A panel of 56 B. anthracis strains from the Biological Defense Research Directorate's strain collection (see Additional data file 1) was resequenced using Affymetrix resequencing arrays (RAs) and base calls determined using the ABACUS software package 32. Each RA was capable of resequencing 29,212 base-pairs (bp) or about 0.5% of the B. anthracis genome from a single isolate sample (see Additional data file 2). Long PCR sample preparation and chip processing was conducted for 118 RAs. Analysis of these 118 RAs with the ABACUS software package shows that 115 are successful (97.5%). Experimental failure occurs when less than 60% of the total possible bases fail to achieve quality scores exceeding the ABACUS user-defined threshold. For this study, the total threshold was set at 31 and a strand minimum of -2 32, as determined from analysis of a replication experiment described below.

The 115 successful RAs call 92.6% of the possible bases (3,109,539 bp out of a total possible of 3,359,380 bp). Figure 1 shows the distribution of quality scores across all 3,359,380 base calls. Amplicon failure, typically arising from long PCR (LPCR) failure, accounts for 1.1% of the uncalled bases. The remaining base-calling failure (6.3%) consists of features on the RAs that fail to generate quality scores exceeding the experimental threshold.

Previous results demonstrated that base-calling failure was concentrated among RA oligonucleotide probes containing multiple purines. Purine-rich probes were observed to have lower hybridization intensities at identical positions across multiple RAs. Guanine-rich probes, in particular, showed the greatest reduction in hybridization intensity (see Figure 6 in 32). Consequently, total quality scores at these sites frequently failed to exceed the quality-score threshold and they remained uncalled. To determine if probe sequence composition, specifically purine and guanine content, contributed to the 6.3% of bases not called, the sequence composition of the purine-rich oligonucleotide probes at 4,209 sites successfully called on all 115 RAs (484,035 total sites) was compared to that at the 886 sites that failed to be called on any RA (101,890 total sites). These failed sites account for 3.0% of the total base calling failure in the experiment. Uncalled sites are composed of oligonucleotide probes with a significantly higher purine composition (P < 10^-22). A similar pattern is detected if we limit our analysis to guanine-rich probes (P < 10^-9). This latter result is surprising given that the B. anthracis genomic sequences we examined have a low G+C content (~34%). Nevertheless, these analyses demonstrate that both purine-rich and guanine-rich oligonucleotide probes are significantly more likely to fail to generate quality scores exceeding the experimental threshold.

Building on the recognition of the importance of automated algorithms to assess data quality 4243, we used two methods to assess the quality of microarray resequencing data 32. The first consisted of a replicate experiment where 51 samples were independently hybridized on 102 RAs. A parameter search that optimized the percentage of called bases, while minimizing the number of discrepancies between replicates was then performed. A total of 1,489,812 bases could have been called in each replicate experiment. At the optimal parameter values (total threshold of 31, strand minimum of -2 see Cutler et al. 32), 90.6% (1,349,178) of sites are called in both replicates. Other parameter values provide similar levels of base calling and discrepancy rates. The optimal parameter values are similar to those previously used by Cutler et al. 32. Of the bases called in both replicates, 1,349,177 are called identically. Only one site is called differently. This corresponds to a replication discrepancy rate of 7.4 × 10^-7 (Table 1). If repeatability could be related to accuracy, then this level of repeatability would correspond to a phred score of at least 61 4243. This calculation assumes that the discrepancy rate corresponds to a binomial error probability of P, where phred = -10 log₁₀P. These replication levels and discrepancy rates are consistent with those previously reported 32, providing further evidence for the ability of RAs analyzed with ABACUS to produce highly replicable data.

While RA data is highly replicable, repeated systematic errors would not be detected in a replicate experiment. To obtain an independent estimate of RA sequence accuracy, we compared the sequence data from 30 RAs where the same B. anthracis strain had been sequenced using the random shotgun approach and deposited in GenBank (B. anthracis: strain Ames, NC_003997 8, Vollum, NZ_AAEP00000000, 4 June 2004 update, strain Australia 94, NZ_AAES00000000, 7 June 2004 update, strain Kruger B NZ_AAEQ00000000, 7 June 2004 update (J Ravel, DA Rasko, MF Shumway, L Jiang, RZ Cer, NB Federova, M Wilson, S Stanley, S Decker, TD Read, et al., unpublished work). In a comparison of 398,467 bp of RA- and shotgun-generated sequence, we observed 15 discrepancies occurring at six sites. This corresponds to a discrepancy rate of 3.8 × 10^-5. If we make the conservative assumption that all discrepancies lie in the RA-generated sequence, this level of accuracy would correspond to a phred score of at least 44.

To determine if this conservative assumption is warranted, we examined in greater detail the nature of the RA/shotgun sequence discrepancies. Five of the discrepant sites, accounting for 10 discrepancies total (twofold RA replication at each site), were found in Kruger B strain sequences. The one remaining site, accounting for five discrepancies (fivefold RA replication at this site), was found in Vollum strain sequences. At all 15 discrepancies, the RA called a base identical to the Ames reference sequence 8, while the Kruger B/Vollum shotgun sequence called a new SNP. The fact that the shotgun sequence called a SNP at every discrepancy was surprising, leading us to examine more closely the level of shotgun coverage and assembly at each discrepant site. A comparison of the latest shotgun assembly of the Kruger B strain (J Ravel, et al., unpublished work) with the RA Kruger B strain base calls agreed with the RA base calls. The latest Vollum shotgun assembly (J Ravel, et al., unpublished work) still disagreed at the one site (five discrepancies total), but this discrepancy was based on a single shotgun sequencing read with a phred score of 7 at the discrepant base. Clearly, the shotgun coverage lacks sufficient depth at this site to make a reliable base call and it seems far more likely that the fivefold RA base call is correct. Hence, the RA sequence data has less than one discrepancy per 398,467 bases called, or a discrepancy rate of < 2.5 × 10^-6 (Table 1). This observed level of sequencing accuracy corresponds to a phred score of 56. These data demonstrate that our conservative assumption is not warranted. Resequencing array data quality from a single experiment matches, and in some cases perhaps exceeds, that obtained by multiple DNA sequencing reads using conventional DNA sequencing technologies 4243.

We identify 37 SNPs among 56 B. anthracis strains. The SNP location, base-call, and position relative to the respective GenBank reference sequences 678 are contained in Additional data file 3. Twenty-four of the 37 SNPs, including two singletons, were independently confirmed in identical strains where whole-genome random shotgun sequence was available (A0039, A4088 and A0442 in Additional data file 1 (J Ravel, et al., unpublished work)). Of the remaining 13 SNPs not independently verified by The Institute of Genomic Research (TIGR), 11 were seen only once in our collection of strains and two SNPs were seen three times.

Population genetic inference typically assumes that study samples are selected without prior knowledge of their patterns of genetic variation. For this study, we selected diverse strains from widely distant geographic regions in an attempt to sample the full extent of genetic variation in B. anthracis. The number of SNPs identified, the amount of sequence generated and the nucleotide diversity 44 of the 56 strains is contained in Table 2. We performed analyses for sequences comprising the total dataset, for each genomic region separately, and for the total dataset with each resequenced base assigned into an annotated SNP class. We report three main findings. First, the total average level of DNA sequence variation in B. anthracis is very low. This finding is in agreement with previous studies 1128. This level of genetic variation is much lower than that seen in commonly studied bacterial species 14, roughly half of that observed in the human genome and 25-fold lower than that observed in D. melanogaster 383945464748. Second, the B. anthracis chromosome appears less variable than either the pXO1 or pXO2 plasmids, although this difference is not statistically significant. Third, the patterns of genetic variation by SNP class (see Table 2 and Additional data file 4) are similar to that seen in other well studied bacterial 14 and eukaryotic genomes 45. Silent sites, those sites that when mutated do not alter the protein primary structure, are significantly more variable than are amino acid altering replacement sites (P = 0.0011). Intergenic regions are observed to have intermediate levels of genetic variation, whereas replacement sites, those sites that when mutated alter the protein primary structure, are the least variable. Replacement sites are marginally significantly less variable than intergenic sites (P = 0.039) whereas silent sites are not significantly more variable than intergenic sites (P = 0.22).

The neutral theory of molecular evolution predicts a characteristic frequency spectrum of SNPs, or segregating sites, for populations at equilibrium 49. Deviations from this expected distribution are observed when an experimental population sample contains an excess of low frequency, rare SNPs, or an excess of high frequency, common SNPs, relative to the neutral expectation. These deviations can arise as a consequence of demographic history and/or the action of natural selection 50. Figure 2 compares the observed and expected percent of SNPs in four allele-frequency classes. The data suggest an observed excess of rare SNPs as compared to that expected under the neutral theory. For example, while the neutral theory predicts that approximately 60% of SNPs should have minor allele frequencies less than or equal to 0.25, we observe that more that 92% of the B. anthracis SNPs we discovered have minor allele frequencies that fall into this class, a statistically significant difference (Figure 2).

We used the Tajima's D statistic 50 to further assess this pattern for the entire dataset, for SNPs from each genomic region and for each SNP class (Table 2). Tajima's D is a summary statistic for the site (or SNP) frequency spectrum, whose value is negative when there is an excess of rare variants and positive when there is an excess of common variants, relative to the neutral expectation. The test statistic is calculated from two different estimates of levels of genetic variation, the number of segregating sites 44 and the average number of nucleotide differences estimated from pairwise comparisons 50. We observe that Tajima's D is negative for SNPs comprising the total dataset, each genomic region and each SNP class. While none of the individual test statistics is statistically significant, they collectively suggest an excess of rare variants in B. anthracis. If we scale our variation estimates drawn from the 0.5% resequenced in 56 B. anthracis genomes, we can estimate a range around the total number of SNPs that one would detect upon sequencing two random B. anthracis isolates, sampled in the same fashion as isolates in this study were chosen. Our results indicate that we should expect to find, on average, between 944 (standard deviation (SD) 454) 50 and 1,586 SNPs (SD 762) 44. A substantial proportion of these SNPs, probably more than expected under the neutral theory, would be rare.

Using multiple sequence alignments of 17 genes from B. anthracis (NC_003997, Ames) and B. cereus (NC_004722, ATCC 14579 9 and NC_003909, ATCC 10987 10) the patterns of genetic polymorphism and divergence at silent and replacement sites was assessed. The raw counts are presented in Table 3. It is striking that two B. cereus strains exhibit more polymorphism at silent and replacement sites than divergence from B. anthracis. This result confirms, at the DNA sequence level, previous results suggesting that the B. cereus species group is diverse and polyphyletic in origin. B. anthracis then appears to be a clonal lineage derived from, and nested within, a diverse species. In other words, the species names do not encompass or reflect the evolutionary history of the species 105152.

The 37 SNPs discovered on the B. anthracis chromosome and plasmids pXO1 and pXO2, possess in total, 636 pairs of sites where two alleles are observed. In principle, the alleles at each pair of sites could form four distinct haplotypes. Plasmid transfer between different B. anthracis strains would affect physically unlinked site pairs resulting in four distinct haplotypes. Homologous recombination or gene conversion between physically linked site pairs is also expected to produce all four haplotypes. The straightforward counting of the number of haplotypes that one detects in a large population sample, such as the one used in this study, is often referred to as the four-gamete test 53.

Among the 636 site pairs in our sample, we observe 26 pairs of sites with two haplotypes, 610 pairs of sites with three haplotypes, and no pairs of sites with four haplotypes. This striking result implies that the value of D', the standardized measure of linkage disequilibrium (LD) 54, is equal to 1, its maximum value, for all site pairs that we observe. Among the 137 site pairs where we could have detected statistically significant LD at P < 10^-3, we observe that 52 site pairs exhibit statistically significant LD. Four of the six site pairs showing significant LD on the B. anthracis main chromosome are over 500 kb apart.

Because of the low level of genetic variation in B. anthracis (2829 and this study), determining the phylogenetic relationship among B. anthracis strains has proven difficult. Twenty-four B. anthracis strains characterized with a single fluorescent AFLP primer combination were reported to be monomorphic 27. One recent MLST study sequenced seven housekeeping genes (approximately 3 kb total) in 5 B. anthracis strains and reported that the strains were monomorphic at the sites examined. Another recent MLST study sequenced seven genes (approximately 3 kb total) in 11 diverse B. anthracis strains finding three polymorphic nucleotides 55. Neither the AFLP nor the MLST studies discover and genotype sufficient genetic variation to distinguish between B. anthracis strains.

The most successful marker-based approach used to date, MLVA, determined the genotypes at eight VNTR loci in 426 B. anthracis isolates, enabling the construction of a phylogenetic tree of B. anthracis strains 29. We sought to determine if our resequencing of 0.5% of each of 56 B. anthracis genomes is capable of confirming the major phylogenetic groupings determined by MLVA. To test this, we concatenated the 37 variant positions for all strains in this study, calculated a distance matrix using a simple Kimura substitution model, and generated an Unweighted Pair Group Method Arithmetic Mean (UPGMA) tree (see methods 56; Figure 3). The strains group together in a manner broadly similar to that found by Keim et al. 29 with B strains forming an outgroup and most A strains being found together in the same subgroups (Figure 3). There are exceptions: one group in Figure 3 contains a mix of A3a, A1a, A1b and A2 strains. This anomaly is probably due to the relatively few SNPs that effectively distinguish these groups when only 0.5% of the genome is sampled. All B. anthracis Ames strains but ASC394 correctly cluster in an A3b group. B. anthracis ASC394 may be a case of an originally mistyped or mislabeled strain. Nevertheless, our data suggest that limited, random resequencing of 0.5% of the 56 B. anthracis genomes discovers and genotypes sufficient genetic variation to determine the major phylogenetic relationships among B. anthracis strains.

We selected a geographically diverse panel of 56 B. anthracis strains from the Biological Defense Research Directorate collection (see Additional data file 1). Twenty-four of the strains originated from the Louisiana State University collection 29. These have been typed by MLVA 29 and in order to sample diversity, we chose a group that had representatives of the A1a, A1b, A2, A3a, A3b, A3d, A4, B1 and B2 lineages. The remaining 35 strains originate from a UK collection and were chosen to represent geographical variation as well as unusual phenotypes such as gamma phage and penicillin resistance. Six of the UK strains were reisolates of the Ames strain 11, which allowed us to test the reproducibility of resequencing.

Unique genomic sequences were identified using Miropeats 68 at the default thresholds from among the B. anthracis Ames chromosome (5.2 megabase-pair (Mb), NC_003997) and plasmids pXO1 (181.6 kb, NC_001496) and pXO2 (96.2 kb, NC_002146). The genomic regions that we resequenced included at least one gene of interest (pXO1: toxin lethal factor precursor lef, toxin moiety, protective antigen pagA; pXO2: encapsulation protein gene CapC; Ames chromosome: vrrA, DNA-directed RNA polymerase rpoB, yfhp protein), but also included many surrounding loci (see Additional data file 4 for complete listing). The total chip design consisted of 6,191 bp from pXO1, 6,725 bp from pXO2, and 16,584 bp from the Ames chromosome (total submitted sequence 29,500 bp). From these unique sequences, a single 20 × 25 μm RA design capable of resequencing 29,212 bp or 0.5% of the B. anthracis genome was fabricated by Affymetrix (see Additional data file 3). The final sequences submitted for RA design are contained in Additional data file 5.

Five milliliters of brain heart infusion (BHI) was inoculated and grown 12-16 h at 37°C. One-ml aliquots of cells were centrifuged for 10 min at 5,000-7,500g. Pellets were resuspended in 720 μl enzymatic lysis buffer (20 mM Tris-Cl pH 8.0, 2 mM EDTA, 1.2% Triton X-100, 20 mg/ml lysozyme) and incubated at 37°C for 1 h. After incubation 100 ml of Proteinase K was added along with 800 ml of Qiagen buffer AL, and incubated at 70°C for an additional 30 min. Then, 800 ml of 100% ethanol was added and this was split onto four of the Qiagen DNAeasy tissue kit. The DNA was then washed and eluted according to the Qiagen protocol. After the DNA was eluted, it was passed through a 0.22 mm filter. Sterility was confirmed by plating 10% DNA preparation directly on SBA plates with a second 10% inoculated into a 5 ml broth culture. The plate and the broth were allowed to incubate for 7 days. Two hundred milliliters of the broth culture was subcultured onto SBA at day 4. If there was no growth on any of these cultures the DNA was considered sterile and removed from the BSL-3 lab for subsequence analyses.

Genomic DNA was amplified using Long PCR (LPCR) protocols described in Cutler et al. 32. The primers that amplified each RA fragment are shown in Additional data file 3. The primer sequences were:

ant8 AAAAAGACGAGATGCGTCAACATCCCGTCCCA,

ant9 TCAACTAAATCCGCACCTAGGGTTGCTGTAAG,

ant10 ATTACTTTGAGTGGTCCCGTCTTTATCCCCCT,

ant11 ACATTAGCAGGCAAGGACAGTGGTGTTGGAGA,

ant14 ATTCACGCTCTCCCACCCAGATATTCCTACAT,

ant15 GTCCTAATATCGGTGAGCAACGCAGGGTAGTT,

ant20 GAGAAGAACCCCTACTACACGCATTGATACTG,

ant21 TTTAGTAGCGAGGGTACAGGCGCGTTTATACC,

ant26 TGGAAGCAGGCTTCGTAAGTGTAGGCGACGTT,

ant27 GTTGCATGTTCGCTCCCATAAGTGCGCGGTTA,

ant 32 AATGGGTGTATAGGGGTGATCTGTTGTGATGG,

ant33 TCCATGTTCGGCCATCTGATTCCGTCACTACT.

Long PCR product concentration was determined by using Pico-Green (Molecular Probes, Inc.) with lambda DNA standards (Invitrogen). The LPCR products were then pooled, DNAse digested, biotin endlabelled and hybridized to individual RAs overnight following established protocols contained in 32. Subsequent washes and stains were carried out as described in Cutler et al. 32 and were only washed and not antibody stained. RAs were scanned at 570 nm, with a pixel size of 3 μm/pixel averaged over two scans. Automated grid alignment and base calling was performed for the .DAT files on a Mac G5 computer with the ABACUS software suite.

An ABACUS parameter search was employed to determine those parameters that called the maximal number of bases while minimizing discrepancies 32. This total experiment consisted of 118 RAs, of which three failed (< 60% base calling). Of the remaining 115 RAs, 8 were used to sequence individual strains once. Of the remaining 107 RAs, 96 were used to replicate hybridize 48 B. anthracis strains, while the remaining 11 RAs were used as additional multiple replicates of these same strains. In total, sequence data was generated from 56 unique B. anthracis strains (see Additional data file 1 for strain listing). In order to obtain the most complete data possible, for those strains with replicate RA sequences, a single composite strain sequence was generated for subsequent population genetic analyses. The current version of ABACUS algorithm is not designed to detect insertion/deletion variation.

The effect of oligonucleotide probe composition was determined by choosing for each base, the probe with the most purines or the most guanines. The number of times that a given base was called was tabulated across all 115 successful RAs. The mean purine and guanine composition was determined for the classes that were called in all 115 RAs and uncalled in all 115 RAs. A Student's t test with unequal variances was used to test for difference in mean sequence composition (purines/guanines) between the always called and never called classes. The DNA sequence files for the 115 RAs and the original RA image files (.DAT files) are available from the authors and will be made available through the NCBI Trace Archive.

All population genetic analyses were calculated using the popgen_fasta2.0.c code (Cutler DJ, unpublished work) on the collection of 56 sets of B. anthracis fasta files. The fasta files were analyzed in total and separately for the main chromosome and plasmids pXO1 and pXO2. The identification of genes was taken from publicly available annotation contained in the relevant GenBank refseq files (B. anthracis str. Ames NC_003997; pXO1, NC_001496; pXO2 NC_002146). The statistical significance of linkage disequilibrium between site pairs was performed by using the Fisher's Exact Test at P < 10^-369.

To account for missing data, θ is estimated by [Σ_n(S_n/a_n)]/L, where S_n is the number of observed segregating sites at positions with exactly n alleles sequenced (n is a maximum of 56, fewer with missing data), a_n = Σ_{i = 1..n-1} 1/i, and L is the total length of the sequence examined. Var{θ} is estimated by [Σ_n (L_nθ/a_n + (L_n)²b_nθ²/(a_n)²]/L², where L_n is the number of sites with data from exactly n alleles, and b_n = Σ_{i = 1..n-1} 1/i². With missing data π is estimated by [Σ_i 2p_iq_in_i/(n_i - 1)]/L, where the sum is taken over all sites i, p_i and q_i are the allele frequencies at site i, and n_i is the number of alleles sequenced at site i.

To determine if the estimates of theta between SNP types (silent, replacement, intergenic) are significantly different, we used the number of samples sequenced, the number of segregating sites, and the length of the region to find a maximum-likelihood estimate of theta per site for each SNP type using equations 11 and 12 in Hudson 70. We compared all possible SNP types against each other (silent vs replacement, silent vs intergenic, replacement vs. intergenic). For a given pair of SNP types, we first determined the maximum-likelihood estimator of theta for each type individually. We then determined the maximum-likelihood estimator of theta, assuming both types had identical theta per site. We ask whether the model with different thetas for each type fits significantly better than the model with a single theta through a likelihood ratio test. Reported significances are the p-values from the likelihood ratio test.

Comparing the observed site frequency spectrum with that expected under the neutral theory is a powerful approach to detect unusual patterns of genetic diversity. We employed two different approaches for this analysis. First, we calculated the expected number of sites with minor allele frequency i as Σ_nθL_n [1/i + 1/(n - i)] and from this determine the expected percent of sites under the neutral expectation. This is directly compared with the observed percent of SNPs in Figure 2. Confidence intervals for the sample proportion of each SNP minor allele frequency classes as

where N is the number of SNPs observed for each class, is their observed frequency, and .

As a second method, we employed Tajima's D statistic 50, estimated as (π - θ)/Var(π - θ). Under the neutral model, π and θ have the same expectation, hence Tajima's D is expected to be 0. Since π is a function of site heterozygosities and θ is a function of the total number of segregating sites, Tajima's D is negative (positive) with an excess (deficit) of rare sites. We use our estimated values of π 51 and θ 45, multiplied by the total genome B. anthracis genome length (5,505,178), to determine the expected number of SNPs that we would expect to observe among two B. anthracis strains sampled in same random fashion as isolates in this study were chosen. Using Equations 6-9 in 51, we calculated the variance of and θ estimators. The one standard deviation (SD) that we report is the square root of this variance.

The 37 variable positions identified in this study were concatenated together to create artificial sequence types. A DNA distance matrix was created using DNADIST, plotted as a UPGMA tree using NEIGHBOR and the tree plotted using DRAWGRAM 57.