Do You Want to Build a Genome? Benchmarking Hybrid Bacterial Genome Assembly Methods

Long read sequencing technologies now allow routine highly contiguous assembly of bacterial genomes. However, because of the lower accuracy of some long read data, it is often combined with short read data (e.g. Illumina), to improve assembly quality. There are a number of methods available for producing such hybrid assemblies. Here we use Illumina and Oxford Nanopore (ONT) data from 49 natural isolates of Escherichia coli to characterise differences in assembly accuracy for five assembly methods (Canu, Unicycler, Raven, Flye, and Redbean). We evaluate assembly accuracy using five metrics designed to measure structural accuracy and sequence accuracy (indel and substitution frequency). We assess structural accuracy by quantifying (1) the contiguity of chromosomes and plasmids; (2) the fraction of concordantly mapped Illumina reads withheld from the assembly; and (3) whether rRNA operons are correctly oriented. We assess indel and substitution frequency by quantifying (1) the fraction of open reading frames that appear truncated and (2) the number of variants that are called using Illumina reads only. Applying these assembly metrics to a large number of E. coli strains, we find that different assembly methods offer different advantages. In particular, we find that Unicycler assemblies have the highest sequence accuracy in non-repetitive regions, while Flye and Raven tend to be the most structurally accurate. In addition, we find that there are unidentified strain-specific characteristics that affect ONT consensus accuracy, despite individual reads having similar levels of accuracy. The differences in consensus accuracy of the ONT reads can preclude accurate assembly regardless of assembly method. These results provide quantitative insight into the best approaches for hybrid assembly of bacterial genomes and the expected levels of structural and sequence accuracy. They also show that there are intrinsic idiosyncratic strain-level differences that inhibit accurate long read bacterial genome assembly. However, we also show it is possible to diagnose problematic assemblies, even in the absence of ground truth, by comparing long-read first and short-read first assemblies. Author NotesAll supporting data, code and protocols have been provided within the article or through supplementary data files. The supporting code is available from the GitHub repository https://github.com/GeorgiaBreckell/assembly_pipeline. nine supplementary figures and three supplementary tables are available with the online version of this article. Data summarySequence data and genome assemblies for the natural isolates are available at https://www.ebi.ac.uk/ena/browser/view/PRJEB36951. Genome assemblies for additional E. coli strains used here are available from NCBI: (MG1655, SE11, REL606, CFT073, W, IA136, O157:H7-EDL933)