Prioritizing peptides for targeted mass spectrometry experiments using deep learning

Area of Science:

• Proteomics
• Analytical Chemistry
• Computational Biology

Background:

• Selecting optimal peptides is crucial for accurate protein quantification in targeted mass spectrometry.
• Current methods for peptide selection have limitations, including reliance on prior empirical data or sequence-based machine learning models that ignore precursor charge state.

Purpose of the Study:

• To introduce Bromo, a novel transformer-based deep learning model for ranking peptide precursors based on their relative mass spectrometry response.
• To address limitations of existing methods by incorporating precursor charge state and training on large-scale data.

Main Methods:

• Developed Bromo, a deep learning model utilizing a transformer architecture.
• Trained Bromo on millions of annotated peptide pairs from public data-independent acquisition mass spectrometry datasets.
• Evaluated Bromo's performance against existing sequence-based methods on diverse, independent datasets.

Main Results:

• Bromo consistently outperforms existing sequence-based methods in predicting peptide response for targeted mass spectrometry.
• The model effectively accounts for precursor charge state in peptide ranking.
• Fine-tuning Bromo on experiment-specific data improves target peptide selection across different sample preparation, matrix, and instrument conditions.

Conclusions:

• Bromo offers a robust and adaptable solution for selecting optimal target peptides in targeted mass spectrometry.
• The model's performance and adaptability make it a valuable tool for developing assays for selected reaction monitoring and parallel reaction monitoring.
• This approach has the potential to significantly enhance the reliability and efficiency of protein quantification experiments.