Open ashipde opened 4 years ago
When applied to real data, it doesn't seem to be working very well. It is very difficult to adjust the hyper parameters, especially the part involving Maximum Mean Discrepancy. If the parameters change even slightly, the low-dimensional representation collapses (i.e., it becomes a y=x kind of line, or in the worst case, it converges to a single point). The results are also not very consistent. It seems to require some delicate tweaking of the hyper parameters to match the real data set.
Thank you for your opinion. I am seeing similar issues in my evaluation of SAUCIE for batch-correction. Batch-corrected data was much worse than original data with all of the 9-10 lambda_b settings that I tried (batch effect and biological signal were both gone). May be this is why there are very few published studies that use SAUCIE.
I have CyTOF data from multiple batches without a batch control, and I am struggling to find a good way to perform batch correction. SAUCIE appears to be good, based on the paper published by its developers, but I don't see it used that much in other studies or mentioned in online communities.
I came across your SAUCIE_PyTorch project, and its test_saucie webpage. It seems you have played with SAUCIE a little bit. I am curious about your impression. Will you recommend it for batch correction, or for clustering?
Thanks.