Systematic Assessment of Deep Learning-Based Predictors of Fragmentation Intensity Profiles

Area of Science:

• Proteomics
• Computational Biology
• Mass Spectrometry

Background:

• Deep learning methods are increasingly used for predicting peptide fragment intensities in mass spectrometry.
• Accurate prediction of these intensities is crucial for peptide identification and quantification.
• Several deep learning models have been developed, but a systematic comparison is needed.

Purpose of the Study:

• To comprehensively assess and compare the performance of six prominent deep learning-based peptide fragment intensity prediction methods.
• To evaluate prediction accuracy and speed across a large and diverse dataset of mass spectrometry spectra.
• To identify the most effective methods for specific experimental conditions and inform future research.

Main Methods:

• Evaluated six deep learning methods: Prosit, DeepMass:Prism, pDeep3, AlphaPeptDeep, Prosit Transformer, and Guan et al.'s method.
• Utilized a dataset of ~1.7 million precursors and >18 million experimental spectra from the PRIDE repository.
• Assessed prediction accuracy using Pearson's correlation and normalized angle for experimental b and y fragment intensities.

Main Results:

• Prosit Transformer and pDeep3 demonstrated superior prediction accuracy across various experimental conditions.
• Prediction accuracy was influenced by precursor charge, length, and collision energy.
• Performance varied among methods depending on the specific mass spectrometry data characteristics.

Conclusions:

• The study provides a systematic benchmark for deep learning-based peptide fragment intensity prediction tools.
• Prosit Transformer and pDeep3 are recommended for high-accuracy MS/MS spectra prediction.
• This comparative analysis aids researchers in selecting appropriate prediction methods for their specific proteomic studies.