Refinement of AlphaFold2 Models against Experimental Cryo-EM Density Maps at 4-6Å Resolution

This work provides new evidence of the utility of deep learning-based protein structure prediction approaches, specifically AlphaFold2, in the interpretation of 4-6 Å resolution cryo-EM maps. We describe the dependencies, as well as the strengths and limitations, of integrating experimental and AI-based approaches to building accurate models, even from poorly resolved density maps. The test followed recent work that implemented a refinement protocol in the Phenix program, which successfully refined AlphaFold2 models in high-resolution maps but which at lower resolution relied on simulated "hybrid density maps". To study the noise and imperfections present in experimental cryo-EM maps more realistically, in this work, we selected only experimental map/model pairs in the 4-6 Å resolution range where refinement performance starts to degrade. Most of the AlphaFold2 predicted models are highly accurate, particularly for the 9 larger chains (226-373 residues long) of the 10 cases, exhibiting TM-scores above 0.9. A small chain of 115 residues in length containing three helices was poorly predicted, with a TM-score of 0.52. The observed success of the subsequent refinement step depends significantly on the quality of the AlphaFold2 prediction, the quality of the experimental cryo-EM data, and the quality of the alignment of the model with the density.