YuhanQiu/ESC_GPT_model_for_scLAMBDA

scLAMBDA — hESC Perturbation Model (PARK7 example)

This repository hosts a filtered training dataset and model files for running scLAMBDA on hESC (human embryonic stem cells), based on Perturb-seq dataset GSE217460.
The goal is to enable zero-shot or fine-tuned prediction of gene perturbations (e.g., PARK7 overexpress) on hESC transcriptomes.

---

📂 Contents

• GSE217460_GPT.h5ad
  Single-cell RNA-seq dataset after filtering, retaining only genes covered by GenePT text embeddings (text-embedding-3-large).
  This ensures full compatibility with scLAMBDA training and inference.

• scLAMBDA_hESC_GPT.ipynb
  Full training and inference pipeline (environment setup → model initialization → training → perturbation prediction → visualization).

• embedding_coverage_report.csv
  Report showing which genes in the hESC dataset are covered by GenePT embeddings, and the overall coverage rate.

• ckpt.pth
  Checkpoint files from fine-tuning scLAMBDA on the hESC dataset.
  If included, they can be directly loaded for downstream prediction.

• GenePT_gene_protein_embedding_model_3_text.pickle
  Pretrained gene/protein embeddings generated by GenePT (dimension = 3072). Used as input embeddings for perturbation modeling and prediction.

• Fast_prediction.ipynb
  Simple prediction script without training.

---

🧬 Dataset details

• Source: GSE217460 (Perturb-seq, hESC)
• Cells: ~200k hESC cells (CTRL + ~1700 perturbations)
• Genes: ~20k original → filtered to ~20k genes with valid embeddings
• Format: .h5ad (AnnData, Scanpy-compatible)

---

⚙️ Training workflow (summary)

Steps to reproduce the model training (see scLAMBDA):

1. Environment

   • GPU: RTX4090 (24GB)
   • PyTorch 2.1.2 (CUDA 11.8)
   • Python 3.10

2. Filter dataset

   adata = sc.read_h5ad("raw_data.h5ad")
   covered = [g for g in adata.var_names if g in gene_embeddings]
   adata = adata[:, covered].copy()
   adata.write_h5ad("adata_filtered_for_embedding_covered_genes.h5ad")
   

3. Train scLAMBDA

   model = sclambda.model.Model(
       adata,
       gene_embeddings,
       training_epochs=500,
       batch_size=800,
       latent_dim=50,
       hidden_dim=1024,
       lambda_MI=500,
       multi_gene=False
   )
   model.train()
   

4. Predict perturbations (example: PARK7)

   res = model.predict(["PARK7"], return_type="cells")
   

---

🔗 References

• scLAMBDA: https://github.com/gefeiwang/scLAMBDA
• Perturb-seq dataset (GSE217460): NCBI GEO
• GenePT embeddings: Zenodo resource

---

📝 Citation

If you use this resource, please cite:

• Wang et al. scLAMBDA: Language Models for Predicting Gene Perturbation Effects in Single Cells (GitHub: gefeiwang/scLAMBDA)

• Gefei Wang, Tianyu Liu, Jia Zhao, Youshu Cheng, Hongyu Zhao. Modeling and predicting single-cell multi-gene perturbation responses with scLAMBDA. bioRxiv 2024; doi: https://doi.org/10.1101/2024.12.04.626878.