Large variation in rbfe results

[tomsyme]

I am experiencing large variation in hybrid topology RBFE results (OpenFE version 1.7.0). For example

ligand_i ligand_j DDG(i->j) (kcal/mol) uncertainty (kcal/mol)
17 10 -4.0 3.0

When reporting raw results, the above corresponds to

leg ligand_i ligand_j DG(i->j) (kcal/mol) MBAR uncertainty (kcal/mol)
complex 17 10 6.0 1.0
complex 17 10 6.0 1.0
complex 17 10 5.0 1.0
solvent 17 10 5.0 1.0
solvent 17 10 11.1 0.9
solvent 17 10 13.0 1.0

I re-ran the transformation and got the same issue.

ligand_i ligand_j DDG(i->j) (kcal/mol) uncertainty (kcal/mol)
17 10 0.0 4.0

leg ligand_i ligand_j DG(i->j) (kcal/mol) MBAR uncertainty (kcal/mol)
complex 17 10 11.2 0.9
complex 17 10 11.8 0.9
complex 17 10 5.4 0.9
solvent 17 10 11.9 0.8
solvent 17 10 5.0 1.0
solvent 17 10 11.0 1.0

This is using openff-2.2.1 small molecule FF.

I have run with identical inputs apart from using gaff-2.2.20, and this returned

ligand_i ligand_j DDG(i->j) (kcal/mol) uncertainty (kcal/mol)
17 10 0.4 0.4

leg ligand_i ligand_j DG(i->j) (kcal/mol) MBAR uncertainty (kcal/mol)
complex 17 10 11.0 2.0
complex 17 10 11.0 1.0
complex 17 10 11.0 1.0
solvent 17 10 11.0 1.0
solvent 17 10 11.0 1.0
solvent 17 10 10.4 1.0

The transformation is Ph-I to Ph-OPr and is not destabilising the binding pose. Experimental data show they have similar Kd. I can share more info if needed.

[IAlibay]

The large difference in solvent results is particularly interesting.

Couple of sanity checks:

• Are you assigning partial charges ahead of time & using the same ones throughout?
• Is your ligand particularly large/floppy, could it undergo large rearrangements?
• How does your mapping look like? Is there a chance you might have mismatched atoms or a mapping that allows for ring breaking or hybridisation changes?

[tomsyme]

I assign charges ahead of time, and use the same charges for both gaff and openff calculations.

Below are the molecules and the mapping. It is not very flexible apart from the propyl chain

[IAlibay]

That is really strange - given that gaff2 doesn't have the large standard deviation but openff-2 does, it looks like it might indeed be a forcefield issue.

I know it's a bit hard to extract, but have you had a look at the trajectory for the solvent simulation? Especiallly the ones that have vastly different estimates than the others. Are there odd deformations happening for some reason?