The structural context of mutations in proteins predicts their effect on antibiotic resistance

eLife Assessment

This important study leverages a large global dataset of tens of thousands of tuberculosis samples to place recurrent protein-coding mutations into their three-dimensional structural context, offering an expanded view of how antibiotic resistance emerges compared to traditional genetic analyses alone. The strength of evidence is compelling, supported by the scale and breadth of the dataset and the systematic structural analysis, although some of the assumptions made in the modeling approach are only partially supported. Overall, the work will be of broad interest to researchers studying microbial evolution, antibiotic resistance, and structure-function relationships in pathogens.

https://doi.org/10.7554/eLife.109450.3.sa0

Significance of the findings:

Important: Findings that have theoretical or practical implications beyond a single subfield

Landmark
Fundamental
Important
Valuable
Useful

Strength of evidence:

Compelling: Evidence that features methods, data and analyses more rigorous than the current state-of-the-art

Exceptional
Compelling
Convincing
Solid
Incomplete
Inadequate

During the peer-review process the editor and reviewers write an eLife Assessment that summarises the significance of the findings reported in the article (on a scale ranging from landmark to useful) and the strength of the evidence (on a scale ranging from exceptional to inadequate). Learn more about eLife Assessments

Abstract
Introduction
Results
Discussion
Methods
Appendix 1
Data availability
References
Article and author information
Metrics

Abstract

In Mycobacterium tuberculosis, a prevalent and deadly pathogen, resistance to antibiotics evolves primarily through non-synonymous mutations in proteins. Sequence-based analyses can uncover the genetic basis of antibiotic resistance, but these methods focus on primary sequence and often neglect other biological signals, such as protein structural information. We hypothesize that integrating the structural context of mutations improves the prediction of effects on function and phenotype. We curate high-confidence structural annotations for the M. tuberculosis proteome from 1350 crystallography and 2337 AlphaFold predictions, and mutations from over 31,000 M. tuberculosis isolates. We demonstrate that mutations in proteins known to cause resistance are clustered in 3D space, even in proteins where inactivating mutations at any position are thought to cause resistance. We find over 450 proteins in the M. tuberculosis proteome that display signals of clustered mutations, many of which have a known relationship with antibiotic resistance. We show that a supervised classifier trained on 3D distance to known resistance sites alone has an F1 score of 96.5% at classifying mutations as resistance-conferring on a held-out test set. This work demonstrates that protein structure provides useful information for categorizing which variants may cause antibiotic resistance, even when the majority of structures are AI-predicted.

Introduction

The increasing prevalence of antibiotic-resistant Mycobacterium tuberculosis is challenging control of tuberculosis (TB), a disease responsible for the highest number of infectious disease-related deaths globally (World Health Organization, 2021b). Currently, diagnosis of antibiotic-resistant TB relies on time-consuming laboratory phenotype testing or the detection of resistance-conferring mutations in the M. tuberculosis genome through molecular assays or genetic sequencing (Global Tuberculosis Programme, 2021; Lange et al., 2018). Although the sensitivity and specificity of these assays are high for several first- and some second-line TB drugs, prediction is less accurate for other second-line antibiotics or novel agents like bedaquiline and pretomanid, which have recently become cornerstones of multi-drug resistant TB treatment. Improving the accuracy of resistance diagnosis relies on new computational approaches that can better link mutations with their functional impact on the resistance phenotype.

Two major computational strategies exist for identifying resistance-conferring variants: supervised approaches which associate genetic variation with resistance (Bush and Moore, 2012), and unsupervised approaches that search for signals of evolutionary adaptation, which can indicate resistance development. Supervised statistical methods such as genome-wide association studies and random forest classifiers have identified mutations associated with antibiotic resistance in M. tuberculosis (Farhat et al., 2019; Gröschel et al., 2021; Conkle-Gutierrez et al., 2022; Coll et al., 2018; Kulkarni et al., 2024), and machine learning approaches have built on this success to identify additional resistance-conferring variants (Green et al., 2022; Chen et al., 2019; Wang et al., 2024; Pruthi et al., 2024; Serajian et al., 2024; Kulkarni et al., 2025; Pal and Mohanty, 2024). But, these approaches require a large number of phenotyped isolates (both resistant and susceptible) to make accurate predictions, which presents a major limitation, especially for recently evolved mutations. Evolution-based approaches are an alternative for finding variants associated with antibiotic resistance by analyzing their mutational frequency and phylogenetic distribution: by searching for convergent positive selection in M. tuberculosis, studies have found mutations in the ald gene associated with D-cycloserine resistance (Desjardins et al., 2016), phase variation associated with virulence (Vargas et al., 2023), and mutations involved in host–pathogen interactions that potentiate the evolution of antibiotic resistance (Green et al., 2023).

Both evolution-based and supervised approaches consider only the DNA or protein sequence as input. They generally assume that all sites are equally likely to affect the phenotype, and consequently, many examples of a mutation are needed to infer significant effects. This assumption is not true: proteins have three-dimensional shapes and functional regions, and not every mutation is equally likely to impact function. Analyzing mutations in their three-dimensional context has uncovered hotspots of mutation in cancer (Miller et al., 2015; Gao et al., 2017; Kamburov et al., 2015; Meyer et al., 2016; Niu et al., 2016), and provided post hoc rationale for resistance-conferring variants in M. tuberculosis (Green et al., 2023; Phelan et al., 2016). Protein three-dimensional structure has shown utility as an input feature for identifying resistance-conferring variants in known resistance-conferring proteins such as RpoB (Lynch et al., 2025; Portelli et al., 2020), PncA (Karmakar et al., 2020; Carter et al., 2024; Dissanayake et al., 2025), and AtpE (Karmakar et al., 2019). While past work has sought to reannotate parts of the M. tuberculosis proteome with computationally predicted protein structures using older structure prediction methods (Modlin et al., 2021a), we can now infer a protein structure for nearly every protein in the proteome using AlphaFold (Jumper et al., 2021), leading to new works examining the 3D location of mutations in known and suspected resistance-conferring proteins (Pal et al., 2025; Wood et al., 2025).

In this paper, we use an unsupervised method to discover clustering of mutations in M. tuberculosis proteins by integrating a proteome-wide structural database with mutations from over 31,000 clinical isolates. We show that mutations display statistically significant clustering in resistance-conferring genes, even in non-essential proteins such as PncA where individually rare inactivating mutations are thought to cause resistance (Farhat et al., 2016). We identify over 450 proteins in the M. tuberculosis proteome that have significant clustering of mutations in their structures. Finally, we show that protein structural information provides a useful feature to predict whether variants are associated with antibiotic resistance across all proteins with resistance variants.

Results

Characterization of protein-modifying mutations in M. tuberculosis

We aimed to study how acquired missense variation in the M. tuberculosis proteome distributes in the structure of each protein. We chose to analyze the number of independent arisals of each mutation (homoplasy) rather than their population-level frequency because analyzing the frequency of alleles in a population can be biased by oversampling of particular lineages and by evolutionary recency. This ensures that more recent evolutionary events are not under-represented due to lack of time to spread in the population. To accomplish this, we used a previously compiled dataset of genomes of 31,428 isolates from the Mycobacterium tuberculosis complex (MTBC), with ancestral sequence reconstruction to determine the number of independent arisals of each mutation (Supplementary file 1; Vargas et al., 2023; Green et al., 2023; Vargas et al., 2021). The dataset represents a diversity of MTBC isolates, with 2815 isolates from Lineage 1; 8090 from Lineage 2; 3398 from Lineage 3; 16,931 from Lineage 4; 98 from Lineage 5; and 96 from Lineage 6.

We filtered the 782,565 unique SNPs and 47,425 insertion/deletion mutations (indels) found in the original dataset to focus only on missense mutations in protein-coding genes, or indels that preserve the original translation frame. We count the number of independent arisals (homoplasy score) for each SNP 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 (Marin et al., 2022; Modlin et al., 2021b), we observe 469,942 total unique missense mutations and 5104 total in-frame indels (Methods, Supplementary file 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- and second-line antibiotics (Figure 1). We also observe a large number of homoplastic mutations in the protein RpoC, in which mutations can compensate for fitness loss after evolution of rifampicin resistance via mutations in RpoB (Comas et al., 2011). Thus, this is a marker of, but not a direct cause of, antibiotic resistance. Another protein with many homoplastic mutations is Cas10, a component of the CRISPR–Cas system. The majority of mutation events (1534 of 1861) in this protein are due to a short in-frame indel at position 3,131,469 in the H37Rv reference genome, found in over 5000 isolates, which, given the repetitive nature of the sequence region, may indicate phase variation (Appendix 1—figure 1). Our analysis does not return ribosomal RNA genes, which are important causes of aminoglycoside resistance, because we are restricted to protein-coding genes. Finally, we note that we do not find large numbers of homoplastic mutations in genes encoding resistance to newly repurposed and introduced antibiotics linezolid, pretomanid, delamanid, bedaquiline, and clofazimine, because the majority of our dataset was sequenced before wide adoption of these drugs.

Figure 1

Download asset Open asset

Proteins with the highest frequency of mutation events are associated with antibiotic resistance.

(A) Workflow used to create our combined dataset of missense substitutions and in-frame indel mutations mapped to protein 3D structures, for 92% of the *M. tuberculosis* H37Rv proteome. Using a dataset of homoplastic mutations from 31,428 *Mycobacterium tuberculosis* complex (MTBC) isolates (Green et al., 2023; Vargas et al., 2023), we mapped mutations to protein sequences and 3D structures based on a combination of experimentally determined (RCSB PDB; Berman et al., 2000) and computationally predicted (AlphaFold; Varadi et al., 2022) structures. (B) The total number of mutation events in our dataset per protein, versus the percent of the amino acids in the protein’s structure that have been mutated at least once. Marginal histograms are displayed along both axes. Proteins with the highest frequency of mutations are those associated with resistance to antibiotics, according to the WHO catalog of resistance-associated mutations (Walker et al., 2022).

Structures of the whole M. tuberculosis proteome

We identified an experimentally measured or predicted structure for each protein in the M. tuberculosis proteome. We used a sensitive pipeline to search the RCSB Protein Data Bank (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 = 1350). 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) (Modlin et al., 2021b; Marin et al., 2022), or with fewer than 30 residues represented in the filtered protein structure, for a final dataset of 3687 proteins out of 3996 annotated ORFs in the M. tuberculosis H37Rv reference proteome (Supplementary file 3).

Antibiotic resistance proteins demonstrate mutational clustering

To calculate clustering of mutations, we use an approach from geographic analysis that calculates spatial autocorrelation between statistics, called the Getis–Ord G-statistic (Getis and Ord, 1992; Ord and Getis, 1995). 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 (Liang et al., 2020).

We examine the raw Getis–Ord score, which can be interpreted as a z-score, on the structures of nine proteins in which mutations are known to associate with antibiotic resistance (Walker et al., 2022) – 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, for example, RpoB, EmbB, InhA, and 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, for example, 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.

Figure 2

Download asset Open asset

G-statistic reveals clustering of mutations in antibiotic-resistance-conferring proteins.

(A) Workflow to compute residue-wise Getis–Ord statistic for proteins in *M. tuberculosis*. (B) Results for proteins: KatG is shown in complex with heme (orange), PDB ID = 4C51 chain A (Zhao et al., 2013). RpoB is shown in complex with rifampin (orange), PDB ID = 5UH6 chain C (Lin et al., 2017) aligned as described in Methods. PncA is shown in complex with Fe²⁺, PDB ID = 3PL1 chain A (Petrella et al., 2011). RsmG (encoded by the *gidB* gene) structure from *Thermus thermophilus* is shown in complex with ligand adenosine monophosphate (please note the streptomycin-binding site is unknown in *M. tuberculosis* and thus is not shown here), PDB ID = 3G8A chain A (Gregory et al., 2009).

A protein-level statistic to test for mutational clustering

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 randomly accumulating substitutions would be evenly distributed in the 3D structure of the protein. Uneven distribution of substitutions in 3D space may indicate selection; purifying selection leads to a depletion of substitutions in the hydrophobic cores of proteins, which must maintain stability (Echave et al., 2016), whereas positive selection may lead to mutations that cluster in protein–protein interaction interfaces or ligand binding pockets.

We evaluated five candidate protein-level statistics to quantify clustering of mutations possibly indicative of positive selection (Methods, Supplementary file 4). We selected as a positive control set nine proteins known to be under selection for antibiotic resistance– RpoB, EmbB, KatG, RpoC, PncA, GyrA, RsmG (GidB), EthA, and RS12. Because each of these proteins is 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).

Figure 3

Download asset Open asset

Benchmarking the ability of G-scores to find significant clustering in protein structures.

The procedure for generating downsampled true positive and true negative examples from real proteins. We then test five scores for their ability to distinguish true positives and negatives.

Significant mutational clustering across the MTBC proteome

We test all 3592 proteins in the merged M. tuberculosis structure and homoplasy dataset, and identify 499 proteins with significant clustering (Figure 4, Supplementary file 5).

Figure 4

Download asset Open asset

Hits of proteome-wide screen for clustering of mutations.

(A) Pipeline for detecting hits in 3687 proteins. (B) GO terms (abbreviated for space) with top fold enrichment (all significant at FDR <0.05). (C) Examples of proteins with significant clustering. All structures shown are from AlphaFold with low-confidence residues filtered out (note that relative domain orientation for PknB and PknH is low confidence).

We assay whether all proteins known to cause antibiotic resistance demonstrate mutational clustering, using the World Health Organization (WHO) catalog of resistance-conferring mutations (Walker et al., 2022, Kulkarni et al., 2024). The majority of proteins with Tier-1 resistance-conferring mutations (8 of 15) demonstrate significant clustering. Proteins without clustering include TlyA and RplC, in which there is only one missense substitution known to confer resistance, MmpR5, a transcriptional regulator protein involved in resistance to the newly administered drug bedaquiline, and RS12 (RS12_MYCTU, encoded by the gene rpsL), whose G-score is driven by just two highly mutated residue positions, K43 and K88 (Appendix 1—figure 2). In a small protein such as RS12 (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 Å in 3D space, confirming that the identified clusters are in a single location within the protein. Decreasing the distance threshold to 8 and 5 Å results in 74.7% (373) and 61.1% (305) of hits whose top pairs are close in 3D, respectively (Appendix 1—figure 3). We find that the distance between top pairs of residues does not increase with protein length, a proxy for size (Appendix 1—figure 4). Proteins with a high number of mutations that are not spatially clustered are not significant hits. This includes Cas10, which has 1534 mutation events 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), 37 GO categories are significantly enriched (FDR <0.05) (Figure 4). The most enriched categories by fold change (Appendix 1—table 1; Supplementary file 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 support that PknA, PknB, and PknH have a role in regulating cellular permeability and thus intrinsic drug resistance, potentially through post-translational modification of proteins involved in cell wall synthesis (Sun et al., 2022; Zeng et al., 2020; Sharma et al., 2006). For all of the Pkn proteins found by our clustering score, the region of significant clustering is not in 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 domain (Cavazos et al., 2012; 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 (Farhat et al., 2013). Amino sugars are a key component of the cell wall, and thus these proteins could be under positive selection for phenotypes including intrinsic drug resistance and adaptation to the host environment.

We find significant clustering in RNase J, ribonuclease J, which has previously been identified in GWAS for antibiotic resistance (Farhat et al., 2019), whose loss is known to lead to multi-drug tolerance (Martini et al., 2022), and 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; Bao et al., 2023).

Structure supports the prediction of resistance-conferring variants

Given that most drug resistance genes demonstrate 3D 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 catalog of resistance-associated mutations from M. tuberculosis (Kulkarni et al., 2024) consisting of 583 unique missense protein-coding variants annotated as associated with resistance (R-assoc) and 58 annotated as not associated, with isolates carrying only these mutations being antibiotic susceptible (S-assoc) (Methods, Supplementary file 7; Walker et al., 2022).

We ask whether protein structure information is useful for 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 catalog. 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 96.5%, Table 1, Methods). Prediction from 1D proximity alone is less accurate at an F1 of 89.1%, 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.

Table 1

Performance of classification models on predicting whether mutations are R-conferring from the mutation catalog.

The 1D-proximity model was trained using just the distance in primary sequence to the nearest known R mutation, the 3D-proximity model was trained using the distance in 3D to the nearest known R mutation, and G-score was trained using the G-score calculated in this manuscript. Reported values are calculated on the held-out test set.

Feature set	F1	Precision/PPV	Sensitivity
3D-proximity	96.5	98.2	94.9
1D-proximity	89.1	98.3	81.3
G-score	80.8	99.2	68.2

While the G-score prediction has the lowest F1 score and sensitivity, it has high precision (99.2%). 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.

Discussion

In this paper, we demonstrate that mutations cluster in three dimensions in the structure of proteins known to confer antibiotic resistance, through a novel application of the Getis–Ord statistic to evolutionary sequence data from clinical TB isolates. We apply our method to all proteins in the M. tuberculosis H37Rv proteome, using homoplasy data generated from a dataset of over 31,000 M. tuberculosis complex isolates. We find significant clustering of mutations in over 450 proteins, including eight known resistance-conferring proteins, and in pathways thought to be important for pathogenesis and antibiotic resistance.

We observe spatial clustering of mutations in known resistance-conferring proteins. For proteins like RpoB, where mutations are known to cluster in the rifampicin binding site, this serves as an internal control for our method. However, for proteins where mutations that lead to loss of function are known to cause resistance, such as PncA and RsmG (GidB), it is not necessarily expected to find clustering of mutations. We suspect that the observed clustering is due to mutations in a certain region of the protein being more likely to cause loss of function.

When running our analysis on the entire proteome, we find a significant enrichment for hits in proteins involved in regulation of fatty acid biosynthetic processes, and in amino sugar metabolism. Both of these pathways are related to cell envelope synthesis, which may be a response to antibiotic pressure or an adaptation to host environments. We also find hits to the proteins MycP1, MycP2, and MycP3, which are secreted from the cell and thought to be involved in host cell entry. We chose to apply our approach to homoplasy data, not allele frequency data, to ensure that we were not biased to overweight more ancient mutations in our analysis. Thus, we believe significant Getis–Ord score clustering indicates protein hotspots of positive and/or diversifying selection, where M. tuberculosis is adapting to antibiotic pressure and host environments.

Our variant effect prediction results support that structural information is a meaningful feature for predicting whether specific genetic variants confer antibiotic resistance. Distance to ligand binding sites has previously been shown to be a meaningful feature for classification of resistance in the proteins RpoB (Lynch et al., 2025; Portelli et al., 2020), PncA (Karmakar et al., 2020), and AtpE (Karmakar et al., 2019), and our work extends this to all proteins with known resistance-conferring mutations. These results on resistance variant classification show potential methods for expanding the catalog of known resistance-conferring mutations. In addition, recent efforts to predict M. tuberculosis antibiotic resistance from sequence data have suggested that our ability to predict antibiotic resistance from sequence alone may be reaching a plateau, and that more data and different data modalities, not more innovative model architectures, are needed in order to further improve predictions (Wang et al., 2024). While we do not suggest that our current model be used in clinical settings, we envision structure being added as an additional feature in future work to predict organismal resistance phenotypes from sequences, as it has already been successfully used in resistance variant classification (Pal et al., 2025; Wood et al., 2025).

One drawback of our approach is that our random shuffling procedure implicitly assumes that all sites in a protein are equally likely to mutate. This may pose a problem for proteins with conserved hydrophobic cores, or multidomain proteins where one domain is more conserved than the other, which could manifest in apparent higher or lower rates of mutation in one region of the protein. We account for this bias when we compute the distance between the top-scoring pair of residues in each protein, finding that for over 90% of the proteomic hits, those two residues are within 15 Ångstroms of one another, indicating that the patch of high-scoring residues is in a single location on the protein, not spread throughout a surface or domain. However, we cannot entirely rule out the effects of differing mutation rates throughout a protein. Future work could consider simulating the accumulation of substitutions according to their predicted effects on protein stability, to better reflect a real-world evolutionary scenario.

The research community has long sought to determine the genetic basis of antibiotic resistance in M. tuberculosis, using both supervised methods that rely on labeled data, and unsupervised methods that seek to find patterns of positive selection indicative of antibiotic resistance. The availability of predicted protein structures has opened a new avenue for understanding the effects of mutations in microbial genomes, and in M. tuberculosis specifically. We hypothesize that protein structural proximity is a useful feature because it captures how likely any given pair of mutations are to have a similar phenotypic effect – mutations in the same functional regions of proteins are more likely to cause the same effect. We show that protein 3D structure can be used to classify resistance-conferring variants across all proteins in a supervised framework, and used in an unsupervised fashion to discover targets of positive selection in the M. tuberculosis genome.

Methods

Key resources table

Reagent type (species) or resource	Designation	Source or reference	Identifiers	Additional information
Other	M. tuberculosis H37Rv	UniProt	UP000001584	Reference proteome
Other	M. tuberculosis H37Rv	Cole et al., 1998.	H37Rv	Reference genome

M. tuberculosis mutation dataset

We use a dataset of SNPs and indels found in M. tuberculosis isolates, mapped to their occurrences on a phylogenetic tree, published by previous work (Vargas et al., 2023; Vargas et al., 2021). This dataset was constructed using a previously validated pipeline for calling variants in M. tuberculosis, by mapping reads to the H37Rv reference genome using BWA-MEM v0.7.17 (Li, 2013; Li and Durbin, 2009) after trimming and filtering with PRINSEQ v0.20.4 (Schmieder and Edwards, 2011), contaminant removal with Kraken v0.10.6 (Wood and Salzberg, 2014), and duplicate read removal with Picard v2.9.2. All isolates were required to have at least 95% of bases in the reference genome with at least 10x coverage. Variant calling was performed with Pilon (Walker et al., 2014), and additional quality control filters were applied to ensure accurate allele calls. The resulting dataset was two matrices: a list of all positions found to have a SNP in any isolate, relative to the H37Rv reference (N = 782,565), and a list of all positions found to have an insertion or deletion in any isolate, relative to the H37Rv reference (N = 47,425) (Vargas et al., 2023).

Phylogeny and ancestral sequence reconstruction were performed by Vargas et al. as described in a previous publication (Vargas et al., 2023). In their procedure, phylogenies were constructed separately for sub-lineages (L1, L2, L3, L4A, L4B, L4C, L5, and L6) for computational feasibility. Phylogenies were constructed from concatenated SNP data using IQ-TREE (Nguyen et al., 2015). SNPPar (Edwards et al., 2021) was used to reconstruct SNPs, and the method previously published by Vargas et al. was used to reconstruct indels (Vargas et al., 2023).

Searching PDB structure database

We search the RCSB Protein Data Bank structure database (download date: February 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 (Hopf et al., 2019), which operates in two stages: first, we construct sequence alignments against the UniProt sequence database (download date February 5, 2021) at bitscores of 0.1, 0.2, and 0.3 times the query sequence length, using jackhmmer with five iterations (Eddy, 2023). Second, we use hmmbuild and hmmsearch to search the RCSB PDB sequence database using the constructed alignments.

Selecting structure hits

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 3996 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 lowest e-value.

Using AlphaFold database

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

Preparing protein structure data for analysis

We exclude from consideration proteins with documented difficulties in mutation calling using short-read sequencing. Using data from Marin et al., 2022, 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.

Preparing mutations for analysis

For SNP mutations, we consider all missense SNPs that occur at least once in our M. tuberculosis dataset, for a total of 477,607 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 missense mutations. We observe 5510 unique in-frame indel positions. Combining the SNP and in-frame 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% (Marin et al., 2022; Modlin et al., 2021b), as well as all mutations occurring in proteins with <90% mean empirical base-pair recall or <90% mean mappability (Marin et al., 2022). After filtering, we have 475,046 unique polymorphisms for downstream analysis (469,942 total unique missense mutations and 5104 indels) (Supplementary file 2).

Merging mutation and structure data

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 3592 proteins with mutations mapped and acceptable structures.

Computing inter-residue distances

The EVcouplings Python package was used to compute the distance between wild-type amino acid residues in all protein structures (Hopf et al., 2019). The package calculates the distance between all heavy (non-hydrogen) atoms in residue i and residue j, then returns the minimum of those distances.

Computing the Getis–Ord score for clustering of homoplastic mutations

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 (Getis and Ord, 1992; Ord and Getis, 1995; Liang et al., 2020). 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 computed from the wild-type amino acids. x is an L × 1 vector where the entry $x_{i}$ contains the homoplasy score for the ith residue in the protein, and W is an L × L matrix with entries defined as:

w_{i, j} = {\begin{matrix} \frac{1}{d_{i, j}} if i \neq j \\ 0 if i = j \end{matrix}

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:

G_{i} = \frac{\sum_{j = 1}^{L} w_{i, j} x_{j} - \bar{X} \sum_{j = 1}^{L} w_{i, j}}{S K}

where $\bar{X}$ is the mean of x, and:

S = \sqrt{\frac{\sum_{j = 1}^{L} x_{j}^{2}}{L} - {\bar{X}}^{2}}

K = \sqrt{\frac{L \sum_{j = 1}^{L} w_{i, j}^{2} - {(\sum_{j = 1}^{L} w_{i, j})}^{2}}{L - 1}}

Preparing GeO score calibration data

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 mutated site). For this, we used the nine proteins known to be involved in antibiotic resistance with demonstrated clustering of mutations (RpoB, EmbB, KatG, RpoC, 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. Note that we do not re-compute inter-residue distances when simulating mutations in an amino acid, as the distances used as input to the GeO score are the wild-type inter-residue distances.

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 from the one expected by chance. We compare five 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 median p-value. The other two features are based on the raw G-score for the real distribution versus 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.

Running on whole proteome

We then run the G-score calculation on the whole proteome using the above pipeline. We apply our 95% precision threshold to find 499 initial hits.

GO enrichment analysis

We test whether our gene hits are enriched in particular GO functional categories (Ashburner et al., 2000; Aleksander et al., 2023). 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 reference 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.

AlphaFill

We downloaded hits from the AlphaFill v1 database (access date: October 7, 2024) to find potential ligand-binding locations in our proteins (Hekkelman et al., 2023). For RNase J, two ligands are in proximity to the high GeO score regions: ‘U5P’ and ‘C5P’, uridine-5′-monophosphate and cytidine-5′-monophosphate.

Parsing the WHO mutation catalog

The World Health Organization (WHO) Mutation Catalog 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 antibiotics (World Health Organization, 2021a). 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 labeled variants using regression-based grading based on the frequency of mutations in resistant and susceptible isolates (Kulkarni et al., 2024).

We limit our analysis to the variants observed in the following 15 protein-coding genes: Rv0678, atpE, ddn, embB, ethA, gid, gyrA, gyrB, inhA, katG, pncA, rplC, rpoB, rpsL, and tlyA. After removing duplicates – because multiple genomic variants can cause the same missense mutation – 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 S-assoc (category 4 or 5).

Fitting a classifier on the WHO mutation catalog

For each of the 9365 unique missense variants in the catalog (Kulkarni et al., 2024), 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 (Hopf et al., 2019).

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 (Pedregosa et al., 2011). 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 (Ruopp et al., 2008).

Appendix 1

Appendix 1—figure 1

Download asset Open asset

Homoplastic inframe insertion at position 3131469 in the Cas10 gene (Rv2823c).

Screenshot from Mycobrowser (https://mycobrowser.epfl.ch/) of relevant genomic region beginning at 3131469. Table showing the number of mutation events per lineage and total in the dataset, as well as total number of isolates with the alternate allele. Note that the inserted sequence is similar to but not exactly the same as the H37Rv reference sequence at that location, and that similar motifs recur throughout the sequence region.

Appendix 1—figure 2

Download asset Open asset

Mutations and GeO clustering for RS12 (RS12_MYCTU).

Two residues, shown in orange (K43 and K88) are highly mutated in the protein RpsL, leading to significant G-score clustering in that region of the protein.

Appendix 1—figure 3

Download asset Open asset

Distance between top pairs of high G-score residues in proteins with significant clustering.

Of the 499 proteins with significant hits, we analyze what number are still significant after filtering to ensure that the minimum inter-atom distance between the top 2 residues with high G-score is less than a defined distance threshold. 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. Decreasing the distance threshold to 8 and 5 Ångstroms results in 74.7% (373) and 61.1% (305) hits whose top pairs are close in 3-D, respectively.

Appendix 1—figure 4

Download asset Open asset

Relationship between protein length and distance between top pair of residues.

Of the 499 proteins with significant hits, we analyze whether there is a statistically significant relationship between the length of the protein (filtered for residues that pass our structure quality thresholds) and the distance between the two residues with highest G-score. Using Ordinary Least Squares regression implemented with default parameters in statsmodels v0.14.4, we find a weak negative relationship: R2 = 0.008, beta = -3.1806, p-value = 0.048, indicating that distance between top pairs does not increase with protein length, hence our method is likely capturing real signal for 3-D clustering.

Appendix 1—table 1

Top 10 GO categories significantly enriched in the clustered protein set.

Categories with identical members and FDR (e.g., GO:0071103 DNA conformation change and GO:0006265 DNA topological change) have only one representative category shown. See Supplementary file 6 for complete table.

GO ID	GO term label	UniProt identifiers	Fold enrich.	FDR	Category
GO:0035635	Entry of bacterium into host cell	Q6MX51_MYCTU, GLMU_MYCTU, SAHH_MYCTU	11.04	0.01	Host cell entry
GO:0006265	DNA topological change	GYRA_MYCTU, GYRB_MYCTU, TOP1_MYCTU	11.04	0.01	DNA topology
GO:0016539	Intein-mediated protein splicing	DNAB_MYCTU, RECA_MYCTU, Y1461_MYCTU	11.04	0.01	Protein maturation
GO:0046349	Amino sugar biosynthetic process	GLMM_MYCTU, MURA_MYCTU, GLMU_MYCTU	11.04	0.01	Amino sugar
GO:0006047	UDP-N-acetylglucosamine metabolic process	GLMM_MYCTU, GLMS_MYCTU, GLMU_MYCTU	11.04	0.01	Amino sugar
GO:0062014	Negative regulation of small molecule metabolic process	PKNB_MYCTU, GARA_MYCTU, PKNA_MYCTU, PKNE_MYCTU, PKND_MYCTU	9.2	<0.01	Fatty acid
GO:0042304	Regulation of fatty acid biosynthetic process	PKNB_MYCTU, PKNA_MYCTU, PKNE_MYCTU, PKND_MYCTU	8.83	0.01	Fatty acid
GO:0006040	Amino sugar metabolic process	GLMM_MYCTU, GLMS_MYCTU, MURA_MYCTU, GLMU_MYCTU	8.83	0.01	Amino sugar
GO:0015990	Electron transport coupled proton transport	NUOM_MYCTU, COX1_MYCTU, NUOL_MYCTU	8.28	0.04	ETC
GO:0046890	Regulation of lipid biosynthetic process	PKNB_MYCTU, PKNA_MYCTU, PKNE_MYCTU, P71814_MYCTU, PKND_MYCTU	7.88	<0.01	Fatty acid

Data availability

Code is available on GitHub at https://github.com/aggreen/MTB_Mut_Clust, copy archived at Tasmin and Green, 2026; and the modified EVcouplings pipeline for structure search is found at https://github.com/aggreen/EVcouplings/tree/feature/structure_finder, copy archived at Green, 2026. All strains used in our analyses are publicly available, and the raw read data are available for download from the NCBI using accession codes found in the isolate annotation table. Precalculated inter-residue distances for all proteins are available at: https://doi.org/10.5281/zenodo.20766452. Data newly generated by this paper, including tables of all mutations and all clustering scores, are found in the Supplementary files. WHO mutation classification data 2nd edition is publicly available here: https://www.who.int/publications/i/item/9789240082410.

The following data sets were generated

1. Green AG
(2026) Zenodo
Supporting protein structure data for "The structural context of mutations in proteins predicts their effect on antibiotic resistance".

https://doi.org/10.5281/zenodo.20766453

References

1. Aleksander SA
2. Balhoff J
3. Carbon S
4. Cherry JM
5. Drabkin HJ
6. Ebert D
7. Feuermann M
8. Gaudet P
9. Harris NL
10. Hill DP
11. Lee R
12. Mi H
13. Moxon S
14. Mungall CJ
15. Muruganugan A
16. Mushayahama T
17. Sternberg PW
18. Thomas PD
19. Van Auken K
20. Ramsey J
21. Siegele DA
22. Chisholm RL
23. Fey P
24. Aspromonte MC
25. Nugnes MV
26. Quaglia F
27. Tosatto S
28. Giglio M
29. Nadendla S
30. Antonazzo G
31. Attrill H
32. Dos Santos G
33. Marygold S
34. Strelets V
35. Tabone CJ
36. Thurmond J
37. Zhou P
38. Ahmed SH
39. Asanitthong P
40. Luna Buitrago D
41. Erdol MN
42. Gage MC
43. Ali Kadhum M
44. Li KYC
45. Long M
46. Michalak A
47. Pesala A
48. Pritazahra A
49. Saverimuttu SCC
50. Su R
51. Thurlow KE
52. Lovering RC
53. Logie C
54. Oliferenko S
55. Blake J
56. Christie K
57. Corbani L
58. Dolan ME
59. Drabkin HJ
60. Hill DP
61. Ni L
62. Sitnikov D
63. Smith C
64. Cuzick A
65. Seager J
66. Cooper L
67. Elser J
68. Jaiswal P
69. Gupta P
70. Jaiswal P
71. Naithani S
72. Lera-Ramirez M
73. Rutherford K
74. Wood V
75. De Pons JL
76. Dwinell MR
77. Hayman GT
78. Kaldunski ML
79. Kwitek AE
80. Laulederkind SJF
81. Tutaj MA
82. Vedi M
83. Wang S-J
84. D’Eustachio P
85. Aimo L
86. Axelsen K
87. Bridge A
88. Hyka-Nouspikel N
89. Morgat A
90. Aleksander SA
91. Cherry JM
92. Engel SR
93. Karra K
94. Miyasato SR
95. Nash RS
96. Skrzypek MS
97. Weng S
98. Wong ED
99. Bakker E
100. Berardini TZ
101. Reiser L
102. Auchincloss A
103. Axelsen K
104. Argoud-Puy G
105. Blatter M-C
106. Boutet E
107. Breuza L
108. Bridge A
109. Casals-Casas C
110. Coudert E
111. Estreicher A
112. Livia Famiglietti M
113. Feuermann M
114. Gos A
115. Gruaz-Gumowski N
116. Hulo C
117. Hyka-Nouspikel N
118. Jungo F
119. Le Mercier P
120. Lieberherr D
121. Masson P
122. Morgat A
123. Pedruzzi I
124. Pourcel L
125. Poux S
126. Rivoire C
127. Sundaram S
128. Bateman A
129. Bowler-Barnett E
130. Bye-A-Jee H
131. Denny P
132. Ignatchenko A
133. Ishtiaq R
134. Lock A
135. Lussi Y
136. Magrane M
137. Martin MJ
138. Orchard S
139. Raposo P
140. Speretta E
141. Tyagi N
142. Warner K
143. Zaru R
144. Diehl AD
145. Lee R
146. Chan J
147. Diamantakis S
148. Raciti D
149. Zarowiecki M
150. Fisher M
151. James-Zorn C
152. Ponferrada V
153. Zorn A
154. Ramachandran S
155. Ruzicka L
156. Westerfield M
157. Gene Ontology Consortium
(2023) The gene ontology knowledgebase in 2023
GENETICS 224:iyad031.

https://doi.org/10.1093/genetics/iyad031
- PubMed
- Google Scholar
1. Ashburner M
2. Ball CA
3. Blake JA
4. Botstein D
5. Butler H
6. Cherry JM
7. Davis AP
8. Dolinski K
9. Dwight SS
10. Eppig JT
11. Harris MA
12. Hill DP
13. Issel-Tarver L
14. Kasarskis A
15. Lewis S
16. Matese JC
17. Richardson JE
18. Ringwald M
19. Rubin GM
20. Sherlock G
(2000) Gene Ontology: tool for the unification of biology
Nature Genetics 25:25–29.

https://doi.org/10.1038/75556
- Google Scholar
1. Bao L
2. Hu J
3. Zhan B
4. Chi M
5. Li Z
6. Wang S
7. Shan C
8. Zhao Z
9. Guo Y
10. Ding X
11. Ji C
12. Tao S
13. Ni T
14. Zhang X
15. Zhao G
16. Li J
(2023) Structural insights into RNase J that plays an essential role in Mycobacterium tuberculosis RNA metabolism
Nature Communications 14:2280.

https://doi.org/10.1038/s41467-023-38045-z
- PubMed
- Google Scholar
1. Berman HM
2. Westbrook J
3. Feng Z
4. Gilliland G
5. Bhat TN
6. Weissig H
7. Shindyalov IN
8. Bourne PE
(2000) The protein data bank
Nucleic Acids Research 28:235–242.

https://doi.org/10.1093/nar/28.1.235
- PubMed
- Google Scholar
1. Bush WS
2. Moore JH
(2012) Chapter 11: Genome-wide association studies
PLOS Computational Biology 8:e1002822.

https://doi.org/10.1371/journal.pcbi.1002822
- PubMed
- Google Scholar
1. Carter JJ
2. Walker TM
3. Walker AS
4. Whitfield MG
5. Morlock GP
6. Lynch CI
7. Adlard D
8. Peto TEA
9. Posey JE
10. Crook DW
11. Fowler PW
(2024) Prediction of pyrazinamide resistance in Mycobacterium tuberculosis using structure-based machine-learning approaches
JAC-Antimicrobial Resistance 6:dlae037.

https://doi.org/10.1093/jacamr/dlae037
- PubMed
- Google Scholar
(2012) Structure of the sensor domain of Mycobacterium tuberculosis PknH receptor kinase reveals a conserved binding cleft
Journal of Molecular Biology 422:488–494.

https://doi.org/10.1016/j.jmb.2012.06.011
- PubMed
- Google Scholar
1. Chen ML
2. Doddi A
3. Royer J
4. Freschi L
5. Schito M
6. Ezewudo M
7. Kohane IS
8. Beam A
9. Farhat M
(2019) Beyond multidrug resistance: Leveraging rare variants with machine and statistical learning models in Mycobacterium tuberculosis resistance prediction
EBioMedicine 43:356–369.

https://doi.org/10.1016/j.ebiom.2019.04.016
- PubMed
- Google Scholar
1. Cole ST
2. Brosch R
3. Parkhill J
4. Garnier T
5. Churcher C
6. Harris D
7. Gordon SV
8. Eiglmeier K
9. Gas S
10. Barry CE
11. Tekaia F
12. Badcock K
13. Basham D
14. Brown D
15. Chillingworth T
16. Connor R
17. Davies R
18. Devlin K
19. Feltwell T
20. Gentles S
21. Hamlin N
22. Holroyd S
23. Hornsby T
24. Jagels K
25. Krogh A
26. McLean J
27. Moule S
28. Murphy L
29. Oliver K
30. Osborne J
31. Quail MA
32. Rajandream MA
33. Rogers J
34. Rutter S
35. Seeger K
36. Skelton J
37. Squares R
38. Squares S
39. Sulston JE
40. Taylor K
41. Whitehead S
42. Barrell BG
(1998) Deciphering the biology of Mycobacterium tuberculosis from the complete genome sequence
Nature 393:537–544.

https://doi.org/10.1038/31159
- PubMed
- Google Scholar
1. Coll F
2. Phelan J
3. Hill-Cawthorne GA
4. Nair MB
5. Mallard K
6. Ali S
7. Abdallah AM
8. Alghamdi S
9. Alsomali M
10. Ahmed AO
11. Portelli S
12. Oppong Y
13. Alves A
14. Bessa TB
15. Campino S
16. Caws M
17. Chatterjee A
18. Crampin AC
19. Dheda K
20. Furnham N
21. Glynn JR
22. Grandjean L
23. Minh Ha D
24. Hasan R
25. Hasan Z
26. Hibberd ML
27. Joloba M
28. Jones-López EC
29. Matsumoto T
30. Miranda A
31. Moore DJ
32. Mocillo N
33. Panaiotov S
34. Parkhill J
35. Penha C
36. Perdigão J
37. Portugal I
38. Rchiad Z
39. Robledo J
40. Sheen P
41. Shesha NT
42. Sirgel FA
43. Sola C
44. Oliveira Sousa E
45. Streicher EM
46. Helden PV
47. Viveiros M
48. Warren RM
49. McNerney R
50. Pain A
51. Clark TG
(2018) Genome-wide analysis of multi- and extensively drug-resistant Mycobacterium tuberculosis
Nature Genetics 50:307–316.

https://doi.org/10.1038/s41588-017-0029-0
- PubMed
- Google Scholar
1. Comas I
2. Borrell S
3. Roetzer A
4. Rose G
5. Malla B
6. Kato-Maeda M
7. Galagan J
8. Niemann S
9. Gagneux S
(2011) Whole-genome sequencing of rifampicin-resistant Mycobacterium tuberculosis strains identifies compensatory mutations in RNA polymerase genes
Nature Genetics 44:106–110.

https://doi.org/10.1038/ng.1038
- PubMed
- Google Scholar
(2022) Distribution of common and rare genetic markers of second-line-injectable-drug resistance in Mycobacterium tuberculosis revealed by a genome-wide association study
Antimicrobial Agents and Chemotherapy 66:e02075.

https://doi.org/10.1128/aac.02075-21
- Google Scholar
1. Desjardins CA
2. Cohen KA
3. Munsamy V
4. Abeel T
5. Maharaj K
6. Walker BJ
7. Shea TP
8. Almeida DV
9. Manson AL
10. Salazar A
11. Padayatchi N
12. O’Donnell MR
13. Mlisana KP
14. Wortman J
15. Birren BW
16. Grosset J
17. Earl AM
18. Pym AS
(2016) Genomic and functional analyses of Mycobacterium tuberculosis strains implicate ald in D-cycloserine resistance
Nature Genetics 48:544–551.

https://doi.org/10.1038/ng.3548
- PubMed
- Google Scholar
Preprint
(2025) Predicting pyrazinamide resistance in Mycobacterium tuberculosis using a graph convolutional network
bioRxiv.

https://doi.org/10.1101/2025.10.28.685176
- Google Scholar
(2016) Causes of evolutionary rate variation among protein sites
Nature Reviews. Genetics 17:109–121.

https://doi.org/10.1038/nrg.2015.18
- PubMed
- Google Scholar
Software
1. Eddy S
(2023) HMMER user’s guide: biological sequence analysis using profile hidden markov models, version v3.4
HMMER.

http://hmmer.org
1. Edwards DJ
2. Duchene S
3. Pope B
4. Holt KE
(2021) SNPPar: identifying convergent evolution and other homoplasies from microbial whole-genome alignments
Microbial Genomics 7:000694.

https://doi.org/10.1099/mgen.0.000694
- PubMed
- Google Scholar
1. Farhat MR
2. Shapiro BJ
3. Kieser KJ
4. Sultana R
5. Jacobson KR
6. Victor TC
7. Warren RM
8. Streicher EM
9. Calver A
10. Sloutsky A
11. Kaur D
12. Posey JE
13. Plikaytis B
14. Oggioni MR
15. Gardy JL
16. Johnston JC
17. Rodrigues M
18. Tang PKC
19. Kato-Maeda M
20. Borowsky ML
21. Muddukrishna B
22. Kreiswirth BN
23. Kurepina N
24. Galagan J
25. Gagneux S
26. Birren B
27. Rubin EJ
28. Lander ES
29. Sabeti PC
30. Murray M
(2013) Genomic analysis identifies targets of convergent positive selection in drug-resistant Mycobacterium tuberculosis
Nature Genetics 45:1183–1189.

https://doi.org/10.1038/ng.2747
- PubMed
- Google Scholar
1. Farhat MR
2. Sultana R
3. Iartchouk O
4. Bozeman S
5. Galagan J
6. Sisk P
7. Stolte C
8. Nebenzahl-Guimaraes H
9. Jacobson K
10. Sloutsky A
11. Kaur D
12. Posey J
13. Kreiswirth BN
14. Kurepina N
15. Rigouts L
16. Streicher EM
17. Victor TC
18. Warren RM
19. van Soolingen D
20. Murray M
(2016) Genetic determinants of drug resistance in Mycobacterium tuberculosis and their diagnostic value
American Journal of Respiratory and Critical Care Medicine 194:621–630.

https://doi.org/10.1164/rccm.201510-2091OC
- PubMed
- Google Scholar
1. Farhat MR
2. Freschi L
3. Calderon R
4. Ioerger T
5. Snyder M
6. Meehan CJ
7. de Jong B
8. Rigouts L
9. Sloutsky A
10. Kaur D
11. Sunyaev S
12. van Soolingen D
13. Shendure J
14. Sacchettini J
15. Murray M
(2019) GWAS for quantitative resistance phenotypes in Mycobacterium tuberculosis reveals resistance genes and regulatory regions
Nature Communications 10:2128.

https://doi.org/10.1038/s41467-019-10110-6
- PubMed
- Google Scholar
1. Gao J
2. Chang MT
3. Johnsen HC
4. Gao SP
5. Sylvester BE
6. Sumer SO
7. Zhang H
8. Solit DB
9. Taylor BS
10. Schultz N
11. Sander C
(2017) 3D clusters of somatic mutations in cancer reveal numerous rare mutations as functional targets
Genome Medicine 9:4.

https://doi.org/10.1186/s13073-016-0393-x
- PubMed
- Google Scholar
1. Getis A
2. Ord JK
(1992) The analysis of spatial association by use of distance statistics
Geographical Analysis 24:189–206.

https://doi.org/10.1111/j.1538-4632.1992.tb00261.x
- Google Scholar
Website
1. Global Tuberculosis Programme
(2021) WHO consolidated guidelines on tuberculosis. Module 3: Diagnosis - rapid diagnostics for tuberculosis detection 2021 update
Accessed July 7, 2021.

https://www.who.int/publications/i/item/9789240029415
1. Green AG
2. Yoon CH
3. Chen ML
4. Ektefaie Y
5. Fina M
6. Freschi L
7. Gröschel MI
8. Kohane I
9. Beam A
10. Farhat M
(2022) A convolutional neural network highlights mutations relevant to antimicrobial resistance in Mycobacterium tuberculosis
Nature Communications 13:3817.

https://doi.org/10.1038/s41467-022-31236-0
- PubMed
- Google Scholar
1. Green AG
2. Vargas R
3. Marin MG
4. Freschi L
5. Xie J
6. Farhat MR
(2023) Analysis of genome-wide mutational dependence in naturally evolving Mycobacterium tuberculosis Populations
Molecular Biology and Evolution 40:msad131.

https://doi.org/10.1093/molbev/msad131
- PubMed
- Google Scholar
Software
1. Green AG
(2026) EVcouplings, version swh:1:rev:bdcf459ec56d958d143651ce03ba3a489c8c8fac
Software Heritage.

https://archive.softwareheritage.org/swh:1:dir:bbf051ef069c7e56666bd3c4d96f39fc9b44fabe;origin=https://github.com/aggreen/EVcouplings;visit=swh:1:snp:c7f35c464ff8711fc2d516add4579a1790cd22a2;anchor=swh:1:rev:bdcf459ec56d958d143651ce03ba3a489c8c8fac
(2009) Structural and functional studies of the Thermus thermophilus 16S rRNA methyltransferase RsmG
RNA 15:1693–1704.

https://doi.org/10.1261/rna.1652709
- PubMed
- Google Scholar
1. Gröschel MI
2. Owens M
3. Freschi L
4. Vargas R Jr
5. Marin MG
6. Phelan J
7. Iqbal Z
8. Dixit A
9. Farhat MR
(2021) GenTB: A user-friendly genome-based predictor for tuberculosis resistance powered by machine learning
Genome Medicine 13:138.

https://doi.org/10.1186/s13073-021-00953-4
- PubMed
- Google Scholar
(2023) AlphaFill: enriching AlphaFold models with ligands and cofactors
Nature Methods 20:205–213.

https://doi.org/10.1038/s41592-022-01685-y
- PubMed
- Google Scholar
1. Hopf TA
2. Green AG
3. Schubert B
4. Mersmann S
5. Schärfe CPI
6. Ingraham JB
7. Toth-Petroczy A
8. Brock K
9. Riesselman AJ
10. Palmedo P
11. Kang C
12. Sheridan R
13. Draizen EJ
14. Dallago C
15. Sander C
16. Marks DS
(2019) The EVcouplings Python framework for coevolutionary sequence analysis
Bioinformatics 35:1582–1584.

https://doi.org/10.1093/bioinformatics/bty862
- PubMed
- Google Scholar
1. Jumper J
2. Evans R
3. Pritzel A
4. Green T
5. Figurnov M
6. Ronneberger O
7. Tunyasuvunakool K
8. Bates R
9. Žídek A
10. Potapenko A
11. Bridgland A
12. Meyer C
13. Kohl SAA
14. Ballard AJ
15. Cowie A
16. Romera-Paredes B
17. Nikolov S
18. Jain R
19. Adler J
20. Back T
21. Petersen S
22. Reiman D
23. Clancy E
24. Zielinski M
25. Steinegger M
26. Pacholska M
27. Berghammer T
28. Bodenstein S
29. Silver D
30. Vinyals O
31. Senior AW
32. Kavukcuoglu K
33. Kohli P
34. Hassabis D
(2021) Highly accurate protein structure prediction with AlphaFold
Nature 596:583–589.

https://doi.org/10.1038/s41586-021-03819-2
- PubMed
- Google Scholar
1. Kamburov A
2. Lawrence MS
3. Polak P
4. Leshchiner I
5. Lage K
6. Golub TR
7. Lander ES
8. Getz G
(2015) Comprehensive assessment of cancer missense mutation clustering in protein structures
PNAS 112:E5486–E5495.

https://doi.org/10.1073/pnas.1516373112
- PubMed
- Google Scholar
(2019) Empirical ways to identify novel Bedaquiline resistance mutations in AtpE
PLOS ONE 14:e0217169.

https://doi.org/10.1371/journal.pone.0217169
- PubMed
- Google Scholar
(2020) Structure guided prediction of Pyrazinamide resistance mutations in pncA
Scientific Reports 10:1875.

https://doi.org/10.1038/s41598-020-58635-x
- PubMed
- Google Scholar
Preprint
(2024) Regression for accurate and sensitive grading of mutations diagnostic of antibiotic resistance in Mycobacterium tuberculosis
medRxiv.

https://doi.org/10.1101/2024.07.01.24309598
- Google Scholar
Preprint
1. Kulkarni SG
2. Green AG
3. Mann BC
4. Malatesta S
5. Goodwin SK
6. Cesare N
7. Mulaudzi S
8. Rawoot N
9. Warren R
10. Jacobson KR
11. Farhat MR
12. MIC-ML Consortium
(2025) Convolutional neural networks quantify antibiotic resistance in Mycobacterium tuberculosis with diagnostic grade accuracy and predict treatment response
medRxiv.

https://doi.org/10.1101/2025.08.05.25333066
- Google Scholar
1. Lange C
2. Chesov D
3. Heyckendorf J
4. Leung CC
5. Udwadia Z
6. Dheda K
(2018) Drug-resistant tuberculosis: An update on disease burden, diagnosis and treatment
Respirology 23:656–673.

https://doi.org/10.1111/resp.13304
- PubMed
- Google Scholar
1. Li H
2. Durbin R
(2009) Fast and accurate short read alignment with Burrows-Wheeler transform
Bioinformatics 25:1754–1760.

https://doi.org/10.1093/bioinformatics/btp324
- PubMed
- Google Scholar
Preprint
1. Li H
(2013) Aligning sequence reads, clone sequences and assembly contigs with BWA-MEM
arXiv.

https://doi.org/10.48550/arXiv.1303.3997
- Google Scholar
Preprint
1. Liang S
2. Mort M
3. Stenson PD
4. Cooper DN
5. Yu H
(2020) PIVOTAL: prioritizing variants of uncertain significance with spatial genomic patterns in the 3D proteome
bioRxiv.

https://doi.org/10.1101/2020.06.04.135103
- Google Scholar
1. Lin W
2. Mandal S
3. Degen D
4. Liu Y
5. Ebright YW
6. Li S
7. Feng Y
8. Zhang Y
9. Mandal S
10. Jiang Y
11. Liu S
12. Gigliotti M
13. Talaue M
14. Connell N
15. Das K
16. Arnold E
17. Ebright RH
(2017) Structural basis of Mycobacterium tuberculosis transcription and transcription inhibition
Molecular Cell 66:169–179.

https://doi.org/10.1016/j.molcel.2017.03.001
- PubMed
- Google Scholar
(2025) Predicting rifampicin resistance in Mycobacterium tuberculosis using machine learning informed by protein structural and chemical features
ERJ Open Research 11:00952–02024.

https://doi.org/10.1183/23120541.00952-2024
- PubMed
- Google Scholar
1. Marin M
2. Vargas R Jr
3. Harris M
4. Jeffrey B
5. Epperson LE
6. Durbin D
7. Strong M
8. Salfinger M
9. Iqbal Z
10. Akhundova I
11. Vashakidze S
12. Crudu V
13. Rosenthal A
14. Farhat MR
(2022) Benchmarking the empirical accuracy of short-read sequencing across the M. tuberculosis genome
Bioinformatics 38:1781–1787.

https://doi.org/10.1093/bioinformatics/btac023
- PubMed
- Google Scholar
1. Martini MC
2. Hicks ND
3. Xiao J
4. Alonso MN
5. Barbier T
6. Sixsmith J
7. Fortune SM
8. Shell SS
(2022) Loss of RNase J leads to multi-drug tolerance and accumulation of highly structured mRNA fragments in Mycobacterium tuberculosis
PLOS Pathogens 18:e1010705.

https://doi.org/10.1371/journal.ppat.1010705
- PubMed
- Google Scholar
1. Meyer MJ
2. Lapcevic R
3. Romero AE
4. Yoon M
5. Das J
6. Beltrán JF
7. Mort M
8. Stenson PD
9. Cooper DN
10. Paccanaro A
11. Yu H
(2016) mutation3D: Cancer gene prediction through atomic clustering of coding variants in the structural proteome
Human Mutation 37:447–456.

https://doi.org/10.1002/humu.22963
- PubMed
- Google Scholar
1. Miller ML
2. Reznik E
3. Gauthier NP
4. Aksoy BA
5. Korkut A
6. Gao J
7. Ciriello G
8. Schultz N
9. Sander C
(2015) Pan-cancer analysis of mutation hotspots in protein domains
Cell Systems 1:197–209.

https://doi.org/10.1016/j.cels.2015.08.014
- PubMed
- Google Scholar
1. Modlin SJ
2. Elghraoui A
3. Gunasekaran D
4. Zlotnicki AM
5. Dillon NA
6. Dhillon N
7. Kuo N
8. Robinhold C
9. Chan CK
10. Baughn AD
11. Valafar F
(2021a) Structure-Aware Mycobacterium tuberculosis functional annotation uncloaks resistance, metabolic, and virulence genes
mSystems 6:e00673.

https://doi.org/10.1128/mSystems.00673-21
- Google Scholar
(2021b) Exact mapping of Illumina blind spots in the Mycobacterium tuberculosis genome reveals platform-wide and workflow-specific biases
Microbial Genomics 7:mgen000465.

https://doi.org/10.1099/mgen.0.000465
- PubMed
- Google Scholar
(2015) IQ-TREE: a fast and effective stochastic algorithm for estimating maximum-likelihood phylogenies
Molecular Biology and Evolution 32:268–274.

https://doi.org/10.1093/molbev/msu300
- PubMed
- Google Scholar
1. Niu B
2. Scott AD
3. Sengupta S
4. Bailey MH
5. Batra P
6. Ning J
7. Wyczalkowski MA
8. Liang W-W
9. Zhang Q
10. McLellan MD
11. Sun SQ
12. Tripathi P
13. Lou C
14. Ye K
15. Mashl RJ
16. Wallis J
17. Wendl MC
18. Chen F
19. Ding L
(2016) Protein-structure-guided discovery of functional mutations across 19 cancer types
Nature Genetics 48:827–837.

https://doi.org/10.1038/ng.3586
- PubMed
- Google Scholar
1. Ord JK
2. Getis A
(1995) Local Spatial autocorrelation statistics: distributional issues and an application
Geographical Analysis 27:286–306.

https://doi.org/10.1111/j.1538-4632.1995.tb00912.x
- Google Scholar
1. Pal A
2. Mohanty D
(2024) Machine learning-based approach for identification of new resistance associated mutations from whole genome sequences of Mycobacterium tuberculosis
Bioinformatics Advances 5:vbaf050.

https://doi.org/10.1093/bioadv/vbaf050
- Google Scholar
1. Pal A
2. Pal S
3. Mahapatra S
4. Pandey A
5. Mohanty D
(2025) In silico analysis of the functional implications of drug resistance associated mutations in Mycobacterium tuberculosis
Computational and Structural Biotechnology Journal 27:5425–5440.

https://doi.org/10.1016/j.csbj.2025.11.054
- PubMed
- Google Scholar
(2011)
Scikit-learn: machine learning in python

Journal of Machine Learning Research: JMLR 12:2825–2830.
- Google Scholar
(2011) Crystal structure of the pyrazinamidase of Mycobacterium tuberculosis: insights into natural and acquired resistance to pyrazinamide
PLOS ONE 6:e15785.

https://doi.org/10.1371/journal.pone.0015785
- PubMed
- Google Scholar
1. Phelan J
2. Coll F
3. McNerney R
4. Ascher DB
5. Pires DEV
6. Furnham N
7. Coeck N
8. Hill-Cawthorne GA
9. Nair MB
10. Mallard K
11. Ramsay A
12. Campino S
13. Hibberd ML
14. Pain A
15. Rigouts L
16. Clark TG
(2016) Mycobacterium tuberculosis whole genome sequencing and protein structure modelling provides insights into anti-tuberculosis drug resistance
BMC Medicine 14:31.

https://doi.org/10.1186/s12916-016-0575-9
- PubMed
- Google Scholar
1. Portelli S
2. Myung Y
3. Furnham N
4. Vedithi SC
5. Pires DEV
6. Ascher DB
(2020) Prediction of rifampicin resistance beyond the RRDR using structure-based machine learning approaches
Scientific Reports 10:18120.

https://doi.org/10.1038/s41598-020-74648-y
- PubMed
- Google Scholar
1. Pruthi SS
2. Billows N
3. Thorpe J
4. Campino S
5. Phelan JE
6. Mohareb F
7. Clark TG
(2024) Leveraging large-scale Mycobacterium tuberculosis whole genome sequence data to characterise drug-resistant mutations using machine learning and statistical approaches
Scientific Reports 14:27091.

https://doi.org/10.1038/s41598-024-77947-w
- PubMed
- Google Scholar
(2008) Youden Index and optimal cut-point estimated from observations affected by a lower limit of detection
Biometrical Journal. Biometrische Zeitschrift 50:419–430.

https://doi.org/10.1002/bimj.200710415
- PubMed
- Google Scholar
1. Schmieder R
2. Edwards R
(2011) Quality control and preprocessing of metagenomic datasets
Bioinformatics 27:863–864.

https://doi.org/10.1093/bioinformatics/btr026
- PubMed
- Google Scholar
1. Serajian M
2. Marini S
3. Alanko JN
4. Noyes NR
5. Prosperi M
6. Boucher C
(2024) Scalable de novo classification of antibiotic resistance of Mycobacterium tuberculosis
Bioinformatics 40:i39–i47.

https://doi.org/10.1093/bioinformatics/btae243
- PubMed
- Google Scholar
1. Sharma K
2. Gupta M
3. Pathak M
4. Gupta N
5. Koul A
6. Sarangi S
7. Baweja R
8. Singh Y
(2006) Transcriptional control of the mycobacterial embCAB operon by PknH through a regulatory protein, EmbR, in vivo
Journal of Bacteriology 188:2936–2944.

https://doi.org/10.1128/JB.188.8.2936-2944.2006
- PubMed
- Google Scholar
1. Sun M
2. Ge S
3. Li Z
(2022) The role of phosphorylation and acylation in the regulation of drug resistance in Mycobacterium tuberculosis
Biomedicines 10:2592.

https://doi.org/10.3390/biomedicines10102592
- PubMed
- Google Scholar
Software
1. Tasmin M
2. Green AG
(2026) MTB_Mut_Clust, version swh:1:rev:e0ec365f30686fba727fbb20a8a5fbc9bee34eb7
Software Heritage.

https://archive.softwareheritage.org/swh:1:dir:520da0a3a8476fe406b0129ac24f5f43ea4bcd7a;origin=https://github.com/aggreen/MTB_Mut_Clust;visit=swh:1:snp:8e191d962092991ba726fd957e75efd8a1893a28;anchor=swh:1:rev:e0ec365f30686fba727fbb20a8a5fbc9bee34eb7
1. Varadi M
2. Anyango S
3. Deshpande M
4. Nair S
5. Natassia C
6. Yordanova G
7. Yuan D
8. Stroe O
9. Wood G
10. Laydon A
11. Žídek A
12. Green T
13. Tunyasuvunakool K
14. Petersen S
15. Jumper J
16. Clancy E
17. Green R
18. Vora A
19. Lutfi M
20. Figurnov M
21. Cowie A
22. Hobbs N
23. Kohli P
24. Kleywegt G
25. Birney E
26. Hassabis D
27. Velankar S
(2022) AlphaFold protein structure database: massively expanding the structural coverage of protein-sequence space with high-accuracy models
Nucleic Acids Research 50:D439–D444.

https://doi.org/10.1093/nar/gkab1061
- PubMed
- Google Scholar
1. Vargas R Jr
2. Freschi L
3. Spitaleri A
4. Tahseen S
5. Barilar I
6. Niemann S
7. Miotto P
8. Cirillo DM
9. Köser CU
10. Farhat MR
(2021) Role of epistasis in amikacin, kanamycin, bedaquiline, and clofazimine resistance in Mycobacterium tuberculosis complex
Antimicrobial Agents and Chemotherapy 65:01164-21.

https://doi.org/10.1128/AAC.01164-21
- Google Scholar
1. Vargas R Jr
2. Luna MJ
3. Freschi L
4. Marin M
5. Froom R
6. Murphy KC
7. Campbell EA
8. Ioerger TR
9. Sassetti CM
10. Farhat MR
(2023) Phase variation as a major mechanism of adaptation in Mycobacterium tuberculosis complex
PNAS 120:e2301394120.

https://doi.org/10.1073/pnas.2301394120
- PubMed
- Google Scholar
1. Walker BJ
2. Abeel T
3. Shea T
4. Priest M
5. Abouelliel A
6. Sakthikumar S
7. Cuomo CA
8. Zeng Q
9. Wortman J
10. Young SK
11. Earl AM
(2014) Pilon: an integrated tool for comprehensive microbial variant detection and genome assembly improvement
PLOS ONE 9:e112963.

https://doi.org/10.1371/journal.pone.0112963
- PubMed
- Google Scholar
1. Walker TM
2. Miotto P
3. Köser CU
4. Fowler PW
5. Knaggs J
6. Iqbal Z
7. Hunt M
8. Chindelevitch L
9. Farhat M
10. Cirillo DM
11. Comas I
12. Posey J
13. Omar SV
14. Peto TE
15. Suresh A
16. Uplekar S
17. Laurent S
18. Colman RE
19. Nathanson C-M
20. Zignol M
21. Walker AS
22. Crook DW
23. Ismail N
24. Rodwell TC
25. CRyPTIC Consortium
26. Seq&Treat Consortium
(2022) The 2021 WHO catalogue of Mycobacterium tuberculosis complex mutations associated with drug resistance: A genotypic analysis
The Lancet. Microbe 3:e265–e273.

https://doi.org/10.1016/S2666-5247(21)00301-3
- PubMed
- Google Scholar
1. Wang Y
2. Jiang Z
3. Liang P
4. Liu Z
5. Cai H
6. Sun Q
(2024) TB-DROP: deep learning-based drug resistance prediction of Mycobacterium tuberculosis utilizing whole genome mutations
BMC Genomics 25:167.

https://doi.org/10.1186/s12864-024-10066-y
- Google Scholar
1. Wood DE
2. Salzberg SL
(2014) Kraken: ultrafast metagenomic sequence classification using exact alignments
Genome Biology 15:R46.

https://doi.org/10.1186/gb-2014-15-3-r46
- PubMed
- Google Scholar
Preprint
(2025) Risk‑based prediction of novel AMR variants using protein language models
bioRxiv.

https://doi.org/10.1101/2025.09.12.672331
- Google Scholar
Website
1. World Health Organization
(2021a) Catalogue of mutations in Mycobacterium tuberculosis complex and their association with drug resistance
Accessed November 15, 2023.

https://www.who.int/publications/i/item/9789240082410
Report
1. World Health Organization
(2021b)
Global Tuberculosis Report

World health Organization.
- Google Scholar
1. Zeng J
2. Platig J
3. Cheng T-Y
4. Ahmed S
5. Skaf Y
6. Potluri L-P
7. Schwartz D
8. Steen H
9. Moody DB
10. Husson RN
(2020) Protein kinases PknA and PknB independently and coordinately regulate essential Mycobacterium tuberculosis physiologies and antimicrobial susceptibility
PLOS Pathogens 16:e1008452.

https://doi.org/10.1371/journal.ppat.1008452
- PubMed
- Google Scholar
(2013) Access channel residues Ser315 and Asp137 in Mycobacterium tuberculosis catalase-peroxidase (KatG) control peroxidatic activation of the pro-drug isoniazid
Chem. Commun 49:11650–11652.

https://doi.org/10.1039/C3CC47022A
- Google Scholar

Article and author information

Author details

Anna G Green
1. Department of Biomedical Informatics, Harvard Medical School, Boston, United States
2. Manning College of Information and Computer Sciences, University of Massachusetts, Amherst, United States
Contribution
Conceptualization, Resources, Data curation, Software, Formal analysis, Supervision, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing – original draft, Project administration, Writing – review and editing

For correspondence
annagreen@umass.edu

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0001-7548-3682
Mahbuba Tasmin

Manning College of Information and Computer Sciences, University of Massachusetts, Amherst, United States

Contribution
Data curation, Software, Investigation, Methodology, Writing – original draft

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0003-1884-8838
Roger Vargas Jr

Department of Biomedical Informatics, Harvard Medical School, Boston, United States

Contribution
Resources, Data curation, Writing – review and editing

Competing interests
No competing interests declared
Maha Reda Farhat
1. Department of Biomedical Informatics, Harvard Medical School, Boston, United States
2. Division of Pulmonary & Critical Care, Massachusetts General Hospital, Boston, United States
Contribution
Conceptualization, Resources, Data curation, Supervision, Funding acquisition, Validation, Investigation, Writing – original draft, Project administration, Writing – review and editing

For correspondence
Maha_Farhat@hms.harvard.edu

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0002-3871-5760

Funding

National Institutes of Health (F32AI161793)

Anna G Green

National Science Foundation Graduate Research Fellowship Program (DGE1745303)

Roger Vargas Jr

The funders had no role in study design, data collection, and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank members of the Farhat lab at Harvard Medical School and the SAGE lab at the University of Massachusetts Amherst for valuable discussion about the project. Computational resources and support were provided by the Orchestra High Performance Compute Cluster at Harvard Medical School, which is funded by the NIH (NCRR 1S10RR028832-01). AGG was supported by NIH/NIAID F32AI161793. RVJ was supported by the National Science Foundation Graduate Research Fellowship under Grant No. DGE1745303.

Version history

Preprint posted: September 25, 2025
Sent for peer review: October 17, 2025
Reviewed Preprint version 1: December 12, 2025
Reviewed Preprint version 2: September 1, 2026
Version of Record published: October 9, 2026

Cite all versions

You can cite all versions using the DOI https://doi.org/10.7554/eLife.109450. This DOI represents all versions, and will always resolve to the latest one.

Copyright

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.