Fig. 1:

GNNRefine for protein model refinement. a. The flowchart includes feature extraction from the starting model, refined distance prediction using GNNs, and refined model building based on the refined distance prediction; b. the network with 10 message passing layers and 256 hidden neurons for both node and edge features. The atom embedding is concatenated with other residue features to form the node feature. The edge feature is derived for a pair of residues with Euclidean distance less than 10Å, including spatial distance, orientation and sequential separation of the two residues. The refined distance prediction is based on the final edge feature. PyMol 2.3.0 is used for structure visualization.