Improving Long-read Consensus Sequencing Accuracy with Deep Learning

The PacBio HiFi sequencing technology combines less accurate, multi-read passes from the same molecule (subreads) to yield consensus sequencing reads that are both long (averaging 10-25 kb) and highly accurate. However, these reads can retain residual sequencing error, predominantly insertions or deletions at homopolymeric regions. Here, we train deep learning models to polish HiFi reads by recognizing and correcting sequencing errors. We show that our models are effective at reducing these errors by 25-40% in HiFi reads from human as well as E. coli genomes.