A Data Set of Paired Structural Segments Between Protein Data Bank and AlphaFold DB for Medium-Resolution Cryo-EM Density Maps: A Gap in Overall Structural Quality.

Abstract

The recent publication of the AlphaFold Protein Structure Database (AlphaFold DB) offers an opportunity to revisit and refine many existing structural models in computational biology. We tested the hypothesis that models determined from less-than-ideal cryo-electron microscopy (cryo-EM) data could benefit from more recent AlphaFold predictions. Ideally, atomic structures in the Protein Data Bank (PDB) are directly solved from experimental densities with a resolution higher than 4 Å. However, medium-resolution (5-10 Å) cryo-EM maps are also increasingly deposited in the Electron Microscopy Data Bank (EMDB), many of which have associated atomic models of varying quality. It is difficult to solve atomic structures from such medium-resolution maps directly, so modeling approaches often rely on the indirect fitting of known templates. Therefore, we hypothesize that early structural interpretations of medium-resolution cryo-EM maps in the EMDB could benefit from potentially more reliable AlphaFold models derived later after more structural templates become available in the PDB. To study the utility of AlphaFold-predicted models, we conducted systematic mapping between the PDB and AlphaFold DB for structures derived from medium-resolution cryo-EM density maps. A dataset of 918 nonredundant pairs of structural segments was established. Using MolProbity, a structural validation method, we observed a significant difference in the distributions of MolProbity scores between paired structural segments in the PDB and AlphaFold DB. The structural segments in the AlphaFold DB exhibit a unimodal distribution, with an average of 0.96 MolProbity score, which is better than the average (1.98) of their corresponding segments in the PDB. The MolProbity scores of structural segments in the PDB vary significantly more than those in the AlphaFold DB and exhibit a bimodal distribution with a longer tail, indicating a wider range of model quality owing to the diverse structure fitting and refinement strategies used in the past.