jahilton
commented
1 year ago it is unclear how the gene subset was determined
This is actually true for all CELLxGENE datasets. Our our contributing guidelines include "preference is that gene have not been filtered in order to maximize future data integration efforts" and schema states "genes SHOULD NOT be filtered from either dataset". But we do not track any provenance of the gene set.
Given that we only accept human or mouse genes, I think the ask here is that if a contributor wishes to submit mouse & human cells in 1 Dataset, then we should STRONGLY RECOMMEND that the Collection includes at least 1 Dataset with the unfiltered human gene set and 1 Dataset with the unfiltered mouse gene set in order to enable reuse of either.
BAevermann
pablo-gar
brianraymor
@Yanay1 commented on Wed May 10 2023
Description
There are a few datasets that have multiple organisms in them, and this may increase over time.
Context
Here are two examples of datasets with multiple organisms in them:
1: https://cellxgene.cziscience.com/collections/367d95c0-0eb0-4dae-8276-9407239421ee
2: https://cellxgene.cziscience.com/collections/0a77d4c0-d5d0-40f0-aa1a-5e1429bcbd7e
In the first example, the species are mixed together in the anndata and it is unclear how the gene subset was determined.
In the second example, the species are separate and downstream users can do any kind of multi-species integration they want using these anndatas, rather than having to find the raw, un-joined data.
Impact
Each anndata should have one species only. This would improve the ability to do multi-species analysis using the census.
Alternatives you've considered
When the species are joined and some kind of common name (or ortholog) gene subset is used, there is no way to reverse that. The only alternative is to go to the original paper's data source.
Ideal behavior
Each anndata should have one species only.
@pablo-gar commented on Sun May 14 2023
Hi @Yanay1
This relates to the data schema we have for the source h5ads.
What you are seeing is a direct consequence of certain level of flexibility we have in the schema. We allow data contributors to submit datasets in any form that fits their publication needs so long as:
This has lead t the following cases:
We try to stay away from datasets for which the gene space and the cells are from different species, and therefore we discourage submission of multi-species dataset. But since we don't enforce this we do have a few of these cases.
I'm tagging @brianraymor since he is the owner of our source h5ad schema @brianraymor.
Please note that Census data (not the source h5ads) does enforce that only cells that have a matching species in their genes are included, so your issue does not extend to the Census data itself.