No bifurcation in phlower.ext.plot_stream_sc

[kaitflynn]

Hi PHLOWER Team,

I'm using the software on a subset of cell clusters from a single-cell sequencing experiment we've run. When I get to the output from the phlower.ext.plot_stream_sc I don't see any bifurcation in my output. I'm not very familiar with this approach though and wondering if that is possible and should be interpreted as cell states along a continuum, as opposed to an error in my coding?

As background the subsetted cells included encompass fibroblasts, osteoblasts, and osteochondral progenitors from fish scales. All advice would be appreciated. Kind Regards.

Output here:

Code followed here:

Create stream tree

ph.tl.harmonic_stream_tree(adata,
trajs_use=10000,
trajs_clusters='trajs_clusters',
min_bin_number=20,
cut_threshold=1,
layout_name='cumsum',
eigen_n = -1)

Create group column

adata.obs['group'] = adata.obs['seurat_clusters'].astype(str)
adata.obs['group'] = adata.obs['group'].astype('category')

Color dictionary

celltype_colors = {
"0": "#0b871c",
"11": "#fea541",
"13": "#d50a11",
"16": "#8b3df9",
"2":"#573b0a",
"5":"#fe7fce",
"3":"#6a034d",
}

adata.uns['group_colors'] = [celltype_colors[g] for g in adata.obs['group'].cat.categories]

adata.uns['group_colors'] = celltype_colors

Plot stream tree

figs = ph.ext.plot_stream_sc(
adata,
color=['group'],
show_text=True,
show_graph=True,
show_legend=True,
fig_size=(8,5),
s=10,
return_fig=True,
dist_scale=1
)

[chengmingbo]

Hi,

Thanks for using our tools. I've reviewed your results, and the clustering output appears to be suboptimal, as it's identifying only two clusters—one of which is a group of outliers (cluster -1).

Since PHLOWER uses DBSCAN as its default method for clustering trajectory paths, I recommend adjusting its core parameter to improve the result. Specifically, you could try decreasing the eps value in the function phlower.tl.trajs_clustering to a smaller number (e.g., less than 0.5). Alternatively, if DBSCAN's density-based approach isn't ideal for your data, Leiden clustering is available as another option and may yield better results in some cases.

A few other suggestions that might help refine your analysis:

Handling Cluster Output: After clustering, you might sometimes end up with many small clusters or clusters that essentially represent noise. In such cases, you can use phlower.tl.select_trajectory_clusters to filter them out. Additionally, if you notice duplicate clusters pointing to the same terminal state, the function phlower.tl.unique_trajectory_clusters can help unique them.

Sampling More Paths: If adjusting eps doesn't yield better results, consider sampling a broader set of trajectory paths. You can do this by using phlower.tl.random_climb_knn with a larger value for knn_edges_k (e.g., >9) to explore more potential terminal states. The parameter roots_ratio (default 0.1, selecting the top 10% of nodes as roots) can also be tuned based on your actual number of root cells.

Optimizing the Layout: Finally, remember that PHLOWER relies on ddhodge for layout and pseudotime calculation. A well-optimized layout is foundational for good trajectory results. If the steps above don't fully resolve the issue, revisiting the layouting parameters might be necessary.

Feel free to try these steps and let me know if you continue to face any issues.

[kaitflynn]

Thank you for the response. I've attempted your suggestions. I saw no change by lowering the eps value in phlower.tl.trajs_clustering. I also made adjustments to knn_edges_k and roots_ratio with no luck. I then attemped the Louvain clustering method and achieved what appears to be a better output:

If you have any further suggestions I would appreciate them.

Kind Regards.

[chengmingbo]

Hi,

I see that Louvain did generate a tree, but it looks like all branches converge to the same terminal state. I think the result can still be improved.

You may want to use plot_density_grid to check the spatial coverage of each trajectory path cluster. For example:

fig, ax = phlower.pl.plot_density_grid(
    adata,
    figsize=(10, 7),
    cluster_name='trajs_clusters',
    bg_alpha=0.05,
    return_fig=True
)

This visualization can help us manually select trajectory path clusters if needed.

In addition, the function plot_fate_tree_embedding can be used to inspect the tree structure directly on the 2D embedding:

phlower.pl.plot_fate_tree_embedding(
    adata,
    node_size=5,
    with_labels=True
)

Hopefully, these steps will help you obtain a better result. Please let me know if you still run into any issues.