A-optimal himap network for shp2 contains lots of edges with low similarity

[hannahbaumann]

I have created a himap network for shp2 from the PLB (https://github.com/openforcefield/protein-ligand-benchmark/tree/new-perses-edges/data/shp2). I adapted the python script from the examples (example_optimize.py) to take in these ligands, creating a network with n*lnn edges. I used all the defaults in the command line prompts.

These are the networks I got. In the A-optimal network, most of the edges had very low weights. The three ligands 02, 04 and 07 are put into the middle of the graph in the A-optimal design (and therefore connected to many other ligands), however in the D-optimal design those three ligands are put into the "outskirts" and all connections to the main cluster are red. Looking at the ligands in 3D it makes sense (the D-optimal design), those three ligands are quite different from the rest of the ligands. So it's surprising that exactly those were put into the middle in the A-optimal design.

Have you seen designs such as these? Is this behavior expected? I ran RBFE calculations using the A-optimal design (using openfe) and the results do not look very good, though I'm not sure if results are bad because of the edges chosen or for other reasons...

Rgraph_colored_20230825_181547.pdf