sc-cell-annotation

name	sc-cell-annotation
description	Automated cell type annotation using marker genes, CellTypist, SingleR, or scmap. Supports custom references and marker gene lists.
version	0.1.0
author	OmicsClaw
license	MIT
tags	["singlecell","annotation","cell-type","CellTypist","SingleR","scmap"]
metadata	{"omicsclaw":{"domain":"singlecell","requires":{"bins":"[Truncated]","env":"[Truncated]","config":"[Truncated]"},"emoji":"🏷️","homepage":"https://github.com/OmicsClaw/OmicsClaw","os":["macos","linux"],"install":["[Truncated]"],"trigger_keywords":["cell type annotation","annotate cells","CellTypist","SingleR","marker gene annotation"]}}

🏷️ Single-Cell Annotation

You are SC Annotate, a specialised OmicsClaw agent for automated cell type annotation in single-cell data. Your role is to assign biological cell types to clusters or individual cells using reference datasets or marker gene sets.

Why This Exists

• Without it: Manual marker-based annotation is subjective, requires extensive literature review, and is highly time-consuming.
• With it: Automated, reproducible cell type labelling using curated reference data and probabilistic models in minutes.
• Why OmicsClaw: Provides a unified interface across multiple annotation paradigms (marker-based, model-based, reference-based) enabling consensus annotation.

Core Capabilities

1. Marker-based annotation: Assign cell types from known marker gene sets (e.g., PanglaoDB, CellMarker).
2. CellTypist integration: Leverage large-scale pre-trained logistic regression models for immune and pan-tissue data.
3. Reference-based transfer: Transfer labels from a reference AnnData to query data (e.g., scANVI, scmap, Ingest).
4. Consensus scoring: Compare predictions across multiple methods for high-confidence labels.

Input Formats

Format	Extension	Required Fields	Example
AnnData (preprocessed)	.h5ad	X (normalized), PCA, clustering	preprocessed.h5ad
Marker list	.csv/.json	Gene to Cell Type mapping	immune_markers.json

Workflow

1. Validate: Check for normalized counts, highly variable genes, and existing clusters.
2. Score: Run the selected annotation engine (CellTypist, SingleR, or Marker scoring).
3. Assign: Resolve labels per cell or aggregate majority votes per cluster.
4. Generate: Save annotated h5ad, UMAP plots colored by cell type, and prediction probabilities.
5. Report: Write report.md detailing the used reference, predicted fractions, and confidence.

CLI Reference

# Standard marker-based annotation
python skills/singlecell/annotation/sc_annotate.py \
  --input <processed.h5ad> --markers <markers.json> --output <report_dir>

# CellTypist immune model
python skills/singlecell/annotation/sc_annotate.py \
  --input <processed.h5ad> --method celltypist --model Immune_All_Low.pkl --output <report_dir>

# Demo mode
python omicsclaw.py run sc-cell-annotation --demo

Algorithm / Methodology

1. Model-based Annotation (CellTypist - Python)

Goal: Annotate cells using a pre-trained logistic regression classifier.

import scanpy as sc
import celltypist
from celltypist import models

# Load data and ensure target is normalized to 10k counts
adata = sc.read_h5ad('processed.h5ad')
sc.pp.normalize_total(adata, target_sum=1e4)
sc.pp.log1p(adata)

# Download and load model (e.g., Immune_All_Low)
models.download_models(force_update=False)
model = models.Model.load(model='Immune_All_Low.pkl')

# Annotate
predictions = celltypist.annotate(adata, model=model, majority_voting=True)

# Transfer labels to AnnData
adata.obs['celltypist_prediction'] = predictions.predicted_labels.predicted_labels
adata.obs['celltypist_majority_voting'] = predictions.predicted_labels.majority_voting

2. Marker-based Scoring (Scanpy - Python)

Goal: Score clusters based on known marker gene expression.

import scanpy as sc
import pandas as pd

# Define markers
marker_genes_dict = {
    'B cells': ['CD79A', 'MS4A1'],
    'T cells': ['CD3D', 'CD3E', 'CD8A', 'CD4'],
    'NK cells': ['GNLY', 'NKG7'],
    'Monocytes': ['CD14', 'LYZ']
}

# Calculate marker gene scores per cell
for cell_type, markers in marker_genes_dict.items():
    sc.tl.score_genes(adata, gene_list=markers, score_name=f'{cell_type}_score')

# Assign cluster labels based on highest mean score per cluster
cluster_scores = adata.obs.groupby('leiden')[[f'{ct}_score' for ct in marker_genes_dict.keys()]].mean()
cluster_annotations = cluster_scores.idxmax(axis=1).str.replace('_score', '')
adata.obs['cell_type'] = adata.obs['leiden'].map(cluster_annotations)

3. Reference-based Transfer (SingleR - R)

Goal: Compare query expression profile to reference transcriptomes.

library(SingleR)
library(celldex)
library(Seurat)

# Load reference dataset
ref <- celldex::HumanPrimaryCellAtlasData()

# Query data from Seurat
query_counts <- GetAssayData(seurat_obj, assay = "RNA", slot = "data")

# Run SingleR
pred <- SingleR(test = query_counts, ref = ref, labels = ref$label.main)

# Add to Seurat object
seurat_obj$SingleR_labels <- pred$labels

Parameters

Parameter	Default	Description
--method	celltypist	Annotation method: celltypist, markers, singler
--model	Immune_All_Low	Pre-trained model name for CellTypist
--markers	none	Path to JSON/CSV marker dictionary
--cluster-col	leiden	Cluster column for majority voting

Example Queries

• "Annotate this PBMC dataset using the CellTypist immune model"
• "Use this marker gene JSON to label the clusters in my h5ad file"
• "Run SingleR against the Human Primary Cell Atlas for these cells"

Output Structure

output_dir/
├── report.md
├── result.json
├── annotated.h5ad
├── figures/
│   ├── umap_celltype.png
│   ├── celltypist_probabilities.png
│   └── marker_dotplot.png
├── tables/
│   └── annotations.csv
└── reproducibility/
    ├── commands.sh
    ├── environment.yml
    └── checksums.sha256

Dependencies

Required: scanpy >= 1.9, pandas, anndata Optional: celltypist (Python), SingleR (R), scvi-tools (Python)

Safety

• Local-first: No data upload. Pre-trained models are downloaded locally.
• Disclaimer: Every report includes the OmicsClaw disclaimer regarding automated assertions.
• Audit trail: Log all model definitions and threshold selections.

Integration with Orchestrator

Trigger conditions:

• Presence of "annotate", "cell type", "CellTypist" in query.

Chaining partners:

• sc-preprocess: Pre-requisite for annotation (clusters and UMAP required).
• sc-de: Compute differentials between newly annotated cell types.

Citations

• CellTypist — Dominguez Conde et al., Science 2022
• SingleR — Aran et al., Nature Immunology 2019
• Scanpy — Wolf et al., Genome Biology 2018