Bridging protein structure prediction with molecular simulations via diffusion models for missing protonation states

forthcoming

Cancer and cardiovascular diseases are the leading causes of death in the EU, resulting in 54% of deaths in 2021. There is an urgent need for more effective therapeutics for these diseases, but drug discovery is slow, taking 16 to 20 years from target identification to drug approval. To accelerate drug discovery, pharmaceutical companies use computational tools. Among the most successful are physical molecular simulations, which are limited by the availability of experimental structural data. A new generation of machine learning (ML) powered structure prediction tools, such as AlphaFold, offer the potential to supply structural data suitable for physics-based modeling without the need to experimentally solve structures. However, these tools produce 3D structures missing key physical details, which are vital for accurate molecular modeling. A critical physical detail is the assignment of relevant protein protonation states, where misprediction results in large errors in drug binding affinity predictions, slowing down drug discovery. PROTONIX will bridge this gap between physical molecular simulations and ML structure prediction tools to improve the speed and accuracy of computational drug discovery, by adding protonation details to structure predictions. I will focus on the human kinase superfamily, the main therapeutic target class for cancer and cardiovascular diseases. PROTONIX will contain two open-source ML models. First, PROTONFOLD will use a diffusion model to predict relevant protonation states with their corresponding proton positions from ML protein structure predictions, generating simulation-ready files. Second, PROTONCON will use a flow-matching model to integrate protonation state prediction with multiconformer protein structure generation, providing insights into the coupling between protonation states and protein conformations. PROTONIX will point the way for future AlphaFold-like models to produce structures immediately useful for drug discovery.