Next-generation sequencing (NGS) technologies play a fundamental role in studying microbial genomes (Köser et al., 2014). Nowadays, the whole-genome sequencing (WGS) of pathogens and viruses is routinely exploited in epidemiological outbreak analysis (Ferdinand et al., 2021), to identify and characterize bacterial pathogens and transmission chains. Recently, WGS has emerged as a powerful tool that could help in the battle of the spread of antibiotic resistance for different species (Gladstone et al., 2021). Tuberculosis still constitutes one of the most serious threats to human health, killing nearly 1.5 million of people per year (World Health Organization, 2021a). The higher accuracy of short-reads technology (SRS), such as Illumina, together with the use of a catalog of Mycobacterium tuberculosis (MTB) mutations to interpret drug resistance determinants has significantly improved the interpretation of clinical genomes (Ektefaie et al., 2021; Walker et al., 2022). The same technology has been used to investigate tuberculosis outbreaks and transmission dynamics by adopting whole-genome SNP (wgSNP) or core genome Multi-Locus Sequence Typing (cgMLST) schemes assessing genetic relatedness of MTB genomes (Kohl et al., 2014, 2018). However, short-reads technologies are not able to fully resolve hard-to-sequence regions, because has suboptimal capacity to resolve reliably large structural variations, gene duplications, or variations in repetitive regions (Modlin et al., 2021), thereby reducing coverage depth involving a lack of characterization in terms of drug resistance, virulence, and transmission analysis (Medha et al., 2021; Marin et al., 2022). Accurately resolving such regions becomes critical to close bacterial genomes, obtaining more information about virulence, evolutionary mechanisms of drug resistance, and on strain relatedness. The availability of long reads (LRS) from third-generation sequencing technologies, e.g., Oxford Nanopore Technologies (ONT) or PacBio, can improve the resolution of bacterial genomes at level of gene rearrangement, repetitive regions (proline-glutamate/proline-proline-glutamate, PE/PPE), and long insertions/deletions (InDel), usually neglected by short-read sequencing (SRS) due to their low-complex nature. Notably, ONT is a portable, robust, and low-capital-cost sequencer that could conceivably be utilized to conduct WGS analysis in a rapid manner. Recently, different bioinformatic pipelines have been developed to implement the advantages of SRS and LRS in a single unique approach (Walker et al., 2014; Wick et al., 2017, 2021b). The procedure usually involves using first SRS to make de novo assembly and then LRS to build the bridges between the ambiguous regions, relying mostly on the SRS steps. In this work, we aim to compare the performances of SRS, LRS, and hybrid approach on MTB clinical cluster isolates, which are resistant to first- and second-line drugs. For this purpose, we implemented the use of “hybrid reads” (HYBR), in which we first corrected the long reads with high accurate short reads, and then we used them as input for the downstream analysis, including identification of mutations, drug resistance prediction, transmission analysis, de novo genome assembly, and overall genome coverage. Our reverse HYBR approach outperforms the standard hybrid pipeline. Moreover, we aimed to characterize the repetitive regions of the genome, including PPE and PE genes, which are normally neglected during the SRS analysis. The outcome from this analysis indicates that PE and PPE genes, except PE_PGRS, can be included in the SRS analysis at the cost of increasing the sequencing depth. The study was performed using a subset of M. tuberculosis strains previously characterized in our laboratory (Mustazzolu et al., 2018; Abascal et al., 2020; Villa et al., 2021).

The characterization of MTB strains shows different challenges associated with the aims of the genomic analysis. Several solutions were proposed over the years to optimize the analysis, with a major focus on the use of SRS (Kohl et al., 2018; Phelan et al., 2019). Differently from other prokaryotic pathogens, MTB shows genomic features such as the lack of mobile genetic elements (e.g., plasmids), a high GC-content, and the relevant presence of highly variable repetitive regions. All those features contribute to increase possible biases in the genomic analysis (Li and Wren, 2014; Modlin et al., 2021). Common bioinformatic pipelines usually exclude the so-called ‘biased regions’, assuming that large repeats cannot be mapped onto the reference genome unambiguously, as mappability does not depend on coverage, and the results could decrease the accuracy of the transmission analysis (Kohl et al., 2018; South et al., 2022). This compromises the possibility to identify mutations relevant for the virulence and resistance to the main drugs and to provide a comprehensive analysis of entire genomes.

The introduction of the LRS technology represents a valid alternative to the SRS approaches, because it allowed a better characterization of the MTB genome, e.g., InDel and repetitive regions. Comparing the results between SRS and LRS, different studies highlight genome regions where the SRS lacks accuracy due to limit of the technology (Modlin et al., 2021; Peker et al., 2021; Gómez-González et al., 2022; Marin et al., 2022). Although LRS approaches still present a high-error rate (~5–15%), their random nature allows to improve accuracy with higher coverage (Rhoads and Au, 2015; Athanasopoulou et al., 2021; Amoutzias et al., 2022).

In this study, we compared the performances of LRS, SRS, and our modified version of the hybrid long-short reads, HYBR, on 13 MTB strains previously described, showing MDR and preXDR patterns. We analyzed the coverage and the variant calling along the whole genome. We carried out a comparative analysis between the 3 approaches, by performing genome coverage estimation, cluster analysis, drug resistance detection, and de novo assembly. The results obtained showed that the implementation of the HYBR approach, which has the advantage to include the features of the long and short reads, allows a better description of the study strains in terms of genome breadth coverage and assembly compared to SRS, and variant calling and related downstream analysis compared to LRS. In fact, our hybrid approach relies on long reads first corrected by the short reads and then used them in the downstream analysis: this approach allows to adopt the newly hybrid corrected reads for all the tasks of the investigation, while usually hybrid approaches involve both LRS and SRS only for the assembly step. SRS showed several limitations in terms of coverage along the whole-genome compared to LRS and HYBR. The PE_PGRS genes regions resulted as the more problematic for SRS, although those families of repetitive genes retain an important role in terms MTB pathogenesis, and the low coverage could correspond to a not trivial loss of information in the pathogen characterization (de Maio et al., 2020). In particular, we found that PE_PGRS3 and PE_PGRS4 genes present very low coverage in all approaches. Recently, few studies characterized those specific regions in the genome, showing that they are close to each other and present a homologous sequence (percent identity of 81%) due to gene duplication, indicating that they could potentially present critical issues with every technology (Karboul et al., 2008; Phelan et al., 2016; de Maio et al., 2020). Interestingly, the remaining PE and PPE regions showed an overall acceptable coverage for SRS and as already described in other studies, the common practice of excluding those genes from the analysis, due to the high GC-content and the repetitive sequences, could be overcome by removing only the PE_PGRS genes (Modlin et al., 2021; Marin et al., 2022).

The variant calling showed how, with low depths, the high-error rate of the LRS technology masks the variants detected with the random noise produced by the basecalling step. Nevertheless, this issue could be addressed with an enhancement of the sequencing depth, differently from SRS technologies where the error is due to systematic biases (Cabibbe et al., 2020). The HYBR approach outperformed both SRS and LRS, the latter missing few mutations due to coverage issues in those regions. The hybrid reads approach requires a good sequencing depth from LRS otherwise it will inherit the same issues of the parental LRS in terms of signal/noise ratio, especially in those regions where SRS correction does not perform optimally. In fact, most of the undetected mutations were due to frequency threshold (75%), especially in those regions not well covered by SRS. Nevertheless, considering the repetitive regions, this result indicates that HYBR approach can reveal a great number of mutations compared to SRS and LRS, due a better coverage.

In the de novo assembly evaluation, the three approaches were compared among each other and to the widely used hybrid assembler Unicycler (de Maio et al., 2019; Wick et al., 2021a). As stated by the developer, the hybrid assembly executed by Unicycler corresponds to a “short-read-first” approach in which the short reads assembly graph is scaffolded to completion by the long reads (Wick et al., 2017). This approach was proposed with the assumption that LRS presents low depth and accuracy. The improvement of the ONT technology claiming to lower error rate at 1% with the introduction of the new 10.x flow cell chemistry, allowed to rely on the opposite “long-reads-first” approaches as Trycicler (Wick et al., 2021b). In the current comparison, the LRS still relies on the previous technology presenting low depth and accuracy. Nevertheless, the HYBR and the LRS showed the best results, confirming Flye as one of the best-performing assemblers for long reads (Wick et al., 2021a). Interestingly, in our dataset, the hybrid assembler Unicycler performed poorly than Flye, especially considering the NG50 metrics (the length of the shortest contig at 50% of the total genome length), presenting a mean of 1.6 ± 0.3 Mb, lower than the HYBR with 4.3 ± 0.1 Mb (p = 0.02), indicating that our HYBR can better assembly the genomes. As expected for this task, SRS performed very poorly emphasizing its inadequacy for the de novo assembly.

This study presents some limitations: the limited number of samples considered for the analysis despite the deep investigation conducted on each genome and the adoption of the 9.x flow cells technology for LRS bearing a higher error rate compared to the new 10.x as the latter was not available at the time of the study.

This study outlines the strengths and the weaknesses of three approaches. The repetitive regions of the PE_PRGS genes represent a source of blind spots for the SRS, while the remaining PE/PPE regions, usually neglected as well, could be safely included in the analysis, showing good coverage. The LRS shows issues in terms of signal-to-noise ratio but still can correctly identify genetically closed strains and drug resistance-associated mutations, and the increase of sequencing depth enables usually to fix the issue. The HYBR approach overcomes the limitations of both SRS and LRS, showing the best results in all the considered tasks. Although hybrid reads approach suffers from the relative higher cost compared to the single sequencing run of SRS and LRS, it could offer the advantage to better evaluate problematic regions in variant calling, where LRS presents critical issue, and in de novo assembly, where SRS cannot compete with LRS.

In conclusion, depending on the aim of the investigation, both SRS and LRS present complementary advantages and limitations implying that for a full resolution of MTB genomes, where all the mentioned analyses and both technologies are needed, the use of the hybrid reads approach represents a valid option and a well-rounded strategy (Figure 6).

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found at: https://www.ncbi.nlm.nih.gov/, PRJNA903660. The codes used for the analysis presented in the study are deposited in the Github repository, accessible at: https://github.com/Allen13x/MTB_LRSvsSRS.

FDM, AS, SB, AMC, and DMC conceived and supervised the study. SB and VB performed the sequencing. FDM and AS performed the bioinformatics analysis. FDM and AS wrote the draft manuscript. FDM, AS, AMC, and DMC revisioned the draft manuscript. All authors contributed to the article and approved the submitted version.

This study was partially supported by the 2nd ERANet-LAC Transnational Joint Call on Research and Innovation (grant: TRANS-TB-TRANS PER-2012-ELAC2015/T08-0664).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.