prokia / drugvqa Goto Github PK

Hi I was wondering if you can clarify something.
I see in your code that you use BatchNorm2D several times. However at test time you do not switch to model.eval(). Why is that ?

Furthermore, what is the reason for using BatchNorm2D when your batch size is 1?

Hi,
Can you please share the exact values of metrics on IBM-BindingDB dataset?
Thank you

Thanks for sharing such great work! I have some questions and would appreciate your kind help:

(1) The original metrics[1] of different methods on BIndingDB dataset are fuzzy histogram. How can we get the precise values to draw new comparison as your paper?

(2) I am confused about the DUDE dataset and can not find reference about the datasets. Could you kindly explain how to get the detailed DTI records "drug(smiles) protein(amino_acid) label(0,1)" from original dataset?

Thanks for your help!

[1] Gao, K. Y., Fokoue, A., Luo, H., Iyengar, A., Dey, S. & Zhang, P. Interpretable drug target prediction using deep neural representation. In Int. Joint Conf. on Artificial Intelligence 3371–3377 (IJCAI, 2018).

Hi, first of all congrats on the paper.

For the BindingDB dataset. They provide linear(1D) representations of both protein (amino acid seq) and ligand (SMILES). How do you map 1D amino acid seq to a fully connected 2D distance matrix ? I assume some type of pBLAST against PDB but I am interested in what parameters you use for it.

Hi,

Thanks for sharing such a great work.
Is it possible to release the code for implementing the attentions visualisation as shown in the paper as well? If not, could you share me some guidelines on how to replicate that part of your work?

Thanks very much for your help.

hi, i see in your model_pkl/DUDE folder, the size of DUDE30Res-fold3-1.pkl is 1.69MB, while others are 2.19MB

According to your distance equation, the distance between neighboring residues should be 0.5, calculated by:

where both $d(r_i,r_j$ and d_0 are 3.8.
However, this distance in your data (contactMap) is 1.0.
Correct me if I am wrong!

line 90 in model.py

original : self.linear_final_step = torch.nn.Linear(self.lstm_hid_dim2+d_a,args['dense_hid'])
==>
fixed : self.linear_final_step = torch.nn.Linear(self.lstm_hid_dim2+args['d_a'],args['dense_hid'])

Hi, thank you for sharing your excellent work, may I ask for the split 5-fold human dataset please? Or could you provide the specific random seed that you use for 5 fold splitting please?

Sorry to interrupt you. I have encountered many problems when I processed the PDB data into the contactMap. For example, there are more than 20 kinds of aminos in PDB, and two CA atoms in a residue. I have tried many packages to parse the PDB, including Biopython and RDKIT, but still can not solve the problems and do not kown the right process way. I would be most grateful if you could release the code of PDB data preprocessing.
Thanks!