Introduction

Tuberculosis (TB) is among the ten top causes of death worldwide. In 2018 more than ten million people fell ill and 1.5 million died from TB (https://www.who.int/tb/global-report-2019). During the last two decades, significant efforts have been made to understand strain-level genetic diversity in the TB bacillus Mycobacterium tuberculosis (Mtb) and its geographic distribution. A robust classification of Mtb strains into evolutionarily meaningful sub-lineages is important both for taxonomic purposes and because sub-lineages can differ in virulence or antibiotic resistance1. Mtb was first classified into three principal genetic groups in 1997, based on two neutral single nucleotide substitutions (SNSs) in the antibiotic resistance genes katG (codon 463) and gyrA (codon 95)2. Since then several studies have attempted higher resolution classification, using large genomic deletions3, spoligotyping4 and SNSs5,6,7,8,9,10,11. There are currently nine recognized Mtb lineages (L1–9), with two lineages (L2, L4) that have been well represented in taxonomic and phylogeographic evaluations, while two others (L1 and L3) were the subject of more recent works6,7,8,9,12,13,14,15,16,17. Within the major lineages, 53 sub-lineages have been described and are definable with a SNS “barcode”6: 7 for L1, 6 for L2, 4 for L3, and 36 for L4. These 53 sub-lineages best characterize diversity within L2 and L4. The population structure of L1 and L3 is less understood as these lineages are most prevalent in countries where pathogen sequencing had been less widely applied. Recently, studies fueled by increasing sequencing capacity in high-burden TB settings have begun to evaluate the evolutionary history of L1 and L3 including the role of migration and dispersal in driving their prevalence in different parts of the world15,16,17,18.

Host-pathogen co-evolution and more recent host-related selective pressures have been postulated to drive genetic diversity in Mtb19,20. Of high relevance to public health, this genetic diversity is thought to underlie the observed differences in Mtb sub-lineage transmissibility21,22 or host specificity8. For example, of the Mtb sub-lineages prevalent in Ho Chi Minh City, Vietnam, one modern L2 sub-lineage (2.2.1) was found to be more transmissible relative to other L2 sub-lineages, L4 or L121. Another study reported reduced transmissibility for L3 compared to the other Mtb major lineages in Montreal, Canada23. Finally, the phylogeography of L5 and L6 (also known as M. africanum), L7, L8 and even some L4 sub-lineages (e.g., 4.6.1/Uganda, 4.6.2/Cameroon, 4.1.3/Ghana, and 4.5) has demonstrated that these groups are more geographically restricted than L2 and other L4 sub-lineages. These geographically restricted groups are associated with lower variation in Mtb T-cell epitopes and support the notion that Mtb sub-lineages have a spectrum of pathogenic strategies and maybe niche specialists infecting preferentially humans of a specific population or ancestry8. However, more evidence is needed to support this hypothesis given the relatively few genetic differences that define Mtb sub-lineages or human ancestry.

Here, we sought to expand our understanding of the Mtb population structure by leveraging 9584 genomes sampled from 49 countries, including 738 L1 and 1104 L3 isolates. We identify and validate 22 novel sub-lineages and 8 additional internal groups (i.e., genetically divergent groups found in sub-lineages that cannot be further partitioned in a hierarchical fashion according to our criteria), including 7 in L1 and 4 in L3, and expand the SNS typing barcode to 95 sites. We find that L3 and L2 have similar phylogenetic structure, yet L2 along with L4 manifest a phylogenetic signal of increased transmissibility compared with L1 or L3. These findings, along with our observations of novel geographically restricted sub-lineages, expand the evidence supporting the Mtb human co-evolution hypothesis.

Results

Detailed population structure of L1–4 and a hierarchical sub-lineage naming system

We assembled a high-quality data set of whole genomes, antibiotic resistance phenotypes, and geographic sites of isolation for 9584 clinical Mtb samples (“Methods” section and Supplementary Data 1). Of the total, 4939 (52%) were pan-susceptible, i.e., susceptible to at least isoniazid and rifampicin (and all other antibiotics when additional phenotypic data were available), and 4645 (48%) were resistant to one or more antibiotics (Supplementary Fig. 1a). Using the 62 SNS lineage barcode6, 738 isolates were classified as L1 (8%), 2193 as L2 (22%), 1104 as L3 (12%) and 5549 as L4 (58%, Supplementary Fig. 1b). Among the 4939 pan-susceptible isolates, we identified high-quality genome-wide SNSs (83,735 for L1, 56,736 for L2, 76,817 for L3, and 185,622 for L4) that we used in building maximum-likelihood phylogenies for each major lineage (L1–4, “Methods” section). We computed an index of genetic divergence (FST) between groups defined by each bifurcation in each phylogeny. Sub-lineages were defined as monophyletic groups that had high FST (>0.33) and were also clearly separated from other groups in principal component analysis (PCA, see “Methods” section). We also defined internal groups to sub-lineages (see “Methods” section): an internal group is a monophyletic group genetically divergent (by FST and PCA) from its neighboring groups, but has one or more ancestral branches that show a low degree of divergence or low support (bootstrap values). Internal groups do not represent true sub-lineages in a hierarchical fashion, but defining them allows us to further characterize the Mtb population structure. We provide code to automate all the steps described above. Our approach is scalable and can be used on other organisms (see “Methods” section).

To better classify Mtb isolates in the context of the global Mtb population structure, we developed a hierarchical sub-lineage naming scheme (Supplementary Data 2) where each subdivision in the classification corresponds to a split in the phylogenetic tree of each major Mtb lineage. Starting with the global Mtb lineage numbers (e.g., L1), we recursively introduced a subdivision (e.g., from 1.2 to 1.2.1 and 1.2.2) at each bifurcation of the phylogenetic tree whenever both subclades sufficiently diverged. Formally, we defined these splits using bootstrap criteria, and independent validations by FST and PCA (see “Methods” section). Internal groups were denoted with the letter “i” (e.g., 4.1.i1). This proposed system overcomes two major shortcomings of the existing schemas: same-level sub-lineages are never overlapping (unlike the system of Stucki et al.8 sub-lineage 4.10 includes sub-lineages 4.7–4.9), and the names reflect both phylogenetic relationships and genetic similarity (unlike semantic naming such as the “Asia ancestral” lineage in the system of Shitikov et al.7). Further, this naming system can be standardized to automate the process of lineage definition. These advantages come at the price of long sub-lineage names in the case of complex phylogenies (e.g., for L4, sub-lineage 4.10 gets the lineage designation 4.2.1.1.1.1.1.1). For compatibility with naming conventions already in use and to keep names as short as possible, we designed a second, shorthand, naming system that expands the Coll et al. lineage schema by adding new subdivisions and differentiating between sub-lineages and internal groups. For instance, sub-lineage 4.3.1 is designated as 4.3.i1, informing the user that this is an internal group of sub-lineage 4.3. To simplify the use of the hierarchical naming schema and the updated shorthand schema, we provide a table that compares them side by side along with naming systems currently in use (Supplementary Data 2).

Using the sub-lineage definition rules and the sub-lineage naming scheme described above, we characterized six previously undescribed sub-lineages of L1 (Fig. 1 and Supplementary Fig. 2); five of which expand the current description of 1.2. We also detected an internal group of 91 isolates (1.1.3.i1) characterized by a long defining branch in the phylogeny (corresponding to 82 SNSs), a high FST (0.48), and geographically restricted to Malawi (85/91, 93% isolates, Fig. 1 and Supplementary Fig. 3). We estimated the date of the emergence of the MRCA of such a group (see “Methods” section) and we found it to be between 1497 and 1754. We found four previously undescribed sub-lineages of L3 (Fig. 2 and Supplementary Fig. 4), revising L3 into four main groups, whereas previously only two partitions of one sub-lineage were characterized (3.1). We found that the latter two partitions are in fact internal groups of the largest sub-lineage (3.1.1) in our revised classification.

Fig. 1: Phylogenetic tree reconstruction of lineage 1 (binary tree).
figure 1

Gray circles define splits where the FST (fixation index) calculated using the descendants of the two children nodes is greater than 0.33. The sub-lineages are defined by colored areas (blue: sub-lineages already described in the literature; green: sub-lineages described here; purple: internal sub-lineages). Source data are provided as a Source Data file.

Full size image
Fig. 2: Phylogenetic tree reconstruction of lineage 3 (binary tree).
figure 2

Gray circles define splits where the FST (fixation index) calculated using the descendants of the two children nodes is greater than 0.33. The sub-lineages are defined by colored areas (green: sub-lineages described here; purple: internal sub-lineages). Source data are provided as a Source Data file.

Full size image

L2 is divided into two groups: proto-Beijing and Beijing with the latter in turn partitioned into two groups: ancient- and modern-Beijing7. Each one of these groups is characterized by further subdivisions (three for the ancient-Beijing group and seven for the modern-Beijing group; see Supplementary Fig. 4). We found a new sub-lineage (2.2.1.2, Fig. 3, and Supplementary Fig. 5) within the previously characterized ancient-Beijing group. However, genetic diversity within the modern-Beijing group (2.2.1.1.1) was lower than in the other L2 sub-lineages and the tree topology and FST calculations did not support further hierarchical subdivisions. Although we did find three internal groups of modern-Beijing: two undescribed and one that corresponds to the Central Asia group7. For L4, our results support a complex population structure with 21 sub-lineages and 15 internal groups. In particular, we found 11 previously undescribed sub-lineages and 5 internal groups that expand our understanding of previously characterized sub-lineages (e.g., 4.2.2; 4.2 in the Coll et al. classification) or that were not characterized since these isolates were simply classified as L4 (e.g., 4.11, Fig. 4, and Supplementary Fig. 6) using the other barcodes.

Fig. 3: Phylogenetic tree reconstruction of lineage 2 (binary tree).
figure 3

Gray circles define splits where the FST (fixation index) calculated using the descendants of the two children nodes is greater than 0.33. The sub-lineages are defined by colored areas (blue: sub-lineages already described in the literature; green: sub-lineages described here; purple: internal sub-lineages). Source data are provided as a Source Data file.

Full size image
Fig. 4: Phylogenetic tree reconstruction of lineage 4 (binary tree).
figure 4

Gray circles define splits where the FST (fixation index) calculated using the descendants of the two children nodes is greater than 0.33. The sub-lineages are defined by colored areas (blue: sub-lineages already described in the literature; green: sub-lineages described here; purple: internal sub-lineages). Source data are provided as a Source Data file.

Full size image

A new barcode to define L1–4 Mtb sub-lineages and a software package to type Mtb strains from WGS data

We defined a SNS barcode for distinguishing the obtained sub-lineages (Supplementary Data 3). We characterized new synonymous SNSs found in 100% of isolates from a given sub-lineage, but not in other isolates from the same major lineage, compiling 95 SNSs into an expanded barcode (Supplementary Data 3). We validated the barcode by using it to call sub-lineages in the hold-out set of 4645 resistant isolates and comparing the resulting sub-lineage designations with maximum-likelihood phylogenies inferred from the full SNS data (Supplementary Figs. 7–10). A sub-lineage was validated if it was found in the hold-out data and formed a monophyletic group in the phylogeny. Considering the “recent” sub-lineages, i.e., the most detailed level of classification in our system, we were able to validate eight out of nine L1 sub-lineages including five out of six of the new sub-lineages described here, with the exception of 1.1.1.2. We validated all four new L3 sub-lineages, all five L2 sub-lineages including the one previously undescribed, and 16 of the 21 L4 sub-lineages including two described here. The sub-lineages we could not confirm were not represented by any isolate in the validation phylogenies. We did not observe any paraphyletic sub-lineages in the revised classification system.

We developed fast-lineage-caller, a software tool that classifies Mtb genomes using the SNS barcode proposed above. For a given genome, it returns the corresponding sub-lineage as output using our hierarchical naming system in addition to four other existing numerical/semantic naming systems, when applicable (see “Methods” section). The tool also informs the user on how many SNSs support a given lineage call and allows for filtering of low-quality variants. The tool is generalizable and can manage additional barcodes defined by the user to type the core genome of potentially any bacterial species.

Geographic distribution of the Mtb sub-lineages

Next, we examined whether certain sub-lineages were geographically restricted, which would support the Mtb-human co-evolution hypothesis, or whether they constituted prevalent circulating sub-lineages in several different countries (i.e., geographically unrestricted)8. We used our SNS barcode to determine the sub-lineages of 17,432 isolates (see “Methods” section) sampled from 74 countries (Supplementary Fig. 11 and Supplementary Data 4, 5). We computed the Simpson diversity index (Sdi) as a measure of geographic diversity that controls for variable sub-lineage frequency (see “Methods” section) for each well-represented sub-lineage or internal group (n > 20). We hypothesized that geographically unrestricted lineages would have a higher Sdi. We found Sdi to correlate highly (⍴ = 0.68; p-value = 5.7 × 10−7) with the number of continents from which a given sub-lineage was isolated (Supplementary Fig. 12). The Sdi ranged between a minimum of 0.05 and a maximum of 0.72, with a median value of 0.46 (Fig. 5). The known geographically restricted sub-lineages8 had an Sdi between 0.28 and 0.5 (Fig. 5 and Supplementary Table 1), while the known geographically unrestricted sub-lineages8,9 had an Sdi between 0.55 and 0.61 (Fig. 5 and Supplementary Table 2). We found 11 sub-lineages/internal groups with Sdi <0.28 (Supplementary Table 3), and 11 sub-lineages/internal groups with Sdi >0.61 (Supplementary Table 4), i.e., more extreme than previously reported geographically restricted or unrestricted sub-lineages, respectively.

Fig. 5: Histogram of the Simpson diversity index calculated for sub-lineages of lineages 1–4.
figure 5

A data set of 17,432 isolates from 74 countries was used to perform this analysis. Yellow triangles designate the Simpson diversity index values of sub-lineages designated as geographically restricted by Stucki et al. Light gray circles designate the Simpson diversity index values of sub-lineages designated as geographically unrestricted by Stucki et al. Source data are provided as a Source Data file.

Full size image

While the currently known geographically restricted sub-lineages are all in L4, we found evidence of geographic restriction for two sub-lineages/internal groups of L1. The first, the L1 internal group 1.1.3.i1, showed a very low Sdi (0.06) and was only found at high frequency among the circulating L1 isolates in Malawi (Fig. 6). This finding is also in agreement with the L1 phylogeny (Fig. 1) that shows a relatively long (82 SNS) branch defining this group. The second geographically restricted L1 sub-lineage is 1.1.1.1 (Sdi = 0.12) that was only found at high frequency among circulating L1 isolates in South-East Asia (Vietnam and Thailand, Fig. 7). To exclude the possibility that these two groups appeared geographically restricted as a result of oversampling transmission outbreaks, we calculated the distribution of the pairwise SNS distance for each of these two sub-lineages. We measured a median SNS distance of 204 and 401, respectively, refuting this kind of sampling error for these groups (typical pairwise SNS distance in outbreaks <15–20 SNS24) (Supplementary Fig. 13).

Fig. 6: Geographic distribution of internal sub-lineage 1.1.3.i1.
figure 6

Colors represent the percentage of 1.1.3.i1 strains isolated in a given country with respect to all lineage 1 strains isolated in such country. Source data are provided as a Source Data file.

Full size image
Fig. 7: Geographic distribution of internal sub-lineage 1.1.1.1.
figure 7

Colors represent the percentage of 1.1.1.1 strains isolated in a given country with respect to all lineage 1 strains isolated in such country. Source data are provided as a Source Data file.

Full size image

The geographically unrestricted sub-lineages spanned several recognized generalist sub-lineages or internal groups (e.g., 4.1.i1.1.1.1, 4.3, 4.10, and 2.2.1.1.1, which correspond to 4.1.2/Haarlem, 4.3/LAM, 4.10/PGG3, modern-Beijing in the Stucki et al.8 or the Shitikov et al.7 classification). In addition, we found a candidate geographically unrestricted sub-lineage of L1 (1.1.2). L1.1.2 spanned 253 isolates from 7 countries and 4 continents and its Sdi was 0.61 (Fig. 8). Overall we found a low, but significant correlation between Sdi and the number of isolates sampled from each sub-lineage (⍴ = 0.34; p-value = 0.03, Supplementary Fig. 14) indicating that the most prevalent sub-lineages were more likely to be geographically unrestricted.

Fig. 8: Geographic distribution of internal sub-lineage 1.1.2.
figure 8

Colors represent the percentage of 1.1.2 strains isolated in a given country with respect to all lineage 1 strains isolated in such country. Source data are provided as a Source Data file.

Full size image

To validate the geographic distribution of sub-lineages, we curated a data set of 3791 Mtb isolates sampled to reliably represent the entire population of TB patients in five countries (Azerbaijan, Bangladesh, Pakistan, South Africa, and Ukraine, Supplementary Data 6, Zignol et al. data set). In this data set, we were able to identify 9 of the 30 new groups/sub-lineages described above. L1 and 1.1.2 isolates were found in three out of five countries (Azerbaijan, Pakistan, and Ukraine) spanning two continents2 (Supplementary Figs. 15 and 16) supporting the hypothesis that L1.1.2 is a geographically unrestricted L1 sub-lineage. The candidate geographically restricted group 1.1.3.i1 from Malawi, was only additionally found in South Africa (Supplementary Fig. 16). L2 isolates were found in all five countries with 98% belonging to the modern-Beijing group (Supplementary Figs. 15 and 17). The modern-Beijing internal group 2.2.1.1.1.i3, corresponding to the Central Asia group7, was the most prevalent L2 group in Ukraine and Azerbaijan and the second most prevalent group in Pakistan. This confirms 2.2.1.1.1.i3’s observed high Sdi (0.66) in the convenience training data and is in line with modern-Beijing’s transmissibility (next paragraph and ref. 21). L3 isolates were found in four (Azerbaijan, Bangladesh, Pakistan, South Africa) of the five countries (Supplementary Fig. 15). Sub-lineage 3.1.1 was the most prevalent sub-lineage (Supplementary Fig. 18), but the two new L3 sub-lineages we describe (3.2.1 and 3.1.2) were also observed at low frequency (2.4-4.8%) in Bangladesh and Pakistan. L4 isolates and most commonly 4.1, 4.3, and 4.10 (that correspond to 4.1, 4.3/LAM, and 4.10/PGG in the Stucki et al.8 classification; Supplementary Fig. 19) were found in all five countries in line with their Sdi >0.67 and results on L4 transmissibility below.

Differences in transmissibility between the Mtb global lineages

The observation that some lineages/sub-lineages are more geographically widespread than others raises the question of whether this results from differences in marginal transmissibility across human populations. On a topological level, we observed L2 and L3 phylogenies to be qualitatively different from those of L1 and L4 (Figs. 1–4): displaying a star-like pattern with shorter internal branches and longer branches near the termini. We confirmed this quantitatively by generating a single phylogenetic tree for all 9584 L1–4 isolates and plotting cumulative branch lengths from root to tip for each main lineage (Supplementary Fig. 20). Star-like topologies have been postulated to associate with rapid or effective viral or bacterial transmission e.g., a “super-spreading” event in outbreak contexts25. To compare transmissibility between the four lineages, we compared the distributions of terminal branch lengths expecting a skew toward shorter terminal branch lengths supporting the idea of higher transmissibility. We found L4 to have the shortest median terminal branch length, followed in order by L2, L3, and L1 (medians: 6.2 × 10−5, 8.2 × 10−5, 10.2 × 10−5, 17.5 × 10−5, respectively; all pairwise two-sided Wilcoxon rank-sum tests significant p-value < 0.001; Fig. 9). Shorter internal node-to-tip distance is a second phylogenetic correlate of transmissibility; the distribution of this measure across the four lineages revealed a similar hierarchy to the terminal branch length distribution (Supplementary Fig. 21). We also computed the cumulative distribution of isolates separated by increasing total pairwise SNS distance (Supplementary Fig. 22). The proportion of L4 isolates separated by <5 SNS was highest and followed by L2, L3, and L1, respectively. Thus, despite the topological similarities between the L3 and L2 trees, L3 was measured to be less transmissible by interrogating the topology of the trees.

Fig. 9: Distributions of terminal branch lengths for the four global Mtb lineages (L1–L4).
figure 9

Two-sided Wilcoxon rank sum tests were performed to test that two distributions were significantly different. Medians: 6.2 × 10−5 (L4), 8.2 × 10−5 (L2), 10.2 × 10−5 (L3), 17.5 × 10−5 (L1). Comparisons: L1 vs L2, L3 or L4 (p-value < 2.2 × 10−16); L2 vs L3 (p-value = 3.6 × 10−6), L2 vs L4 (p-value < 2.2 × 10−16); L3 vs L4 (p-value < 2.2 × 10−16). Description of the distributions (L1: n = 739, Min: 0.5 × 10−5, 1st Quartile: 6.7 × 10−5, Median: 17.5 × 10−5, 3rd Quartile: 28 × 10−5, Max: 120 × 10−5; L2: n = 2193, Min: 0.7 × 10−5, 1st Quartile: 5.3 × 10−5, Median: 8.2 × 10−5, 3rd Quartile: 12 × 10−5, Max: 110 × 10−5; L3: n = 1103, Min: 0.5 × 10−5, 1st Quartile: 4.5 × 10−5, Median: 10.2 × 10−5, 3rd Quartile: 20 × 10−5, Max: 80 × 10−5; L4: n = 5514, Min: 0.2 × 10−5, 1st Quartile: 2.6 × 10−5, Median: 6.2 × 10−5, 3rd Quartile: 13 × 10−5, Max: 70 × 10−5). Source data are provided as a Source Data file.

Full size image

To confirm that the measured transmissibility differences are not due to sampling bias in the source data, we compared the distributions of terminal branch lengths for the four major lineages using the susceptible isolates only (n = 4939/9584; data set with curated phenotypes) and using the Zignol et al. data set, where the isolates have been randomly sampled in five countries. In both cases, we found that L4 and L2 have the shortest median terminal branch lengths and L1 the longest. In the Zignol et al. data set, we also found that L4 and L2 terminal branch lengths were shorter than L3’s (Supplementary Information, Supplementary Figs. 23 and 24).

Discussion

Our results provide the most detailed picture of Mtb’s population structure to date and extend previous schemas6,8, particularly for L1 and L3. We describe 7 and 4 new sub-lineages/internal groups, respectively, for these two lineages. The largest improvement was in increasing the resolution of sub-lineage 1.2 (5 subdivisions instead of 2), and in partitioning L3 into 4 groups since previously only a single group with one sub-lineage was characterized (3.1) for L3. We found an internal sub-lineage of 1.1 (1.1.3.i1) that was found almost exclusively in Malawi (85/91 isolates) with nearest neighbors isolated from India. The tMRCA of this group dates back to a point in time between c1497 and c1754. Recent work examining the evolutionary history of L1 concluded that its origin was most likely in South Asia with a tMRCA estimated in the twelfth century AD16. Dissemination of L1 out of South Asia may have been related to increase in maritime trade between the continents in this era including seasonal trade following the monsoon season between South Asia and East/Southern Africa. In this study, a group of isolates belonging to sub-lineage 1.1.3 was defined, the vast majority of which are from Malawi, consistent with our results. European contact with the autochthonous populations in South Eastern Africa is estimated to have taken place in the late 15th century, around the time of origin of sub-lineage 1.1.3.i1. Despite the opportunity for dissemination mediated by trade and colonization into Europe and other continents, we observe an unusual pattern of geographic restriction of this group of isolates, consistent with a specialist phenotype. This supports the idea that this is a candidate lineage with adaptation to a specific human genetic background in this region of Africa through co-evolution. This observation can be confirmed as more extensive and systematic pathogen whole genome sequencing becomes available from Sub-saharan Africa.

For L2, we find limited internal diversity with the exception of major splits between the proto-, ancient and modern-Beijing groups. Our schema supports 3 out of 4 groups previously described as either proto-Beijing or ancient-Beijing; the fourth unsupported group in our schema is Asia ancestral 3 previously proposed as a subdivision of ancient Beijing7. We also identify a separate new sub-lineage of the ancient-Beijing group (2.2.1.2). Our results show no evidence for sub-lineages inside the modern-Beijing group, but support three internal groups (2.2.1.1.1.i1, 2.2.1.1.1.i2, 2.2.1.1.1.i3) one of which (Central Asia) was previously described by Shitikov et al. Our analysis supports the existing schemas for L46,8 and extends them by defining 11 new sub-lineages (and 5 internal groups). In particular, we improve the resolution of known sub-lineages, for instance, 4.2, 4.6.1, and 4.6.2, and define three new sub-lineages (4.11–4.13) which increase the total number of L4 major sub-lineages from 7 (4.1–4.6 and 4.10 in the Stucki et al. classification) to 10.

We define a SNS barcode (95 SNS) that allows the rapid assignment of Mtb sub-lineage designations with detailed semantic and hierarchical sub-lineage naming schema from genomic data (.vcf files)6,7,8. Along with this, we provide a software package and comparative tables/figures to facilitate the interchange between five SNS schemas and three Mtb sub-lineage naming systems6,7,8,26. We expect these tools to facilitate the use of the expanded Mtb classification by scholars of Mtb evolution and public health practitioners alike for applications such as the rapid assessment of potential outbreaks and isolate triage for detailed phylogenetic analyses. Further, low-level susceptibility to specific Mtb drugs has been reported for some lineages27,28. Improving the resolution of classification at the sub-lineage level can allow for a better understanding of these findings and potentially a more refined prediction of drug resistance profiles.

We find evidence supporting the Mtb-human co-evolution hypothesis and its corollary of lineage differential adaptation8, including a spectrum of transmissibility across the four major Mtb lineages. We characterize L4 and L2 as the most transmissible, L1 as the least transmissible one, and L3 showing an intermediate level of transmissibility using different phylogenetic metrics. This is consistent with previous studies that have identified L2 sub-lineages as more transmissible than L1 in Vietnam21 and Malawi22. This result also supports several reports which have failed to identify differences in transmissibility between L2 and L429,30,31. In order to test the robustness of our findings and minimize the potential sampling bias, we also compared the transmissibility of the major Mtb lineages using a the pan-susceptible isolates form our data set with curated phenotypes (n = 4939/9584) and using a data set where isolates were randomly sampled in five countries and we got similar results. Finally, our results are also in line with a larger number of geographically unrestricted “generalist” L4 and L2 sub-lineages observed (compared with L3 or L1).

We find two new candidates with geographically restricted sub-lineages/internal groups, one of them from Malawi and one from South-East Asia (Vietnam and Thailand). We also report evidence of a geographically unrestricted “generalist” sub-lineage in L1 (1.1.2). L1 is the most ancestral of the four main Mtb lineages (L1–4), and given the evidence of low transmissibility on average across countries and sub-lineages, the finding of a potential generalist L1 sub-lineage requires further study and validation with additional data.

Our results raise the question of whether the geographically restricted sub-lineages represent cases of specialization or adaptation to specific human populations and whether the geographically unrestricted sub-lineages might have genetic features that allow them to spread efficiently in many different human populations. Although the phylogenies and validated geographic distribution are suggestive of differential adaptation, confirmation of this observation requires control for TB exposure and differences in contact networks using epidemiological data. A limitation of our analysis is the lack of this data and other patient and country-level data, including human ancestry, recent migration, and HIV prevalence. Nevertheless, we find large differences in the geographic distribution of more than 20 Mtb sub-lineages across 51 countries and validated using systematically sampled data from these respective geographies. In the future, as additional patient metadata and Mtb sequence data become available, follow-up analyses can confirm the geographic distribution of these sub-lineages. In summary, this work provides new insights into the population structure, biogeography, and transmissibility of Mtb and demonstrates the use of this information to classify Mtb strains and investigate the relationships between population structure, pathogenicity, and transmission.

Methods

Phenotypic and geographic location data

We compiled a first data set of antibiotic susceptibility and resistance data for 11,349 Mtb isolates using public databases (Patric)32 and literature curation9,33,34,35,36,37,38,39,40,41,42,43,44. A summary table of the data and the scripts used to generate it are available at https://github.com/farhat-lab/resdata (v1.0). The isolate metadata including the country of isolation were downloaded using SRAtools v2.9.1 (https://github.com/ncbi/sra-tools). Isolates were filtered according to the procedures described below, resulting in a data set of 9584 isolates. The full list of isolates and countries of isolation are available on Supplementary Data 1. We used this data set to determine the sub-lineages definitions and study the differences in transmissibility between the major Mtb lineages. In order to determine the geographic distribution of the Mtb sub-lineages we used a second data set of 17,431 Mtb isolates. The full list of isolates and countries of isolation are available on Supplementary Data 4. Finally, in order to validate the geographic distribution of sub-lineages we used a third data set of 3791 Mtb isolates from five countries and three continents (Azerbaijan, Bangladesh, Pakistan, South Africa, and Ukraine, Supplementary Data 6). These isolates come from epidemiological surveys designed to be representative of the entire population of TB patients of each one of these countries18. A summary table containing the breakdown by country and lineage of all these three datasets is available on Supplementary Data 7.

Sequencing data analysis and lineage calling

Sequence read data for the isolates were downloaded from NCBI (the list of BioSamples is available on Supplementary Data 1). We used an implementation of the pipeline proposed by Ezewudo et al.45 to get the genetic variants that characterize any strain with respect to the Mtb reference strain H37Rv. For each isolate, identified by a BioSample, we downloaded all the associated Illumina sequencing runs and we ran PRINSEQ v0.20.446 to trim and filter the reads (using an average phred score threshold of 20). Then we ran Kraken v0.10.647 and discarded the isolates where <90% of the reads were assigned to Mtb. For this purpose, we set up a custom Kraken database, to reduce the memory requirements of the default database (Reference sequences: NC_009565.1 / L4, NC_000962.3 / L4, NC_017524.1 / L4, NC_002755.2 / L4, NC_021054.1 / L2).

Reads were then aligned to the H37Rv (NC_000962.3) reference genome using “BWA mem” v0.7.1748. Duplicate reads were removed with Picard v2.9.2 (http://broadinstitute.github.io/picard/). As an additional quality check the coverage of the Mtb genome was evaluated using samtools v1.949 and all isolates having a coverage of less than 95% of the Mtb genome with a depth of at least 10X were dropped. Genetic variants were called using Pilon v1.2250. Lineage calls were made using the fast-lineage-caller v1.0 (https://github.com/farhat-lab/fast-lineage-caller), using the Coll et al.6 and Shitikov et al.7 SNS schemes.

Phylogenetic trees

In order to reduce the probability of having mixed isolates (e.g., isolates from different lineages/sub-lineages that infected the same patient), that could affect the determination of the tree of a given Mtb lineage, we computed the F2 lineage-mixture metric51 and excluded the isolates that had scores greater than 0.5. We separately determined the phylogenetic trees for the pan-susceptible isolates (isolates that were susceptible to both isoniazid and rifampicin as well as all other antibiotics they were tested) and the resistant ones (isolates that were resistant to one or more antibiotics). To generate the trees we merged the .vcf files of the isolates of a given lineage with bcftools v1.952. Then we removed the repetitive, antibiotic resistance and low coverage regions (Supplementary Data 8). We generated a multi-sequence.fasta alignment from the merged.vcf with vcf2phylip v1.5 (https://doi.org/10.5281/zenodo.1257057) and we determined the phylogenetic tree with iqtree v1.6.10 (automatic model selection, 1000 bootstraps).

F ST and PCA

The FST was calculated using the PopGenome R package v2.6.153. PCA analysis was performed using the ade4 R package54 v1.7-8 (R v3.5.1).

Sub-lineage definition

We defined the sub-lineages by evaluating: (1) the information of the tree topology, (2) the support (bootstrap), (3) the FST, and (4) the results of a PCA analysis performed at each node. We considered a node as an informative one if: the support was >95 (iqtree ultrafast bootstrap), the FST was >0.33 and in the PCA analysis, we were able to identify two groups (since our trees were binary trees). The threshold of the FST (0.33), we choose was already used to characterize sub-lineages in the previous works8 and the distributions of the FST calculated at each internal node of all lineages suggest it is a reasonable value to discriminate between Mtb sub-lineages (Supplementary Fig. 25). In a few situations (n = 3) we relaxed one of these constraints either because the node was the first split within a major lineage (split between 1.1 and 1.2, split between 4.1 and 4.[2–6,10–13]) or the group was already identified by previous studies and the results taken together were indicating that the group is valid (4.3.i1). The scripts used to define the sub-lineages/internal groups are available at https://github.com/farhat-lab/mtb-popstruct-2020.

Hierarchical sub-lineage naming system

We assigned to all the sub-lineages their major lineage number (1, 2, 3, or 4). For each of the informative nodes in the tree (it is important to notice that we generated binary trees), we assigned a “.1” for the subtree having most of the descendants, while we assigned a “.2” for the subtree having the least descendants (exceptions to this rule were made for 4.1 and 2.1 to retain compatibility with the Coll et al.6 SNS scheme). We stopped defining groups when a split was not sufficiently supported (ultrafast bootstrap < 95), the FST was <0.33, or the results of a PCA analysis were not conclusive (see also the Sub-lineage definition section). We also allowed for internal groups, i.e., if one or more ancestors of a given informative node were not informative we considered all the descendants of the informative node as members of an internal group of a given sub-lineage. We, therefore, added a “.i <n> ” suffix to design such groups of strains as internal groups (e.g., 4.1.i1).

Comparison of tree topologies

In order to compare the topologies of L1–4 trees, we calculated the proportion of tree length as a function of the node level. We used a phylogenetic tree containing all L1–4 isolates. Starting from the MRCA node of each one of the L1–4 sets of isolates we iteratively looked for the children nodes (target nodes) and determined the average patristic distances between the target nodes and the MRCA node. Each one of the iterations represented a node level. We then normalized the resulting averaged distances by the maximum value of the average patristic distance obtained. This algorithm was implemented in R using the ape v5.355 and phangorn v2.5.356 packages.

A python lineage caller

We developed a python module (https://github.com/farhat-lab/fast-lineage-caller) that takes as input a.vcf file and returns the lineage / sub-lineage calls. We provide five SNS schemes for Mtb: Coll et al.6 (all lineages), the Shitikov et al.7 (L2), Stucki et al.8 (L4, geographically bounded and unbounded sub-lineages), Lipworth et al.26 (Mtb complex species, Mtb main lineages) and the one proposed here (L1–4). Version 1.0 was used for the analyses described in this study.

Dating the MRCA of internal group 1.1.3.i1

We calculated the median branch length from the MRCA node of all 1.1.3.i1 isolates and converted it to the number of SNS using the alignment length. We then considered an evolutionary rate bracket of 0.3 to 0.6 substitutions per genome per year and multiplied it for the number of SNS to get a time range corresponding to such branch length.

Simpson diversity index

The Simpson diversity index was calculated using the vegan (version 2.5.6; https://cran.r-project.org/package=vegan) R package. The correlations between the Simpson diversity index and the number of continents where a given sub-lineage has been found or the number of isolates of each sub-lineage (Supplementary Figs. 12 and 14) were determined using the cor function of the stats57 (version 3.5.1) R package.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.