Improved prediction of peptide detectability for targeted proteomics using a rank-based algorithm and...

Area of Science:

• Proteomics
• Bioinformatics
• Computational Biology

Background:

• Accurate prediction of proteotypic peptides is crucial for targeted proteomics and systems biology.
• Current prediction algorithms often rely on physicochemical parameters and require positive/negative training sets.
• A significant portion of proteomes lacks experimental proteomics evidence for guiding peptide selection.

Purpose of the Study:

• To introduce PeptideRank, a novel learning-to-rank approach for predicting peptide detectability.
• To eliminate the need for negative datasets in training peptide prediction models.
• To enable informed selection of proteotypic peptides for targeted quantification of observed and unobserved proteins.

Main Methods:

• Utilized a learning-to-rank algorithm for peptide detectability prediction from shotgun proteomics data.
• Employed a wide range of peptide properties to train ranking models.
• Evaluated performance using rank accuracy metrics.

Main Results:

• PeptideRank complements existing prediction algorithms.
• Optimal performance was achieved when trained on organism-specific shotgun proteomics data.
• PeptideRank demonstrated high accuracy for short to medium-sized, abundant proteins, including membrane proteins.

Conclusions:

• PeptideRank offers a novel, rank-based approach for predicting suitable proteotypic peptides for targeted proteomics.
• The method is inspired by information retrieval techniques, circumventing the need for curated training sets.
• Enables prediction of proteotypic peptides for unobserved proteins and non-model organisms lacking prior proteomics data.