Application of MALDI-TOF MS and machine learning for the detection of SARS-CoV-2 and non-SARS-CoV-2 respiratory infections

Abstract

Matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) could aid the diagnosis of acute respiratory infections (ARIs) owing to its affordability and high-throughput capacity. MALDI-TOF MS has been proposed for use on commonly available respiratory samples, without specialized sample preparation, making this technology especially attractive for implementation in low-resource regions. Here, we assessed the utility of MALDI-TOF MS in differentiating severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) vs non-COVID acute respiratory infections (NCARIs) in a clinical lab setting in Kazakhstan. Nasopharyngeal swabs were collected from inpatients and outpatients with respiratory symptoms and from asymptomatic controls (ACs) in 2020-2022. PCR was used to differentiate SARS-CoV-2+ and NCARI cases. MALDI-TOF MS spectra were obtained for a total of 252 samples (115 SARS-CoV-2+, 98 NCARIs, and 39 ACs) without specialized sample preparation. In our first sub-analysis, we followed a published protocol for peak preprocessing and machine learning (ML), trained on publicly available spectra from South American SARS-CoV-2+ and NCARI samples. In our second sub-analysis, we trained ML models on a peak intensity matrix representative of both South American (SA) and Kazakhstan (Kaz) samples. Applying the established MALDI-TOF MS pipeline "as is" resulted in a high detection rate for SARS-CoV-2+ samples (91.0%), but low accuracy for NCARIs (48.0%) and ACs (67.0%) by the top-performing random forest model. After re-training of the ML algorithms on the SA-Kaz peak intensity matrix, the accuracy of detection by the top-performing support vector machine with radial basis function kernel model was at 88.0%, 95.0%, and 78% for the Kazakhstan SARS-CoV-2+, NCARI, and AC subjects, respectively, with a SARS-CoV-2 vs rest receiver operating characteristic area under the curve of 0.983 [0.958, 0.987]; a high differentiation accuracy was maintained for the South American SARS-CoV-2 and NCARIs. MALDI-TOF MS/ML is a feasible approach for the differentiation of ARI without specialized sample preparation. The implementation of MALDI-TOF MS/ML in a real clinical lab setting will necessitate continuous optimization to keep up with the rapidly evolving landscape of ARI.IMPORTANCEIn this proof-of-concept study, the authors used matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) and machine learning (ML) to identify and distinguish acute respiratory infections (ARI) caused by SARS-CoV-2 versus other pathogens in low-resource clinical settings, without the need for specialized sample preparation. The ML models were trained on a varied collection of MALDI-TOF MS spectra from studies conducted in Kazakhstan and South America. Initially, the MALDI-TOF MS/ML pipeline, trained exclusively on South American samples, exhibited diminished effectiveness in recognizing non-SARS-CoV-2 infections from Kazakhstan. Incorporation of spectral signatures from Kazakhstan substantially increased the accuracy of detection. These results underscore the potential of employing MALDI-TOF MS/ML in resource-constrained settings to augment current approaches for detecting and differentiating ARI.

Keywords: COVID-19; MALDI-TOF MS; SARS-CoV-2; acute respiratory infection; machine learning.

Fig 1

Overall study workflow and description of the analyses. AC: asymptomatic controls; MALDI-TOF-MS: matrix-assisted laser desorption ionization mass spectrometry; ML: machine learning; NCARI: non-COVID acute respiratory infections; NPS: nasopharyngeal swab.

Fig 2

MALDI-TOF MS peak data generated using nasopharyngeal swabs and processed following the MALDI-TOF MS/ML pipeline developed by Nachtigall and colleagues (4). (A–C) representative MALDI-TOF MS spectra from symptomatic SARS-CoV-2+ (A), symptomatic non-SARS-CoV-2 (B), and a healthy control sample from Kazakhstan (C). The central line indicates median value of the spectra, while the shaded region on either side represents the interquartile interval. Insets depict a range from 3,000 to 5,500 m/z encompassing 70% (62/88) of the identified peaks. (D) PCA of the combined data set incorporating MALDI-TOF MS data both from Kazakhstan and South America (2020 SARS-CoV+ and symptomatic SARS-CoV-2-negative) (3). (E) Dendrogram of the mass spectra stratified by sub-group from the combined dataset based on the peak intensity matrix for Analysis I.

Fig 3

Classification accuracy of the MALDI-ML algorithms assessed on the data from Kazakhstan and South America. (A) Accuracy metrics for each of the seven ML models trained on the South American MALDI-TOF MS data (Analysis I in the current study) for the differentiation of study sub-groups. (B) ROC curves of the top-performing RF and SVM-L algorithms (Analysis I). (C) Accuracy metrics for each of the seven ML models trained on the combined South America-Kazakhstan data set (Analysis II in the current study) for the differentiation of study sub-groups. (D) ROC curves for the top-performing SVM-R and DT algorithms (Analysis II).