STEP analysis

This repository contains the notebooks, analysis scripts, and result figures for Decoding Spatial Microarchitectures in Complex Tissues.

Start here

• docs/analysis_index.md maps each analysis to its code and results.
• docs/reproduction_guide.md describes the repository layout and execution conventions.
• docs/environment.md defines the locked Python environment.
• docs/data_accessions.md lists source accessions and expected local paths.
• docs/zenodo_manifest.tsv lists the raw, intermediate, and output artifacts for deposition.
• docs/zenodo_deposition.md describes deterministic packaging and draft upload.
• notebooks/ contains dataset workflows and tutorials.
• scripts/ contains simulations, benchmarks, and figure generation.
• workflows/ contains compact figures, tables, and summary values.

Analysis topics

• Spatial-domain benchmarks on 10x Visium, MERFISH, STARmap, Stereo-seq, Visium HD, and Slide-seq V2.
• CRC and prostate microarchitecture identification and interpretation.
• Liver zonation and cross-modal deconvolution.
• Spatial coherence simulations and STEP module ablation.
• Graph construction, sampling, threshold, and runtime analyses.
• Whole-brain MERFISH integration across 239 sections.
• Matched MERFISH and STARmap PLUS cross-technology mouse-brain integration.

Data sources

Dataset	Public source
Human colorectal cancer, 10x Visium HD	10x Genomics datasets
Human prostate, Slide-seq V2	GEO GSE181294
Human DLPFC, 10x Visium	spatialLIBD HumanPilot
Mouse hypothalamus, MERFISH	Dryad doi:10.5061/dryad.8t8s248
Mouse whole brain, MERFISH	Zhang et al. whole-brain atlas
Cross-technology mouse-brain reference	NicheCompass reproducibility repository
Mouse medial prefrontal cortex, STARmap	STARmap Resources
Mouse embryo E16.5, Stereo-seq	CNGBdb CNP0001543
Human normal and biliary-atresia liver	Transcriptomics technical optimization
Human normal liver, 10x Visium	Figshare doi:10.6084/m9.figshare.22321447.v1
scRNA-seq integration benchmarks	scib-reproducibility
Mouse small intestine, 10x Visium HD	10x Genomics datasets

Expected local paths and artifact provenance are provided in docs/data_accessions.md.

Environment

The analyses use Python 3.11, step-kit==0.3, PyTorch 2.3.1 with CUDA 12.1, and DGL 2.5.0 built for PyTorch 2.3 and CUDA 12.1. Create the Python environment with:

uv sync --extra benchmark --extra notebook --extra microarchitecture

See docs/environment.md for the resolved package-source and build-system details.

Place input H5AD files under data/ or link that directory to the local data store. Each runnable script accepts explicit input and output paths.

Public source datasets are not duplicated in this repository. Their expected locations are documented by the relevant notebook or command-line interface. Generated data and model checkpoints are written under the ignored results/ directory.

Results

Notebook outputs show the main analysis steps for each dataset. Workflow directories provide compact result bundles with figures, metric tables, and the settings needed to reproduce them.

Citation

Citation metadata are provided in CITATION.cff. The source code and documentation in this repository are distributed under the Apache License 2.0.