Genomic variant-driven prediction of azole resistance in Aspergillus fumigatus using GWAS and machine learning

Azole resistance in Aspergillus fumigatus , a major cause of invasive aspergillosis, threatens public health. Known drivers include cyp51A/B mutations (e.g., TR34/L98H and TR46/Y121F/T289A), yet existing studies have largely focused on clinical isolates and known genetic determinants, leaving gaps in understanding broader genomic contributions. This study aimed to develop a machine learning–based framework to overcome limitations of traditional GWAS and identify novel resistance loci beyond cyp51A . A global collection of 590 A. fumigatus strains was analyzed, including whole-genome sequencing (WGS) data from 15 countries and resistance phenotypes using CLSI/EUCAST guidelines. Phylogenetic analysis revealed four clades without geographic clustering. Clade III harbored the highest proportion of resistant strains (ITR: 51.89%, POS: 50.48%, VOR: 38.68%), predominantly linked to cyp51A tandem repeats. In contrast, Clade IV strains frequently carried point mutations but showed lower resistance rates. GWAS was performed using PLINK and GAPIT frameworks, and 7,098 high confidence SNPs were selected for ML modeling. Ten classifiers were evaluated using repeated random 80:20 train-test splits, with five-fold cross-validation used for RFECV-based feature selection and model tuning where applicable. RF and XGBoost achieved superior performance, with mean AUCs > 95% and accuracy > 88% across all azoles. Penalized logistic regression outperformed SVM and AdaBoost. Decision trees exhibited the lowest accuracy. The SNP SCM000172.1_1781459 was identified as a key predictor for all three azoles. Cross-resistance analysis revealed significant overlap between ITR and POS resistance loci, whereas VOR-associated loci were distinct, suggesting divergent mechanisms. The findings provide actionable insights for resistance surveillance, antifungal development, and tailored treatment strategies.