pDeep: Predicting MS/MS Spectra of Peptides with Deep Learning

Area of Science:

• Proteomics
• Computational Biology
• Biochemistry

Background:

• Tandem mass spectrometry (MS/MS) proteomics relies on comparing experimental spectra to theoretical peptide spectra.
• Accurate theoretical spectrum prediction is vital for reliable peptide identification in MS/MS-based proteomics.

Purpose of the Study:

• To introduce pDeep, a novel deep neural network model for peptide spectrum prediction.
• To evaluate pDeep's performance in predicting various MS/MS fragmentation spectra.
• To explore the capability of pDeep in distinguishing challenging isobaric peptides.

Main Methods:

• Development of pDeep, a deep neural network utilizing bidirectional long short-term memory (BiLSTM).
• Prediction of higher-energy collisional dissociation (HCD), electron-transfer dissociation (ETD), and integrated (EThcD) MS/MS spectra.
• Analysis of intermediate neural network layers to understand amino acid physicochemical properties and fragmentation behavior.

Main Results:

• pDeep achieved median Pearson correlation coefficients greater than 0.9 for MS/MS spectrum prediction.
• The model demonstrated the ability to reveal physicochemical properties of amino acids from its intermediate layers.
• pDeep showed significant potential in differentiating highly similar peptides, including those with isobaric amino acids (e.g., N/G, Q/G, I/L).

Conclusions:

• pDeep offers a powerful deep learning approach for accurate peptide theoretical spectrum prediction in proteomics.
• The model enhances the identification of peptides, particularly those with isobaric amino acids, overcoming limitations of traditional search engines.
• pDeep contributes to advancing MS/MS-based proteomics by improving spectral prediction accuracy and peptide discrimination capabilities.