Run scTransform from scanpy

dmckinnon4 · January 20, 2022, 8:11pm

There has been some discussion of implementing scTransform in scanpy but I cannot figure out if this actually happened. See:

github.com/theislab/scanpy

Any interest in getting SCTransform up on Scanpy?

opened 08:24PM - 10 Feb 21 UTC

atarashansky

Enhancement ✨

- [ ] Additional function parameters / changed functionality / changed defaults?… - [ ] New analysis tool: A simple analysis tool you have been using and are missing in `sc.tools`? - [ ] New plotting function: A kind of plot you would like to seein `sc.pl`? - [ ] External tools: Do you know an existing package that should go into `sc.external.*`? - [x] Other? I recently ported [SCTransform](https://github.com/ChristophH/sctransform) from R into python. Any interest in getting it onto Scanpy? The original paper is [here](https://genomebiology.biomedcentral.com/articles/10.1186/s13059-019-1874-1). It's a variance-stabilizing transformation that overcomes some key drawbacks of previous, similar methods (e.g. overfitting caused by building regression models from individual genes as opposed to groups of similar genes). It also eliminates the need for pseudocounts, log transformations, or library size normalization. My code is [here](https://github.com/atarashansky/SCTransformPy). Implementation notes (from the SCTransformPy README): - Poisson regression is done using the `statsmodels` package and parallelized with `multiprocessing`. - Improved Sheather & Jones bandwidth calculation is implemented by the `KDEpy` package. - Estimating the negative binomial dispersion factor, `theta`, using MLE was translated from the `theta.ml` function in R. - Pearson residuals are automatically clipped to be non-negative. This ensures that sparsity structure can be preserved in the data. Practically, the results do not change much when allowing for dense, negative values. Anecdotally, it produces very similar results to the R implementation, though the code itself is still a little rough around the edges. I also have to do more formal quantitative benchmarking to ensure results are similar to those of the original package. I thought I'd gauge interest here prior to working on making it `scanpy`-ready.

I cannot figure out how to get scTransform to work calling the R routine, although there is some discussion of this as well.

Does anyone know the status of running scTransform from Python/scanpy? Thanks.