C106-16 Proteomic and Machine Learning Approaches for Predicting Drug Resistance in Nontuberculous Mycobacteria

Abstract Recent advances in mass spectrometry-based proteomics and machine learning have enabled new avenues for antimicrobial resistance prediction in nontuberculous mycobacteria (NTM). This study optimized peptide-based prediction models and proteomic workflows for M. abscessus and related species using clinical isolates from NIH and UNC cohorts. A hybrid database (“4DR + UniProt Reference Proteome”) was employed to benchmark peptide-level features influencing drug-resistant (DR) and drug-susceptible (DS) phenotypes. Confusion matrix analyses demonstrated an overall model accuracy of 84% (sensitivity 0.85, specificity 0.83), with false predictions largely associated with low peptide counts or borderline scores (∼0.5). SHAP feature analysis revealed that peptide-level contributions did not always align with positive predictive values (PPV), highlighting the complexity of peptide decision weighting. Complementary strong cation exchange (SCX) fractionation improved peptide coverage, though collection efficiency averaged 20%, prompting optimization of sample input and fraction inclusion. For the UNC cohort, varying PPV cut-offs (0.6-0.7) did not significantly enhance model discrimination, suggesting robustness across thresholds. Subsequent analyses expanded to 50 new NTM clinical isolates using the PEP-TORCH algorithm. Among these, 27 samples (54%) were accurately classified, including 20 M. abscessus strains. However, M. avium predictions remained inconsistent, likely reflecting limitations in reference proteome completeness and sample quality. Circos visualization of ∼650,000 peptides mapped to 4,092 reference proteins identified 51 DR-associated proteins with 20 supporting peptides each. Collectively, these results establish a scalable pipeline integrating LC-MS/MS peptide discovery, machine learning-based classification, and interactive R-based data management. Ongoing efforts aim to expand the peptide-protein reference atlas and enhance prediction reliability across NTM species. This abstract is funded by: NIH