CellRank for multiome blood development

[mikemartinez99]

Hello @Marius1311,

I have a zebrafish multiome dataset studying developmental hematopoiesis. The dataset consists of 6 developmental timepoints (12–30 hpf) spanning blood and vascular populations. The data were integrated in Seurat using WNN, and I have used MultiVelo to infer latent times based on coupled RNA and ATAC dynamics.

My goal is to use CellRank2 to estimate fate probabilities of two early progenitor populations and determine whether they show a bias toward alternative developmental lineages.

I am currently using the RealTimeKernel, with developmental ordering informed by MultiVelo latent time. I wanted to ask whether this type of setup is compatible with CellRank assumptions, given that:

• the latent time is derived from RNA + chromatin dynamics rather than RNA velocity alone,
• the underlying neighborhood structure comes from a multimodal integration approach,
• the system is a strongly time-dependent developmental process rather than a steady-state differentiation system.

Some additional context:

• We have only one sample per developmental timepoint, so it is possible that some biological variation was smoothed during integration. However, MultiVelo latent times correlate strongly with known developmental timepoints.
• The RealTimeKernel random walk behaves as expected, with starting and ending regions matching our known cell type annotations.
• I have tested both GPCCA and CFLARE estimators, but I do not observe a clear eigengap. My assumption is that this may be expected for a rapidly progressing developmental system rather than a system with well-separated metastable states.

My main questions are:

1. Is CellRank2 appropriate for this type of developmental multiome system?
2. Is it reasonable to use a multimodal connectivity graph (e.g., Seurat WNN/WSNN) with a RealTimeKernel, or should CellRank transitions generally be based on an RNA-derived neighborhood graph?
3. In a system where developmental endpoints are biologically known but do not form strong metastable states, is manually specifying initial and terminal states an appropriate approach for fate probability estimation?
4. If macrostate identification is not appropriate due to the transient nature of the system, are there recommended alternatives within CellRank for estimating lineage biases?

Thank you!