Grape-Pi: Graph-Based Neural Networks for Enhanced Protein Identification in Proteomics Pipelines

GRAph-neural-network using Protein-protein-interaction for Enhancing Protein Identification (Grape-Pi) is a deep learning framework for predict protein existence based on protein feature generated from Mass spectrometry (MS) instrument/analysis software and protein-protein-interaction (PPI) network.

The main idea is to promote proteins with medium evidence but are supported by protein-protein-interaction information as existent. Unlike traditional network analysis, PPI information is used with strong assumptions and restricted to specific sub-network structures (e.g. clique), Grape-Pi model is a fully data-driven model and can be much more versatile.

The contribution of Grape-Pi comes in threefold. First, we developed a dataloader module designed for loading MS protein data and protein-protein-interaction data into dataset format that can be readily used by torch-geometry. Second, we customized the graphgym module for the purpose of supervised learning in proteomics data. Third, we explored the design space and discussed caveats for training such a model for the best performance.