A direction of pseudotime reversed: how to correct it ?

[akhst7]

I've been trying to construct a pseudotime trajectory from HSC to downstream lineages from Human BM scRNAseq data . This scRNAseq was done with the CITESeq, and annotation was done with CITESeq and transcriptome data (and initially analyzed by Seurat and available from https://www.icloud.com/iclouddrive/0fbp9OjMNaiqUjcWI0cQOGYTw#palantir%5F1), and resulting UMAP is shown below;

scv.pl.scatter(
    an,
    basis='X_sct1_umap.np',
    #components=["1, 2"],
    color="Level3M_noNA_copy",
    colorbar=True,
    figsize=(10,10),
    legend_fontsize=7,
    legend_fontoutline=2
    )

To test Palantir PsuedoTime, I selected the root/early_cell (HSC) and some of the terminal lineage cells from regions of interest on UMAP as follows;

sc.pl.embedding(
    basis='X_sct1_umap.np', 
    adata=an, 
    color=["Level3M_noNA_copy"],
    palette=palette72, size=5, 
    add_outline=False, 
    groups=['HSC', 'pDC', 'ERP', 'Pro-B', 'Myeloid intermediate 1', 'immNeu']  )

I used pal.utils.early_cell and pal.utils.find_terminal_states to find the pseudotime origin and desinations as follows;

early_cell = pal.utils.early_cell(an, celltype="HSC", celltype_column="Level3M_noNA_copy", eigvec_key="X_palantir_multiscale", fallback_seed=123)

terminal_cells=pal.utils.find_terminal_states(an, celltypes=["pDC", "ERP", "Pro-B", "Myeloid intermediate 1", "immNeu"], celltype_column="Level3M_noNA_copy", eigvec_key="X_palantir_multiscale", fallback_seed=123)

Quick check of locatioins of those cells on UMAP;

early_cell=pd.Series("Early_Cell", index="TATCTTGCAGTAGAAT-1_6")
combo_states=pd.concat([early_cell, terminal_cells])
pal.plot.highlight_cells_on_umap(an, cells=combo_states, embedding_basis="X_sct1_umap_np_small")
plt.show()

The pseudotime was generated by using these cells as follows;

pr_res = pal.core.run_palantir(data=an, early_cell=early_cell, terminal_states=terminal_states, eigvec_key="X_palantir_multiscale")

Follwoing plots shows the result and diffustion components;

pal.plot.plot_palantir_results(an, embedding_basis="X_sct1_umap_np_small", s=3)
plt.show()

pal.plot.plot_diffusion_components(an, dm_res='X_palantir_diff_comp', embedding_basis="X_umap")

Looking at these plots, immature and mature states are reversed. The problem appears to be that Palantir is picking Myeloid intermediate 1, and immNeu as the start of pseudotime, insted of HSC.

As a comparison, I run scanpy dpt on the same data, which gave me somewhat a expected result.

sc.pl.embedding(
    an,
    basis="X_sct1_umap.np",
    color=["dpt_pseudotime", "palantir_pseudotime"],
    color_map="viridis",
)

I've tried selecting different early-cells but no avail. At this point, I'd really appreciate any suggestions or inputs.

Thanks in advance.

[katosh]

Hi @akhst7, thank you for reporting this issue.

Quick fix

Try passing use_early_cell_as_start=True to run_palantir. This will ensure your specified HSC is used as the starting cell:

pal.core.run_palantir(adata, early_cell, ..., use_early_cell_as_start=True)

Why this happens

By default, Palantir doesn't use your early_cell directly. Instead, it searches for a cell on the boundary of diffusion space (technically, the axis-aligned bounding box) to ensure the starting cell represents an extreme state. Since none of your HSCs happen to lie on this boundary, Palantir picks a different cell as the starting point.

This doesn't mean your HSCs aren't a valid starting population. They may simply be extreme along a dimension not captured by the default number of diffusion components. If you'd like to explore this further, you can try increasing n_components in run_diffusion_maps (default is 10) and n_eigs in determine_multiscale_space to better resolve the state space and potentially capture a "stemness" axis.

Additional observation

Your UMAP may have been computed using a different PCA than what was used for pal.utils.run_diffusion_maps(adata, pca_key="X_pca") (or after additional filtering/batch correction). This can cause inconsistencies between the UMAP visualization and diffusion map distances, resulting in a noisier image. To avoid this, verify that the same PCA representation is used for both.

Let me know if this resolves your issue!

[akhst7]

@katosh
So I tried the quick fix you suggested but I dont think use_early_cell_as_start=True does not change anything. Any suggestions ?

pr_res = pal.core.run_palantir(data=an, early_cell="TATCTTGCAGTAGAAT-1_6", terminal_states=terminal_states, eigvec_key="X_palantir_multiscale",use_early_cell_as_start=True )

pal.plot.plot_palantir_results(an, embedding_basis="X_sct1_umap_np_small", s=3) plt.show()

'pal.plot.plot_diffusion_components(an, dm_res='X_palantir_diff_comp', embedding_basis="X_umap")'

[katosh]

It sounds like you're running into an inconsistency between the embedding used as input to your UMAP and the embedding used for the diffusion maps. This can lead to confusing results where the visual representation doesn't align with the computed pseudotime.

Looking at your plots, I notice a few things:

1. Only 5 diffusion components - This is quite small. None of them seem to align well with the "stemness" of your annotated HSCs.

2. Noisy diffusion components on the UMAP - The diffusion components appear quite noisy when projected onto the UMAP, which suggests the embedding used as input to your UMAP differs from the one used for diffusion maps. To ensure consistency, pass the same embedding to Palantir's diffusion maps via the pca_key parameter:

   pal.utils.run_diffusion_maps(
       adata,
       pca_key='X_pca_harmony',  # use the same embedding as input to your UMAP
       n_components=10,          # increase from 5
   )

   See the documentation for more details.

3. Blue dots standing out in the HSC cluster - Those few low-pseudotime cells in the HSC cluster are likely positioned at extreme locations in diffusion map space. The fallback algorithm in pal.core.run_palantir selects the starting cell from the axis-aligned bounding box of the diffusion space, which may be picking these outlier cells rather than representative HSCs.

As a quick workaround, you could manually select the HSC cell with the smallest value in component 3 (DC3), which seems to best align with stemness:

hsc_mask = adata.obs['cell_type'] == 'HSC'  # adjust to your annotation column/label
dc3_values = adata.obsm['DM_EigenVectors'][hsc_mask, 3] 
start_cell = adata.obs_names[hsc_mask][dc3_values.argmin()]

Then use this cell as your early_cell parameter and make sure to set use_early_cell_as_start=True, since this cell won't be on the bounding box.

For a more robust solution, I'd recommend:

1. Ensuring consistency between the embedding used as input to your UMAP and the diffusion maps
2. Increasing n_components in run_diffusion_maps (try 10-15 instead of 5)

This should give you diffusion components that better capture the developmental axes and lead to more sensible starting cell selection.

[akhst7]

@katosh,
So I was using the pca created by scanpy' for running the palantir diffusion map, which was a mistake. UMAP originally created by Seuratwas based on thepcacreated in R. I replace the pca with the originalSeurat pca to generate the 'palantir diffusion map (n=15) and picked early as well as terminal cell types.

Pseudotime and entropy with terminal state probabilities' as well as trajectoriesnow look what would be expected, although thetrajectoryofProB` needs to be improved. But overall it works really well now.

Thanks.