Investigating the mitochondrial genomic landscape of Arabidopsis thaliana by long-read sequencing

Plant mitochondrial genomes have distinctive features compared to those of animals; namely, they are large and divergent, with sizes ranging from hundreds of thousands of to a few million bases. Recombination among repetitive regions is thought to produce similar structures that differ slightly, known as “multipartite structures,” which contribute to different phenotypes. Although many reference plant mitochondrial genomes represent almost all the genes in mitochondria, the full spectrum of their structures remains largely unknown. The emergence of long-read sequencing technology is expected to yield this landscape; however, many studies aimed to assemble only one representative circular genome, because properly understanding multipartite structures using existing assemblers is not feasible. To elucidate multipartite structures, we leveraged the information in existing reference genomes and classified long reads according to their corresponding structures. We developed a method that exploits two classic algorithms, partial order alignment (POA) and the hidden Markov model (HMM) to construct a sensitive read classifier. This method enables us to represent a set of reads as a POA graph and analyze it using the HMM. We can then calculate the likelihood of a read occurring in a given cluster, resulting in an iterative clustering algorithm. For synthetic data, our proposed method reliably detected one variation site out of 9,000-bp synthetic long reads with a 15% sequencing-error rate and produced accurate clustering. It was also capable of clustering long reads from six very similar sequences containing only slight differences. For real data, we assembled putative multipartite structures of mitochondrial genomes of Arabidopsis thaliana from nine accessions sequenced using PacBio Sequel. The results indicated that there are recurrent and strain-specific structures in A. thaliana mitochondrial genomes.