May 27, 2021
The goal of Menten GCN is to create GCN tensors from protein models (poses). We are aligning with Spektral’s vocabulary style when talking about GCNs and Rosetta’s vocabulary when talking about poses.
Each node (vertex) in our graph represents a single residue position. Edges connect nodes that are close in 3D space. Our goal in Menten GCN is to analyze small pockets of residues at a time, though the size of each pocket is entirely up to the user and can encompass the entire protein if you wish.
We generate a graph by first declaring one or more “focus” residues. These residues will be at the center of our pocket. Menten GCN will automatically select the residue positions closest in space to the focus residues and will use them to build neighbor nodes. Menten GCN will also automatically add edges between any two nodes that are close in space.
We have 3 primary parameters in this system:
These parameters are used to define 3 input tensors:
One nuance of the “E” tensor is that edges can have direction. Every pair of residues has room for two edge tensors in our system. Some of our edge features are symmetric (like distance) so they will have the same value going in both directions. Other edge tensors are asymmetric (like relative geometries) so they will have different values for each of the two slots in “E”.
See the DataMaker class and examples for more details.