All benchmarking was conducted on a Intel® Core™ i7-10700K CPU @ 3.80GHz on a machine running Ubuntu 20.04.6 LTS.

The following Tables show the summary statistics of plassembler (with Flye and Raven) and Unicycler run from real read sets simulated from 6 samples from Wick, Judd, Wyres et al 2021. These were compared against the independent Trycyler derived and manually curated ground truth you can find here.

You can find the full pipeline used to generate these results here and the results themselves here.

You can find the full details on how the real read sets were obtained and subsampled here.

Time and Memory Usage

As can be seen below, plassembler was consistently faster than Unicycler in terms of wall clock time, with the largest increase coming single-threaded. plassembler with Raven was also faster than with Flye, but less so than in the simulated benchmarking.

Threads	Program	Median Wall Clock Time (s)	Min Wall Clock Time (s)	Max Wall Clock Time (s)	Median Max Memory (MB)	Min Max Memory (MB)	Max Max Memory (MB)
1	Plassembler with Flye	7063	4559	7865	9921	9116	10024
1	Plassembler with Raven	5918	3430	7028	2623	2214	3039
1	Unicycler	48325	37282	58823	4671	3583	5832
8	Plassembler with Flye	1493	1250	1688	10172	9710	10897
8	Plassembler with Raven	1126	709	1356	6626	2872	8374
8	Unicycler	7500	4509	9659	7535	7003	8128
16	Plassembler with Flye	975	779	1163	11803	10344	13467
16	Plassembler with Raven	699	397	840	6574	3343	14784
16	Unicycler	3876	2944	5036	14041	13647	14062

Assembly Accuracy

All results below were taken from each program run with 8 threads.

plassembler and Unicycler had identical genome fractions and low indel and mismatch rates. Similarly to the simulated benchmarking, plassembler recovered 2 additional small plasmids missed by Unicycler of lengths 1934 bp _(K. variicola INF345) and 10697 bp (K. oxytoca MSB1 2C).

It turned out that the 10697 bp plasmid recovered in Klebsiella oxytoca MSB1 2C was not recovered using the Trycycler based long-read first assembly method in any sample by Wick, Judd, Wyers, et al., 2021, and annotation with Bakta v1.7.0 (Schwengers et al., 2021) revealed that this plasmid contains a Type III toxin-antitoxin system (Supplementary Table 9) along with other plasmid replication genes (repA etc) - it is possibly a real plasmid missed in that study!

Brackets indicate (minimum and maximum). Indels and Mismatches are expressed as median, and per 100 kbp from QUAST.

Program	Additional Plasmids Recovered	Total Missed Plasmids	Total Fragmented Plasmids	Total Misassemblies	Genome Fraction	Indels (per 100kbp)	Mismatches (per 100kbp)
Plassembler with Flye	2	2	1	0	97.02 (mean), 100 (median), (82.14, 100)	0.46 (0 min, 8.6 max)	0.2 (0 min, 135.23 max)
Plassembler with Raven	2	2	1	0	97.02 (mean), 100 (median), (82.14, 100)	0.64 (0 min, 1.67 max)	0.58 (0 min, 7.94 max)
Unicycler	0	2	1	0	97.02 (mean), 100 (median), (82.14, 100)	0.18 (0 min, 1.67 max)	0.18 (0 min, 2.65 max)