EXP: "meaningfulness" of CCPAs low-dimensional embeddings - Githubissues

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

EXP: "meaningfulness" of CCPAs low-dimensional embeddings #82

Closed siboehm closed 2 years ago

siboehm commented 2 years ago

So far just a collection of ideas for a section in our paper, where we try to show that the low-dimensional embeddings (after the drug embedder transformation) are useful for later downstream tasks.

Including:

low-dimensional Embeddings cluster meaningfully (#80)
Good embeddings (GROVER, seq2seq) correlate with low disentanglement loss #66
(rough idea) Good embeddings allow the training of accurate molecular property predictors (toxicity, molecular size, ...) which are useful later for optimization.