Machine Learning for PROTAC Decomposition and Enhanced Degradation Prediction

Abstract

PROTACs (Proteolysis Targeting Chimeras) are bifunctional molecules composed of three components that mediate the degradation of target proteins, and are widely used in drug discovery. This project explores the application of machine learning in two key aspects: splitting PROTAC molecules into their three components (E3 ligase, linker, and POI), and predicting the degradation potential of PROTACs on target proteins. We evaluated an existing splitting model using internal data from AstraZeneca. Given recent updates to the public PROTAC dataset, we retrained the degradation prediction model on the expanded data. Additionally, we are transitioning the model from a binary classification task to a regression approach to directly predict degradation-related values such as DC50 and Dmax. We also investigated whether the solvent-accessible surface area (SASA) of lysine residues on the target protein influences degradation outcomes, though no clear relationship was observed.