Fig. 3. Current automatic map interpretation tools work poorly with an unclear map but can be improved upon by iterative AlphaFold prediction and model rebuilding.
a, Machine-learning method for automatic map interpretation (DeepTracer22) applied to the SARS-Cov-2 structure shown in Fig. 1a. Deposited model is in brown and DeepTracer model is in blue. b, Comparison of DeepTracer model with density map. c,d, As in a and b except model building carried out with the Phenix tool map_to_model and map_to_model structure is in magenta18. The unoccupied density in b and d that does not correspond to the brown deposited model in a and c corresponds to an antibody heavy chain that is part of this structure. e, Progress of automated model building for structures shown in Fig. 1 using AlphaFold prediction iterated with model rebuilding on the basis of a density map. The resolution of the map and the PDB identifier for each structure is listed. The vertical bars show the percentage of Cα atoms in the deposited structure that are within 3 Å of any Cα atom in the corresponding model. The purple bars represents initial AlphaFold models, superimposed on the deposited structure. The salmon, gray, yellow and red bars represent the rebuilt model in cycles 1, 2, 3 and 4 of iterative AlphaFold modeling and rebuilding, respectively.