Dynamo Pseudotime Velocity

Dynamo Pseudotime Velocity

Goal

Turn pseudotime plus a valid neighbor graph into low-dimensional and gene-wise velocity with the current dynamo API, then optionally branch into vector-field reconstruction, topology / potential analysis, fate prediction, and presentation-only animation without treating the bone marrow notebook as the workflow identity.

Quick Workflow

1. Inspect the user's AnnData for a pseudotime column, a usable embedding, and whether preprocessing still needs to be run.
2. Preprocess with Preprocessor.preprocess_adata(..., recipe='monocle') unless the user explicitly wants another recipe.
3. Materialize the low-dimensional basis and neighbor graph you will actually use. Do not assume reduceDimension(...) already left obsp['distances'] behind.
4. Create a sparse-backed M_s layer, then run dyn.tl.pseudotime_velocity(...) with the right method and dynamics_info choice.
5. Add dyn.vf.VectorField(...) only when the user needs topology, potential, curl / divergence, or downstream fate prediction.
6. Treat topography, StreamFuncAnim, and animate_streamplot as optional presentation steps after the analytical outputs already exist.

Interface Summary

• Preprocessor.preprocess_adata(adata, recipe='monocle', tkey=None, experiment_type=None) is the preprocessing wrapper.
• dyn.tl.reduceDimension(..., reduction_method='umap') is the default low-dimensional embedding path, but current source also exposes trimap, diffusion_map, tsne, psl, and sude.
• dyn.tl.pseudotime_velocity(..., pseudotime='pseudotime', basis='umap', adj_key='distances', ekey='M_s', method='hodge', dynamics_info=False, unspliced_RNA=False) is the core pseudotime-to-velocity entrypoint.
• dyn.vf.VectorField(..., basis='umap', method='SparseVFC', pot_curl_div=False, **kwargs) reconstructs the low-dimensional vector field.
• dyn.pd.fate(..., direction='both', average=False, sampling='arc_length', inverse_transform=False) predicts trajectories from chosen initial cells after vector-field reconstruction.
• dyn.pl.topography(...), dyn.mv.StreamFuncAnim(...), dyn.pl.compute_velocity_on_grid(...), and dyn.pl.animate_streamplot(...) are optional notebook-style visualization / animation entrypoints.

Read source-grounding.md before documenting any narrower parameter behavior than the live source currently supports.

Branch Selection

• Use the fresh-embedding branch by default: preprocess, run reduceDimension(...), then explicitly run dyn.tl.neighbors(...) if obsp['distances'] is missing.
• Use the existing-embedding parity branch only when the user already has a trusted 2D embedding such as X_tsne and wants notebook-like plotting parity more than a fresh UMAP.
• Use method='hodge' for the normal pseudotime-to-velocity path. It is the default and was empirically checked through vector field, potential, and fate.
• Use method='gradient' when the user explicitly wants the alternate graph-gradient construction. This branch was also empirically checked and still populates the same core velocity outputs.
• Use method='naive' only when the user explicitly asks for that older branch or you are debugging differences between graph-construction strategies.
• Use dynamics_info=True when downstream tools or checks need adata.uns['dynamics']; notebook parity leaves it at the default False, but the reusable skill should prefer the more complete branch when compatibility matters.
• Use unspliced_RNA=True only when the user explicitly wants the pseudo-unspliced output stored under add_ukey such as M_u_pseudo.
• Use VectorField(..., method='SparseVFC') by default. Switch to method='dynode' only when the environment has dynode and the user explicitly wants that branch.
• Use pd.fate(..., direction='forward', average=False, inverse_transform=False) for notebook-like future-state prediction from progenitor cells. Keep sampling='arc_length' unless the user explicitly wants another path.

Read branch-selection.md before choosing non-default recipe, reduction_method, method, direction, or sampling branches.

Input Contract

• Expect an AnnData with a pseudotime column such as adata.obs['palantir_pseudotime'].
• Expect either a valid embedding already stored under the basis-specific key such as adata.obsm['X_umap'], or enough information to compute one.
• Expect a materialized adjacency matrix in .obsp under the chosen adj_key. The default is distances.
• Expect the expression layer passed as ekey to support .toarray(). If you alias from dense adata.X, wrap it with scipy.sparse.csr_matrix(...) first.
• Treat notebook-specific columns such as clusters, palantir_pseudotime, and X_tsne as worked-example inputs, not universal defaults.

Minimal Execution Patterns

For the default reusable path:

import dynamo as dyn
from scipy import sparse

adata = dyn.sample_data.bone_marrow()
adata.obs_names_make_unique()

pre = dyn.pp.Preprocessor(cell_cycle_score_enable=False)
pre.preprocess_adata(adata, recipe="monocle")

dyn.tl.reduceDimension(adata)
if "distances" not in adata.obsp:
    dyn.tl.neighbors(adata, basis="pca")

adata.layers["M_s"] = adata.X.copy() if sparse.issparse(adata.X) else sparse.csr_matrix(adata.X)

dyn.tl.pseudotime_velocity(
    adata,
    pseudotime="palantir_pseudotime",
    basis="umap",
    method="hodge",
    dynamics_info=True,
)

For notebook-style embedding reuse when an existing 2D embedding already exists:

if "X_tsne" in adata.obsm and "X_umap" not in adata.obsm:
    adata.obsm["X_umap"] = adata.obsm["X_tsne"].copy()

if "distances" not in adata.obsp:
    dyn.tl.neighbors(adata, basis="pca")

For the optional vector-field, topology, and fate branch:

dyn.vf.VectorField(adata, basis="umap", M=100, pot_curl_div=True, cores=1)

progenitor = adata.obs_names[adata.obs[group_key].isin(progenitor_labels)][:20]

dyn.pd.fate(
    adata,
    basis="umap",
    init_cells=progenitor,
    interpolation_num=30,
    direction="forward",
    inverse_transform=False,
    average=False,
    cores=1,
)

For the non-default gradient branch:

dyn.tl.pseudotime_velocity(
    adata,
    pseudotime="palantir_pseudotime",
    basis="umap",
    method="gradient",
    dynamics_info=False,
)

Validation

After preprocessing, check these items:

• adata.obsm["X_pca"] exists.
• adata.var["use_for_pca"] exists.
• adata.uns["pp"] exists.

Before pseudotime_velocity(...), check these items:

• the chosen pseudotime key exists in adata.obs
• adata.obsm["X_umap"] exists for the default basis='umap' path
• adata.obsp["distances"] exists if adj_key='distances'
• adata.layers["M_s"] exists and is sparse-backed or otherwise .toarray()-compatible

After pseudotime_velocity(...), check these items:

• adata.obsm["velocity_umap"] exists
• adata.layers["velocity_S"] exists
• adata.obsp["pseudotime_transition_matrix"] exists
• adata.var["use_for_dynamics"] and adata.var["use_for_transition"] exist
• adata.uns["dynamics"] exists only if dynamics_info=True

After VectorField(..., basis='umap', pot_curl_div=True), check these items:

• adata.uns["VecFld_umap"] exists
• adata.obs["obs_vf_angle_umap"] exists
• adata.obs["umap_ddhodge_potential"] exists
• adata.obs["curl_umap"] and adata.obs["divergence_umap"] exist

After pd.fate(..., basis='umap'), check these items:

• adata.uns["fate_umap"] exists
• adata.uns["fate_umap"]["t"] exists
• adata.uns["fate_umap"]["prediction"] exists

Constraints

• Do not assume dyn.tl.reduceDimension(...) alone leaves a usable obsp['distances'] matrix in this runtime. Check and materialize it explicitly with dyn.tl.neighbors(...) when needed.
• Do not treat adata.obsm['X_umap'] = adata.obsm['X_tsne'] as the default workflow. It is a demo-specific parity shortcut from the notebook.
• Do not set adata.layers["M_s"] = adata.X blindly when adata.X is dense. Current source calls .toarray() on ekey.
• Do not imply that M=1000 is a public VectorField signature parameter. Current source forwards it through **kwargs.
• Do not let animation setup block the analytical path. StreamFuncAnim and GIF export may require external tooling such as imagemagick.

Resource Map

• Read branch-selection.md for decision rules across recipe, reduction_method, method, and fate branches.
• Read visualization-and-animation.md when the user explicitly wants notebook-style plots, topology figures, or animations.
• Read source-grounding.md for inspected signatures, source-level branch evidence, and empirical execution notes.
• Read source-notebook-map.md to trace 201_dynamo_beyondvelo.ipynb into the reusable skill layout.
• Read compatibility.md when notebook prose and current source behavior diverge.