spatial-transcriptomics-tutorials-with-omicverse

Spatial Transcriptomics with OmicVerse

This skill covers spatial analysis workflows organized into three stages: Preprocessing, Deconvolution, and Downstream Analysis. Each stage includes the critical function calls, parameter guidance, and common pitfalls.

Defensive Validation: Always Check Spatial Coordinates First

Before ANY spatial operation, verify that spatial coordinates exist and are numeric:

# Required check before spatial analysis
assert 'spatial' in adata.obsm, \
    "Missing adata.obsm['spatial']. Load with ov.io.spatial.read_visium() or set manually."
# Cast to float64 to prevent coordinate precision issues during rotation/cropping
adata.obsm['spatial'] = adata.obsm['spatial'].astype('float64')

Stage 1: Preprocessing

Crop, Rotate, and Align Coordinates

Load Visium data and manipulate spatial coordinates for region selection and alignment:

import scanpy as sc, omicverse as ov
ov.plot_set()

adata = sc.datasets.visium_sge(sample_id="V1_Breast_Cancer_Block_A_Section_1")
library_id = list(adata.uns['spatial'].keys())[0]

# Cast coordinates before manipulation
adata.obsm['spatial'] = adata.obsm['spatial'].astype('float64')

# Crop to region of interest
adata_crop = ov.space.crop_space_visium(adata, crop_loc=(0, 0), crop_area=(1000, 1000),
                                         library_id=library_id, scale=1)

# Rotate and auto-align
adata_rot = ov.space.rotate_space_visium(adata, angle=45, library_id=library_id)
ov.space.map_spatial_auto(adata_rot, method='phase')
# For manual refinement: ov.space.map_spatial_manual(adata_rot, ...)

Visium HD Cell Segmentation

Segment Visium HD bins into cells using cellpose:

adata = ov.space.read_visium_10x(path="binned_outputs/square_002um/",
                                  source_image_path="tissue_image.btf")
ov.pp.filter_genes(adata, min_cells=3)
ov.pp.filter_cells(adata, min_counts=1)

# H&E-based segmentation
adata = ov.space.visium_10x_hd_cellpose_he(adata, mpp=0.3, gpu=True, buffer=150)
# Expand labels to neighboring bins
ov.space.visium_10x_hd_cellpose_expand(adata, labels_key='labels_he',
                                        expanded_labels_key='labels_he_expanded', max_bin_distance=4)
# Gene-expression-driven seeds
ov.space.visium_10x_hd_cellpose_gex(adata, obs_key="n_counts_adjusted", mpp=0.3, sigma=5)
# Merge labels and aggregate to cell-level
ov.space.salvage_secondary_labels(adata, primary_label='labels_he_expanded',
                                   secondary_label='labels_gex', labels_key='labels_joint')
cdata = ov.space.bin2cell(adata, labels_key='labels_joint')

Xenium Preprocessing

10x Genomics Xenium output (cell × gene matrices with polygon segmentation already done by the instrument). Read with ov.io.read_xenium; cache the parsed AnnData on disk for fast reuse:

import omicverse as ov
ov.style(font_path='Arial')
ov.settings.cpu_gpu_mixed_init()        # optional: enable mixed CPU/GPU acceleration

# Load — set load_image=False to skip morphology image when not needed
adata = ov.io.read_xenium('data/xenium_breast_rep1', load_image=False)

# Cache for fast re-load (cold parse → write cache; warm reads are 10–50x faster)
adata = ov.io.read_xenium(
    'data/xenium_breast_rep1',
    cache_file='data/xenium_breast_rep1_cache.h5ad',
)

# Standard scanpy preprocessing
ov.pp.normalize_total(adata, target_sum=1e4)
ov.pp.log1p(adata)
ov.pp.scale(adata)
ov.pp.pca(adata, layer='scaled', n_pcs=50)
ov.pp.neighbors(adata, n_neighbors=15, use_rep='scaled|original|X_pca', n_pcs=50)
ov.pp.leiden(adata, resolution=0.5)

# Spatial visualisation — Xenium coordinates have inverted y; remember to invert axis
ov.pl.embedding(adata, basis='spatial', color='leiden',
                palette=ov.pl.palette_112, legend_fontsize=8)

For cell-segmentation overlay (polygon-aware figures with optional H&E / DAPI background), use ov.pl.spatialseg:

library_id = list(adata.uns['spatial'].keys())[0]

ov.pl.spatialseg(
    adata, color='leiden',
    library_id=library_id,
    edges_color='white', edges_width=0.3,
    alpha=1.0, legend_fontsize=8,
    palette=ov.pl.palette_112,
    crop_coord=(2000, 3200, 2500, 3700),    # x0, x1, y0, y1 in spatial coords
    figsize=(7, 6),
)

# With morphology image overlay (DAPI / H&E behind cells)
adata_img = ov.io.read_xenium('data/xenium_breast_rep1', load_image=True)
# Re-attach the same processed obs / obsm before plotting
ov.pl.spatialseg(
    adata_img, color='KRT7',
    library_id=library_id,
    edges_color='white', edges_width=0.4,
    alpha=0.65, alpha_img=1.0,        # 0.45–0.65 keeps morphology visible
    cmap=ov.pl.create_custom_colormap('#a51616'), vmax=10,
    seg_contourpx=1.5,                # dashed cell-outline contour
    crop_coord=(2000, 3200, 2500, 3700),
    figsize=(7, 6),
)

Xenium-specific gotchas:

• vmax='p99.2' on ov.pl.embedding clips the long-tailed expression distribution that's typical of Xenium probes; without it a few hot cells dominate the colour scale.
• alpha=0.45–0.65 on spatialseg plots when alpha_img=1.0 keeps DAPI morphology readable through cluster fills. Higher alpha hides the background.
• Spatial axis: Xenium spatial coordinates use image-pixel convention (y increases downward); call ax.invert_yaxis() on ov.pl.embedding to match the morphology image orientation.
• Cache the parsed AnnData with cache_file= — the cold parse of a 100k-cell Xenium run is 10–60 s; warm reads are sub-second.

Stage 2: Deconvolution

Critical API Reference: Method Selection

# CORRECT — Tangram: method passed to deconvolution() call
decov_obj = ov.space.Deconvolution(adata_sc=sc_adata, adata_sp=sp_adata,
                                    celltype_key='Subset', result_dir='result/tangram')
decov_obj.preprocess_sc(max_cells=5000)
decov_obj.preprocess_sp()
decov_obj.deconvolution(method='tangram', num_epochs=1000)

# CORRECT — cell2location: method passed at INIT time
cell2_obj = ov.space.Deconvolution(adata_sc=sc_adata, adata_sp=sp_adata,
                                    celltype_key='Subset', result_dir='result/c2l',
                                    method='cell2location')
cell2_obj.deconvolution(max_epochs=30000)
cell2_obj.save_model('result/c2l/model')

Note: For cell2location, the method parameter is set at initialization, not at the deconvolution() call. For Tangram, it's passed to deconvolution().

FlashDeconv — Atlas-Scale Sketching-Based Deconvolution

For Visium HD / Slide-seq / Stereo-seq (10⁵–10⁶ spots) where Tangram and cell2location become impractical, use method='FlashDeconv'. It runs on CPU, finishes in minutes for 10k spots vs. 60+ for cell2location, and ships built-in spatial regularisation:

decov_obj = ov.space.Deconvolution(
    adata_sc=sc_adata,
    adata_sp=sp_adata,
    celltype_key='Subset',
    result_dir='result/flashdeconv',
    method='FlashDeconv',                 # set at INIT time, like cell2location
)
decov_obj.preprocess_sc(max_cells=5000)
decov_obj.preprocess_sp()
decov_obj.deconvolution(
    sketch_dim=512,           # randomized-sketch dimension; raise to 1024 for HD-2µm
    lambda_spatial=5000,      # spatial smoothness; raise to 10000 for sparse / noisy data
    n_hvg=2000,               # raise to 3000 when accuracy matters more than speed
)

# Results — three accessors (same payload, different wrappings):
decov_obj.adata_cell2location               # AnnData with proportions in .X
decov_obj.adata_sp.obsm['flashdeconv']      # DataFrame of proportions
decov_obj.adata_sp.obs['flashdeconv_dominant']  # dominant cell type per spot

# Spatial visualisation reuses the same plot stack as cell2location
ov.pl.plot_spatial(
    adata=decov_obj.adata_cell2location,
    color=clust_labels, labels=clust_labels,
    show_img=True, style='fast',
    max_color_quantile=0.992,
    circle_diameter=4,
    colorbar_position='right',
    palette=color_dict,
)

When to pick FlashDeconv over the others:

Feature	FlashDeconv	Tangram	cell2location
GPU required	No	Optional	Recommended
10k spots wall-time	~2 min	~15 min	~60 min
Native Visium HD support	Yes	Limited	Limited
Built-in spatial regularisation	Yes (lambda_spatial)	No	No

Parameter tuning quick reference (from the FlashDeconv tutorial):

• Noisy / sparse data → raise lambda_spatial to 10000.
• Visium HD 2 µm → raise sketch_dim to 1024.
• Accuracy-critical → raise n_hvg to 3000.

Starfysh Archetypal Deconvolution

Treat the file names below as placeholders for your own local exports; the matching public spatial tutorials describe the expected AnnData counts matrix and signature table formats.

from omicverse.external.starfysh import AA, utils, plot_utils

visium_args = utils.prepare_data(adata_path="data/counts.h5ad",
                                  signature_path="data/signatures.csv",
                                  min_cells=10, filter_hvg=True, n_top_genes=3000)
adata, adata_normed = visium_args.get_adata()
aa_model = AA.ArchetypalAnalysis(adata_orig=adata_normed)
aa_model.fit(k=12, n_init=10)
visium_args = utils.refine_anchors(visium_args, aa_model, add_marker=True)
model, history = utils.run_starfysh(visium_args, poe=False, n_repeat=5, lr=5e-3, max_epochs=500)

Stage 3: Downstream Analysis

Spatial Clustering

# GraphST + mclust
ov.utils.cluster(adata, use_rep='graphst|original|X_pca', method='mclust', n_components=7)

# STAGATE
ov.utils.cluster(adata, use_rep='STAGATE', method='mclust', n_components=7)

# Merge small clusters
ov.space.merge_cluster(adata, groupby='mclust', resolution=0.5)

Algorithm choice: GraphST and STAGATE require precalculated latent spaces. For standard clustering without spatial-aware embeddings, use Leiden/Louvain directly.

Multi-Slice Integration (STAligner)

import anndata as ad
Batch_list = [ov.read(p) for p in slice_paths]
adata_concat = ad.concat(Batch_list, label='slice_name', keys=section_ids)
STAligner_obj = ov.space.pySTAligner(adata=adata_concat, batch_key='slice_name',
                                      hidden_dims=[256, 64], use_gpu=True)
STAligner_obj.train_STAligner_subgraph(nepochs=800, lr=1e-3)
STAligner_obj.train()
adata_aligned = STAligner_obj.predicted()
sc.pp.neighbors(adata_aligned, use_rep='STAligner')

Spatial Trajectories

SpaceFlow — pseudo-spatial maps:

sf_obj = ov.space.pySpaceFlow(adata)
sf_obj.train(spatial_regularization_strength=0.1, num_epochs=300, patience=50)
sf_obj.cal_pSM(n_neighbors=20, resolution=1.0)

STT — transition dynamics:

STT_obj = ov.space.STT(adata, spatial_loc='xy_loc', region='Region', n_neighbors=20)
STT_obj.stage_estimate()
STT_obj.train(n_states=9, n_iter=15, weight_connectivities=0.5)

Cell Communication (COMMOT + FlowSig)

df_cellchat = ov.external.commot.pp.ligand_receptor_database(species='human', database='cellchat')
df_cellchat = ov.external.commot.pp.filter_lr_database(df_cellchat, adata, min_expr_frac=0.05)
ov.external.commot.tl.spatial_communication(adata, lr_database=df_cellchat,
                                              distance_threshold=500, result_prefix='cellchat')
# FlowSig network
adata.layers['normalized'] = adata.X.copy()
ov.external.flowsig.tl.construct_intercellular_flow_network(
    adata, commot_output_key='commot-cellchat',
    flowsig_output_key='flowsig-cellchat', edge_threshold=0.7)

Structural Layers (GASTON) and Slice Alignment (SLAT)

GASTON — iso-depth estimation:

gas_obj = ov.space.GASTON(adata)
A = gas_obj.prepare_inputs(n_pcs=50)
gas_obj.load_rescale(A)
gas_obj.train(hidden_dims=[64, 32], dropout=0.1, max_epochs=2000)
gaston_isodepth, gaston_labels = gas_obj.cal_iso_depth(n_layers=5)

SLAT — cross-slice alignment:

from omicverse.external.scSLAT.model import Cal_Spatial_Net, load_anndatas, run_SLAT, spatial_match
Cal_Spatial_Net(adata1, k_cutoff=20, model='KNN')
Cal_Spatial_Net(adata2, k_cutoff=20, model='KNN')
edges, features = load_anndatas([adata1, adata2], feature='DPCA', check_order=False)
embeddings, *_ = run_SLAT(features, edges, LGCN_layer=5)
best, index, distance = spatial_match(embeddings, adatas=[adata1, adata2])

Troubleshooting

• ValueError: spatial coordinates out of bounds after rotation: Cast adata.obsm['spatial'] to float64 BEFORE calling rotate_space_visium. Integer coordinates lose precision during trigonometric rotation.
• Cellpose segmentation fails with memory error: For large .btf images, use backend='tifffile' to memory-map the image. Reduce buffer parameter if GPU memory is insufficient.
• Gene ID overlap failure in Tangram/cell2location: Harmonise identifiers (ENSEMBL vs gene symbols) between adata_sc and adata_sp before calling preprocess_sc/preprocess_sp. Drop non-overlapping genes.
• mclust clustering error: Requires rpy2 and the R mclust package. If R bindings are unavailable, switch to method='louvain' or method='leiden'.
• STAligner/SpaceFlow embeddings collapse to a single point: Verify adata.obsm['spatial'] exists and coordinates are scaled appropriately. Tune learning rate (try lr=5e-4) and regularisation strength.
• FlowSig returns empty network: Build spatial neighbor graphs before Moran's I filtering. Increase bootstraps or lower edge_threshold (try 0.5) if the network is too sparse.
• GASTON RuntimeError in training: Provide a writable out_dir path. PyTorch nondeterminism may cause variation between runs—set torch.manual_seed() for reproducibility.
• SLAT alignment has many low-quality matches: Regenerate spatial graphs with a higher k_cutoff value. Inspect low_quality_index flags and filter cells with high distance scores.
• STT pathway enrichment fails: gseapy needs network access for gene set downloads. Cache gene sets locally with ov.utils.geneset_prepare() and pass the dictionary directly.

Dependencies

• Core: omicverse, scanpy, anndata, squidpy, numpy, matplotlib
• Segmentation: cellpose, opencv-python/tifffile, optional GPU PyTorch
• Deconvolution: tangram-sc, cell2location, pytorch-lightning; Starfysh needs torch, scikit-learn
• Downstream: scikit-learn, commot, flowsig, gseapy, torch-backed modules (STAligner, SpaceFlow, GASTON, SLAT)

Examples

• "Crop and rotate my Visium slide, then run cellpose segmentation on the HD data and aggregate to cell-level AnnData."
• "Deconvolve my lymph node spatial data with Tangram and cell2location, compare proportions, and plot cell-type maps."
• "Integrate three DLPFC slices with STAligner, cluster with STAGATE, and infer communication with COMMOT+FlowSig."

References

• Quick copy/paste commands: reference.md