conos
conos copied to clipboard
Integration of multi-modal data based on a joint space
Working with multi-modal data (e.g., 'CITE-seq + scRNA-seq' or 'TCR-seq + scRNA-seq') often produces joint dimensionality reductions (e.g., MOFA2, mvTCR or totalVI). However, most integration methods can't work with those (to my knowledge, only scVI family can). It seems like our graph approach should be easily adjustable to integrate these reductions into the graph construction process.
The way I see it, we need:
- Adjust graph construction function to use a custom existing reduction instead of computing a new joint reduction. Having just this might be enough to get decent performance.
- Make a wrapper and example for computing new joint embedding. Currently the recipe is simple: (i) for samples
X1
andX2
train a dimensionality reduction on each of them (get reduction functionsr1
andr2
), (ii) get reductionsY11 = r1(X1); Y12 = r1(X2); Y21 = r2(X1); Y22 = r2(X2)
, (iii) obtain a joint reductionY = cbind( rbind(Y11, Y21), rbind(Y12, Y22) )
. - Test it on several modalities.
This is a long shot, but I want the idea to be written somewhere :)
Update: perhaps, the easiest solution is to build two Conos objects on independent embeddings and then do WNN on Conos graphs