Accurate imputation of inversions in human genomes using different algorithms and data sources

Abstract

Complex genomic regions harbor different structural arrangements that can mutate quite rapidly, which makes determining their functional effects very difficult. Characterization of inversions originated by homologous mechanisms is especially challenging due to the presence of inverted repeats at the breakpoints and the fact that most of them are recurrent. Imputation is a useful method to infer missing genotypes, but it has been mainly limited to simple variants and little is known about how well it works for human inversions. Here, we tested five common imputation programs to impute a set of 52 inversions, which have been experimentally genotyped in multiple samples and lacked perfectly linked single nucleotide polymorphisms (SNPs). Using whole-genome sequencing data and simulated microarrays with variable SNP density, we found that 40.4%–75.5% of inversions could be accurately imputed in three human populations by at least one program, with results depending mostly on inversion recurrence and the number of available SNPs and genotyped samples. Besides, genotype probability filtering was a key factor for inversion imputation accuracy. In particular, Minimac4 and IMPUTE5 showed more accurately imputed inversions and less poorly imputed individuals with respect to the other methods. This work therefore contributes to optimizing inversion imputation in order to study their functional impact.

Introduction

During the last decades, there has been a great interest in the complete characterization of human genomic variation and its association with phenotypic traits and disease susceptibility. Traditionally, single nucleotide polymorphisms (SNPs) and small indels have been the main target, with many large-scale projects across multiple individuals and populations, such as the 1000 Genomes Project (1KGP) [1, 2] or the Genome Aggregation Database (gnomAD) [3]. In addition, several studies using diverse genomic techniques are finally enabling the identification of all the other more-complex structural variants (SVs) [2, 4–9]. However, due to the technical challenges of the detection of these variants, in most cases the number of reliable genotypes is limited, which complicates determining their functional impact.

Specifically, inversions are a type of SVs that reverse the orientation of a DNA segment. This confers unique biological properties to inversions, and they have been associated with phenotypic traits and adaptation in multiple species, but it also makes them particularly difficult to study [10]. Moreover, while inversions generated by nonhomologous (NH) mechanisms typically have nearby variants in perfect linkage disequilibrium (LD) that facilitate estimating accurately the inversion status [11], inversions mediated by highly identical inverted repeats (IRs) often undergo recurrent toggling between the original and inverted orientations via nonallelic homologous recombination (NAHR) [11–13]. This dynamic process increases haplotype diversity and, as a consequence, there is no single variant or set of variants in high LD with the inversion that can reliably predict its orientation. Therefore, low-throughput experimental techniques are required to determine the genotype of the inversions, and their effects have been likely missed in existing genome-wide association studies (GWAS), which are based mainly on SNPs [11].

GWAS is the standard procedure to check the association of genetic variants to phenotypic traits in large cohorts. These analyses were performed extensively with microarray data, that allow a cost-efficient genotyping of a reduced amount of common SNPs in a high number of samples, and they are now increasingly performed by whole-genome sequencing (WGS) at lower or deeper coverage [14, 15]. Imputation is widely used to infer missing variant genotypes for downstream analyses, and multiple imputation algorithms have been developed, each one with its strengths and weaknesses [16–18]. Some of the most popular tools are Beagle, which uses a haplotype cluster graph model [19] and IMPUTE2, that applies a Markov chain Monte Carlo framework, which scales poorly [20]. Moreover, there are other methods, such as IMPUTE5, that enhances this process with positional Burrows–Wheeler transform for fast, accurate imputation from large panels [21]; or Minimac4, which employs a state space reduction while using a Markov chain haplotyping procedure [22]. In addition, scoreInvHap was designed to infer genotypes of a few inversions by creating haplotype clusters with flanking SNP LD data, and it has been used to examine inversion phenotypic associations [23–25]. However, there is still very little information on the performance of different imputation methods for SVs, given the specific characteristics of these variants [26, 27].

In this study, we took advantage of the generation of a valuable data set of experimentally validated genotypes of 134 inversions in 95–550 diverse samples [11, 12, 28] to evaluate the performance of human inversion imputation by benchmarking five widely used imputation tools, focusing on those inversions lacking variants in perfect LD. Our primary aim was to assess the best strategy to impute inversions using current state-of-the-art methods and to identify the main factors affecting imputation reliability, in order to be able to determine their functional effects.

Materials and methods

Small variant and inversion data

All analyses were carried out using the GRCh38 (hg38) human genome reference assembly. The small variant information (including SNPs and small indels), population of origin and sex of each individual for the reference and target panels used for phasing and imputation were extracted in VCF format from the 1KGP high-coverage WGS dataset (1KGP-HC), which can be accessed at https://ftp.1000genomes.ebi.ac.uk/vol1/ftp/data_collections/1000G_2504_high_coverage/working/20220422_3202_phased_SNV_INDEL_SV/ [2]. The genetic maps required for phasing and imputation were obtained from https://bochet.gcc.biostat.washington.edu/beagle/genetic_maps/. VCF files were filtered to contain only biallelic variants and exclude singletons using BFCtools 1.19 (with arguments view –min-ac 2:minor -M2 -m2 -v snps) [29]. No additional minor allele frequency (MAF) filter was applied.

Highly-accurate experimental inversion genotypes to carry out and validate the imputation procedure were obtained in previous studies by different PCR-based methods [11, 12, 28]. The dataset includes 52 inversions without known tag SNPs in perfect LD (r² = 1) across African (AFR), East-Asian (EAS), and European (EUR) super-populations based on a previous analysis of the original set of 134 inversions with 1KGP-HC data (Supplementary Table S1). Inversions had been genotyped in a variable number of 86–438 AFR, EAS, and EUR 1KGP individuals depending on the technique used (Supplementary Fig. S1). Detailed characterization of the inversions was done in the initial studies where the genotypes were described, including the analysis of the unique origin and recurrence events for most of them, based on haplotype differentiation [11, 12] (Supplementary Table S1). Singleton inversions with <2 minor alleles in the genotyped individuals across populations were not considered in the analysis.

Variant phasing

Most imputation methods, except for scoreInvHap, use haplotype information to impute unknown variants and they require the input data to be phased. Thus, we re-phased the 1KGP-HC VCF files to assign each of the experimental inversion alleles to the corresponding haplotype. For this, we extracted the variants of an extended region including the inversion and up to 200 kb upstream and downstream from its breakpoints (Supplementary Table S1) with BCFtools [29]. Available inversion genotypes were added as another variant (0/0, 0/1, or 1/1) in the VCF in the central position of the inversion region, using bash and vcf-sort (VCFtools 0.1.16) [29]. Males were treated as diploid for chromosome X, transforming hemizygous 0 and 1 genotypes to 0/0 and 1/1. Next, we tested phasing with Beagle v5.4 (22Jul22.46e) and SHAPEIT5 (5.1.1) [19, 30] and the results were compared using a Wilcoxon signed-rank test. For Beagle, phasing was conducted using default parameter settings. For SHAPEIT5, the phase_common and the phase_rare steps were run respectively with the arguments –filter-maf 0.05, and –pbwt-modulo 0.005. Then, the processed VCF files were divided into the different super-populations using BCFtools [29].

Inversion imputation

To benchmark inversion imputation performance, we used Beagle 5.4 (22Jul22.46e), IMPUTE2 (2.3.2), IMPUTE5 (1.2.0), Minimac4 (4.1.6), and scoreInvHap (1.30.0) [19–23]. All programs were run independently for different super-populations with default parameters, specifying human effective population size (N_e) as 20 000 when possible and adding –pbwt-cm 0.005 to IMPUTE5. To determine the imputation accuracy of most programs, a leave-one-out (LOO) strategy was employed. In this approach, an individual with masked inversion genotype is the target for the imputation, using the rest as the reference data, and the process is repeated for all the individuals of the super-population. Imputation accuracy of each method for the different inversions and super-populations was determined by calculating the LD metric r² between all imputed and experimentally obtained inversion genotypes, which ranges from 0 to 1 (perfect imputation), using PLINK (1.90b7.2), with parameters –ld-window-kb 1 000 000 –ld-window 999 999 –ld-window-r2 0 [31]. The imputation process was repeated five times for each method with different random seeds, and the median r² value of the five replicates was taken as imputation accuracy. For scoreInvHap [23], the input LD values (r²) of the inversion with small variants within the 200 kb extended region were calculated from the VCFs of each super-population using PLINK [31], with the same parameters as before. The remaining input information was obtained by processing the data according to scoreInvHap documentation (https://bioconductor.org/packages/release/bioc/html/scoreInvHap.html). As scoreInvHap depends only on the LD values and does not consider the real inversion genotypes, imputation accuracy was obtained directly by calculating the r² between experimental and imputed genotypes in each super-population, without requiring LOO processing. In all cases, for inversions in which the imputation results only showed one of the two orientations or the genotype probability (GP) filtering removed all samples containing one of the orientations, r² could not be calculated.

In addition, to check if imputation improves when increasing the number of available genotypes, imputation for all methods except scoreInvHap was repeated for the global (GLB) population, combining together the AFR, EAS, and EUR individuals. The imputation, LOO and r² calculation were done as previously described by using as reference the unsplit VCF files. All imputation results are provided in Supplementary Table S2.

GP filtering

For each imputed variant, imputation algorithms report GP, with values ranging from 0 to 1, representing the likelihood of the three possible genotypes (reference homozygote, heterozygote and alternative homozygote, respectively). In all analyses, the inversion genotype with the highest GP was selected as the most probable one, and, for simplicity, GP refers to the maximum value observed for each inversion and individual. GP values were extracted from the genotype field for the imputed variants of the VCF file. To determine the effect on imputation accuracy of eliminating low confidence imputed genotypes, different GP filters, ranging from 0 to 0.95, were applied to remove those individual genotypes that did not meet the minimum GP criterion prior to the calculation of the r² in the LOO process. Alternatively, to compare the effect of filtering the same number of low confidence imputed genotypes across programs, we also removed 5%–25% of samples with lowest GP for a given inversion and imputation method prior to r² calculation. Ties with the same GP values for an inversion were resolved by random selection of the filtered samples.

All follow-up comparisons, except for when specifically mentioned, have been carried out applying a GP filtering threshold of ≥0.9. Imputable inversions in a population were considered those with an imputation accuracy of r² ≥ 0.8 and >50% remaining samples after the 0.9 GP filtering.

Imputation in SNP arrays

To check inversion imputation performance from SNP genotyping array data, two representative arrays with different SNP densities were simulated: the high-density Illumina Infinium Omni 2.5–8 v1.3 array (Omni), interrogating approximately 2.37 million SNPs; and the low-density Illumina Infinium Global Screening Array v2.0 (GSA), with approximately 650 000 SNPs. To simulate these arrays, CSV files containing the SNP information in the GRCh38 assembly were downloaded from https://support.illumina.com/array/array_kits/humanomni2_5-8_beadchip_kit/downloads.html and https://support.illumina.com/array/array_kits/infinium-global-screening-array/downloads.html. Then, the position of the array SNPs was retrieved and those within the inversion plus 200 kb flanking region were selected from the filtered 1KGP-HC VCFs using BCFtools [29]. Finally, splitting of samples per population, imputation, LOO, and calculation of imputation accuracy were done as described above (Supplementary Table S2).

Statistics and linear model analysis

Comparisons between phasing and imputation methods or reference panels were performed using a nonparametric paired Wilcoxon signed-rank test, including always the set of inversions in common to the conditions compared. Comparisons between populations were done with a nonparametric unpaired Wilcoxon rank-sum test, since the inversions polymorphic in each population are not the same. The Pearson correlation coefficient (R) was used to compute correlations between programs or variables. Linear regression models were fitted to the imputation accuracy (r²) using different biological features as independent variables, such as number of genotyped individuals, ratio between IR and inversion length, MAF, density of SNPs in the region (number of SNPs divided by the size of the region used for imputation) and being in the chromosome X or not (Supplementary Table S3), with the formula r² ∼ Predictors in R. Imputation results from the three populations were combined and, with the exception of scoreInvHap, one linear model per imputation method was fitted with and without GP < 0.9 filtering.

Results

Comparison of inversion imputation methods

As already mentioned, we took advantage of the available highly-accurate inversion genotypes from multiple 1KGP individuals [11, 12, 28], together with their high-coverage WGS data [2], to evaluate how well different imputation strategies and methods infer the genotypes of 52 inversions (45 generated by NAHR and 7 by NH methods) without any tag SNP in complete LD (r² = 1) across the three studied super-populations (AFR, EUR, and EAS) (Supplementary Table S1). Inversion genotypes available in each population varied from 27 to 179 (with EAS having in most cases the lowest number of samples), and MAF of polymorphic variants in a population ranged between 0.006 and 0.5, although 90% of the variants (47/52) had a MAF higher than 0.1 in at least one population (Supplementary Fig. S1). Inversion imputation accuracy was assessed using the LOO strategy by calculating the r² between the real and imputed genotypes in all the available samples where the inversions had been experimentally genotyped (see the ‘Materials and methods’ section). Specifically, we tested the performance of inversion phasing, five common imputation programs, filtering of imputed genotypes, and using a population-specific or global reference for imputation.

Phasing

Since variant phasing is required for three of the imputation programs (Beagle, IMPUTE5, and Minimac4), and recommended for IMPUTE2, first we assessed whether imputation results differ between phasing methods. We found significantly higher imputation accuracy across inversions and populations when inversion phasing was done with Beagle compared to SHAPEIT5 in the four imputation programs tested (Supplementary Fig. S2). Therefore, only Beagle was used for phasing in subsequent analyses.

Imputation programs

Next, we tested inversion imputation performance of five imputation programs: Beagle, IMPUTE2, IMPUTE5, Minimac4, and scoreInvHap [19–23]. When considering all the inversions and populations together, imputation accuracy varied considerably between methods. The best performing programs were IMPUTE5 and Minimac4, followed closely by IMPUTE2, while Beagle, and particularly scoreInvHap, showed significantly worse results than the other three in all populations, except EAS (Fig. 1 and Supplementary Table S4). These differences could also be appreciated in the inversion imputation accuracy distribution, with IMPUTE2, and, especially, IMPUTE5 and Minimac4 generating sharper peaks closer to r² = 1 in most populations, whereas Beagle and scoreInvHap were able to impute perfectly a smaller number of inversions (Fig. 1 and Supplementary Fig. S3).

Figure 1.

Inversion imputation results in human populations using five programs. Density plots show the distribution of imputation accuracy values (r²) of the 52 analyzed inversions for all samples, divided by imputation program, with the three super-populations represented together in different colors (purple for AFR, blue for EAS, and yellow for EUR). Colored dots below the graphs indicate the actual values for every inversion, with boxplots at the bottom representing the median (black line) and quartile values of each population and method.

The distribution of inversion imputation accuracy also differed between populations, although, except for Beagle, no significant differences were found (Table 1 and Fig. 1). In general, inversions were imputed more accurately in EUR, followed by AFR. In contrast, EAS showed a wider r² range that was biased to lower values in all programs (Supplementary Fig. S3), which is probably due to the reduced number of genotyped samples in this case (Supplementary Fig. S1).

Table 1.

Summary of inversion imputation results for different programs in three populations

	AFR (49 inv.)		EAS (47 inv.)		EUR (52 inv.)		Wilcoxon rank-sum test P-value
Program	n	r 2	n	r 2	n	r 2	AFR-EUR	AFR-EAS	EUR-EAS
Beagle	49	0.640	47	0.501	50	0.744	0.018	0.164	0.004
IMPUTE2	48	0.780	42	0.635	51	0.798	0.468	0.220	0.121
IMPUTE5	49	0.848	47	0.645	52	0.858	0.656	0.141	0.086
Minimac4	49	0.846	47	0.665	52	0.858	0.668	0.129	0.098
scoreInvHap	49	0.472	47	0.557	51	0.638	0.560	0.076	0.284

For each method and population, the median imputation accuracy (r²) of all inversions for which both orientations were present in real and imputed genotypes and the r² can be calculated (indicated by n) is shown. Inversion imputation between populations was compared using an unpaired Wilcoxon rank-sum test and significant differences without multiple comparison correction are shown in boldface. The number of polymorphic inversions in each population is indicated within parenthesis.

GP filtering

The GP is a score generated by each software that estimates the degree of confidence of the imputed genotypes. Except for scoreInvHap, the GP values for every method were usually high, with 83.7% (IMPUTE2) to 96.0% (Minimac4) of them exceeding 0.9 (Supplementary Fig. S4A).

To check the effect of GP filtering on inversion imputation, we screened a range of GP thresholds from 0.5 to 0.95 in the different programs (except for scoreInvHap, for which no filtering was applied due to its peculiar GP distribution). When considering the common set of inversions that can be analyzed by all programs, Minimac4 and IMPUTE5 consistently showed the highest median imputation accuracy, with r² values improving slightly with increasing GP (Fig. 2A). This effect is more drastic for IMPUTE2, in which median imputation accuracy in the three populations reached values closer to or even higher (e.g. EAS) than the other two programs as GP threshold increased (Fig. 2A). However, IMPUTE2 improvement came at a cost of excluding 8.4–17.4% of low-confidence genotypes in the three populations at GP ≥ 0.90, comprising most samples for some inversions, which reduced the number of inversions with reliable results (Fig. 2A). Conversely, despite having high GP values similar to the other programs, median imputation accuracy in Beagle remained lower than in other methods across all GP filter thresholds (Fig. 2A).

Figure 2.

Effect of GP filtering on inversion imputation accuracy (r²) with different imputation programs. (A) Median r² (solid line) and sample loss (dashed line) in each population across increasing GP filtering thresholds for all inversions together. (B) Median r² (solid line) and lowest GP value of the remaining genotypes (dashed line) in each population for all inversions, after removing progressively the worst imputed genotypes of every inversion (measured by the lowest GP) in 5% intervals (up to 25%). The number of analyzed inversions per population (n) are indicated on top of each graph. To make the results comparable, in each analysis and population we included only inversions that retain samples with both orientations at the strictest filtering criteria (GP ≥ 0.95 or 25% filtered samples), in which r² with experimental genotypes can be calculated, across all the programs. In addition, those with <50% of samples remaining at the highest GP filter were excluded from the analysis.

As observed, GP values of the imputed genotypes varied among programs. Therefore, we also assessed the effect of filtering a fixed fraction of genotypes. For each inversion, 0%–25% of the samples with the lowest GP were removed in increments of 5% and those inversions retaining both orientations at the strictest GP filter across all programs in each population were included in the comparison (Fig. 2B). When the same proportion of samples was eliminated, Minimac4 and IMPUTE5 were again the best performing inversion imputation software for all filters, whereas IMPUTE2 showed lower median r² in AFR and EUR and similar values in EAS for ≥15% of filtered genotypes. On the other hand, Beagle and scoreInvHap had the lowest median r², although filtering equal number of samples made Beagle results more comparable to those of IMPUTE2, even surpassing it in EUR. Similarly, sample filtering improved considerably scoreInvHap results, especially in EAS, in which it showed higher imputation accuracy than Beagle (Fig. 2B).

Population versus global reference panel

Since some inversions were originally genotyped in a small number of samples of certain populations, like EAS, we also tested the effect on inversion imputation of using a global reference panel of all the samples of the three populations together (GLB), instead of just those from the population of the individual. Using a GP filtering of 0.9, the expansion of the reference size by combining different populations had overall a small effect on the results of the four programs that rely on a reference panel rather than on LD with SNPs (scoreInvHap), although it improved inversion imputability in certain situations (Supplementary Table S2). When looking at the distribution of the r² difference between the GLB and population-specific reference, Beagle displayed considerable variation, followed by IMPUTE2, in which several inversions showed clearly better imputation accuracy with the GLB panel (Fig. 3 and Supplementary Fig. S5). In contrast, for IMPUTE5 and Minimac4, a large proportion of inversions were imputed equally well in the GLB compared to the AFR or EUR population panels, although imputation accuracy with the GLB reference increased significantly in EAS (Fig. 3 and Supplementary Fig. S5). As expected, inversions with higher imputation improvement in the GLB panel tended to be those with lower population sample size (Supplementary Table S2).

Figure 3.

Comparison of imputation accuracy (r²) obtained by using different reference panels. The graphs show the density distribution across all analyzed inversions (Y-axis) of the difference in r² of the global minus the population panel (X-axis) for four imputation programs, with the three populations represented in different colors. Individual graphs for each population and additional data are shown in Supplementary Fig. S5.

Inversion imputation results

Next, we assessed how many of the analyzed inversions could be imputed accurately from WGS data in each population [defined as those with good imputation accuracy (r² ≥ 0.8) and >50% of samples remaining after filtering]. When using the population-specific reference panel, excluding low-confidence genotypes generally increased the number of imputable inversions, with the GP ≥ 0.9 filter typically outperforming the results with GP ≥ 0.8 or without GP filtering in most programs (Fig. 4A). However, in Minimac4 and IMPUTE2 the distribution of GP values is skewed towards higher values (Supplementary Fig. S4), and the number of imputable inversions were usually very similar across GP thresholds. Conversely, scoreInvHap accurately imputed very few inversions without filters (12.8%–23.1%) and GP filtering removed virtually all the imputed genotypes (Fig. 4A). Therefore, except for scoreInvHap, a GP ≥ 0.9 filter was established as the standard criteria for inversion imputation.

Figure 4.

Inversion imputability with different programs in three populations. (A) Effect of GP filtering (including no filter and GP ≥ 0.8 or 0.9) on the number of inversions that can be imputed by each program (r² ≥ 0.8 and ≥ 50% remaining samples), expressed as the percentage of polymorphic inversions having at least two chromosomes with the minor orientation in the analyzed samples of that population. (B) UpSet plots showing the intersection of polymorphic inversions that are imputable across different imputation tools after applying a GP ≥ 0.9 filter (except for scoreInvHap) in the three populations. Each column represents a unique combination of programs that successfully imputes a given inversion and the number of imputed inversions in that combination (bar chart). Left bars indicate the total amount of inversions imputed by every method.

Considering the 0.9 GP filter, imputable inversions in each population with the four best performing programs ranged from 27.7% (13) in EAS for Beagle to 63.5% (33) in EUR for Minimac4. Minimac4 and IMPUTE5 were the programs imputing the highest number of inversions in EUR and AFR or in EAS, respectively, with IMPUTE2 following closely (Fig. 4A). Beagle, on the other hand, could only impute reliably 27.7%–50.0% of the inversions. When combining the results of all the programs, the total number of imputable inversions increased to 75.5% (37) in AFR and 65.4% (34) in EUR, while, consistent with the lower imputation accuracy (Figs 1 and 2), only 46.8% of inversions (22) could be accurately imputed in EAS (Fig. 4B). Looking into the imputability at inversion level, a high concordance was found between methods, and most inversions could be accurately imputed by at least two of them, with Minimac4 and IMPUTE5 being the biggest contributors and imputing basically the same inversions (Fig. 4B). Nevertheless, a population bias was observed in the distribution of imputation methods across inversions. Whereas in EAS and EUR the vast majority of imputable inversions (86.4% and 88.2%, respectively) could be imputed by >2 methods (and many by even four or five), in AFR there was more variation between programs and this number reached only 73.0% (Fig. 4B), probably due to the higher haplotype diversity that makes imputation more challenging.

Individual results of the different imputation methods are shown in Supplementary Table S2 and Supplementary Fig. S6. In general, the majority of programs obtained comparable imputation accuracies for each inversion and population, with the same inversions having similar high or low imputation accuracy across methods, and small variations resulting in inversions being imputable in some cases but not in others (Supplementary Fig. S7). Consistent with this, the different methods showed quite high pairwise correlations, with the average Pearson correlation coefficient (R) of the r² values of Beagle, IMPUTE2, IMPUTE5 and Minimac4 compared to the other four methods reaching 0.73, 0.74, 0.78, and 0.77, respectively, and IMPUTE5 and Minimac4 results being the most similar (R = 0.98) (Supplementary Fig. S8). The main exception was again scoreInvHap, which generated the most different imputation results with respect to the rest of methods (average R between r² values of 0.42) (Supplementary Fig. S7 and S8).

When considering the best method for the imputable inversions, three tools clearly stood out, with more inversions being imputed most reliably by IMPUTE2 in AFR, IMPUTE5 and Minimac4 in EAS, and Minimac4 in EUR (Supplementary Table S5), although differences in imputation accuracy across the three methods were generally small. Specifically, in each population, between 22.4% and 31.9% of the inversions were imputed with r² = 1 by at least one method (11 in AFR, 15 in EAS, and 16 in EUR), with IMPUTE2, IMPUTE5, or Minimac4 always imputing perfectly the highest number of inversions (Supplementary Table S5). However, IMPUTE2 results came at the cost of retaining less samples (Supplementary Table S2). Using instead a global reference panel, similar patterns were observed and improvements in imputation accuracy tended to be minor, but there were three extra inversions showing r² = 1 imputation accuracy in certain populations, including two in EAS that could not be imputed accurately before. Interestingly, two inversions (HsInv0290 and HsInv1111) could be imputed well with scoreInvHap in AFR, while the other methods performed quite poorly (Supplementary Fig. S6). These inversions have in common that they span a relatively large region including big IRs, although scoreInvHap results varied considerably among populations.

Factors affecting imputation accuracy

Besides the different programs, several factors can affect the imputation of individual inversions. In general, a good correlation was found between inversion imputation accuracy values in the three populations, although this correlation was slightly lower than across methods (Supplementary Fig. S9). As already mentioned, EAS was the population with worse imputation results due to the smaller number of genotyped samples (Supplementary Fig. S1). Moreover, the most important factor for imputation appears to be the inversion generation mechanism. For example, all NH inversions, which have a unique origin [11, 28], could be accurately imputed (usually with r² = 1) by most programs in the analyzed populations using the local or global reference panel (Supplementary Table S2 and Supplementary Fig. S6).

On the other hand, NAHR inversions, most of which have been previously identified as recurrent [11, 12], showed worse imputation results, with 46.7% (21/45) not accurately imputed in any or just one of the studied populations (Supplementary Table S2 and Supplementary Fig. S6). However, another 31.1% (14/45) could be imputed well in all the populations in which the inversion is present using at least one program, although imputation accuracies were much lower than for NH inversions (Supplementary Table S2 and Supplementary Fig. S6). Actually, the maximum LD with other variants in each population, which is a good approximation of how many chromosomes are affected by recurrence events [11, 12], was positively correlated with NAHR inversion imputation r² values for IMPUTE5 and Minimac4 (IMPUTE5: R = 0.504, n = 124, P = 2.46 × 10⁻⁹; Minimac4: R = 0.502, n = 125, P = 2.49 × 10⁻⁹). That was the case for HsInv0061, with a unique origin, and HsInv0209, with only a few recurrent chromosomes [11], which had high LD with SNPs and good imputation accuracy in all the populations where the inversion is present (Supplementary Fig. S6). In contrast, HsInv0344 or HsInv0608 have suffered several recurrence events that affect multiple chromosomes [11, 12], resulting in relatively low LD with SNPs, and they were not imputable in any population (Supplementary Fig. S6). In addition, some NAHR inversions were imputable in certain populations in which there are no or few recurrent chromosomes, but not in others that have higher levels of recurrence and lower LD with SNPs (e.g. HsInv0395 and HsInv0605 in EUR versus AFR) (Supplementary Table S1 and Supplementary Fig. S6) [11, 12], although other genetic exchange mechanisms like gene conversion could also play a role. One last factor with a significant effect on imputability was the variant frequency, especially for recurrent inversions, with worse imputation accuracy in populations with low versus high frequency (e.g. HsInv0030 in EUR versus AFR or HsInv0341 in AFR versus EUR). Nevertheless, imputation results likely depend also on other characteristics related to the genetic structure of the region, with some inversions showing good imputation accuracy despite having low LD levels, like HsInv0390 in EUR or HsInv0114, HsInv0266, HsInv0396, and HsInv0403 in AFR.

To better quantify the influence of several biological features on imputability of NAHR inversions, we fitted linear models of each program imputation results (except for scoreInvHap) using as variables the inversion MAF, density of SNPs in the region, type of chromosome (Chr. X versus autosomes), number of genotyped samples, and ratio between the IR and the inversion size (IR/inversion length), which has been previously associated with inversion recurrence [11, 12]. Without GP filtering, which allowed us to compare basically the same inversions, the different programs gave similar results, with variables explaining between 14.2% and 23.0% of the variance in imputation accuracy values (Supplementary Table S3). Filtering imputed genotypes with GP < 0.9 improved a little the variance in imputation accuracy explained by most models (15.2%–26.1%), especially for IMPUTE2 (Supplementary Table S3), probably due to removing some of the most complex inversions. Apart from this, the only variable that had a consistent significant effect was the number of genotyped samples, increasing imputation accuracy in all cases (Fig. 5). Other variables that have been associated with higher inversion recurrence levels [11, 12] showed consistent effects across models, but they were only significant in a few of them. For example, in most cases, IR/Inversion length ratio and Chr. X inversions were negatively correlated with imputation accuracy, although results were only significant for some of the IMPUTE2 models (Fig. 5 and Supplementary Table S3). Finally, MAF and the density of SNPs typically had small, nonsignificant positive effects on imputation.

Figure 5.

Results of linear models of NAHR inversion imputation using different programs. Graphs show the model coefficient estimates (X-axis) of the different variables in each program model (Y-axis), with and without removing unreliable imputed genotypes having GP values lower than 0.9 (labeled as GP ≥ 0.9 and no filter, respectively). Error bars represent the 95% confidence interval and color indicates the significance of the variable effects according to the P-value scale (right side). Dashed line marks the 0 value (no effect).

Imputation from SNP array data

As the last step, we checked how well inversions are imputed with SNP genotyping arrays and compared the results with those from WGS. As mentioned in the introduction, arrays interrogate fewer variants, which could significantly affect imputation. To test this, we simulated two types of commonly used SNP arrays with different density: Omni high-resolution array (∼2.4 million SNPs) and GSA low-resolution array (∼650 000 SNPs). Reliability of the genotypes obtained with the different imputation programs was largely proportional to the number of SNPs available, which reduced the total fraction of imputed inversion genotypes with GP ≥ 0.9 in the different programs to 77.5%–85.8% for the Omni and 66.6%–78.6% for the GSA arrays (Supplementary Fig. S4). As there are less variants in the reference panel, the total number of imputable inversions in the three populations also decreased to 42.6%–65.3% for Omni and 40.4%–59.5% for GSA (Fig. 6). In addition, median imputation accuracy with most programs was clearly lower for GSA compared to Omni or WGS data in EAS and AFR, although in EUR the differences were smaller (Fig. 6A). However, Omni and WGS tended to behave more similarly, with median imputation accuracy being equal or even higher for Omni in some populations and programs (Fig. 6A), suggesting a limited benefit of a greater density of SNPs. The main exception was Beagle, which, surprisingly, in the three populations showed higher imputation accuracy and number of imputed inversions with both arrays than with WGS data (Fig. 6).

Figure 6.

Inversion imputation results from SNP array data. (A) Median imputation accuracy (r²) across inversions from Omni and GSA array data compared to WGS using four imputation programs after GP ≥ 0.9 filtering in the three analyzed populations. To make the results comparable, within each population we included only inversions that retained samples with both orientations in all the analyses: AFR = 35; EAS = 28; EUR = 37. (B) Percentage of inversions that can be imputed by each program (r² ≥ 0.8 and ≥ 50% remaining samples) using the different types of array and WGS data, expressed with respect to the polymorphic inversions having at least two chromosomes with the minor orientation in the analyzed samples of that population. The total fraction of inversions imputable by any of the four programs is shown in the Combined category.

When considering the different populations, results were somewhat variable between the tested arrays and programs (excluding Beagle). As before, EUR showed the highest imputation accuracy and there was little variation between the two arrays and WGS, with high inversion imputability rates even in the smaller array (Fig. 6). In EAS, imputation results from the three types of data were also relatively similar, although they were much lower than in the other populations, as happened for WGS data (Fig. 6). Conversely, in AFR, imputation accuracy and the number of imputable inversions were more or less equivalent between WGS and the Omni array, excluding Minimac4, but GSA exhibited a considerably smaller amount of imputable inversions (Fig. 6), which suggests that the array SNPs do not properly capture African genetic diversity. Despite this variation, IMPUTE5 was overall the best performing method using array data, with IMPUTE2 closely behind, whereas Minimac4 appeared to be more hindered by the SNP sparsity, especially in EUR and AFR.

Discussion

Accurate genotypes are crucial to determine the functional and phenotypic effects of genetic variants. By performing for the first time an exhaustive comparison of the imputation of challenging inversions using different methods, we found that 46.8%–75.5% of the studied inversions in each population can be imputed with high accuracy from WGS data, despite our relatively strict (r² ≥ 0.8) imputability criteria. Therefore, this work paves the way to generate reliable information for a large fraction of inversions that goes beyond previous analyses based on a single imputation tool [11, 24, 25]. Nevertheless, there are still many inversions that cannot be accurately imputed with any program in either all or certain populations. In addition, we have found differences in imputation accuracy between populations, with worse imputation in AFR and especially EAS, which suggest that imputation success relies heavily on the haplotype diversity of the population in the affected regions and the number of available reference genotypes to properly capture this diversity.

Imputation algorithms differ in their approaches to modeling haplotype inheritance [19–23]. Inversions represent a special type of variant since they suppress recombination [10], which may disrupt these calculations, affecting their own imputation. Here, we have shown that IMPUTE5 and Minimac4 outperformed the other tools, with IMPUTE5 performing somewhat better in the array data, making them our recommended choices for inversion imputation. Although IMPUTE2 achieved comparable imputation results for many of the inversions (Fig. 4 and Supplementary Table S3), it does so at a significant cost in sample loss following GP filtering. This effect is particularly pronounced for low MAF inversions (Supplementary Table S2). Moreover, IMPUTE2 has much longer running times and requires more computing resources than the other tools, limiting its utility. In contrast, the poor results of scoreInvHap may stem from its direct reliance on LD between inversions and surrounding variants, which likely undermines its performance in this dataset. Nevertheless, there are a few large inversions that could be imputed only with scoreInvHap, as seen before for the 4.2 Mb 8p23.1 inversion [32]. This suggests that, for large inversions, the other methods may be negatively affected by positioning the inversion in the reference panel as its middle point, since they take into account the local variation around the variant, which could be quite distant from the actual breakpoints. Besides, there is no clear explanation for the relatively low inversion imputation of Beagle, although it could be related to the higher DNA sequence fragmentation of this method, which could compromise the imputation of larger SVs compared to SNPs. In any case, as illustrated by scoreInvHap, there is no single best imputation method across all inversions, and it might be necessary to use several of them to impute the maximum number of variants as accurately as possible.

Aside from testing the different programs, we also assessed several other alternative parameters that can affect imputation. In particular, GP filtering markedly improved imputation accuracy, while minimizing sample loss (Fig. 2). Since imputation is typically a preparatory step for functional analyses, excessive sample filtering can reduce statistical power or introduce bias toward specific haplotypes. The top-performing methods, IMPUTE5 and Minimac4, exhibited acceptable sample retention, with only 5.9% and 4% of samples lost, respectively, in WGS data after a GP ≥ 0.9 filtering, compared with a much higher 16.3% for IMPUTE2 (Supplementary Fig. S4A). Therefore, the 0.9 GP criterion seems to be a good compromise between imputation accuracy and sample size. On the other hand, using a global reference panel by combining multiple populations could help imputation in some specific inversions with low sample size (such as in EAS), but gains compared to a population specific panel tend to be small, and imputation accuracy can decrease in certain cases due to including recurrence events from different populations.

Among the biological factors examined, recurrence represents clearly the biggest challenge for imputation, since both inversion orientations can coexist in exactly or virtually identical haplotypes [11–13]. In fact, despite showing good imputation accuracy, in some NAHR inversions, genotypes filtered out due to low GP are likely caused by recurrent events (such as in HsInv0193, HsInv1117, or HsInv0286). Apart from this, as observed for EAS, the number of samples was the only variable significantly affecting imputation in all models (Fig. 5). Other factors also had some influence on imputation, although in most cases they were not significant (Fig. 5). For example, the density of SNPs and MAF usually had a positive effect, which might be related to a finer resolution of haplotypes and the need of a minimum amount of haplotypes of both orientations for the programs to be able to detect and impute them correctly. In the case of chromosome type, Chr. X inversions exhibited lower imputation accuracy across all methods, but the difference does not appear to stem from disparities in imputation performance between autosomal and sex chromosomes [17]. Instead, the explanation likely lies in NAHR-mediated inversion hotspots created by the elevated frequency of segmental duplications and the high inversion recurrence rate observed in Chr. X [11–13, 33]. Nevertheless, the ratio between the size of the inversion and the IRs, which is a good predictor of recurrence levels [11, 12], showed only a moderate negative effect on imputation in most methods. This is probably due to the opposite effects of inversion length, both hampering imputation and decreasing recurrence. Moreover, there are other factors that could affect imputation accuracy, like inversion evolutionary history, with older inversions accumulating more genetic differentiation and suffering greater LD erosion though recurrence and further recombination between chromosomes with the same orientation, or the number of individuals affected by each recurrence event. Thus, incorporating inversion age and better recurrence rate estimates in future analyses could help disentangle these confounded effects.

However, despite the inversion imputation improvements, there are some important limitations. First, for some inversions the number of available genotypes is quite low, which reduces the accuracy of the imputation, especially in certain populations. More importantly, a remarkable proportion of inversions are still not well imputed with any of the methods, mainly due to high recurrence levels. Resolving such complexity requires a larger reference panel to include enough representatives of all haplotype combinations. Current efforts to identify IR-mediated inversions using ultra-long reads are expected to generate more exhaustive genotype data for most human inversions, including many of those most challenging [8, 9, 34, 35], which will result in better imputation of a larger number of inversions. In addition, future methodological developments, such as the application of machine learning strategies, could also help impute recurrent events. Therefore, our results shed light on the capabilities and limitations of existing tools for handling SVs, setting the stage for further exploration of imputation strategies in complex genomic contexts. Such advances should ultimately allow us to investigate in detail the functional effects and phenotypic associations of the whole set of human genomic variation.

Supplementary data

Supplementary data is available at NAR Genomics & Bioinformatics online.

Conflict of interest

None declared.

Funding

This work was supported by research grants PID2019-107836RB-I00 and PID2022-137615OB-I00 to M.C. and Predoctoral Contract PRE-2020-092440 to I.Y., funded by the Agencia Estatal de Investigación of the Ministerio de Ciencia, Innovación y Universidades (MICIU/AEI/10.13039/501100011033, Spain) and the European Regional Development Fund (ERDF, EU), and grant 2021 SGR 00526 to support the activities of the research groups in Catalonia to M.C. and FI-STEP predoctoral grant 2025 STEP 00107 to A.M. from the Departament de Recerca i Universitats (Generalitat de Catalunya, Spain), co-funded by theEuropean Social Fund Plus (ESF+).

Data availability

All the data generated in this article are available in the article and in its online supplementary material. Inversion genotypes are available in dbVar (https://www.ncbi.nlm.nih.gov/dbvar/, study accession number nstd169, nstd185, and nstd255) and 1000 Genomes High Coverage data can be obtained from https://ftp.1000genomes.ebi.ac.uk/vol1/ftp/data_collections/1000G_2504_high_coverage/working/20220422_3202_phased_SNV_INDEL_SV/.