omics-multiome

name	omics-multiome
description	Single-cell multiome (paired RNA + ATAC) — MuData assembly, per-modality preprocess, joint embedding (WNN / MultiVI), joint clustering & annotation, chromatin-aware velocity (MultiVelo), eGRN (SCENIC+), cross-modal interpretation.
requiredTools	["run_python","create_notebook","add_cell","observe_figure","omics_preflight","omics_runtime"]
evidencePolicy	required
outputSchema	grounded_response
minConfidence	medium
tags	["multiome","multimodal","rna","atac","scverse","mudata","muon","single-cell"]
extends	omics-shared

Single-Cell Multiome (RNA + ATAC) Analysis

Paired RNA + ATAC from the same nuclei. Reuses the scRNA and scATAC per-modality recipes, adds joint modeling. Run compute through the omics_runtime tool with modality="multiome" (dispatches into the pinned task3 env, records evidence).

Prerequisites

1. Preflight: omics_preflight(modality="multiome") must pass
2. Shared foundation: Load omics-shared first
3. Data format: MuData with paired cells, or separate RNA/ATAC AnnData to assemble
4. Per-modality knowledge: Familiar with the scRNA and scATAC recipes

Capability Menu

Capability	When to Use	Documentation
Load & Assemble	Build validated paired MuData from various sources	method/load_multiome.md
Per-Modality Preprocess	Reuse P1/P3 recipes on each modality before joining	method/per_modality_preprocess.md
Joint Embedding	Combine modalities into one cell representation (WNN/MultiVI)	method/joint_embedding.md
Joint Cluster & Annotate	Cluster on joint rep; label via two routes (marker+LLM / reference)	method/joint_cluster_annotate.md
Multiome Velocity	Chromatin-informed RNA velocity (MultiVelo, external env)	method/multiome_velocity.md
eGRN & Regulation	TF→region→gene eGRN + region-gene linkage (SCENIC+, external pipeline)	method/regulation.md
Cross-Modal Interpretation	Modality weights (which modality drives a population); peak-gene & eGRN → regulation.md	method/cross_modality.md

Global Rules

1. Preflight first: omics_preflight(modality="multiome") before compute
2. Reuse scRNA/scATAC: Don't re-derive uni-modal steps - point to those recipes
3. Ground every claim: All quantitative results → EvidenceRecord
4. Observe figures: observe_figure on every plot before conclusions
5. Earn joint embedding: Compare against single-modality baselines (ARI/NMI) - only claim improvement if measured
6. Abstain over fabricate: Missing data/deps → structured blocker, not guess
7. Heavy methods expensive: MultiVI/MultiVelo/SCENIC+ - start small, sanity-check, scale
8. Write to workdir: runs/omics/<id>/ for reproducibility

Container Structure

MuData with modalities under .mod:

mdata.mod['rna']   # RNA AnnData
mdata.mod['atac']  # ATAC AnnData
mdata.obs          # Shared cell metadata

Per conventions (§6): counts in layers["counts"], embeddings in obsm["X_*"], clusters in obs["leiden"], labels in obs["cell_type"]

Standard Workflow

1. Load/Assemble → validated MuData with shared barcodes
2. Per-Modality Preprocess → RNA (QC/norm/HVG/PCA), ATAC (QC/features/embedding)
3. Joint Embedding → WNN or MultiVI latent space
4. Joint Cluster → Leiden on joint representation
5. Annotate → marker+LLM or FM pipeline
6. Cross-Modal Analysis → peak-gene links, TF activity, regulatory interpretation

Runtime Dispatch

Run compute through the omics_runtime tool with modality="multiome":

# Available subcommands:
#   - load_multiome: Assemble/validate paired MuData

Joint embedding (WNN / MultiVI) is hand-rolled muon in run_python; velocity (MultiVelo) and eGRN (SCENIC+) run as external pipelines in isolated envs — see the method docs.