Identifying contamination with advanced visualization and analysis practices: metagenomic approaches for eukaryotic genome assemblies

Genome assemblies, and raw sequencing data for DNA and RNA

Boothby et al. (2015) constructed three paired-end Illumina libraries (insert sizes of 0.3, 0.5 and 0.8 kbp) for 2 × 100 paired-end sequencing on a HiSeq2000, and six single-end long-read libraries (five Illumina Moleculo libraries sequenced by the Illumina “long read” DNA sequencing service, and one PacBio SMRT library sequenced using the P6-C4 chemistry and a 1 X 240 movie), which altogether provided a co-assembly of 252.5 Mbp. The tardigrade genome released by Boothby et al. (2015), along with the nine sequencing data used for its assembly, are available at http://weatherby.genetics.utah.edu/seq_transf. Independently, Koutsovoulos et al. (2016) generated a 0.3 kbp insert library and a 1.1 kbp insert mate-pair library for 2 × 100 paired end sequencing on a HiSeq2000 that provided a co-assembly of 185.8 Mbp (nHd.1.0). These authors subsequently curated a 135 Mbp draft genome (nHd.2.3) by removing potential contamination and re-assembling filtered short reads (Koutsovoulos et al., 2016). The tardigrade raw assembly and curated draft genome released by Koutsovoulos et al. (2016) are available at http://badger.bio.ed.ac.uk/H_dujardini, and their two sequencing datasets are available from the ENA, under study accession PRJEB11910.

RNA-seq data

We obtained the RNA-seq data using the NCBI accession id PRJNA272543(Levin et al., 2016). Briefly, Levin et al. isolated RNA from H. dujardini using the Trizol reagent (Invotrogen), constructed paired-end Illumina libraries according to the TruSeq RNA-seq protocol, and sequenced their cDNA libraries with a read length of 100 bp.

Quality filtering and read mapping

We used illumina-utils (Eren et al., 2013) (available from http://github.com/meren/illumina-utils) for quality filtering of short Illumina reads using ‘iu-filter-quality-minoche’ script with default parameters, which implements the quality filtering described by Minoche, Dohm & Himmelbauer (2011). Bowtie2 v2.2.4 (Langmead & Salzberg, 2012) with default parameters mapped all reads to the scaffolds, and we used samtools v1.2 (Li et al., 2009) to convert reported SAM files to BAM files.

Overview of the anvi’o workflow

Our workflow with anvi’o to identify and remove contamination from a given collection of scaffolds consists of four main steps. The first step is the processing of the FASTA file of scaffolds to create an anvi’o contigs database (CDB). The resulting database holds basic information about each scaffold in the assembly (such as the k-mer frequency, or GC-content). The second step is the profiling of each BAM file with respect to the CDB we generated in the previous step. Each anvi’o profile describes essential statistics for each scaffold in a given BAM file, including their average coverage, and the portion of each scaffold covered by at least one read. The third step is the merging of all anvi’o profiles. The merging step combines all statistics from individual profiles, and uses them to compute hierarchical clusterings of scaffolds. The default organization of scaffolds is determined by the average coverage information from individual profiles, and the sequence composition information from the CDB. This organization makes it possible to identify scaffolds that distribute similarly across different library preparations. The final step is the visualization of the merged data on the anvi’o interactive interface. The anvi’o interactive interface provides a holistic perspective of the combined data, which allows the identification of draft genome bins, and removal of contaminants.

Processing of scaffolds, and mapping results

We used anvi’o v1.2.2 (available from http://github.com/meren/anvio) to process scaffolds and mapping results, visualize the distribution of scaffolds, and identify draft genomes following the workflow outlined in the previous section, and detailed in Eren et al. (2015). We created an anvi’o contigs database CDB for each scaffold collection using the ‘anvi-gen-contigs-database’ program with default parameters (where k equals 4 for k-mer frequency analysis). We then annotated scaffolds with myRAST (available from http://theseed.org/) and imported these results into the CDB using the program ‘anvi-populate-genes-table’ to store the information about the locations of open reading frames (ORFs) in scaffolds, and their taxonomical and functional inference. We profiled individual BAM files using the program ‘anvi-profile’ with a minimum contig length of 1 kbp, and the program ‘anvi-merge’ combined resulting profiles with default parameters. For the analysis of Boothby et al. (2015) assembly, we also profiled the RNA-Seq data published by Levin et al. (2016) to identify scaffolds with transcriptomic activity, and exported the table for proportion of each scaffold covered by transcripts using the script ‘get-db-table-as-matrix.’ We used the supplementary material published by Boothby et al. (2015) (“Dataset S1” in the original publication) to identify scaffolds with proposed HGTs. Finally, we used the program ‘anvi-interactive’ to visualize the merged data, and identify genome bins. We included RNA-Seq results and scaffolds with HGTs into our visualization using the ‘--additional-layers’ flag. To finalize the anvi’o generated SVG files for publication, we used Inkscape v0.91 (available from https://inkscape.org/).

Predicting the number of bacterial genomes in an assembly

We used the occurrence of bacterial single-copy genes as a proxy to the expected number of bacterial genomes in a raw assembly or in a curated genome bin. First, we ran on each CDB generated in this study the anvi’o program ‘anvi-populate-search-tables’ to search using HMMer v3.1b2 (Eddy, 2011) for bacterial single-copy genes Campbell et al. (2013) published. Then, we used the anvi’o script ‘gen-stats-for-single-copy-genes’ to report the number of hits per single-copy gene as an array of integers from each CDB. We finally used mode (i.e., the most frequently occurring number) of this array as the expected number of complete bacterial genomes in a given collection of scaffolds. For additional discussion regarding the relevance of this metric to predict the number of bacterial genomes in an assembly, see the Supplemental Information 1. The script ‘gen-stats-for-single-copy-genes’ also used the R library ‘ggplot’ v1.0.0 (R Development Core Team R, 2011; Ginestet, 2011) to plot the occurrence of single-copy genes.

Taxonomical and functional annotation of bacterial genomes

We uploaded bacterial draft genomes identified from the raw tardigrade genomic assembly results into the RAST server (Aziz et al., 2008), and used the RAST best taxonomic hits and FigFams to infer the taxonomy of genome bins and functions they harbor.

Data availability

The URL http://merenlab.org/data/ reports (1) anvi’o files to regenerate Figs. 1 and 2, (2) our curation of the tardigrade genome from Boothby et al.’s assembly (which is also available through the NCBI under the bioproject ID PRJNA309530), and (3) the FASTA files for bacterial genomes we identified in the raw assemblies from Boothby et al. and Koutsovoulos et al.

A holistic view of the data

The use of multiple library preparations and sequencing strategies is likely to result in more optimal assembly results (Gnerre et al., 2010). Hence, we focused on the scaffolds generated by Boothby et al. (2015) as a foundation to maximize the recovery of the tardigrade genome. To provide a holistic understanding of the composite sequencing data generated by the two teams, we mapped the raw data from the nine DNA sequencing libraries from Boothby et al., and the two Illumina libraries from Koutsovoulos et al. (2016) on this assembly. Anvi’o generated a hierarchical clustering of scaffolds by combining the tetra-nucleotide frequency and coverage of each scaffold across the 11 DNA sequencing libraries (Eren et al., 2015). Besides visualizing the coverage of each scaffold in each sample, we highlighted scaffolds with HGTs identified by Boothby et al. on the resulting organization of scaffolds, and visualized RNA-seq mapping results. Figure 1 displays the anvi’o merged profile that represents all this information in a single display.

A draft genome for H. dujardini

Through the anvi’o interactive interface we selected 14,961 scaffolds from the Boothby et al. assembly that recruited large number of short-reads in a consistent manner (Fig. 1). This 182.2 Mbp selection with consistent coverage (#1 in Fig. 1) represents our curation of the tardigrade draft genome from Boothby et al.’s assembly. The remaining 7,535 scaffolds, which total about 70 Mbp of the assembly, harbored 96.1% of HGTs identified by Boothby et al. These scaffolds recruited only 0.05% of the reads from the RNA-Seq data, highlighting the extent of contamination in the original assembly. This finding is in agreement with Koutsovoulos et al.’s findings; however, our curated draft genome from the Boothby et al.’s assembly is 47 Mbp larger than the draft genome released by Koutsovoulos et al. (2016), most probably due to Boothby et al.’s inclusion of longer reads from Moleculo libraries. While the portion of scaffolds covered by RNA-Seq data suggests that this additional 47 Mbp still originate from the tardigrade genome, the biological relevance of this information (or lack thereof) for the characterization of the tardigrade genome falls outside of the scope of our study.

The origin of bacterial contamination

Our mapping results indicate the presence of non-target sequences in the assembly that recruit reads only from long-read libraries. One interpretation could be that most of the contamination in Boothby et al.’s assembly originated from Moleculo libraries, post DNA-extraction (Fig. 1). However, while a recent study shows that the majority of long reads from Moleculo libraries originated from low-abundance organisms in the analyzed samples (Sharon et al., 2015), another study suggests relatively more sequencing bias in Moleculo library preparation results (Kuleshov et al., 2015). Therefore, an alternative interpretation of the mapping results can be that the bacterial contaminants were present in the sample pre-DNA extraction at very low abundances, and each Moleculo library preparation included long reads originating from different parts of this rare community. Regardless, long reads considerably improved Boothby et al.’s assembly, which resulted in a larger tardigrade genome following the removal of non-target sequences. While these results reiterate that the use of long-read libraries is essential to generate more comprehensive assemblies, they also suggest that extra care should be taken to better mitigate the presence of non-target sequences in assembly results when long-read libraries are used for sequencing.

We identified three near-complete bacterial genomes affiliated to Chitinophaga and Thermosinus in Boothby et al.’s assembly (Fig. 1). Surprisingly, Boothby et al. identified only a small portion of these complete bacterial genomes as sources of HGTs while applying a metric specifically designed to detect foreign DNA in eukaryotic genomes. For instance, none of the 4,459 genes in bacterial draft genome #2 (selection #3 in Fig. 1) were reported in Boothby et al.’s findings as HGTs. We also processed and visualized the raw assembly (nHd.1.0) from Koutsovoulos et al. (2016) using anvi’o (Fig. S1), and recovered eight bacterial genomes. However, we found no taxonomical overlap between high-completion bacterial genomes from the two sequencing projects (Table S1).

Interestingly, one bacterial genome (selection #2 in Fig. 1) was detected in DNA libraries from both groups, as well as in the RNA-seq data, suggesting that the related bacterial population was in all samples prior to the DNA/RNA extraction step. This genome is affiliated to Chitinophaga, and harbors genes coding for chitin degradation and utilization (Table S2). Chitin occurs naturally in the feeding apparatus of tardigrades (Guidetti et al., 2015), and might be a source of carbon for its microbial inhabitants. The genome also harbors genes coding for the biosynthesis of proteorhodopsin, host invasion and intracellular resistance, dormancy and sporulation, oxidative stress, and tryptophan, which is an essential amino acid for animals (Crawford, 1989; Zelante et al., 2013). Although this genome may belong to a tardigrade symbiont, the generation of the data does not allow us to rule out the possibility that it may be associated with the food source. Nevertheless, this finding suggests that there may be cases where non-target genomes in an assembly can provide clues about the lifestyle of a given host.

Best practices to assess bacterial contamination

Initial assessment of the occurrence of bacterial single-copy genes can provide a quick estimation of the number of bacterial genomes that occur in assembly results (Supplemental Information 1). The use of bacterial single-copy genes can give much more accurate representation of potential bacterial contamination than screening for 16S rRNA genes alone, as they are less likely to be found in co-assembly results (Miller et al., 2011; Delmont et al., 2015). Although Boothby et al. (2015) reported the lack of 16S rRNA genes in their assembly, anvi’o estimated that it contained at least 10 complete bacterial genomes (Fig. 2) using a bacterial single-copy gene collection (Campbell et al., 2013). This simple yet powerful step could identify cases of extensive contamination, and alert researchers to be diligent in identifying scaffolds originating from bacterial organisms. Figure 2 also summarizes the HMM hits in scaffolds found in curated tardigrade genomes from our analysis and Koutsovoulos et al.’s study. We observed that the average significance score for the remaining HMM hits for bacterial single-copy genes in curated genomes was 4.2 times lower in average compared to the HMM hits in assembly results (Table S3). The decrease in the significance scores, and the very similar patterns of occurrence of HMM hits between the two curation efforts suggest that some of the HMM profiles may not be specific enough to be identified only in bacteria.

Two-dimensional scatterplots have a long history of identifying distinct genomes in assembly results (Tyson et al., 2004) and continue to be used for delineating microbial genomes in metagenomic assemblies (Albertsen et al., 2013; Cantor et al., 2015), as well as detecting contamination in eukaryotic assembly results (Kumar et al., 2013). Although scatterplots can describe the organization of assembled contigs, they suffer from limited number of dimensions they can display, and their inability to depict complex supporting data that can improve the identification of individual genomes. These limitations are particularly problematic in sequencing projects covering multiple sequencing libraries, where displaying mapping results from each library can help detecting sources of contaminants. Despite their successful applications, two dimensional scatter plots limit researchers to the use of simple characteristics of the data that can be represented on an axis (such as GC-content). In contrast, clustering scaffolds, and overlaying multiple layers of independent information produce more comprehensive visualizations that display multiple aspects of the data.