omicverse/omicverse-skills/single-to-spatial-mapping

Single2Spatial spatial mapping

Overview

Apply this skill when converting single-cell references into spatially resolved profiles. It follows `t_single2spatial.ipynb`, demonstrating how Single2Spatial trains on PDAC scRNA-seq and Visium data, reconstructs spot-level proportions, and visualises marker expression.

Instructions

1. Import dependencies and style

• Load omicverse as ov, scanpy as sc, anndata, pandas as pd, numpy as np, and matplotlib.pyplot as plt.
• Call ov.utils.ov_plot_set() (or ov.plot_set() in older versions) to align plots with omicverse styling.

1. Load single-cell and spatial datasets

• Read processed matrices with pd.read_csv(...) then create AnnData objects (anndata.AnnData(raw_df.T)).
• Attach metadata: single_data.obs = pd.read_csv(...)[['Cell_type']] and spatial_data.obs = pd.read_csv(... ) containing coordinates and slide metadata.

1. Initialise Single2Spatial

• Instantiate ov.bulk2single.Single2Spatial(single_data=single_data, spatial_data=spatial_data, celltype_key='Cell_type', spot_key=['xcoord','ycoord'], gpu=0).
• Note that inputs should be normalised/log-scaled scRNA-seq matrices; ensure spot_key matches spatial coordinate columns.

1. Train the deep-forest model

• Execute st_model.train(spot_num=500, cell_num=10, df_save_dir='...', df_save_name='pdac_df', k=10, num_epochs=1000, batch_size=1000, predicted_size=32) to fit the mapper and generate reconstructed spatial AnnData (sp_adata).
• Explain that spot_num defines sampled pseudo-spots per iteration and cell_num controls per-spot cell draws.

1. Load pretrained weights

• Use st_model.load(modelsize=14478, df_load_dir='.../pdac_df.pth', k=10, predicted_size=32) when checkpoints already exist to skip training.

1. Assess spot-level outputs

• Call st_model.spot_assess() to compute aggregated spot AnnData (sp_adata_spot) for QC.
• Plot marker genes with sc.pl.embedding(sp_adata, basis='X_spatial', color=['REG1A', 'CLDN1', ...], frameon=False, ncols=4).

1. Visualise proportions and cell-type maps

• Use sc.pl.embedding(sp_adata_spot, basis='X_spatial', color=['Acinar cells', ...], frameon=False) to highlight per-spot cell fractions.
• Plot sp_adata coloured by Cell_type with palette=ov.utils.ov_palette()[11:] to show reconstructed assignments.

1. Export results

• Encourage saving generated AnnData objects (sp_adata.write_h5ad(...), sp_adata_spot.write_h5ad(...)) and derived CSV summaries for downstream reporting.

1. Defensive validation

``python # Before Single2Spatial: verify spatial coordinates exist for col in spot_key: assert col in spatial_data.obs.columns, f"Spatial coordinate column '{col}' not found in spatial_data.obs" # Verify scRNA-seq is log-normalized (max should be <~15, not hundreds/thousands) if single_data.X.max() > 50: print("WARNING: scRNA-seq data may not be log-normalized. Raw counts cause scale mismatches.") # Verify cell type column exists assert celltype_key in single_data.obs.columns, f"Cell type column '{celltype_key}' not found" ``

1. Troubleshooting tips

• If training diverges, reduce learning_rate via keyword arguments or decrease predicted_size to stabilise the forest.
• Ensure scRNA-seq inputs are log-normalised; raw counts can lead to scale mismatches and poor spatial predictions.
• Verify GPU availability when gpu is non-zero; fallback to CPU by omitting the argument or setting gpu=-1.

Examples

• "Train Single2Spatial on PDAC scRNA-seq and Visium slides, then visualise REG1A and CLDN1 spatial expression."
• "Load a saved Single2Spatial checkpoint to regenerate spot-level cell-type proportions for reporting."
• "Plot reconstructed cell-type maps with omicverse palettes to compare against histology."

References

• Tutorial notebook: `t_single2spatial.ipynb`
• Example datasets and models: see the tutorial page for linked PDAC datasets and pretrained models: t_single2spatial
• Quick copy/paste commands: `reference.md`