SCOUBI

SCOUBI is a computational framework for analyzing extrasomatic transcripts in imaging-based spatial transcriptomics data. Rather than discarding unassigned RNA as background, SCOUBI uses marker-gene expression of neurites and spatial ligand–receptor colocalization to classify extrasomatic regions as axon- or dendrite-enriched and to identify putative “interfaces” enriched for intercellular signaling programs. These interfaces provide a transcriptomic window into neuronal communication sites that are typically inaccessible to cell-centric spatial analyses. By preserving anatomical context, SCOUBI enables regional and molecular characterization of putative synaptic signaling environments directly from intact tissue.

Preprint: add preprint link here

Overview

Imaging-based spatial transcriptomics often detects a substantial fraction of transcripts outside segmented cell bodies. SCOUBI treats this extrasomatic signal as a biologically informative layer of tissue organization.

The workflow has two main stages:

1. Neurite identity classification
   Extrasomatic spatial bins are classified as axon- or dendrite-enriched using curated marker genes and a signaling-informed optimization objective.

2. Interface identification
   SCOUBI scans for local regions where axonic and dendritic bins converge and where significant ligand–receptor pairs are colocalized.

The resulting interface map can be used to study:

• Spatial distribution of putative neuronal communication regions
• Neurite-enriched genes
• Interface-enriched genes
• Regional variation in interface transcriptomes
• Interface signatures compared with nearby cell bodies

Installation

pip install scoubi

For local development from a checkout:

pip install -e .

Quickstart

import torch
import scoubi

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

adata = scoubi.io.load_data(
    "data/data.parquet",
    cell_type="data/cell_types.csv",
    region="data/regions.csv",
)

adata = scoubi.pp.bin_data(adata, binsize=2)
adata = scoubi.md.train(adata, axon_markers, dendrite_markers, device=device)
adata = scoubi.tl.overview(adata, device=device)
adata = scoubi.tl.axon_dendrite_enrichment(adata)
adata = scoubi.tl.distance(adata)
adata = scoubi.tl.expression_profile(adata, key="region")
adata = scoubi.tl.communication_profile(adata, key="region")

adata.summarize()

Module Map

Data and Runtime Notes

• scoubi.md.train() can use a bundled ligand-receptor reference table (from NeuronChat) when pairs=None.
• GPU acceleration is optional; most workflows can run on CPU, although training and convolution-heavy steps are faster on CUDA when available.

Tutorial

The guided walkthrough lives in tutorial.ipynb. It covers data loading, model training, spatial overview generation, enrichment analysis, interface profiling, communication analysis, and save/load workflows.

License

Released under the MIT License. See LICENSE.

Citation

If you use SCOUBI in your work, please cite:

@article{scoubi,
  title = {SCOUBI: Signaling-informed Characterization of Unresolved Biological Interfaces},
  author = {TODO},
  journal = {TODO},
  year = {TODO},
  doi = {TODO}
}