Genomic diversity in autopsy samples reveals within-host dissemination of HIV-associated M. tuberculosis

Abstract

Mycobacterium tuberculosis remains a leading cause of death worldwide, especially among individuals infected with HIV¹. While phylogenetic analysis has revealed M. tuberculosis spread throughout history^2–5 and in local outbreaks^6–8, much less is understood about its dissemination within the body. Here, we report genomic analysis of 2693 samples collected postmortem from lung and extrapulmonary biopsies of 44 subjects in KwaZulu-Natal, South Africa who received minimal antitubercular treatment and most of whom were HIV seropositive. We found that purifying selection acted within individual patients, without the need for patient-to-patient transmission. Despite negative selection, mycobacteria diversified within individuals to form sub-lineages that co-existed for years. These sub-lineages, as well as distinct strains from mixed infections, were differentially distributed throughout the lung, suggesting temporary barriers to pathogen migration. As a consequence, samples taken from the upper airway often captured only a fraction of the population diversity, challenging current methods of outbreak tracing and resistance diagnostics. Phylogenetic analysis indicated that dissemination from the lungs to extrapulmonary sites was as frequent as between lung sites—supporting similar migration routes within and between organs, at least in subjects with HIV. Genomic diversity therefore provides a record of pathogen diversification and repeated dissemination across the body.

Mutations that pathogens accumulate over time can be used to reconstruct their transmission history. In theory, the same principles used to track spread across the globe and between individuals can inform our understanding of how M. tuberculosis spreads within the lungs and throughout the body^9,10. While M. tuberculosis primarily causes lung infections, it can disseminate to extrapulmonary organs, particularly in subjects co-infected with HIV. Understanding how M. tuberculosis diversifies and spreads within individuals is crucial to outbreak tracing¹¹, clinical diagnosis of resistance^12–14, and designing strategies to minimize de novo emergence of antibiotic resistance^15–17.

Here, we conducted a postmortem study in KwaZulu-Natal, South Africa to capture the spatial diversity of M. tuberculosis within each person (Fig. 1). Inclusion criteria included tuberculosis as cause of death, suspected disseminated infection, and fewer than 4 days of anti-tubercular treatment (Online Methods). Ninety-six of the 100 subjects enrolled were HIV positive, reflecting the high rate of HIV positivity among those dying from tuberculosis in the province. For each subject, a minimally invasive autopsy was performed, collecting specimens from each of six lung sites (3 regions within each lung, Supplementary Table 1), as well as from the liver and spleen. Each lung biopsy site was processed individually, while multiple biopsies from the other organs were pooled to form a single site per organ. Endotracheal aspirate, a proxy for what might have been detectable from a sputum specimen, was also collected. Specimens were cultured on solid media to select for M. tuberculosis. We focused on 44 subjects that had positive cultures for at least one lung site and at least one extrapulmonary site. Of these, all but two were HIV positive (Supplementary Table 2).

Figure 1. Postmortem analysis of 2693 M. tuberculosis samples from 329 sites across the body of 44 subjects.

Subjects who died in the emergency, admitting, or inpatient wards of Edendale Hospital in Kwazulu-Natal, South Africa were eligible to be enrolled in our study if they had been on anti-tubercular therapy for fewer than 4 days before death. For each subject, postmortem, biopsies were taken from each of six locations within the lung, each treated as a separate site and from multiple biopsies within the liver and spleen, pooled to form one site per organ. Aspirates of endotracheal secretions were taken, as were aspirates from clinically apparent ascitic fluid, pleural fluid, or pus collections when detected. Lung sites 1–3 were taken from the right lung and lung sites 4–6 were taken from the left lung (top to bottom, Supplementary Table 1). For some subjects, samples from ascitic fluid, pleural fluid, or pus collections were obtained. Each site was cultured separately for M. tuberculosis. DNA from one or more samples (up to 15 per site) was sequenced at the whole genome level.

A total of 2693 M. tuberculosis samples from 329 sites were analyzed for both fixed and polymorphic single-nucleotide polymorphisms (SNPs) at the whole-genome level (Supplementary Table 3). For the first 12 subjects, the diversity on solid media was scraped into a single tube, resulting in one DNA sample per site. For the remaining subjects, up to 15 individual colonies from each site were processed as separate samples (in many cases colonies were found to be formed by multiple cells). Genomic libraries were sequenced on the HiSeq platform to an average of 178x for scrape samples and 59x for colony samples. Reads were aligned to an approximate ancestral genome of the M. tuberculosis complex³, allowing inference of derived and ancestral alleles. Co-variation of mutation frequencies across samples from each subject distinguished pre-existing polymorphisms due to mixed infection from de novo mutations (Online Methods, Supplementary Fig. 1). We then used a constraint-based method to infer individual genotypes in non-clonal samples (Supplementary Fig. 2).

Four subjects had mixed infections, providing a prevalence estimate (9%) consistent with high endemic areas^15,18,19. These subjects were identified by the average mutational distance of their M. tuberculosis population to its most recent common ancestor (〈dMRCA〉, Online Methods, Fig. 2a). Given the M. tb molecular clock rate of about 0.3–0.5 SNPs/year^5,8,20, these subjects had values of 〈dMRCA〉 consistent with mixed infection (201–439 SNPs/cell), while all others had values of fewer than 5 SNPs/cell. In 2 of these 4 cases, the minority strain was detected at low frequency across samples (5% and 11%), highlighting the importance of deep sampling for detection of mixed infections.

Figure 2. Genomic sequencing reveals variation due to mixed infection and de novo mutation.

(a) A histogram of the average value of 〈dMRCA〉 across subjects, the average number of single nucleotide mutations per M. tuberculosis cell in that subject—relative to each subject’s population’s most recent common ancestor (MRCA). (b) A histogram of the number of de novo mutations within each subject’s population, after removal of polymorphisms arising due to mixed infection (Supplementary Fig. 1). (c) Two examples of within-subject phylogenies created from inferred genotypes. Genotypes were called if they were found at least 20% frequency within a sample (each sample can contain multiple genotypes); colored bars indicate site.

Mutations appearing de novo showed evidence of purifying selection. The number of de novo mutations detected varied considerably across subjects (median 8.5 SNPs, range 0–59; Fig. 2b) and was somewhat dependent on the number of specimens (Supplementary Fig. 3a), but was generally too low for selection to be analyzed on a subject-by-subject basis. Considering all de novo mutations together, the relative proportions of different nucleotide-to-nucleotide mutations closely matched that of mutations between subjects(Supplementary Fig. 4). De novo mutations had fewer amino-acid changing mutations than expected under a model of neutral evolution (dN/dS = 0.78, p<.01, Online Methods), similar to mutations observed between subjects (dN/dS = 0.80, p<.001). These data suggest that purifying selection known from between patient comparisons^21,22 does not require patient-to-patient transmission, but can act within each subject.

De novo mutations led to sub-lineages that co-existed for many years in some patients. Focusing on single-strain infections, 〈dMRCA〉 estimates varied considerably across subjects (range 0–5 SNPs/cell, 0–16 years; Fig 2a) and were not sensitive to the number of samples (Supplementary Fig. 3c–d). Some subjects had values of 〈dMRCA〉 consistent with many years of within-patient diversification, yet most subjects had much lower 〈dMRCA〉 estimates (median 0.24 SNPs/cell, < 1 year), consistent with the fact that immunocompromised subjects are likely to have fulminant courses of disease. Subjects who were HIV seronegative or diagnosed with HIV infection only postmortem had higher values 〈dMRCA〉 on average (6 subjects, median 1.0 SNPs/cell, 2.2–3.4 years; p<.02, Wilcoxon rank sum test), consistent with a longer duration of infection in patients who are not immunocomprised^23,24. For all subjects with de novo mutations, even when 〈dMRCA〉 was small, diversification led to coexisting short-branched sub-lineages (Fig. 2c).

The diversity arising from de novo mutation and mixed infections was heterogeneously distributed across lung sites. In subjects with mixed infections, strain distribution varied considerably across sites, often with only one strain detected per lung site (Fig. 3a). Considering only de novo mutations, cells from the same site were significantly more similar to one another than cells from different sites (one-sided, two-sample t-test, p<0.05 for 17/30 subjects with multiple sites and at least one de novo mutation; Fig. 3b). Inspection of intrapersonal phylogenies further revealed that minority genotypes can be localized to a single site and that some sites were dominated by minority sub-lineages not seen elsewhere within the lungs (e.g. Fig. 3c). This spatial heterogeneity was neither universal nor long-lived; many new mutations were shared across lung sites and no long branches were unique to a single lung or site. These results are consistent with reports of spatial heterogeneity in other lung infections⁹ and in M. tuberculosis infections of HIV negative individuals^25–27. Whatever the mechanism, this heterogeneity meant that many de novo mutations and strains infecting subjects’ lungs were undetected in tracheal aspirate samples (Fig. 3d; Supplementary Fig. 5). While sputum samples were not available, this suggests that spatial heterogeneity can compromise the reliability of antibiotic resistance profiling and other assays performed even on the entire diversity within respiratory samples²⁸.

Figure 3. M. tuberculosis diversity within the lungs is spatially structured.

For each subject, we compared the distribution of strains, mutations, and genotypes across the lungs and tracheal aspirate, averaging across samples from each site. (a) The relative abundance of both strains in each respiratory site (Tracheal aspirate and lung sites 1–6, left to right) is indicated for subjects with mixed infection. Different shades of grey represent the two strains. Error bars indicate the standard error of the mean (s.e.m.) for subjects with multiple samples per site. Thin black lines indicate the average across sites from each subject. (b) For each subject with a single strain infection and multiple samples per site, the average distance between two cells from different lung sites is plotted against the average distance between two cells from the same lung site (averaging across sites). Error bars indicate the s.e.m. across sites or pairs of sites. (c) For two example subjects, the relative abundance of each genotype within each respiratory site is indicated (Tracheal aspirate and lung sites 1–6, left to right). Color represents genotype; the phylogenetic relationships of these genotypes are shown at left. (d) For each subject, the number of de novo mutations detected within the subject is plotted against the number of de novo mutations detected only within the tracheal aspirate. Low abundance mutations were the most likely to be missing from sputum (Supplementary Fig. 5a–b). Jitter added for visibility (Up to 0.5 mutations; same value added in x and y).

Surprisingly, the genetic distance between cells from different sites, both within the lungs and between organs, did not depend on spatial proximity. Cells from sites in different lungs (right vs. left) were, on average, no more different than cells taken from sites in the same lung (Fig. 4a). Further, two cells from different organs were no more different than two cells taken from different lung sites (Fig. 4b; Supplementary Fig. 6). Moreover, genotypes from the liver or spleen sites did not cluster phylogenetically, suggesting dissemination of multiple genotypes to each of these organs (Fig. 4c). To compare rates of transmission between organs and within the lungs, we estimated the number of transmitted genotypes between the lungs and extrapulmonary organs inferred by their joint phylogeny (Fig. 4c, Online Methods). We inferred multiple transmitted genotypes between the lungs and extrapulmonary organs for many subjects (Fig. 4c–d). These multiple genotypes might result from a single transmission event involving multiple genotypes or repeated transmission events²⁹; we note that the detection of different genotypes in the liver and spleen of the same individual suggests independent multiple transmission events. Interestingly, the number of genotypes transmitted from the lungs to extrapulmonary sites was as high as the number of transmissions between different lung sites (Fig. 4d).

Figure 4. M. tuberculosis dissemination within lungs, between lungs, and between organs follows similar dynamics.

For each subject with a single strain infection: (a) the average distance between cells from the right lung and left lung is plotted against the average distance between cells from the same lung (right lung, dark blue; left lung, light blue); and (b) the average distance between a cell from a lung site and a cell from an extrapulmonary organ is plotted against the average distance between cells from the different lung sites. Error bars indicate s.e.m. across sites or pairs of sites. (c) Examples of intrasubject phylogenies showing shared genotypes. Genotypes are indicated as within an organ with a colored box in the corresponding column if they were found above 20% frequency in at least two samples from that organ. Potential transmission events between the lungs and extrapulmonary organs are identified by mutations shared across organs and indicated with a double-sided arrow of corresponding color. (d) For each subject, the minimum number of transmitted genotypes to the liver and/or spleen from the lungs was inferred (y-axis) and compared to mean of the minimum number of transmitted genotypes to each lung site from the rest of the lungs (averaging across sites, x-axis). Simulations were used to normalize for sampling efforts (Online Methods). Error bars indicate the s.e.m. across lung sites. Jitter added for visibility (Up to 10%; Same value added in x and y). Magenta indicates liver and purple indicates spleen.

The similar patterns of migration within and between organs suggest that dissemination of M. tuberculosis across the lungs of these subjects is no easier than dissemination across organs. These results support the hypothesis that both intralung and interorgan spread are mediated by similar dissemination mechanisms, such as by macrophage trafficking through the bloodstream or other mechanisms of hematogenous spread^30–32. It would be interesting to investigate the generality of these results in subjects with less advanced disease or without HIV.

The diversity of genotypes within individuals and their stratification across the body have implications for outbreak tracing. A major impetus for whole-genome sequencing of M. tuberculosis, as well as other pathogens, has been to identify transmission links between individuals. While future approaches will incorporate the pathogen diversity within each person^33–35, current approaches establish an epidemiological linkage between people if the mutational distance between single samples from each patient is below a set threshold^7,8,22,36. Considering all genotypes within a subject, 20% of subjects harbored genotypes separated by more than a common threshold of 5 SNPs⁸ (Supplementary Table 2; 9% emerging from multiple-strain infection) and 11% harbored genotypes separated by more than an alternative threshold of 12 SNPs³⁷. These estimates may be even greater when considering longer duration infections, such as those of HIV negative individuals. While the thresholds chosen will depend on the mutation detection sensitivity unique to each study, the frequency of substantial within-patient distances adds weight to the mounting evidence^11,33,38,39 against ruling out epidemiological links on the basis of mutational thresholds between single samples.

This study represents the first large-scale genomic investigation of Mycobacterium tuberculosis diversity across the human body. Extensive sampling from multiple body sites and detection of minor alleles within each sample shows that even the notoriously slowly evolving M. tuberculosis presents substantial within-patient diversity. The distribution of this diversity across the body suggests that, at least in people with HIV, dissemination within the lungs follows a similar mechanism as dissemination from the lungs to extra-pulmonary organs. In total, these findings highlight the potential of whole-genome sequencing to make inferences about the history of individual bacterial infections from the diversity surveyed at a single time point. We anticipate that future studies will extend these approaches to address open questions about M. tuberculosis and other pathogens, including the contribution of spatial heterogeneity to evolution of antibiotic resistance and the dynamics of dormancy and reactivation.

Methods

Study cohort

Subjects were eligible for inclusion in our study if they died while at Edendale Hospital, were age 20 or over, were on anti-tubercular treatment for fewer than 4 days before death, and were either smear positive for acid fast bacilli or had at least one focal process compatible with tuberculosis without an alternative diagnosis. The number of subjects studied was determined based on feasibility. Subjects were de-identified prior to DNA analysis and mutation calling was performed while blind to subject information.

Sample collection

Following informed consent by family members under a protocol approved by the institutional review board at University of KwaZulu-Natal, a minimally invasive autopsy was performed for each subject. Briefly, three lung sites were collected from each side, at the second, fourth and sixth intercostal spaces. A new sterile needle was used at each site (see Supplementary Note for more detail). Specimens were stored on ice, and transported to the University of KwaZulu-Natal in Durban. Participants not documented in the case notes to be HIV seropositive were tested post-mortem using a lateral flow ELISA assay. Subject information is listed in Supplementary Table 2. Cause of death for all subjects was culture confirmed disseminated tuberculosis.

Biopsy specimens were cut into small pieces and decontaminated using Petroff’s method⁴⁰ (to enrich for M. tuberculosis). Endotracheal aspirate specimens were decontaminated with a 4% NALC/NaOH solution. 500μl of the decontaminated specimen was inoculated into a MGIT tube and 100ul on 3 separate 7H11 plates⁴¹ to select for M. tuberculosis. For the first 12 subjects, the diversity on these 3 plates was scraped into a single tube for DNA analysis. For the remaining cultures, up to 5 single-colonies, demonstrating morphological variation, were selected from each plate (for a maximum of 15 colonies per specimen). Each colony was then sub-cultured onto dedicated 7H11 plates and incubated for three weeks at 37°C, under aerobic conditions.

Antibiotic resistance testing

Phenotypic antibiotic susceptibility testing was performed on all decontaminated specimens. The modified proportion method was used for antibiotic susceptibility testing. Positive MGIT tubes from each specimen were used to set up the antibiotic susceptibility tests. Tubes were confirmed to be uncontaminated and incubated for a further 24–48 hours after becoming instrument positive before the test was performed. Three week 7H11 cultures were used in the event the MGITs were found to be contaminated or to confirm resistance.

Quadrant 7H10 agar plates that comprised a control (antibiotic free) quadrant and antibiotic containing quadrants were used. The following concentrations were used (μg/mL): Isoniazid: 1, Rifampicin: 1, Ethambutol: 7.5, Streptomycin: 2, Ofloxacin: 2, Kanamycin: 5, Ethionamide: 5, Capreomycin: 10. Cultures were diluted so as to obtain approximately 100 colony forming units (CFU) on each quadrant and seeded onto each quadrant. The plates were sealed in CO2 permeable bags and incubated at 37°C, 5% CO2 for one week and under aerobic conditions for a further two weeks. Control strains (H37Rv and A169) were set up in parallel. Antibiotic containing quadrants with CFU > 1% of the CFU on the control quadrant was interpreted as resistant. No growth or <1% was interpreted as susceptible. Cultures were read in a double blind manner.

No significant variation was found across specimens from a given subject. Specimens from subjects P4, P9, and P12 were multi-drug-resistant, specimens from P16 were rifampicin resistant, and all others were pan-sensitive (Supplementary Table 2).

DNA extraction

M. tuberculosis colonies were scraped from agar and inoculated into 500μl of water in an eppendorf tube and heat killed for 30 minutes at 80°C. DNA was extracted using the CTAB method⁴². 70μl 10% sodium dodecyl sulfate and 50μl proteinase K (10 mg/ml stock solution) were added and the solution incubated for 1 hour at 60°C using a thermomixer on low mode shaking. Samples were removed from the thermomixer, and 100μl 5M NaCl (preheated to 60°C) was added and mixed thoroughly by hand inversion.. 100μl of 10% CTAB (N-acetyl-N,N,N-trimethyl ammonium bromide, preheated to 60°C) was added and the solution mixed by hand inversion. Tubes were then incubated at 60°C for 15 minutes (thermomixer on low-mode shaking). 700μl chloroform/isoamyl alcohol (24:1) was added and the solution mixed by hand inversion. Tubes were then centrifuged for 10 min at 13,000rpm. The upper (aqueous) phase was transferred to tube with 700μl of a cold isopropanol and mixed by hand inversion. DNA was precipitated by freezing –20°C for at least 30 minutes and centrifuged for 10 minutes at 13,000rpm. Tubes were drained and the pellet washed with 80% ethanol followed by centrifugation for 5–10min. These were then drained and the pellet was allowed to dry. DNA was suspended in 55μl of TE.

Illumina sequencing and mutation identification

Genomic libraries were constructed and barcoded using a previously described modified version of the Illumina Nextera protocol⁴³. Genomic libraries were sequenced on the Illumina HiSeq 2000 platform by Macrogen using paired-end 100 bp reads. Reads were aligned to a modified H37rv reference genome in which each nucleotide reflects this ancestor of the M. tuberculosis complex³ (provided by Iñaki Comas). The limited mutation distance of all extant TB to this ancestor (< .05% of the genome) and limited recombination in M. tuberculosis³ gives us confidence that each difference from this reference that passed our conservative filters (Supplementary Note) represents a newly derived mutation. Standard approaches were used for read filtering and alignment⁴⁴. Average coverage for each sample is listed Supplementary Table 4.

Potential cross-contamination or mislabeling events were identified by the presence of multiple mutations not found in any other sample from that subject and/or multiple mutations found in samples from different subjects processed on the same plate. We identified 68 such potentially contaminated samples. New repeated genomic libraries were prepared for 40 of these 68 samples. In 31 cases, the duplicate matched more closely with other samples from that subject and was used for analysis. In 9 cases, the duplicate also showed evidence of contamination and the two replicates of the sample were discarded (for a total of 37 discarded samples out of 2730, 1.4%). In our final data, subjects had distinct strains (Supplementary Fig. 7). As some of these samples may have been discarded in error, some true mixed infections may have been called as simple.

Many single-colony samples from subjects with mixed infections had hundreds of mutations at intermediate frequencies—suggesting that cells from different strains were comingled in a single colony (Supplementary Fig. 1c; consistent with the cording phenotype of M. tb⁴⁵). We therefore searched for both fixed and polymorphic mutations within each sample, using a series of filters and statistical tests to distinguish false positives (e.g. alignment and sequencing errors) from real polymorphisms⁴⁴. Additional filters regarding the proportion of samples in each subject with a polymorphism were also used to remove notoriously problematic sites, such as those in the proline-glutamate and proline-proline-glutamate genes⁴⁶ (see Supplementary Note for details). Among these filters, polymorphic mutations were required to be at least 10% frequency and supported by at least 4 reads on each strand. As a consequence, rare variants may have not been detected, particularly in cases where few samples were available per site.

For each polymorphic position that met the filtering criteria in at least one sample, raw reads were used to determine the allele frequency across samples from that subject. The 518 de novo mutations found and their distribution across samples are listed in Supplementary Table 4. Mutations and genes were annotated according to H37rv (NC_000962.3). None of these mutations are known to be associated with antibiotic resistance⁴⁷.

dN/dS

Calculations for dN/dS, the ratio of nonsynonymous mutations to synonymous mutations divided by the ratio under a neutrality, were performed normalizing for the spectrum of mutations observed as previously described⁴⁴. P-values for depletion of non-synonymous mutations were calculated according to the binomial cumulative distribution function.

Identification of polymorphisms arising from multiple-strain infection

Mutations carried by each of the infecting strains covary across samples with one another and anti-covary with mutations from the other strain⁴⁸. Mutations arising from multiple strain infection are defined by their contribution to the primary principal component (PC1) in a Principal Component Analysis of the mutation frequency matrix. De novo mutations are identified as those without a significant contribution to PC1 (absolute value < 0.015, cutoff determined empirically; Supplementary Fig. 1). Previously described SNP markers were used to assign each strain to a global lineage⁴⁹.

Estimation of 〈dMRCA〉

The average mutational distance across cells in a subject’s M. tuberculosis population to its most recent common ancestor, 〈dMRCA〉, was calculated as the sum of the mutation frequencies at each polymorphic position called within the subject⁴⁴. To normalize for sampling efforts, 〈dMRCA〉 was first calculated across samples from each specimen, and these values were averaged for each subject. For subjects previously diagnosed with HIV, the mean value of 〈dMRCA〉 was 0.42 SNPs/cell and for subjects who were HIV negative or diagnosed post-mortem, this mean was 1.8 SNPs/cell (Fig. 2a, Supplementary Table 2).

Interpretation of 〈dMRCA〉 as time of infection assumes that these subjects were infected with a single genotype; if subjects were infected by multiple closely related genotypes, these estimates are inflated. Alternatively, these estimates may be underestimations if mutations have swept the subject’s diversity since infection due to adaptation or bottlenecks (drift). Estimates of 〈dMRCA〉 for subjects with multiple-strain infection are lower bounds because the strict mutation caller was optimized for detecting de novo mutations. Other potential sources of error include Poisson error in the number of mutations accumulated in each lineage since each subject’s MRCA, underestimation due to limited sensitivity in detecting mutations, overestimation due to false positives, incorrect designation of ancestral versus derived alleles, and underestimation due to incomplete sampling of the diversity within each subject.

Genotype identification and phylogenetic inference

We used a conservative algorithm to infer a minimal set of distinct genotypes within each subject based the assumptions that genotypes are shared across samples from each subject and that each SNP occurred only once within a subject (strict parsimony). A direct consequence of the strict parsimony assumption is that two mutations at high frequency (majority) in the same sample must coexist on the same genotype^50,51. For each sample within a subject, starting with the most homogenous and covered samples, high frequency polymorphisms were matched with previously identified genotypes, or if no such match was made, the polymorphisms found at high frequency were used to define a new genotype (Supplementary Fig. 2). Some de novo mutations could not be confidently assigned to a genotype and were omitted; combined with our stringent mutation caller, this approach produced conservative phylogenies in which some genotypes were not identified or called with fewer mutations than they really have. After genotypes were identified, each sample was then assigned to one or more genotypes. We note that our algorithm will not work well in cases where parallel nucleotide evolution is expected or where pure samples are not available and that the cutoffs in our implementation may need to be adjusted for different use cases (Supplementary Note).

Pairwise genetic distances

We calculated the average pairwise distance between cells in different samples, considering the frequency of each mutation f at each confident polymorphic position x on the genome, in samples i and j as:

$$\sum _x{f}_{xi}+f_{xj}-2f_{xi}{f}_{xj}$$

Only lung, liver, and spleen sites with more than 2 samples per site were included in the analysis. To normalize for different sampling efforts across sites, we first calculate the mean at each site (x-axis, Fig. 3b) or pairs of sites and then average across site or pairs. Error bars indicate the s.e.m. of the means across sites or pairs of sites.

Transmission analysis

We developed an algorithm based on the number of shared mutations between sites (Fig 4c). Only genotypes found in two separate samples within an organ or lung site were considered to reduce the effects of potential cross-contamination. We iterated through mutations shared between the destination site and the rest of the lung. On each iteration, the genotype closest to the subject’s MRCA containing a shared mutation was chosen and counted as a transmission event. All mutations in that genotype were removed from the list of unaccounted for shared mutations, and the next iteration was performed. In this way, groups of mutations appearing on the same branch of the phylogeny were only counted as a single transmitted genotype. Lastly, the presence of the subject’s MRCA in both locations suggests a transmission of this ancestral genotype (Fig 4c.). Simulations were performed to normalize for different sampling efforts across sites and organs, omitting one site from the lung each trial (for estimations of intraorgan transfer, this omitted site was then treated as destination). Only sites with more than 6 samples were included and each site was randomly subsampled, taking the minimum number of samples across sites with >6 samples. For each omitted lung site, 1,000 trials were performed and averaged.

Code availability

Custom MATLAB code used for calling mutations and analyzing the data can be found at: https://github.com/kishonylab/TB-diversity-across-organs.