Closed igordot closed 2 years ago
SC3 scales poorly with the number of cells in your sample and after 5k cells it is quite slow and requires significant amounts of memories. Seurat is significantly faster, but according to our benchmarks it is less accurate for small datasets. If you have a fast machine or little patience you may want to adjust the recommended threshold.
Thank you for clarifying. I also noticed poor clustering with low cell numbers, but it could be just due to the data itself. Are your benchmarks internal or is it possible to see the differences between SC3 and Seurat clustering somewhere?
Some of the benchmarks are available in the SC3 paper, see: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5410170/ Though Seurat has been updated a few times since then, but our piecemeal checking of the above on newer versions suggest the updates haven't radically improved its performance on small datasets.
In section 10.4, you have the following notes:
I haven't seen that elsewhere. Is there a specific reason for the difference?