Antibiotic-resistant tuberculosis remains a substantial global public health problem. Rapid genomic prediction of resistance could shorten time to effective therapy, but widespread clinical adoption is limited because most observed variants in resistance-associated genes are of uncertain significance. The authors developed a forecasting framework, FARM, to predict whether newly observed mutations in Mycobacterium tuberculosis resistance genes are likely to cause phenotypic resistance and to prioritize variants for follow-up.
FARM combines multiple, complementary feature types to represent mutations. The framework incorporates the local structural context of the amino-acid change, biophysical energy features that reflect energetic consequences of substitution, representations derived from protein language models, and mutational physicochemical descriptors from the AAIndex. By integrating these modalities, FARM aims to capture both mechanistic and learned sequence-level signals that may indicate resistance mechanisms.
Model training used 345 mutations labeled in the World Health Organization 2021 catalogue as resistance-associated or non-resistance-associated. The authors emphasize interpretability in their modeling approach, training models that can be examined to understand which features drive classification decisions. Precise details of model types, hyperparameters, feature importance metrics, and interpretable methods are reported in the source materials (and code is linked in the GitHub repository noted in the article). The work is presented as a forecasting tool to inform prioritization and experimental follow-up rather than as definitive clinical decision guidance.
Across holdout evaluations, models trained on the multimodal input achieved strong discrimination between resistance-associated and non-resistance-associated variants. Reported area under the receiver operating characteristic curve (AUC) values ranged from 0.843 to 0.943 depending on model configuration. These performance metrics indicate good separation on the curated labeled set, supporting the approach’s potential utility in predicting variant effects when labeled data are available for training.
To assess prospective performance, the authors conducted a temporal evaluation on 62 mutations that were reclassified after the training dataset release. Using this held-back set, the selected Combined model recovered 80.7% of resistant reclassifications. For resistant-class detection in this temporal set the model’s F1 score was reported as 86.8 and the AUC as 0.735. This temporal test evaluates the model’s ability to forecast resistance assignments that were not present in the original training catalogue.
FARM was applied to 4,525 mutations currently listed as uncertain significance (variants of unknown significance, VUS). The framework prioritized 696 candidate resistance mutations from this pool for further consideration and experimental validation. Prioritization is intended to guide future laboratory work and inform updates to variant catalogues rather than to serve as definitive reclassifications absent experimental evidence.
Among the prioritized candidates, the authors identified mutations in genes implicated in resistance to newly used antitubercular agents, specifically bedaquiline, delamanid, and pretomanid. The authors propose that such forecasts can help focus experimental resources and support timely updates to reference catalogues like the WHO database as new evidence accumulates.
This work is reported as a preprint on bioRxiv and has not undergone peer review. The authors declared no competing interests. The article references a public code repository for data and models (GitHub link in the source). The framework is presented as a tool for prioritization and to support future catalogue updates and experimental follow-up; the source does not claim that FARM outputs should replace laboratory phenotyping or formal catalogue reclassification without confirmatory evidence.
References and resources
The preprint provides a DOI and points to an external GitHub repository for code and data. For full methodological details, model specifications, and to access the code used in training and evaluation, consult the linked repository and the preprint PDF as cited in the source.