LLM-Derived Gene Semantic Embeddings Enable Spatiotemporal Functional Mapping of Mouse Embryonic Development

Overview

This repository contains the code and analysis pipeline for the paper "LLM-Derived Gene Semantic Embeddings Enable Spatiotemporal Functional Mapping of Mouse Embryonic Development". The study proposes a knowledge-driven computational framework that leverages large language model (LLM) embeddings to map the spatiotemporal functional organization of mouse embryonic development from spatial transcriptomics data.

Environment

• Python 3.9+
• Install dependencies: pip install -r requirements.txt
• R packages (for GO enrichment analysis): ClusterProfiler, org.Mm.eg.db

Data Sources

The spatial transcriptomics data used in this study is from:

Zhong et al., A single-cell Stereo-seq transcriptomic atlas of mouse organogenesis, 2023.

Raw data files are not included in this repository due to size (~5GB+). To reproduce the analysis, you need to obtain:

1. Expression matrices (per timepoint, 2000 HVGs): Place CSV files in data/expression_matrices/ as E9.5_matrix.csv ... E16.5_matrix.csv.
2. Gene descriptions: Download gene functional description texts and save as data/gene_descriptions.json.
3. LLM embeddings: Generate 2048-D embeddings using Qwen Embedding-v4 and save as data/gene_embeddings_2048.json.
4. Full h5ad file (~7.3GB): Obtain from the original Stereo-seq data source.

Repository Structure

NC-code/
├── README.md
├── requirements.txt
├── src/                                # Analysis scripts
│   ├── pipeline.py                     # Gene2Role pipeline orchestrator
│   ├── spearman.py                     # Spearman correlation → signed edgelist
│   ├── eeisp.py                        # EEISP edge inference
│   ├── split_cells.py                  # Split expression matrix by cell type
│   ├── merge_edgelist.py               # Merge multiple edgelists
│   ├── core50run.py                    # Core50 batch embedding runner
│   ├── run_embedding_generation.py     # Embedding generation wrapper
│   ├── run-14.5-16.5.py               # Cosine similarity & embedding analysis
│   └── submit_gene2role.sh             # HPC job submission script
├── notebooks/                          # Jupyter analysis notebooks (in order)
│   ├── 00_generate_expression_matrices.ipynb
│   ├── 01_data_preprocessing.ipynb
│   ├── 02_generate_topological_embeddings.ipynb
│   ├── 03_community_detection.ipynb
│   ├── 04_spatiotemporal_analysis.ipynb
│   └── 05_gene2role_embedding.ipynb
├── figures/                            # Paper figures
│   ├── fig1.png ~ fig11.png
│   └── supplementary/fig8.jpg
├── results/                            # Analysis output CSVs
│   ├── communities_for_analysis.csv
│   ├── communities_for_analysis_FIXED.csv
│   ├── module_sizes.csv
│   ├── cdf_plot_data.csv
│   ├── community_topological_evolution_distance.csv
│   └── gene2role_rankings/             # Per-timepoint gene centrality rankings
├── tools/
│   └── SignedS2V/                      # Signed network embedding tool
│       ├── README.md
│       └── src/
└── data/                               # (gitignored) Raw data, see Data Sources

Pipeline Workflow

1. Embedding Generation (notebooks/00-02): Generate 2048-D gene embeddings using Qwen LLM, compute cosine similarity, construct gene functional network.

2. Network Construction & Module Detection (notebooks/03): Optimize similarity threshold (0.76), detect communities via hybrid strategy (connected components + Leiden), annotate with GO enrichment.

3. Spatial Reconstruction (notebooks/04): Score cells across 92 functional modules, perform Leiden clustering on functional score vectors, map back to physical coordinates.

4. Stability Analysis (notebooks/04): Jaccard index tracking across resolution parameters to recapitulate developmental hierarchy.

5. Temporal Network Analysis (notebooks/05): Gene2Role signed network embedding on liver-specific modules to track hub switching and dynamic turnover.

Citation

If you use this code, please cite our paper:

@article{llmspatioemb2026,
  title={LLM-Derived Gene Semantic Embeddings Enable Spatiotemporal Functional Mapping of Mouse Embryonic Development},
  author={Anonymized for review}, % TODO: restore author list before publication
  year={2026}
}

License

This code is provided for academic research purposes.