Sequencing of natural strains of Arabidopsis thaliana with short reads

Results

We employed the Illumina SBS technology to sequence two naturally inbred, and thus homozygous, A. thaliana accessions, Bur-0 and Tsu-1. To characterize the method, and to identify errors in the reference genome sequence (The Arabidopsis Genome Initiative 2000), we included the Col-0 accession in our study. The genetic distances of the two wild accessions to Col-0 are representative of much of the global population (Nordborg et al. 2005). Recombinant inbred lines (RILs) from crosses between Col-0 and Bur-0 (Simon et al. 2008) as well as Tsu-0, which is likely nearly identical to Tsu-1 as deduced from a genome-wide scan with 149 markers (Warthmann et al. 2007; http://naturalvariation.org/hapmap; http://dbsgap.versailles.inra.fr/vnat/Fichier_collection/Rech_rils_pop.php), have been established, making the Bur-0 and Tsu-1 sequences immediately useful in quantitative trait locus (QTL) cloning efforts with these RILs.

Bur-0, Tsu-1, and Col-0 have been sampled before by PCR and Sanger sequencing at 1213 fragments spaced throughout the genome (a total of 612–654 kb per accession), and the entire genomes have been analyzed with Perlegen resequencing arrays (Nordborg et al. 2005; Clark et al. 2007). We used these two data sets, the “dideoxy” and “resequencing array” data, in combination with the reference sequence as resources for our study. The Bur-0 and Tsu-1 seed stocks were the same as those used to generate the dideoxy and resequencing array data. The Col-0 lines (CS22625 and CS22681) examined in these studies were at most four generations apart.

Generation and alignment of SBS data to the reference genome

From genomic DNA libraries, we produced 120–173 million SBS reads per accession (Supplemental Methods). Most, 88%, were 36 bp in length, and the rest were 31 or 35 bp. Across all accessions, 75% of these raw reads were retained after applying a custom quality filter (Table 1). We aligned the filtered reads to the 119-Mb euchromatic Col-0 reference genome using the enhanced suffix array implementation Vmatch (http://www.vmatch.de; see Methods). We allowed up to four mismatches or indels of up to 3 bp, and reads were aligned through a series of increasing edit distances. A read that had been mapped at a given distance was removed from subsequent alignment attempts; where reads matched multiple genomic locations at a given distance, e.g., in repetitive regions, all alignments were recorded. For Col-0, 89% of filtered reads could be aligned to the reference sequence (Table 1).

Based on the reference sequence, we classified as repetitive any position for which all overlapping 36-bp sequences matched perfectly to multiple genome locations, and as moderately repetitive those for which at least one overlapping 36-bp sequence mapped to multiple locations. These sequences constitute 5% and 8% of the reference genome, respectively, leaving 87% (or 104 Mb) as nonrepetitive. For Col-0, 99.0% of nonrepetitive positions were covered by at least one read, and 95% were covered by three or more reads. Coverage depths by accession ranged from 15- to almost 25-fold (Table 1; Supplemental Methods).

We observed that read coverage was from about five- to 9500-fold higher than expected for 345–385 kb of the reference genome in a given accession (see Supplemental Methods). These regions, which corresponded to rDNA, centromeric, and other highly repetitive sequences whose copy number is not reflected by the reference genome assembly, were excluded from subsequent analyses. Disregarding these regions as mapping targets is expected to significantly reduce the runtime of our (or any related) read mapping method.

Read quality and coverage characteristics

For Col-0, nonperfect alignments resulted predominantly from sequencing errors. As assessed with uniquely aligned reads, incorrect base calls were infrequent at the 5′ end of reads, but increased to ∼5% at position 36 (Fig. 1A). In total, 67% of aligned reads were error-free, and 93% had at most two errors (Supplemental Fig. S1). By error type, insertion or deletion of bases accounted for <1% of disagreements with the reference (Table 2). Incorrectly called bases generally had low quality as measured by the probability (“prb”) and “chastity” values calculated by Illumina’s SolexaPipeline (Fig. 1B; Supplemental Fig. S2; see Methods).

We also assessed coverage uniformity in nonrepetitive regions. Compared with the random (Poisson) expectation, a broader distribution in per-position coverage was observed in the Col-0 data (Supplemental Fig. S3). Coverage was positively correlated with local GC content, and the strongest correlation was observed with window sizes of ∼100 bp (Supplemental Fig. S4). This is longer than the read lengths we used, suggesting a potential bias during library construction using the protocol recommended by Illumina at the time (see Hillier et al. 2008 for a similar finding). Biases in base calling, although minor, may also have contributed to the observed patterns (Table 2). Similar relationships between local GC content and coverage were also apparent in the Bur-0 and Tsu-1 data (Supplemental Figs. S4, S5).

Optimization of sequence predictions

To optimize parameters for identification of polymorphic and conserved bases from the aligned reads, we used the genome-wide Bur-0 dideoxy data set, which covers 604 kb of the Bur-0 genome and includes 2806 SNPs and 326 indels of 1- to 3-bp in length at nonrepetitive positions (Nordborg et al. 2005; Clark et al. 2007). We examined the predictive power at coverage depths from about three- to 25-fold by subsampling the Bur-0 data (Table 1; Fig. 2).

To lessen the ambiguities resulting from repeats, we predicted sequences only at nonrepetitive sites for which per-position read coverage was less than fivefold the expected sequencing depth (∼87% of the reference genome; see Supplemental Methods). Briefly, a sequence was predicted at a position subject to minimum read support and concordance of bases in the aligned reads (80% for base calls, 67% for indels). Minimum read support (Fig. 2) is the lowest number of reads required to accept a prediction for a given position, after reads with an aligned base of low quality at a position had been excluded, based on empirically determined prb and chastity score thresholds that removed ∼79% of sequencing errors while retaining ∼93% of all bases (see Methods). Moreover, as ambiguous alignments are inherently more likely at read ends (Supplemental Fig. S6), and because error rates are higher at the 3′ end of reads (Fig. 1A), we required all predictions to be supported by the core, which excluded four bases each at either end, of at least one read. When predictions were compared with the dideoxy data set, 16 false-positive SNP calls remained. The primary reason for these false predictions was ambiguous alignments at the edge of indel polymorphisms, and not the accumulation of sequence errors.

We used the dideoxy data set to examine the relationships between prediction specificity and sensitivity at different genome-wide coverage depths and minimum read support at a position (see Supplemental Table S1 for details). Even at about threefold coverage, specificity for SNPs was >99% when requiring a minimum read support of two, and more than half of all SNPs could be detected. Above ∼11-fold coverage, sensitivity increased only moderately, and reached 85% at ∼25-fold coverage. For 1-bp indels, which constitute half of the indels in the dideoxy data (Nordborg et al. 2005; Clark et al. 2007), the specificity and sensitivity of predictions was lower, with higher coverage needed to attain similar performance estimates (see also Supplemental Fig. S6). For 2- to 3-bp indels, 24 of 100 indels in the dideoxy data were correctly predicted with only two false predictions when employing a minimum read support of three and all data. For nonrepetitive bases identical to the Col-0 reference sequence, nearly 98.9% were predicted at a specificity of 99.985% when using all data and a minimum read support of three.

We also determined performance separately for coding and noncoding sequences (see Supplemental Methods). Coding regions are on average less polymorphic than noncoding regions (Nordborg et al. 2005; Clark et al. 2007), simplifying read alignment. Additionally, because our read coverage was higher in GC-rich regions (Supplemental Fig. S4), we would expect better performance in coding sequences, which tend to have higher GC content than noncoding sequences in A. thaliana (The Arabidopsis Genome Initiative 2000). Indeed, sensitivity for SNP prediction was higher for coding than for noncoding sequences (91% vs. 79%), with little difference in specificity (99.7% vs. 99.0%).

We obtained very similar performance estimates when applying our pipeline and parameter sets to the lower coverage Tsu-1 data (Supplemental Fig. S7).

Comparison with MAQ

For the Bur-0 data, we compared SHORE with the Mapping and Assembly with Quality (MAQ) software package (http://maq.sourceforge.net/), which also performs read alignment and SNP prediction, but identifies indels only from paired-end reads (Li et al. 2008). As input for MAQ, we supplied all Bur-0 reads together with their transformed prb values in FASTQ format (see Supplemental Information). Employing similar read mapping parameters as we used for SHORE (Supplemental Table S2), we predicted SNPs for regions of overlap to the Bur-0 dideoxy data, with and without MAQ’s SNPfilter option, which maximizes specificity (Table 3; see Supplemental Information; Supplemental Table S3).

The error rate of MAQ-unfiltered predictions was >18-fold higher than those of SHORE, with only a small increase in sensitivity (Table 3; Supplemental Table S1). The MAQ SNPfilter option significantly decreased error rates, but they were still four times higher than those of SHORE, while sensitivity was then substantially lower (Table 3). SHORE therefore compares favorably with MAQ, at least for A. thaliana, which is considerably more polymorphic than human. One reason might be that SHORE, but not MAQ, allows for gapped alignments and prioritizes read core information. These optimizations reduce alignment ambiguities, e.g., mismatches where gaps should be introduced, which can lead to false SNP predictions. In addition, the standard implementation of SHORE does not rely on single read coverage, even if base qualities are optimal, thereby reducing the potential of false positives introduced by undetected sequencing errors.

Genome-wide predictions in the reference accession

Applying a minimum read support of three, we predicted SNPs, 1- to 3-bp indels, and reference bases genome-wide for each accession. For our Col-0 accession, which is closely related to the strain used for the reference sequence (The Arabidopsis Genome Initiative 2000), we predicted 1172 SNPs and 1287 indels. Although some of these predictions might be false, many identify potential errors in the reference sequence or reflect differences among the “Col-0” accessions used in various laboratories. Consistent with the latter possibilities, seven SNP and 11 indel predictions had also been identified in the 654 kb of Col-0 dideoxy data (Nordborg et al. 2005; Clark et al. 2007), for which we made only a single false prediction. Moreover, for 472 of the Col-0 SNP predictions, hybridization data were available from resequencing arrays (Clark et al. 2007). Here, the concordance with SBS calls was 97%. Finally, the majority of the Col-0 SNPs (82%) and indels (91%) were also predicted in either Bur-0 or Tsu-1. We conclude that many of these cases reflect errors in the reference sequence (Supplemental Table S4). To facilitate exclusion of potential reference errors in future sequencing projects, we provide these data in a dedicated Col-0 SNP database at http://1001genomes.org.

Genome-wide predictions in the divergent accessions

For Bur-0 and Tsu-1, we predicted 823,325 nonredundant SNPs and 79,961 nonredundant 1- to 3-bp indels after accounting for potential errors in the reference sequence (Table 4). In earlier work, we used resequencing array data for genome-wide SNP discovery at a specificity of ∼98% (Clark et al. 2007), with the resulting MBML2 data set serving as the basis for the A. thaliana 250 k SNP array design (Kim et al. 2007). The MBML2 data set contained SNP predictions for 153,218 and 120,524 positions in Bur-0 and Tsu-1, respectively, that we could also analyze with the Illumina SBS approach. At these positions, the concordance between the two methods was 99.0%.

With a modification to our base-calling algorithm (see Methods), we predicted 20,795 and 18,047 nonredundant SNPs in moderately repetitive positions for Bur-0 and Tsu-1, respectively, along with 1209 and 1026 1- to 3-bp indels (Supplemental Table S5). Moderately repetitive positions are largely absent from the dideoxy data obtained after PCR amplification (Nordborg et al. 2005), and the repeat composition in the dideoxy data differs from the genome at large, with a lower proportion of long repetitive tracks. Nevertheless, 56 of the 126 SNPs and indels in moderately repetitive regions that did overlap the dideoxy data were predicted correctly, with only three false predictions.

Identifying highly polymorphic regions and targets for de novo assembly

For Bur-0 and Tsu-1, 4.3 Mb of nonrepetitive or moderately repetitive regions in the reference genome were not covered by reads. In part, this is due to stochastic differences in coverage, although systematic biases, such as the reduced coverage of GC poor regions, likely contribute (Supplemental Fig. S3). However, low or no coverage can also be a consequence of sequences being highly dissimilar or deleted relative to the reference. To detect such polymorphic regions (PRs), we identified tracts of at least eight contiguous nonrepetitive or moderately repetitive positions covered by one or no read. Because of reduced read coverage for GC-poor regions, we made predictions only when GC content within PRs was at least 25% (see Methods).

For Col-0, 28 kb of the reference sequence was included in PRs (Table 5), and of this, 22 kb localized to eight genomic locations for which reduced or absent hybridization was apparent in the resequencing array data (Clark et al. 2007). Each of these regions overlapped PRs in Bur-0 and Tsu-1, and five of them overlapped known misassemblies in the reference genome that contain cloning vector sequences (Supplemental Table S6). In contrast to Col-0, 3.25 (3.13) Mb of sequences was included in PRs in Bur-0 (Tsu-1) (Table 5). As expected, these PRs strongly overlapped with a similar set of predictions made based on strongly reduced hybridization signal for Bur-0 or Tsu-1 on resequencing arrays (Supplemental Fig. S8; Zeller et al. 2008).

We also identified short regions with fewer than eight contiguous positions with markedly reduced read coverage (see Methods). Many of these low-coverage tracks are expected to result from stochastic dips in coverage, and 80,912 were found in Col-0. Nonetheless, ∼1.6-fold more were identified in Bur-0 (134,544) and Tsu-1 (127,262). This suggests that many correspond to dissimilar sequences, such as clusters of SNPs, or insertions of any size. These regions, along with PRs, are therefore candidates for targeted de novo assembly (see below).

Targeted de novo assembly of dissimilar sequences

To characterize polymorphic sequences inaccessible to our mapping approach, we developed a targeted de novo assembly strategy that incorporates the Velvet short-read assembler (Zerbino and Birney 2008). As input, we included unmapped (leftover) reads of high quality and all reads that mapped to 100 bp of 5′ and 3′ flanking sequences of either PRs or low-coverage regions (Fig. 3; see Methods). To reduce the potential for incorrect assemblies, we excluded instances where >20% of flanking sequences were repetitive or had higher-than-expected coverage (see Methods). After assembly, we used several criteria to retain only high-confidence contigs spanning polymorphic regions. A contig was rejected if (1) it was <150 bp; (2) it harbored <80% of the 5′ and 3′ flanking reads for a given target; (3) it contained reads from unlinked flanking regions; or (4) <5% of the reads came from the pool of leftover reads (see Methods). The last criterion helped to identify contigs that resulted specifically from sequence dissimilarities, as it should exclude instances where reduced read coverage (one of the criteria used to select targets for de novo assembly) resulted solely from stochastic effects on coverage.

We aligned each contig to the homologous target in the reference genome. At this step, we further removed contig sequences for which alignments did not span the PR or low-coverage-region endpoints (see Methods). Genome-wide, 7396 (3525) contigs passed all filter criteria for Bur-0 (Tsu-1). The longest deletion was 641 bp, and the longest insertion was 408 bp. As assessed against the Bur-0 dideoxy data (Nordborg et al. 2005; Clark et al. 2007), 27% of the regions selected for targeted de novo assembly were closed, and for each of 80 contigs that overlapped the dideoxy data by at least 100 bp, the contig sequence was identical to the dideoxy sequence. Two additional sets of 96 short (≤250 bp) and 96 long (>250 bp) randomly chosen assemblies were validated by PCR and dideoxy sequencing (Supplemental Methods). All 188 successfully amplified and sequenced assemblies were found to be completely correct (Supplemental Table S7). Combined, the 268 validated assemblies revealed 287 indels. The longest validated deletion was 222 bp, and the longest insertion was 124 bp.

Genome-wide, for all but one of the 10,921 Bur-0 and Tsu-1 contigs, one or more sequence differences were observed relative to the reference genome. A large fraction of the targeted assemblies harbored polymorphisms inaccessible to our mapping strategy; in Bur-0 (Tsu-1) we identified 3597 (1690) deletions and 2989 (1257) insertions >3 bp (Supplemental Table S8). As expected, more deletions were identified, as de novo assembly of inserted sequences is a more complex task. In addition, the assemblies contained 13,599 (6997) SNPs and 4214 (2033) 1- to 3-bp indels that had not been predicted from the aligned reads alone (Supplemental Table S8).

The success of targeted assembly is expected to depend on coverage depth. This likely explains the higher number of contigs and polymorphisms detected for Bur-0, with ∼25-fold coverage, than for Tsu-1, with ∼18-fold coverage (Table 1).

When the targeted assembly method was applied to the Col-0 data, only 20 contigs passed the quality filters. The “assembly target regions” for Col-0 likely resulted from stochastic differences in coverage (i.e., no reads were present to contribute to assemblies). The coverage depths for Col-0 and Tsu-1 were similar, but 170 times as many assemblies were obtained for Tsu-1, indicating the robustness of our method.

Detection of duplications

Earlier work with microarrays has suggested that A. thaliana accessions vary substantially in copy number for many sequences (Borevitz et al. 2003, 2007). These finding are generally consistent with the high fraction of tandemly duplicated sequences observed in the Col-0 reference (The Arabidopsis Genome Initiative 2000), and with targeted sequencing of genomic clones from divergent accessions (e.g., Noël et al. 1999; Kroymann et al. 2003; Bomblies et al. 2007). In short-read data, higher-than-expected coverage can indicate such variants, and read coverage has been used as a metric for region-specific copy number in C. elegans (Hillier et al. 2008).

To identify sequences present in two or more copies relative to the reference, we identified regions of higher-than-expected coverage in which several reads supported more than one base at one or more positions (Fig. 4; Emrich et al. 2007). First, we estimated the expected coverage per position based on sequencing bias, local repetitiveness, and sequencing depth (see Supplemental Methods). The correlation between the expected coverage and the observed coverage is significant for Col-0 (Spearman’s test, ρ = 0.82, P < 10⁻¹⁰). Second, we used this estimated value to segment the genome into regions of contiguous, nonrepetitive positions of at least 250 bp for which the ratio of observed to expected coverage at all positions was between 1.2 and 3 (see Supplemental Methods). For Bur-0, 554 kb were included in these regions, and 499 kb for Tsu-1. A fraction of these is likely to result from stochastic effects or unrecognized biases in the method, as our criteria also identified a total of 199 kb in Col-0 (Supplemental Table S9). We therefore applied a second step to identify duplicated regions by searching for apparent “heterozygous” positions (Emrich et al. 2007). Here, reads from duplicated sequences map to the same location in the reference sequence, and the presence of two alternative high-quality bases at the same position defines sites that have diverged between duplicated sequences. Such copy variable positions (CVPs; see Supplemental Methods) were strongly overrepresented in the regions of higher-than-expected coverage in both Bur-0 and Tsu-1 (χ²-test, P < 10⁻¹⁰ for both Bur-0 and Tsu-1). The presence of two alternative bases within regions of elevated coverage was also readily apparent in resequencing array data (Fig. 4; Supplemental Fig. S9; Clark et al. 2007). While 332 and 364 kb of elevated coverage regions with one or more CVPs were identified in Bur-0 and Tsu-1, 30-fold less sequence was identified in Col-0, an indication of the high specificity of our method (Supplemental Table S9).

Effects on genes

Our SBS-based predictions allow a finer scale characterization of the types of sequences and genes that vary among accessions than was previously possible. As assessed with polymorphisms predicted from aligned reads (Table 4), sequence variation is common within translated sequences. For example, we found 168,866 nonredundant SNPs in coding sequences, of which 80,660 were predicted to change amino acids, and 651 to introduce premature stop codons (Supplemental Table S10). Consistent with the expectation of selective constraint in coding regions, 1- or 2-bp indels, which introduce frameshifts, were about 12-fold underrepresented in coding relative to noncoding sequences. However, 1839 potential frameshift variants were identified in Bur-0 or Tsu-1 relative to the Col-0 reference (Supplemental Table S10). Within genes, the relative position of premature stops or frameshift variants was strongly biased to the 3′ end of open reading frames, which would be expected to have less dramatic effects on gene products (Fig. 5).

We also found that 1282 (1098) of Bur-0 (Tsu-1) PRs >99 bp overlapped coding regions. Many of these likely reflect genic deletions, and in total, 1871 (1694) of 26,819 protein-coding genes were predicted to be affected by major-effect changes (premature stops, frameshifts, or long PRs) (Supplemental Table S10). The distribution of these polymorphisms is highly nonrandom, with many genes harboring two or more such variants (χ²-tests, P < 10⁻¹⁰ for Bur-0 and Tsu-0). Although we expect predictions of major-effect changes to have moderately higher error rates (see discussion in Clark et al. 2007), our observations indicate that many genes annotated in Col-0 are either pseudogenes in other accessions, or alternatively vary substantially in structure. Consistent with earlier studies, the distribution of major-effect changes by gene family was also nonrandom, with NBS-LRR genes involved in pathogen response harboring exceptional polymorphism levels (Supplemental Tables S11, S1; Clark et al. 2007).