Feature vector and Data distribution in Training, Validation and Development.

[Rajratnpranesh]

Thank you for open sourcing your work. I am working on an implementation of GNN on protein PDB files. What I am looking for is the feature vector of each protein generated before the input in the dense layer of GNN. I am able to save the files in a pickle file. But the issue is if I put 20 PDB data then it shows 39 training samples,3 validation samples and 4 test samples. I can't figure out why and how this is happening as split = 0.7, 0.1, 0.2.
An explanation would be very helpful. Also if you can suggest me a way to save the feature vector of each protein generated before the input in the dense layer of GNN. I need to use the feature vector for further analysis.
Thank you for your time and consideration. I will wait for your response.

[rz4]

Hi Raj, Thanks for looking through our project. Sorry for the late reply, this week has been very busy. To answer your question without going into too much detail, each PDB file can contain a number of different protein chains. These can also be chains from other protein structures or ligands found at the site during imaging. It's best to indicate which protein chains you'd like to use as to not pollute your training set with incorrect structures. Let me know if you have any other questions,

…

On Mon, Jan 27, 2020, 2:26 PM Rajratnpranesh ***@***.***> wrote: Thank you for open sourcing your work. I am working on an implementation of GNN on protein PDB files. What I am looking for is the feature vector of each protein generated before the input in the dense layer of GNN. I am able to save the files in a pickle file. But the issue is if I put 20 PDB data then it shows 39 training samples,3 validation samples and 4 test samples. I can't figure out why and how this is happening as split = 0.7, 0.1, 0.2. An explanation would be very helpful. Also if you can suggest me a way to save the feature vector of each protein generated before the input in the dense layer of GNN. I need to use the feature vector for further analysis. Thank you for your time and consideration. I will wait for your response. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1?email_source=notifications&email_token=ADSO5FL2Z3I2ZYSNV6GXOFLQ75NPDA5CNFSM4KMJELV2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4IJB5MWA>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADSO5FMDLU3XR6KA56C4N7TQ75NPDANCNFSM4KMJELVQ> .

[Rajratnpranesh]

Hi Rafael,

Thank you for your kind reply. As you have mentioned in the README file I have already specified the chain for each of the protein PDB files in the data.csv. For instance- "1a2k protein and chain C". So, The extracted protein graph is of a specific protein chain i.e. "1a2k_c". So, I have already indicated the protein chain I want to use. I am still not sure what is the reason for the more no. of training data.

Please tell me if I am missing something to understand. I need one feature for each protein in the training data.
I will wait for your reply.