Species-specific basecallers improve actual accuracy of nanopore sequencing in plants.

Background: Long-read sequencing platforms offered by Oxford Nanopore Technologies (ONT) allow native DNA containing epigenetic modifications to be directly sequenced, but can be limited by lower per-base accuracies. A key step post-sequencing is basecalling, the process of converting raw electrical signals produced by the sequencing device into nucleotide sequences. This is challenging as current basecallers are primarily based on mixtures of model species for training. Here we utilise both ONT PromethION and higher accuracy PacBio Sequel II HiFi sequencing on two plants, Phebalium stellatum and Xanthorrhoea johnsonii, to train species-specific basecaller models with the aim of improving per-base accuracy. We investigate sequencing accuracies achieved by ONT basecallers and assess accuracy gains by training single-species and species-specific basecaller models. We also evaluate accuracy gains from ONT's improved flowcells (R10.4, FLO-PRO112) and sequencing kits (SQK-LSK112). For the truth dataset for both model training and accuracy assessment, we developed highly accurate, contiguous diploid reference genomes with PacBio Sequel II HiFi reads.

Results: Basecalling with ONT Guppy 5 and 6 super-accurate gave almost identical results, attaining read accuracies of 91.96% and 94.15%. Guppy's plant-specific model gave highly mixed results, attaining read accuracies of 91.47% and 96.18%. Species-specific basecalling models improved read accuracy, attaining 93.24% and 95.16% read accuracies. R10.4 sequencing kits also improve sequencing accuracy, attaining read accuracies of 95.46% (super-accurate) and 96.87% (species-specific).

Conclusions: The use of a single mixed-species basecaller model, such as ONT Guppy super-accurate, may be reducing the accuracy of nanopore sequencing, due to conflicting genome biology within the training dataset and study species. Training of single-species and genome-specific basecaller models improves read accuracy. Studies that aim to do large-scale long-read genotyping would primarily benefit from training their own basecalling models. Such studies could use sequencing accuracy gains and improving bioinformatics tools to improve study outcomes.