Recently, I found the article that explained the mechanism of graph convolutional neural network ( Questions about GraphConv layer - DeepChem and Understanding Attention in Graph Networks for Chemical Structures - Questions - Deep Graph Library (dgl.ai)).
I am also curious about how the GCN take in different molecules. If I understand correctly in that posts, the input should be something like the following: (a disconnected graph)
For a batch of molecules of 2, the adjacency matrix should look like this:
[Info of 1st mol, 0]
[0, Info of 2nd mol]
But this raises questions,
(1) If I want to take two molecules as a pair to predict a class, how do the GCN gather the info for each molecules pair and extract the features to correspond to a specific class (if the task is classification)?
Won’t the GCN turns out to output the features over the batch according to the above understanding?
(2) If I would like to pass a pair of molecule together as a disconnected graph to classify whether they bind to each other, how will the input of GCN be? Since the input dimension will vary, how to preprocess the input and how to build the neural network with the GCN?
(3) I found that most GCN models are used to predict the properties of molecules (graph). Is it possible to use GCN to solve the task of classifying whether two molecules bind? What will be the practical way to accomplish this? I only know that generating the disconnected graph for the pair to input would be helpful.
Thank you very much for your great help.