Development of Rosaceae Crop-Specific Nanopore Models for Community Use Through the Genome Database for Rosaceae

Oxford Nanopore Technologies (ONT) sequencing platforms have enabled scientists to generate sequence reads of 20 kilobase or more, which has facilitated the assembly of complete genomes. Nanopore sequencing involves recording the change in electrical signals as the DNA molecule transverses the pore. Converting this signal information into bases – a process called basecalling – is challenging and require the implementation of machine learning. The models developed for ONT basecallers have continually improved. However, the mean quality of the reads generated fall below a Phred score of 20 (99% accuracy), a relative target for high-quality assembly and variant calling, when used on challenging plant samples. These low-quality base calls limit the ability to conduct de novo genome assembly, accurate phasing of haplotypes, and long-read genotyping. To overcome these shortcomings, we fine-tuned ONT’s high accuracy basecalling model using Bonito, an (Artificial Intelligence (AI)-based deep learning basecaller, to develop crop-specific models for five important species in the Rosaceae family. These species include highly valuable tree crops such as apple, peach, pear, plum, and sweet cherry. Through the development of an automated training pipeline, we were able to achieve ∼14% higher mean (Q15.3 vs Q17.7) and median (Q16.6 vs Q19.5) basecall quality, and a ∼20% increase in total reads with Phred scores >Q15. These results were achieved without significantly affecting the read length N50s and total reads called. As a result, these new crop-specific models will enable members of the Rosaceae genomics community to produce higher quality ONT sequencing data, nearing the desired threshold of Q20 or more. Furthermore, we are releasing our automated pipeline to facilitate others to train their own organism or crop-specific models.