Early Release – Alignment-Free Tool for Assemblies and Reads in Genomic Surveillance of Tuberculosis – Volume 32, Number 11—November 2026 – Emerging Infectious Diseases journal

Figure 1

Figure 1. Development and implementation of the lineage framework for Pathotypr, an alignment-free tool for assemblies and reads in genomic surveillance of tuberculosis. A) Maximum-likelihood phylogeny of 26,813 Mycobacterium tuberculosiscomplex…

To extract lineage phylomarkers, we obtained a maximum-likelihood phylogeny of 26,813 MTBC genomes and 609,003 polymorphic sites, which placed L1–L10 and A1–A4 within 1 evolutionary backbone and exposed the marked underrepresentation of certain lineages (Figure 1, panel A). We then mapped substitutions to branches and selected single-nucleotide polymorphisms (SNPs) on the stem branch leading to each monophyletic lineage or clade. After removing 3,039 missense variants and 454 homoplastic nonmissense variants, 2,145 of 5,638 candidates remained as lineage-defining markers. A variable overlap of lineage markers occurred with other tools, ranging from no overlap to 698 (62.8%) with TB-Profiler (https://tbdr.lshtm.ac.uk) and 1,750 (12.9%) with SNP-IT (https://github.com/samlipworth/snpit) (Appendix Table 1). Pathotypr lineage-defining markers comprised 33%–46% of total SNPs across lineages, supporting the filtering approach (Figure 1, panel B). However, the small number of genomes for L7–L10 and A1–A3 might be missing diversity not sampled to date.

Figure 2

Figure 2. Schematic overview of the 5 modules grouped into 3 functional categories—machine learning, genotyping, and utilities—in developing Pathotypr, an alignment-free tool for assemblies and reads in genomic surveillance of tuberculosis. A)…

Pathotypr uses 3 modules for lineage assignment (Figure 2). The Predict module assigns lineages to genome assemblies using the k-mer–based Random Forest algorithm generated by the Train module, and the Classify and Split-FASTQ modules assign lineages on the basis of SNP markers to assemblies (Classify) and raw sequencing reads (Split-FASTQ). First, we compared lineage assignments by Predict and Classify results with the phylogenetic placement of 498 complete RefSeq genomes: Predict matched the phylogenetic lineage in all 498 genomes, and Classify assigned 497. The unassigned genome did not contain the complete marker path required for a call, which could reflect assembly gaps or errors at diagnostic marker positions. Next, Predict and Classify successfully returned the same major-lineage assignment for all 254 chromosome-level assemblies that were not used to train the Random Forest. Finally, Split-FASTQ assigned all 162 clinical read datasets to the expected lineage. Those results show the specific role of each module: Classify provides traceable SNP-based assignments, Predict offers an alternative when the marker path is incomplete, and Split-FASTQ extends the lineage framework directly to short reads (Appendix, Table 2, Figure 1).

Because the reference choice for read mapping can affect genome coverage and variant calls, we developed and tested Pathotypr’s fourth module, Match Ref, a lineage-aware reference selection (9,10). Across 162 clinical read sets, mapping to the reference selected by Pathotypr reduced nonfixed SNPs by a median of 89% relative to the MTBC ancestral reference and increased median coverage breadth by 0.84 percentage points (10) (Appendix Figure 2).

To test whether the new lineage framework could be applied to surveillance-scale datasets, we typed 88,071 UShER-TB samples (11; K.M. Karim et al., unpub. data, https://www.medrxiv.org/content/10.1101/2025.07.22.25331806v1) with Pathotypr module Split-FASTQ (≈24 hours using 4 threads). To compare typing results with TB-Profiler, we excluded 1,290 genomes with multiple lineage calls from the comparison. Although only 698 (62%) of the TB-Profiler lineage markers were shared with Pathotypr, both tools assigned the same major lineage to every directly comparable sample (Appendix Table 3). Pathotypr additionally identified 2 L10, 6 A1, and 23 A2 isolates that were not identified with the TB-Profiler marker set. TB-Profiler misclassified 1 L10 isolate as La1.5/M. bovis and left the other L10 genome and A1 and A2 genomes unassigned. Compared with 7 other MTBC genomic analysis tools, Pathotypr uniquely combined rapid k-mer/Random Forest classification (≈1 s), support for both FASTA and FASTQ file type inputs, drug-resistance prediction, desktop and command-line interfaces, and the broadest lineage coverage (L1–L10 and A1–A4), and still maintained sublineage resolution (Table).

Figure 3

Figure 3. Resistance-informed surveillance analysis of 7,148 CRyPTIC isolates in developing Pathotypr, an alignment-free tool for assemblies and reads in genomic surveillance. A) Performance of WHO confidence group 1–2 resistance markers compared…

Pathotypr’s Split-FASTQ complements lineage assignment by reporting World Health Organization–cataloged resistance-associated variants alongside the lineage call in assemblies with different coordinates, as illustrated across the 498-genome validation phylogeny (Appendix Figure 1). For 7,148 CRyPTIC isolates, we generated categorical resistance profiles from World Health Organization confidence groups 1 and 2 and compared them with phenotypic drug-susceptibility results. Genotypic and phenotypic resistance profiles agreed for 86.4% of isolates, and severe disagreement, defined as a difference of >3 levels between the genotypic and phenotypic profiles, occurred in 1.4%. Multidrug-resistant or pre–-extensively drug-resistant status was assigned with 93.5% sensitivity and 92.3% positive predictive value. Agreement was lowest for bedaquiline, clofazimine, delamanid, and linezolid, for which most phenotypically resistant isolates carried variants of uncertain significance or no cataloged variant (Figure 3, panel A).

We next used the 7,148-isolate CRyPTIC dataset to examine how Pathotypr–Split-FASTQ outputs could contribute to public health surveillance. Lineage assignments formed monophyletic groups in the annotated phylogeny (Appendix Figure 3), and ancestral-state reconstruction inferred 135 probable transitions from non-Europe ancestors into Germany, Italy, or Ukraine, comprising 590 descendant isolates (Figure 3, panels B and C). Among these 590 isolates, 195 (33.1%) had multidrug-resistant or pre–extensively drug-resistant genotypes, compared with 5.5% of the remaining Europe isolates in the dataset. Because CRyPTIC is enriched for drug-resistant isolates, is not population-based, and, after restriction to isolates with a documented country of origin and collection date, includes only Germany, Italy, and Ukraine as recipient countries in Europe, those proportions should not be interpreted as national estimates of importation or transmission. Rather, the analysis illustrates how standardized lineage and resistance outputs can be combined with other tools to identify cross-border clusters that merit epidemiologic follow-up (12).

The central scientific advance of this study is a phylogeny-derived lineage framework that represents the recognized diversity of the MTBC, with Pathotypr as the software system we developed to make that framework usable in genomic surveillance. This system’s 2,145 markers provide a traceable and extensible basis for assigning common and uncommon, human-adapted and animal-adapted groups, and the independent k-mer model preserves classification when direct marker detection is incomplete. By analyzing assemblies and reads without a conventional mapping workflow, reporting multiple supported lineages in mixed samples, and selecting a closely related reference genome, Pathotypr reduces unresolved or inappropriate assignments and places resistance results within a consistent evolutionary context.

Broad lineage coverage, rapid processing, standardized resistance reporting, and lineage-aware reference selection make Pathotypr a practical route from sequence data to comparable surveillance results. Our platform’s low memory requirements, desktop graphical interface, and command-line implementation make the same workflow accessible to laboratories with limited bioinformatics infrastructure while supporting automated analysis of large collections. Those features can improve reporting across laboratories, recognition of uncommon introductions, outbreak review, and prioritization of resistant cross-border clusters.