acceleration plot behaviour differences

[chansigit]

I tried several types of acceleration plots and found their behaviours puzzling.

0.heatmap

sc.pl.heatmap(adata, markers, groupby='field_louvain', swap_axes=True, layer='acceleration', dendrogram=True)

The heatmap looks fine to me

1. sc.pl.umap

sc.pl.umap(adata, layer='acceleration', color='Cndp2')

It also looks fine to me as it is consistent with the heatmap

---

2. dyn.pl.umap

dyn.pl.umap(adata, layer='acceleration', color='Cndp2', pointsize=0.05, alpha=1,frontier=False)

I thought this plot would grab the embedding from the "X_umap" but it probably grab embeddings from the "acceleration_umap"
Question 1: What is "acceleration_umap"? Can I use "X_umap" with this function?

---

3. dyn.pl.acceleration

dyn.pl.acceleration(adata, basis='umap', color='Cndp2', pointsize=0.05, alpha=1,frontier=False)

This plot looks weird to me because the colorbar ranges from 0 to 0.4 .
But the previous heatmap and plots told me the acceleration ranges from 0 to 0.0002
Question 2: Why did I got acceleration values of quite different scales from different plots?

---

4. another dyn.pl.acceleration

dyn.pl.acceleration(adata, basis='pca', color='Cndp2', pointsize=0.05, alpha=1,frontier=False)

This plot looks weird to me too because the colorbar ranges from 0 to 0.03 when I changed the basis to 'pca' .
But the previous heatmap and plots told me the acceleration ranges from 0 to 0.0002
Question 3: Does the basis parameter here changes the embedding coordinates or the actual acceleration values computed with different methods?

---

my anndata:

AnnData object with n_obs × n_vars = 4284 × 31053
    obs: 'Clusters', 'ann0608', 'phase', 'nGenes', 'nCounts', 'pMito', 'pass_basic_filter', 'unspliced_Size_Factor', 'initial_unspliced_cell_size', 'spliced_Size_Factor', 'initial_spliced_cell_size', 'Size_Factor', 'initial_cell_size', 'ntr', 'cell_cycle_phase', 'jaccard_velocity_confidence', 'umap_ddhodge_div', 'umap_ddhodge_potential', 'curl_umap', 'divergence_umap', 'control_point_umap', 'inlier_prob_umap', 'obs_vf_angle_umap', 'control_point_pca', 'inlier_prob_pca', 'obs_vf_angle_pca', 'speed_pca', 'divergence_pca', 'acceleration_pca', 'curvature_pca', 'pca_ddhodge_div', 'pca_ddhodge_potential', 'speed_umap', 'acceleration_umap', 'curvature_umap', 'field_louvain', 'field_leiden'
    var: 'Accession', 'Chromosome', 'End', 'Start', 'Strand', 'nCells', 'nCounts', 'pass_basic_filter', 'log_m', 'log_cv', 'score', 'frac', 'use_for_pca', 'ntr', 'beta', 'gamma', 'half_life', 'alpha_b', 'alpha_r2', 'gamma_b', 'gamma_r2', 'gamma_logLL', 'delta_b', 'delta_r2', 'bs', 'bf', 'uu0', 'ul0', 'su0', 'sl0', 'U0', 'S0', 'total0', 'use_for_dynamics', 'use_for_transition', 'avg_prog_confidence', 'avg_mature_confidence', 'avg_confidence', 'confident_genes'
    uns: 'ann0608_colors', 'pp', 'velocyto_SVR', 'PCs', 'explained_variance_ratio_', 'pca_mean', 'pca_fit', 'feature_selection', 'cell_phase_genes', 'dynamics', 'neighbors', 'grid_velocity_umap', 'gene_wise_confidence', 'VecFld_umap', 'grid_velocity_pca', 'VecFld_pca', 'field_louvain', 'field_louvain_colors', 'field_leiden', 'field_leiden_colors', 'rank_velocity_S', 'rank_abs_velocity_S'
    obsm: 'X_pca', 'X_umap', 'X_umap_rna', 'X', 'cell_cycle_scores', 'velocity_umap', 'velocity_umap_SparseVFC', 'X_umap_SparseVFC', 'velocity_pca', 'velocity_pca_SparseVFC', 'X_pca_SparseVFC', 'acceleration_pca', 'curvature_pca', 'acceleration_umap', 'curvature_umap'
    layers: 'ambiguous', 'matrix', 'spliced', 'unspliced', 'X_unspliced', 'X_spliced', 'M_u', 'M_uu', 'M_s', 'M_us', 'M_ss', 'velocity_S', 'acceleration', 'curvature'
    obsp: 'moments_con', 'distances', 'connectivities', 'pearson_transition_matrix', 'umap_ddhodge', 'pca_ddhodge', 'vf_feature_knn'

Thank you for your answers!

[Xiaojieqiu]

I am sorry so I cannot really answer each question closely in the near future because from now on I need to prepare my job interviews and other stuff stuff these days. My general answer is that there is a risk to mix tools together and you tend to use other tool's design principle over dynamo which is at times not reasonable.

1. the sc heatmap and umap should really use a divergent colormap (such as bwr instead of viridis) because acceleation values can be negative

2. dyn.pl.umap(adata, color=['Cndp2'], layer='acceleration', frontier=True) uses acceleration_umap (as it is shown from the running message) that is because you also calculate acceleration magnitude for the umap embedding. if you delete or change adata.obsm['acceleration_umap'], it will use the umap. I will think the best way to avoid using this later but probably the running message is already enough. Note that acceleration_umap is the acceleration magnitude for the acceleration vectors on umap space. I have explained this to you in another github issue.

3. dyn.pl.acceleration is used to plot the acceleration magnitude (not the acceleration value) for each cell

[chansigit]

Thank you so much Xiaojie. Your agile responses really impressed me a lot. Wish everything goes smoothly to you!
I now realized I messed the value and the magnitude.