AttnPep: A Self-Attention-Based Deep Learning Method for Peptide Identification in Shotgun Proteomics

Area of Science:

• Proteomics
• Bioinformatics
• Computational Biology

Background:

• Shotgun proteomics identifies peptides via search engines, generating peptide-spectra matches (PSMs).
• Many identified PSMs are incorrect, necessitating postprocessing software for reranking.
• Existing reranking methods face limitations like distribution dependency and shallow models.

Purpose of the Study:

• To introduce AttnPep, a deep learning model for rescoring PSM scores.
• To leverage the Self-Attention module for improved feature focus and PSM classification.
• To enhance the accuracy of peptide identification in shotgun proteomics.

Main Methods:

• Developed AttnPep, a deep learning model incorporating a Self-Attention module.
• Trained and evaluated AttnPep on PSM data, comparing it against PeptideProphet, Percolator, and proteoTorch.
• Assessed performance based on the identification of correct PSMs (q-value <0.01) and synthetic peptides.

Main Results:

• AttnPep demonstrated an average increase of 9.29% in correct PSMs compared to existing methods.
• The Self-Attention module enabled AttnPep to focus on relevant features, improving PSM discrimination.
• AttnPep showed superior ability in distinguishing correct from incorrect PSMs and identified more synthetic peptides in complex SWATH data.

Conclusions:

• AttnPep offers a significant advancement in peptide-spectra match scoring accuracy.
• The deep learning approach with Self-Attention effectively addresses limitations of current reranking software.
• AttnPep provides a more robust and accurate method for peptide identification in proteomics research.