Accelerated rational PROTAC design via deep learning and molecular simulations

Proteolysis-targeting chimeras (PROTACs) have emerged as effective tools to selectively degrade disease-related proteins by using the ubiquitin-proteasome system. Developing PROTACs involves extensive tests and trials to explore the vast chemical space. To accelerate this process, we propose a novel deep generative model for the rational design of PROTACs in a low-resource setting, which is then guided to sample PROTACs with optimal pharmacokinetics through deep reinforcement learning. Applying this method to the bromodomain-containing protein 4 target protein, we generated 5,000 compounds that were further filtered through machine learning-based classifiers and physics-driven simulations. As a proof of concept, we identified, synthesized and experimentally tested six candidate bromodomain-containing protein 4-degrading PROTACs, of which three were validated by cell-based assays and western blot analysis. One lead candidate was further tested and demonstrated favourable pharmacokinetics in mice. This combination of deep learning and molecular simulations may facilitate rational PROTAC design and optimization.