AlphaPeptDeep: a modular deep learning framework to predict peptide properties for proteomics.

Area of Science:

• Proteomics
• Computational Biology
• Machine Learning

Background:

• Deep learning (DL) models are increasingly vital in mass spectrometry (MS)-based proteomics for predicting peptide properties.
• The rapid evolution of DL architectures presents challenges for proteomics researchers seeking to integrate these tools.
• Existing DL models offer good accuracy in predicting peptide retention time, ion mobility, and fragment intensities from amino acid sequences.

Purpose of the Study:

• To introduce AlphaPeptDeep, a modular Python framework designed to simplify the application of deep learning in proteomics.
• To enable proteomics researchers, including non-specialists, to easily create and utilize predictive models for peptide properties.
• To provide a flexible platform for predicting various sequence-based peptide properties, including post-translational modifications.

Main Methods:

• Developed AlphaPeptDeep as a modular Python framework utilizing the PyTorch deep learning library.
• Implemented a 'model shop' feature allowing users to create custom predictive models with minimal coding.
• Employed transfer learning to enable model refinement with limited data for specific experimental conditions.
• Incorporated a generic representation for post-translational modifications based on chemical composition.

Main Results:

• AlphaPeptDeep models demonstrate performance on par with existing tools for predicting retention time, collisional cross sections, and fragment intensities.
• The framework successfully predicts additional sequence-based properties, as shown by a model for HLA peptide prediction.
• The modular design and model shop facilitate the integration of advanced DL techniques for proteomics researchers.

Conclusions:

• AlphaPeptDeep lowers the barrier for proteomics researchers to leverage advanced deep learning for peptide property prediction.
• The framework's flexibility and transfer learning capabilities enhance its applicability across diverse experimental settings.
• AlphaPeptDeep facilitates improved peptide identification and characterization in MS-based proteomics, including specialized applications like HLA peptide analysis.