Genus-level transfer learning of Matrix-assisted laser desorption/ionization time-of-flight mass spectrometry data predicts antibiotic resistance with greater accuracy

Bacterial resistance, driven by excessive antibiotic use, has rendered many traditional antibiotics ineffective. Despite the advantages of applying matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) to predict bacterial antimicrobial resistance, limited databases and a lack of high-quality data hinder this effort. This study aimed to address this lack of data by integrating MALDI-TOF MS technology with transfer learning approaches, utilizing genus-level data, to predict species-specific bacterial antimicrobial resistance. The data were retrieved from the DRIAMS dataset. A multilayer perceptron deep neural network was applied for pretraining and fine-tuning, integrating genus-level data to enhance the accuracy of the species-specific predictions. Notably, fine-tuned models enhanced antibiotic resistance prediction. In Staphylococcus aureus, the area under the receiver operating characteristic curve (AUROC) improved from 0.95 to 0.96 and the area under the precision-recall curve (AUPRC) improved from 0.85 to 0.88 for oxacillin; for fusidic acid, the AUROC improved from 0.80 to 0.81 and the AUPRC from 0.32 to 0.36; and for ciprofloxacin, the AUROC improved from 0.80 to 0.82 while the AUPRC remained 0.57. In Klebsiella pneumoniae, the AUROC held steady at 0.75, while the AUPRC declined from 0.45 to 0.41 for ciprofloxacin, indicating potential limitations for certain pathogen-antibiotic combinations. Despite such variability, the overall improvements across multiple strains highlight the potential of transfer learning in antimicrobial resistance prediction and underscore the need to expand genus-level databases to improve species-level diagnostic accuracy, thereby enhancing clinical diagnostics.