iPSC_GPT_model_for_STATE

A SE-conditioned ST (State Transition) model trained on the GSE213069 iPSC perturbation dataset, designed for zero-shot virtual gene knockout (KO) prediction.

---

🧬 Model Overview

• Framework: ArcInstitute/STATE
• Architecture: SE-conditioned State Transition (SE → ST)
• Dataset: GSE213069 iPSC perturbation scRNA-seq
• Feature space: Top-5000 selected genes (variance + DE task score)
• Conditioning: Gene embeddings from SE-600M
• Task: Predict transcriptional changes following unseen gene perturbations (virtual KO)

This model enables zero-shot predictions for genes not explicitly perturbed in training, by conditioning on SE embeddings.

---

📂 Files in this repository

• ST-iPSC_GPT.ckpt
  Trained STATE checkpoint (~289 MB, best validation loss)
• config.yaml
  Training snapshot (Hydra/Lightning config)
• var_dims.pkl
  Feature order mapping (Top-5000 gene list used in training)
• SE-GSE213069.h5ad
  AnnData with WT/control cells and precomputed SE embeddings (obsm["X_se"])
• STATE_SE_ST_training.ipynb → Jupyter Notebook (task-oriented Top-5000 feature selection, splits, training pipeline)
• selected_top5000_scores.csv → Table of the top-5000 selected genes with their task-oriented scores (used for training feature space)
• Virtual_KO_Inference_SE_ST.ipynb → Jupyter Notebook for virtual gene knockout analysis
• README.md (this file)

---

📊 Downstream analysis

• Compute differential expression (DEG) between WT and KO cells

• Export ranked gene list

• Plot pluripotency marker genes (POU5F1/OCT4, SOX2, NANOG, KLF4, ESRRB, ZFP42) before vs after KO

Example scripts and notebooks are provided in the repo.

---

📖 Citation

If you use this model, please cite:

• STATE framework: Arc Institute (https://github.com/ArcInstitute/state)
• Dataset: GSE213069 (NCBI GEO)

---

⚠️ Notes

• This checkpoint is SE-conditioned. Ensure your input AnnData includes obsm["X_se"].
• Works best with genes present in the Top-5000 training feature set.
• Intended for research use (virtual gene perturbation, proof-of-concept) only.