Finding the Peak in the Mass-Stack: Rapid and Accurate Detection of Virulence and Resistance in Clinical Routine Diagnostics with MALDI-TOF Mass Spectrometry

Infectious diseases are amongst the most common causes of morbidity and death. Moreover, the global increase of antimicrobial resistance (AMR) threatens to undo earlier achievements in modern medicine. Accurate and fast bacterial identification in clinical diagnostics is key, as it forms the basis of a tailored treatment. The most commonly used tool for bacterial species identification in clinical routine diagnostics is Matrix-Assisted Laser Desorption Ionization-Time of Flight Mass Spectrometry (MALDI-TOF MS). MALDI-TOF MS is fast, accurate, and costeffective. However, three key problems remain: (i) not all clinically relevant bacterial species can reliably be identified using MALDI-TOF MS; (ii) virulent strains within a species are not routinely identified; and (iii) the time from sample collection until an AMR profile is available still requires 48-72 hours. How can these critical aspects in diagnostics be overcome? And what would be the potential impact for the patient?
In this thesis, I aimed to increase the resolution of bacterial identification by MALDI-TOF MS in clinical routine diagnostics with a focus on the genus Klebsiella (Chapter I) and the species E. coli (Chapter II) and to predict AMR from MALDI-TOF mass spectra, using machine learning approaches (Chapter V).
In the first part of my thesis (Chapter I), I established a ribosomal marker-based approach to distinguish the species within the genus Klebsiella with MALDI-TOF MS. Next, I applied this to a large, international dataset of mass spectra (n=33,160) and AMR profiles (n=7,876) to identify species-specific trends in AMR profiles. Further, I linked the species classification to clinical outcomes compiled from a single healthcare centre (n=957 clinical cases). I found that strains of the K. oxytoca complex to be significantly more likely causes invasive infections than strains of the K. pneumoniae complex.
To anticipate the course of an infection, it is necessary to know which bacterial strains have the potential to cause life-threatening diseases, such as sepsis. In the second project of my thesis (Chapter II), I, therefore, isolated over 1,000 E. coli strains from urinary tract- and bloodstream infections and compiled patient characteristics and outcomes of the respective clinical cases (n=831). Applying a bacterial genome wide association study (bGWAS), I substantiated papGII as an important, patient-independent bacterial factor for causing invasive infections (bacteraemia). However, I could not identify MALDI-TOF MS peaks specific for E. coli strains carrying this virulence factor and rapid sequence amplification-based diagnostics might be more suitable and faster.
MALDI-TOF mass spectral quality (MSQ) is crucial for accurate species identification. While analysing MALDI-TOF mass spectra from different healthcare centres, I observed differing numbers of detected ribosomal marker masses. MSQ is currently not precisely defined nor regularly assessed in diagnostic laboratories. I, therefore, sought to identify mass spectral features, which can be used to precisely describe MSQ and identify simple protocols yielding the highest MSQ for varying bacterial strains (Chapter III).
Further, I identified a large heterogeneity of MSQ from mass spectra acquired in 36 international diagnostic laboratories, mainly driven by a few particularly well or poorly performing MALDI-TOF MS devices and likely linked to sample preparation practices (Chapter IV). Applying the simple protocols identified in Chapter III improved MSQ for previously poorly performing devices/laboratories. The resolution of MALDI-TOF MS based bacterial identification can be improved with a high MSQ in clinical routine diagnostics. A standardised MSQ will likely benefit direct phenotype prediction by supervised classification algorithms (i.e. machine learning).
To assess the potential of machine learning based analysis approaches to predict AMR from MALDI-TOF mass spectra, we compiled an extensive dataset of over 300,000 routinely acquired MALDI-TOF mass spectra with matching AMR data from four healthcare centres (Chapter V). We yielded accurate predictions for important species and clinically relevant antibiotic drugs: for Ceftriaxone (as an indicator for ESBL) for E. coli and K. pneumoniae, we yield an area under the receiver operating characteristic curve (AUROC) of 0.74 for both and for Oxacillin (as an indicator for MRSA) for S. aureus with an AUROC of 0.80. The classification was most accurate if the classifiers were trained on spectra of a single species acquired at the same centre and in close temporal proximity to the test set.
Overall, my thesis showed (i) the potential of MALDI-TOF MS to identify bacteria with higher resolution, (ii) that AMR can accurately be predicted from MALDI-TOF mass spectra and (iii) high MSQ is essential to translate these advances into clinical routine diagnostics.