Appraisal of AlphaFold2-Predicted Models in Cryo-EM Map Interpretation

Abstract

Highly accurate predicted models from deep-learning-based protein structure prediction approaches, specifically AlphaFold2, are now readily available to aid structural biologists and computational modelers. We study the dependencies, as well as the strengths and limitations, of integrating cryo-EM maps and AlphaFold2 models into building accurate models, even from poorly resolved density maps. In our previous work [1, 2], we refined AlphaFold2 models systematically using Phenix software against experimental and “hybrid” cryo-EM maps at resolution values ranging from 4–12Å. Hybrid maps are experimental maps filtered to a lower-than-experimental resolution using a Gaussian convolution in real space; they provide a convenient variable resolution that can be tuned to find the refinement approach’s breaking point. Starting with 13 high-resolution experimental maps, we extended the resolution to values of 5Å, 6Å, 8Å, 10Å, and 12Å [1]. Our results suggested that cryo-EM density maps at resolution values as low as 6Å can be used to improve the AlphaFold2 models, which exceeded the resolution limit of 4.5Å suggested by the Phenix developers. We also explored the refinement success for AlphaFold2 models using only experimental cryo-EM maps in the 4–6Å resolution range [2]. While most of the AlphaFold2 predicted models were outright successful (TM scores above 0.9), the observed success of the subsequent refinement depended significantly on the qualities of the AlphaFold2 prediction, the experimental cryo-EM data, and the alignment of the model with the density [2]. Our experiences and observations in [1, 2] are summarized as a schematic in Fig. 1 that provides a blueprint for our ongoing work. We start the evaluation of the refinement strategy by assessing the match between the AlphaFold2 model and the cryo-EM map. In fitting atomic structures to cryo-EM maps, there is a widely used score called cross-correlation (implemented in our Situs collage or colores tools). However, it has some limitations. For example, the absolute value depends on the resolution of a map on any cropping, masking, or zero padding or on sequence insertions or deletions in the structure. It would be worth developing a more robust score like TM in an atomic structure alignment that is invariant under all these changes, so it can be used objectively to compare map/model pairs. However, often, in practice, only relative comparisons are required, for which the cross-correlation remains adequate. If the match is good and the cryo-EM map exhibits a resolution above 4.5Å, the AlphaFold2 model can be refined (left scenario in Fig. 1). For map resolutions below 6Å, the refinement performance with Phenix declines significantly (the center scenario in Fig. 1). In such a case, it might be best to leave the AlphaFold2 model alone. An intriguing alternative hypothesis is to ask if one can improve the map faithfully and sufficiently (e.g., with deep learning trained on known map-model pairs). Making low-resolution maps suitable for refinement is challenging because maps cannot typically be sharpened beyond their experimental resolution. If there is no suitable map-model match to initiate the refinement (right branch in Fig. 1), it could be due to an imperfect AlphaFold2 prediction (5 of 13 cases in [1] and 1 of 10 cases in [2] exhibited a TM score below 0.8), or the cryo-EM map might be of low quality. The actual reason for the mismatch can be determined by visual inspection of the map-model pair and by local resolution measures that assess local cryo-EM map quality. If the local map quality is low, one arrives again at the center scenario in Fig. 1 described above. A more intriguing case is the right scenario in Fig. 1, where the cryo-EM map is deemed more reliable than the AlphaFold2 prediction. Straightforward refinement of the AlphaFold2 model will then fail because of the tendency of refinement to find the nearest optimum [2]. One could overcome the local trap by generating a large number of structurally perturbed decoys from the AlphaFold2 model and then screening them against the EM data using cross-correlation or secondary structure agreement scores for ranking. The best decoy would then be used for the Phenix refinement. The structural perturbation can be implemented with a dimensionality reduction like PCA or normal mode analysis or with a reverse kinematic tracing inspired by robotics [3] to avoid the curse of dimensionality in the perturbation since an exhaustive search for more than a few degrees of freedom in the flexibility model is prohibitively expensive. Or one could use molecular dynamics simulations, but large deformations are often not exhaustively sampled in short simulation runs. Finally, one could develop an alternative flexible refinement tool that uses helical or sheet propensities to guide the flexible fitting (in addition to mere fitting to density), but this would require a more substantial development effort. We will present the