Speaker
Description
Previously, protein function was traditionally understood through the structure-function relationship and biological activity. However, a substantial portion of the human proteome contain intrinsically disordered proteins and regions that lack stable, globular folds. Instead these proteins function through a combination of “fly-casting” effects and transient binding structures which become ordered upon binding. Their structural heterogeneity make them difficult to characterize experimentally (CryoEM, X-Ray) or computationally (molecular dynamics). These proteins are of particular importance as they are involved in major disease pathways, including cancer and neurodegenerative diseases.
Recent advances in generative artificial intelligence have enabled protein structures and conformational ensembles to be generated at extraordinary speeds compared with conventional molecular simulations, but pharmaceutical drug discovery pipelines are hesitant to implement these techniques due to concerns surrounding their ability to reproduce unbound thermodynamic ensembles of highly flexible proteins. Building on our recently published work in PLOS Computational Biology, we benchmarked six state-of-the-art generative-AI models (AlphaFold3, Boltz-2, AlphaFlow, BioEmu, idpSAM, and IDPFold) using the disordered transactivation domain 1 of the tumor suppressor protein, p53, as a model system. The generative-AI were compared against solvent all-atom MD across four different forcefields, coarse-grained and CALVADOS simulations, temperature replica-exchange molecular dynamics (T-REMD), and multimodal experimental measurements from NMR, SAXS, circular dichroism and smFRET.
Our structural and thermodynamic analyses reveal substantial differences and challenges between the generative-AI models. General-purpose models, e.g. AlphaFold3 and Boltz-2x, strongly restrict the conformational ensemble and favor compact, highly helical states resembling experimentally determined p53-MDM2-bound states. In contrast, dynamic and IDP-oriented models, including BioEmu and idpSAM, more effectively reproduce extended conformations and broad free-energy basins observed in physical simulations. Nevertheless, reduced representation generative approaches remail less successful at recovering the atomic environments required to reproduce experimental NMR chemical shifts. Despite these initial successes, proper modeling of the protein through choice selection of forcefield and water models in an all-atom representation MD still produced the most accurate to experimental conformational ensembles.
Together, these results show that generative-AI can provide rapid exploration of the global conformational landscape of intrinsically disordered proteins, but that physics-based molecular dynamics simulations and high-performance computing remain essential for resolving their thermodynamics, local molecular structure, and therapeutically relevant interaction mechanisms. The best results were obtained from a combination of the generative-AI’s ability to freely jump across free energy barriers, while still restricted by short MD-physics based trajectories.