LLM-scCurator

Data-centric feature distillation for robust zero-shot cell-type annotation.

• Docs: https://llm-sccurator.readthedocs.io/
• Paper reproducibility: Zenodo v0.1.0 (DOI: 10.5281/zenodo.17970494).
• Cite: bioRxiv preprint (DOI: 10.64898/2025.12.28.696778). Newer tags (see GitHub Releases) add usability features without changing paper benchmarks.

---

🚀 Overview

LLM-scCurator is a Large Language Model–based curator for single-cell and spatial transcriptomics. It performs noise-aware marker distillation—suppressing technical programs (e.g., ribosomal/mitochondrial), clonotype signals (TCR/Ig), and stress signatures while rescuing lineage markers—and applies leakage-safe lineage filters before prompting an LLM. It supports hierarchical annotation (coarse-to-fine clustering and labeling) for single-cell and spatial transcriptomics data.

Key Features

• 🛡️ Noise-aware filtering: Automatically removes lineage-specific noise (TCR/Ig) and state-dependent noise (ribosomal/mitochondrial).
• 🧠 Context-aware inference: Automatically infers lineage context (e.g., "T cell") to guide LLM reasoning.
• 🔬 Hierarchical discovery: One-line function to dissect complex tissues into major lineages and fine-grained subtypes.
• 🌍 Spatial ready: Validated on scRNA-seq (10x) and spatial transcriptomics (Xenium, Visium).
• 🔒 Privacy-first, institutional-ready: Feature distillation runs locally; annotation works with cloud or local LLM backends, or institution-approved chat UIs (no tool-side API calls).

Fig. 1a. Overview of LLM-scCurator: data-centric, adaptive feature distillation recovers identity markers despite biological-noise programs (e.g., ribosomal, cell-cycle, stress). (bioRxiv preprint)

---

📦 Installation

• Option A (recommended): Install from PyPI

  pip install llm-sc-curator
  

  (See PyPI project page: https://pypi.org/project/llm-sc-curator/)

• Option B: Install from GitHub (development)

  # 1. Clone the repository
  git clone https://github.com/kenflab/LLM-scCurator.git
  
  # 2. Navigate to the directory
  cd LLM-scCurator
  
  # 3. Install the package (and dependencies)
  pip install .

Notes:

If you already have a Scanpy/Seurat pipeline environment, you can install it into that environment.

---

🐳 Docker (official environment)

We provide an official Docker environment (Python + R + Jupyter), sufficient to run LLM-scCurator and most paper figure generation.
Optionally includes Ollama for local LLM annotation (no cloud API key required).

• Option A: Prebuilt image (recommended)

  Use the published image from GitHub Container Registry (GHCR).

  # from the repo root (optional, for notebooks / file access)
  docker pull ghcr.io/kenflab/llm-sc-curator:official

  Run Jupyter:

  docker run --rm -it \
    -p 8888:8888 \
    -v "$PWD":/work \
    -e GEMINI_API_KEY \
    -e OPENAI_API_KEY \
    ghcr.io/kenflab/llm-sc-curator:official
  

  Open Jupyter: http://localhost:8888
  
  (Use the token printed in the container logs.)
  
  Notes:

  For manuscript reproducibility, we also provide versioned tags (e.g., :0.1.0). Prefer a version tag when matching a paper release.

• Option B: Build locally (development)

  • Option B-1: Build locally with Compose (recommended for dev)

    # from the repo root
    docker compose -f docker/docker-compose.yml build
    docker compose -f docker/docker-compose.yml up

    B-1.1) Open Jupyter

    • http://localhost:8888 Workspace mount: /work

    B-1.2) If prompted for "Password or token"

    • Get the tokenized URL from container logs:

      docker compose -f docker/docker-compose.yml logs -f llm-sc-curator

    • Then either:
      • open the printed URL (contains ?token=...) in your browser, or
      • paste the token value into the login prompt.

  • Option B-2: Build locally without Compose (alternative)

    # from the repo root
    docker build -f docker/Dockerfile -t llm-sc-curator:official .

    B-2.1) Run Jupyter

    docker run --rm -it \
      -p 8888:8888 \
      -v "$PWD":/work \
      -e GEMINI_API_KEY \
      -e OPENAI_API_KEY \
      llm-sc-curator:official

    B-2.2) Open Jupyter

    • http://localhost:8888 Workspace mount: /work

---

🖥️ Apptainer / Singularity (HPC)

• Option A: Prebuilt image (recommended)

  Use the published image from GitHub Container Registry (GHCR).

  apptainer build llm-sc-curator.sif docker://ghcr.io/kenflab/llm-sc-curator:official

• Option B: a .sif from the Docker image (development)

  docker compose -f docker/docker-compose.yml build
  apptainer build llm-sc-curator.sif docker-daemon://llm-sc-curator:official

Run Jupyter (either image):

apptainer exec --cleanenv \
  --bind "$PWD":/work \
  llm-sc-curator.sif \
  bash -lc 'jupyter lab --ip=0.0.0.0 --port=8888 --no-browser 

---

🔒 Privacy

We respect the sensitivity of clinical and biological data. LLM-scCurator is designed so that raw expression matrices and cell-level metadata can remain within your local environment.

• Local execution: Preprocessing, confounder masking, and feature ranking run locally on your machine.
• Minimal transmission (optional): If you choose to use an external LLM API, only anonymized, cluster-level marker lists (e.g., top 50 gene symbols) and minimal tissue context are sent.
• User control: You decide what additional context (e.g., disease state, treatment, platform) to include. Always follow institutional policy and the LLM provider’s terms before sharing any information.

Institutional-friendly workflows

Many institutions restrict which AI tools can be used with internal clinical or research datasets. To support these real-world constraints, we provide two end-to-end workflows that keep raw matrices and cell-level metadata local and avoid external LLM API calls unless explicitly permitted:

• Fully local LLM (Ollama): Curate features and optionally annotate clusters using a local LLM backend (no external transmission). examples/local/local_quickstart_ollama.ipynb

• Local feature distillation → Approved chat LLM annotation (no external LLM API calls): Curate features locally, export a curated cluster→genes table, then annotate it via an institution-approved chat interface (e.g., Microsoft Copilot “Work”) by uploading the CSV/Excel or pasting markers. examples/local/local_quickstart_approved_ai_workflow.ipynb

---

⚡ Quick Start

Documentation (—Getting started, Concepts, User guide, and the full API reference—): https://llm-sccurator.readthedocs.io/

🐍 For Python / Scanpy Users

1. Set your API key (simplest: paste in the notebook)

import scanpy as sc
from llm_sc_curator import LLMscCurator

GEMINI_API_KEY = "PASTE_YOUR_KEY_HERE"
# OPENAI_API_KEY = "PASTE_YOUR_KEY_HERE"  # optional

# Load your data
adata = sc.read_h5ad("my_data.h5ad")
  
# Initialize with your API Key (Google AI Studio)
curator = LLMscCurator(api_key=GEMINI_API_KEY)
curator.set_global_context(adata)

2. Run LLM-scCurator

• Option A: hierarchical discovery mode(iterative coarse-to-fine clustering and labeling)

  # Fully automated hierarchical annotation (includes clustering)
  adata = curator.run_hierarchical_discovery(adata)
  
  # Visualize
  sc.pl.umap(adata, color=['major_type', 'fine_type'])

• Option B: Annotate your existing clusters (cluster → table/CSV → per-cell labels)
  Use this when you already have clusters (e.g., Seurat seurat_clusters, Leiden, etc.) and want to annotate each cluster once, then propagate labels to cells.

  # v0.1.1+
  from llm_sc_curator import (
      export_cluster_annotation_table,
      apply_cluster_map_to_cells,
  )
  
  cluster_col = "seurat_clusters"  # change if needed
  
  # 1) Annotate each cluster (once)
  clusters = sorted(adata.obs[cluster_col].astype(str).unique())
  cluster_results = {}
  genes_by_cluster = {}
  
  for cl in clusters:
      genes = curator.curate_features(
          adata,
          group_col=cluster_col,
          target_group=str(cl),
          use_statistics=True,
      )
      genes_by_cluster[str(cl)] = genes or []
  
      if genes:
          cluster_results[str(cl)] = curator.annotate(genes, use_auto_context=True)
      else:
          cluster_results[str(cl)] = {
              "cell_type": "NoGenes",
              "confidence": "Low",
              "reasoning": "Curated gene list empty",
          }
  
  # 2) Export a shareable cluster table (CSV/DataFrame)
  df_cluster = export_cluster_annotation_table(
      adata,
      cluster_col=cluster_col,
      cluster_results=cluster_results,
      genes_by_cluster=genes_by_cluster,
      prefix="Curated",
  )
  df_cluster.to_csv("cluster_curated_map.csv", index=False)
  
  # 3) Propagate cluster labels to per-cell labels
  apply_cluster_map_to_cells(
      adata,
      cluster_col=cluster_col,
      df_cluster=df_cluster,
      label_col="Curated_CellType",
      new_col="Curated_CellType",
  )

Notes:

Manuscript results correspond to v0.1.0; later minor releases add user-facing utilities without changing core behavior.

📊 For R / Seurat Users

You can use LLM-scCurator in two ways:

• Option A (recommended): Export → run in Python We provide a helper script examples/R/export_to_curator.R to export your Seurat object seamlessly for processing in Python.

  source("examples/R/export_to_curator.R")
  Rscript examples/R/export_to_curator.R \
    --in_rds path/to/seurat_object.rds \
    --outdir out_seurat \
    --cluster_col seurat_clusters

  Output:

  • counts.mtx (raw counts; recommended)
  • features.tsv (gene list)
  • obs.csv (cell metadata; includes seurat_clusters)
  • umap.csv (optional, if available)

  Notes:

  • The folder will contain: counts.mtx, features.tsv, obs.csv (and umap.csv if available).
  • Then continue in the Python/Colab tutorial to run LLM-scCurator and write cluster_curated_map.csv,
  • which can be re-imported into Seurat for plotting.

• Option B: Run from R via reticulate (advanced)

  If you prefer to stay in R, you can invoke the Python package via reticulate (Python-in-R). This is more sensitive to Python environment configuration, so we recommend Option A for most users.

  • Use the official Docker (Python + R + Jupyter) and follow the step-by-step tutorial notebook: 📓 examples/R/run_llm_sccurator_R_reticulate.ipynb
    The notebook includes:
    • Use LLM-scCurator for robust marker distillation (no API key required)
    • 🔑 Optional: For annotation, use Gemini/OpenAI APIs (API key required) or Ollama (no API key) .

---

📄 Manuscript reproduction

For manuscript-facing verification (benchmarks, figures, and Source Data), use the versioned assets under paper/. See paper/README.md for the primary instructions.

Notes:

• Figures are supported by exported Source Data in paper/source_data/ (see paper/FIGURE_MAP.csv for panel → file mapping).
• Re-running LLM/API calls or external reference annotators is optional; LLM API outputs may vary across runs even with temperature=0.
• For transparency, we include read-only provenance notebooks with example run logs in paper/notebooks/

---

📓 Colab notebooks

• Python / Scanpy quickstart (recommended: colab_quickstart.ipynb)

  • 
    ☝️ Runs end-to-end on a public Scanpy dataset (no API key required by default).

    • 🔑 Optional: If an API key is provided (replace GEMINI_API_KEY = "YOUR_KEY_HERE"), the notebook can also run LLM-scCurator automatic hierarchical cell annotation.

  • OpenAI quickstart (OpenAI backend: colab_quickstart_openai.ipynb)

  • 
    ☝️ Same workflow as the Python / Scanpy quickstart, but configured for the OpenAI backend.

    • 🔑 Optional: If an API key is provided (replace OPENAI_API_KEY= "YOUR_KEY_HERE"), the notebook can also run LLM-scCurator automatic hierarchical cell annotation. OPENAI_API_KEY requires OpenAI API billing (paid API credits).

• R / Seurat quickstart (export → Python LLM-scCurator → back to Seurat: colab_quickstart_R.ipynb)

  • 
    ☝️ Runs a minimal Seurat workflow in R, exports a Seurat object to an AnnData-ready folder, runs LLM-scCurator in Python, then re-imports labels into Seurat for visualization and marker sanity checks.
    • 🔑 Optional: Requires an API key for LLM-scCurator annotation (same setup as above).
    • Recommended for Seurat users who want to keep Seurat clustering/UMAP but use LLM-scCurator for robust marker distillation and annotation.

---

🔑 Backends setup (API keys or local Ollama)

LLM-scCurator supports both cloud LLM APIs (Gemini / OpenAI) and a local LLM backend (Ollama).
No manual installation is required: the official Docker environment already includes LLM-scCurator and its dependencies. If you use the local Ollama backend, no API key is needed.

Set your provider API key as an environment variable (Cloud LLM APIs):

• GEMINI_API_KEY for Google Gemini
• OPENAI_API_KEY for OpenAI API

See each provider’s documentation for how to obtain an API key and for current usage policies.

• Option A (Gemini steps):
  A-1. Go to Google AI Studio.
  A-2. Log in with your Google Account.
  A-3. Click Get API key (top-left) $\rightarrow$ Create API key.
  A-4. Copy the key and use it in your code.
  

• Option B (OpenAI steps):
  B-1. Go to OpenAI Platform.
  B-2. Log in with your OpenAI Account.
  B-3. Click Create new secret key $\rightarrow$ Create secret key.
  B-4. Copy the key and use it in your code.

Notes:

Google Gemini can be used within its free-tier limits.
OpenAI API usage requires enabling billing (paid API credits); ChatGPT subscriptions (e.g. Plus) do NOT include API usage.

---

Citation

• bioRxiv preprint: 10.64898/2025.12.28.696778
• Zenodo archive (v0.1.0): 10.5281/zenodo.17970494
• GitHub release tag: v0.1.0