Background

The Mycobacterium genus comprises more than 190 species and subspecies, including the Mycobacterium tuberculosis complex (MTBC) and non-tuberculosis mycobacteria (NTM) [1]. Tuberculosis (TB), a bacterial infection caused by MTBC, is one of the leading causes of death worldwide. Millions of newly diagnosed TB cases and 1.6 million TB-related deaths were reported in 2021 [2]. The effective treatment and management of TB is challenging owing to the emergence of multidrug-resistant TB. Globally, the treatment success rate for patients newly diagnosed with TB is 86%, whereas that for patients diagnosed with multidrug-resistant TB is only 60% [2]. Thus, timely detection of drug resistance is crucial to guide treatment options and prevent further transmission [3].

In addition to MTBC, NTM has gained recognition as an important human pathogen, as its incidence and prevalence continue to increase worldwide [4, 5]. NTM can be broadly categorized into two groups (rapidly- and slow-growing mycobacteria), of which Mycobacterium avium complex (MAC) and Mycobacterium abscessus complex are the most important species frequently isolated from patients with NTM infection [5, 6]. MTBC and NTM share similar microbiological properties and can lead to infections with overlapping clinical symptoms. However, they exhibit distinct disease characteristics and respond to different treatment options, and their susceptibility patterns to antimicrobial drugs vary depending on the species causing the infection [4].

Culture-based methods have traditionally been used as the gold standard for bacterial identification and diagnosis of drug-resistant TB. However, they are labor-intensive, require specialized infrastructure, and can take weeks to months [7]. Although molecular-based methods, such as GeneXpert and line probe assays, are more rapid than culture-based methods, they only target the most common resistance-associated variants (RAV) for a limited number of drugs [7, 8]. Whole-genome sequencing (WGS) is a useful tool for species identification, drug resistance prediction, and transmission tracing [9, 10]. In particular, WGS not only allows for the screening of well-known RAVs but also presents opportunities to uncover novel genetic alterations for both new and repurposed drugs. For example, WGS identified rplC p.Cys154Arg as a dominant mutation in linezolid-resistant isolates [11]. Mutations in mmpR5 (Rv0678), atpE, and pepQ, and those in ddn, fbiA, fbiB, fbiC, fbiD, and fgd1 have also been reported as potential mechanisms underlying bedaquiline and delamanid resistance, respectively [12, 13]. More recently, the World Health Organization (WHO) published the first comprehensive catalogue of mutations in MTBC, applying rigorous classification criteria to assign a confidence grade to each variant associated with drug resistance [10].

Despite the many advantages of WGS, its application in routine clinical settings has been limited primarily owing to the requirement of bioinformatics expertise and high-performance computing systems [14]. In addition, the lack of expertise in command line bioinformatics among biologists has hinder the widespread use of WGS data. To overcome these challenges, it is necessary to develop analytical tools that meet the following functionalities: 1) identification of both MTBC and NTM at the species level; 2) prediction of drug-resistant TB based on an endorsed knowledge base, such as the WHO mutational catalogue; 3) user-friendliness and low system requirements, including graphical user interface (GUI) or web-based software, and 4) data quality control and visualization. Given that mutation-calling algorithms can yield inaccurate results depending on the quality of the read and mapping accuracy, visualizing and reviewing detected mutations before the final report is crucial [15]. Although several tools for analyzing mycobacterial genomes have been developed in recent years, including KvarQ, PhyResSE, Mykrobe, TBProfiler, ReSeqTB-UVP, and SAM-TB [16,17,18,19,20,21], software that integrates all these functionalities is scarce (Additional file 1: Table S1). Additionally, variations in drug resistance prediction results can occur between software packages owing to differences in the analysis pipelines and knowledge bases.

In this study, we developed GenoMycAnalyzer, a web-based software program that integrates functions for the identification of MTBC and NTM species and their lineages and spoligotype prediction, variant calling, annotation, drug resistance prediction, and visualization. The performance of the GenoMycAnalyzer pipeline was evaluated using publicly available WGS data.

Implementation

Quality control of raw sequence reads

Sequencing adapters and low-quality bases were trimmed using Cutadapt version 4.2 [22]. Sequencing reads with a Phred base quality score greater than 20 and lengths longer than 50 bp were retained. The base quality of the fastq files before and after trimming was assessed using FastQC version 0.11.9 [23]. To remove the sequencing reads originating from the host genome, the trimmed reads were aligned to the human reference genome (hg38) using BWA MEM version 0.7.17 [24]. The mapping status, including the total number of bases and reads, duplicate rate, mapping rate, and sequencing depth, was calculated using SAMtools version 1.16.1 [25], and only unmapped reads were extracted for downstream analyses.

Species identification and spoligotyping

The identification of MTBC and NTM at the species level was performed using Kraken2 version 2.1.2, a taxonomic classification system based on k-mer matches [26]. To improve species predictions, a custom-built database containing 107 sequences from 18 NTM species in addition to the Kraken2 built-in database version 2023–8-18 was constructed (Additional file 1: Table S2). Subsequently, Braken software was used to estimate the abundance of species using taxonomy labels assigned by Kraken2 [26]. The most abundant species estimated by Braken was assigned to the species in the sample. Samples with fewer than 200 sequence reads mapped to the Mycobacterium genus or with a mapping rate of less than 10% to a predicted Mycobacterium species were excluded from further analyses. For MTBC, SpoTyping version 2.1 was used to determine the spoligotype from trimmed sequence reads [27]. The output is presented as an octal code.

Variant calling and lineage prediction

For isolates identified as MTBC, sequencing reads were mapped to the H37Rv reference genome (GenBank accession number: NC_000962) using BWA MEM, as previously described [28]. Duplicate reads were marked and de-duplicated using SAMtools version 1.16.1 [25]. Single-nucleotide variants (SNVs) and small indels were identified using BCFtools version 1.16 with the ‘ploidy’ option set to 1 and the ‘consensus-caller’ model [29]. SnpEff version 5.1d was used to define mutations in genomic sequences and predict their functional consequences [30]. To obtain a reliable and robust mutation calling, the following variants were eliminated: (i) read depth < 30, (ii) Phred quality score < 30, (iii) mapping quality score < 30, and (iv) variant allele frequency (VAF) < 5%. Lineages were predicted based on a previously reported single-nucleotide polymorphisms (SNP) barcoding assay using an in-house Python script [31, 32].

Detection of a large deletion

A large deletion profile at the gene level (> 50 bp) was estimated using the WGS data. The target genes were limited to four (pncA, katG, gid, and ethA) known to be associated with drug resistance [33]. Large deletions were defined as regions encompassing at least 50 bp with a sequencing depth of less than three standard deviations from the mean or less than ten mapped reads [34]. For comparison, large deletions were also predicted using the DELLY software version 1.1.7 with default parameters [35]. All identified large deletions were manually inspected using the Integrative Genomics Viewer (IGV) browser version 2.15.5 [15].

Genotypic drug susceptibility testing

The annotated and filtered variants were classified into one of the following five groups according to the WHO mutational catalogue [10]: 1, associated with resistance; 2, associated with resistance-interim; 3, uncertain significance; 4, not associated with resistance-interim; and 5, not associated with resistance (neutral mutations). In brief, group 1 mutations were defined when the following five criteria were met: 1) a sum of resistant and susceptible isolates with a solo mutation ≥ 5, 2) a lower bound of the 95% confidence interval of the positive predictive value (PPV) conditional on being solo ≥ 25%, 3) odds ratio (OR) > 1, 4) OR of solo mutations > 1, and 5) false discovery rate-corrected P ≤ 0.05. Group 2 mutations were present as solo in pncA in at least two resistant isolates with a PPV of ≥ 50%, whereas group 4 mutations were present as solo in pncA with a PPV ≤ 40% and an upper bound of the 95% confidence interval of ≤ 75%. Group 3 comprised mutations that did not meet the criteria for inclusion in groups 1, 2, 4, or 5. We defined genotypic resistance to 15 antituberculosis drugs (isoniazid, rifampicin, ethambutol, pyrazinamide, levofloxacin, moxifloxacin, linezolid, bedaquiline, clofazimine, delamanid, amikacin, capreomycin, kanamycin, streptomycin, and ethionamide) as SNPs or indels overlapped with group 1 or 2 variants of the WHO mutational catalogue at either the nucleotide or amino acid level (Additional file 1: Table S3). In addition, large deletions of pncA, katG, gid, and ethA were considered resistant to pyrazinamide, isoniazid, streptomycin, and ethionamide, respectively. Variants falling into groups 3, 4, and 5 in the WHO mutational catalogue or mutations that did not meet the criteria outlined in the WHO mutational catalogue were considered genotypically susceptible. Detected RAVs were visualized in a Circos plot with the sequencing depth of the corresponding gene [36].

Statistical analysis

The sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were determined. The receiver operating characteristic (ROC) curve and area under the curve (AUC) were used to compare the predictive values of the WHO mutational catalogue and GenoMycAnalyzer datasets. All statistical analyses were performed using SPSS version 29 (IBM Corp., Armonk, NY, USA).

Dataset collection and availability of the GenoMycAnalyzer

The inclusion criteria encompassed Mycobacterium spp. meeting the following conditions: isolates not used in the training of the WHO mutational catalogue, isolates with sequencing depth of 100 or more, and isolates with species information or pDST results. The pDST measurement method was not considered. According to these criteria, we collected a dataset of 8,259 Illumina raw sequence reads for which laboratory-based metadata were available from public datasets. Raw sequence data were downloaded from the European Nucleotide Archive (ENA), Sequence Read Archive (SRA), or DDBJ Sequence Read Archive. Among the datasets, those of 1,284 isolates were used to evaluate the accuracy of species identification, and those of 5,473 isolates with pDST results for at least one antituberculosis drug were used to evaluate the performance in predicting drug resistance. In addition, the datasets of 1,935 and 54 isolates were used to evaluate the accuracy of lineage and spoligotype predictions, respectively. Therefore, the data used for validation of each functionality comprise different datasets. Accession numbers and corresponding metadata are listed in Additional file 2: Table S4.

The GenoMycAnalyzer is accessible at https://www.mycochase.org/. This software is compatible with Windows, macOS, and Linux operating systems and is released under the GNU General Public License 3.0. The manual for the software can be found via the “help” tab on the GenoMycAnalyzer website and is accessible without the need to register or log in. Additionally, the source code for the data processing pipeline is available in the GitHub repository at https://github.com/IRCGP-Lab/GenoMycAnalyzer_Source.

Data encryption and privacy

The GenoMycAnalyzer server encrypts all transmissions using the Fernet module from the cryptography package (https://github.com/pyca/cryptography). Fernet employs the Advanced Encryption Standard (AEC) algorithm for encryption, ensuring that encrypted data cannot be accessed without the key. This provides users with robust, secure communication with the web server.

Results

Pipeline overview

GenoMycAnalyzer is a freely available web-based pipeline for users to analyze Mycobacterium genomes. The unique features of the GenoMycAnalyzer are listed in Table 1. The GenoMycAnalyzer consists of five sequential steps: 1) pre-processing and quality control, 2) species identification and molecular typing, 3) variant calling and annotation, 4) gDST, and 5) visualization and report generation (Fig. 1). GenoMycAnalyzer ingests single-end or paired-end fastq files generated by Illumina instruments as input, and supports the multiple file uploads and batch analyses. In the pre-processing step, GenoMycAnalyzer removes sequencing adaptors, low-quality bases, and sequencing reads originating from the host genome. Subsequently, the filtered reads are classified to the species level using a custom-built database that includes both MTBC and NTM sequences. If the isolate is predicted to be MTBC, GenoMycAnalyzer is used to perform subsequent analyses, such as lineage and spoligotype prediction, variant calling, annotation, and RAV identification. Notably, detected variants can be interactively inspected using the IGV browser [15], which helps determine the false-positive or false-negative calls. Finally, the analysis results are summarized in a report and downloaded in PDF file format. The run time for a sample with a sequencing depth of 300-fold is approximately 15 min. The uploaded data and processed files are retained on the server for three months. The GenoMycAnalyzer is available at http://www.mycochase.org.

Table 1 Features of GenoMycAnalyzer
Full size table
Fig. 1
figure 1

GenoMycAnalyzer pipelines. A The GenoMycAnalyzer platform includes five modules: pre-processing, identification, variant calling, genotypic DST, and visualization and report. (B-D) Example of WGS analyzed by GenoMycAnalyzer. B Partial result of the 'Analysis Summary' is displayed. The strain is identified as M. tuberculosis, with lineage (2.2.1) and spoligotype (000000000003771). Genotypic DST for 15 types of antimicrobial drugs is color-coded (resistant: red, susceptible: green). A Circos plot represents the sequencing coverage depth and detected variants for 17 genes harboring group 1 or 2 mutations. C Among the detected variants, overlapped variants with the WHO mutational catalogue are shown in ‘RAV (resistance-associated variants)’. D The reads mapping status of gyrA p.D94G, which confers resistance to fluoroquinolone, is shown in IGV implemented in GenoMycAnalyzer

Full size image

SNP-based genotypic DST performance of GenoMycAnalyzer

The GenoMycAnalyzer knowledge database comprises the 13,221 unique mutations reported in the WHO mutational catalogue, including 1,149 group 1 and 2 mutations known to confer drug resistance [10]. To evaluate the predictive performance of the GenoMycAnalyzer for gDST, we analyzed 5,473 MTBC genomes with pDST results for at least one antituberculosis drug. All isolates used in this evaluation were independent datasets not included in the training data of the WHO mutational catalogue. The mean coverage of the sequencing depth was 193.6x, with an average of 97.6% of bases covered by at least 50 reads in each isolate. The completeness of the pDST results differed depending on the drug, with the highest rates observed for the first-line drugs, rifampicin (94.4%), isoniazid (94.2%), ethambutol (84.0%), and pyrazinamide (60.4%) (Additional file 2: Table S4). The pDST data for second-line drugs ranged from 59.6% (streptomycin) to 1.9% (levofloxacin). Among the 5,473 isolates, 3,094 (56.5%) were pan-susceptible, 1,599 (29.2%) were multidrug-resistant TB (MDR-TB), and 77 (1.4%) were extensively drug-resistant TB (XDR-TB). The remaining 703 isolates (12.8%) were resistant to at least one drug. Four drugs were excluded from the evaluation because of the lack of RAVs (bedaquiline and clofazimine) or phenotypically resistant isolates (linezolid and delamanid).

Overall, the GenoMycAnalyzer gDST for first-line drugs exhibited excellent predictive values (Fig. 2 and Table 2): 95.9% sensitivity and 97.3% specificity for rifampicin, 91.0% sensitivity and 97.7% specificity for isoniazid, and 89.9% sensitivity and 91.9% specificity for ethambutol. The sensitivity for pyrazinamide was lower than that for other first-line drugs (79.6% sensitivity and 96.2% specificity), which is consistent with a previous report [10]. The predictive values for second-line drugs varied for each drug, with sensitivities ranging from 98.2% (levofloxacin) to 74.9% (capreomycin) and specificities ranging from 98.7% (amikacin) to 79.5% (ethionamide) (Fig. 2 and Table 2). The sensitivity for assigning MDR-TB and XDR-TB was 88.2% (95% CI 86.5–89.7%) and 83.1% (95% CI 72.9–90.7%), respectively.

Fig. 2
figure 2

Genomic variants associated with 11 antimicrobial drugs. Genomic alterations and phenotypic DST to rifampicin, isoniazid, ethambutol, pyrazinamide, levofloxacin, moxifloxacin, amikacin, capreomycin, kanamycin, streptomycin, and ethionamide are shown. Each row and column represents the resistance-associated gene and strain, respectively

Full size image
Table 2 Performances of GenoMycAnalyzer for genotypic drug susceptibility testing
Full size table

Next, we calculated the differences in AUC between the GenoMycAnalyzer and WHO mutational catalogue datasets and found that the predictive values for pyrazinamide, levofloxacin, and kanamycin were significantly better than those reported in the WHO mutational catalogue (Additional file 1: Table S5). In contrast, the predictive values for streptomycin and ethionamide were poorer than those of the WHO mutational catalogue. The predictive values for the remaining six drugs were not significantly different from those reported in the WHO mutational catalogue, suggesting the non-inferiority of the GenoMycAnalyzer (Additional file 1: Table S5).

Discordance analysis in resistance predictions

To gain insights into the discrepancies between pDST and gDST, we examined RAVs for rifampicin and isoniazid, for which genotypic resistance mechanisms are well understood. To this end, we further analyzed isolates for which GenoMycAnalyzer predicted resistance while pDST was reported as susceptible (false positive; FP), and vice versa (false negative; FN) using TBProfiler (Additional file 3: Table S6) [19]. GenoMycAnalyzer called 74 FP predictions for isoniazid, and all of them had RAVs corresponding to group 1 or 2 mutations: katG p.S315T (n = 24), fabG1 c.-15C > T (n = 22), fabG1 c.-8 T > C (n = 14), fabG1 p.L203L (n = 7), fabG1 c.-8 T > A (n = 3), katG p.P429fs (n = 2), katG p.199 fs (n = 1), and katG p.Q471* (n = 1). Notably, all variants were consistently detected using TBProfiler except for two isolates. Similarly, all 101 FP predictions for rifampicin by GenoMycAnalyzer had RAVs corresponding to group 1 or 2 mutations, of which 88 (87.1%) were concordant with those of TBProfiler. The most frequent RAV for rifampicin FP predictions were rpoB p.S450L (n = 26) and p.I491F (n = 26), followed by p.L452P (n = 12), p.L430P (n = 8), and p.H445N (n = 5).

Among the 208 FN predictions for isoniazid, 130 isolates (62.5%) were identified as resistant, whereas the remaining 78 isolates (37.5%) were consistently predicted as susceptible (false negative) by TBProfiler (Additional file 3: Table S6). Notably, most variants detected in isolates predicted to be resistant by TBProfiler were also detected using GenoMycAnalyzer (n = 105, 80.8%); however, they were group 3 (uncertain significance) mutations (n = 98) or mutations not included in the WHO mutational catalogue (n = 7). Overall, 88.0% of the isoniazid FN predictions were concordant with TBProfiler. Regarding rifampicin, 62 of the 73 FN predictions made by GenoMycAnalyzer (84.9%) were concordant with those made by TBProfiler.

Detection of large deletions increases the sensitivity of gDST

GenoMycAnalyzer identified 25 large deletions in katG (0.5%), 58 in pncA (1.8%), 69 in gid (2.1%), and nine in ethA (0.8%). In contrast, when large deletions were predicted using DELLY, the number of large deletions detected was higher than that identified using the GenoMycAnalyzer: 114 in katG (2.2%), 36 in pncA (1.1%), 90 in gid (2.8%), and 33 in ethA (2.9%). Most large deletions detected by the GenoMycAnalyzer (63.0%) overlapped with those detected by DELLY, whereas 60 large deletions were specific to the GenoMycAnalyzer (Fig. 3A). Manual inspection using IGV confirmed that all deletions detected by the GenoMycAnalyzer were true positives; however, only 35.1% of the DELLY-specific deletions were true positives (Fig. 3B-D and Additional file 4: Table S7). Comparisons of large deletions with the pDST results revealed that GenoMycAnalyzer had a PPV of 100% for isoniazid, 93.1% for pyrazinamide, 79.7% for streptomycin, and 44.4% for ethionamide, whereas DELLY had a PPV of 71.9% for isoniazid, 83.3% for pyrazinamide, 75.6% for streptomycin, and 24.2% for ethionamide (Additional file 1: Table S8). The sensitivity of the GenoMycAnalyzer gDST, including large deletions, improved to 92.0, 84.5, 83.3, and 80.2% for isoniazid, pyrazinamide, streptomycin, and ethionamide, respectively.

Fig. 3
figure 3

Evaluation of large deletions detected by GenoMycAnalyzer and DELLY. A Venn diagram of the large deletions detected by GenoMycAnalyzer and DELLY. Large deletions confirmed as true positive by manual inspection are displayed in gray. B-D Examples of large deletions detected in gid (B), pncA (C), katG (D), and ethA (E). In each example, the top (1), middle (2), and bottom (3) panels represent large deletions detected by GenoMycAnalyzer only, both tools, and DELLY only, respectively. The x-axis represents genomic position, and the y-axis represents sequencing depth. The black arrows indicate the breakpoint of a large deletion. GenoMycAnalyzer detects large deletions at the gene level. Accession numbers for the B-1 to E-3 are SRR958195, ERR718365, SRR6824567, SRR6824340, SRR6824578, ERR040134, SRR6824287, ERR040137, ERR867540, SRR924218, ERR038737, and SRR6824300, respectively

Full size image

Species identification performance of GenoMycAnalyzer

To identify MTBC and NTM at the species level, GenoMycAnalyzer was integrated with the Kraken2 and Braken software [26]. To evaluate the accuracy of species identification, we analyzed 1,284 mycobacterial genomes, including 69 isolates of MTBCs and 1,215 isolates of 39 NTM species [20, 37]. The GenoMycAnalyzer reliably distinguished between MTBC and NTM at the species level, achieving an accuracy of 98.8% (1,268 concordant results out of 1,284 isolates), whereas the accuracy of the Kraken2 built-in database was 90.0% (1,156 concordant results out of 1,284 isolates) (Additional file 1: Table S9). Furthermore, the GenoMycAnalyzer successfully identified isolates that had not previously been identified at the species level in the NCBI database. This included an isolate labeled M. avium complex sp. CF00315-00498, in which the GenoMycAnalyzer was identified as M. marseillense. However, at the subspecies level, the identification performance of GenoMycAnalyzer was significantly lower than its species-level identification performance (508 concordant results out of 803 isolates; 63.3%) (Additional file 1: Table S10). For example, of the 494 isolates reported as M. abscessus subsp. abscessus, 296 (59.9%) were identified as the same subspecies, whereas the remaining 198 isolates were identified as M. abscessus subsp. massiliense within the same species. Similarly, three out of the 37 isolates reported as M. intracellulare subsp. chimaera were identified as M. intracellulare. Although incorrect classifications at the subspecies level were identified within the same species (Additional file 1: Table S10), these results suggest that GenoMycAnalyzer subspecies identification may not be optimal. Therefore, GenoMycAnalyzer reports MTBC and NTM at the species level without further subspecies identification.

Lineage and spoligotype predictions

The lineage of MTBC isolates can be determined based on SNP differences [31, 32]. To evaluate the accuracy of the lineage-prediction module implemented in GenoMycAnalyzer, we analyzed 1,935 MTBC isolates with phylogenetically determined lineages [31]. A total of 84 lineages/sub-lineages were identified, all of which were concordant with previously reported lineages (Additional file 5: Table S11).

GenoMycAnalyzer predicts spoligotypes based on 43 spacer sequences of the direct repeat locus in the MTBC genome. We analyzed 54 MTBC isolates with experimentally determined spoligotypes [38] and identified 20 different spoligotypes. Notably, spoligotype predictions using the GenoMycAnalyzer were concordant with the laboratory-determined spoligotypes for all isolates, resulting in an accuracy of 100% (Additional file 1: Table S12).

Discussion

In this study, we present a high-confidence automated bioinformatic tool we dubbed the GenoMycAnalyzer. We present extensive validation of this tool, which analyzes MTBC WGS data to provide gDST for 15 antituberculosis drugs using the WHO mutational catalogue [10] and large deletions [33]. GenoMycAnalyzer also offers species-level identification of MTBC and NTM, prediction of the MTBC lineage and spoligotype, and visualization of the detected variants. In a clinical setting, WGS data obtained after patient sample processing could be automatically processed through this pipeline which would present a user-friendly interface and deliver results in standard lab format.

The WHO mutational catalogue is the most comprehensive knowledge base of mutations associated with MTBC drug resistance and assigns a confidence grade to one of five groups for each variant [10]. Group 1 and 2 mutations present in the catalogue accurately predicted resistance and strongly correlated with pDST for most drugs [10], and GenoMycAnalyzer reports strains harboring group 1 or 2 as resistant. One of the limitations of the WHO mutational catalogue is that it has not been validated using an independent dataset [10]. Using the GenoMycAnalyzer, we confirmed the excellent discriminative ability of drug resistance for most antituberculosis drugs, even in an independent large dataset. Specifically, the AUC of GenoMycAnalyzer showed equivalent (rifampicin, isoniazid, ethambutol, moxifloxacin, amikacin, and capreomycin) or better (pyrazinamide, levofloxacin, and kanamycin) prediction performances than those reported by the WHO mutational catalogue, thus confirming the non-inferiority of the GenoMycAnalyzer.

In some cases, it is noteworthy that mutations reported in the WHO mutational catalogue and those detected in this study overlap at the amino acid level but are discordant at the nucleotide level. For example, nucleotide change of embB p.Met306Ile conferring ethambutol resistance is ‘c.918G > A’ in the WHO catalogue, whereas ‘c.918G > C’ causing the identical p.Met306Ile was recurrently detected in our ethambutol-resistant isolates. Likewise, katG c.944G > C (p.Ser315Thr), rpsL c.128A > G (p.Lys43Arg), and rpoB c.1349C > T (p.Ser450Leu) identified in this study overlapped at the amino acid level, but were discordant at the nucleotide level. This observation could be related to the inherent redundancy in the genetic code, suggesting that clinicians assessing or diagnosing gDST using the present version of the WHO mutational catalogue should exercise caution to prevent misdiagnoses.

Large deletions lead to the loss of genetic material, can disrupt protein function, and are sporadically observed in drug resistance-related genes such as katG and pncA [39, 40]. Using WGS data, GenoMycAnalyzer successfully detected large deletions with varying frequencies in katG (0.5%), pncA (1.8%), gid (2.1%), and ethA (0.8%). Notably, large deletions were mutually exclusive to resistance-conferring mutations. When large deletions were included, the gDST sensitivity to pyrazinamide and streptomycin increased by 4.9% and 5.4%, respectively. In addition, more than half of the isolates with pncA large deletions belonged to the East Asian lineage (lineage 2, 38 of 58 isolates with large pncA deletion, 65.5%), consistent with previous findings [28]. These results suggest that the use of WGS data in clinical settings may provide advantages for diagnosing drug-resistant MTBC by simultaneously detecting large deletions and resistance-conferring variants that are difficult to detect using conventional PCR.

Identifying NTMs at the species level is important for effective treatment [41]. Kraken2 is effective in identifying mixed NTM infections, with the highest sensitivity and specificity among the tested analytical tools [42,43,44]. However, species prediction by Kraken2 using a built-in database revealed that only 90% of the tested isolates matched their known species, and most of the mismatches were due to the absence of NTM sequences within the database (Additional file 1: Table S9). These include M. colombiense, M. asiaticum, and M. malmoense, which cause human infections [45,46,47]. In contrast, GenoMycAnalyzer uses a custom-built database that adds 107 NTM sequences, and its species identification accuracy is as high as 98.8%. The GenoMycAnalyzer has two additional functionalities that are important for mycobacterial analysis. First, the GenoMycAnalyzer includes a molecular typing function, and large-scale validation confirmed accurate lineage and spoligotype predictions. These features may be useful for epidemiological surveillance of mycobacteria [31, 48]. Second, the GenoMycAnalyzer integrates the IGV browser to visualize the detected variants. Manual inspection using IGV may allow investigators to assess variants efficiently, serving as a potent method for validating variant calls and reducing false positives while confirming true findings [15].

An inherent limitation of this study is that the gDST predictions between software programs may be inconsistent. These discrepancies could arise from the cohort used for the analysis, incorrect pDST data, lack of standardization of the bioinformatics pipeline, or different knowledge databases. Given that most of the differences between GenoMycAnalyzer and TBProfiler results were due to group 3 variants, further evaluation of group 3 mutations is required. In addition, we did not evaluate the four drugs because of a lack of resistance-conferring mutations or phenotypically resistant isolates. Further studies in a larger cohort are required to discover novel resistance-conferring mutations and elucidate the underlying resistance mechanisms. GenoMycAnalyzer includes a function for users to edit the knowledge database, which serves as a valuable bridge until regular database updates occur. Finally, although the GenoMycAnalyzer accurately distinguished clinically important mycobacteria at the species level, further efforts are needed to improve the accuracy of subspecies identification.

Conclusions

GenoMycAnalyzer software provides flexible and rapid analysis of WGS data from the Illumina platform to predict species, drug resistance, and molecular profiles with high accuracy. GenoMycAnalyzer also offers both web-based and GUI interfaces, which can help biologists with limited access to high-performance computing systems or limited bioinformatics skills. Ultimately, by streamlining the interpretation of WGS data, the GenoMycAnalyzer has the potential to significantly impact TB management and contribute to global efforts to combat this infectious disease.

Availability and requirements

Project name: GenoMycAnalyzer.

Project home page: https://www.mycochase.org/

Operating system(s): Windows, Linux, MacOS.

Programming language: Python, SQL, JavaScript, HTML.

Other requirements: None.

License: GNU General Public License.

Any restrictions to use by non-academics: None.