search

njmarko / gnn-polypharmacy-ddi

Graph Neural Network (GNN) Transformer for predicting drug-drug interactions (ddi, polypharmacy) from drug molecules.
1 stars 0 forks source link

Application to multiple classifications #2

Open Orange2107 opened 10 months ago

Orange2107 commented 10 months ago

Hi @njmarko Have you tried applying the model to a multiclassification of 964 side effects?

njmarko commented 10 months ago

Hi @Orange2107

I haven't tried that yet because I currently don't have the hardware to perform such experiments.

It is possible to do something like that, and the authors of the Drug-Drug Adverse Effect Prediction with Graph Co-Attention paper tried it with their model on the same dataset.

I tried my model on the same dataset with their dataloader, and with some modifications, I believe it could be tweaked to work for multiclassification of side effects. You can check out how the authors of that paper managed to do it for their model in their code.

Orange2107 commented 10 months ago

Hi @njmarko I'm glad you were able to reply. I couldn't find a section on multi- classification after looking up her code. Thank you

njmarko commented 10 months ago

I did a quick search in that repo, and I also couldn't find it.

I only found cross entropy in the embedded repo.

It is mentioned in the paper though.

image image

Orange2107 commented 10 months ago

@njmarko Yeah, I followed this and turned the output of your model into 964 dimensions. I removed the neg set and used the pos set for multiclassification. I get high AUROC but very low AUPRC. I'm trying to figure that out now.