Hello, The bioinformatics core at MDIBL is incredibly interested in using

scdownstream

for the cell type detection and annotation prediction modules. At MDIBL, we work primarily with non-mammalian animal models (zebrafish, axolotl, drosophila, etc.), and although the custom gene list parameter is available for cell cycle scoring, the cell type prediction algorithms are entirely human/mouse centric models trained on human/mouse datasets and often the orthology mapping from human to say, axolotl, is not complete. My question is, are there any people in the community who are working on scaling this pipeline’s capabilities for more species models to be trained or included? Or is this a task that has already been solved? If not, I would like to work with the community to add native organism annotation prediction modules for non-mammalian animal models using several publicly available single cell expression atlases as training reference datasets. I am a complete Nextflow beginner and would like to work together with anyone who may be interested in this project on developing such modules.

Hey, as the maintainer of this pipeline I can quite confidently say that no one is working on making the cell type annotation work for more 'exotic' species. From what I know, most people use it for human and mouse, that's why the development focus has been these species so far. However, I am quite motivated to collaborate with you to make the pipeline fit your needs. Maybe we can set up a meeting to clarify what you need and how we can efficiently collaborate?

Hey! I have been workiing with cross-species analyeses for the last two years now. I think the orthogene packadge I have been using will be a pretty good one for this use case as it supports 700+ species including the aforementioned axolotl. It also has a very comprehensive suite of mapping strategies you can use for handling thigns like many to one etc mappings. I have some scripts I would be willing to share for adapation into a module for the pipeline :)

As I don't have much experience with cell type annotations from 'exotix' species, is it really sufficient to map the genes to e.g. human orthologs and then use human cell type annotation models for annotating? I can imagine that the biological programs, markers, etc. have diverged to a degree that could lead annotation tools astray. Has this been investigated before? The alternative of having dedicated cell type annotation models per species however is surely also not very tractable

Very true, this is quite a tough problem

As MDIBL brought this up, I am sure they will already have a bit of an idea on how this can be handled. Curious to hear that @mcox

Hi all, as a member of the CGDS core at MDIBL, I can confirm this is one of our biggest challenges. Hallmark marker genes for a given cell type in human and mouse don't always translate cleanly to organisms like axolotl. On top of that, axolotl genome annotation is still fairly naive, and gene nomenclature frequently doesn't align with mouse/human conventions. Our current approach leans heavily on the knowledge from our faculty to manually label cell types, using marker genes generated by our in-house scRNA-seq downstream Nextflow pipeline. Longer-term, I see two main paths forward: (1) orthology-based label transfer from the large existing human/mouse reference databases, which comes with its own nontrivial mapping challenges, or (2) building organism-specific cell atlas dictionaries by aggregating as much annotated single-cell data as possible, which is powerful but compute and labor-intensive.