BMC Bioinformatics (Jun 2021)

PlasForest: a homology-based random forest classifier for plasmid detection in genomic datasets

  • Léa Pradier,
  • Tazzio Tissot,
  • Anna-Sophie Fiston-Lavier,
  • Stéphanie Bedhomme

DOI
https://doi.org/10.1186/s12859-021-04270-w
Journal volume & issue
Vol. 22, no. 1
pp. 1 – 17

Abstract

Read online

Abstract Background Plasmids are mobile genetic elements that often carry accessory genes, and are vectors for horizontal transfer between bacterial genomes. Plasmid detection in large genomic datasets is crucial to analyze their spread and quantify their role in bacteria adaptation and particularly in antibiotic resistance propagation. Bioinformatics methods have been developed to detect plasmids. However, they suffer from low sensitivity (i.e., most plasmids remain undetected) or low precision (i.e., these methods identify chromosomes as plasmids), and are overall not adapted to identify plasmids in whole genomes that are not fully assembled (contigs and scaffolds). Results We developed PlasForest, a homology-based random forest classifier identifying bacterial plasmid sequences in partially assembled genomes. Without knowing the taxonomical origin of the samples, PlasForest identifies contigs as plasmids or chromosomes with a F1 score of 0.950. Notably, it can detect 77.4% of plasmid contigs below 1 kb with 2.8% of false positives and 99.9% of plasmid contigs over 50 kb with 2.2% of false positives. Conclusions PlasForest outperforms other currently available tools on genomic datasets by being both sensitive and precise. The performance of PlasForest on metagenomic assemblies are currently well below those of other k-mer-based methods, and we discuss how homology-based approaches could improve plasmid detection in such datasets.

Keywords