first of all let me congratulate you on the great work!
I would like to ask you about the use of the ligand hydrogens - in the paper, you say the final model does not use hydrogens for the score model - did it bring a significant improvement? And does not this improvement come just from the fact that there is less atoms to align in the RMSD computation?
Just to make myself clear on where the ligand hydrogens are lost - they are used just in the node features on the input, but the network does not predict their poses at all, right? So to obtain them from DiffDock the best I can do is to run Diffdock and then run some external protonation tool?
We realized that not using the hydrogens (as done in previous work) was giving slightly better performance and improving the runtime early on in the project, so I would not be able to tell you how much difference it makes in the final version & hyperparameters.
No, all the methods even in previous papers are evaluated with heavy-atoms RMSD therefore the presence or not of hydrogens does not affect it.
We remove the hydrogens during the preprocessing therefore they are neither used as features nor the network predicts their poses, it is however possible to get both by not using the --remove_hs flag (however this requires retraining the model).
Dear authors,
first of all let me congratulate you on the great work!
I would like to ask you about the use of the ligand hydrogens - in the paper, you say the final model does not use hydrogens for the score model - did it bring a significant improvement? And does not this improvement come just from the fact that there is less atoms to align in the RMSD computation?
Just to make myself clear on where the ligand hydrogens are lost - they are used just in the node features on the input, but the network does not predict their poses at all, right? So to obtain them from DiffDock the best I can do is to run Diffdock and then run some external protonation tool?
Thank you very much in advance for any reply!
Petr