Questions regarding residue-level prediction

[thomasgho]

Hello, thanks for the work.

I was wondering about residue-level prediction - PredictionPoint2Residue.py is using a different method to that outlined by the paper. It seems like its finding the closest three nearest neighbor point cloud points to the average residue position below a cutoff distance dt. The paper describes a more hierarchical way (first find nearest neighbors to atoms, then nearest neighbors of those to the average residue position, with no mention of cutoff).

I'm not sure the details of how the point cloud mesh is created (from PyMol as suggested in the paper), but I'm suspecting I'm getting artificially good results due to its relationship with dt. Increasing dt drastically reduces results, and not using dt is not reproducing the results outlined in the paper. I was wondering what method was used to produce the results in the paper, and what the motivation for a particular dt value was if used.

Below is a quick plot of the minimum distance between point cloud points and residue atoms (epipred dataset):