which molecular embedding was ultimately chosen?

[bhomass]

Out of the many drug embedding methods, which was behind you published results in fig 2 and 4?

[MxMstrmn]

I am citing this passage from the "Perturbation Network" section of the paper:

The molecule encoder G can be any encoding network that maps molecular representations to fixed-size embeddings. Due to the limited number of drugs available in scRNA-seq HTSs, we propose to rely on a pretrained encoding model and freeze G during training. We tested multiple different options for G and include a detailed benchmark in the Appendix A.1. We found that RDKit features performed well in our setting and report all following results for chemCPA with RDKit as the molecule encoder G.

[bhomass]

Thank you!