Plans for antibody-derived tags (CITE-seq) and other data modalities

[PeteHaitch]

We're finding kana really useful in our collaborative work and my colleagues are starting to use it instead of or to supplement interactive analyses they usually perform with 10x Genomics' Loupe Browser.
One feature that I think kana currently lacks, in comparison to the Loupe Browser, is the ability to analyse (or even visualise?) data from other modalities, such as antibody-derived tags (ADTs) used in CITE-seq.
Is this something that you are considering?

An issue I foresee is how these data are stored:

• 10x: these features are simply additional rows of the count matrix that are annotated as Antibody Capture rather than Gene Expression
• H5AD: I don't think H5AD files have the notion of an 'alternative experiment', which is where such data would typically live in a SingleCellExperiment-based workflow. I might be wrong, but it seems like they (the Python/AnnData/scanpy single-cell world) are using MuData, so that would be yet another format to support (although it at least seems to be similar to.h5ad).

Keen to hear your thoughts on this.

Cheers,
Pete
CC: @ashsea1

[LTLA]

Uh... well, no plans, really.

Right now, if you're using the 10X formats, they'll just get treated as extra genes, which may or may not be sensible. They should at least be visualize-able somewhere, you can just search for them in the marker table.

Actually making the analysis aware of ADTs would be a bit of effort. Not impossible, but it would require some thought. The cheapest approach would be to weight each feature somehow (e.g., using mumosa's neighbor-based weighting scheme) and let the PCA handle the rest. Weights could be adjusted to be gene-only, ADT-only, equivalent weights, or something in between, if users want to toy around with the effects of changing the contributions of the two modalities on the clustering/t-SNE/UMAP output.

[PeteHaitch]

From what I can tell, support has been added for the 10x and mtx readers.
Would it be feasible to add support to the h5ad reader?

[LTLA]

Hard to say. I don't believe there's a consistent field for the feature type in their row annotations, so we would have to guess. We could hope that whoever generated the H5AD would have named it feature_type, but that's not very reliable. I suppose it's no less reliable than our guessing of the gene symbols, but even that just involves checking if there's a name or sym(bol) column somewhere.

[PeteHaitch]

Yeah ... I realised that after posting. I'll ponder it over the weekend.

[jkanche]

@PeteHaitch our current release (v2.0.0) might be of interest to you. we now fully support multi-modal analysis of ADT + RNA

[PeteHaitch]

Awesome, thank you both! And just in time for a talk I'm giving today where I'm promoting kana.