High-throughput pooled resequencing offers significant potential for whole genome population sequencing. However, its main drawback is the loss of haplotype information. In order to regain some of this information, we present LDx, a computational tool for estimating linkage disequilibrium (LD) from pooled resequencing data. LDx uses an approximate maximum likelihood approach to estimate LD (r2) between pairs of SNPs that can be observed within and among single reads. LDx also reports r2 estimates derived solely from observed genotype counts. We demonstrate that the LDx estimates are highly correlated with r2 estimated from individually resequenced strains. We discuss the performance of LDx using more stringent quality conditions and infer via simulation the degree to which performance can improve based on read depth. Finally we demonstrate two possible uses of LDx with real and simulated pooled resequencing data. First, we use LDx to infer genomewide patterns of decay of LD with physical distance in D. melanogaster population resequencing data. Second, we demonstrate that r2 estimates from LDx are capable of distinguishing alternative demographic models representing plausible demographic histories of D. melanogaster.
Citation: Feder AF, Petrov DA, Bergland AO (2012) LDx: Estimation of Linkage Disequilibrium from High-Throughput Pooled Resequencing Data. PLoS ONE 7(11): e48588. doi:10.1371/journal.pone.0048588
Editor: Rongling Wu, Pennsylvania State University, United States of America
Received: July 6, 2012; Accepted: October 3, 2012; Published: November 9, 2012
Copyright: © 2012 Feder et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: The research was supported by the National Institutes of Health (NIH) grants 1R01GM089926 and P50HG002568 to DAP and the NIH National Research Service Award fellowship (F32 GM097837-01) to AOB. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Linkage disequilibrium (LD) is a measure of the association between alleles at two loci encapsulating how often these alleles are observed together. LD is an important statistic because it reflects the historical rates of recombination between loci and thus forms the basis for many tests of selection  and the estimation of demography , . Measurement of LD fundamentally requires knowledge of multi-locus haplotype frequencies within a species and these frequencies have been traditionally obtained through direct observation of haplotypes or statistical inference of haplotypes from unphased genotype data , . While these approaches are feasible for single locus studies, they can become logistically and computationally difficult when applied genomewide.
Here, we present a simple and cost effective method to directly measure short-scale LD genomewide using pooled next-generation resequencing data without any prior knowledge of genotype frequencies or of the haplotypes present in the population. Pooled resequencing data is generated by anonymously mixing DNA from multiple individuals from a population or species followed by massively parallel sequencing. Pooled resequencing occurs naturally when sequencing intrinsically heterogeneous samples (e.g., tissue samples from one individual or microbe communities) and is becoming a common experimental technique for quantitative  and population genetic ,  analyses. Pooled resequencing is a highly accurate method to estimate SNP – frequencies and has also been used to estimate haplotype frequencies from pooled samples when haplotypes are known a priori .
While there is some debate concerning the use of pooled resequencing versus simply sequencing strains individually, both methods have merits in different situations and certain scenarios necessitate or benefit from the use of pooled sequencing (see Futschik and Schlotterer  and Cutler and Jensen  for extensive discussion). For instance, in some cases individual genomes cannot be isolated (e.g., tissue samples). In other circumstances, often encountered in evolutionary applications, sampling many individuals of a population is easy but sequencing them is labor-intensive or prohibitively expensive. Although pooled resequencing has proved useful in measuring allele frequencies to assess population differentiation  and summary statistics based on the site frequency spectrum,  researchers often forfeit estimates of linkage between polymorphic loci because of the limited haplotype information available in an experiment utilizing pooled resequencing.
We demonstrate that some of this haplotype information can be reclaimed on a short scale which nonetheless allows genomewide patterns of linkage to be observed. Our approach, called LDx, directly estimates LD from pooled samples by measuring two-locus haplotype frequencies across short sequence reads that tile any particular genomic region. We test the accuracy of our technique empirically by estimating r2, a common measure of LD, in a pooled sample of 92 wild type Drosophila melanogaster with individually sequenced genomes . We find that our technique accurately estimates r2 across the genome and that the correlation between the pooled and actual estimates of r2 is in the expected range given the sampling variance determined by the read depth of our samples. Finally, we show two applications of LDx: first, we demonstrate that estimates of r2 based on pooled samples show a classic signature of decay with physical distance and that the rate of decay is negatively correlated with recombination rate; second, we use LDx to investigate two alternative demographic histories of D. melanogaster. LDx is implemented as an open-source Perl script available via sourceforge (https://sourceforge.net/projects/ldx/).
Calculation of haplotype tables
To generate two-locus haplotype tables, LDx takes a list of sites that are polymorphic within the pooled sample and a file containing the positional mapping information of each read, specified in the SAM format . The position of polymorphic sites can be inferred directly from the pooled sequence data using a variety of techniques  or can be a list of polymorphisms known a priori. LDx then finds all reads that cover pairs of polymorphic sites whose distance apart is less than the maximum insert size of the sequencing library. As is shown in Figure 1, the count for each two locus haplotypes is computed, where xij is the number of genotypes observed with allele i at the first locus and allele j at the second locus. We refer to the number of reads that cover both polymorphic sites as the intersecting read depth. r2 is calculated between pairs of sites with intersecting read depth greater than a minimum threshold, by default ten. In the case of loci with more than two alleles, LDx takes the two most frequent alleles and reports r2 estimates with reference to those.
Figure 1. Cartoon depicting information leveraged from pooled paired end reads.
The cartoon represents an example observation between two loci. Although many reads hit one locus or the other, only five reads cross both loci. In this example, pA, computed only from intersecting reads, is 3/5, while pA′, computed from all available reads is 4/8.doi:10.1371/journal.pone.0048588.g001
Method 1 – direct inference
LDx reports the r2 value that would be calculated by naive observation of the haplotype table. That is, it is computed as
where pA and pB are the allele frequencies computed only from the intersecting reads:
Method 2 – approximate maximum likelihood
To estimate r2 using maximum likelihood, LDx uses the observed haplotype table and allele frequency estimates derived from all reads covering the two loci. We estimate allele frequencies pA′ and pB′ using total read depth rather than from the marginal allele frequencies calculated from the haplotype table because estimates made from all reads will be more accurate than estimates just made from intersecting reads.
For each pair of sites, we estimate r2 by computing the maximally likely r2 conditional on the observed allele frequencies. While the observed frequencies represent only an approximation to the true frequencies, they act as a useful proxy for the purpose of evaluating the likelihood of r2. The likelihood of the observed haplotype table conditional on r2 and the observed allele frequencies is,
where fij is the expected proportion of haplotype ij given r2 and is computed
and the allele frequencies are estimated as follows:
Using this approach, the most likely linkage disequilibrium estimate can only be computed for SNP pairs where the allele frequencies estimated across all reads are congruent with the haplotype table estimated from the intersecting reads. Because the intersecting reads are a subset of the total number of reads, such incongruent estimates are likely to occur when the true r2 is high, but not equal to one. When allele frequency estimates are incongruent with the haplotype table, the maximum likelihood is undefined and the reported maximum likelihood is at the boundary of the likelihood surface. LDx reports information on whether the r2 estimate for a particular pair of sites is likely to be undefined.
This method is labeled as approximate, because it assumes the observed allele frequencies as true, instead of simultaneously maximizing the probabilities of the observed allele frequencies and r2. Our implementations of the simultaneous three variable maximization frequently failed to converge. In scenarios in which our estimates did converge, the true MLE and approximate MLE yielded similar results (results not shown). We therefore report the approximate MLE r2 as a faster and more reliable proxy estimate.
Accounting for the experimental design
In pooled resequencing experiments, the binomial (multinomial) variance associated with esimates of allele (haplotype) frequency are a function of the number of chromosomes sampled and the number of reads at any locus (supplemental equation 3 in ). The variance of frequency estimates can be easily approximated by calculating the effective number of observations at a given locus, conditional on read depth and number of chromosomes in the sample, as
We use this formula to calculate the effective number of observations for each two-locus genotype when calculating the approximate maximum likelihood estimates of r2 above. LDx uses the effective number of observations to estimate the 95% confidence intervals surrounding the approximate MLE estimate. Confidence intervals are calculated as ±1.96 log-likelihood units away from the MLE (see the users guide).
To test the accuracy of r2 estimation from pooled resequencing, we used short read data described elsewhere ( SRA accession SRR353365.1). Briefly, this library is a pool of 92 highly inbred D. melanogaster strains derived from a natural population in Raleigh, North Carolina representing a subset of the 162 strain Drosophila Genetic Reference Panel (DGRP, ). Average autosomal coverage in this library is ~40× and average coverage of the X-chromosome is ~20×. Only reads with base quality scores >20 were used. We identified all biallelic SNPs in the DGRP population that are fixed within each strain (i.e., sites with no residual heterozygosity) using precomputed SNP tables (https://www.hgsc.bcm.edu/content/drosphila-genetic-reference-panel). Of those, we only considered sites in which the total read depth in the pooled sample was less than twice the chromosomal average in order to exclude potential copy number variants from the analysis. Our analysis also includes investigation of the accuracy of r2 estimates based on the number of intersecting reads and the observed minor allele frequency.
To test whether the observed correction between r2 estimated from pooled data and the DGRP is expected given binomial sampling, we generated simulated reads from the DGRP data. To generate simulated pooled paired end reads, we used wgsim (23). wgsim accepts a FASTA file listing full haplotypes from multiple individuals and simulates the pooling process as if sampling from a population composed of these individuals at user specified read depths, read lengths and gap sizes. We used wgsim to simulate a population composed of the 85 DGRP strains with 93 bp paired end reads at ~10×, 40×, 100× and 200× coverage. Note, we simulated a pooled population of 85 that are a perfect subset of the 92 strains used in the experimental pooled resequencing study; we were unable to simulate pooled resequencing for all 92 because 7 strains were not sequenced to sufficiently high coverage.
We generated estimates of r2 from these libraries as described above, with a minor allele frequency cutoff of 1%, and a minimum intersecting read depth of 10, except that for the simulated 10× library, in which we only required a minimum of 5 intersecting reads.
LDx represents, to our knowledge, the first effort to estimate levels of linkage disequilbrium from pooled resequencing data directly with no prior information of haplotype frequencies. The one existing method to infer haplotype frequencies and levels of LD from pooled data  requires prior knowledge of haplotype frequencies in the population. Obtaining prior knowledge of genomic haplotypes can be difficult, expensive and labor intensive. Moreover, the method presented in Long et al.  as well as analogous methods to phase di- and polyploid sequence data (e.g., , ) likely perform best when prior haplotypes are drawn directly from the population in question. This requirement limits the utility of these approaches. Through bypassing the haplotyping step, LDx can be applied to populations for which only pooled resequencing data exist.
Two-locus haplotype reconstruction
LDx recovers sufficient data from the pooled paired end resequencing data to make inferences of linkage disequilibrium through identifying SNP pairs with many intersecting reads. LDx is able to detect SNP pairs that fall both on a single read and across paired end reads, creating a bimodal distribution on the distances between two SNPs of an identified SNP pair (Figure 2A). As read depth increases in our simulations, we find that the proportion of SNP pairs where r2 can be estimated by the approximate maximum likelihood methods increases (Figure 2B).
Figure 2. Identification of SNP pairs.
A) The distance between component SNPs of a SNP pair are bimodally distributed, reflecting the frequency of pairs that fall within a single read or across paired end reads. B) Increasing the read depth increased the proportion of pairs it was possible to locate in the pooled paired-end read data with a 0.01 allele frequency cutoff. This proportion of estimable pairs is calculated by counting the number of SNPs in a moving window of length 300 bp and using that to compute the number of possible SNP pairings (n choose 2). This is then compared to the number of SNP pairs identified at a given read depth.doi:10.1371/journal.pone.0048588.g002
r2 estimates from pooled samples were highly correlated with estimates from the actual haplotype data (p-values for all correlation coefficients <<0.001, Figure 3AB). For the direct estimation method, we observed a small amount of upward bias in our observed estimates of r2 due to sparse sampling of the haplotype tables, leading to r2 estimates at 1. This upwards bias was not present in the method since estimates integrated both the allele frequencies and the observed haplotype tables. We observed a small amount of downward bias in our approximate MLE estimates, because incongruities between allele frequency estimates and observed haplotype frequencies caused r2 estimates of zero when only a subset of the haplotype table was sampled. The accuracy of r2 estimates by LDx increases with higher minor allele frequency (Figure 3D). r2 is more accurately estimated for these pairs because there is a high probability of observing all possible haplotypes.
Figure 3. Method performance of LDx in predicting linkage.
r2 measured from the DGRP haplotypes is strongly correlated with estimates from A) the direct observation method and B) the maximum likelihood method. In A), observing only a sparse sampling of the haplotypes creates the overabundance of observed r2 estimates of 1. We determined the correlation between our r2 estimates and r2 values derived from haplotype data provided by the DGRP (Mackay et al 2012). We restricted the DGRP dataset to those strains present within our sample (92 of 162 strains). C) Increasing the simulated read depth increased the correlation between the true r2 and the r2 estimated by the direct observation (red) and maximum likelihood (blue) methods. Estimates in these figures have minor allele frequency cutoff of 1%. D) Filtering based on minor allele frequency leads to more accurate r2 estimates for the direct observation (red) and maximum likelihood (blue) methods. Points represent r2 estimates made from pooled resequencing of the DGRP.doi:10.1371/journal.pone.0048588.g003
Dependence on read depth, read length and insert size
In simulations of different read depths, we found that increasing read depth leads to an increase in the correlation between DGRP r2 and r2 estimated by the direct estimation and approximate MLE methods (Figure 3C). The observed correlation estimate between the DGRP r2 and both direct estimation and aproximate MLE r2 from the NC92 data fell within the range of correlation estimates produced by our simulations. This serves as a validation of our simulation procedure.
Given these results, increasing read length (and keeping the number of reads constant) is expected to increase the accuracy of r2 estimates because read depth at any given locus will increase (results not shown). However, increasing insert size will generally decrease the accuracy of r2 estimates because the average intersecting read depth for any two SNP pairs will be lower. To see this, note that the variance of insert size scales proportionally to the average insert size. Thus, increasing the insert size will decrease the intersecting read depth particularly for pairs of SNPs that are at the average distance between the paired end reads.
Decay of LD with distance
To test that estimates of r2 made by LDx are biologically meaningful, we measured the decay of r2 with physical distance in our pooled resequencing data. LDx estimates of r2 show the classic pattern of decay with physical distance (Figure 4) and the rate of decay varies as a function of recombination rate in a pattern highly congruent with the decay rate of true r2 estimates (Table 1). In regions of low recombination, the rate of decay of LD is higher than in regions of high recombination. This is because at very short physical distance (e.g., less than approximately 100 bp), loci in regions of low recombination are highly linked (high r2) whereas loci in regions of high recombination are less tightly linked (lower r2). However, by ~300 bp, loci in regions of both low and high recombination have similar patterns of linkage.
Figure 4. LDx predictions decay at a biologically plausible rate.
r2 decays in a similar pattern among the direct estimation (red), maximum likelihood (blue) and DGRP (green) r2 measures. Points represent average r2 within distance classes. Averages were applied only to pairs that had minor allele frequency >0.1. Lines represent predicted decay or r2 with physical distance. Decay models were fit in R 2.13 (R core Development Team 2012).doi:10.1371/journal.pone.0048588.g004
Table 1. Comparison of the decay of r2 with distance and recombination rate as estimated by different methods.doi:10.1371/journal.pone.0048588.t001
Use of LDx in differentiating between demographic events
Estimates of the site frequency spectrum and their deviation from the expectation under neutrality can be useful for identifying demographic events . However, in some situations, alternative demographic events can result in populations with very similar levels of polymorphism. For instance, following a population bottleneck we expect a reduction in heterzygosity that is proportional to the the duration and the magnitude of the bottleneck. To see this, note that expected heterozygosity following a bottleneck can be computed as,
, where Ht is the post-bottleneck estimate of heterozygosity, H0 is the initial heterozygosity, t is the duration of the bottleneck and Nb is the size of the bottleneck population. Therefore, a population with a bottleneck half as severe but with a duration twice as long as some original population will have an identical estimate of heterozygosity, measured as π. However, the LD between sites in these two populations may not necessarily the same. In these situations, LDx can be used to distinguish these models.
We measured π using Variscan  and r2 in a forward-simulated population run in SFS_code  for an out of Africa bottleneck in D. melanogaster  (see figure 5). We then repeated the simulations in two additional simulated populations – one with a bottleneck twice as large, but lasting half as long (severe), and one with a bottleneck half as large but twice as long in duration (mild). The average r2/bp estimated both by the approximate MLE method and the direct computation are reported in table 2.
Figure 5. Reference Demographic Model.
Following Table 2 in Thornton & Andolfatto's out of Africa model  at ρ/θ = 7, the population reaches equilbrium at population size N0, contracts to a size of Nb, and then expands back to N0 after 4N0t generations. The population then continues another 4N0 (.048) generations before sampling. In our model, we used N0 = 1000 and sampled 20 individuals.doi:10.1371/journal.pone.0048588.g005
Table 2. Comparison of r2 values in population with bottlenecks producing similar average pairwise differences (π).doi:10.1371/journal.pone.0048588.t002
LDx estimated a significantly higher average r2/bp for both the approximate MLE and direct estimation r2 values for the severe model when compared to the original model (p-values 0.014 and 0.007, respectively). While LDx did not report a significantly lower r2/bp for the mild bottleneck model, it was significantly different from the severe model (p-values 0.0013 and 0.0012, respectively).
LDx represents, to our knowledge, the first effort to directly estimate levels of linkage disequilibrium from high-throughput pooled resequencing data with no prior knowledge of haplotype structure in the target population. It provides an accurate estimate of linkage over hundreds of basepairs genomewide, and suggests that important information on linkage can be retrieved from populations sequenced using pooled sequencing. Note, however, that our ability to estimate LD accurately between any two specific points is low even at reasonably high sequencing depths and even if they are physically close to each other, because the number of reads that overlap any two particular SNPs is much lower than the coverage at any one specific SNP (Fig. 2B).
Certain conditions make the extraction of useful LD information from pooled data very difficult. For example, if the read length of the pooled sequences is much shorter than the length at which linkage decays to background levels in the genome, LDx will not provide informative output concerning r2. Further, linkage cannot be calculated beyond the length of a read pair, as haplotyping is impossible with pooled data. Indeed, those researchers interested in identifying faint signals at long distances may have better success with individual strain haplotyping. Additionally, if genomic polymorphisms are very sparse, LDx will estimate linkage based on a small number of pairs. Such limitations make it unlikely that LDx or similar methods will useful for humans or other organisms with low levels of polymorphism per basepair.
Despite these limitations, we imagine estimates of r2 made by LDx will be useful in understanding how patterns of LD change genomewide due to selection and demography. For instance, strong bottlenecks are expected to dramatically increase pairwise LD genomewide and the average change in LD before and after a bottleneck could be used to estimate the severity of the bottleneck . As demonstrated above, certain disparate demographic effects will leave similar imprints in the site frequency spectrum. LDx offers the potential to differentiate these scenarios by detecting differences in linkage. LDx could also be useful for identifying previously unannotated paralogs as these regions should have aberrantly high estimates of LD.
As sequencing technology continues to improve, read depth and fragment length will increase. This will result in a higher accuracy of r2 estimation and an increase in the probability that r2 can be estimated between two SNPs. While these improvements will only marginally increase the accuracy of allele frequency estimation, they will dramatically increase the accuracy of LD estimation from pooled data.
Conceived and designed the experiments: AFF AOB DAP. Performed the experiments: AFF AOB. Analyzed the data: AFF AOB DAP. Contributed reagents/materials/analysis tools: AFF AOB. Wrote the paper: AFF AOB DAP.
- 1. Nielsen R (2005) Molecular signatures of natural selection. Annu Rev Genet 39: 197–218. doi: 10.1146/annurev.genet.39.073003.112420
- 2. Beerli P, Felsenstein J (1999) Maximum-Likelihood Estimation of Migration Rates and Effective Population Numbers in Two Populations Using a Coalescent Approach. Genetics 152: 763–773.
- 3. Pritchard JK, Stephens M, Donnelly P (2000) Inference of Population Structure Using Multilocus Genotype Data. Genetics 155: 945–959.
- 4. Stephens M, Smith NJ, Donnelly P (2001) A New Statistical Method for Haplotype Reconstruction from Population Data. Am J Hum Gen 68 (4) 978–989. doi: 10.1086/319501
- 5. Clark AG (1990) Inference of haplotypes from PCR-amplified samples of diploid populations. Mol Biol Evol 7: 111–122.
- 6. Magwene PM, Willis JH, Kelly JK (2011) The Statistics of Bulk Segregant Analysis Using Next Generation Sequencing. PLoS Comput Biol 7 (11) e1002255. doi: 10.1371/journal.pcbi.1002255
- 7. Futschik A, Schlötterer C (2010) Massively Parallel Sequencing of Pooled DNA Samples–The Next Generation of Molecular Markers. Genetics 186: 207–218. doi: 10.1534/genetics.110.114397
- 8. Kofler R, Orozco-terWengel P, De Maio N, Pandey RV, Nolte V, et al. (2011) PoPoolation: A Toolbox for Population Genetic Analysis of Next Generation Sequencing Data from Pooled Individuals. PLoS ONE 6 (1) e15925. doi: 10.1371/journal.pone.0015925
- 9. Shaw SH, Carrasquillo MM, Kashuk C, Puffenberger EG, Chakravarti A (1998) Allele frequency distributions in pooled DNA samples: applications to mapping complex disease genes. Genome Res 8 (2) 111–123.
- 10. Hajirasouliha I, Hormozdiari F, Sahinalp SC, Birol I (2008) Optimal pooling for genome re-sequencing with ultra-high-throughput short-read technologies. ISMB 24: i32–i40. doi: 10.1093/bioinformatics/btn173
- 11. Van Tassel CP, Smith TP, Matukumalli LK, Taylor JF, Schnabel RD, et al. (2008) SNP discovery and allele frequency estimation by deep sequencing of reduced representation libraries. Nature Methods 5 (3) 247–252. doi: 10.1038/nmeth.1185
- 12. Holt KE, Teo YY, Li H, Nair S, Dougan G, et al. (2009) Detecting SNPs and estimating allele frequencies in clonal bacterial populations by sequencing pooled DNA. Bioinformatics 25 (16) 2074–2075. doi: 10.1093/bioinformatics/btp344
- 13. Out AA, van Minderhout IJ, Geoman JJ, Ariyurek Y, Ossowski S, et al. (2009) Deep sequencing to reveal new variants in pooled DNA samples. Hum Mutat 30 (12) 1703–1712. doi: 10.1002/humu.21122
- 14. Bansal V (2010) A statistical method for the detection of variants from next-generation re-sequencing of DNA pools. ISMB 26: i318–i324. doi: 10.1093/bioinformatics/btq214
- 15. Amaral AJ, Ferretti L, Megens HJ, Crooijmans RPMA, Nie H, et al. (2011) Genome-wide footprints of pig domestication and selection revealed through massive parallel sequencing of pooled DNA. PLoS ONE 6 (4) e14782. doi: 10.1371/journal.pone.0014782
- 16. Zhu Y, Bergland AO, Gonzalez-Perez J, Petrov DA (2012) Empirical validation of pooled whole genome population re-sequencing in Drosophila melanogaster.. PLoS One Iin press. doi: 10.1371/journal.pone.0041901
- 17. Long Q, Jeffares DC, Zhang Q, Ye K, Nizhynska V, et al. (2011) PoolHap: Inferring Haplotype Frequencies from Pooled Samples by Next Generation Sequencing. PLoS ONE 6 (1) e15292. doi: 10.1371/journal.pone.0015292
- 18. Futschik A, Schlotterer C (2010) The next generation of molecular markers from massively parallel sequencing of pooled DNA samples. Genetics 186: 207–218. doi: 10.1534/genetics.110.114397
- 19. Cutler DJ, Jensen JD (2010) To pool, or not to pool? Genetics 186: 41–43. doi: 10.1534/genetics.110.121012
- 20. Magwene PM, Willis JH, Kelly JK (2011) The Statistics of Bulk Segregant Analysis Using Next Generation Sequencing. PLoS Comput Biol 7 (11) e1002255. doi: 10.1371/journal.pcbi.1002255
- 21. Kofler R, Orozco-terWengel P, De Maio N, Pandey RV, Nolte V, et al. (2011) PoPoolation: A Toolbox for Population Genetic Analysis of Next Generation Sequencing Data from Pooled Individuals. PLoS ONE 6 (1) e15925. doi: 10.1371/journal.pone.0015925
- 22. Mackay TFC, Richards S, Stone EA, Barbadilla A, Ayroles JF, et al. (2011) The Drosophila melanogaster Genetic Reference Panel. Nature 482: 173–178. doi: 10.1038/nature10811
- 23. Li H, Handsaker B, Wysoker A, Fennell T, Ruan J, et al. (2009) The Sequence alignment/map (SAM) format and SAMtools. Bioinformatics 25: 2078–2079. doi: 10.1093/bioinformatics/btp352
- 24. Bansal V, Harismendy O, Tewhey R, Murray SS, Schork NJ, et al. (2010) Accurate detection and genotyping of SNPs utilizing population sequencing data. Genome Res 20 (4) 537–545. doi: 10.1101/gr.100040.109
- 25. Kolaczkowski B, Kern AD, Holloway AK, Begun DJ (2011) Genomic differentiation between temperate and tropical Australian populations of Drosophila melanogaster. Genetics 187 (1) 245–260. doi: 10.1534/genetics.110.123059
- 26. Su SY, White J, Balding DJ, Coin LJM (2008) Inference of haplotypic phase and missing genotypes in polyploidy organisms and variable copy number genomic regions. BMC Bioinformatics 9: 513. doi: 10.1186/1471-2105-9-513
- 27. Tajima F (1989) Statistical methods for testing the neutral mutation hypothesis by DNA polymorphism. Genetics 123: 585–595.
- 28. Gillespie JH (2004) Population Genetics: A Concise Guide, 2nd ed. Baltimore, MD: The Johns Hopkins University Press.
- 29. Vilella AJ, Blanco-Garcia A, Hutter S, Rozas J (2005) VariScan: Analysis of evolutionary patterns from large-scale DNA sequence polymorphism data. Bioinformatics 21: 2791–2793. doi: 10.1093/bioinformatics/bti403
- 30. Hernandez RD (2008) A flexible forward simulator for populations subject to selection and demography. Bioinformatics 24 (23) 2786–7. doi: 10.1093/bioinformatics/btn522
- 31. Thornton KR, Andolfatto P (2006) Approximate Bayesian Inference Reveals Evidence for a Recent, Severe Bottleneck in a Netherlands Population of Drosophila melanogaster.. Genetics 172: 1607–1619. doi: 10.1534/genetics.105.048223
- 32. Itoh M, Nanba N, Hasegawa M, Inomata N, Kondo R, et al. (2010) Seasonal changes in the long-distance linkage disequilibrium in Drosophila melanogaster. Journal of Heredity 101: 26–32. doi: 10.1093/jhered/esp079
- 33. Fiston-Lavier AS, Singh ND, Lipatov M, Petrov DA (2010) Drosophila melanogaster recombination rate calculator. Gene 463: 18–20. doi: 10.1016/j.gene.2010.04.015