Generation of peptide detectability datasets from single DIA experiment for prediction model fine-tuning

Accurate prediction of peptide detectability in mass spectrometry-based proteomics is critical for improving both protein identification and quantification. Current models generally estimate detectability from amino acid sequences; however, peptide detectability is influenced by the instruments, acquisition methods, and experimental conditions, limiting the applicability of sequence-based models. State-of-the-art approaches mitigate this issue by fine-tuning models for each experimental setup, yet this strategy demands extensive training datasets-often comprising up to 300 000 peptides-incurring substantial experimental and computational costs.

In this study, we present a complementary approach for generating peptide detectability datasets directly from a single DIA experiment. These datasets enable fine-tuning of prediction models with minimal raw data, while improving adaptation to specific experimental conditions. This strategy substantially reduces both the data and cost requirements typically associated with model training. Furthermore, we show that filtering search libraries based on predicted detectability increases peptide identification rates and decreases computational time.

Code is available on https://github.com/leoschn/Detectability. Data are available via ProteomeXchange with identifier https://proteomecentral.proteomexchange.org/cgi/GetDataset?ID=PXD076276PXD076276.