variant pairs for which corresponding WGS and WES aggregate genotypes were available. Counts are stratified by variant allele frequency (≤1% and >1%) and aggregate copy number for the locus (only 4 and 6 shown).

(B) Comparison of allele frequencies estimated from WGS and WES aggregate genotypes, with each point representing a single variant.

(C and D) Comparison of short-read aggregate variant calls with HPRC haplotype assemblies for 30 European-ancestry individuals across 15 paralogous genes (NAIP was excluded due to discordant copy number). (C) Per-gene precision and recall of aggregate variant calls, using HPRC assembly-based calls as ground truth. (D) Per-gene genotype concordance between aggregate genotypes and HPRC assembly-based genotypes.

To validate aggregate variant calls using an orthogonal approach, we utilized high-quality diploid de novo haplotype assemblies for 30 European-ancestry individuals from the HPRC²⁸ (see material and methods). We compared Parascopy aggregate variant calls to HPRC assembly-based calls across the same set of 15 paralogous genes used for WGS-WES concordance. The recall (fraction of HPRC-identified variants also identified using aggregate variant calls) was high across all genes, with a median of 0.994 and a mean of 0.990 (gene-level statistics in Figure 3C). Similarly, precision was also high, with a median of 0.993 and a mean of 0.992. The genotype concordance at shared variant sites was 0.996 and 0.985 for two-copy and three-copy genes, respectively (Figure 3D). The fraction of genotypes excluded due to low quality (genotype quality score < 10 for Parascopy aggregate genotypes) was higher for three-copy genes (11.1%) than for two-copy genes (4.0%; Table S4), reflecting the greater challenge of accurate genotyping at higher ploidy.¹⁵

Having established the accuracy of aggregate variant calling and genotyping, we analyzed genetic variation in four disease-associated paralogous genes (SMN1, NCF1, CFC1, and IKBKG) using aggregate variant calling applied to WGS and WES data from the 1KGP and disease-relevant cohorts. Each gene was chosen to showcase a specific type of population or disease association analysis (illustrated in Figures 1C–1F) enabled by aggregate calling but not feasible with standard diploid variant calling due to low recall/precision. For SMN1, we demonstrate population-level analysis of aggregate copy-number and variant genotypes—localizing rare missense variants and identifying tag-SNPs for the two-copy SMN1 haplotypes. For NCF1, we demonstrate the genotyping and association analysis for a well-characterized pathogenic variant that is difficult to genotype using standard calling due to pseudogene interference (Figure 1C). For CFC1, we re-examined the association of rare protein-coding variants with heterotaxy using case-control WES data (Figure 1E), made possible by the high recall of aggregate calling compared to standard calling (recall < 0.25) in this gene. Finally, for the X-linked IKBKG gene, we demonstrate haplotype phasing of aggregate genotypes using both trio-based and statistical approaches (Figure 1F).

SMN1 and population-level analysis of coding variation

The SMN1 and SMN2 genes are nearly identical paralogs located on chromosome 5q13. SMA, a severe neuromuscular disorder, is primarily caused by bi-allelic loss or deletion of SMN1.^39,40 Missense and loss-of-function variants in SMN1 are known to contribute to SMA pathogenesis in an estimated 5% of individuals.⁴¹ Since SMN2 has >99.9% sequence identity with SMN1, diploid variant calling using short reads has very low recall. Therefore, we used aggregate copy-number and variant genotypes obtained from 1KGP samples to analyze coding variants. We used Parascopy to obtain copy-number and aggregate variant calls for SMN1 from 2,504 individuals representing five continental superpopulations in the 1KGP (material and methods). Two loss-of-function variants and 26 missense variants were detected across the five superpopulations (Table S5). Both loss-of-function variants were rare and detected in one or two individuals. Notably, all three individuals with the loss-of-function variants were estimated to have 4 copies of SMN1 and 1 copy of SMN2 using Parascopy.¹² Among the 26 missense variants, 17 were singletons, and two variants had two carriers. Notably, only a single missense variant (c.865T>A [p.Cys289Ser]) was observed in the AMR superpopulation (n = 352 individuals). Exon 7 is the only exon that has sequence differences (PSVs) between SMN1 and SMN2 located within or near the coding sequence. The c.859G>C (p.Gly287Arg) missense variant, detected in four individuals, was located less than 20 bp away from the splice-site PSV (c.840C>T [p.Phe280=]), and Parascopy was able to determine that the variant allele was present on the SMN2 haplotype.¹¹ This was consistent with previous studies showing that this variant reduces the severity of SMA by increasing the expression of the full-length transcript from SMN2.⁴²

Three missense variants were observed in 6 or more individuals (Table S5), almost all in individuals of AFR ancestry. In each carrier, the variant allele was estimated to have an allele count of one. In the absence of neighboring PSVs, it was not feasible to directly determine whether these variants are present on SMN1 or SMN2 haplotypes. One of these variants (c.5C>G [p.Ala2Gly]) has been found in individuals with SMA in combination with an SMN1 deletion and is likely pathogenic.⁴³ Another variant (c.796T>C [p.Ser266Pro])—with 13 carriers—is also known to be pathogenic for SMA.⁴⁴ Analysis of the distribution of SMN1 copy number for the carriers of these missense variants in 661 AFR-ancestry individuals revealed that all three variants were present only in individuals with an SMN1 copy number ≥ 3 (Table 1, n = 358) and absent from all 300 individuals with an SMN1 copy number of 2. The complete absence of these variants in individuals with two copies of SMN1 (Table 1) indicates that the variant allele is carried exclusively on SMN1-SMN1 haplotypes (chromosomes with two copies of SMN1). This was confirmed by EM-based haplotype frequency estimation (material and methods). This likely explains the relatively high frequency for these missense variants in the AFR superpopulation since only one of the two SMN1 copies on a haplotype is affected by the rare missense variants.

Table 1.

Distribution of aggregate variant genotypes and SMN1 copy number for three low-frequency SMN1 missense variants identified using aggregate variant calling for 661 AFR individuals from 1KGP

Variant	Aggregate genotype	SMN1 copy number
		1	2	3	4
c.5C>G (p.Ala2Gly)	K//K	3	300	258	94
	K-1//K	0	0	3	3
c.278A>C (p.Lys93Thr)	K//K	3	300	244	82
	K-1//K	0	0	17	15
c.796T>C (p.Ser266Pro)	K//K	3	300	254	91
	K-1//K	0	0	7	6

Each cell in the table gives the number of individuals with a missense variant genotype (K//K is the reference genotype and K-1//K is the genotype with one non-reference allele; K is the aggregate copy number) and SMN1 copy number (1–4). SMN1 CN was estimated from WGS data using Parascopy. Variant annotation is based on the canonical RefSeq transcript NM_000344.4.

Using aggregate copy-number genotypes across five continental superpopulations and an EM algorithm assuming Hardy-Weinberg equilibrium (material and methods; Table S6), we estimated that SMN1-SMN1 (two-copy SMN1) haplotypes occur at a frequency of 34.4% in the AFR superpopulation—substantially higher than in the SAS (8.43%) or European (2.53%) superpopulation and consistent with previous estimates of 30%–35%^45,46—and a frequency of 35.6% from a long-read sequencing study.⁴⁷ Given this high frequency, we searched for sequence variants that tag the SMN1-SMN1 haplotype configuration in the AFR superpopulation (material and methods). In 661 AFR ancestry individuals, 81 had a (4,0) SMN1/SMN2 copy-number profile, and 115 had a (2,2) profile; SNVs enriched in the (4,0) group were identified as SMN1-SMN1 tag-SNPs (Figure 4). An SNV (chr5:70951020G>A) was present on 55.5% of two-copy SMN1 haplotypes and completely absent on other haplotypes, with an aggregate allele frequency of 11.36% and 0.73% in AFR and American superpopulations, respectively. Another SNV (chr5:70923922G>C), with a frequency of 57.6% on two-copy SMN1 haplotypes vs. 1.75% on others, was independently identified in the long-read sequencing study, where it marked S1-9d haplotypes over-represented in two-copy SMN1 alleles.⁴⁷ The SNP at chr5:70952074—designated g.27134T>G and previously used in clinical testing to identify silent carriers in Ashkenazi Jewish populations⁴⁸—was enriched in AFR individuals with a (4,0) copy number (75.4% haplotype frequency) but also present on 29.5% single-copy SMN1 haplotypes (Figure 4).

Figure 4.

Frequency distribution of aggregate genotypes for three SNVs that tag two-copy SMN1 haplotypes in African-ancestry samples

For each of the three SNVs, 661 African samples from 1KGP were partitioned into four groups based on (SMN1,SMN2) copy number estimated using the paralogous sequence variant c.840C>T. Each sub-plot displays the frequencies (absolute count is shown on top of each bar) of different aggregate genotypes within that copy-number group. Individuals with (4,0) and (3,1) copy numbers show substantially higher non-reference genotype frequencies compared to individuals with (2,2) and (2,1) copy numbers for each of the three SNVs. Each SNV is represented by two positions (first in SMN1 and second in SMN2).

To further analyze the variants in individuals with SMN1-SMN1 haplotypes, we computed pairwise correlation (r²) between aggregate variant genotypes for the 81 individuals with a (4,0) SMN1/SMN2 copy-number profile using an EM algorithm (material and methods) and grouped variants into clusters using a pairwise r² threshold of 0.6. Four clusters of correlated variants were identified (Figure S2). The largest cluster (34 variants) included all three tag-SNPs described above. Comparing the variant clusters with long-read-based SMN1/SMN2 haplotypes from Chen et al.,⁴⁷ we found that one cluster of eight variants matched the S1-9c haplotype (only observed on SMN1-SMN1 alleles in long-read data) exactly, while the tag-SNP cluster corresponded to the S1-9 haplotype lineage (over-represented among two-copy SMN1 alleles).

NCF1 variants and association with immune disease

Mutations in the NCF1 (MIM: 608512) gene that encodes p47-phox are known to cause chronic granulomatous disease (CGD [MIM: 233700])—a monogenic immune disorder.⁴⁹ NCF1 shares high homology with two pseudogenes, NCF1B and NCF1C, located on the same chromosome. A 2 bp coding deletion (c.75_76delGT [p.Ala26ProfsTer22]) is the most common pathogenic variant in NCF1 and results from gene conversion between the NCF1 and the pseudogenes that carry the deletion allele (illustrated in Figure 5A).^49,50 As a result, reads originating from NCF1 can mismap to the pseudogenes, making it difficult to accurately genotype the variant using standard variant calling. We used aggregate variant calling to analyze the population frequencies for this pathogenic ΔGT variant and perform case-control association analysis in an immune disease cohort.

Figure 5.

Aggregate variant calling for the triplicated NCF1 locus

(A) Genomic structure of the NCF1 locus with the NCF1 gene and the two pseudogenes NCF1B and NCF1C. Arrows indicate the direction of transcription. The three copies reside within a 105 kb segmental duplication with >99.5% sequence similarity (yellow). The duplication between NCF1 and NCF1C extends for an additional 37 kb (orange). A key paralogous sequence variant that differentiates NCF1 (GTGT sequence) from the two pseudogenes (GT sequence) is shown in red.

(B) Frequencies of aggregate genotypes for the GTGT/GT (ΔGT) variant across 2437 1KGP samples from five superpopulations with high-quality genotypes (using NCF1 as the primary copy). Aggregate copy number (K) varies across individuals: 96.4% have CN = 6, with 1.1%, 2.5%, and 0.05% showing CN = 5, 7, and 8, respectively.

(C) Frequencies of aggregate genotypes for the ΔGT variant excluding the genotype 2//K (same data as in B but without the bars for the reference genotype).

(D) Distribution of WES reads supporting the two alleles for the ΔGT variant across 487 individuals of European ancestry with an estimated CN = 6. Points are colored based on the predicted GTGT aggregate genotype; low-quality calls (genotype quality < 10) are shown with a cross. Lines indicate expected allele count ratios for different aggregate genotypes.

Analysis of aggregate genotypes for the ΔGT variant across the 1KGP samples showed that more than 80% of individuals in all superpopulations carry exactly two copies of the functional NCF1 allele (Figure 5B). The frequencies of individuals with 1 or 3 copies of the GTGT allele were highly variable across 1KGP superpopulations (Figure 5C), with the EAS superpopulation having the lowest frequency, while European ancestry samples had the highest frequency. In the European superpopulation, 14.3% and 1.4% of individuals had 3 and 4 copies of the functional GTGT allele, respectively. These estimates were highly concordant with previously reported ddPCR-based frequencies in European individuals (14.7% and 1.6% for 3 and 4 copies, respectively).⁴⁹ 4 individuals in the European superpopulation were estimated to have only one copy of the functional NCF1 allele using WGS data. The genotypes for 3/4 individuals were concordant with WES data, with the only discordant sample showing a low genotype quality of 4 in the WES data (Figure 5D). Therefore, we estimate that approximately 0.6% of European-ancestry individuals harbor a single functional copy of NCF1 and are carriers for CGD. Comparison of the GT/GTGT allele counts for the ΔGT variant in WES data showed that the observed ratios were consistent with expected ratios for different aggregate genotypes, supporting the reliability of the estimated genotypes from WES data.

Next, we analyzed genotypes for the ΔGT variant in a cohort of 1,880 individuals (1,545 affected individuals with complex immune abnormalities and 335 control subjects) that underwent WES.³⁴ Notably, four individuals were reported to have a prior molecular diagnosis with pathogenic NCF1 mutations; however, NCF1 mutations were not identified from WES using standard variant calling. We performed aggregate variant calling for the NCF1 gene using a ploidy of 6 for all individuals (material and methods). The count distribution of the number of GTGT alleles is shown in Table S7. Notably, 10 individuals had a genotype equal to 0//6 corresponding to zero copies of the functional NCF1 allele. All 10 individuals were affected(no additional phenotype information was available to confirm the presence of CGD). Similar to the 1KGP data, the carrier frequency (genotype = 1//6) was 0.80%. An exact binomial test indicated that observing 10 or more homozygotes in 1545 affected individuals (given a carrier allele frequency equal to 0.8%) is highly unlikely (p = 1.4 × 10⁻¹¹). For comparison, we performed standard variant calling using GATK HaplotypeCaller (material and methods) using aligned reads for the 1,880 samples and analyzed genotypes for the ΔGT variant. 6 of the 10 individuals were correctly called as 2//2, which corresponds to zero copies of the functional GTGT allele (consistent with the aggregate genotypes). However, the read depth for all 6 samples at this position was <20 (range = 3–19), most likely because most of the reads with the GT deletion from the NCF1 copy were incorrectly mapped to the pseudogene copies. In contrast, the mean aggregate read depth (summed across the three copies) for this variant for the 10 individuals was 545x. The remaining four individuals had genotypes equal to 0//2 or 1//2 with very low read depth (<5 reads). We performed association analysis for the ΔGT variant and a disease-associated missense variant (rs201802880) in the immune disease cohort (see material and methods) but did not observe a significant association (Table S7). We also identified four additional loss-of-function variants (2 stop-gain, 1 frameshift, and 1 splice-donor) in the immune disease cohort (Table S8), including a known founder mutation in Jewish populations.⁵¹ All four carriers of the Jewish founder mutation were affected individuals with complex immune phenotypes.

CFC1 variants and association with heterotaxy

The CFC1 gene, spanning 7.4 kb of DNA sequence, lies in a human-specific segmental duplication on chromosome 2 and shares high sequence homology with CFC1B,⁵² resulting in a recall of less than 25% for standard variant calling (Figure 2). As a result, CFC1 has been largely excluded from sequencing studies despite being a Mendelian disease gene for heterotaxy syndrome (MIM: 605376).^53,54 We analyzed aggregate variant calls from 1KGP superpopulations and a pediatric disease cohort to analyze genetic variation in CFC1 and examine whether rare variants in this gene are associated with heterotaxy.

Analysis of PSV genotypes for CFC1 in the 1KGP European superpopulation revealed that none of the 14 PSVs correspond to fixed differences that can differentiate CFC1 and CFC1B (Table S9), making paralog-specific variant calling infeasible. Analysis of the aggregate variant calls revealed two frameshift and two stop-gain variants (Figure 6; Table S10). One of the frameshift variants (c.522delC [p.Ala175fs]) was the same variant that was previously identified in two individuals with heterotaxy.⁵³ This variant was detected in 13 individuals in the 2,504 1KGP samples, including 6 of European ancestry (see Figure S3 for visualization of reads for one individual). The genotypes for the six European-ancestry samples with the frameshift variant were concordant between WGS and WES data.

Figure 6.

Protein-coding variants identified using aggregate variant calling in the paralogous genes CFC1 and CFC1B

Aggregate variant calling was performed using WGS data for 2,504 samples from 1KGP, WES data for 331 individuals with heterotaxy (affected individuals), and WES data for 438 control subjects (without heterotaxy). Protein-level variant annotations are with respect to RefSeq transcript NM_032545.4; single-letter amino acid codes are used for brevity. Loss-of-function (stop-gained, frameshift) variants identified from 1KGP samples are shown at the bottom (see Table S10 for a complete list). Three common missense variants observed in all cohorts (Table S11) are highlighted by boxed labels. Exons are shown with black rectangles and are scaled according to their size; introns are not shown to scale. Variants are labeled with either a left semicircle (appear only in the affected individuals), a right semicircle (only in control subjects or 1KGP), or a full circle (appear in both groups).

Next, we analyzed WES data from PCGC,⁵⁵ comprising 331 affected individuals with congenital heart disease and heterotaxy and 438 control subjects (without congenital heart disease or heterotaxy). aggregate variant calling for WES data was performed using a ploidy of 4. Four common missense variants (allele frequency > 0.005) were identified in the affected individuals (Figure 6). Each of these four variants was also detected in the control subjects at a similar frequency and the 1KGP samples (Table S11). Three of the four variants showed strong pairwise correlation with each other (r² > 0.68 using aggregate genotypes). Since the affected individuals and control subjects were drawn from multiple continental populations, we did not carry out single-variant association analysis for common variants. An additional four missense variants had low allele frequency in the individuals with heterotaxy and control subjects but showed a high MAF (>1%) in the AFR superpopulation from 1KGP (Table S12). Excluding these variants, 10 of the 331 (3.02%) probands with heterotaxy were carriers of rare missense (with respect to the canonical transcript) variants in CFC1. In comparison, the frequency of rare missense variant carriers in the controls was 1.14% (two-sided Fisher’s exact test p = 0.053). For the five missense variants (c.482A>C [p.Asp161Ala], c.461C>T [p.Pro154Leu], c.388G>C [p.Ala130Pro], c.218C>T [p.Pro73Leu], and c.172T>C [p.Phe58Leu]) that were observed in a single individual with heterotaxy, we assessed if the variant was de novo by analyzing the WES data for the parents. For each of the five missense variants, one of the parents had clear evidence of the variant allele with an aggregate genotype equal to 1//4, ruling out the possibility that any of these variants are de novo. Finally, we searched for individuals that harbored two deleterious coding variants, a single variant with two allele copies, or known pathogenic CFC1 variants. One individual harbored a rare variant with a genotype of 2//4; however, the variant was inherited from an unaffected parent (see the supplemental note for details).

IKBKG gene and pseudo-diploid genotypes via phasing

IKBKG encodes NEMO, an essential modulator of NF-κB signaling, and mutations or partial gene deletions cause X-linked diseases (incontinentia pigmenti and hypohidrotic ectodermal dysplasia [MIM: 300291]).^56,57 Exons 3–10 of IKBKG span a 36 kb segmental duplication with 99.9% sequence similarity that harbors the IKBKGP1 pseudogene (Figure 7A). Hemizygous males typically carry one copy each of IKBKG and IKBKGP1, providing aggregate diploid genotypes that can be used as a reference panel for phasing the aggregate tetraploid genotypes observed in females.

Figure 7.

Aggregate genotype-based phasing of the X chromosome gene IKBKG

(A) The protein-coding gene IKBKG and its pseudogene IKBKGP1 (cyan) are located on the X chromosome and overlap a segmental duplication with ≈99.9% sequence identity (yellow). Genomic coordinates correspond to the GRCh38 reference assembly.

(B) A typical male possesses a single copy of the IKBKG locus. Aggregate variant calling produces a collapsed IKBKG+IKBKGP1 haplotype, consisting of haploid genotypes in the non-duplicated part of IKBKG (two variants shown) and diploid genotypes within the duplicated region (four variants shown). Black diamonds and red squares denote the reference (“A”) and variant (“a”) alleles, respectively.

(C), Trio phasing with a female child. The daughter inherits one X chromosome from each parent. The daughter’s aggregate variant calls can be decomposed into two collapsed IKBKG+IKBKGP1 haplotypes, which in turn enable phasing of the mother. For trios with a male child, the mother can be phased using the child’s collapsed haplotype.

(D) For population-based statistical phasing, frequencies of all collapsed haplotypes are estimated from male individuals and used to find the most likely pair of collapsed haplotypes that explain a tetraploid genotype for each female individual.

We utilized two strategies to phase aggregate variants and genotypes in females that are similar to trio-based and statistical phasing approaches used for phasing diploid genotypes (illustrated in Figures 7C and 7D; see material and methods). In both cases, diploid genotypes in non-duplicated exons are phased completely, while tetraploid genotypes in duplicated exons are split into a pair of diploid genotypes, each representing one IKBKG + IKBKGP1 haplotype (Figure 7B). Analysis of aggregate variant calls and genotypes across the 1KGP samples for the five PSVs located in the duplicated region (chrX:154555685–154565100) revealed that all of these PSVs are polymorphic (see Table S13 for summary statistics) and cannot be used to distinguish IKBKG from IKBKGP1. Using the aggregate variant calls, we selected 11 variants that were common in all continental superpopulations (MAF ≥ 5%) for phasing. 2 of these variants lie in a unique region (exons 1 and 2 and upstream of IKBKG), and the remaining 9 variants lie in the segmental duplication region (see Table S14 for a list of these variants). Across the 1KGP cohort, 813 females had a reference copy number at the IKBKG locus and were part of a trio. 810 of these females were phased successfully using trio phasing. Three individuals were not phased since other individuals in the trio had copy-number changes.

For statistical phasing, we selected 1,458 female samples (total of 1,604 samples) that exhibited no CNVs in the IKBKG region. Of these, 1,311 individuals (89.9%) were successfully phased using statistical phasing, and no suitable haplotypes were found for the remaining 147 samples. 723 females could be phased using both trio phasing and statistical phasing. Across 723 individuals phased with the two methods, the phased genotypes matched in 99.2% of comparisons (7,953 heterozygous variants); this fraction increased to 99.5% when considering 7,733 (97.2%) genotypes with high genotype qualities (≥10).

Discussion

Advances in high-throughput sequencing technologies over the past two decades have resulted in rapid progress in identifying new genes and mutations for rare diseases. There is increasing recognition that variants in genes that overlap segmental duplications are either missed or remain difficult to detect using standard variant calling pipelines for short-read WGS and WES,^3,4,18 and new computational methods continue to be developed to address this problem.^13,58,59 In this paper, we showed that aggregate variant calling using our bioinformatics tool Parascopy significantly improves variant recall—without sacrificing precision—for paralogous genes with high sequence similarity such as SMN1 and CFC1. Comparison with haplotype-resolved de novo assemblies for 30 samples showed precision and recall above 99% and high genotype concordance (>98%) across 14 paralogous genes, providing independent validation of the accuracy of short-read aggregate variant calling. By analyzing four different disease-associated paralogous genes, we were able to show that tag-SNP analysis, case-control rare variant association analysis, and haplotype phasing (for an X-linked gene) can be performed using aggregate variant calls and genotypes. Each of these applications was showcased using a specifically chosen paralogous gene; systematic evaluation across diverse paralogous genes remains future work.

For the well-studied SMN1 gene, analysis of aggregate genotypes converged on two-copy SMN1 haplotypes—common in AFR populations. First, we showed that low-frequency missense variants observed in AFR-ancestry individuals were carried exclusively on two-copy SMN1 haplotypes, explaining their appreciable allele frequencies despite putative pathogenicity. Second, we identified tag-SNPs for two-copy SMN1 haplotypes using aggregate genotype frequencies. Finally, pairwise correlation analysis of aggregate variants recovered variant clusters that correspond to long-read-defined haplotypes,⁴⁷ including the S1-9c haplotype that occurs exclusively within two-copy SMN1 alleles. The tag-SNP analysis has direct clinical relevance, as silent carriers (individuals with two copies of SMN1 on one chromosome and zero copies on the other) test negative on standard copy-number-based carrier screening assays. Our results show that correlation analysis of aggregate genotypes can complement long-read approaches in identifying population-specific tag-SNP panels for clinical use in populations with a high frequency of two-copy SMN1 haplotypes.

Heterotaxy is a congenital disorder of left-right asymmetry, with congenital heart disease present in over 70% of affected individuals.⁶⁰ Mutations in CFC1 were first associated with heterotaxy syndrome in 2001⁵³ using candidate-gene sequencing. Recent large-scale sequencing studies of congenital heart disease^61,62 have likely been unable to examine the CFC1 gene thoroughly because standard variant calling has low recall (<25%) in this human-specific segmental duplication. Analysis of 1KGP WGS data revealed that a previously reported pathogenic frameshift mutation in CFC1 is actually a rare variant in human populations. Additionally, several loss-of-function mutations were detected in 1KGP samples, making high-penetrance pathogenicity for CFC1 loss of function unlikely. Analysis of WES data did not reveal a statistically significant burden of rare missense variants in individuals with heterotaxy compared to control subjects. We were also able to rule out the possibility that rare CFC1 missense variants in affected individuals are de novo. However, two previously reported pathogenic missense variants were not detected in the >3,000 individuals analyzed in our study. Collectively, our results show that the association of CFC1 mutations with heterotaxy syndrome should be re-examined.

Haplotype phasing was demonstrated specifically for the X-linked IKBKG gene, where aggregate diploid genotypes in hemizygous males were used as a reference panel of collapsed haplotypes for phasing aggregate tetraploid genotypes in females. Extending this approach to autosomal paralogous genes would require an equivalent database of diploid-collapsed genotypes, which is not currently available from short-read data. Long-read sequencing and haplotype-resolved assemblies such as those from the HPRC^28,63 can potentially be used to generate the reference panel and enable phasing of autosomal paralogous genes.

A recent study used long-read HiFi sequencing and haplotype phasing to analyze genetic diversity in paralogous genes with high sequence similarity that are challenging even for long reads.^47,64 Notably, the method used in this study was also based on an aggregate read-mapping approach where long reads are realigned to a single locus. This demonstrates that aggregate read analysis is a natural framework for analyzing paralogous genes using both short-read and long-read sequencing. Compared with short reads, the long-range information present in HiFi reads improves the mappability of reads in paralogous genes and enables haplotype reconstruction. Nevertheless, our results show that aggregate copy number and variant calling provide valuable information using short-read sequencing.

Limitations of the study

aggregate variant calling and our study have several limitations. First, aggregate genotypes collapse variant information across paralog copies, and paralog-specific genotypes cannot be determined for most variants—a key constraint for genes such as SMN1 and NCF1. Second, aggregate genotypes provide lower resolution for distinguishing genotypes that differ by a single allele (e.g., 1//4 vs. 2//4), which can be important for distinguishing between heterozygous and recessive states. Third, our analyses were restricted to genes with two or three paralogous copies, and accuracy was generally lower for three-copy genes than for two-copy genes. Paralogous genes reside in segmental duplications that vary widely in copy number, sequence similarity, and gene conversion^5,30; aggregate variant calling using short reads is most useful for genes or genic subregions with high sequence similarity or those with gene conversion that limit standard calling. Long-read sequencing is needed to analyze paralogous genes with higher copy number or complex duplication structures. Finally, aggregate variant calling requires the aggregate copy number to be specified as the ploidy for variant calling, and we did not systematically evaluate the impact of incorrect ploidy on accuracy—a concern most relevant for WES data. A recent computational method has shown that aggregate copy number can be estimated for paralogous genes from WES data with high accuracy (∼97%)⁶⁵; therefore, aggregate variant calling from WES data can also be performed using appropriate ploidy information.

Recent advances in long-read sequencing and haplotype-resolved de novo assemblies of human genomes^2,28,63,66 have enabled detailed analysis of segmentally duplicated regions of the human genome. Several recent studies have applied long-read sequencing to specific paralogous disease genes such as SMN1/SMN2,^47,67 IKBKG,⁶⁸ and HYDIN.¹⁰ Nevertheless, for the vast amounts of short-read WGS and WES data that have already been generated for hundreds of thousands of human samples, aggregate variant calling in combination with downstream analysis approaches—such as those explored in this paper—has the potential to identify disease-associated variants in known disease genes as well as interrogate paralogous genes that are not yet disease-associated. Developing computational and statistical methods that can analyze aggregate genotypes and also leverage information from long-read sequencing will be important to fully utilize the information available in short reads. These advances will allow paralogous genes to be incorporated into population genetics and disease association studies at scale.

Data and code availability

WGS and WES data for 1KGP samples are available from the 1KGP FTP site. All other sequence datasets analyzed in this paper are available from NCBI SRA or dbGaP as detailed in the material and methods section. The code used for analyzing these datasets is available at github.com/tprodanov/agvc-paper. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This work was supported by a grant from the National Human Genome Research Institute (NHGRI; R01 HG010149). The WES data (dbGaP: phs000571) used for association analysis of heterotaxy were generated by the PCGC, under the auspices of the National Heart, Lung, and Blood Institute’s Bench to Bassinet Program. This manuscript was not prepared in collaboration with investigators of the PCGC, has not been reviewed and/or approved by the PCGC, and does not necessarily reflect the opinions of the PCGC investigators or the NHLBI. The sequence data for the cohort with complex immune phenotypes (dbGaP: phs001899.v2.p1) was funded by the NIAID Intramural Research Program (NIH Office of Intramural Research) and submitted by Morgan Similuk. We greatly appreciate the funding sources and investigators of these studies for sharing the data without which the current study would not have been feasible.

Author contributions

T.P. implemented the methods, performed the benchmarking experiments, and analyzed 1000 Genomes datasets. V.B. designed and supervised the work. V.B. analyzed the 1000 Genomes and dbGaP datasets. S.Y.B. analyzed the dbGaP data for heterotaxy. T.P. and V.B. wrote the manuscript. All authors reviewed and approved the final version of the manuscript.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used Claude Code to help clarify and revise individual paragraphs. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Web resources

• GenBank, https://www.ncbi.nlm.nih.gov/genbank/

• GIAB, https://ftp-trace.ncbi.nlm.nih.gov/giab/ftp/

• GIAB (GRCh38_AllTandemRepeatsandHomopolymers_slop5.bed.gz), ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/release/genome-stratifications/v3.0/GRCh38/LowComplexity/

• HPRC, https://data.humanpangenome.org/assemblies

• Online Mendelian Inheritance in Man, http://www.omim.org

• Parascopy, github.com/tprodanov/parascopy

• The International Genome Sample Resource, https://www.internationalgenome.org/data-portal/data-collection/30x-grch38

Supplemental information

Table S2. Per-gene precision and recall of aggregate variant calling (Parascopy) and standard variant calling (GATK HaplotypeCaller) across simulated and GIAB benchmarks

The file contains two tabs: one for the simulations and one for the GIAB genomes. It is provided as a separate Excel file.