Reference-guided assembly of four diverse Arabidopsis thaliana genomes

Abstract

We present whole-genome assemblies of four divergent Arabidopsis thaliana strains that complement the 125-Mb reference genome sequence released a decade ago. Using a newly developed reference-guided approach, we assembled large contigs from 9 to 42 Gb of Illumina short-read data from the Landsberg erecta (Ler-1), C24, Bur-0, and Kro-0 strains, which have been sequenced as part of the 1,001 Genomes Project for this species. Using alignments against the reference sequence, we first reduced the complexity of the de novo assembly and later integrated reads without similarity to the reference sequence. As an example, half of the noncentromeric C24 genome was covered by scaffolds that are longer than 260 kb, with a maximum of 2.2 Mb. Moreover, over 96% of the reference genome was covered by the reference-guided assembly, compared with only 87% with a complete de novo assembly. Comparisons with 2 Mb of dideoxy sequence reveal that the per-base error rate of the reference-guided assemblies was below 1 in 10,000. Our assemblies provide a detailed, genomewide picture of large-scale differences between A. thaliana individuals, most of which are difficult to access with alignment-consensus methods only. We demonstrate their practical relevance in studying the expression differences of polymorphic genes and show how the analysis of sRNA sequencing data can lead to erroneous conclusions if aligned against the reference genome alone. Genome assemblies, raw reads, and further information are accessible through http://1001genomes.org/projects/assemblies.html.

Fig. 1.

Illustration of reference-guided assembly. Reads and their alignments are shown in blue. Regions of constant coverage were defined as blocks. Adjacent blocks were combined into superblocks until they reached a minimal length of 12 kb. Superblocks were defined in an overlapping fashion, such that blocks could belong to several superblocks. All reads of a superblock were assembled with reads that had not been aligned. Resulting contigs (dark blue) were merged into a nonredundant set of supercontigs (green). Short read alignments against the supercontigs allowed for error correction and scaffolding. Short read alignments against the scaffolds (red) enabled a final quality assessment and filtering.

Fig. 2.

Frequency of indel lengths and variation of coding sequence lengths in Ler-1. Indels with lengths that are a multiple of three are enriched in coding regions (yellow), but not in noncoding regions (gray). This is even more apparent when comparing total length differences between orthologous coding sequences between Col-0 and Ler-1 (green). This trend can only be explained by complex changes in coding sequences that together restore the frame use. See Fig. S3 for other accessions.