Introduction

With the introduction of massively parallel high-throughput next generation sequencing (NGS) platforms, fast and cost-effective whole genome shotgun sequencing of the full variety of organisms has been enabled. The rapid advancement in this field is best represented by the recently introduced small scaled benchtop sequencers (BS), e.g., the MiSeq by Illumina (San Diego, California) and the Ion Torrent PGM by Life Technologies (Carlsbad, California). Albeit providing a lower throughput than their conventional non-benchtop counterparts (e.g., Illumina's HiSeq and Ion Torrent's Proton), they still provide sufficient genomic coverage and sequencing accuracy to be efficiently used for sequencing bacterial genomes [1]–[2].

One crucial step in genome based analysis is the attempt to de novo assemble raw sequencing reads into a bacterial chromosome. De novo genome assembly is the process of reconstructing a whole genome sequence from short sequencing reads by finding common subsequences and assembling overlapping reads to longer continuous sequences, i.e. contigs, under the assumption that such reads originate from the same genomic location. If special pairs of reads with a known pairing distance are available, i.e. mate-pair (MP) or paired-end (PE) reads, this information can be used to arrange individual contigs in an ordered sequence consisting of contigs and gaps of known sizes (scaffolds). This is in particular useful to span gaps related to long repetitive elements which are hard to be resolved solely by overlapping reads of limited length. In general, most current assembly algorithms can be assigned to one of two classes based on their underlying data structure: de Bruijn graph (DBG) and overlap layout consensus (OLC) assemblers. Both approaches utilize a graph structure built upon the sequencing reads and algorithms for graph traversal in order to deduce overlapping sequences and to generate contigs. Very briefly, OLC assemblers build a graph by connecting nodes which represent the sequencing reads by edges representing the specific overlaps. For the DBG approach reads are initially partitioned into substrings of the reads of a fixed length (k-mers) and a graph is built by connecting nodes symbolizing sub-reads that share a specific prefix and suffix, respectively. See Compeau et al. 2011 [3], Li et al. 2012 [4] and Nagarajan and Pop 2013 [5] for detailed descriptions on the principles of DBG and OLC assemblers.

The rapid progress in the field of NGS as well as diverse sequencing procedures and protocols had also impacts on genome assembly algorithms and assembly software solutions. The number of sequence assemblers steadily increased in recent years with currently several dozens of assemblers available (22 genome assemblers compared in 2008 [6], additional 13 assemblers in 2010 [7], further ten assemblers in 2011 [8]). Although this variety becomes somewhat more limited as some assemblers require a particular kind or set of sequencing data as an input, researchers are still confronted with a wide range of assembler candidates. The decision which de novo assembler to use is conditional to several aspects, notably the specification of the applied sequencing platform and protocol (e.g., single-end versus PE reads), characteristics of the sequencing results (read lengths and error profiles) and, if available, characteristics of the sequenced genome (e.g., number of repetitive genomic elements or the genomic %GC-content).

Therefore, systematic evaluations of assemblers are necessary to provide the research community with scientifically sound decision-making support. In the past, several efforts have been made to assess genome assembler efficiency on different scales and for different application scenarios [8]–[11]. Most prominently, two comparative studies with recently introduced remakes contributed greatly to this field, i.e. Assemblathon and Genome Assembly Gold-Standard Evaluations (GAGE). For the first large scale assembler competition Assemblathon 1 [12] and its recent successor Assemblathon 2 [13] assemblies were performed and submitted by different institutions and thereafter evaluated. In the first study, 17 participants generated 41 assemblies based on simulated short Illumina HiSeq reads covering 16 different genome assemblers. In Assemblathon 2, 21 participants assembled 43 genomes of three vertebrate species sequenced on different instruments while using eight genome assemblers. In contrast, for the GAGE [14] and GAGE-B [15] competitions all assemblies were generated under equal conditions. Whereas GAGE compared eight different genome assemblers using multiple Illumina libraries of four different pro- and eukaryotic data sets, GAGE-B concentrated on single library assemblies of nine Illumina sequenced bacterial strains using also eight different assemblers.

Albeit these studies give a very comprehensive picture of the efficiency and applicability of state-of-the-art assembler algorithms, several questions remain unanswered. First and foremost, NGS is evolving fast and consumables, protocols, and technical specifications of BS differ in comparison to conventional NGS instruments. For instance, the MiSeq offers an improved read length of 2×300 base pairs (bp) compared to the maximum 2×150 bp of the HiSeq system. The Ion Torrent PGM, although comparable to Roche's 454 in terms of the error profile, produces a different read length distribution (current maximum read length of about 400 bp for the PGM compared to the 1,000 bp at max. for the 454 GS−FLX+ system). Up to now, only the GAGE-B evaluation took into account MiSeq data sets while evaluations of assemblies originating from PGM instruments are missing. Therefore, methods need to be reconsidered for deciding how to assemble data originating from BS instruments. Secondly, bacterial genomes are underrepresented, particularly in the Assemblathon surveys. Even if bacterial genomes are lightweights regarding their assembly difficulty compared to e.g. vertebrates, they nevertheless have different requirements on sequencing procedures and the assembler algorithms. GAGE-B recently tried to close this gap, yet for the majority of the used data sets a high quality genome reference was not available. A third aspect of an assembler evaluation is more of a practice-oriented nature, e.g. the run time and memory usage of an assembly and the demands each assembler has on the compute infrastructure. An assembler recommendation would have little practical value if best performing assemblers cannot be operated due to impractical hardware requirements. However, this aspect was covered insufficiently in the past. Finally, a comparison study should be transparent in order to sustain reproducibility. This means that all steps beginning from the raw sequencing data to the final results should be sufficiently documented. Especially the Assemblathon competitions could not fully satisfy this requirement owed to the fact that the assemblies were performed at different institutions and documented at varying extent.

Here, we present a de novo assembler evaluation with a strong focus on practical aspects and which addresses the aforementioned unanswered questions. To this end, the main objectives of this Genome Assembly Benchmark Tuned On Bacteria and benchtop sequencers (GABenchToB) are:

• to use real (no synthetic) data originating from benchtop sequencers,
• to use single libraries only of bacterial genomes with an available high-quality reference,
• to consider open licensed as well as commercial assemblers,
• to select assemblers covering different assembly strategies,
• to not perform extensive assembly fine tuning but to rest on default parameters for the different sequencing platforms where possible and to use unprocessed raw reads,
• to include run time benchmark parameters, and
• to ensure for equal executing conditions by using dedicated computing hosts and to deposit all information necessary for reproduction at an open repository.

Results

Evaluation data sets

We have chosen three different bacterial strains for our assembly evaluation: Escherichia coli O157:H7 Sakai (American Tissue Culture Collection [ATCC] accession no. BAA-460), Staphylococcus aureus COL (Network on Antimicrobial Resistance in Staphylococcus aureus [NARSA] accession no. NRS100), and Mycobacterium tuberculosis H37Rv (ATCC accession no. 25618). For all three bacteria full reference sequences are available: NCBI Reference Sequence (RefSeq) accession no. NC_002695.1, NC_002128.1 and NC_002127.1 with modifications as described previously [2] were used as a reference for the E. coli genome and plasmids; NC_002951.2 and NC_006629.2 for the S. aureus genome and plasmid, and NC_000962.2 for the M. tuberculosis genome, respectively. The genomes cover a GC-content between 33 and 66% and have a genome size between 2.81 and 5.59 megabases (Mb). Samples of the three bacterial strains were sequenced on BS platforms Ion Torrent PGM (PGM) by Life Technologies and MiSeq (MIS) by Illumina. Both platforms are not only represented by all three organisms but also with different chemistries, i.e. 2×150 base pairs (bp) PE and 2×250 bp PE sequencing for the MiSeq as well as 200 bp and 400 bp sequencing for the PGM, resulting in a total of ten single library data sets (Table 1). Except for the four E. coli data sets, which were used and published previously [2], all other libraries were generated newly for the purpose of this study.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Overview of the data sets used in this study and their sequencing yield.

https://doi.org/10.1371/journal.pone.0107014.t001

De novo assembler selection

To enable an assembler evaluation we aimed to select a set of de novo assemblers regarding the following criteria: assemblers which are (i) representing DBG or OLC approaches; (ii) free-to-use or open source as well as commercial products; (iii) unbiased in terms of the supported sequencing technology or the required sequencing library (processing single and paired-end reads, not relying on mate-pair libraries and no requirement on multiple libraries); (iv) up-to-date regarding to the date of their release or the latest software update; and (v) established and widely used either in recent de novo sequencing projects or in other assembler evaluations. Based on these criteria we selected nine representative assemblers: AbySS [16], Celera [17], CLC Assembly Cell (CLC bio A/S, Aarhus, Denmark) [18], GS De Novo Assembler (454 Life Sciences, Branford, CT) [19], MIRA [20], SeqMan Ngen (DNASTAR Inc., WI, USA) [21], SOAPdenovo2 [22], SPAdes [23], and Velvet [24] (Table 2). Some assemblers we did not include in our study, but which performed best in either one or several categories in previous evaluations [11]–[15] are: Allpaths-LG [25] because its algorithm requires input data consisting of at least two different libraries (one PE and one MP), ARACHNE [26] because it combines only long Sanger with MP reads, MaSuRCA [27] and SGA [28] because both can not handle single-end data, and Phusion [29] because it builds solely upon MP reads.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. De novo assemblers used for comparison.

https://doi.org/10.1371/journal.pone.0107014.t002

Genome coverage

One major parameter that influences the result of a genome assembly is the amount of available data to cover the whole genome, i.e. the depth of coverage. Insufficient coverage may lead to an inferior assembly result either due to uncovered genomic regions or due to the impossibility to overrule randomly distributed sequencing errors. In contrast, each increase of the target coverage involves a higher load of the sequencing platform capacity, which comes at the price of an increased sequencing cost per library as well as of a growing computational effort due to the increased data volume. Also, non-randomly distributed sequencing errors (e.g., systematic errors related to homopolymer regions) can not be ruled out by redundant sequencing and accumulate with increasing coverage which can trick assemblers to handle them as true biological events [32]. Therefore, finding a balance between sequencing effort, cost, assembly turn-around time and assembly quality is crucial in advance to the sequencing itself. The observation that an ad infinitum increased depth of coverage does not necessarily improve assembly results was already reported by Lin et al. [10]. They found that for seven assemblers operated on six simulated data sets based on a eukaryotic chromosome the depth of coverage at which the N50 length plateau was reached never exceeded values of 50. Consequentially, an upper coverage threshold of 70x was used to compare the assembly results. Similar examinations in further studies yielded comparable results [9], [15], [33]. However, common to all these studies were shortcomings regarding the resolution, the range of assemblers, the amount of data at which this effect was studied, and most of all only Illumina reads were used. To address the question to what extent depth of coverage affects assembly results we have randomly sub-sampled all PGM and MiSeq data sets into a fix range of subsets. All subsets were subsequently assembled using three DBG (CLC, ABYSS, and SPADES) and three OLC assemblers (NEWBLER, CELRA, and MIRA). This reduced assembler set was chosen to cope with the high computational effort while maintaining meaningful results with regard to our assembler selection criteria.

Independent of the data set and the assembler, we found out that for each assembler and data set a window of reasonable coverage can be identified which is confined by a lower and upper bound, which mark the areas of inappropriate coverage (Figure 1). Given the E. coli Sakai data sets, too low coverages result in inferior NGA50 lengths progressively improving with increasing coverage, whereas above the upper bound a further increase in the depth of coverage does not necessarily effects superior assemblies. For the PGM platform oversampling, i.e. too high coverage, can even have a negative impact on the assembly especially if combined with OLC assemblers. Here, NGA50 lengths are constantly falling after reaching a specific maximum. This effect is apparent for both the PGM 200 bp and PGM 400 bp data, albeit the global coverage maxima are higher for the 400 bp (between 40- and 100-fold coverage) than for the 200 bp data sets (between 40- and 50-fold coverage). The MiSeq platform is not susceptible to oversampling. After reaching sufficient coverage, NGA50 lengths are mostly saturated and show only moderate improvement or degradation, marking the upper bound of appropriate coverage for the MiSeq data sets. These findings are in good accordance with the other data sets and with other metrics, as e.g. the number of mis-assemblies (local and non-local mis-assemblies) or the number of assembly errors (insertions, deletions and substitutions) and is further supported by another sub-sampling approach (see Supporting Information Text S1 and Figures S1–S5). In order to achieve comparability between the data sets and the assemblers, we have selected two different cutoffs, reflecting the described upper coverage bounds, at which we sub-sample our data sets prior to any consecutive analysis. The 40-fold coverage limit was chosen for the PGM 200 bp data, whereas all MiSeq and the remaining PGM 400 bp data sets were sub-sampled at 75-fold depth of coverage. Even though these thresholds do not always represent individual global optima, they are not in favor or disfavor of any particular assembler or sequencing platform and in good accordance with previous findings [9], [15], [33].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Effect of the depth of coverage on the assembly efficiency measured by NGA50 sizes based on randomly sub-sampled E. coli Sakai data sets.

The coverage is referring to the average depth each genomic position is covered by the sequencing reads and not to the average depth of coverage the assemblies are actually reaching. The fitted average is, for each data set, the mean of all NGA50 lengths at each coverage fitted to a nonlinear local regression model. Sub-sampling was done in steps as a percentage of the original full sample size; hence, the x-axis ranges of the four sub-plots differ. The dotted vertical lines mark the finally used 40-fold (PGM 200 bp) and 75-fold coverage limits (PGM 400 bp, MiSeq 2×150 bp and MiSeq 2×250 bp).

https://doi.org/10.1371/journal.pone.0107014.g001

Evaluation of de novo assemblies

Combining ten data sets and nine assemblers we have generated 90 de novo assemblies. For every single assembly, the assembly effort was measured using QUAST [31]. Results were transformed into a spread-sheet readable tabulator separated file format and stored in addition to the complete assembly output, log files, and run time benchmarking results. In Figure 2 we present the results of each assembly using the following two key metrics: the NGA50 length and the number of mis-assemblies. As mentioned before, these figures show either scaffolds or contigs according to the data, i.e. scaffolds for all MiSeq assemblies except for MIRA and SEQMAN and contigs for all others including the PGM assemblies. Similarly, Figure 3 shows the central aspect of the running time benchmark, i.e. the wall clock time and the average system utilization. Details about specific assembly parameters and the full execution pipeline are given in the Supporting Information (Text S1) or can be looked up at the following project site: ftp://ftp.cebitec.uni-bielefeld.de/pub/GABenchToB.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Comparison between the de novo genome assemblies based on the NGA50 length and the number of mis-assemblies.

The NGA50 length (A, in kilobases) and the number of mis-assemblies (B, combining local and non-local mis-assemblies) on the y-axis are either contig or scaffold based, respectively. Scaffolds for MiSeq 2×150 bp and MiSeq 2×250 bp assemblies obtained by ABYSS, CELERA, CLC, NEWBLER, SOAP2, SPADES, and VELVET; contigs for MiSeq assemblies obtained by MIRA and SEQMAN as well as for all PGM assemblies. The second plot (B) is further divided into two plot rows where the upper row has an altered y-axis scale only showing high rates of mis-assemblies ranging from two hundred up to thousand.

https://doi.org/10.1371/journal.pone.0107014.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Computing time of de novo genome assemblies.

Based on the elapsed wall clock time (A, in hours) and the total CPU utilization (B, in percent and relative to the 48 available CPU cores of the executing compute host). With regard to the CPU utilization, all assemblies have been instructed via proper parameterization to make maximal use of the 48 available CPU cores. The only exceptions to this were SEQMAN, which does not support parallelization, and CELERA, which due to configuration constraints has altering concurrency and multi-threading parameters for different internal processes. For DBG assemblers only run time and CPU utilization of the single assemblies with the best performing k-mer parameter are shown and not the summation of the full k-mer optimization procedure (for SPADES and CLC this is equivalent).

https://doi.org/10.1371/journal.pone.0107014.g003

At this repository, also the 90 assembly results together with all computed assembly metrics can be accessed. In addition to this core set, each of the more than two thousand assemblies used for the coverage and k-mer parameter optimization are deposited here, with the limitation that not for all of those assembly and run time metrics are available.

Discussion

One of our main goals was to provide the research community with a practice-oriented de novo assembly evaluation of bacterial genomes sequenced by benchtop instruments. In this spirit, our study is intended to be more than a pure assembler comparison following the question which assembler performs best. Also, it is not a sequencing platform comparison trying to answer which platform allows for the best assemblies. Instead, given a diverse set of sequencing data, we identified those assemblers, which are best suited to handle individually different data sets and meet specific requirements such as the least amount of mis-assemblies or the fastest compute time.

Consequently, given the assembly results in Figure 2 and Figure 3, none of the assemblers emerges as the overall winner. The individual assembler performance as given by the NGA50 length strongly depends on the nature of the data (Figure 2A). For MiSeq reads the assembler selection is less restricted than it is for PGM reads. Both, DBG and OLC assemblers are generally applicable on the MiSeq data sets. The E. coli MiSeq 2×150 bp and 2×250 bp assemblies, for example, yield NGA50 values of the same magnitude (about 155 kb) for both approaches. The highest MiSeq 2×150 bp S. aureus NGA50 lengths originate from DBG assemblers, whereas OLC assemblers performed better on the MiSeq 2×250 bp S. aureus and M. tuberculosis data. In contrast, assembled PGM data yielded better results more often when using OLC than DBG assemblers. Comparing the sum of all NGA50 lengths of the DBG assembled PGM 200 bp and 400 bp data sets with those using OLC, a more than 2-fold increase can be observed (719 megabases (Mb) compared to 1518 Mb). The only exception to this is the DBG SPADES assembler, which achieves NGA50 lengths on a par with the OLC assemblers. Omitting SPADES from this comparison, the proportion changes from a 2-fold increase to an over 6-fold increase (220 Mb compared to 1518 Mb), clearly showing the difficulties DBG assemblers have while dealing with PGM data. Considering the NGA50 length, we want to highlight the DBG assembler SPADES and the OLC assemblers NEWBLER and MIRA, as they repeatedly perform above average in their category and are more robust with respect to the kind of sequencing data.

Laying focus on the number of mis-assemblies, a different picture arises (Figure 2B). The NGA50 is the NG50 length after contigs have been split at each observed mis-assembly position. Resting on this relation, the naïve assumption would be that lots of mis-assemblies are directly reflected by a low NGA50 and vice versa. However, here a high and low mis-assembly rates could be observed independent of the NGA50 length for both assembler approaches across all data sets. The highest rate of mis-assemblies originates from an assembly that can be, measured against the NGA50, declared as failed (S. aureus PGM200 bp SEQMAN with NGA50 of 2292 bp). In contrast, some assemblies with also a very low NGA50 show very few mis-assemblies (e.g., all PGM ABYSS assemblies). However, discussing rates of mis-assemblies for assemblies with extremely low NGA50 values has little meaning. Assemblies with a reasonable NGA50 length again show both, i.e. a high (e.g., the SOAP2 and VELVET MiSeq E. coli and M. tuberculosis assemblies) and low (e.g., the SPADES MiSeq E. coli assemblies) rate of mis-assemblies. Of those, three assemblers show the most consistent pattern of mis-assemblies. While SPADES produced the least mis-assemblies in total across all data sets and sequencing platforms, NEWBLER produced high rates of mis-assemblies only when dealing with MiSeq data, and MIRA had one of the most miss-assembled contigs across all platforms. For all other assemblers, the number of mis-assemblies depended on the processed data and varies from very high to very low. Moreover, the NGA50 metric shows no significant correlation with the mis-assembly rates (data not shown). This implies, that a low NGA50 value is not necessarily equivalent to many mis-assemblies. Instead, it simply indicates a comparably inferior assembly, either due to many mis-assemblies or low assembly contiguity. These findings are also consistent with other metrics, e.g. the amount of fully covered genes and the number of assembly errors (insertions, deletions, and substitutions; Figure S8).

Genomes of extreme GC-rich or GC-poor bacteria are known to be challenging for genome assemblers as amplification biases of GC-poor or GC-rich regions can result in uneven genome coverage. Here, a weak relationship between the genomic GC-content and assembly contiguity could be found. Assemblages of the GC-poor S. aureus genome (33% GC) tend to reach higher NGA50 lengths than the E. coli (51% GC) and the GC-rich M. tuberculosis (66% GC) genome, for which on average the lowest NGA50 lengths were achieved. Likewise, the S. aureus assemblies exhibited fewer mis-assemblies than those of the other two genomes. However, the relatively inferior M. tuberculosis assemblies cannot be entirely explained by increased mis-assemblies, as they are comparable to the ones based on the E. coli genome. Instead, this discrepancy might be explained by the finding that extreme GC-rich regions are especially difficult to amplify [34], possibly lowering the assembly completeness and by that the contiguity. Despite these findings, a general conclusion on the effect of the GC-content on genome assemblies cannot be drawn, as this would require a broader range of differing genomes to be analyzed. In addition, the differences shown here are not clear enough and for all genomes a successful assembly could be generated. This is also supported by a previous study, which reports that the degree of a GC bias, the factor most influencing the assembly contiguity, correlates neither with the mean GC content nor with the standard deviation of GC content of a genome [35]. In the same study it could also be shown that a sufficient depth of coverage can compensate for a GC bias, which may explain the comparatively low differences observed here with regard to the sub-sampling optimized approach used in our study.

One neglected aspect in preceding assembler evaluation studies has been the computational cost of an assembly. By measuring the wall clock time of all assembly processes we observed a great discrepancy between DBG and OLC assemblers (Figure 3A). With the exception of SPADES, all DBG assemblers finished within less than 20 minutes, the majority even within less than five minutes, whereas OLC assemblers took from eight minutes up to 20 hours. NEWBLER is to be highlighted positively by showing consistently the shortest run time of the OLC assemblers. CELERA and MIRA, on the contrary, repeatedly exceeded run times of three hours. However, for all DBG assemblers only the run time of a single assembly effort is shown, i.e. of the single assembly call resulting in the highest N50 length (Table S1). So the total wall clock time for an entire assembly project needs to be adjusted with respect to the chosen k-mer optimization strategy and may markedly reduce the outstanding performance of the DBG assemblers. Exceptions to this are the assemblers CLC and SPADES as both do not rely on an external k-mer optimization step. Thus, their run times are equivalent to the overall needed run times. In this sense, CLC outperforms all other assemblers many times over. SPADES prolonged execution times, in contrast, can be explained by its operating procedures, as in the case of MiSeq PE data six and for PGM data five consecutive assemblies were performed internally.

A good indicator showing to which extent an assembler benefits from a parallelized computing environment is the average CPU utilization (Figure 3B). Surprisingly, only ABYSS is nearly full parallelized. All other assemblers have an average CPU utilization below 50%, which means that of the 48 available and requested CPUs, on average, only half of them were used during the full assembly procedure. Next to ABYSS, CLC and SOAP2 show the second and third best CPU utilizations, respectively. Also CLC and SOAP2 have a higher utilization if operating on MiSeq data than on PGM, showing that they are obviously optimized for Illumina data. All other assemblers have a total CPU utilization below 20%. This implies that the running time of those assemblers cannot be increased considerably by assigning more CPU cores. However, a low CPU utilization does not necessarily result in long running times, as demonstrated by VELVET, which always finished in less than four minutes but never exceeded a comparably low workload of 13.3%. One aspect negatively influencing the CPU utilization are input and output (I/O) operations, causing processes to enter a waiting state. Here, most critical are I/O operations caused by swapping when system memory is insufficient. The memory usage of our assemblies was very different but system memory was always sufficient and no assembly process was waiting because of swapping (Figure S9). However, some assemblers (e.g., MIRA, SPADES and CELERA) utilize or demand considerably more memory than others, which should be considered before an assembly attempt in order to circumvent swapping. One possible solution, besides the extension of the system memory, is to make use of specific memory constraining parameters that memory intensive assemblers usually offer. With regard to the average system utilization, we want to clarify that the assemblers are different in terms of the implemented parallelization and the internal assembly pipeline. Thus, a low CPU utilization does not mean that the entire assembly is inadequately parallelized, but that, for instance, parts of the assembly are constrained to system I/O (e.g., because of data pre- and post-processing), resulting in wait times of the depending processes. Therefore, this metric should not be considered as a fixed upper limit of the parallelization capabilities of each assembler. Instead, it reflects to what extent an assembler scales, in terms of run time, with the provided computing power, and helps to find a suitable combination between assemblers and the available hardware infrastructure.

One central aspect of this study was to compare the assemblies from a practical point of view. By that, we omitted additional data processing steps like, for instance, error correction methods. However, to respond to the question whether upstream error correction is reasonable and which assembler would profit from such methods we have exemplary used two different read correction methods on our data sets (BayesHammer from SPAdes [23] for MiSeq data and Coral version 1.4 [36] for PGM data) and compared the consecutive assemblies with those using uncorrected data (Figure S10). In essence, the observed differences are marginal. Even for those assemblers that do not include an error correcting pre-processing step (e.g., SOAP2, VELVET and NEWBLER) the beneficial effects are, albeit partially present, not altering their individual performance.

In this evaluation, we did not perform extensive parameter optimization, i.e. whenever possible we used default parameters. For mandatory or data specific parameters without default values we have chosen appropriate values (details given in Text S1). The only exception to this was the k-mer parameter optimization. We have shown that an optimal k-mer highly depends on both the assembler as well as the given data set. This implies that unless the assembler itself offers a suitable k-mer parameter estimation (as e.g., CLC) or default values (as e.g., SPADES), currently the best solution is to pursue a trial-and-error approach. Of course, this problem is not described for the first time and algorithms have been developed to predict optimal k-mer parameters a priori. KmerGenie, for instance, uses a heuristic to generate k-mer abundance histograms in order to estimate the best possible k-mer value [37]. However, for our data sets the optimal k-mer as predicted by KmerGenie did not matched the k-mer length at which the best assembly was achieved (data not shown). This may be caused by parameter estimations inferred only from the input data, which cannot take into account assembler peculiarities. Still, k-mer estimation may prove useful for other data sets and, especially due to quick heuristics, it is to be preferred over randomly chosen or alleged established k-mer lengths. Parameters proven successful in the past may not be adequate for new assembly problems. The downsides of the trial-and-error approach, in turn, are drastically increased running times countering the speed advantage of DBG assemblers. Besides the k-mer parameter, testing and comparing various parameter settings for each assembler is possible beyond the scope of every evaluation effort simply due to the high computational cost. Moreover, results of comparisons in which assemblies were highly optimized are hard to transfer to other application scenarios, weakening the overall conclusions of such a comparison. Therefore, good default parameters and standardized recipes are needed to support unbiased and useful comparisons. Every evaluation is affected by the applied evaluation procedure and the used metrics. Therefore, the results of this study cannot be interpreted as final principles to rule in or out individual assemblers, but to give a general advice which assemblers to shortlist in consideration of an upcoming bacterial genome assembly. Nevertheless, given the scope of our evaluation we want to highlight some promising combinations between assemblers and bacterial BS data. The most obvious recommendation is to pair different BS platforms with specific assembler approaches, i.e. PGM data is better combined with OLC assemblers whereas MiSeq data shows stronger preferences towards DBG assemblers. The only exception here is the CELERA assembler, which consistently performs better on MiSeq than on PGM data. For MiSeq data, promising DBG assemblers to begin with are SPADES and CLC. Both assemblers offer good performing default k-mer parameters, are generally easy to execute and show one of the highest NGA50 and lowest mis-assembly rates among the DBG assemblers. In addition, SPADES also performed best on PGM data among DBG assemblers. The CLC assembler shows the least memory footprint and the notably fastest running time of all assemblers, which qualifies it for quick and reasonable draft assemblies. Given MiSeq data and a highly parallelized computing infrastructure, the ABYSS assembler might also be an option, as it generates good assembly results while offering the best CPU utilization. For PGM data, researchers should consider to sub-sample their data to a coverage between 40 and 80-fold prior to an assembly in order to prevent negative oversampling effects. The two OLC assemblers that are a good entry point for PGM data sets are MIRA and NEWBLER. MIRA is able to achieve very high NGA50 values particular at higher coverages but comes at the price of more mis-assemblies and a longer running time. NEWBLER, in contrast, convinces with the shortest running time of all OLC assemblers and a low rate of mis-assemblies. Finally, it is to be noted that there is still a great discrepancy between researchers who are developing or evaluating genome assemblers and researchers who simply want to use them. The former ones take great care to avoid exaggerated generalizations, like to conclude that a particular assembler provides the best assemblies. The latter ones, on the other hand, are confronted with concrete application scenarios and therefore require decision-making support without needing to perform extensive evaluations personally. Therefore, evaluations should consider realistic assembly scenarios and include assembly metrics that summarize several assembly features into easy to communicate metrics.

Methods

Supporting Information

Acknowledgments

The authors gratefully thank the system administrators of the Bioinformatics Resource Facility at the CeBiTec for general technical support and in particular for providing the dedicated computing infrastructure. Furthermore, the authors thank DNASTAR Inc. (Madison, Wisconsin, USA), CLC bio A/S (Katrinebjerg, Aarhus, Denmark) and 454 Life Sciences (Branford, CT, USA) for providing extended trail or free licenses of their assembly software products.