PB-Net: Automatic peak integration by sequential deep learning for multiple reaction monitoring

Area of Science:

• Biochemistry
• Computational Biology
• Analytical Chemistry

Background:

• Mass spectrometry (MS) based proteomics is crucial for molecular and cellular biochemistry.
• Multiple reaction monitoring (MRM) is a key MS technique for molecule detection and quantification.
• Accurate analysis of MRM data, particularly peak integration, often requires expert annotation, limiting throughput.

Purpose of the Study:

• To develop a computational framework for fully automatic chromatographic peak integration in MRM data.
• To introduce PB-Net (Peak Boundary Neural Network), a deep learning method for accurate peak integration.
• To reduce the reliance on manual expert annotation in MS-based proteomics.

Main Methods:

• Developed PB-Net, a deep learning model based on sequential neural networks.
• Generated a large dataset of over 170,000 expert-annotated peaks from MS transitions.
• Trained and validated the model on peptides and intact glycopeptides across a wide dynamic range.

Main Results:

• PB-Net achieved near-perfect agreement (Pearson's r = 0.997) with human annotated ground truth.
• Demonstrated substantially improved robustness and accuracy compared to existing state-of-the-art peak integration software.
• The model performs accurately on unseen test samples, including peptides and glycopeptides.

Conclusions:

• PB-Net offers a highly accurate and robust solution for automatic chromatographic peak integration in MRM experiments.
• The method significantly enhances the efficiency and throughput of mass spectrometry data analysis.
• PB-Net benefits researchers in high-throughput MS applications, providing results comparable to human annotators.