oganm
commented
5 years ago I examined the code from DSA that is used here and the official DSA package and it seems the difference stems from usage of non negative matrix factorization by in the dipenps/declassification package. Based on my results, DSA implementation that is used here seems to perform better though I am still unsure if one can call it DSA as described by the original paper. The original implementation happily gives negative results for endothelial cells for the dataset in this study for instance. Even resulting in a negative correlation for them
I was fiddling with your code here and noticed that you used a third party implementation of DSA to calculate proportions rather than the functions provided by the DSA package itself. I was trying to do something similar with the DSA function from https://github.com/zhandong/DSA and it seems like the results are very different. Was wondering if you had any insight into why since you do appear to be loading the DSA package as well