Dynamical model returned unexpected results and missing lots of key DEGs?

[denvercal1234GitHub]

Hi there,

Thanks for the tools.

My workflow to merge Seurat object (with computed UMAP and clustering) and loom objects in Python as below; unfortunately, the results don't look right (high unspliced counts, phase portraits look unexpected, lots of key genes were not fitted).

Q1. Does my workflow make sense? In #215 & #192, @VolkerBergen suggested instead to subset the concatenated loom object via adata = adata[list_of_barcodes].copy() and read the cluster annotations via clusters = scv.load('filename'), which can stored under adata.obs['clusters']

Q2. Of 10K+ genes, only 40+ genes fitted to dynamical? When I checked adata.var['fit_alpha'], there were 1,000+ rows but only 40 or so had some values for fit_alpha (most are quite low). Indeed, adata.var["fit_likelihood"].sort_values(ascending=False).index[:1000] would only return 40 genes with values. Many of the key genes that were identified in Seurat as DEGs for given clusters (were not hittable?) can not be plotted in scv.pl.velocity(adata_obj, var_names=['gene_of_interest']).

In #309, @WeilerP kindly advised and so I included retain_genes argument in scv.pp.filter_and_normalize. Should I also set var_names = 'all' (#858) or keep var_names='velocity_genes' in scv.tl.recover_dynamics()?

Q3. The velocity fields and the phase portraits don't look as expected. Can I gain any insights from RNA Velocity for this dataset, or I need to use CellRank?

Thank you so much for your help!

My workflow:

1. Read in each of the loom objects with scv.read() as adata
2. Rename the .obs.index of these loom adata to match the cell barcode naming of my Seurat object
3. Do .var_names_make_unique() to the loom objects
4. Do .concatenate to concatenate all the loom object adata
5. Read in my Seurat object (that was fully processed and clustered in R, and converted into h5d file - below) with sc.read_h5ad
6. Check the .obs_names for the concatenated loom match with the .obs_names of the Seurat object
7. Do scv.utils.merge to merge the Seurat object and the concatenated loom
8. Perform scv.pp.filter_and_normalize(min_shared_counts=30, n_top_genes=3000, retain_genes = clusteringFineTune_genes_list)
9. Perform scv.pp.moments(n_pcs=30, n_neighbors=30)
10. Perform scv.tl.recover_dynamics and scv.tl.velocity( mode='dynamical') with default parameters as in the tutorial

This is my Seurat object:

This is my loom object:

This is my merged object of the Seurat and concatenated loom:

Some QC plots:

Result of scv.pl.velocity_embedding_stream:

Result of scv.pl.velocity_graph default parameters don't match?

Result of scv.pl.scatter of top genes do not have almond shape (top_genes = adata.var['fit_likelihood'].sort_values(ascending=False).index)):