Predicting Proteolysis in Complex Proteomes Using Deep Learning

Area of Science:

• Biochemistry
• Bioinformatics
• Proteomics

Background:

• Proteolysis, mediated by proteases and reactive oxygen species (ROS), is crucial for tissue remodeling.
• Predicting protein susceptibility to ROS is feasible via amino acid composition, but protease susceptibility prediction remains challenging.
• Existing tools like PROSPER and DeepCleave have limitations in predicting endogenous protease cleavage sites.

Purpose of the Study:

• To develop bioinformatics tools for predicting protease cleavage sites and comparing protein vulnerabilities to protease- and ROS-mediated proteolysis.
• To evaluate existing protease cleavage site prediction models (PROSPER, DeepCleave) against experimental data for MMP9.
• To create a novel web application, the Manchester proteome susceptibility calculator (MPSC), integrating new models.

Main Methods:

• Experimental evaluation of PROSPER and DeepCleave using MMP9 cleavage sites in purified proteins and ECM proteome.
• Development of deep bidirectional recurrent neural network (BRNN) models for predicting cleavage sites of 14 tissue proteases.
• Integration of BRNN predictions and amino acid composition analysis (for ROS susceptibility) into the MPSC web app.

Main Results:

• BRNN models demonstrated superior performance in predicting cleavage sites within native dermal ECM proteins compared to PROSPER and DeepCleave.
• Application of MPSC to the skin proteome indicated fibrillar collagens are primarily susceptible to protease-mediated proteolysis, unlike the elastic fiber network.
• Identified novel putative targets for oxidative damage (dermatopontin, fibulins, defensins) and protease action (laminins, nidogen).

Conclusions:

• The developed BRNN models and MPSC offer improved prediction of protein susceptibility to proteolysis.
• MPSC provides a valuable tool for identifying potential proteolysis targets across various tissues and disease states.
• Findings highlight differential susceptibility of skin ECM components to protease versus ROS-mediated degradation.