dynamo-one-shot-total-rna-velocity

name

dynamo-one-shot-total-rna-velocity

description

Run or adapt a one-shot total RNA velocity workflow in `dynamo` for metabolic-labeling or scNT-seq `AnnData`, including monocle preprocessing with an optional curated gene list, grouped moments by labeling time, Model-2 `dynamics`, `calculate_velocity_alpha_minus_gamma_s`, low-dimensional projection with `cell_velocities`, and optional streamline or phase-portrait plotting. Use when converting tutorials such as `301_tutorial_hsc_velocity.ipynb`, or when choosing between `one_shot_method` branches like `sci_fate` and `combined` and projection `method` branches like `cosine` and `pearson`.

Dynamo One-Shot Total RNA Velocity

Goal

Run the stable execution spine behind the hematopoiesis scNT-seq tutorial: preprocess a one-shot labeling dataset, smooth by labeling time, force the Model-2 no-splicing branch for kinetics, compute total RNA velocity with alpha - gamma * s, then project that custom velocity to a low-dimensional embedding without treating the tutorial dataset as the workflow identity.

Quick Workflow

1. Inspect the input AnnData for new, total, and time, plus any reusable helpers such as uns["genes_to_use"] and obsm["X_umap"].
2. Preprocess with the monocle path. If a curated gene list exists, preserve it explicitly and use Preprocessor(force_gene_list=...) plus config_monocle_recipe(..., n_top_genes=len(gene_list)).
3. Materialize grouped moments with dyn.tl.moments(..., group='time') so the one-shot fit sees time-specific neighborhoods and creates M_n, M_t, and M_s.
4. Force the no-splicing Model-2 branch by setting adata.uns["pp"]["has_splicing"] = False, then run dyn.tl.dynamics(..., group='time', one_shot_method='sci_fate', model='deterministic') unless the user explicitly wants another branch.
5. Convert the fitted kinetic parameters into total RNA velocity with dyn.tl.calculate_velocity_alpha_minus_gamma_s(...).
6. Project the custom total RNA velocity with dyn.tl.cell_velocities(..., X=adata.layers["M_t"], V=adata.layers["velocity_alpha_minus_gamma_s"], method='cosine').
7. Treat streamline_plot and phase_portraits as optional visualization steps after velocity_umap already exists.

Interface Summary

• Preprocessor.config_monocle_recipe(adata, n_top_genes=2000) configures the notebook-style monocle preprocessing path.
• Preprocessor.preprocess_adata_monocle(adata, tkey=None, experiment_type=None) is the exact notebook entrypoint; Preprocessor.preprocess_adata(..., recipe='monocle', ...) is the higher-level wrapper.
• dyn.tl.moments(..., group=None, use_gaussian_kernel=False, use_mnn=False, layers='all') creates smoothed first and second moments and stores obsp["moments_con"].
• dyn.tl.dynamics(..., group=None, model='auto', est_method='auto', one_shot_method='combined', ...) estimates one-shot kinetic parameters and writes velocity-like layers such as velocity_N and velocity_T.
• dyn.tl.calculate_velocity_alpha_minus_gamma_s(adata, gene_subset_key='use_for_pca', velocity_layer_name='velocity_alpha_minus_gamma_s') converts fitted rates plus M_n and M_s into the total RNA velocity layer used in the notebook.
• dyn.tl.cell_velocities(..., X=None, V=None, basis='umap', method='pearson', adj_key='distances', enforce=False) projects the custom velocity to low-dimensional space and stores a kernel-specific transition matrix.
• dyn.pl.streamline_plot(..., basis='umap', method='gaussian') and dyn.pl.phase_portraits(...) are optional presentation entrypoints after the analytical path is complete.

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

Branch Selection

• Use the curated-gene-list branch when uns["genes_to_use"] or a user-supplied marker/HVG list exists. Preserve that list explicitly before subsetting or concatenating because anndata.concat(...) can drop it.
• Use recipe='monocle' for this workflow family. Other recipe branches are real source-level options, but the total-RNA one-shot path here was empirically checked only on the monocle branch.
• Use group='time' on moments(...) and dynamics(...) unless the user explicitly wants to smooth and fit without time-specific neighborhoods.
• Use one_shot_method='sci_fate' for notebook parity and one_shot_method='combined' when the user explicitly wants the steady-state-plus-absolute-gamma branch. Both were empirically checked.
• Use model='deterministic' by default for this total-RNA workflow. If the user wants a splicing-aware or stochastic branch, stop and switch to a different workflow instead of silently mixing models.
• Use cell_velocities(..., method='cosine') for notebook-style total-RNA projection. Use method='pearson' when the user wants the alternate kernel with a different transition-matrix key. Treat kmc, fp, and transform as advanced branches with extra assumptions.
• Use streamline_plot(..., method='gaussian') for default streamline rendering. Switch to method='SparseVFC' only after a vector field has been reconstructed and the user explicitly wants that rendering path.

Read references/branch-selection.md before choosing non-default recipe, one_shot_method, model, method, or plotting branches.

Input Contract

• Expect a metabolic-labeling AnnData with usable new and total layers and a labeling-time column.
• Expect the time column to be accessible as adata.obs["time"] before calling calculate_velocity_alpha_minus_gamma_s(...), or rename / copy the user's time key to time.
• Treat spliced and unspliced as optional inputs for the raw object, not as the target model branch. The notebook-compatible total-RNA path forces has_splicing=False before dynamics(...).
• Expect M_n, M_t, and M_s to be created by moments(...) before calling calculate_velocity_alpha_minus_gamma_s(...).
• Expect a valid embedding such as X_umap plus a matching adjacency matrix under obsp["distances"] before calling cell_velocities(...).
• Treat tutorial-specific fields such as genes_to_use, cell_type, and the hematopoiesis sample layout as worked examples, not universal defaults.

Minimal Execution Patterns

For the notebook-compatible one-shot total RNA workflow:

import dynamo as dyn

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

selected_genes = list(adata.uns["genes_to_use"])

pre = dyn.pp.Preprocessor(force_gene_list=selected_genes)
pre.config_monocle_recipe(adata, n_top_genes=len(selected_genes))
pre.preprocess_adata_monocle(adata, tkey="time", experiment_type="one-shot")

dyn.tl.reduceDimension(adata)
dyn.tl.moments(adata, group="time")

adata.uns["pp"]["has_splicing"] = False
dyn.tl.dynamics(
    adata,
    group="time",
    one_shot_method="sci_fate",
    model="deterministic",
    cores=1,
)

dyn.tl.calculate_velocity_alpha_minus_gamma_s(adata)
dyn.tl.cell_velocities(
    adata,
    enforce=True,
    X=adata.layers["M_t"],
    V=adata.layers["velocity_alpha_minus_gamma_s"],
    method="cosine",
)

For a generic one-shot labeling dataset without a baked-in gene list:

pre = dyn.pp.Preprocessor()
pre.preprocess_adata(adata, recipe="monocle", tkey="time", experiment_type="one-shot")

For a non-default combined one-shot branch:

dyn.tl.dynamics(
    adata,
    group="time",
    one_shot_method="combined",
    model="deterministic",
    cores=1,
)

For a non-default pearson projection branch:

dyn.tl.cell_velocities(
    adata,
    enforce=True,
    X=adata.layers["M_t"],
    V=adata.layers["velocity_alpha_minus_gamma_s"],
    method="pearson",
)

Validation

After preprocessing, check these items:

• adata.uns["pp"]["experiment_type"] == "one-shot"
• adata.var["use_for_pca"] exists
• adata.obsm["X_pca"] exists

After moments(..., group='time'), check these items:

• adata.obsp["moments_con"] exists
• adata.layers["M_n"], adata.layers["M_t"], and adata.layers["M_s"] exist

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

• adata.obs["time"] exists
• adata.uns["pp"]["has_splicing"] is False for the Model-2 branch

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

• adata.uns["dynamics"]["model"] == "deterministic" for the default branch
• adata.layers["velocity_T"] exists
• adata.var["use_for_dynamics"] exists

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

• adata.layers["velocity_alpha_minus_gamma_s"] exists
• the function reports each unique time point it processed

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

• adata.obsm["velocity_umap"] exists for the default UMAP branch
• adata.obsp["cosine_transition_matrix"] exists for method='cosine'
• adata.obsp["pearson_transition_matrix"] exists for method='pearson'

Constraints

• Do not treat velocity_T from dynamics(...) as the notebook endpoint. The tutorial's total-RNA visualization path depends on the extra calculate_velocity_alpha_minus_gamma_s(...) plus cell_velocities(...) steps.
• Do not forget to rename or copy the time column to adata.obs["time"] before calculate_velocity_alpha_minus_gamma_s(...); current source hard-codes that key.
• Do not assume uns["genes_to_use"] survives AnnData concatenation or subsetting workflows. Preserve it explicitly when reconstructing subsets.
• Do not keep has_splicing=True if the goal is Model-2 total RNA velocity. Preprocessing auto-detects splicing and labeling together on the hematopoiesis raw object.
• Do not interpret one_shot_method from adata.uns["dynamics"]; current runtime does not store it there.
• Do not make small per-time groups the default. Reviewer runs with fewer than 50 cells per time group emitted NaN-velocity warnings and filtered genes before projection.

Resource Map

• Read references/branch-selection.md for decision rules across recipe, grouped smoothing, one_shot_method, and projection method branches.
• Read references/visualization-and-phase-portraits.md when the user explicitly wants notebook-style streamlines or total-RNA phase-portrait adaptation.
• Read references/source-grounding.md for inspected signatures, source-level branch evidence, and empirical execution notes.
• Read references/source-notebook-map.md to see how docs/tutorials/notebooks/301_tutorial_hsc_velocity.ipynb was converted into this reusable skill.
• Read references/compatibility.md when notebook wording and current source behavior diverge.