On the (im)possibility of reconstructing plasmids from whole-genome short-read sequencing data

To benchmark algorithms for automated plasmid sequence reconstruction from short-read sequencing data, we selected 42 publicly available complete bacterial genome sequences spanning 12 genera, containing 148 plasmids. We predicted plasmids from short-read data with four programs (PlasmidSPAdes, Recycler, cBar and PlasmidFinder) and compared the outcome to the reference sequences. PlasmidSPAdes reconstructs plasmids based on coverage differences in the assembly graph. It reconstructed most of the reference plasmids (recall=0.82), but approximately a quarter of the predicted plasmid contigs were false positives (precision=0.75). PlasmidSPAdes merged 84% of the predictions from genomes with multiple plasmids into a single bin. Recycler searches the assembly graph for sub-graphs corresponding to circular sequences and correctly predicted small plasmids, but failed with long plasmids (recall=0.12, precision=0.30). cBar, which applies pentamer frequency analysis to detect plasmid-derived contigs, showed a recall and precision of 0.76 and 0.62, respectively. However, cBar categorizes contigs as plasmid-derived and does not bin the different plasmids. PlasmidFinder, which searches for replicons, had the highest precision (1.0), but was restricted by the contents of its database and the contig length obtained from de novo assembly (recall=0.36). PlasmidSPAdes and Recycler detected putative small plasmids (< 10 kbp), which were also predicted as plasmids by cBar, but were absent in the original assembly. This study shows that it is possible to automatically predict small plasmids. Prediction of large plasmids (> 50 kbp) containing repeated sequences remains challenging and limits the high-throughput analysis of plasmids from short-read whole-genome sequencing data.