Protein Lifting (Pointcloud to Graph)

Proteins are represented as point clouds where each point is a residue in the protein, setting CarbonAlpha as its centers. The graph is created by connecting residues that are close to each other in the 3D space or that appear in a sequential order.

The point cloud 3D domain is lifted to a graph defining the edges between its nodes in the following way: - Sequentialwise: Connecting residues that appear in a sequential order (one after another). This approach is based on the presence of peptide bonds, which link the amino acids in a protein chain in a specific sequence. - KNN: Connecting residues that are close to each other in the 3D space. This approach is based on the physical proximity of the residues in the protein structure.

Additionally, we have also included the following features to the graph:

• Node key: One-hot encoding per each residue type.
• Pos: position of the residues in the 3D space.
• Edge features: angle (direction between CarbonAlpha (CA) and CarbonBeta (CB) atoms of each residue between different residues) and distances between the connected residues.

---

From https://github.com/pyt-team/challenge-icml-2024/pull/42