A cost-sensitive online learning method for peptide identification.

Area of Science:

• Proteomics
• Computational Biology
• Bioinformatics

Background:

• Post-database search is crucial for refining peptide-spectrum matches (PSMs) in tandem mass spectrometry (MS/MS).
• Existing methods struggle with large-scale and unbalanced datasets, necessitating more efficient learning strategies.
• Kernel methods offer potential for simplifying complex data relationships but face computational challenges.

Purpose of the Study:

• To develop an efficient learning strategy for accurate peptide identification on challenging datasets.
• To address the computational complexity and overfitting issues associated with complex learning models in proteomics.
• To improve the accuracy and stability of peptide identification, particularly for unbalanced datasets.

Main Methods:

• An online learning algorithm, OLCS-Ranker, was developed, processing one training sample at a time to reduce memory requirements.
• A cost-sensitive learning approach was implemented, assigning higher loss to decoy PSMs than target PSMs.
• Kernel methods were utilized to map data into high-dimensional spaces for simplified modeling.

Main Results:

• OLCS-Ranker significantly reduces memory requirements for computation.
• The cost-sensitive model effectively lowers the false discovery rate on unbalanced PSM datasets.
• OLCS-Ranker demonstrates superior accuracy and stability compared to existing methods.

Conclusions:

• OLCS-Ranker provides a more accurate and stable solution for peptide identification, especially on unbalanced datasets.
• The algorithm achieves substantial speed improvements, being 15-85 times faster than CRanker.
• This method offers a computationally efficient and effective approach for large-scale proteomics data analysis.