ProPickML: Advancing Clinical Diagnostics with Automated Peak Picking in Label-Free Targeted Proteomics.

Abstract

In targeted proteomics utilizing Selected Reaction Monitoring (SRM), the precise detection of specific peptides within complex mixtures remains a significant challenge, particularly due to noise and interference in chromatograms. Existing methodologies, such as isotopic labeling and scoring algorithms, offer partial solutions but are constrained by high run times and elevated false discovery rates. To address these limitations, we have developed ProPickML a machine learning-based tool designed to accurately identify peptide peaks across diverse data sets, independent of the assumed presence of the peptide. This model was trained on a manually labeled data set and subsequently validated to assess its predictive accuracy. The results demonstrate that the model reliably identifies peptide peaks in the presence of noise, achieving a Matthews correlation coefficient (MCC) of 0.81 on an independent test data set, surpassing mProphet's MCC of 0.71. Implemented in R as ProPickML, this tool offers a competitive, cost-effective alternative to existing techniques, significantly reducing reliance on isotopic labeling and enhancing the accuracy of peptide identification in SRM workflows.