PeptideForest: Semisupervised Machine Learning Integrating Multiple Search Engines for Peptide Identification.

Area of Science:

• Proteomics
• Bioinformatics
• Computational Biology

Background:

• Bottom-up proteomics relies on accurate peptide spectrum matching (PSM).
• Existing algorithms have varying strengths and weaknesses, posing a challenge for users.
• There is a need for integrated approaches to improve PSM accuracy and reliability.

Purpose of the Study:

• To introduce PeptideForest, a semisupervised machine learning approach for integrating multiple PSM algorithms.
• To enhance the number of high-quality peptide-to-spectrum matches.
• To validate the quality of spectral assignments using TMT quantification.

Main Methods:

• Developed PeptideForest, a random forest classifier trained on assignments from multiple algorithms.
• Integrated PeptideForest into the Ursgal pipeline framework.
• Utilized TMT quantification of samples with known ground truths to assess spectral assignment quality.

Main Results:

• PeptideForest increased peptide-to-spectrum matches with q-value < 1% by 25.2 ± 1.6% compared to MS-GF+.
• The improvement in PSM quantity did not compromise spectral assignment quality, as validated by TMT quantification.
• Demonstrated that PeptideForest provides deeper insights into bottom-up proteomics data.

Conclusions:

• PeptideForest effectively integrates multiple algorithms to improve peptide spectrum matching in bottom-up proteomics.
• The method enhances the identification of peptides without sacrificing quantification accuracy.
• PeptideForest offers a valuable tool for gaining deeper insights in proteomics research.