gplas: a comprehensive tool for plasmid analysis using short-read graphs

Abstract

Summary

Plasmids can horizontally transmit genetic traits, enabling rapid bacterial adaptation to new environments and hosts. Short-read whole-genome sequencing data are often applied to large-scale bacterial comparative genomics projects but the reconstruction of plasmids from these data is facing severe limitations, such as the inability to distinguish plasmids from each other in a bacterial genome. We developed gplas, a new approach to reliably separate plasmid contigs into discrete components using sequence composition, coverage, assembly graph information and network partitioning based on a pruned network of plasmid unitigs. Gplas facilitates the analysis of large numbers of bacterial isolates and allows a detailed analysis of plasmid epidemiology based solely on short-read sequence data.

Availability and implementation

Gplas is written in R, Bash and uses a Snakemake pipeline as a workflow management system. Gplas is available under the GNU General Public License v3.0 at https://gitlab.com/sirarredondo/gplas.git.

Supplementary information

Supplementary data are available at Bioinformatics online.

1 Introduction

A single bacterial cell can harbor several distinct plasmids; however, current plasmid prediction tools from short-read WGS often have a binary outcome (plasmid or chromosome). To bin predicted plasmids into discrete entities, we built a new method based on the following concepts: (i) contigs of the same plasmid have a uniform sequence coverage (Antipov et al., 2016; Rozov et al., 2016), (ii) plasmid paths in the assembly graph can be searched for using a greedy approach (Müller and Chauve, 2019) and (iii) removal of repeat units from the plasmid graphs disconnects the graph into independent components (Vielva et al., 2017).

Here, we refined these ideas and introduce the concept of unitigs co-occurrence to create a pruned plasmidome network. Using an unsupervised approach, the network is queried to find highly connected nodes corresponding to sequences belonging to the same discrete plasmid unit, representing a single plasmid. We show that our approach outperforms other de novo and reference-based tools and fully automates the reconstruction of plasmids from short reads.

2 Materials and methods

2.1 Gplas algorithm

Given a short-read assembly graph (gfa format), segments (nodes) and edges (links) are extracted from the graph. Gplas uses mlplasmids (version 1.0.0, prediction threshold = 0.5) or plasflow (version 1.1, prediction threshold = 0.7) to classify segments as plasmid- or chromosome-derived and selects segments with an in- and out-degree of 1 (unitigs) (Arredondo-Alonso et al., 2018; Krawczyk et al., 2018). The k-mer coverage SD of the chromosome-derived unitigs is computed to quantify the fluctuation in the coverage of segments belonging to the same replicon unit. Plasmid-derived unitigs are considered to search for plasmid walks with a similar coverage and composition using a greedy approach (Supplementary Methods S1). Gplas creates a plasmidome network (undirected graph) in which nodes correspond to plasmid unitigs and edges are created and weighted based on the co-existence of the nodes in the solution space of the computed walks. Modularity values computed using a selection of partitioning algorithms (Blondel et al., 2008; Newman, 2006; Pons and Latapy, 2005) are considered to perform a voting decision regarding the split of the components into different bins (subcomponents) in the undirected network (Supplementary Methods S1). These bins represent the set of plasmids present in the bacterial isolate and are plotted in the plasmidome network using igraph R package (Csardi et al., 2006). The pseudocode and formalization of the algorithm are available in Algorithm 1 and Supplementary Methods S1, respectively.

Algorithm 1.

Gplas pseudocode

2.2 Benchmarking dataset

Gplas was benchmarked against current existing tools to bin plasmid contigs from short-read WGS: (i) plasmidSPAdes (de novo-based approach, version 3.12) (Antipov et al., 2016), (ii) mob-recon (reference-based approach, version 1.4.9.1) (Robertson and Nash, 2018) and (iii) hyasp (hybrid approach, version 1.0.0) (Müller and Chauve, 2019). To evaluate the binning tools, we selected a set of 28 genomes with short- and long-read WGS available including 106 plasmids from 9 different bacterial species, which were not present in the databases or training sets of the tools (Supplementary Methods S3 and Table S1) (Arredondo-Alonso et al., 2020; De Maio et al., 2019; Decano et al., 2019; Wick et al., 2017).

Let n_bin be the total number of nodes present in the predicted bin and define ref as the reference replicon sequence with a highest number of nodes in each bin. Let n_ref be the total number of nodes comprised in ref. We then define two metrics commonly used in metagenomics for binning evaluation: (i) precision and (ii) completeness (Supplementary Methods S4).

$$\begin{array}{c}\mathrm{precision}=\frac{n_{\mathrm{bin}}\in n_{\mathrm{ref}}}{n_{\mathrm{bin}}}\\ \mathrm{completeness}=\frac{n_{\mathrm{bin}}\in n_{\mathrm{ref}}}{n_{\mathrm{ref}}}.\end{array}$$

3 Results

Gplas in combination with mlplasmids obtained the highest average precision (0.88) indicating that the predicted components were mostly formed by nodes belonging to the same discrete plasmid unit (Table 1 and Supplementary Fig. S1). The reported average completeness value (0.79) showed that most of the nodes from a single plasmid were recovered as a discrete plasmid bin by gplas (Table 1 and Supplementary Fig. S2). We observed a decline in the performance of gplas in combination with mlplasmids (precision = 0.82, completeness = 0.72) when considering uniquely bins with a size larger than one which indicated merging problems of large plasmids with a similar k-mer coverage (Supplementary Fig. S3 and Results S2). However, in all cases, the performance of gplas in combination with mlplasmids performed better than other de-novo and reference-based tools tested here (Table 1). To show the potential of gplas in combination with mlplasmids, we showcase the performance of our approach in two distinct bacterial isolates (Supplementary Results S1 and S2).

Table 1.

Gplas benchmarking

Tool	Precision	Completeness	Bin size
gplas–mlplasmids	0.88/0.82a	0.79/0.72a	6.02/10.9a
gplas–plasflow	0.62/0.45a	0.52/0.32a	7.17/11.1a
hyasp	0.64/0.56a	0.36/0.30a	3.84/5.65a
mob-recon	0.79/0.71a	0.56/0.51a	3.4/7.22a
plasmidSPAdes	0.52/0.27a	0.56/0.38a	6.99/13.7a

^aComponents >1 node.

Mlplasmids only contains a limited range of species models (Supplementary Methods). For other bacterial species, we observed that plasflow probabilities in combination with gplas performed similar than the other de-novo approaches but also introduced bias when wrongly predicting chromosome contigs as plasmid nodes (Table 1 and Supplementary Fig. S1), thereby creating bins corresponding to chromosome and plasmid chimeras (precision = 0.62).

4 Discussion

We present a new tool called gplas, which enables the binning and a detailed analysis workflow of binary classified plasmid contigs into discrete plasmid units by relying on the structure of the assembly graph, k-mer information and partitioning of a pruned plasmidome network. A limitation of the presented approach is the generation of chimeras resulting from plasmids with similar k-mer profiles, k-mer coverage and sharing repeat unit(s), such as a transposase or an IS element. These cases cannot be unambiguously solved. Here, we integrated and extended upon features to predict plasmid sequences and exploit the information present in short-read graphs to automate the reconstruction of plasmids.