We aimed to study how acquired missense variation in the M. tuberculosis proteome distributes in the structure of each protein. To do this, we compiled a dataset of genomes of 31,428 isolates from the Mycobacterium tuberculosis complex (MTBC). We performed ancestral sequence reconstruction to determine the number of independent arisals of each mutation (homoplasy) (Supplementary Data 1).^16,17 The dataset represents a diversity of MTBC isolates: with 2,815 isolates from lineage 1; 8,090 from Lineage 2; 3,398 from Lineage 3; 16,931 from Lineage 4; 98 from Lineage 5; and 96 from Lineage 6.

We filtered the 782,565 unique SNVs and 47,425 insertion/deletion mutations (INDELs) found in the original dataset to focus only non-synonymous mutations in protein coding genes, or INDELS that preserve the original translation frame. We count the number of independent arisals (homoplasy score) for each SNV and INDEL. As we are looking for positive selection that alters but does not completely ablate protein function, we do not consider frameshift mutations. After excluding mutations occurring in regions where variant calling is computationally challenging with short-read sequencing,^27,28 we observe a total of 469,942 total unique non-synonymous SNVs and 5,104 total inframe indels (Methods, Supplementary Data 2). On a per-protein basis, a mean of 32.0% of sites have at least one mutation across the dataset of 31,428 isolates, with a mean of 1.4 mutations per mutated site. We observe that proteins with the highest total number of mutations are those known to be involved in resistance to first line antibiotics (Figure 1).

We identified a real or predicted structure for each protein in the M. tuberculosis proteome. We used a sensitive pipeline to search the RCSB Protein structure databank (Methods) for structures homologous to M. tuberculosis proteins. We find a protein structure that represents at least 90% of the protein sequence for 34% of proteins (N = 1,371). For the remainder of proteins, we use the structure downloaded from the AlphaFold protein structure database, removing residues where the structure prediction is uncertain (pLDDT < 70). We exclude proteins with known low accuracy of variant calling (Methods),^24,27 or with fewer than 30 residues represented in the filtered protein structure, for a final dataset of 3,687 proteins out of 3,996 annotated ORFs in the M. tuberculosis H37Rv reference proteome (Supplementary Data 3).

To calculate clustering of mutations, we use an approach from geographic analysis that calculates spatial autocorrelation between statistics, called the Getis-Ord G-statistic.^33,34 Briefly, for each residue i in the protein, it calculates a z-score that computes if residue i is more proximal to highly mutated residues than would be expected by chance. This approach was previously applied to three-dimensional structure data by PIVOTAL to prioritize mutations associated with human disease.³⁵

We examine the raw Getis-Ord score, which can be interpreted as a z-score, on the structures of 9 proteins in which mutations are known to confer antibiotic resistance³² – RpoB, EmbB, KatG, InhA, PncA, GyrA, RsmG (GidB), EthA, and Rs12 (Methods). For each protein, we observe a region with residue G-scores greater than 5 that are clustered in a single location (Figure 2). This finding is expected for essential drug target proteins where resistance-conferring mutations preserve protein function while occluding antibiotic binding, e.g. RpoB, EmbB, InhA, GyrA, as mutations will tend to cluster in the antibiotic binding sites of those proteins. Notably we also identify significant clustering for proteins that are not essential to the cell, where any mutation that disrupts the protein can be resistance-conferring, e.g. PncA. We observe that the residue G-score distribution demonstrates a periodic pattern across the gene length in which higher scores alternate with lower scores even when the high G-score residues cluster in a single protein location (Figure 2), a signal which emerges due to the 3D structure of the protein.

The G-score provides a residue-level z-score measuring three-dimensional proximity to other residues with high numbers of amino acid substitutions. We postulate that neutral amino-acid substitutions are more evenly distributed in 3-D space than functional substitutions that indicate selection, due to neutral substitutions being more likely to occur at the protein surface or flexible regions of the protein (e.g. loops and termini)⁴⁰, whereas amino acid substitutions that indicate selection on the protein are more likely to cluster in protein-protein interaction interfaces or protein cores.

From these postulates, we evaluated five candidate protein-level statistics to quantify clustering of mutations possibly indicative of positive selection (Methods, Supplementary Data 4). We selected as a positive control set of proteins the nine above proteins known to be under selection for antibiotic resistance, with demonstrated residue-level clustering– RpoB, EmbB, KatG, InhA, PncA, GyrA, RsmG (GidB), EthA, and Rs12. Because each of these proteins in an outlier in terms of the overall number of mutations observed (Figure 1B), we reasoned that downsampling the number of mutations would produce controls that were more similar to the average protein in the proteome. Hence for each of the nine proteins, we generate 100 positive control examples by downsampling the number of mutations observed while keeping the same relative probabilities of mutations at each site (Figure 3, Methods), thus preserving the signal for clustering while forcing their total number of mutations to be similar to that observed across all proteins in the proteome. We generated negative controls for the same nine proteins by simulating mutations from a uniform distribution across all residues, setting the mutation rate to the mean per-site mutation rate across the proteome, and generating 100 negative controls per protein. We find the Kolmogorov-Smirnov statistic to have the highest discrimination between positive and negative controls (95% precision, 68% recall, at significance level p < 0.01).

We test all 3,687 proteins in the M. tuberculosis structure and homoplasy dataset, and identify 499 proteins with significant clustering (Figure 4, Supplementary Data 5).

We assay whether all proteins known to cause antibiotic resistance demonstrate mutational clustering, using the WHO catalog of resistance-conferring mutations³². The majority of proteins with Tier-1 resistance conferring mutations (8 of 12, 67%) demonstrate significant clustering. The four proteins without clustering include TlyA and RplC, in which there is only one non-synonymous substitution known to confer resistance, and MmpR5 a transcriptional regular protein involved in resistance to the newly administered drug bedaquiline. The fourth protein is RpsL (RS12_MYCTU), whose G-score is driven by just two highly mutated residue positions, K43 and K88 (Supplementary Figure S1). In a small protein such as RpsL (124 amino acids), having two highly mutated residues close in 3D structure occurs with some frequency in the negative control examples.

For 90.6% (452 of 499) of the significant hits, the top G-score pair of residues are within 15 Ångstroms in 3-D space, confirming that the identified clusters are in a single location within the protein. Proteins with a high number of mutations that are not spatially clustered are not significant hits. This includes Cas10, which has 1861 independent mutation events, 1489 of which occur in the same amino acid position.

We perform a Gene Ontology Enrichment analysis on the 452 hits with a close proximity top pair to understand their possible functional implications (Methods). After removing the less specific gene categories (with >20 genes), 54 GO categories are significantly enriched (FDR < 0.05) (Figure 4). The most enriched categories by fold change (Supplementary Table 1, Supplementary Data 6) are regulation of fatty acid metabolism and biosynthesis; UDP-N-acetylglucosamine/amino sugar metabolic processes; entry of bacterium into host; protein maturation; DNA topological change; and energy-coupled proton transport. We find clustering in the DNA topological change proteins GyrA and GyrB, both known to confer resistance to fluoroquinolones, and Top1, DNA topoisomerase I, which is not known to be associated with resistance.

A second notable GO category is composed primarily of the Pkn family of protein kinases – PknA, PknB, PknD, PknE, and PknH – thought to negatively regulate fatty acid biosynthesis. Data supports that PknA, PknB, and PknH have a role in intrinsic drug resistance, potentially through post-translational modification of proteins involved in cell wall synthesis.^41–43 For all of the Pkn proteins found by our clustering score, the region of significant clustering is not the protein kinase domain but instead in other functional domains of the protein. For example, the significant clustering in PknB is in the extracellular domains, and in PknH is in the putative transmembrane helix and sensor domain⁴⁴ (Figure 4). Therefore, we suspect that the mutations may relate to regulation of activity rather than the kinase reactions themselves.

A third notable group of proteins is involved in amino sugar metabolic processes: GlmM, GlmS, GlmU, and MurA. The related protein MurD was previously identified in a screen for targets of convergent positive selection in the M. tuberculosis proteome.⁴⁵

We find significant clustering in RnJ, Ribonuclease J, which has previously been identified in GWAS for antibiotic resistance,⁴ and whose loss is known to lead to multi-drug tolerance.⁴⁶ which has been shown to have both ribonuclease and beta lactamase activity. We find significant clustering around the AlphaFill-inferred binding site of ribonucleotides (Methods), which corresponds to the known active site for RNase activity (Figure 4).⁴⁷

Given that most drug resistance genes demonstrate 3-D mutational clustering, as detailed above, we tested if structural proximity to known mutations carries information on the functional impact of new mutations on phenotype. We use a previously compiled catalogue of resistance-associated mutations from M. tuberculosis,⁸ consisting of 583 unique non-synonymous protein-coding variants annotated as associated with resistance (R-assoc) and 58 as not associated and isolates carrying only these mutation are antibiotic susceptible (S-assoc) (Methods, Supplementary Data 7)³².

We ask whether protein structure information is useful at distinguishing resistance-conferring variants. In addition to the G-score computed using the homoplasy data from the ∼31,000 isolate dataset, we compute the distance in 3D between any variant and the nearest resistance-conferring variant in that protein from the catalogue. We benchmark both metrics against distance in 1D (on the protein sequence). We trained a logistic regression classifier to predict whether a mutation is resistance associated (R-assoc, vs. S-assoc, susceptibility associated), using a 70-30 train-test split of the cataloged mutations. 3D structural proximity alone best predicted resistance association (F1 score of 94.6%, Table 1, Methods). Prediction from 1D proximity alone is less accurate at F1 of 92.8%, and prediction from G-score had the lowest F1 at 80.8%. The relatively strong performance of 1D proximity prediction is due to the fact that most, but not all, mutations in our dataset that have close 3D proximity to a mutated residue also have close 1D proximity.

While the G-sore prediction has the lowest F1 score and sensitivity, it has the highest precision (98.3%). This may be because the G-score is a combination of structural information and mutational information, as a residue will only have a high G-score if it is proximal to at least one residue with a high number of mutations (presumably due to being causal of resistance).

This means that the G-score model is more conservative at calling R-assoc variants as truly R-assoc.

We search the RCSB Protein Data Bank structure database (download date: Feb 5, 2021) for experimentally determined structures with sequence similarity to the Uniprot Proteome of Mycobacterium tuberculosis (ID = UP000001584). In order to execute a sensitive search for homologous structures, we built a modified version of the EVcouplings pipeline,⁵⁰ which operates in two stages: first, we construct sequence alignments against the Uniprot sequence database (download date Feb 2021) at bitscores of 0.1, 0.2, and 0.3 times the query sequence length, using jackhmmer from hmm-suite v3.1 with five iterations⁵¹. Second, we use the hmmbuild and hmmsearch tools of hmm-suite to search the RCSB PDB sequence database using the constructed alignments.

We then aim to select high-coverage protein structures from the experimental structure database to represent the M. tuberculosis proteins. We define coverage as whether an amino acid residue in the M. tuberculosis protein is represented by a resolved amino acid in the protein structure. Of the 3993 proteins in the proteome, 2984 (75%) have any hits to the structure database. 1509 (38%) of these hits have at least 90% coverage. In cases where more than one hit had 90% coverage, we select the representative hit in the following way: if a protein structure is from M. tuberculosis, we select that structure, otherwise we select the hit with the highest e-value.

Predicted structures for the Mycobacterium tuberculosis reference proteome (ID = UP000001584) were downloaded from the AlphaFold Protein Structure Database on Feb 15, 2023. Residues with pLDDT < 70 were removed from each structure.

We exclude from consideration proteins with documented difficulties in mutation calling using short-read sequencing. Using data from Marin et al²⁷, proteins with <90% mean empirical base pair recall (across all residues in the protein) or <90% mean residue mappability were removed from consideration. Proteins with fewer than 30 residues with resolved structure coordinates were removed, for a final total of 3687 proteins, 1350 of which are experimentally determined structures and the remaining 2337 of which are from AlphaFold.

For SNP mutations, we consider all non-synonymous genic SNPs that occur at least once in our M. tuberculosis dataset, for a total of 476,607 unique polymorphisms. For insertion and deletion (indel) mutations, we exclude frameshift mutations as these are likely to ablate protein function and thus are not analogous to non-synonymous mutations. We observe 5,678 unique in-frame indels. Combining the SNP and inframe indel mutations, we observe 483,117 unique mutations. To account for possible artifacts from short-read sequencing base calling errors, we exclude mutations in positions designated as blindspots by Modlin et al or with empirical base-pair recall <90%,^27,28 as well as all mutations occurring in proteins with <90% mean empirical base-pair recall or <90% mean mappability.²² After filtering, we have 475,046 unique polymorphisms for downstream analysis (469,942 total unique non-synonymous SNVs and 5,104 INDELs) (Supplementary Data 2).

Upon merging our structure dataset with our mutation dataset, 335,224 mutations can be mapped to a residue in a protein with high-quality structure information. We exclude proteins with fewer than two mutations mapped to their structure from the clustering calculation. Indel mutations were treated as occurring at the first position in the sequence. We proceed with clustering analysis on 3657 proteins with mutations mapped.

Our implementation of the Getis-Ord score for protein three-dimensional structure was inspired by PIVOTAL, which applied the Getis-Ord score to clustering of mutations in human disease.^{33– 35} The two values input to the Getis-Ord statistic computation are a per-residue score x, here the per-amino acid homoplasy score, and a weight matrix W that contains the inverse of the inter-residue distances. x is an L x 1 vector where the entry x_i contains the homoplasy score for the i-th residue in the protein, and W is an L x L matrix with entries defined as: Where d_i,j is the minimum inter-residue atomic distance, in Å, between i and j. Thus, residues that are closer in three dimensions will have higher weights.

The Getis-Ord statistic for residue i is calculated as: Where is the mean of x, and:

We sought to generate a dataset of positive controls – proteins known to have clustering of mutations – with a frequency of mutation similar to the average protein in our dataset (mean of 1.4 mutations per site). For this, we used the nine proteins known to be involved in antibiotic resistance with demonstrated clustering of mutations (RpoB, EmbB, KatG, InhA, PncA, GyrA, RsmG (GidB), EthA, and Rs12). For each of these nine proteins, we generate 100 positive control examples.

Because the total number of mutations in a site, as well as the 3D configuration of sites, contributes to the G-score, we chose to downsample these proteins to generate positive controls that are more similar to other proteins in the total number of mutations. To generate positive control examples, we build an empirical distribution based on the observed number of mutations per site in the protein, divided by the total mutations. We add a pseudocount of one to all residues with zero mutations. We then sample M mutations from the empirical mutation distribution, where M = 0.42 times the number of residues in the protein with structure data, because each site is mutated on average 0.42 times. This sampling strategy preserves the relative frequencies of each mutation, and their 3D location, while greatly reducing the number of total mutations observed in a protein. We generate the negative controls by sampling from a uniform distribution over all residues in a protein with structure data.

Then, we build a heuristic based on the Getis-Ord score to determine if we observe significant clustering in a protein. For each control example, we perform 10,000 random permutations of the mutations, shuffling their configuration in space but keeping the number of mutation events the same. For positive controls, we expect the real configuration of mutations to produce a significantly different configuration of G scores than random permutations. For negative controls, we do not expect the real configuration of mutations to produce a significantly different configuration of G scores than random permutations.

Next, for each generated control, we test the ability of various scores to distinguish between the real example and a random reshuffling. By comparing the true distribution with the empirical random distribution, we compute a p-value for whether the overall distribution of Getis-Ord statistics for residues in the protein is different than the one expected by chance. We compare six scores on the basis of their precision recall curves for recovering the positive controls. The first three scores are based on comparing the p-value of the Kolmogorov-Smirnov test when comparing the real distribution to 10,000 random shuffles: the number of significant p-values (p < 0.01), the smallest p-value, and the mean p-value. The other two features are based on the raw G-score for the real distribution vs. the 10,000 random shuffles: how often the max G-score is greater for the real distribution, and how often the median G-score is greater for the real distribution.

We then run the G-score calculation on the whole proteome using the above pipeline. We apply our 95% precision threshold to find 451 initial hits.

We test whether our gene hits are enriched in particular GO functional categories.^52,53 We use the online GO enrichment tool (https://geneontology.org/) to search for enriched biological processes among our 452 hit proteins. We downloaded the.json file from the GO enrichment tool, and filtered for GO categories with fewer than 20 members in the M. tuberculosis H37Rv refence genome, to remove overly broad categories like “biological process,” and terms with identical constituent genes. We retained categories with FDR < 0.05, for a total of 37 enriched terms.

We downloaded hits from the AlphaFill v1 database (access date: 7 October 2024) to find potential ligand-binding locations in our proteins. For RnJ, two ligands are in proximity of the high GeO score regions: “U5P” and “C5P”, uridine-5’-monophosphate and cytidine-5’-monophosphate.

The World Health Organization (WHO) Mutation Catalogue maintains a list of over 30,000 unique variants observed in M. tuberculosis genomes and whether those mutations are diagnostic of antibiotic resistance against 13 antibiotics⁵⁴. Mutations are graded with five confidence categories: (1) Associated with Resistance, (2) Associated with Resistance - Interim, (3) Uncertain Significance, (4) Not Associated with Resistance - Interim, and (5) Not Associated with Resistance. Kulkarni et al. provided an update to this catalog which increased the number of labelled variants using regression-based grading based on frequency of mutations in resistant and susceptible isolates.⁸

We limit our analysis the non-synonymous variants observed in the following 15 protein-coding genes: Rv0678, atpE, ddn, embB, ethA, gid, gyrA, gyrB, inhA, katG, pncA, rplC, rpoB, rpsL, tlyA. After removing duplicates – because multiple genomic variants can cause the same non-synonymous substitution – we were left with 9365 variants, composed mostly of uncertain variants. After filtering for variants with structure information, we have 583 R-assoc (category 1 or 2), and 58 as S-assoc (category 4 or 5).

For each of 9365 unique non-synonymous variants in the catalogue,⁸ we extracted the following features: The minimum coordinate difference to the nearest non-self R variant along the amino-acid sequence (“1D proximity”). The inter-atomic distance to the nearest non-self R variant in Ångstroms (“3D proximity”), the G-score of the residue as computed in our previous analysis. 3D proximity was calculated using the EVcouplings Python package.⁵⁰

We used a 70-30 train-test split to construct a dataset. We employed weighted sampling to address the class imbalance between resistance-conferring and non-resistance-conferring variants. We employed a logistic regression classifier in scikit-learn, and hyperparameters were tuned by grid search.⁵⁵ Models were evaluated for F1-Score, precision, and recall. To select an optimal model threshold, we choose the threshold which maximizes the sum of sensitivity and specificity.⁵⁶