random state about harmony

[ilyakorsunsky]

Dear @wangjiawen2013,

Great suggestions!

There have been several requests for scanpy support. I have not started this; I will look into what it would take.

For reproducibility, I'm working on this feature now. In the meantime, I've found that setting the random seed in R (e.g. set.seed(1)) before running Harmony gets you 90% of the way there. This is because Harmony now initializes clusters with base R kmeans. When these initial clusters are the same, the final Harmony answer is also often consistent.

Thanks for your patience!

Best,
Ilya

[wangjiawen2013]

Thanks for you powerful tool. I integrated harmony into my scanpy pipeline in a customized manner now, however, I hope you can release a version that compatible well with scanpy.

[ilyakorsunsky]

Dear @wangjiawen2013,

In more recent tests, I've found that set.seed() works both initialization and Harmony clustering. This code demonstrates that. When you change either one of the seeds in cluster initialization or cluster updates, the result will be slightly different.

library(harmony)
data(cell_lines_small)

with(cell_lines_small, {    
    ## seed affects cluster initialization
    set.seed(1)
    res1 <<- HarmonyMatrix(
        scaled_pcs, meta_data, 'dataset', 
        do_pca=FALSE, verbose=FALSE, return_object=TRUE, max.iter.harmony=0
    )

    ## seed affects cluster initialization
    set.seed(1)
    res2 <<- HarmonyMatrix(
        scaled_pcs, meta_data, 'dataset', 
        do_pca=FALSE, verbose=FALSE, return_object=TRUE, max.iter.harmony=0
    )

})
max(abs(res1$Z_corr - res2$Z_corr))


## seed affects harmony updates in `update_R()`
set.seed(1)
.x <- harmony:::harmonize(res1, 10, FALSE)

## seed affects harmony updates in `update_R()`
set.seed(1)
.x <- harmony:::harmonize(res2, 10, FALSE)

max(abs(res1$Z_corr - res2$Z_corr))

Best,
Ilya