HyPo: Super Fast & Accurate Polisher for Long Read Genome Assemblies

Key results

1. 01

   ShownOn HG002 assemblies, Hypo-polished contigs had the fewest base-level errors and the fewest missed true variants of all polishers; adding long reads lowered Hypo's errors further, whereas a long-read round left Racon's errors unchanged.

   Against gold-standard high-confidence small variants, the runner-up Racon had more than three fold more InDel errors and about 50% more SNP errors than Hypo and missed almost double the variants. Canu drafts and their polished versions had fewer errors than Wtdbg2 ones. The counts are in figures missing from the extracted text. · 4.2 Comparison of Errors; 3.2 Accuracy Evaluation

2. 02

   ShownHypo ran faster than Racon and wtpoa-cns in both short-read-only and hybrid modes; ntEdit, which checks draft k-mers against a Bloom filter instead of building alignment consensus, was the fastest short-read polisher, with Hypo described as not much slower.

   Short reads only: Hypo took less than one-third of Racon's time and less than half of wtpoa-cns's. Two-round polishing (long then short reads): Hypo finished in under one-third the time of Racon or wtpoa-cns. Absolute timings are in a table not included in the extracted text. · 4.1 Performance comparison; Abstract

3. 03

   ShownHypo needed about half of Racon's memory and polished the whole HG002 genome on a 512 GB server, where neither Racon nor Pilon could run; wtpoa-cns and ntEdit used little memory.

   Racon failed on the whole genome on the 512GB machine even with target splitting and fasta input, so its run was measured on an 8TB RAM machine; Pilon was infeasible on the whole genome. Hypo can process fewer contigs per batch to fit smaller machines at some cost in speed. Memory values are in a table missing from the text. · 4.1 Performance comparison; 3.3 Performance Evaluation; Abstract

4. 04

   ShownHypo-polished assemblies contained the most complete genomic features (genes, coding sequences, exons) on every data-set, while wtpoa-cns polishing removed features that the draft already had.

   Quast feature counts on Chr21, Chr1 and whole-genome assemblies; with Illumina-only polishing the runner-up was Pilon on Canu drafts and Racon on Wtdbg2 drafts. Long reads slightly improved Hypo (not on the Chr21 Canu assembly) and left Racon unchanged. Counts are in a figure not in the extracted text. · Genomic features; 3.2 Accuracy Evaluation

5. 05

   ShownBUSCO completeness did not separate the polishers clearly: Hypo scored at or near the best on every data-set, wtpoa-cns ranged from best to worst, and Racon scored worst on the whole genome.

   BUSCO with the eukaryote ortholog set, computed through Quast, on the Chr21, Chr1 and whole-genome HG002 assemblies; the scores are in a figure that the extracted text does not include. · BUSCO Score; 3.2 Accuracy Evaluation; 1 Introduction