Predicting Pseudomonas aeruginosa drug resistance using artificial intelligence and clinical MALDI-TOF mass spectra

Abstract

Matrix-assisted laser desorption/ionization-time of flight mass spectrometry (MALDI-TOF MS) is widely used in clinical microbiology laboratories for bacterial identification but its use for detection of antimicrobial resistance (AMR) remains limited. Here, we used MALDI-TOF MS with artificial intelligence (AI) approaches to successfully predict AMR in Pseudomonas aeruginosa, a priority pathogen with complex AMR mechanisms. The highest performance was achieved for modern β-lactam/β-lactamase inhibitor drugs, namely, ceftazidime/avibactam and ceftolozane/tazobactam. For these drugs, the model demonstrated area under the receiver operating characteristic curve (AUROC) of 0.869 and 0.856, specificity of 0.925 and 0.897, and sensitivity of 0.731 and 0.714, respectively. As part of this work, we developed dynamic binning, a feature engineering technique that effectively reduces the high-dimensional feature set and has wide-ranging applicability to MALDI-TOF MS data. Compared to conventional feature engineering approaches, the dynamic binning method yielded highest performance in 7 of 10 antimicrobials. Moreover, we showcased the efficacy of transfer learning in enhancing the AUROC performance for 8 of 11 antimicrobials. By assessing the contribution of features to the model's prediction, we identified proteins that may contribute to AMR mechanisms. Our findings demonstrate the potential of combining AI with MALDI-TOF MS as a rapid AMR diagnostic tool for Pseudomonas aeruginosa.IMPORTANCEPseudomonas aeruginosa is a key bacterial pathogen that causes significant global morbidity and mortality. Antimicrobial resistance (AMR) emerges rapidly in P. aeruginosa and is driven by complex mechanisms. Drug-resistant P. aeruginosa is a major challenge in clinical settings due to limited treatment options. Early detection of AMR can guide antibiotic choices, improve patient outcomes, and avoid unnecessary antibiotic use. Matrix-assisted laser desorption/ionization-time of flight mass spectrometry (MALDI-TOF MS) is widely used for rapid species identification in clinical microbiology. In this study, we repurposed mass spectra generated by MALDI-TOF and used them as inputs for artificial intelligence approaches to successfully predict AMR in P. aeruginosa for multiple key antibiotic classes. This work represents an important advance toward using MALDI-TOF as a rapid AMR diagnostic for P. aeruginosa in clinical settings.

Keywords: MALDI-TOF MS; Pseudomonas aeruginosa; antimicrobial resistance; machine learning.

Fig 1

Summary of antimicrobial susceptibility testing phenotypes. Number and proportion of non-resistant and resistant isolates of each antimicrobial. Broth microdilution assays were conducted for all 360 isolates to obtain MIC values for 11 antipseudomonal agents.

Fig 2

Schematic workflow of the study. Spectra preprocessing: raw spectra were preprocessed with a five-step pipeline (18). To generate the benchmarking feature sets, we conducted fixed-length binning (3 Da, 20 Da, and 30 Da) of preprocessed spectra, resulting in feature vectors of 6,000, 900, and 600 dimensions, respectively. Dynamic binning: we determined the mean number of intensity peaks in 500 Da intervals from 2,000 to 20,000 Da. The bin widths used for each region (ranging from 1 Da to 100 Da) are inversely proportional to the corresponding mean peak count of that region. Latent representation learning: a vision transformer model was trained to differentiate high-risk bacterial pathogens (20). The final hidden layer of the model was used as embedding to extract additional features from the MALDI-TOF spectra. Model development: we used 5-fold cross-validation with 80%–20% training–testing split to train models. Performance evaluation: AUROC was the key performance metric. Model interpretation: the Shapley additive explanations (SHAP) algorithm was used to identify feature bins with high importance (21). We then analyzed the regions of interest to interpret models and identify potential biological AMR determinants. Abbreviations: MALDI-TOF MS, matrix-assisted laser desorption/ionization–time of flight mass spectrometry; AST, antimicrobial susceptibility testing; BMD, broth microdilution; MIC, minimum inhibitory concentration; m/z, mass-to-charge ratio.

Fig 3

The predictive performance of dynamic binning across all tested machine learning models. The mean metric scores across 10 random splits are reported. The vertical black lines at the tip of each bar represent the 95% CI (95% CI). Abbreviations: AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve; AZT, aztreonam; IMI, imipenem; MER, meropenem; TAZ, ceftazidime; P/T, piperacillin/tazobactam; C/T, ceftolozane/tazobactam; CZA, ceftazidime/avibactam; CIP, ciprofloxacin; AMI, amikacin; TOB, tobramycin; COL, colistin.

Fig 4

Comparison of dynamic binning and fixed-length binning approaches. The binning approaches are ordered from left to right by the size of the feature set in descending order. Markers and lines indicate mean and 95% CI of the metrics, respectively. Abbreviations: AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve; AZT, aztreonam; IMI, imipenem; MER:, meropenem; TAZ, ceftazidime; P/T, piperacillin/tazobactam; C/T, ceftolozane/tazobactam; CZA, ceftazidime/avibactam; CIP, ciprofloxacin; AMI, amikacin; TOB, tobramycin.

Fig 5

Cross-institution performance of dynamic binning. Different training (y-axis) and testing (x-axis) sets using DRIAMS and Alfred Hospital data were used to train and evaluate the models. Data are reported as mean AUROC (95% CI) across 10 random splits. Abbreviations: AUROC, area under the receiver operating characteristic curve; DRIAMS, Database of Resistance Information on Antimicrobials and MALDI-TOF Mass Spectra.

Fig 6

Model interpretation analysis using the SHAP algorithm for (a) ceftazidime/avibactam, (b) ciprofloxacin, and (c) tobramycin. The left panel for each antimicrobial displays the top 10 contributing features in descending order of the mean absolute SHAP value (from top to bottom). The y-axis labels include the spectral ranges, with the asterisks indicating the presence of reviewed proteins within the region. The right panels display the SHAP value of each individual sample for the corresponding features on the left. Each sample dot is color-coded by the magnitude of the feature value.