Training Configuration of pre-trained MPNN_CNN

[pykao]

Hi Kexin Huang,

I am using the provided pre-trained MPNN_CNN model. When I looked into its model configuration file, it looks wired to me.

{'input_dim_drug': 1024,
'input_dim_protein': 8420,
'hidden_dim_drug': 128,
'hidden_dim_protein': 256,
'cls_hidden_dims': [1024, 1024, 512],
'batch_size': 16,
'train_epoch': 1,
'LR': 0.001,
'drug_encoding': 'MPNN',
'target_encoding': 'CNN',
'result_folder': './result/',
'binary': False,
'mpnn_hidden_size': 128,
'mpnn_depth': 3,
'cnn_target_filters': [32, 64, 96],
'cnn_target_kernels': [4, 8, 12],
'num_workers': 0,
'decay': 0}

Did you only train this model for only 1 epoch with batch size 16?

Best regards,
Po-Yu Kao

[kexinhuang12345]

That's weird, I must have stored the wrong model. Let me double-check and I will upload the correct model.

[kexinhuang12345]

Hey it seems this model is wrong. You can use "MPNN_CNN_BindingDB_IC50" instead. It is trained on a much larger training set (~10^5 -> 10^6) and should have higher quality. Do note that the units now switches from Kd to IC50.

[pykao]

Did you use the latest BindingDB to train this model?

[kexinhuang12345]

Hey, it is using the past version 2020m2. There should be some minor difference with the current most up to date version regarding the number of training points.

[pykao]

Thank you for your reply 👍🏽 Please let me know if you want the trained MPNN_CNN on BindingDB using Kd.

[kexinhuang12345]

No problem! Did you mean you are managed to train the model? If so, would be great to share with me ([email protected]), thanks!

[kexinhuang12345]

You can simply use the model.save('XXX') function and then send me the model file; i will upload to the server and update the link, thanks again!

[chemlove]

Hi Kexin,
It seems that the pre-trained model MPNN_CNN downloaded using pretrained_dir = download_pretrained_model('pretrained_models') in the oneliner.py still showing the old configuration:

{'input_dim_drug': 1024,
'input_dim_protein': 8420,
'hidden_dim_drug': 128,
'hidden_dim_protein': 256,
'cls_hidden_dims': [1024, 1024, 512],
'batch_size': 16,
'train_epoch': 1,
'LR': 0.001,
'drug_encoding': 'MPNN',
'target_encoding': 'CNN',
'result_folder': './result/',
'binary': False,
'mpnn_hidden_size': 128,
'mpnn_depth': 3,
'cnn_target_filters': [32, 64, 96],
'cnn_target_kernels': [4, 8, 12]}

Maybe you need to update the model file on the https://dataverse.harvard.edu/api/access/datafile/

Maybe the configure files corresponding to pretrained_dir = download_pretrained_model('models_configs') also need a update.

[kexinhuang12345]

Sounds good, do you want to contribute and train a new model for it?

[chemlove]

I'd like to have a try. Could you please give me the dataset of BindDB Kd? And what preproccess or data cleaning is needed before I start the train?
Subsequent help may be needed since I am a complete newbie for ML :)

[kexinhuang12345]

Sounds good, it should be the one in the https://github.com/kexinhuang12345/DeepPurpose/blob/master/DEMO/Transformer%2BCNN_BindingDB.ipynb

simply replacing the model and parameter should be good

[Jameel9]

Thank you for your fruitful discussion and big thank-you to the developers of this library. My question is:
In the latest release od DeepPurpose, was the MPNN_CNN model corrected and it works fine now?