search

immunogenomics / harmony

Fast, sensitive and accurate integration of single-cell data with Harmony
https://portals.broadinstitute.org/harmony/
Other
528 stars 99 forks source link

different clustering results #259

Open pikapika505 opened 2 months ago

pikapika505 commented 2 months ago

Hello,

I integrated some data using Harmony back in 2022 (on Seurat V4), but I didn’t save the Seurat object at that time. When I needed to perform further analysis, I reran Harmony on the same data on Seurat V5, but the clustering results were quite different. Specifically, the number of clusters decreased from 20 to 18, and most importantly, I can’t find cluster 17, which is crucial for our analysis.

Rplot Kim_brain_harm_clusters

Is there any way to recover the clustering results from 2022? If reproducing the previous results isn't possible, I plan to increase the resolution parameter until I see the cluster of interest and annotate the new clusters. How can I justify that the new annotation closely matches the previous one?

Can it be because of Seurat version?

Thank you for your help!

pati-ni commented 1 month ago

It could be several things. I am afraid all of the algorithms have some stochasticity, so it is unlikely you can recreate the vectors you need to reproduce the analysis.

Besides harmony, which has some random elements in the initialization, the PCA can also have some stochasticity since it uses truncated randomized SVD, and clustering also may have some stochasticity. Do you have the PC embeddings that you used for harmony somewhere?