cell type-aware embedding

[Nusob888]

Description of feature

Dear DRVI team,

This is a great tool and am enjoying using it a lot. I am just wondering if there are any plans to incorporate cell-type awareness into the model, similar to SCANVI and MRVI?

Many thanks

[moinfar]

Dear Bo,

We're delighted to hear that you're enjoying it.

Incorporating supervised information, such as cell types, is indeed an interesting idea. However, DRVI is fundamentally designed for unsupervised discovery in single-cell omics. One of its key strengths is its ability to remain unbiased with respect to potentially imperfect supervised annotations, allowing DRVI to uncover patterns and insights that go beyond existing annotations.

Extending the model to incorporate supervised information is a lower priority than other directions outlined in the discussions section of the paper. Users often incorporate cell-type information to improve integration performance —when the integration performance is measured by the same annotations— introducing a circular logic that we aim to avoid. We’re always open to revisiting our priorities based on compelling use cases. Do you have a reasonable example or scenario where embedding such supervised information could significantly enhance users' analyses? We'd love to hear your perspective.

[Nusob888]

Dear Amir,

Thanks for the great explanation. Agree re: the circular logic, I guess the case use I was imagining was something akin to Spectra from the Pe'er lab. Where cell type specific programs may be of interest. I guess the easiest way to get around this might be to re-train a model for a specific lineage e.g. Myeloid vs. lymphoid.

[moinfar]

Dear Bo,

Thanks for your valuable suggestion.

To acquire lineage-specific programs, one can train DRVI on the whole data and link (not a 1-1 correspondence just relevance) programs identified by DRVI to known variations such as cell types and lineages (this assignment, for example, can be done by mutual information). This approach has the potential to identify not only lineage-specific programs but also shared programs. We expect DRVI to uncover all the mentioned programs as we do not expect irrelevant programs to interfere in DRVI. Alternatively, one can follow your suggested approach and consider it as a nonlinear, batch-corrected version of Spectra.