Microbial communities from the environment such as the sea, soil, and human gastrointestinal tract often constitute various microorganisms with uneven abundances, thus showing high diversity. The rapid variation of microorganisms may lead to multiple strains of the same species, which further elevates the biodiversity and complexity of microbial communities. Related studies have demonstrated that different strains may show different phenotypes or play different biological functions in the environment (Marx, 2016; Van Rossum et al., 2020). For example, Escherichia coli (E. coli) is usually harmless and common in the gut of mammals, but unusual strains such as E. coli (STEC) O26:H11 and E. coli (EHEC) O104:H4 are highly pathogenic and cause bloody diarrhea or even death (Burger, 2012; Bonanno et al., 2015); different strains of pathogen Staphylococcus aureus and Streptococcus pyogenes could induce significantly varying human immune responses (Sela et al., 2018); some strains of Helicobacter pylori (e.g., cagA-positive strains) are correlated with a higher risk of gastric cancer (Cover, 2016). Therefore, deciphering the strain diversity of microbial communities plays a crucial role in various applications.

In general, amplicon sequencing (e.g., 16S/18S rRNA or ITS) and traditional cultivation-based methods overlook the vast majority of genetic material, whereas metagenomic sequencing enables genome-scale assembly and strain-level analysis. Strain-resolved de novo genome assembly is the ultimate solution to decipher strain diversity in metagenomic samples; however, this problem has largely remained unresolved so far. Notably, here we define the concept of “strain” to agree with that of a “haplotype”, which is a non-redundant genome sequence with sufficient sequencing coverage. This is in accordance with the definition of a strain in the related work (Vicedomini et al., 2021).

Next-generation sequencing (NGS) has been routinely used in de novo metagenome assembly, and many related tools have been developed such as SPAdes (Bankevich et al., 2012), IDBA-UD (Peng et al., 2012), and MEGAHIT (Li et al., 2015). These tools only generate squashed assemblies at the species level without considering strain diversity. In addition, VG-flow (Baaijens et al., 2020) performs strain-aware metagenome assembly at the whole genome scale, whereas Snowball (Gregor et al., 2016) and DESMAN (Quince et al., 2017) do so only at the gene scale. These de novo assembly methods struggle to generate long-strain-specific genome sequences because NGS reads are too short (hundreds of bp) to span inter-/intra-genomic repeats.

Third-generation sequencing (TGS) technologies such as Pacific Biosciences (PacBio) and Oxford Nanopore Technologies (ONT) have revolutionized the method of DNA sequencing because they generate considerably longer reads (tens of Kbps to Mbp). Hence, they have recently provided unprecedented opportunities to carry out haplotype- or strain-resolved de novo genome assembly. In particular, PacBio high-fidelity (HiFi) long reads are highly accurate (sequencing error rate < 1 % ) such that haplotype- or strain-resolved genome assembly can be achieved relatively easily. Related works such as HiCanu (Nurk et al., 2020), hifiasm (Cheng et al., 2021), and hifiasm-meta (Feng et al., 2021) are designed to assemble high-fidelity long reads without losing haplotype information. Nevertheless, other types of long reads (i.e., PacBio CLR, ONT) usually suffer from high sequencing error rate (5–15%). This makes it more difficult to perform strain-resolved metagenome assembly mainly due to the hindrance in distinguishing low frequent strain variations from sequencing errors. Related approaches such as Canu (Koren et al., 2017) and metaFlye (Kolmogorov et al., 2020) as the most dominant metagenome assemblers are unable to separate strains of divergence vary by less than 5%; note that the majority of different strains from one species vary by less than 5%. Later, Strainberry (Vicedomini et al., 2021) was proposed as a post-hoc strain resolver based on pre-assembled contigs from Canu or metaFlye. However, it is only suitable to deal with low-complexity samples and the final result largely depends on the performance of the pre-assemblers used. Overall, there is plenty of room to improve the strain-aware metagenome assembly from error-prone long reads. In addition, previous studies (Luo et al., 2021; Luo et al., 2022) have demonstrated that performing haplotype-aware error correction in noisy long reads prior to genome assembly can significantly improve haplotype awareness.

Here, we propose MetaBooster and MetaBooster-HiFi, as two pipelines for strain-aware metagenome assembly from PacBio CLR and Oxford Nanopore long-read sequencing data. First, MetaBooster/MetaBooster-HiFi performs error correction on raw reads using VeChat (Luo et al., 2022), a new variation graph-based standalone tool to correct errors in long reads. Subsequently, reads corrected by VeChat are combined with reads corrected by Canu, a state-of-the-art long-read assembler (Koren et al., 2017). A crucial insight is that joining the two read sets, and thereby joining the (obviously complementary) virtues of VeChat and Canu creates maximum synergy, which considerably “boosts” the performance of assemblers that one subsequently uses. Here, we focus on assembling the joined reads using either Canu itself (MetaBooster), or the PacBio HiFi mode of Canu (MetaBooster-HiFi).

In this study, we do not present a new method per se; however, we do present a pipeline that is superior over all conceivable de novo long-read-based metagenome assembly strategies presented so far, by fairly large margins. In other words, the primary novelty of this work is the superior performance of the “boosted” error correction—assembly pipeline. An additional novelty is the introduction of VeChat, as a variation graph-based error correction tool for long reads into the world of de novo metagenome assembly.

Benchmarking experiments on both simulated and real sequencing data demonstrate that either the MetaBooster or the MetaBooster-HiFi pipeline drastically outperforms the existing de novo metagenome assemblers in terms of relevant assembly evaluation criteria such as genome fraction, contig length, and error rate. Of further interest, applying Strainberry as a post-hoc tool for revisiting the resulting assemblies, improves the assemblies of MetaBooster and MetaBooster-HiFi by further small amounts. In comparison with the strategies originally suggested in Strainberry, one achieves further increases in performance.

We presented MetaBooster and MetaBooster-HiFi as two novel error correction and assembly pipelines by which to perform de novo strain aware metagenome assembly. In experiments, we observed that MetaBooster and MetaBooster-HiFi introduce novel qualities in that respect, in terms of genome fraction (strain-specific genome coverage), error rates or contiguity of assemblies.

MetaBooster and MetaBooster-HiFi combine the error-corrected reads from VeChat, which is based on a variation graph-based method, and the ones one can retrieve from Canu’s error correction module, as an overlap graph-based method. This way, our pipelines create synergy by exploiting the advantages of the two strategies, on the one hand. To be specific, VeChat generally achieves better error correction for most reads than that of Canu as documented by Luo et al., 2022. However, VeChat may not perform very well in some particular regions, for example in highly divergent genomic regions where different strains substantially vary (see Supplementary Figure S1). Canu may achieve better error correction in such highly variable regions, because it uses a consensus-based strategy for correction. From this point of view, Canu complements the generally superior, but in difficult regions incomplete reads of VeChat wherever necessary. This shows that the joined corrected reads from Vechat and Canu can generate better assembly (see bottom two panels in Supplementary Figure S1). On the other hand, the combination itself increases the sequencing coverage. Both of these points lead to enhanced strain-aware metagenome assemblies (see Supplementary Table S4 for additional arguments about merging error correction modules).

Evaluating benchmarking results in detail indicates that in most cases, either MetaBooster or MetaBooster-HiFi pipeline outperforms Canu and metaFlye, as the state-of-the-art strain-aware de novo metagenome assembly tools on most datasets, in one or more of the most relevant categories, such as genome fraction (coverage of individual strains), contig contiguity and error rate. As a general trend, MetaBooster-HiFi appears to outperform MetaBooster on higher-quality data such as PacBio CLR data. MetaBooster, on the contrary, shows better performance on lower-quality data such as ONT reads. Note that the evolution of nanopore sequencing technologies is fast, with the most recently published platform being able to generate (long) reads of error rates as low as < 5 % . This motivates the parallel development of the MetaBooster (for today’s low-quality reads) and MetaBooster-HiFi (future long reads of higher quality), providing flexibility relative to future progress in sequencing technology. In rather unclear situations, a guideline for real practice is to run both pipelines and opt for the better one relative to the practically relevant criteria one can determine.

We also evaluated Strainberry as a post-hoc strain resolver that has the potential to enhance assemblies further. Clearly, Strainberry’s performance depends on the assemblies that one feeds into it, which are generated by independent software, where Canu and metaFlye were suggested by the developers of Strainberry. Using MetaBooster or MetaBooster-HiFi as the assembler whose output is used for Strainberry, one can achieve further improvements over the assembly strategies originally suggested by Strainberry.

Overall, while exhibiting various favorable features, a conclusive insight into the why and how of how advantages (and on rare occasions also disadvantages) is achieved by the “booster” pipelines has not yet been reached. In future work, we are planning to study how to reach optimal synergy when crafting strain aware metagenome assembly pipelines from approaches presented in the literature.