scPilot: Large Language Model Reasoning for Automated Single-Cell Analysis

The first systematic framework for omics-native reasoning.

📖 Overview

scPilot operates as a true scientific assistant that automates core single-cell analyses—Cell-Type Annotation, Trajectory Inference, and GRN Prediction.

Unlike traditional "tool agents" that simply write code, scPilot performs Omics-Native Reasoning: it directly inspects data summaries, explicitly articulates biological hypotheses, and iteratively refines its conclusions using on-demand bioinformatics tools.

Key Features

• 🔬 Biological Context First: Incorporates tissue, species, and experimental metadata into reasoning.
• 🔄 Iterative Refinement: Self-corrects hypotheses based on computational evidence (e.g., dotplots, marker genes).
• 📊 Transparent & Auditable: Generates full reasoning traces, not just black-box vectors.

---

🚀 Performance Highlights

Evaluated on scBench (9 expertly curated datasets), scPilot demonstrates superior performance over direct LLM prompting and traditional pipelines.

Task	Metric Improvement	Key Result
Cell-Type Annotation	+11% Accuracy	Iterative reasoning lifts average accuracy by 11% compared to one-shot methods.
Trajectory Inference	-30% Graph Error	Cuts trajectory graph-edit distance by 30% (using Gemini-2.5-Pro).
GRN Prediction	+0.03 AUROC	Improves Gene Regulatory Network prediction AUROC over baseline baselines.

💰 Cost Efficiency

scPilot is designed to be efficient. Average costs per run (using Gemini-2.5-Pro) are minimal:

• Cell-type annotation (Retina): ~$0.03 / run
• Trajectory inference (Neocortex): ~$0.04 / run
• GRN TF-gene prediction: ~$0.12 / run

---

🛠️ Installation & Setup

1. API Key Configuration

Replace OPENAI_API_KEY and GOOGLE_API_KEY in /config/settings.py.

2. Environment Setup

Refer to requirements.txt.

Note: numpy < 2.0 is required for py-Monocle.

3. Data Preparation

Download large file dependencies from Google Drive and place them in the scPilot/uploads/ folder.

---

🏃 Running scPilot Tasks

0. Configure LLM

In your config, set model_provider (openai/google) and model_name (e.g., gpt-4o, gemini-2.5-pro).

1. Cell Type Annotation

• Run: Task1_scPilot.py (scPilot version) or Task1_direct.py (Direct prompting).
• Config: Update CellTypeAnnotationDatasets.xlsx with your dataset details (grouping, species, etc.).

2. Trajectory Inference

• Run: Traj_scPilot_1.ipynb through 3.ipynb for scPilot; Traj_Direct_1.ipynb through 3.ipynb for Direct prompting.
• Note: Direct version notebooks may require manual tree copying.

3. GRN TF-Gene Prediction

• Run: Task3_combined.py (Generates both direct and scPilot predictions).
• Options: Change PREDICT_CONTEXT to "Liver", "Stomach", or "Kidney" in the config.

Current Results

Results used in the paper can be found in the google drive.

This result includes 3 tasks (Task1_results/, Task2_results/ and Task3_results/), along with additional experiment, ablation study, etc) in the rebuttal stage.

---

🔗 Citation

If you use scPilot, please cite our NeurIPS 2025 paper:

@inproceedings{
gao2025scpilot,
title={scPilot: Large Language Model Reasoning Toward Automated Single-Cell Analysis and Discovery},
author={Yiming Gao and Zhen Wang and Jefferson Chen and Mark Antkowiak and Mengzhou Hu and JungHo Kong and Dexter Pratt and Jieyuan Liu and Enze Ma and Zhiting Hu and Eric P. Xing},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year={2025},
url={https://openreview.net/forum?id=Vzi96rTe4w}
}