Machine Learning-Based Fragment Selection Improves the Performance of Qualitative PRM Assays

Targeted mass spectrometry-based platforms have become
a valuable
tool for the sensitive and specific detection of protein biomarkers
in clinical and research settings. Traditionally, developing a targeted
assay for peptide quantification has involved manually preselecting
several fragment ions and establishing a limit of detection (LOD)
and a lower limit of quantitation (LLOQ) for confident detection of
the target. Established thresholds such as LOD and LLOQ, however,
inherently sacrifice sensitivity to afford specificity. Here, we demonstrate
that machine learning can be applied to qualitative PRM assays to
discriminate positive from negative samples more effectively than
a traditional approach utilizing conventional methods. To demonstrate
the utility of this method, we trained an ensemble machine learning
model using 282 SARS-CoV-2 positive and 994 SARS-CoV-2 negative nasopharyngeal
swabs (NP swab) analyzed using a targeted PRM method. This model was
then validated using an independent set of 200 positive and 150 negative
samples and achieved a sensitivity of 92% relative to results obtained
by RT-PCR, which was superior to a traditional approach that resulted
in 86.5% sensitivity when analyzing the same data. These results demonstrate
that machine learning can be applied to qualitative PRM assays and
results in superior performance relative to traditional methods.