cuby4 sample input files

[selimsami]

Hi @Honza-R ,

Thanks for this very detailed data you have shared.

If possible, could you please add or share a sample input file for cuby4 for single point calculations on PL-REX structures? I'm using the exact suggested method as in the paper (with 2.20 D3H4X corrections) but the correlations I get are somewhat lower by about 0.1.

Thanks a lot,
Selim

[georgebyron95]

Hi @Honza-R

I am pushing up this issue because I am trying as well to reproduce the SQM2.20 scores on the PL-REX dataset but I have some pretty bad numbers, like an average error of 1-2 on 004-AR instances, which I don't think can be associated with hardware differences. If you could share the Cuby4 YAML input files for the various terms (interaction energy, ligand conformational energy) and some brief comments on how to include the entropic contribution it would be great!

Thank you very much for the job and the accurate data set
Best,
Piero

[Honza-R]

We cannot share the inputs; they fall under an institutional IP protection. Reproducing the exact results is also complicated by poor reproducibility of MOPAC calculations themselves.

The exact values of the score may not be important as long as you can achieve good correlation between the score and the experimental affinities.

To provide some additional clues, I added more results to the repo (in the other_results folder): the individual components of the score and interaction energies computed with the default PM6-D3H4X/COSMO parameters.

[georgebyron95]

Thank you very much for the additional clues, they will be very useful.

I have noticed the declaration into the paper right after I sent the message here, so I understand the situation.

I had only two concerns:

1. Cuby4 doesn't implement COSMO2 gradients, so how did you run the geometry optimization on the free ligand? I tried to do a numerical gradient calculation on the original ligand.sdf file with PM6-D3H4X/COSMO2 method (using cuby4 corrections) but I hardly get convergence of the ligand file in 200 cycles. I also tried to do a geometry optimization in vacuum and then compute the energy by including COSMO2 but of course this is physically unreasonable because in some cases the free optimized has a larger energy than the ligand within the receptor when including the solvent

2. the other question was about the lm5_entropy_jk field into the ligand.sdf file, which I assume to be the entropy in J/(mol K) that must be multiplied for the temperature. This seems to be true since the additional data you provided show me the right order of magnitude.

If you can disclose some further information on point 1) it will be great, otherwise I will use the data and try to reproduce the correlation coefficient. Thank you very much for the quick response and the support! And congratulation for the very nice work :)

Best,
Piero