Introduction

Vast decreases in the cost of exome sequencing have allowed for major advancements in the identification of causal variants for rare monogenic traits and disorders.1, 2, 3, 4 Although each individual carries 20 000–24 000 single-nucleotide variants, most are common in the population and are unlikely to explain a rare monogenic trait. Variants that are too common to be consistent with the prevalence of a rare disorder are removed from consideration2 and the remaining variants are prioritized based on functional, structural and conservation properties.5, 6, 7 Recent prioritization approaches use cross-species comparisons8 or a combination of scores from several stand-alone methods for increased performance.9, 10, 11, 12 Although such techniques are very powerful when family data are available,2, 3, 4, 13, 14, 15, 16, 17 hundreds of variants often remain for follow-up validation when only a single case individual is sequenced.11, 18, 19

Variant filtering in exome-sequencing studies is usually performed using frequencies that are estimated across large databases of human variation either by ignoring ancestry, or by matching at the level of continental ancestry (eg, the Exome Variant Server (EVS)20)21 thus ignoring sub-continental ancestry. Although FST values calculated within continental populations are usually low (mostly due to the dependency of FST on allele frequency),22, 23 detectable population structure still exists.24, 25 Population genetic models predict that rare variants show greater clustering within continental populations than more common variants26 and empirical studies support this prediction.27, 28, 29, 30, 31, 32, 33 Therefore, a variant might appear rare (<1%) when its frequency is estimated across many populations, when in reality it is only rare in most populations and less rare or even common (>1%) in a one or more clustered sets of populations (see Supplementary Figure 1). For example, variant rs17046386 is common in Africans (therefore unlikely to be pathogenic) and generally rare or absent in non-Africans34, 35 (see Figure 1). However, this variant would not be discarded in the filtering step based on frequency estimates from European reference panels thus increasing the validation burden in the subsequent steps. In addition, the limited size of existing reference panels, especially when defining ancestry at the level of a country (often <100 individuals), induces significant statistical variance in allele frequency estimates that needs to be accounted for (eg, a variant with true frequency of 0.5% has 9.0% probability of being observed with a frequency >1% in a sample of 100 individuals and thus erroneously discarded).

Figure 1
figure 1

Geographic distribution of rs17046386 across the Human Genome Diversity Panel CEPH data. The minor allele is rare in non-African populations, but not rare in African populations.

Full size image

In this work we investigate the use of sub-continental allele frequencies (typically estimated at the level of a country36, 37) for a discrete frequency-based filtering step in exome-sequencing studies but such ideas can also be applied to general statistical methodologies aimed at finding causal genes in exome scans. We propose approaches that leverage the frequency estimates across all sub-continental populations in filtering while accounting for the statistical noise introduced by the smaller number of individuals used to estimate frequencies. We use simulations starting from the 1000 Genomes,28 the NHLBI GO Exome Sequencing Project Exome Variant Server (EVS)20 and the ClinVar38 data to show that our approach improves efficacy of filtering (eg, a reduction of 16% in the number of variants to be followed up in case-only simulations). Importantly, we show that the standard approach that ignores statistical noise in the allele frequency estimation is miscalibrated with respect to the false-negative rate (FNR) (ie, the probability of filtering out a true causal variant). Finally, we validate our approach using exome-sequencing data from 20 real individuals with monogenic disorders for which the true causal variants are known. Here we successfully reduce the number of variants to be functionally tested (a 38% reduction from 750 to 468 in the heterozygous case), while never discarding the known causal variants. Our results show that existing filtering pipelines for studies of monogenic traits can be significantly improved by incorporating ancestry while accounting for statistical noise in the filtering step. Interestingly, utilizing sub-continental population reference panels overcomes the reduction in performance due to higher statistical noise from the smaller panels.

Materials and Methods

Data sets

The 1000 Genomes Project28 has produced a public catalog of human genetic variation through sequencing from several populations: Han Chinese in Beijing (CHB), Japanese in Tokyo (JPT), Southern Han Chinese (CHS), Utah Residents with Northern and Western European ancestry (CEU), Toscani in Italia (TSI), Finnish in Finland (FIN), British in England and Scotland (GBR), Iberian population in Spain (IBS), Yoruba in Ibadan (YRI), Luhya in Webuye (LWK), Americans of African Ancestry in SW USA (ASW), Mexican ancestry from Los Angeles (MXL), Puerto Ricans from Puerto Rico (PUR) and Colombians from Medellin (CLM). We use the 1000 Genomes data (with the exception of IBS individuals, only 14 in total) to evaluate the effectiveness of various filtering approaches. As the vast majority of causal variants for monogenic traits are located in the exome,39 we restrict our analysis to coding regions of autosomal chromosomes. For admixed individuals we downloaded and used the 1000 Genomes Project local ancestry calls (the consensus calls from four inference methods40, 41, 42, 43). Damaging scores for each single-nucleotide variant were estimated using KGGSeq with default parameters11 that combines the functional scores from dbNSFP44 v2.0.

The EVS has released allele counts from 4300 European-Americans and 2203 African Americans20 along with PolyPhen2 scores for missense variants and we used those in our analyses. A set of 1395 pathogenic variants (as reported by multiple submitters) was obtained from the ClinVar database38 (accessed 4 December 2014).

To compare simulations to real data, we used exomes of 101 individuals with self-reported countries of origin including Turkey, Jordan, Tunisia, Egypt, Israel, Iran, Syria and Palestine. We grouped these individuals into a single supplemental population for estimating best matching allele frequencies. Of the 101 individuals, 9 were known to harbor heterozygous variants in genes causing autosomal dominant disorders, 10 had homozygous variants and 1 had compound heterozygous variants in a gene causing an autosomal recessive disorder (see Supplementary Information and Supplementary Tables S1 and S2).

FNR estimation

We estimate the probability of filtering out a true causal variant (FNR) at a given frequency threshold as a function of a reference panel and the maximum true allele frequency of the causal variant. The filtering threshold can be adjusted to provide a desired FNR. Let t be the nominal frequency threshold that is used for filtering. We define the corresponding FNR at this threshold as:

where f is the frequency of the variant in the population, max(fc) is the maximum assumed frequency of the causal variant in the population, P(f) is the proportion of variants with frequency f in the population and P(fref,N >t|f) is the probability that a variant with frequency f is observed at a frequency greater than t in the reference panel of N individuals randomly drawn from the population.

The FNR computation requires knowledge about the distribution of variants across all frequencies in the population; this can be estimated from population genetic theory under various demographic assumptions29, 45, 46, 47, 48, 49 or empirically from the data. In this work, we estimate the distribution P(f) from reference panel allele counts and perform the above integration across the observed site frequency spectrum as follows:

Here fi represents each of the unique allele frequencies observed in the reference panel of N individuals and P'(fi) represents the proportion of variants in the reference panel that have estimated frequency fi. P(fref,N>t|fi) is modeled as a binomial draw with the frequency of success equal to fi and the number of draws equal to the number of allele counts (2N). As the integration is over a discrete space we calculate the probability that the number of success is greater than the threshold times 2N. We propose to filter variants using the minimum frequency threshold t such that FNR(t)<0.05. If multiple populations are used in filtering (ie, removing variants that are common in any population, see below), we employ a Bonferroni correction for the threshold; that is, we require FNR(t) <0.05/K in each of the K considered populations.

Leveraging population structure for improved filtering

We simulate individuals with monogenic disorder by drawing two individuals from a specific 1000 Genomes population and then simulating an offspring assuming Mendelian inheritance and independence between SNPs. We compare three possible disease scenarios (Case-Only, Trio-Dominant and Trio-Recessive) using 40 simulated individuals per scenario and population. The Case-Only scenario assumes there is no information on parental genotypes. The Trio-Dominant scenario assumes that both parental exomes are sequenced and the offspring and one of the parents has the disorder. The Trio-Recessive scenario assumes that both parents are exome sequenced and heterozygous for the causal allele and that the offspring has two copies of the causal allele. Prior to frequency filtering, we remove all variants that do not result in an amino-acid change or do not create or remove a stop codon. In addition, we remove variants inconsistent with the disease scenario.

We consider multiple frequency-filtering approaches. The NoAncestry, f>1% and NoAncestry, FNR<5% approaches estimate allele frequencies and FNRs across all 1000 Genomes individuals. The key intuition here is that statistical noise is decreased with large reference panels but at the cost of ignoring population structure. NoAncestry, f>1% filters out variants with allele frequency >1% without regard for the FNR; NoAncestry, FNR<5% filters out variants above a threshold determined to ensure a 5% FNR. The PopMatched, FNR<5% approach uses only the reference individuals from the sub-continental population (country-level, see 1000 Genomes28) of the simulated individual. The AllPop, FNR <5% approach filters out variants observed in any population above a conservative Bonferroni-corrected FNR <5% threshold. We assume that populations are independent and set the desired FNR for each population to 0.05 divided by the number of populations used in filtering (eg, 14 for simulation results). MaxPopFreq filters variants if observed above 1% allele frequency in any 1000 Genomes continental population and is similar to a strategy implemented by ANNOVAR50 that filters variants if observed above 1% in any 1000 Genomes continental population or the EVS European or African-American populations.

For admixed populations we only simulated the Case-Only scenario by using the genotypes of real admixed individuals from 1000 Genomes as case individuals. In addition to methods above, we considered a method that utilizes local ancestry calls (PopMatched-LA, FNR<5%). In each individual at loci that are homozygous for African, European or Native American ancestry, we used the corresponding continental allele frequency estimates obtained by averaging across all 1000 Genomes individuals from a given continent. In local ancestry heterozygous regions we used a 50–50 weighting of the matching continental frequencies. We use the maximum continental FNR-based frequency threshold from the African, European and Asian continents as the filtering threshold.

Results

Modeling statistical uncertainty increases filtering efficacy

We use simulations from the European-American EVS20 data set to assess filtering based on an FNR as compared with the standard approach of ignoring statistical noise in the allele frequency estimates. We use simulations of various reference panel sizes created with binomial sampling from the frequencies estimated across all the European (or African-American) EVS data. As expected, the frequency threshold that maintains a 5% FNR increases as reference panel size decreases (see Figure 2a). As the maximum frequency of the true causal variant (max(fc)) decreases the number of variants for follow-up per individual also decreases thus increasing filtering performance (see Figure 2b). Overall, we find a diminishing return in performance for reference panels larger than 500 individuals.

Figure 2
figure 2

Reference panel size impacts the efficacy of filtering in exome sequencing in European simulations from the EVS data. We simulated reference panels at various sizes using a binomial sampling from the EVS frequencies. (a) Shows the threshold on the variant frequency needed to achieve a 5% FNR for various assumptions about the maximum frequency of the causal variant in the population (from 0.001 to 0.01). (b) Displays the number of variants that remain to be followed up post-filtering at a 5% FNR rate. As expected with larger reference panel sizes, the estimated frequencies from the reference panel become more accurate making the 5% FNR threshold converge to the maximum assumed frequency of the causal variant (fM), which in turn increases the efficacy of filtering. We observe limited gains in accuracy for reference panels over 500 individuals. Similar results are obtained for simulations of African Americans (see Supplementary Figure 2).

Full size image

Next, we investigated the FNR attained by the standard approach that ignores statistical noise and filters based on the mean frequency estimate. At small reference panels the standard approach is miscalibrated attaining an FNR close to 25% thus removing causal variants from consideration (see Table 1). In contrast, the approach that maintains an FNR<5% significantly increases the number of variants for follow-up from 298 to 724 on average; this is necessary as it reduces the FNR to the desired 5% (see Table 1). With large reference panels the frequency-based approach is conservative (FNR ~0%) yielding twice as many variants for follow-up than the FNR-based approach if the maximum causal frequency is 0.1%. Qualitatively similar results were observed for simulations from the EVS African-American data (see Supplementary Figure 2).

Table 1 Method comparisons for different reference panel sizes and maximum causal allele frequencies
Full size table

Leveraging ancestry to increase filtering performance

Next, we assessed the performance of filtering with or without accounting for the highly structured nature of rare variants.27, 28, 29, 51 Using simulated exome data we investigated the efficacy of filtering across a wide range of methods and sequencing studies. When comparing the methods that do not leverage ancestry, we observe that the NoAncestry, FNR <5% approach leads to a slightly increased number of variants that need to be functionally followed up over the NoAncestry, f>1% approach (Table 2). The increased number of variants is necessary to attain a correct 5% FNR rate (NoAncestry, f>1% attains an FNR of 6%). The MaxPopFreq approach yields the fewest number of variants by filtering at 1% frequency in any continental population, but has a 50% probability of filtering out the true causal variant (FNR=50%) (see Table 2).

Table 2 Average number of variants that remain for follow-up post-filtering in simulations of non-admixed individuals
Full size table

Among all methods that maintain an FNR <5% and use ancestry information, we observe that the method that incorporates data across all populations (AllPop, FNR<5%) attains the best performance across all simulated scenarios (an average 16% reduction across all populations from the NoAncestry, FNR<5% method in the Case-Only scenario, see Table 2). The improvement is likely due to variants common in at least one population that are filtered out as unlikely to be pathogenic. This benefit from assaying variants across many populations comes even at the expense of multiple testing correction (a Bonferroni adjustment is made to the FNR required in each individual population resulting in an average filtering threshold of 3.2%). This demonstrates that the benefit of better population matching outweighs the cost of higher statistical noise from the small reference panels. The greatest improvement from population matching comes with the African populations where there is a 26% decrease in the number of variants remaining for follow-up (see Table 3).

Table 3 Different levels of genetic diversity across populations induce a variation in the average number of variants remaining for follow-up in an individual
Full size table

We also investigated other types of clinical scenarios. As expected, the Trio-Dominant scenario has approximately half as many variants for follow-up as the Case-Only scenario (see Table 2). The Trio-Recessive scenario, simulated without inbreeding, shows less than six variants remaining for all scenarios and methods (see Table 2). Finally, we observe a similar pattern of improved performance when also filtering non-damaging variants as predicted by KGGSeq (Table 2). Therefore, improvements of ancestry-aware filtering do not come preferentially from variants with non-damaging predictions.

Ancestry-aware filtering in admixed individuals

We extend our approach to admixed individuals (eg, African Americans) with genetic ancestry from multiple continents. We incorporate the local ancestry structure in the filtering step with the PopMatched-LA, FNR <5% approach that matches reference panels according to the ancestry at each site in an individual’s genome. This significantly lowers the number of variants for follow-up in the admixed populations as compared with the standard local ancestry naive method (PopMatched, FNR <5%) (see Figure 3). For example, in African-American individuals we observe a reduction from 664 to 487 variants from just matching the local ancestry to continental populations as compared with using all 1000 Genomes data with an FNR <5%. When using information from all populations in the 1000 Genomes data set, there is improvement for all admixed populations over the method that ignores ancestry (NoAncestry, FNR<5%) (see Figure 3).

Figure 3
figure 3

Population matching using local ancestry information improves performance over local ancestry naive population matching in admixed populations. The PopMatched, FNR<5% approach performs poorly because the admixed reference panel sizes are much smaller than non-admixed reference panels leading to increased filtering thresholds. The AllPop, FNR<5% outperforms all other FNR-based approaches.

Full size image

Ancestry-aware filtering in ClinVar data

In the simulations above, we have made the assumption that the frequency distribution of causal variants is well approximated using the rare variants in our data, which may not hold in practice. To investigate deviations from this assumption, we filter the set of ClinVar pathogenic SNPs according to our methods. We find that the AllPop, FNR<5%, PopMatched, FNR<5% and NoAncestry, f>1% approaches filtered out 42 (3.0% FNR), 18 (1.3% FNR) and 38 (2.7% FNR) of the 1395 variants, respectively. This shows that all approaches are conservative with respect to FNR and suggests that by approximating the distribution of frequencies at causal variants from real data we do not artificially increase the FNR in empirical data.

Analysis of 20 exomes of individuals with monogenic traits

To examine the performance of the different filtering strategies when applied to actual data, we used the data from 20 of the 101 real exome sequenced individuals who had monogenic disorders where the causal variants have been previously identified. We assumed a maximum causal allele frequency of 1% for all cases because there was no prevalence data.2 For all modes of inheritance, the number of variants in an individual for follow-up after filtering was lower when filtering with the PopMatched, FNR <5% and AllPop, FNR <5% approaches as opposed to the NoAncestry, f>1% approach that does not account for the FNR (See Supplementary Table 1). We filtered out all variants except those with damaging annotations: splice acceptors, stop gains, frame shifts, stop losses, initiator codon changes, inframe insertions, inframe deletions, missense variants, splice region variants and KGGSeq predicted damaging variants. The 101 individuals form a supplemental population with the test individual held out. We included these individuals when estimating average frequencies across all populations in the 1000 Genomes Project for the real data. Using our AllPop, FNR <5% approach only 468 variants need to be followed up for dominant disorders as compared with 750 for the NoAncestry, f>1% approach (see Table 4). The true causal variant identified in these individuals was never filtered out. This demonstrates that using multiple population frequency estimates significantly reduces the number of variants remaining for follow-up analysis, while still maintaining an appropriate FNR. In Supplementary Table 1 we report the variants remaining in each individual along with country data, presumed inheritance pattern, the zygosity of the causal variant and disease. Removing outliers based on PCA from the 101 self-reported Middle Eastern individuals makes no significant difference in the number of variants remaining for follow-up (see Supplementary Figure S3 and Supplementary Table S3).

Table 4 Average number of variants that remain for follow-up post-filtering in real exome studies of 20 individuals with monogenic disorders
Full size table

Discussion

In this work, we introduce approaches that account for the finite sample size of the existing reference panels used in filtering while leveraging sub-continental ancestry to improve the filtering step in monogenic disease studies. Both the real data analysis of 20 exomes of individuals with known monogenic disorders and the simulations show that our approaches reduce the number of variants that need to be further investigated, thus increasing the effectiveness of identifying causal variants using exome sequencing of unrelated individuals. This work demonstrates that in a clinical setting, even a small reference panel of 100 individuals from a well-matched population can have significant impact on the filtering efficacy.

Our methods are limited in that they do not account for the cases where a second genetic or environmental factor is required for the phenotype to appear and this increases the risk of filtering the true causal if the second factor is rare in some populations. Errors in variant calling in reference populations may also falsely elevate the frequency of a true causal variant and so using multiple technologies for estimating allele frequencies would be a best practice. While our work is presented for use with exome-sequencing studies, its central idea will be extendable to whole genome sequencing, as rare variation both in and out of the exome will show population clustering.

The current bottleneck in using population structure to help identify rare variants is the limited size of the reference panels for narrowly defined sub-continental populations. Large databases such as the EVS could increase their impact if they could report sub-continental allele frequencies in addition to just European and African-American allele frequencies. Recent projects such as the UK10K52 will be extremely valuable as it is a large reference panel of a specific population. The ALFRED database53 will also be a very valuable resource for cross-population work with monogenic diseases when it can provide sequencing level data. Founder populations (eg, Amish or Iceland) where some non-causal variants are pulled to high frequency may be powerful in eliminating non-causal variants if the disease is rare or not present in the founder population.54, 55 Tools such as Kaviar56 will allow researchers to quickly search these emerging sources of population frequency data. Finally, a Bayesian approach to integrate cross-population prevalence, allele frequencies, annotation and functional data in a filter-free probabilistic manner is possible but left to explore in future work.