Existing deep learning techniques for bioactivity prediction require significant prior knowledge of known active molecules for each protein target against which they predict, limiting their applicability in practice. We describe how to incorporate information about the structure of target proteins into the predictions made by deep learning neural networks. We discuss data cleaning, collation, and scaling techniques that are necessary to integrate large structural databases, such as the PDB, with large bioactivity databases, such as ChEMBL and PubChem. Finally, we present case studies where structural target information allowed the successful prediction of hits for targets with no known binders.

Deep learning in the 3rd dimension: Structure-based bioactivity prediction on novel targets

251st National Meeting (2016)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.