TrackCell

A Python package for processing and visualizing single-cell and spatial transcriptomics data. Supports 10x Visium HD (SpaceRanger), 10x Xenium Analyzer and 10x Atera outputs.

Installation

pip install trackcell -i https://pypi.org/simple
# pip install --upgrade trackcell==0.1.9 -i https://pypi.org/simple

Usage

Reading SpaceRanger Output

Reading Cell Segmentation Data

import trackcell as tcl

# Read SpaceRanger cell segmentation output
adata = tcl.io.read_hd_cellseg(
    datapath="SpaceRanger4.0/Cse1/outs/segmented_outputs",
    sample="Cse1"
)

# The resulting AnnData object contains:
# - Expression matrix in .X
# - Cell metadata in .obs
# - Gene metadata in .var
# - Spatial coordinates in .obsm["spatial"]
# - Tissue images in .uns["spatial"][sample]["images"]
# - Scalefactors in .uns["spatial"][sample]["scalefactors"]
# - Cell geometries in .uns["spatial"][sample]["geometries"] (GeoDataFrame)
# - Cell geometries in .obs["geometry"] (WKT strings for serialization)

Subsetting Data and Synchronizing Geometries

Important: When you subset data loaded with read_hd_cellseg(), you must call sync_geometries_after_subset() to synchronize the geometries:

import trackcell as tcl
import numpy as np

# Read data
adata = tcl.io.read_hd_cellseg(
    datapath="SpaceRanger4.0/Cse1/outs/segmented_outputs",
    sample="Cse1"
)

# Subset by spatial region
x_min, x_max = 16000, 18000
y_min, y_max = 14000, 18000

spatial_coords = adata.obsm['spatial']
mask = ((spatial_coords[:, 0] >= x_min) & (spatial_coords[:, 0] <= x_max) &
        (spatial_coords[:, 1] >= y_min) & (spatial_coords[:, 1] <= y_max))

adata_subset = adata[mask].copy()

# IMPORTANT: Synchronize geometries after subsetting
tcl.io.sync_geometries_after_subset(adata_subset, sample="Cse1")

# Now you can safely plot the subset
tcl.pl.spatial_cell(adata_subset, color="classification")

Why this is necessary: When you subset an AnnData object, adata.obs and adata.obsm are automatically subset, but adata.uns["spatial"][sample]["geometries"] (GeoDataFrame) is not. Without synchronization, plotting may fail with errors like ValueError: aspect must be finite and positive.

Reading Xenium Output

import trackcell as tcl

# Read Xenium Analyzer cell segmentation output
adata = tcl.io.read_xenium_cellseg(
    datapath="/path/to/xenium/output",
    sample="sample1"
)

# The resulting AnnData object contains:
# - Expression matrix in .X (CSR, cells × genes)
# - Cell metadata in .obs (centroids, area, counts, segmentation_method, etc.)
# - Gene metadata in .var (gene_ids, feature_types)
# - Spatial coordinates in .obsm["spatial"]
# - Cell polygons in .uns["spatial"][sample]["geometries"] (GeoDataFrame)
# - Cell boundary arrays in .uns["cell_boundaries"] (compact vertex arrays)
# - Nucleus boundary arrays in .uns["nucleus_boundaries"] (if available)
# - WKT geometry strings in .obs["geometry"] (for serialization)
# - Experiment metadata in .uns["experiment"]

The function reads Xenium Analyzer output files:

Key differences from Visium HD:

• Cell boundaries are stored in long-table parquet format (one row per vertex), automatically converted to Shapely polygons
• Expression matrix is in CSC (genes × cells) format, automatically transposed to CSR (cells × genes)
• Cell IDs already include the -1 suffix (no stripping needed)

After loading, the AnnData object is fully compatible with tcl.pl.spatial_cell():

# Visualize Xenium data with cell polygons
tcl.pl.spatial_cell(adata, color="EPCAM", cmap="Reds", edges_width=0.3)

TMA multi-core slice separation: For Xenium TMA data (multiple tissue cores in one region), enable automatic DBSCAN slice separation during load:

adata = tcl.io.read_xenium_cellseg(
    "/path/to/xenium/region", sample="85811_S",
    slice_separate=True, slice_eps=80,
)
print(adata.obs["slice_id"].value_counts())  # S001, S002, ...

See the DBSCAN Slice Separation section below for the full workflow.

Reading Atera Output

import trackcell as tcl

# Inspect a bundle without loading the matrix (cells/genes/polygons/transcripts)
tcl.io.atera_info("/path/to/ROI_bundle")

# Auto-detect official Atera 1.x Zarr vs Xenium-compatible preview
adata = tcl.io.read_atera("/path/to/ROI_bundle")

# var_names are gene symbols; remove negative-control codewords before PCA
adata = tcl.io.filter_atera_genes(adata)

# Stream the single-molecule point cloud (billions of molecules)
for chunk in tcl.io.iter_atera_transcripts(
        "/path/to/ROI_bundle", genes=["NPHS1", "UMOD"], min_qv=20):
    ...

# Q-score / nuclear fraction / density QC (results stored in adata)
summary = tcl.io.atera_transcript_qc("/path/to/ROI_bundle", adata=adata)
tcl.pl.atera_transcript_density(adata)
tcl.pl.atera_nuclear_fraction(adata, genes=["UMOD", "DDX17"])

obsm["spatial"] is in µm (Atera centroids are already µm; do not multiply by pixel_size). Atera is gene-level: there is no isoform, splice-junction, spliced/unspliced or RNA-velocity data. See the Atera documentation for the full schema and multi-ROI handling.

Reading Bin-Level Data (2um/8um/16um)

import trackcell as tcl

# Read SpaceRanger bin-level output (2um/8um/16um bins)
adata = tcl.io.read_hd_bin(
    datapath="SpaceRanger4.0/Cse1/binned_outputs",
    sample="Cse1",
    binsize=16  # Bin size in micrometers (default: 16, common values: 2, 8, or 16)
)

# The function automatically handles:
# - filtered_feature_bc_matrix.h5 (preferred) or filtered_feature_bc_matrix/ directory
# - tissue_positions.parquet or tissue_positions.csv
# - tissue_hires_image.png and tissue_lowres_image.png
# - scalefactors_json.json

# The resulting AnnData object contains:
# - Expression matrix in .X
# - Bin metadata in .obs (with spatial coordinates)
# - Gene metadata in .var
# - Spatial coordinates in .obsm["spatial"]
# - Tissue images in .uns["spatial"][sample]["images"]
# - Scalefactors in .uns["spatial"][sample]["scalefactors"]
# - Bin size in .uns["spatial"][sample]["binsize"] (e.g., 2, 8, or 16)

# Access the bin size information:
print(f"Bin size: {adata.uns['spatial']['Cse1']['binsize']} um")

Visualization

Plotting with Cell Polygons

# Plot cells as polygons (requires data loaded with read_hd_cellseg)
tcl.pl.spatial_cell(
    adata, 
    color="classification",  # Color by cell type
    groups=['Cluster-2', 'Cluster-3'],  # Optional: filter specific groups
    figsize=(10, 10),
    edges_width=0.5,
    edges_color="black",
    alpha=0.8
)

# Plot continuous values (e.g., distance to a label)
tcl.pl.spatial_cell(
    adata,
    color="Cluster-2_dist",  # Distance to Cluster-2
    cmap="Reds",
    figsize=(10, 10)
)

Dual-Color Visualization (Fill + Edge)

Use edge_color to color cell boundaries by a categorical column (e.g., cell type), while color controls the fill (gene expression or continuous value):

# Fill = gene expression, Edge = cell type
tcl.pl.spatial_cell(
    adata,
    color='EPCAM',              # Fill: gene expression (continuous)
    cmap='Reds',
    edge_color='cell_type',     # Edge: cell type (categorical)
    edge_palette={
        'T cell': '#e41a1c',
        'B cell': '#377eb8',
        'Myeloid': '#4daf4a',
    },
    edges_width=1.2,
    alpha=0.7,
)

# Fill = continuous obs, Edge = categorical obs
tcl.pl.spatial_cell(
    adata,
    color='total_counts',       # Fill: UMIs per cell
    cmap='YlOrRd',
    edge_color='cell_type',     # Edge: cell type
    edges_width=1.5,
    alpha=0.7,
)

Plotting Square Bins (Visium HD)

# Read square-bin output
adata_bin = tcl.io.read_hd_bin(
    datapath="SpaceRanger4.0/Cse1/binned_outputs/square_016um",
    sample="Cse1",
    binsize=16
)

# Show only H&E image + coordinate range
tcl.pl.spatial_squarebin(adata_bin, color=None)

# Equivalent alias
tcl.pl.spatial_bin(adata_bin, color=None)

# Plot a gene on square bins
tcl.pl.spatial_squarebin(
    adata_bin,
    color="EPCAM",
    cmap="Reds",
    alpha=0.8,
    alpha_img=0.4,
)

# Use square markers (shape='square') instead of circles (shape='circle', default)
tcl.pl.spatial_squarebin(adata_bin, color="EPCAM", shape="circle")

# Cartesian y-axis (invert_y=False) instead of image coordinates (default True)
tcl.pl.spatial_squarebin(adata_bin, color="EPCAM", invert_y=False)

Marking Regions

Draw rectangular highlights on spatial plots to annotate regions of interest (ROI):

# mark_region works on any matplotlib Axes
ax = tcl.pl.spatial_cell(adata, color="CellType")
tcl.pl.mark_region(ax, xlim=(54500, 56000), ylim=(15000, 16000))

# With semi-transparent fill
tcl.pl.mark_region(
    ax, xlim=(54500, 56000), ylim=(15000, 16000),
    fill_color='red', fill_alpha=0.15, edges_width=3.0,
)

# Multiple regions efficiently (defer refresh)
tcl.pl.mark_region(ax, xlim=(40000, 42000), ylim=(5000, 7000),
                   edges_color='cyan', fill_color='cyan', refresh=False, show=False)
tcl.pl.mark_region(ax, xlim=(55000, 57000), ylim=(15000, 17000),
                   edges_color='yellow', fill_color='yellow', refresh=False, show=False)
tcl.pl.mark_region(ax, xlim=(60000, 62000), ylim=(10000, 12000),
                   edges_color='magenta', fill_color='magenta')  # show=True by default

Interactive ROI Selection (Jupyter)

Select regions of interest interactively in a Jupyter notebook with auto-naming, real-time cell highlighting, toolbar, undo, and keyboard shortcuts.

%matplotlib widget  # required in Jupyter

selector = tcl.pl.select_regions(
    adata, color="CellType", key_added="ROI", inplace=True,
)
# Use toolbar buttons (■ Rect, ● Ellipse, ✎ Lasso) or keyboard r/e/l
# ROIs are auto-named: ROI_1, ROI_2, ...

# Post-hoc renaming
selector.rename_roi("ROI_1", "tumor")
selector.rename_roi("ROI_2", "stroma")

# Results in adata.obs
adata.obs["ROI"].value_counts()

Traditional Point-based Visualization

# Using scanpy (point-based)
sc.pl.spatial(adata, color='classification', size=2, 
              groups=['Cluster-2', 'Cluster-3'],
              legend_fontsize=12, spot_size=10, frameon=True
             )

# Using squidpy (point-based)
sq.pl.spatial_scatter(
    adata, shape=None, color=["classification"], 
    edges_width=0, size=0.1, 
    library_id="spatial", 
    groups=['Cluster-2', 'Cluster-3'],
    figsize=(5, 4), 
    #cmap='Blues'
    #palette = mycolor
    #img_key="0.3_mpp_150_buffer", 
    #basis="spatial_cropped_150_buffer"
)

Converting annohdcell Output to TrackCell Format

TrackCell provides two methods to convert annohdcell's bin2cell output into trackcell-compatible format with polygon geometries for spatial visualization.

Method 1: Convert from 2μm Bin H5AD Only

Create a new cell-level h5ad from annohdcell's 2μm bin h5ad with cell labels:

import trackcell as tcl

# Convert annohdcell 2μm bin h5ad to trackcell format
adata = tcl.io.convert_annohdcell_to_trackcell(
    bin_h5ad_path="b2c_2um.h5ad",
    output_h5ad_path="trackcell_format.h5ad",
    sample="sample1"
)

# Now visualize with trackcell
tcl.pl.spatial_cell(adata, library_id="sample1")

Method 2: Add Geometries to Existing Cell H5AD

Add polygon geometries to annohdcell's final cell h5ad output (preserves exact count aggregation):

import trackcell as tcl

# Add geometries to annohdcell's final cell h5ad
adata = tcl.io.add_geometries_to_annohdcell_output(
    bin_h5ad_path="b2c_2um.h5ad",      # 2μm bin h5ad with cell labels
    cell_h5ad_path="b2c_cell.h5ad",    # Final cell h5ad from annohdcell
    output_h5ad_path="b2c_cell_with_geom.h5ad",
    sample="sample1"
)

# Now visualize with trackcell
tcl.pl.spatial_cell(adata, library_id="sample1")

Key differences:

• Method 1: Quick conversion, simple count summation
• Method 2: Preserves annohdcell's exact count aggregation and all metadata

For detailed documentation, see docs/convert_annohdcell.md

Computing Distances to a Label (10x HD)

# Compute distance to a specific annotation label stored in adata.obs["group_col"]
tcl.tl.hd_labeldist(
    adata,
    groupby="classification",    # obs column containing cell type annotations
    label="Cluster-2",       # target label to measure distances from
    inplace=True          # add "{label}_px" and "{label}_dist" to adata.obs
)

# When inplace=False the function returns a DataFrame with the two columns:
dist_df = tcl.tl.hd_labeldist(adata, groupby="group_col", label="Neuron", inplace=False)

# Visualize distance using cell polygons
tcl.pl.spatial_cell(adata, color='Cluster-2_dist', cmap='Reds', figsize=(10, 10))

# Or using traditional point-based visualization
sc.pl.spatial(adata, color='Cluster-2_dist', size=2,
              legend_fontsize=12, spot_size=10, frameon=True
             )

Multi-Gene Visualization (cell2location-style)

Visualize co-expression of multiple genes in two modes:

Mode 1: Blended Composite (mode='blend')

Maps each gene to a base color, blends by expression level → single hex color per cell. Use with spatial_cell for a composite view:

# Compute blended colors
adata = tcl.tl.multigene_blend(
    adata,
    genes=['EPCAM', 'PECAM1', 'VWF'],
    mode='blend',
)

# Visualize with cell polygons
tcl.pl.spatial_cell(adata, color='multigene_blend')

# Custom colors + gamma correction
tcl.tl.multigene_blend(
    adata,
    genes=['EPCAM', 'PECAM1', 'VWF', 'ACTA2', 'PTPRC'],
    colors=['#e41a1c', '#377eb8', '#4daf4a', '#ff7f00', '#984ea3'],
    vmax_percentile=98, gamma=0.8,
)

Mode 2: Faceted Subplots (mode='facet')

Each gene gets its own panel with a single-hue colormap (white → gene color), matching the cell2location paper style:

tcl.tl.multigene_blend(
    adata,
    genes=['EPCAM', 'PECAM1', 'VWF', 'ACTA2'],
    mode='facet',
    ncols=2,           # 2 columns of subplots
    edges_width=0.3,
)

YardCluster Spatial Clustering

Lightweight CPU spatial clustering that combines BANKSY-style neighborhood features with dual-channel identity/context embeddings and Leiden clustering. One function covers the full pipeline (preprocessing, spatial features, embedding, clustering):

# One-shot YardCluster — includes HVG/normalize/log/scale by default
import trackcell as tcl

tcl.tl.spatial_cluster(adata, mode="auto")

# Visualize tissue domains and cell-type-oriented clusters
tcl.pl.spatial_cell(adata, color="yardcluster_domain", figsize=(10, 10))
tcl.pl.spatial_cell(adata, color="yardcluster_celltype", figsize=(10, 10))

Clustering modes:

• mode="celltype" (λ=0.2): emphasizes cell-intrinsic expression → cell typing
• mode="domain" (λ=0.8): emphasizes neighborhood context → tissue domains
• mode="auto" (default): runs both in one call

Advanced options: multi-sample Harmony integration (integrate="joint"), DE-guided cluster merging (merge_clusters=True), sketch mode for >500k cells, and gradient features (use_gradient=True).

# Multi-sample integration
tcl.tl.spatial_cluster(adata, batch_key="sample_id", integrate="joint", mode="celltype")

# Per-sample separate clustering
tcl.tl.spatial_cluster(adata, batch_key="sample_id", integrate="separate")

# Skip built-in preprocessing if data is already prepared
tcl.tl.spatial_cluster(adata, preprocess=False, mode="auto")

DBSCAN Slice Separation & Colony Clustering

TMA Slice Separation (Xenium)

Split multi-core Xenium TMA regions into physical tissue sections (S001, S002, …) using DBSCAN density clustering on cell centroids:

import trackcell as tcl

# Option A: during load
adata = tcl.io.read_xenium_cellseg("/path/to/xenium/region",
                                    sample="85811_S",
                                    slice_separate=True, slice_eps=80)

# Option B: separate step
adata = tcl.io.read_xenium_cellseg("/path/to/xenium/region")
tcl.tl.spatial_slice_cluster(adata, eps=80, min_cells=1000)

# Summarize and split
summary = tcl.tl.slice_cluster_summary(adata)
slices = tcl.tl.split_by_slice(adata)          # dict: "S001" → AnnData

# Visualize
tcl.pl.spatial_cell(adata, color="slice_id", figsize=(12, 12))

Micro-Region Colony Clustering (GC / Tumor Nests)

Separate spatially disconnected germinal centers or tumor nests within a single tissue section (after BANKSY / YardCluster filtering):

# After clustering with YardCluster, isolate e.g. GC B cells
tcl.tl.spatial_cluster(sub, mode="auto")
gc = sub[sub.obs["yardcluster_domain"] == "Germinal Center B Cells"].copy()

# DBSCAN colony identification
tcl.tl.spatial_colony_cluster(gc, eps=50, min_samples=5, min_cluster_size=40)
tcl.tl.mark_colony_centroids(gc, centroid_label="GCC")

# Distance analysis
tcl.tl.distance_to_nearest_centroids(
    sub, centroid_key="cell_centroid_type",
    distance_key="distance_to_nearest_gcc", centroid_label="GCC",
)
tcl.pl.spatial_cell(sub, color="distance_to_nearest_gcc", cmap="Reds")

Development

License

update

git tag v0.3.8
git push origin v0.3.8

# In GitHub, go to "Releases" → "Draft a new release".