Exploration of hormonal receptor genes in Castaldi et al organoid dataset - Metacell calculation by SeaCells

Reference paper

Upstream Steps

• Assemble adata
• QC filter on cells; Expression filter on genes
• Normalization and log10 transformation by Scanpy functions
• Feature selection (HVG) by Scanpy functions
• Dimensionality reduction
• Cluster identification

This Notebook

Metacell calculation by SEACells (following this tutorial)

---

1. Environment

1.1 Modules

import os
import sys

import numpy as np
import pandas as pd
import scanpy as sc

import pickle

#Plotting
import matplotlib.pyplot as plt
import matplotlib
import seaborn as sns

#utils
import ipynbname
from datetime import datetime

# SeaCell
import SEACells

#import custom functions
sys.path.append('../')
import functions as fn

findfont: Font family ['Raleway'] not found. Falling back to DejaVu Sans.
findfont: Font family ['Lato'] not found. Falling back to DejaVu Sans.

# Some plotting aesthetics
%matplotlib inline

sns.set_style('ticks')
matplotlib.rcParams['figure.figsize'] = [3.5, 3.5]
matplotlib.rcParams['figure.dpi'] = 100

print("Scanpy version: ", sc.__version__)
print("Pandas version: ", pd.__version__)
print("SEACell version: ", SEACells.__version__)

Scanpy version:  1.9.3
Pandas version:  2.0.0
SEACell version:  0.3.3

1.2 Settings

sc.settings.verbosity = 3
sc.settings.set_figure_params(dpi=80)

1.3 Files and parameters

input_file = '../../../../DataDir/ExternalData/SingleCellData/Castaldi_adata.h5ad'

1.4 Start computations

print(datetime.now())

2026-03-05 18:11:47.822083

---

2. Data Load

2.1 Read adata file

• SEACells uses as input filtered unnormalized counts from scRNASeq data, that are subjected to normalization and log-transformation before metacell assignment
• Here we start from data that has already undergone the steps of QC, filtering, normalization, log transformation
• The sum of expression levels from cells to metacells will occur at the count level

adata = sc.read(input_file)
adata

AnnData object with n_obs × n_vars = 14913 × 33538
    obs: 'dataset', 'cellID', 'cellID_newName', 'n_genes_by_counts', 'log1p_n_genes_by_counts', 'total_counts', 'log1p_total_counts', 'total_counts_mt', 'log1p_total_counts_mt', 'pct_counts_mt', 'total_counts_ribo', 'log1p_total_counts_ribo', 'pct_counts_ribo', 'stage', 'type', 'id_stage', 'cellID_newName_type', 'S_score', 'G2M_score', 'phase', 'leidenAnnotated', 'leiden_1.2', 'endpoint_GlutamatergicNeurons_late', 'endpoint_GlutamatergicNeurons_early', 'endpoint_MigratingNeurons', 'endpoint_OuterRadialGliaAstrocytes', 'endpoint_Interneurons', 'endpoint_Interneurons_GAD2', 'endpoint_CajalR_like', 'Exc_Lineage', 'endpoint_GlutamatergicNeurons_both'
    var: 'highly_variable', 'means', 'dispersions', 'dispersions_norm'
    uns: 'cellID_colors', 'cellID_newName_colors', 'cellID_newName_type_colors', 'dataset_colors', 'draw_graph', 'harmony', 'hvg', 'leidenAnnotated_colors', 'log1p', 'pca', 'stage_colors', 'type_colors', 'umap'
    obsm: 'X_draw_graph_fa_harmony', 'X_pca', 'X_pca_harmony', 'X_umap_harmony', 'X_umap_nocorr'
    varm: 'PCs'
    layers: 'counts', 'lognorm'
    obsp: 'harmony_connectivities', 'harmony_distances', 'pca_connectivities', 'pca_distances'

print(adata.X[40:45, 40:45])

  (3, 3)	0.9110727

sc.pl.embedding(adata,  basis="X_umap_harmony", color=['leidenAnnotated', 'dataset'], ncols=1)

/usr/local/lib/python3.8/dist-packages/scanpy/plotting/_tools/scatterplots.py:392: UserWarning: No data for colormapping provided via 'c'. Parameters 'cmap' will be ignored
  cax = scatter(
/usr/local/lib/python3.8/dist-packages/scanpy/plotting/_tools/scatterplots.py:392: UserWarning: No data for colormapping provided via 'c'. Parameters 'cmap' will be ignored
  cax = scatter(

sc.pl.embedding(adata,  basis="X_draw_graph_fa_harmony", color=['leidenAnnotated', 'dataset'], ncols=1)

/usr/local/lib/python3.8/dist-packages/scanpy/plotting/_tools/scatterplots.py:392: UserWarning: No data for colormapping provided via 'c'. Parameters 'cmap' will be ignored
  cax = scatter(
/usr/local/lib/python3.8/dist-packages/scanpy/plotting/_tools/scatterplots.py:392: UserWarning: No data for colormapping provided via 'c'. Parameters 'cmap' will be ignored
  cax = scatter(

adata.obs.head(3)

	dataset	cellID	cellID_newName	n_genes_by_counts	log1p_n_genes_by_counts	total_counts	log1p_total_counts	total_counts_mt	log1p_total_counts_mt	pct_counts_mt	...	leiden_1.2	endpoint_GlutamatergicNeurons_late	endpoint_GlutamatergicNeurons_early	endpoint_MigratingNeurons	endpoint_OuterRadialGliaAstrocytes	endpoint_Interneurons	endpoint_Interneurons_GAD2	endpoint_CajalR_like	Exc_Lineage	endpoint_GlutamatergicNeurons_both
AAACCTGAGAGACTAT-1_DownD250	DownD250	MIFF1	CTL02A	2586	7.858254	6358.0	8.757627	175.0	5.170484	2.752438	...	0	1	0	0	0	0	0	0	GlutamatergicNeurons_late	1
AAACCTGCATGGTTGT-1_DownD250	DownD250	KOLF	CTL08A	1372	7.224753	2540.0	7.840313	73.0	4.304065	2.874016	...	0	1	0	0	0	0	0	0	GlutamatergicNeurons_late	1
AAACCTGTCAGTTAGC-1_DownD250	DownD250	3391B	CTL01	1216	7.104144	2859.0	7.958577	43.0	3.784190	1.504022	...	6	0	0	0	0	1	0	0	Other	0

3 rows × 31 columns

2.2 Pre-processing

Normalization, log-transformation, HVG and dimensionality reduction have been already performed.

3. Run SEACells

We follow the workflow described in the SEACells tutorial

3.1 Define parameters

Parameters to be defined:

• Number of metacells (graining level): choosing one metacell for every 60 single-cells.
• X_pca as the key in .obsm for computing in metacells (to be changed to X_svd in case od ATAC data)
• Number of eigenvalues for metacell initialization

We then employ the SEACells.core.SEACells function to initialize the model.

## Core parameters 
n_SEACells = round(adata.n_obs/60)
print(n_SEACells)

build_kernel_on = 'X_pca_harmony' # key in ad.obsm to use for computing metacells
                          # This would be replaced by 'X_svd' for ATAC data

## Additional parameters
n_waypoint_eigs = 10 # Number of eigenvalues to consider when initializing metacells

249

#help(SEACells.core.SEACells)

model = SEACells.core.SEACells(adata, 
                  build_kernel_on=build_kernel_on, 
                  n_SEACells=n_SEACells, 
                  n_waypoint_eigs=n_waypoint_eigs,
                  convergence_epsilon = 1e-5)

Welcome to SEACells!

3.2 Construct kernel matrix

construct_kernel_matrix function constructs the kernel matrix from data matrix using PCA/SVD and nearest neighbors. Key parameters are:

• n_neighbors: (int) number of nearest neighbors to use for graph construction (default 15). Increasing this number results in a more homogeneous distribution of metacell size, but may impact on the preservation of rare cell types. Usually stable performance in the range 5-30.
• graph_construction: method for graph construction.

model.construct_kernel_matrix(n_neighbors=20)
M = model.kernel_matrix

Computing kNN graph using scanpy NN ...
computing neighbors

2026-03-05 18:18:07.034962: I tensorflow/core/util/port.cc:110] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
2026-03-05 18:18:07.077195: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2026-03-05 18:18:08.073650: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT

    finished: added to `.uns['neighbors']`
    `.obsp['distances']`, distances for each pair of neighbors
    `.obsp['connectivities']`, weighted adjacency matrix (0:00:25)
Computing radius for adaptive bandwidth kernel...

  0%|          | 0/14913 [00:00<?, ?it/s]

Making graph symmetric...
Parameter graph_construction = union being used to build KNN graph...
Computing RBF kernel...

  0%|          | 0/14913 [00:00<?, ?it/s]

Building similarity LIL matrix...

  0%|          | 0/14913 [00:00<?, ?it/s]

Constructing CSR matrix...

3.3 Initialize archetypes

• initialize_archetypes function initializes the B matrix which defines cells as SEACells.
• The function uses waypoint analysis for initialization into to fully cover the phenotype space, and then greedily selects the remaining cells.

# Initialize archetypes
model.initialize_archetypes()

Building kernel on X_pca_harmony
Computing diffusion components from X_pca_harmony for waypoint initialization ... 
Determing nearest neighbor graph...
computing neighbors
    finished: added to `.uns['neighbors']`
    `.obsp['distances']`, distances for each pair of neighbors
    `.obsp['connectivities']`, weighted adjacency matrix (0:00:04)
Done.
Sampling waypoints ...
Done.
Selecting 238 cells from waypoint initialization.
Initializing residual matrix using greedy column selection
Initializing f and g...

100%|██████████| 21/21 [00:00<00:00, 32.11it/s]

Selecting 11 cells from greedy initialization.

adata.obsm['X_umap'] = adata.obsm['X_umap_harmony'].copy()

# Plot the initilization to ensure they are spread across phenotypic space
SEACells.plot.plot_initialization(adata, model)

3.4 Fit Model

We apply the model.fit function.

model.fit(min_iter=10, max_iter=100)
model

Randomly initialized A matrix.
Setting convergence threshold at 0.00239
Starting iteration 1.
Completed iteration 1.
Starting iteration 10.
Completed iteration 10.
Starting iteration 20.
Completed iteration 20.
Starting iteration 30.
Completed iteration 30.
Converged after 35 iterations.

<SEACells.core.SEACells at 0x15544874de50>

4. Access SEACells results

4.1 Check model convergence

# Plot model convergence
model.plot_convergence()

4.2 Inspect SEACells soft assignment

• SEACells analysis returns soft assignments of cells to SEACells (full assignment matrix can be accessed at model.A_)
• The majority of single-cells are assigned to no more than 4 archetypes with non-trivial weight
• The algorithm returns the top 5 metacell assignments as well as the corresponding assignment weights in the function model.get_soft_assignments()

#model.A_

plt.figure(figsize=(4,3))
sns.distplot((model.A_.T > 0.1).sum(axis=1), kde=False)
plt.title(f'Non-trivial (> 0.1) assignments per cell')
plt.xlabel('# Non-trivial SEACell Assignments')
plt.ylabel('# Cells')
plt.show()

plt.figure(figsize=(4,3))
b = np.partition(model.A_.T, -5)    
sns.heatmap(np.sort(b[:,-5:])[:, ::-1], cmap='viridis', vmin=0)
plt.title('Strength of top 5 strongest assignments')
plt.xlabel('$n^{th}$ strongest assignment')
plt.show()

/tmp/ipykernel_2843413/2880491417.py:4: UserWarning: 

`distplot` is a deprecated function and will be removed in seaborn v0.14.0.

Please adapt your code to use either `displot` (a figure-level function with
similar flexibility) or `histplot` (an axes-level function for histograms).

For a guide to updating your code to use the new functions, please see
https://gist.github.com/mwaskom/de44147ed2974457ad6372750bbe5751

  sns.distplot((model.A_.T > 0.1).sum(axis=1), kde=False)

labels,weights = model.get_soft_assignments()
#labels.head()

4.3 Inspect and retrieve SEACells hard assignment

We will use hard assignment of each cell to a single metacell to define our metacells for downstream steps. Hard assingment can be retreived in two ways:

• in the modified anndata object in .obs['SEACell']
• from the model using .get_hard_assignments()

adata.obs[['SEACell']].head()

# Alternatively: 
# model.get_hard_assignments().head()

	SEACell
index
AAACCTGAGAGACTAT-1_DownD250	SEACell-186
AAACCTGCATGGTTGT-1_DownD250	SEACell-28
AAACCTGTCAGTTAGC-1_DownD250	SEACell-109
AAACGGGCAGCTATTG-1_DownD250	SEACell-203
AAACGGGGTGCACGAA-1_DownD250	SEACell-121

5. Summarizing data by hard assignment

• The original dataset can be summarized by SEACell by aggregating cells within each SEACell, summing over all raw data for all cells belonging to a SEACell.
• The output of this function is an anndata object of shape n_metacells x original_data_dimension.
• Data is unnormalized and raw aggregated counts are stored in X.
• Attributes associated with variables (.var) are copied over, but relevant per SEACell attributes must be manually copied, since certain attributes may need to be summed, or averaged etc, depending on the attribute.

# By default, ad.raw is used for summarization. Other layers present in the anndata can be specified using the parameter summarize_layer parameter
SEA_adata = SEACells.core.summarize_by_SEACell(adata, SEACells_label='SEACell', summarize_layer='counts', celltype_label='leidenAnnotated')
SEA_adata

100%|██████████| 249/249 [00:02<00:00, 95.46it/s]

AnnData object with n_obs × n_vars = 249 × 33538
    obs: 'leidenAnnotated', 'leidenAnnotated_purity'
    layers: 'raw'

SEA_adata.obs.head()

	leidenAnnotated	leidenAnnotated_purity
SEACell-186	GlutamatergicNeurons_late	0.934066
SEACell-28	GlutamatergicNeurons_late	0.864865
SEACell-109	Interneurons	0.795181
SEACell-203	MigratingNeurons	0.818182
SEACell-121	RadialGliaProgenitors	0.980392

print(SEA_adata.X[40:43, 40:43])

  (0, 0)	6.0
  (1, 0)	6.0
  (2, 0)	11.0

7. Visualize results

• plot_2D(): visualize metacell assignments on UMAP or any 2-dimensional embedding froma adata.obsm. Plots can also be coloured by metacell assignment. The visualization allows to assess the representativeness of the calculated metacells: ability to represent the global structure of the original dataset. Better representation corresponds to a more uniform coverage of the dataset.

• plot_SEACell_sizes(): distribution of number of cells assigned to each metacell. Given to the non-uniform distribution of cell type density, metacells are expected to vary in size: larger metacells are retrieved from denser regions, and smaller metacells from sparser ones. However outlier values may require further inspection.

SEACells.plot.plot_2D(adata, key='X_umap', colour_metacells=False)

/usr/local/lib/python3.8/dist-packages/seaborn/relational.py:573: UserWarning: No data for colormapping provided via 'c'. Parameters 'cmap' will be ignored
  points = ax.scatter(x=x, y=y, **kws)
/usr/local/lib/python3.8/dist-packages/seaborn/relational.py:573: UserWarning: No data for colormapping provided via 'c'. Parameters 'cmap' will be ignored
  points = ax.scatter(x=x, y=y, **kws)

SEACells.plot.plot_2D(adata, key='X_umap', colour_metacells=True)

/usr/local/lib/python3.8/dist-packages/seaborn/relational.py:573: UserWarning: No data for colormapping provided via 'c'. Parameters 'cmap' will be ignored
  points = ax.scatter(x=x, y=y, **kws)
/usr/local/lib/python3.8/dist-packages/seaborn/relational.py:573: UserWarning: No data for colormapping provided via 'c'. Parameters 'cmap' will be ignored
  points = ax.scatter(x=x, y=y, **kws)

SEACells.plot.plot_SEACell_sizes(adata, bins=10, figsize=(10, 5))

/usr/local/lib/python3.8/dist-packages/SEACells/plot.py:121: UserWarning: 

`distplot` is a deprecated function and will be removed in seaborn v0.14.0.

Please adapt your code to use either `displot` (a figure-level function with
similar flexibility) or `histplot` (an axes-level function for histograms).

For a guide to updating your code to use the new functions, please see
https://gist.github.com/mwaskom/de44147ed2974457ad6372750bbe5751

  sns.distplot(label_df.groupby('SEACell').count().iloc[:, 0], bins=bins)

	size
SEACell
SEACell-0	112
SEACell-1	83
SEACell-10	33
SEACell-100	51
SEACell-101	81
...	...
SEACell-95	73
SEACell-96	76
SEACell-97	21
SEACell-98	99
SEACell-99	33

249 rows × 1 columns

8. Quantifying results

• Purity: compute_celltype_purity(ad, col_name) computes the purity of different celltype labels within a SEACell metacell.

• Compactness: per-SEAcell variance in diffusion components (typically 'X_pca' for RNA). Lower values of compactness suggest more compact/lower variance metacells.

• Separation: distance between a SEACell and its nth_nbr. If cluster is provided as a string, e.g. 'celltype', nearest neighbors are restricted to have the same celltype value. Higher values of separation suggest better distinction between metacells.

8.1 Purity

Cell Population

SEACell_purity = SEACells.evaluate.compute_celltype_purity(adata, 'leidenAnnotated')

plt.figure(figsize=(6,4))
sns.boxplot(data=SEACell_purity, y='leidenAnnotated_purity')
plt.title('leidenAnnotated Purity')
sns.despine()
plt.show()
plt.close()

SEACell_purity.head()

	leidenAnnotated	leidenAnnotated_purity
SEACell
SEACell-0	MigratingNeurons	0.964286
SEACell-1	GlutamatergicNeurons_late	0.939759
SEACell-10	ProliferatingProgenitors	0.909091
SEACell-100	RadialGliaProgenitors	0.764706
SEACell-101	GlutamatergicNeurons_late	0.888889

8.2 Compactness

compactness = SEACells.evaluate.compactness(adata, 'X_pca_harmony')

plt.figure(figsize=(6,4))
sns.boxplot(data=compactness, y='compactness')
plt.title('Compactness')
sns.despine()
plt.show()
plt.close()

compactness.head()

Determing nearest neighbor graph...
computing neighbors
    finished: added to `.uns['neighbors']`
    `.obsp['distances']`, distances for each pair of neighbors
    `.obsp['connectivities']`, weighted adjacency matrix (0:00:07)

	compactness
SEACell
SEACell-0	0.000317
SEACell-1	0.000695
SEACell-10	0.015297
SEACell-100	0.002325
SEACell-101	0.000930

8.3 Separation

separation = SEACells.evaluate.separation(adata, 'X_pca_harmony')

plt.figure(figsize=(6,4))
sns.boxplot(data=separation, y='separation')
plt.title('Separation')
sns.despine()
plt.show()
plt.close()

separation.head()

Determing nearest neighbor graph...
computing neighbors
    finished: added to `.uns['neighbors']`
    `.obsp['distances']`, distances for each pair of neighbors
    `.obsp['connectivities']`, weighted adjacency matrix (0:00:07)

	separation
SEACell
SEACell-0	0.034274
SEACell-1	0.057667
SEACell-10	0.373162
SEACell-100	0.116281
SEACell-101	0.078160

9. Hard assignment: downstream processing

• normalization and log-transformation
• dimensionality reduction: PCA and UMAP on SEACells

SEA_adata.layers['counts'] = SEA_adata.X.copy()

sc.pp.normalize_total(SEA_adata)
sc.pp.log1p(SEA_adata)
SEA_adata.layers['lognorm'] = SEA_adata.X.copy()

sc.pp.highly_variable_genes(SEA_adata, inplace=True)
sc.tl.pca(SEA_adata, use_highly_variable=True)
sc.pp.neighbors(SEA_adata, use_rep='X_pca')
sc.tl.umap(SEA_adata)

normalizing counts per cell
    finished (0:00:00)
extracting highly variable genes
    finished (0:00:00)
--> added
    'highly_variable', boolean vector (adata.var)
    'means', float vector (adata.var)
    'dispersions', float vector (adata.var)
    'dispersions_norm', float vector (adata.var)
computing PCA
    on highly variable genes
    with n_comps=50
    finished (0:00:00)
computing neighbors
    finished: added to `.uns['neighbors']`
    `.obsp['distances']`, distances for each pair of neighbors
    `.obsp['connectivities']`, weighted adjacency matrix (0:00:01)
computing UMAP
    finished: added
    'X_umap', UMAP coordinates (adata.obsm) (0:00:01)

sc.pl.umap(SEA_adata, color=['leidenAnnotated'], s=75)

/usr/local/lib/python3.8/dist-packages/scanpy/plotting/_tools/scatterplots.py:392: UserWarning: No data for colormapping provided via 'c'. Parameters 'cmap' will be ignored
  cax = scatter(

sc.pl.umap(SEA_adata, color=['leidenAnnotated_purity'], s=75)

SEA_adata.obs.columns

Index(['leidenAnnotated', 'leidenAnnotated_purity'], dtype='object')

---

10. Saving

10.1 Save SEACell Adata

SEA_adata

AnnData object with n_obs × n_vars = 249 × 33538
    obs: 'leidenAnnotated', 'leidenAnnotated_purity'
    var: 'highly_variable', 'means', 'dispersions', 'dispersions_norm'
    uns: 'log1p', 'hvg', 'pca', 'neighbors', 'umap', 'leidenAnnotated_colors'
    obsm: 'X_pca', 'X_umap'
    varm: 'PCs'
    layers: 'raw', 'counts', 'lognorm'
    obsp: 'distances', 'connectivities'

SEA_adata.X

<249x33538 sparse matrix of type '<class 'numpy.float64'>'
	with 2976218 stored elements in Compressed Sparse Row format>

SEA_adata.write('../../../../DataDir/ExternalData/SingleCellData/Castaldi_metacells_new.h5ad')

10.2 Save in other formats

nb_fname = ipynbname.name()
nb_fname

'SEACellsCastaldi-NEW'

%%bash -s "$nb_fname"
jupyter nbconvert "$1".ipynb --to="python"
jupyter nbconvert "$1".ipynb --to="html"

[NbConvertApp] Converting notebook SEACellsCastaldi-NEW.ipynb to python
[NbConvertApp] Writing 11413 bytes to SEACellsCastaldi-NEW.py
[NbConvertApp] Converting notebook SEACellsCastaldi-NEW.ipynb to html
[NbConvertApp] Writing 2699077 bytes to SEACellsCastaldi-NEW.html

10.3 Finished computations: timestamp

print(datetime.now())

2026-03-05 20:05:18.276663