Anushriya Subedy and Guillaume Lamoureux. Center for Computational and Integrative Biology, Department of Chemistry, Rutgers University-Camden, Camden, NJ 08102.
Poster # 71
Many diseases (neurodegenerative, infectious, among others) have been linked to improper protein interactions. Predicting protein-protein interactions (PPIs) comprehensively requires a description (explicit or latent) of the specific biological context, including factors such as pH, ionic strength, post-translational modifications, and the availability of interaction partners. However, many of these factors are inadequately represented within the available structural data, and need to be extracted from in vivo functional data.To leverage both sources of data, we are developing SE(3)-equivariant convolutional neural networks that are pre-trained on protein multimer structures from the Protein Data Bank and refined on high-confidence AlphaFold multimer structures for which interaction data from high-throughput assays are available. The addition of interaction data enhances the learned features and better encodes for protein properties affecting PPIs in situ. We show preliminary results for the yeast proteome, using data collected from BioGrid and Yeast Kinome. We investigate the usefulness of both structural and functional data when generating meaningful representations for predicting protein properties associated with PPIs.
Comentarios