In some sense, knowing all coordinates (x_i, y_i, z_i) of an N residue protein generates an N x N distance matrix. Then there are N^2 distance relations known between all residues of one chain, which is incredibly rich information.
I'm wondering how DeepRank-GNN preserves (if at all) the relations between each residue. The bioRxiv seems to suggest that only some of the Internal edges (residues on the same chain if they are within some distance cutoff) are kept. This means only a fraction of the N^2 distance relations are actually encoded as edge features. Does this affect model performance at all, and is this done so graphs fit into memory? My impression of DeepRank (CNN) is that by embedding into a 3-D grid, the spatial relations between all residues are preserved; is this not the case in DeepRank-GNN?
In some sense, knowing all coordinates
(x_i, y_i, z_i)of anNresidue protein generates anN x Ndistance matrix. Then there areN^2distance relations known between all residues of one chain, which is incredibly rich information.I'm wondering how DeepRank-GNN preserves (if at all) the relations between each residue. The bioRxiv seems to suggest that only some of the
Internaledges (residues on the same chain if they are within some distance cutoff) are kept. This means only a fraction of theN^2distance relations are actually encoded as edge features. Does this affect model performance at all, and is this done so graphs fit into memory? My impression of DeepRank (CNN) is that by embedding into a 3-D grid, the spatial relations between all residues are preserved; is this not the case in DeepRank-GNN?