Ad Hoc Antimicrobial Susceptibility Testing from MALDI-TOF MS Spectra in the Clinical Microbiology Laboratory

Introduction of matrix-assisted laser desorption ionization time of flight mass spectrometry (MALDI-TOF MS) in clinical microbiology laboratories over the past decade has revolutionized the way bacteria (and, to a certain extent, fungi) are identified. Gone are the days of relying on a discrete number of microbial phenotypes, measured stepwise over the course of days, to identify microorganisms. MALDI-TOF MS surpasses phenotypic identification methods in terms of cost, speed, and accuracy (1). However, laboratories remain reliant on a microorganism’s phenotype for antimicrobial susceptibility testing (AST). Standard of care methods for AST involve incubation of the microorganism with 2-fold dilutions of antimicrobial and observation of a growth/no growth endpoint [i.e., the minimal inhibitory concentration (MIC)]. MICs are interpreted using clinical cutoffs (or breakpoints) to provide a prediction of clinical outcome: high probability of treatment success (susceptible) vs low probability (resistance). The MIC phenotype, while time-honored, does not generate timely results. A typical MIC requires up to 24 h to observe the phenotype, which can only be performed once the infecting organism has been recovered in pure culture. All told, it is common to wait 72 h or more before final AST results are available. In an era of rapidly emerging antimicrobial resistance (AMR), this is far too long, as patients are either treated too narrowly or too broadly for their infections, fostering further emergence and spread of AMR and negatively impacting patient outcomes. Exaptation of the data generated when identifying a bacterium by MALDI-TOF MS into an AST result is a tantalizing notion. If proven, such an approach would not only improve efficiencies within the laboratory but also reduce time to AST results by 1 to 2 days. Weis et al. (2) explored this approach in a recent study published in Nature Medicine. The authors constructed a machine-learning predictive model by harnessing features of MALDI-TOF mass spectra from >300 000 bacterial isolates identified by 4 medical institutions and comparing these to laboratory-derived antimicrobial susceptibility profiles (2). The authors were able to predict AST results from the mass spectra with surprising accuracy for a proof-of-concept study and furthermore showed these predictions have the potential to benefit patient management and antimicrobial stewardship. At face value, it seems implausible that mass spectra measured for identification purposes could be used to predict an organism’s susceptibility. These spectra do not necessarily include peaks that correspond to individual resistance mechanisms, such as antimicrobial hydrolyzing enzymes or alternative targets. However, AMR is complex and often multifactorial, meaning the presence or absence of individual AMR determinants rarely predicts both AMR and susceptibility. As such, identification of proteins directly linked to specific AMR mechanisms may be less informative exploiting features of the spectrum which correlate with phenotypic AST results. What then might be some limitations of this approach? The spectra routinely captured by MALDI-TOF MS for bacterial identifications include ribosomal proteins, which relate nicely to bacterial strain types. AMR is largely driven by the global expansion of highly successful, multidrug resistant clades of bacteria (3). Strains of bacterial belonging to multidrug resistant clades maintain AMR determinants once acquired with no apparent cost to fitness, a feature that results in close relationship between strain type and resistance. As 1 example, Escherichia coli ST131 is the most common cause of extraintestinal E. coli infections in many regions of the world. Most ST 131 E. coli express the blaCTX-M-15 extended spectrum beta-lactamase gene, rendering them resistant to expanded spectrum cephalosporins, and many are also fluoroquinolone resistant. As such, it is likely that machine learning models predict strain type, which in turn is a correlate of AMR. To be clear, this is not necessarily a disadvantage—rapid differentiation of E. coli ST131 by the laboratory through use of MALDI-TOF MS is of high clinical value, predicting ceftriaxone resistance and enabling more rapid therapeutic escalation for a patient’s infection (4). However, multidrug resistant clades vary both temporally and regionally (3), which means models generated by these approaches require training and validation with local data sets and contemporary isolates. Indeed, Weis et al. (2) demonstrated both: site-specific training data outperformed training on data from all 4 sites for 1 hospital and inclusion of current samples was necessary to predict AMR. Local variability in AMR is inflated across regions and countries: as an example, Swiss hospitals (such as those used in the study by Weis et al.) reported only 9% ceftriaxone resistant E. coli in 2020, whereas neighboring Italian hospitals reported 26% ceftriaxone resistant E. coli in this same timeframe (5). Clearly, model representation is heavily influenced by local epidemiology and will require enrichment with resistant (or susceptible) isolates prior to being portable to regions of higher (or lower) AMR. It is also important to recognize that AMR is also fueled by introduction of new resistance determinants bacterial strains through the horizontal transfer via plasmids, transposons, or integrons. Plasmids carrying AMR genes have been found in bacteria belonging to multiple sequence types, species, and genera circulating in a single hospital (4). Such events would be difficult to identify by a model that is weighted heavily toward strain typing. Quite possibly, horizontal gene transfer could be identified by harnessing other features present in MALDI-TOF MS spectra. This is more than just theoretical, as Weis et al. (2) demonstrate identification of a MALDI-TOF peak in their spectra that corresponds to the peptide PSM-mec, which is encoded on certain SCCmec cassettes that harbor the mecA oxacillin-resistance gene in staphylococci (2, 6). However, the universality of this is requires further study. Finally, a major limitation for the development of any machine learning model for AST relates to source AST data. The MIC cutoffs that define interpretive categories (e.g., susceptible vs resistant) are not static. These breakpoints are routinely adjusted by organizations such as CLSI and the European Committee on Antimicrobial Susceptibility Testing in response to emerging data that show increased risk of treatment failures in isolates that were historically classified as susceptible but with elevated MICs. Clinical laboratories are slow to make these adaptations, and unfortunately, global harmonization of breakpoints has remained elusive. Weis et al. (2) binned MIC data into susceptible vs resistant categories, as reported by the laboratory, in training data sets for prediction models. Depending on the timeframe of when the data were generated for the training set, inclusion of data derived from obsolete breakpoints may significantly impact the model’s reliability. Such models can be updated, provided source MIC data are available, but MICs are unfortunately not always captured in large, publicly available data sets such as the one used by Weis et al., particularly if methods that do not yield an MIC (e.g., disk diffusion) were used for AST. Generation of MIC prediction models is 1 possibility but unlikely due both to the enhanced level of discrimination required of machine learning models and to the inherent variability of MIC testing, not to mention spectra generated by MALDI-TOF MS in clinical practice. AMR is one of the most urgent threats to modern medicine. Major activity is ongoing to improve AST methods in the clinical laboratory, including the generation of rapid and reliable results. Approaches that harness existing technology and data, such as the method described by Weis et al., (2) will expedite uptake of novel approaches. However, continual surveillance of MALDI-TOF spectra, MIC phenotypes, and treatment outcomes will be critical to the further development and sustainability of such diagnostics. Increased availability of paired MALDI-TOF spectra and phenotypic resistance data from a broad range of clinical populations are needed to facilitate expansion of these approaches. MALDI-TOF MS, matrix-assisted laser desorption ionization time of flight mass spectrometry; AST, antimicrobial susceptibility testing; MIC, minimal inhibitory concentration; AMR, antimicrobial resistance. All authors confirmed they have contributed to the intellectual content of this paper and have met the following 4 requirements: (a) significant contributions to the conception and design, acquisition of data, or analysis and interpretation of data; (b) drafting or revising the article for intellectual content; (c) final approval of the published article; and (d) agreement to be accountable for all aspects of the article thus ensuring that questions related to the accuracy or integrity of any part of the article are appropriately investigated and resolved. R.M. Humphries, CLSI, IDSA, CAP, Journal of Clinical Microbiology. R.M. Humphries, Pattern Biosciences, Qiagen, Specific Diagnostics, Accelerate Diagnostics, Torus, Merck, QPex, Roche. R.M. Humphries, Accelerate Diagnostics. R.M. Humphries, IHMA, CLSI, IDSA. R.M. Humphries, bioMerieux, Momentum Biosciences, Pattern Biosciences, Specific Diagnostics, Qiagen (to VUMC). None declared. None declared.