The study team derived lineage markers from a maximum-likelihood phylogeny of 26,813 Mycobacterium tuberculosis complex genomes that contained 609,003 polymorphic sites. That tree placed the 10 human-adapted lineages (L1–L10) and 4 animal-adapted clades (A1–A4) on a single evolutionary backbone and exposed underrepresentation in several groups. The authors mapped substitutions to internal branches and selected single-nucleotide polymorphisms on the stem branch leading to each monophyletic lineage or clade. After filtering out 3,039 missense variants and 454 homoplastic nonmissense variants, 2,145 of 5,638 candidate SNPs remained as lineage-defining markers. Those markers accounted for roughly one-third to nearly half of total SNPs across individual lineages, lending internal consistency to the marker-selection strategy. The investigators also compared marker overlap with other typing tools and reported variable overlap; for example, 698 markers (62.8%) overlapped with TB-Profiler and 1,750 markers (12.9%) overlapped with SNP-IT. The authors cautioned that lineages with few genomes (L7–L10 and A1–A3) might contain unsampled diversity that these markers do not yet represent.
Pathotypr combines three complementary approaches for lineage assignment. The Train and Predict modules implement a k-mer–based Random Forest model trained on the marker set. Predict assigns a lineage to a genome assembly using that k-mer model. Classify makes traceable SNP-based calls on assemblies by following lineage-defining marker paths. Split-FASTQ applies the SNP marker framework directly to short-read data, assigning lineages without full alignment. The training cohort and test comparisons were explicit: Predict matched the phylogenetic lineage in all 498 complete RefSeq genomes used for evaluation, and Classify produced the same call for 497 of those genomes; the single unassigned genome lacked a complete marker path, which the authors attribute to possible assembly gaps or errors at diagnostic marker positions. Predict and Classify returned identical major-lineage calls for all 254 chromosome-level assemblies withheld from Random Forest training. Split-FASTQ assigned expected lineage labels to all 162 clinical read datasets tested. The study team interpreted these results as demonstrating distinct, practical roles for each module—Classify offers SNP-traceable calls, Predict recovers assignments when marker paths are incomplete, and Split-FASTQ extends the framework directly to sequencing reads.
The authors developed a Match Ref module to select a lineage-aware reference genome for read mapping. They reasoned that reference choice affects genome coverage and variant calling. In 162 clinical read sets, mapping to the reference chosen by Pathotypr reduced nonfixed SNPs by a median of 89% relative to mapping against the MTBC ancestral reference. Mapping to the selected reference also increased median coverage breadth by 0.84 percentage points. Those metrics indicate that lineage-aware reference selection can tighten variant calls and modestly improve breadth of coverage, according to the reported comparisons and figures in the Appendix.
To test surveillance-scale performance, the investigators applied Pathotypr’s Split-FASTQ module to 88,071 UShER-TB samples. The authors reported that the run required approximately 24 hours using 4 threads. For direct comparisons with TB-Profiler, they excluded 1,290 genomes that returned multiple-lineage calls. Despite limited marker overlap with TB-Profiler (only 698 shared markers), both tools assigned the same major lineage to every directly comparable sample in this dataset. Pathotypr also identified isolates from lineages and clades that TB-Profiler did not flag: specifically, Pathotypr found 2 L10 isolates, 6 A1 isolates, and 23 A2 isolates that TB-Profiler either misclassified or left unassigned. The authors reported that TB-Profiler misclassified one L10 isolate as La1.5/M. bovis and left several others without lineage assignments. In a multi-tool comparison summarized in the article’s table, Pathotypr was described as the only tool among those compared to combine rapid k-mer/Random Forest classification (≈1 second per call), support for both FASTA and FASTQ inputs, drug-resistance prediction, desktop and command-line interfaces, and the broadest lineage coverage (L1–L10 and A1–A4) while maintaining sublineage resolution.
The tool returns World Health Organization–cataloged resistance-associated variants alongside lineage calls; the study team applied Pathotypr to 7,148 genomes from the CRyPTIC collection to evaluate resistance-informed surveillance analyses. The article includes figures showing performance of WHO confidence group 1–2 resistance markers against phenotypic drug-susceptibility data and maps of probable introductions between source and destination regions, but the main text and abstract do not provide additional numeric performance metrics beyond those visual summaries. The authors therefore reported resistance-calling comparisons and downstream transmission inferences, and they referenced the Appendix for full methods and data tables.
The investigators acknowledged several constraints. First, some lineages and animal clades remain sparsely sampled—specifically L7–L10 and A1–A3—and the marker set may miss diversity not represented in the training collection. Second, one unassigned RefSeq genome illustrated that assembly gaps or errors at marker positions can prevent SNP-path–based calls. Third, the publication is an early-release article; supplementary tables and figures in the Appendix contain expanded results and methodological detail that readers should consult for implementation. The authors did not report additional trial parameters, such as memory usage, exact Random Forest hyperparameters in the main text, or full numerical breakdowns of resistance-call performance in the abstract; those elements appear in the Appendix and referenced datasets. Remaining questions include how expanding representation for under-sampled lineages will alter marker selection and whether the tool’s performance will change as additional, diverse genomes are incorporated into training sets.