Machine learning-assisted genomic profiling to identify differences between Bacillus Calmette-Guérin (BCG) vaccine strains and non-BCG wild-type Mycobacterium bovis

Background Distinguishing Bacillus Calmette-Guérin (BCG) vaccines from pathogenic Mycobacterium bovis is critical in neonatal diagnostics, particularly where polymerase chain reaction (PCR) methods fail to detect key genomic variations in tuberculosis (TB)-endemic regions. Methods We developed a machine learning framework analyzing 72 clinical isolates (28 BCG, 44 non-BCG) using whole-genome sequencing. Two classifiers were implemented: a random forest optimized by out-of-bag error minimization and a one-dimensional convolutional neural network (1D CNN) with dropout regularization (0.3–0.5). Feature selection through permutation testing and gradient activation mapping enhanced interpretability. Results Cross-validation demonstrated robust performance for both models: the random forest achieved 96% accuracy using 47 BCG attenuation-related genes, while the convolutional neural network (CNN) maintained high generalizability with 95.8% (±3.4%) mean accuracy and perfect recall across stratified five-fold validation, supported by strong discriminative capacity (mean area under the curve (AUC): 0.964 ± 0.046). Key biomarkers included metabolic reprogramming (ko01100) and secondary metabolite biosynthesis (ko01110) pathways. Conclusion This genomic approach resolves BCG diagnostic ambiguities through conserved attenuation markers, with the 47-gene panel enabling rapid assays that reduce neonatal overtreatment risks. Future integration of transcriptomic data could optimize these biomarkers for clinical deployment in high-TB-burden settings.