Electrostatics and folding for 5-6 atom long chains

[kvkarandashev]

I was working with qm7 dataset and attempted predicting structure of '0315.xyz' with a model that included '0315.xyz' in the training set only to get linearly placed heavy atoms. The issue disappeared when I excluded '0869.xyz', '1040.xyz', and '0784.xyz' from the training set. (See check_geom_performance.py for the particular procedure for predicting geometry and check_rep_uniqueness.py for how those three particular xyz files were located.) The issue seems to be that '0315.xyz' is an alkene:

The "excluded" xyz files have at least one heteroatom in the heavy atoms chain, resulting in what looks like different degrees of folding due to electrostatic interactions between that heteroatom and hydrogens:

rep_uniqueness.zip