search

ma-compbio / Higashi

single-cell Hi-C, scHi-C, Hi-C, 3D genome, nuclear organization, hypergraph
MIT License
76 stars 10 forks source link

Predicting structures from embedding vector #45

Closed donalday closed 11 months ago

donalday commented 11 months ago

Hi ruochiz,

Thank you for your wonderful tool. I started using the original Higashi API (not Fast-Higashi). I followed the Ramani et al.ipynb tutorial and it runs smoothly. After training, I want to generate predicted contact matrix of new samples. Is there a way to feed coordinates of embedding vector to the higashi_model object to generate the corresponding contact matrices (virtual 'raw', 'impute no nbr' and 'impute with nbr' matrices')?

Donald

ruochiz commented 11 months ago

Hi,

Thank you for your interest! Yes, I think it would be a good idea to generalize prediction to unseen single cells, and it's technically possible with Higashi. But the current implementation of it does not support that (and I don't have a plan to implement that recently).

A hacky way to do it would probably be first create some random contacts for the new samples as place holders, train the model. And then replace the contacts with actual ones... and just do the imputation. That being said, it's not tested, and I don't know if it'll function normally.

donalday commented 11 months ago

Thank you Ruochiz. That's very helpful.