Add prior knowledge in harmony function？

[yxwucq]

Thank you for this fantastic method which fills the blank in group-level integration. I suggest the method uses group-pratition information rather than group annotation information. Honestly, like your article mentioned that sometimes the skewed annotation may get flawed result. But may it still helpful to add the annotation information as prior?

Practically, instead of hard annotation, LLM can be used to convert annotation to embeddings and then contributes to mapping matrix with prior-weights.

Do you think it's worth to be included into the packages?

[ChuanXu1]

@yxwucq, sounds interesting. Can you point me to a paper or document that uses LLM for such purpose? THX!

[yxwucq]

Actually no one has done such things before according to my knowledge lol. I've forked the repo https://github.com/yxwucq/cellhint to implement it. I created this forked version mainly because I wanted it to be the downstream of my project: https://github.com/yxwucq/scExtract. Performance was increased but not significantly, depending on the quality of annotations. Maybe it's better to stay as fork version?