search

mattragoza / LiGAN

Deep generative models of 3D grids for structure-based drug discovery
GNU General Public License v2.0
221 stars 43 forks source link

I want to train a model with the data I have. #43

Closed INHYEOKSONG closed 2 years ago

INHYEOKSONG commented 2 years ago

I'm trying to learn, but there is no problem with the ligand part, but the receptor part is a problem.

Looking at the receptor structure you used, I am not using the entire pdb file, but cutting and inputting only the necessary parts from the entire pdb file.

You only need to solve this part to learn new, so I ask you a question.

And if you want to learn for the purpose of generation, please let me know a few days ago. "{class_label:d} {affinity:f} {RMSD:f} {receptor_path} {ligand_path}" I want to know if affinity and RMSD are required for learning to proceed.

mattragoza commented 2 years ago

The class label, affinity, and RMSD fields are not actually used for the default model architectures in this project. They are a holdover from using the same data file format we use for other projects in Koes Group. You can just set them to zero if you do not know what values to put in those fields in the data file.

mattragoza commented 2 years ago

Regarding your first point: It's ok to use only part of the receptor structure as long as the cutoff size is greater than the grid size, otherwise you might see weird boundary effects in the generated ligands in the regions where the receptor was split.