Clarification on Pocket2Mol model training with CrossDocked2020 dataset

[mahmoud-ekhani]

Hello, I appreciate your excellent work. I am also attempting to train this model on a larger dataset and have started with the CrossDocked2020-v1.3 dataset, which has an RMSD < 2A. This dataset already includes clustered training and test data distributions, so I'm unsure about the need for mmseqs2 clustering. Could you please explain the reason for this step?

Additionally, the CrossDocked2020 dataset has various docking forms, such as Autodock Vina docked poses of ligands in the receptor and the first and second iterations of CNN-optimized poses. Which of these did you use in your training process?

Lastly, I noticed that the .PDB files in your training dataset are smaller than those in the CrossDocked2020 dataset. Did you perform any extra processing steps to obtain these smaller receptor files?

Thank you in advance for your time and assistance!

[daiheng-zhang]

I also noticed on this. I believe the Autodock Vina data was not utilized, and I also found that the training datasets are smaller than those in the CrossDocked2020 dataset. How did you later handle this dataset?