mims-harvard/TDC

bad performance of MLP on drug synergy prediction task

kkhuang1997 opened this issue · 2 comments

kkhuang1997 commented

Describe your question.

Hi, with the purpose to reproduce results of drug synergy prediction task, i.e., https://github.com/mims-harvard/TDC/tree/main/examples/multi_pred/drugcombo, i run through the example code. However, the results look bad in terms of unchanged loss and validation MAE.

(just show the first two rounds)
image

I noticed that another similar issue was proposed, https://github.com/mims-harvard/TDC/issues/169, and suspected bad performance of MLP on this task.

Any advice on how to choose a proper model to deal with this task?

Many thanks!

futianfan commented

thanks for your interest! Maybe you could try other neural architectures to get a more powerful representation, e.g., 1D-CNN or transformer for 1D SMILES string and graph neural network for drug molecular graph.

kexinhuang12345 commented

closing for now, feel free to reopen if question remains!