Validation of genotype imputation in Southeast Asian populations and the effect of single nucleotide polymorphism annotation on imputation outcome

We performed the first part of our analysis using data from PanSNPdb [16]. To illustrate the imputation accuracy in Southeast Asian populations, we selected all available samples from Indonesia (ID, n = 288), Malaysia (MY, n = 217), the Philippines (PI, n = 125), Singapore (SG, n = 90), and Thailand (TH, n = 245). Only SNPs that were polymorphic in all populations were used in this study (n = 52,160).

The second part was imputation accuracy and yield in a patient dataset in which we had access to phenotypes because the phenotypes allowed us to observe the effect of imputation on subsequent association tests. The subjects were 609 Thai dengue patients who were 1–15 years-old from Siriraj, Ramathibodi, and Khon Kaen hospitals. A total of 468,987 SNPs from Illumina Human Hap610 array (Illumina Inc., San Diego, CA) passed the quality control requirements (QC). The accuracy of imputation was tested for each SNP from the dengue dataset by first randomly choosing half of the SNPs from the genotyping panel to create a mutually exclusive set of SNPs: Set 1 and Set 2. Then, Set 1 SNPs were used to impute Set 2 to create a complete SNP panel. Set 2 was also used to impute Set 1 to create a complete SNP panel. Based on our results in the previous section, HMII data were used as a reference for imputing the SNPs from the dengue dataset. The total number of SNPs after imputation was 1,417,081. Post-imputation QC reduced these numbers to 858,480.

For all of the sample datasets, QC was performed in PLINK v1.07 [12] using standard procedures for GWAS [21]. We included all markers with a call rate > 0.95, a minor allele frequency (MAF) > 0.01, and a Hardy–Weinberg equilibrium (HWE) > 10^− 7. Samples with call rates < 0.95 were excluded from the analysis along with samples that had first-degree relationship agreement, as evaluated by expected IBD sharing in PLINK v1.07. Multidimensional scaling (MDS) of Southeast Asian populations from PanSNPdb was performed in PLINK v1.07. This method allowed for visualization of principle components in the admixed population. Plotting of the MDS was conducted in R version 3.0.2 (http://​www.​r-project.​org/​).

In the first part of the study, each population from PanSNPdb was analyzed independently. Five percent of SNPs were randomly selected and removed. The same SNPs set of the removed SNPs were applied to all populations. SHAPEIT version 2 software was used to pre-phase the SNPs [22]. Imputation was accomplished with IMPUTE2 to recover the removed SNPs [23]. Each population was phased and imputed using both references in turn. According to guidelines from IMPUTE2, we imputed each chromosome separately and used windows of 5 Mb with an additional 250 kb buffer region on both sides of the analysis interval. The options used in the program were -buffer 1000, −iter 30, −burnin 10, and -k 80. The processes for random removal, phasing, and imputation were repeated five times.

The second part of our study used all the autosomal SNPs from the dengue dataset. Half of the autosomal SNPs from Thai dengue patients were removed by every second SNP (Set 1). Another dataset (Set 2) was then created in which the other half of the SNPs (Set 1) were removed. In this way, all SNPs were imputed once. Imputation was performed with HMII as described above. After imputation, SNPs were filtered using a QC process similar to the initial filtering of raw genotypes. Post-imputation QC excluded SNPs with MAF < 0.01, call rate < 0.95, and HWE < 5 × 10^− 7. These datasets were used in the GWAS analyses. We then selected only the imputed SNPs from the two datasets and merged them into a single dataset in which all SNPs had been imputed. This dataset was used to compare imputation accuracies of SNPs according to their location relative to known genes.

References used for the imputation were downloaded prior to the imputation process from the Impute website (http://​mathgen.​stats.​ox.​ac.​uk/​impute/​impute.​html). The references were labeled on the website as International HapMap project phase II release #22 (HMII) and 1000G phase I. A total of 1,417,081 SNPs from 90 Chinese and Japanese samples were used from HMII with an additional 39,343,900 SNPs from 1092 worldwide sample populations in the combined reference from 1000G. PanSNPdb shared 47,870 SNPs with the HapMap reference and 51,849 with the 1000G reference. The Thai dengue dataset shared 493,846 SNPs with the HapMap reference and 565,912 with the 1000G reference.

IMPUTE2 gives each imputed genotype a posterior probability score (info score) between zero and one. A higher threshold cut off for the probability score will usually result in higher accuracy but a lower yield. In this study, the posterior probability threshold was set to 0.9 to gain results with a high confidence of accuracy [24]. Genotypes with posterior probabilities < 0.9 were set to missing. Yield was reported as the percentage of non-missing genotypes within the removed SNPs and accuracy as the percentage of imputed, non-missing genotypes that matched the original genotypes.

All SNPs from the dengue dataset were grouped based on their location in genes. Gene annotation was collected from Illumina sample sheets (Illumina Inc., San Diego, CA) and the NCBI database of genetic variation [25]. Targeted gene regions included coding, intergenic, intronic, and untranslated regions (UTR). Some SNPs mapped to more than one location and were marked as being in a complex region. There were 53,277 SNPs that were not in any of these 5 groups and were subsequently discarded from further analysis.

The difference between imputed and genotyped data in the dengue dataset was evaluated by looking at several properties of the SNPs. Differences in accuracy and yield for each gene region were measured by varying the posterior probability threshold from 0.5 to 1.0. MAFs and p-values from chi-square tests for each SNP were compared between imputed and genotyped datasets. Coefficients of determination (r²) were calculated for each comparison to estimate the concordance of imputed and genotyped SNPs. Pairwise LD, which is measured as r-squared, was calculated within a region of < 1 Mb around each SNP. The calculation was performed in PLINK v1.07 using the option --r2 with --ld-window-r2 0 and --ld-window-kb 1000. Average R-squared values for each region were calculated and plotted in Microsoft Excel. R-squared values were also plotted against distance and averaging over 1 kb bins using R software (http://​www.​r-project.​org/​).

The accuracy and variability of imputation for Southeast Asian samples were assessed with five populations downloaded from PanSNPdb and two publicly available references from the HapMap project and the 1000 genomes project. All populations had an average imputation accuracy of more than 92% (Fig. 1), regardless of which reference was used. Imputation with HMII as a reference gave an average accuracy of 96.57%, while for 1000G, the accuracy was 93.98% (Fig. 1a). The Thai (TH) population had the highest accuracy for both reference panels followed by Indonesia (ID), whereas the population from the Philippines (PI) had the lowest accuracy. The yield for each population was lower when imputation was performed with the HMII reference (average = 59.03%) compared to the 1000G reference (average = 68.44%) (Fig. 1b). The yields for all populations were similar when using the same reference. The only exception was the TH population imputed with the 1000G reference, which had a higher yield compared to the other populations.

Next, we looked at the results from each chromosome separately to demonstrate the variability of accuracies and yields (Fig. 2). Most results of imputation with HMII as a reference provided more than 95% accuracy (Fig. 2a). The 1000G reference provided a lower accuracy compared to HMII (Fig. 2b). The most striking result was that there was a change in standard deviation for imputation accuracy between the chromosomes. In particular, chromosomes 19 showed the lowest accuracy and higher variation in accuracy. Plotting the yield by population and chromosome showed no significant differences (Fig. 2c, d). However, chromosomes 19 also showed lower yield and chromosome 22 exhibited the highest level of variability. Because the imputation technique is, to a large extent, based on LD, we wanted to see if the higher variance could be correlated to any differences in the LD-pattern. We calculated the LD for each SNP pair with a distance between 10 kb and 1 MB. The number of these pairs that had an LD > 0.2 was recorded for each chromosome (Additional file 1: Figure S1). Chromosome 19 and 22 had the lowest values.

Trying to explain the differences in imputation accuracy, we investigated the population diversity of the five populations using MDS (Fig. 3). Whereas most samples were grouped together, all populations except TH showed large internal variation along the primary axis (C1) (Fig. 3a). The secondary axis (C2) mainly described the difference between 18 samples of the Thai Mlabri ethnic group (from Nan Province, Thailand) to the rest of the Thai samples. The third axis (C3) also mainly described the variation within Thai samples, whereas the fourth axis (C4) showed variation within PI, ID and SG (Fig. 3b).

Using the dengue GWAS dataset to evaluate the effects of SNP locations within a gene on the quality of imputation, we found 415,710 SNPs located within genes (Additional file 2: Table S1). As expected, intergenic regions contained the most SNPs, while complex regions, where SNPs have been associated with more than one gene, had the fewest. Varying the threshold settings in IMPUTE2 for accepting an imputed genotype showed that increasing the threshold led to an increase in accuracy but a decrease in yield (Additional file 1: Figure S2). Imputation results for SNPs in coding regions showed the highest yield and accuracy. Intronic and intergenic regions led to the second and third highest yield and accuracy, respectively. At a threshold of 0.5, all regions showed a similar yield. However, with an increasing threshold, the yield of coding regions increased compared to other locations. The opposite effect was observed for accuracy, and all locations approached the same level of accuracy when the threshold approached 1.

We further compared the measured MAF from the initial genotyped data to the MAF of the imputed data (Additional file 1: Figures S3-S4). Then, any imputed SNP that did not pass the quality-criteria (call rate > 0.95, MAF > 0.01, HWE > 10^− 7) were removed (Additional file 2: Table S1). The MAF of this reduced set of SNPs were similarly compared to the same set of SNPs from the initial genotyped data (Additional file 1: Figure S4). The correlation for each region between the MAF of the genotyped data and both the imputed data and the imputed and post-imputation filtered data were calculated (Table 1). Coding regions had the fewest SNPs failing the QC, while complex regions had the lowest correlation, followed by untranslated regions (UTR), before filtering of the imputed data. After removing low quality SNPs, all regions showed a high correlation in MAF between imputed and actual genotypes. Looking at the direction of change in MAF showed that SNPs with low initial MAF (MAF < 0.25) predominantly (> 80%) had even lower MAFs after imputation. SNPs with an initial MAF close to 0.5 had an equal distribution of SNPs that obtained higher and lower post imputation MAF. Imputation also appeared to systematically reduce the allele frequency of the initially low MAF for most imputed SNPs (Additional file 1: Figure S5). The same analysis was repeated using p-values from the binary association test instead of the allele frequency for each SNP (Additional file 1: Figures S6 and S7). All regions had a generally low correlation for the p-values before removing low quality SNPs (r < 0.4). After post-imputation QC, all regions showed an increase in correlation. Coding regions had the highest correlation both before (r² = 0.39) and after (r² = 0.81) removing low-quality SNPs. The lowest correlation for the binary association was found in complex regions before post-imputation QC (r² = 0.27) and UTR regions after the QC (r² = 0.76).

To test whether coding regions had a stronger LD compared to the other regions, PLINK v1.07 was used to calculate pairwise LD between each SNP and any other SNP within 1 Mb. LD values, presented as r-squared, are shown as averages for each region (Additional file 1: Figure S8) and as LD vs. distance plot (Additional file 1: Figure S9). SNPs within coding regions had the highest average LD to neighboring SNPs (r² = 0.253), followed by SNPs in intronic regions (r² = 0.232).