feature loadings

[rimmoussa]

Hi,

is it possible to obtain the feature loadings from the PHATE output? I am looking for a way to identify the "important" features in terms of their contribution to the component. the returned object only contains an embedding slot and it would be really helpful to have something similar to feature.loadings which are by default in the seurat object after dimensionality reduction

[kmoon3]

Hi,

There isn't a straightforward way to do this with the current PHATE code. Since PHATE is a nonlinear method, it wouldn't be accurate to obtain loadings via a linear projection. Also, Nystrom-based methods for out of sample extension can be difficult to train. However, we do have a separate method called GRAE that can learn a parametric version of PHATE via an autoencoder. You could then use the decoder to go from the embedding to the original features or you could apply some measure of feature importance on the encoder. You can find the code here: https://github.com/KevinMoonLab/GRAE