RGTFormer: Predicting mutation-associated multi-drug resistance in Mycobacterium tuberculosis using a categorical gated transformer and relational graph convolutional network.

Tuberculosis (TB), caused by Mycobacterium tuberculosis, remains a major global health concern, especially due to multi-drug resistance. Resistance often arises from single nucleotide mutations in drug-target genes, making early prediction vital for effective treatment. Genomic sequencing enables resistance profiling, but accurate prediction requires advanced computational models. This study presents RGTFormer, a novel deep learning model combining a categorical gated transformer with a Relational Graph Convolutional Network (RGCN), to predict whether mutations confer resistance to first-line anti-TB drugs. It utilizes both sequence and structural features from mutations across six key resistance genes. The RGCN captures dependencies between mutations, while the transformer learns complex feature interactions. Evaluated via 10-fold cross-validation and an independent test set, RGTFormer achieved 98.67 % test accuracy and 97.15 % cross-validation accuracy, outperforming traditional machine learning and deep learning baselines. Ablation studies confirmed that the integration of RGCN with gated attention significantly enhances performance. RGTFormer provides a robust, interpretable, and efficient framework for mutation-driven drug resistance prediction in TB. It holds promise for supporting personalized treatment strategies and optimizing drug selection for resistant strains by providing biologically interpretable resistance predictions relevant to clinical decision-making.