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

Spatial Multi-omics Integration - Usage Guide

Overview

Integrate spatial RNA with a second modality -- protein, chromatin accessibility, or a histone mark -- on spatial CITE-seq, DBiT-seq, spatial-ATAC, spatial epigenome-transcriptome, or Visium CytAssist data. The governing decision is whether the modalities were co-measured on the SAME pixels (vertical integration, where joint same-cell methods like MOFA and WNN apply) or come from different serial sections (diagonal integration, where the right operation is spatial registration of DISTINCT cell populations, not joint modeling).

Prerequisites

pip install muon mudata scanpy anndata
pip install mofapy2          # MOFA backend for muon.tl.mofa
pip install paste-bio        # serial-section registration (diagonal)

Quick Start

Example Prompts

Vertical (same-pixel co-profiling)

"I have spatial CITE-seq with whole-transcriptome RNA and an ADT protein panel on the same pixels -- build a joint embedding and find shared factors."

"Run MOFA on my DBiT-seq RNA + protein and tell me which factors are RNA-driven versus protein-driven."

"Cluster my Visium CytAssist Gene+Protein spots on a joint RNA+protein representation, accounting for the spot being multiple cells."

Diagonal (serial sections, different cells)

"My chromatin accessibility was measured on one section and RNA on the adjacent section -- align them into a common coordinate frame."

"Register my two serial spatial sections with PASTE and warn me about what a registered map does and does not mean."

Panel and regime checks

"Before I report that a cell type is absent in my spatial protein data, check whether its markers were even in the antibody panel."

"Tell me whether my platform co-captures both modalities on one pixel or runs adjacent sections, and which integration method that implies."

What the Agent Will Do

  1. Classify the integration regime: same-pixel (vertical) versus serial-section (diagonal).
  2. For vertical data, assemble a MuData with both modalities indexed on identical pixels and intersect observations.
  3. Normalize each modality in its own statistics (log1p for RNA counts, CLR/arcsinh for the targeted protein panel).
  4. Run a joint method (MOFA factors or WNN weighting) and inspect per-modality variance or weights.
  5. For diagonal data, compute an optimal-transport or diffeomorphic registration and stack sections on a shared frame, flagging that the result is a coordinate map of different cells.
  6. Disclose the pixel-is-not-a-cell and bounded-panel caveats in any biological conclusion.

Tips