Computing sc.pp.neighbors without method?

Help scanpy
1

Hello ! I’m new here so not sure if i’m doing this correctly but here we go. I’m building a pipeline to deal with anndata objects (for analysis, visualization etc.)

The input of my pipeline is actually the output of an encoder, of a VAE. Therefore, I already reduced the dimensions of my dataset.

Once i’ve created an anndata object with the encoder’s output I want to use sc.pp.neighbors but without applying any kind of dimensionality reduction’s step. To do so, i’m using adata.X as use_rep but I’m wondering about the parameter method = umap. Is this method going to ‘transform’ my ‘raw’ datas ?
Documentation says it’s an optional parameter but i’m not sure of what sc.pp.neighbors is doing without the method parameter.

sc.pp.neighbors(adata,
n_neighbors=15,
n_pcs=None,
use_rep= X,
knn=True,
random_state=0,
method=‘umap’,
metric=‘euclidean’,
key_added=None,
copy=False)

Also, to be able to run metrics such as Silhouette score, I needed an embedding to be store in adata.obsm. And in the same purpose of not computing any other dimensionality reduction step, I wrote this : adata.obsm[‘embeddings’] = adata.X
Does it make any sense ? Like this I wanted to make sure that everything is computed on the input datas far away from PCA / UMAP etc.

Thank you ! :slight_smile:

2

I’m actually interested in something similar so I can provide custom distances and connectivities to the anndata objects. Did you get any insight on this?

3

It doesn’t have to do anything with UMAP dimensionality reduction. Rather it is using a very efficient nearest neighbor algorithm that is coming from the UMAP package. So no worries. I would not copy the raw expression into obsm. This is rather bad practice as obsm is not meant for sparse data and you are wasting disk space.