| Details: |
Protein structure prediction from sequence has dramatically improved in both accuracy and efficiency in recent years, sparking renewed interest in applying artificial intelligence (AI) approaches to structural biology. However, relying solely on AI predictions is known to be problematic.
In this talk, I will discuss a Bayesian framework developed in our group that provides a powerful way to integrate AI predictions, physics-based simulations, and experimental data, thereby advancing structural biology and drug discovery. I will introduce bAIes, a Bayesian approach that fuses the strengths of AI with molecular dynamics (MD) simulations. I will present two projects that utilise this development to produce useful protein models in different scenarios.
The first project focuses on small-molecule docking. Our framework outperforms traditional MD simulations in effectively sampling ligand-binding pocket conformations. In large virtual screening campaigns against around 50000 small molecules, bAIes shows improved discrimination between binders and non-binders compared to AlphaFold models (pROC higher by 9.4%) or models refined with molecular dynamics (pROC higher by 26.8%).
The second project extends this Bayesian framework to include cryo-electron microscopy (cryo-EM) data. Here, AI-derived structural information and MD simulations are combined with experimental density maps to refine conformations even when only coarse experimental data are available. This approach aims to recover high-resolution details from low-resolution maps by optimally including prior knowledge from AI models.
Together, these two studies underscore how Bayesian frameworks can unify AI, physics, and experiment, providing a principled route toward more accurate and efficient modeling of biomolecular structure and dynamics. |