Fast and Accurate Bacterial Species Identification in Urine Specimens Using LC-MS/MS Mass Spectrometry and Machine Learning*

Abstract

We have developed a new method for the identification of bacterial species causing Urinary Tract Infections. The first training step used DIA analysis on multiple replicates of bacterial inoculates to define a peptide signature by machine learning classifiers. In a second identification step, the signature is monitored by targeted proteomics on unknown samples. This fast, culture-free and accurate method paves the way of the development of new diagnostic approaches limiting the emergence of antimicrobial resistances.

Emerging Options for the Diagnosis of Bacterial Infections and the Characterization of Antimicrobial Resistance

“…As MALDI-TOF MS is performed on an automatable, inexpensive, and rapid platform, several devices have been developed for routine use in clinical micro-biology laboratories. At the current time, the VITEK ® MS (bioMérieux, Inc., Durham, NC, USA) and the MALDI Biotyper CA System (Bruker Daltonics, Inc., Manning Park, MA, USA), are both approved by the US Food and Drug Administration (FDA) for identification of cultured bacteria [ 15 , 38 ]. During a typical analysis, bacteria are identified by their unique mass spectrum via the comparison of findings generated on-site to a database that includes spectra obtained from pure bacterial colonies.…”

“…First, as the database consists of spectra from pure colonies, pathogen identification typically requires a time-consuming culture step to facilitate isolation of bacterial colonies from the clinical sample. As a result, the total time-to-result is typically reduced to a comparatively limited extent, to <50 h, compared to classical biochemical methods which typically require 2–4 days to complete [ 38 ]. Furthermore, in most cases, the characterization of polymicrobial infections can only be achieved by analysis of several colonies that were selected by visual inspection of the culture plate.…”

“…It is important to recognize that several of the aforementioned techniques involve nonselective enrichment of all bacterial cells present in a biological specimen. At the current time, MALDI-TOF-based approaches are not highly effective for assessing polymicrobial samples; positive identification of a single bacterium can be achieved in 64% of the samples and several bacterial species remain unidentified [ 15 , 38 , 44 , 45 , 46 ]. Typical factors leading to false or impossible identification include the use of inappropriate software, lack of unique characters that permit differentiation between spectral signatures, and low abundance of suitable target proteins, which are then masked by high background levels [ 44 ].…”

“…The use of this method also provides an additional measure of confidence, as it facilitates sensitive detection and identification of unique peptide markers in a poly-bacterial mixture [ 56 , 57 , 58 ]. As one example, results from the study of Roux–Dalvai and colleagues [ 38 ] revealed that an HPLC-MS-based assessment of peptides, followed by the establishment of a set of unique peptide signatures using machine learning methods (i.e., top-down approach) resulted in the direct identification of fifteen distinct pathogenic bacterial species in urine specimens. Moreover, if the acquired peptide signatures are systematically evaluated using an in silico library generated from public databases (e.g., Swiss–Prot and Translated EMBL Nucleotide Sequence Data Library (TrEMBL); a bottom-up approach), the number of identifiable isolates can be increased to include more than 12,000 bacterial strains [ 59 ].…”

Increasing Complexity to Simplify Clinical Care: High Resolution Mass Spectrometry as an Enabler of AI Guided Clinical and Therapeutic Monitoring

“…used libraries of peptides obtained from bacterial colonies in DDA mode, which were verified in urine in DIA mode, to train machine learning algorithms to define specific peptidic signatures which could accurately distinguish different microorganisms. [ 153 ] Using this method, the group managed to achieve >95% accuracy in identifying 15 species of bacterial which make up more than 80% of all urinary tract infections. The efficiency in identification (less than 4 h), lack of need for a bacterial culture and transferability of the method makes it ideal for future clinical implementation into diagnostic labs to hasten bacterial identification for targeted therapy and reduce the use of broad‐spectrum antibiotics.…”