scvelo neighbors

[wangjiawen2013]

Dear,
I am trying to get RNA velocity with scvelo and I find that scvelo will search neighbors by default. Now I have an scanpy anndata object, of which the neighbors are made by a customized method. I wanna keep these neighbors, are there any methods to stop scvelo re-searching neighbors automatically ?

[VolkerBergen]

Good point! Neighbors get only updated if..

• not computed yet
• or n_neighbors attribute (default: 30) in scv.pp.moments is larger than the current one given in adata.uns['neighbors']['params']['n_neighbors']

If it got recomputed in your case, it prob traces back to your n_neighbors being less than 30. Just set it to a lower value in scv.pp.moments.

[wangjiawen2013]

Thanks!
Another question, I want to harmonize scanpy with scvelo. In scanpy pipeline, there is a step called sc.pl.scale() to scale the data, which is absent in scvelo. Will it affects sc.tl.louvain and sc.tl.umap or paga when projecting the velocity graph onto an embedding ?

[VolkerBergen]

You can still use sc.tl.scale(). It will only affect the actual data matrix adata.X and thus implicitly any downstream.

The spliced and unspliced contained in adata.layers are not (and should not be) affected by sc.tl.scale()

[wangjiawen2013]

And, during data preprocessing, could I replace scvelo's scv.pp.show_proportions(adata) and scv.pp.cleanup(adata, clean='all') and scv.pp.filter_and_normalize(adata) with scanpy's sc.pp.filter_cells() and sc.pp.filter_genes() and sc.pp.normalize_per_cell() ?

[VolkerBergen]

Sure, you don't need to do any preprocessing within scvelo. Only normalization on the layers is necessary, which is automatically done within scv.pp.moments.

scv.pp.show_proportions does not do anything with your AnnData object; it just return the fractions of spliced/unspliced/ambiguous abundances, which can be useful to know.

scv.pp.cleanup(adata) is not needed at all; only if you have a very large datasets that needs to be cleaned.

scv.pp.filter_and_normalize is equivalent to running sc.pp.filter_genes and sc.pp.normalize_per_cell.

[wangjiawen2013]

please have a look at this link: velocyto-team/velocyto-notebooks#9 (comment)
Could I merge the samples' anndata objects with scanpy ? Then I can minimize the batch effect with BBKNN (https://github.com/Teichlab/bbknn), which serves as an alternative to scanpy.api.pp.neighbors().

[VolkerBergen]

Sure, that is simply adata = adata1.concatenate(adata2, adata3).

https://anndata.readthedocs.io/en/latest/anndata.AnnData.concatenate.html

[wangjiawen2013]

When combining data from multiple 10x genomics libraries, cellranger recommend equalizing the read depth between libraries before merging, to reduce the batch effect introduced by sequencing, the command is "cellranger aggr --nomalize=mapped". Will scvelo equalize the read depth per cell per sample when merging mutliple samples ?

[wangjiawen2013]

And, when should I set the parameter renormalize=True in function scv.pp.moments() ? namesly, scv.pp.moments(adata, renormalize=True). Maybe it's better to have a detailed documents because RNA velocyto is an emerging field which few people are familiar to.

[VolkerBergen]

Once the paper is out, it will shed light on yet unclear statements/attributes.

The renormalize attr just indicates whether to renormalize the first-order moments as explained here: https://scvelo.readthedocs.io/en/latest/api/docs.source.api.pp.moments.html#docs.source.api.pp.moments

It's recommended to use its default value (False). The option is only available as I tried out a couple methodological things. I might even drop this attr later.

[VolkerBergen]

If I understood you correctly (w.r.t equalizing read depth), you can just create an AnnData object for each library, normalize them individually and then concatenate. When having all in one object it does not take difference in libraries into account.

[fanxylab]

Good point! Neighbors get only updated if..

• not computed yet
• or n_neighbors attribute (default: 30) in scv.pp.moments is larger than the current one given in adata.uns['neighbors']['params']['n_neighbors']

If it got recomputed in your case, it prob traces back to your n_neighbors being less than 30. Just set it to a lower value in scv.pp.moments.

Scanpy neighbors use all matrix data (unspliced + spliced + ambiguous) as input data to get neighbor matrixes.
I confuse whether its correct If I dont use the spliced matrix to get neighbor matrixes and return 'Ms' and 'Mu' .

[WeilerP]

@fanxylab, both Scanpy and scvelo use whatever data is stored in .X, i.e., if you store all counts (spliced + unspliced + ambiguous) in .X, scvelo will not calculate the neighbor graph using spliced counts but all instead.
Given that we usually have many more spliced than unspliced reads, I do not expect major differences between the two approaches. This is also what I observed on one or two datasets where the latent space was computed one time using the original count matrix and one time only spliced counts.

[brainfo]

Hi, could we specify "neighbors_key" to where in the anndata.uns to look at by scvelo.pp.moments?