Protein structure sharpens resistance prediction in Mycobacterium tuberculosis
For years, resistance surveillance in Mycobacterium tuberculosis has leaned heavily on sequence alone: a mutation appears in a gene, and its likely effect is inferred from prior catalogues or from whether it disrupts protein function. That approach has practical value, but it leaves out a critical layer of biology. Proteins are not linear strings of amino acids in the cell; they fold into three-dimensional forms, and the position of a mutation within that folded architecture can determine whether it sits near a catalytic pocket, a binding interface, or a residue already linked to drug failure.
This study makes that structural context central to the analysis. Using high-confidence annotations across the M. tuberculosis proteome, the investigators combined 1,350 crystallography-derived structures with 2,337 AlphaFold predictions and over 31,000 bacterial isolates carrying mutations. The result is a large-scale test of a simple but clinically meaningful question: does protein structure improve the ability to distinguish mutations that confer antibiotic resistance from those that do not?
It does.
Clustering in three-dimensional space was not random
A key observation emerged early. Mutations in proteins already known to drive resistance were not scattered evenly across the folded protein surface. They clustered in three-dimensional space. In practical terms, nearby residues in the protein structure tended to accumulate the variants most relevant to resistance, even in proteins where the prevailing assumption has been that loss-of-function at any site could potentially alter susceptibility.
This matters because clustering implies biological selectivity rather than random drift. A resistance-linked mutation often does not need to destroy an entire protein; it may only need to shift a drug-binding region, alter conformational flexibility, or weaken an interaction that controls activity. Structural proximity is therefore a signal of mechanism. For tuberculosis, where resistance evolves under intense antibiotic pressure and where many clinically important variants are non-synonymous changes in proteins, that signal can be more informative than the amino-acid change alone.
The analysis identified more than 450 proteins in the proteome that displayed evidence of clustered mutations. Many already had a recognized relationship with antibiotic resistance. The breadth of that finding is striking because it suggests the pattern is not confined to a handful of canonical resistance genes. Instead, the structural signature reaches across a wider portion of the bacterial proteome, opening a path toward better variant interpretation in genes that may not yet be fully characterized in routine diagnostic pipelines.
A distance-based classifier reached high performance on held-out testing
The most immediate test of utility came from machine learning. Investigators trained a supervised classifier using one structural feature alone: the three-dimensional distance from each mutation to known resistance sites. On a held-out test set, the model achieved an F1 score of 96.5% for classifying mutations as resistance-conferring.
That is a substantial result because the model was not built on a broad panel of molecular descriptors or a complex multi-omic framework. It relied on structural proximity, the kind of feature that can be mapped onto proteins already represented in structural databases or predicted with modern tools. In other words, the model asks a relatively simple question and extracts a surprisingly strong answer from it. For clinical genomics, where the burden is often not discovering a mutation but deciding what it means, this kind of feature may be highly usable.
The performance also reinforces a mechanistic point. If a mutation close to a known resistance site is far more likely to confer resistance than one located elsewhere in the same protein, then structure is acting as a biologically grounded filter. It is capturing the local geography of function: active sites, pockets, channels, and interaction surfaces that are vulnerable to selective pressure from antibiotics.
AlphaFold predictions were useful, not merely supplementary
The dataset depended heavily on structural information from two sources: experimentally solved crystallography structures and AlphaFold-predicted models. The inclusion of more than 2,300 predicted structures is important because many bacterial proteins still lack experimental structural coverage. Here, predicted models were not treated as a stopgap. They were integrated into the analysis and still supported the central conclusion that structural context improves resistance prediction.
That finding has immediate diagnostic implications. If structural information had relevance only for the small fraction of proteins with solved crystal structures, its value would be limited. Instead, the analysis suggests that even largely AI-predicted structure sets can provide actionable information about mutation effect. For pathogens with extensive genomic surveillance but incomplete structural annotation, that opens the door to broader deployment of structure-aware interpretation tools.
The clinical consequence is straightforward: resistance prediction can move beyond cataloguing sequence changes toward a more spatially informed reading of the proteome. For tuberculosis programs that depend on rapid molecular diagnostics, the ability to prioritize suspicious variants more accurately could help distinguish true resistance drivers from incidental polymorphisms.
Why this approach changes the logic of variant interpretation
Sequence-based methods remain essential, but they can treat proteins as if every residue were functionally equivalent. Protein structure breaks that assumption. A mutation buried near a drug-binding pocket may be far more consequential than a mutation of the same type elsewhere. A residue on a flexible loop may alter enzyme dynamics; a change at an interface may disturb substrate access or drug occupancy. These are not subtle distinctions at the bench. They can decide whether a bacterium survives therapy.
The study supports a model in which structural context acts as a biologic coordinate system for mutation effect. That is especially useful in antibiotic resistance, where the same gene may tolerate many variants while a small subset drives clinically meaningful phenotypes. Mapping those variants onto three-dimensional space improves the signal-to-noise ratio in genomic interpretation.
A single mutation can look ordinary in sequence space and decisive in structure space.
What the findings mean for tuberculosis genomics and future resistance calls
The work does not replace functional validation, and it does not claim that every structurally proximal mutation will produce the same phenotype. Resistance remains a property of the organism, the drug, the mutation, and the broader genetic background. Still, the study provides a strong argument for incorporating structure into resistance prediction frameworks, especially when evaluating novel or rare variants discovered in clinical sequencing.
For laboratories and surveillance systems, the practical next step is clear enough: structural annotation should sit closer to the center of mutation interpretation rather than the margins. As more pathogen proteomes receive high-quality predicted structures, the approach may become increasingly feasible in routine analysis. The large number of clustered mutations identified here suggests that structural signals are abundant enough to matter at scale, not just in isolated case reports or a few classic resistance genes.
The main limitation is also the main challenge ahead: translating a powerful structural signal into consistent clinical decision support across diverse bacterial lineages, antibiotics, and testing environments. Even with a 96.5% F1 score in a held-out test set, real-world deployment will require careful validation against phenotypic susceptibility data and ongoing curation as new variants emerge. The study establishes the premise convincingly; the next step is to determine how reliably this structural framework can be used when treatment decisions depend on it.