scGPT Classifier

This project involves using scGPT, a foundation model for single-cell genomics, to generate cell embeddings from single-cell RNA-seq (scRNA-seq) data.

🎯 Goal: Build a classifier capable of annotating unknown liver endothelial cells by mapping them to an annotated reference dataset.

---

📓 Notebooks Overview

This repository contains several Jupyter notebooks designed for preprocessing, embedding generation, and visualization of single-cell data using scGPT. Below is a brief description of each:

---

1️⃣ scGPT_zero_shot_embedding_ramachandran.ipynb

🔍 Goal: Explore scGPT’s ability to annotate unknown cell labels in the endothelial zone of the human liver.

• Uses Ramachandran dataset (GSE136103) as query.
• Uses MacParland dataset (GSE115469) as reference.
• Based on the official scGPT zero-shot tutorial.

---

2️⃣ preprocess_ramachandran.ipynb

⚙️ Goal: Preprocess raw liver samples into usable H5AD format.

• Input: Raw.TAR file with healthy + diseased liver samples.
• Output: Processed H5AD files, ready for scGPT embedding generation.
• Dataset: Ramachandran (GSE136103).

---

3️⃣ scGPT_reference_embedding.ipynb

🧬 Goal: Generate embeddings and evaluate clustering performance.

• Dataset: MacParland (GSE115469).
• Uses cell annotations as labels.
• Tests how well scGPT groups similar cells without fine-tuning.

---

4️⃣ scGPT_UMAPs.ipynb

🗺 Goal: Visualize embeddings.

• Produces UMAP plots for the MacParland dataset embeddings.

---

5️⃣ Rds_to_h5ad.ipynb

🔄 Goal: Convert data formats.

• Converts Seurat .Rds objects → H5AD format (required for scGPT).

---