Feature selection in validating mass spectrometry database search results.

Area of Science:

• Proteomics
• Bioinformatics
• Computational Biology

Background:

• Tandem mass spectrometry (MS/MS) is crucial for protein identification but requires validation to minimize false positives.
• Current validation methods often rely on manual inspection or empirical filters, which can be time-consuming and less accurate.
• Machine learning offers potential for faster and more precise validation of MS/MS data.

Purpose of the Study:

• To identify critical peptide properties for improving validation models in MS/MS-based protein identification.
• To develop and evaluate machine learning models for peptide property selection and classification.
• To assess the performance of models with and without search engine scores.

Main Methods:

• Utilized random forest and support vector machine algorithms for feature selection.
• Developed validation models incorporating optimized peptide features.
• Created classification models based on peptide physicochemical properties and sequence environment, independent of search engine scores.

Main Results:

• An optimized feature set reduced false positives by 58% compared to models using only search engine scores, maintaining 0.8 sensitivity.
• Developed classification models using peptide properties achieved 0.8 AUC, 0.78 accuracy, and 0.7 specificity.
• Identified key peptide properties influencing fragmentation and ionization.

Conclusions:

• Machine learning-based feature selection significantly enhances the accuracy of peptide validation in MS/MS.
• Peptide physicochemical and sequence properties can be effectively used for classification, even without search engine scores.
• The identified properties can improve existing software for peptide sequencing, database searching, and validation.