← Về thư mục
📄 / / profiles / r-and-d / raw_evidence / OpenAI4S / skills / bioskills / bio-spatial-transcriptomics-spatial-communication / references / usage-guide.md

Spatial Cell-Cell Communication - Usage Guide

Overview

This skill infers cell-cell communication from spatial transcriptomics data by scoring ligand-receptor co-expression between sender and receiver cell types against a permutation null. Its core job is decision-making, not recipe-running: choosing a method by whether spatial distance is actually modeled (squidpy ligrec is space-blind, COMMOT is distance-aware, stLearn and CellChat-spatial sit between), choosing the ligand-receptor database deliberately because it drives the result as much as the algorithm, guarding against segmentation spillover that fabricates short-range hits, and treating every call as a co-expression hypothesis on a confidence ladder rather than as validated signaling.

Prerequisites

pip install squidpy scanpy anndata statsmodels matplotlib
pip install commot          # distance-aware optimal-transport inference
# CellChat v2 (spatial) and NicheNet are R/Bioconductor packages

Quick Start

Tell your AI agent what you want to do: - "Map cell-cell communication in my Visium data and tell me which calls are spatially supported" - "Run ligand-receptor analysis but treat the hits as hypotheses, not signaling" - "Use a distance-aware method so adjacency is actually modeled" - "Check whether a short-range interaction is just segmentation spillover"

Example Prompts

Method Choice

"I have Xenium data -- pick a communication method that actually models spatial distance and explain why squidpy ligrec alone is not spatial."

"Run COMMOT on my spatial data and set the signaling range correctly for a secreted chemokine versus a contact-dependent Notch pair."

Honest Inference

"Run ligand-receptor analysis and correct for multiple testing across all pair and cell-type-pair combinations."

"Rank the communication hits as hypotheses and tell me what evidence would move each up the confidence ladder."

Auditing

"A T-cell-to-macrophage interaction came up at their shared boundary -- check whether this is transcript spillover from imperfect segmentation."

"My 300-gene panel returned no communication -- is that a real negative or a panel limitation?"

What the Agent Will Do

  1. Confirm cell-type annotations exist and decide whether the platform even supports the analysis (panel coverage of ligands/receptors).
  2. Choose a method by whether spatial distance must be modeled and by secreted-vs-contact range.
  3. Build or reuse the spatial neighbor graph in the correct coordinate units.
  4. Run the inference (squidpy ligrec baseline and/or distance-aware COMMOT), naming the database and version.
  5. Correct for multiple testing over the full pair-by-cell-type-pair space.
  6. Audit short-range hits for segmentation spillover and place survivors on the confidence ladder.

Tips