search

theislab / chemCPA

Code for "Predicting Cellular Responses to Novel Drug Perturbations at a Single-Cell Resolution", NeurIPS 2022.
https://arxiv.org/abs/2204.13545
MIT License
88 stars 23 forks source link

Add Chemical VAE #93

Closed siboehm closed 2 years ago

siboehm commented 2 years ago

Closes #71. This should be similar to the VAE presented in the original paper, but without jointly training the molecular property predictor. I didn't tune any hyperparameters, but the default HPs seem pretty tuned already.

I achieves pretty low KL div (should be well disentangled) and randomly sampling from the prior yields SMILES that are >95% valid, which is pretty impressive.

@MxMstrmn Can we schedule a few runs with this? I'd hope it's strictly better than the seq2seq model in every way.

review-notebook-app[bot] commented 2 years ago

Check out this pull request on  ReviewNB

See visual diffs & provide feedback on Jupyter Notebooks.

Powered by ReviewNB

MxMstrmn commented 2 years ago

For sure, I will definitely add it to the runs once #86 is closed a we have a reasoable split for it. If we include it for ther workshop, we shall see. But thanks for adding it!