Question about only train CNN_affinity for crossdock_default2018.

[RJ-Li]

Hello developers!

I saw the format of types for training both CNN_score and CNN_affinity needs rmsd and affinity label, but I don't wanna train or use CNN_score in my work, so I am searching for how to make it only for CNN_affinity.

But different papers have different types file format, such as in data/PDBBind2016/Refined_types:

# > but it is different in `data/refined` like: ``` 0 -6.3979 10gs/10gs_rec.gninatypes 10gs/10gs_ligand_0.gninatypes # 5.31559 -8.06592 0 -6.3979 10gs/10gs_rec.gninatypes 10gs/10gs_ligand_1.gninatypes # 9.14515 -8.0171 ``` I want to know how to write my own types file format if I only want to train CNN_affinity? AND what should I do when using training.py if I only train the model `--cnn crossdock_default2018`? It's new to me to use Caffe, so I don't know what kind of file should I use, and where weights_file should be assigned in. (I have all ligand's rmsd so if removing CNN_score is very difficult I can accept it.)

[dkoes]

You can adjust the model file (https://github.com/gnina/models/blob/master/acs2018/default2018.model) to only use the labels you want in the loss function.

[RJ-Li]

Hi, just ask you to make sure the built-in crossdock_default2018.caffemodel is trained on crossdock2020 dataset or PDBbind_v2019_refined-set? I only know its redocking power is tested on PDBbind_v2019_refined-set, and crossdock2020 is generated based on Pocketome.

[dkoes]

The crossdock prefix means it was trained on the crossdocked set.

[Dadiao-shuai]

Have you trained the default2018.model on PDBbind_v2020/v2019? I read the article Three-Dimensional Convolutional Neural Networks and a CrossDocked Data Set for Structure-Based Drug Design , which says you trained them on PDBbind_v2016