scGPT-GBM-classifier

This repository contains two folders (final_analysis & testing_examples) using Richards 2021 data and Neftel 2019 data, respectively. Inside each folder, you will find two Jupyter notebooks for analyzing glioblastoma data using scGPT.

The workflow is as follows:

1. Embedding and Classification
   scGPT_Embedding_Tasks_GBM_Classifier.ipynb
   This notebook:

   • Connects to Google Drive and sets up the environment.
   • Downloads necessary scGPT model checkpoints and data from Google Drive.
   • Loads and preprocesses the GBM datasets.
   • Embeds reference and query data using scGPT.
   • Maps cell type annotations from a reference dataset to a query dataset.
   • Computes performance metrics and plots a confusion matrix of predicted vs. true labels.

2. UMAP Post-Processing
   scGPT_Post_Processing_UMAP_GBM_Classifier.ipynb
   This notebook performs the following:

   • Loads the concatenated embedded AnnData object from the first notebook.
   • Computes neighbourhood graphs, UMAP embeddings, and clustering.
   • Visualizes the results (e.g., UMAP plots with cell type labels).
   • Optionally, computes the Adjusted Rand Index (ARI) to evaluate clustering performance.

---

Requirements

Runtime
We recommend Google Colab with a runtime that supports Python ≤3.10. For best performance:

• Use L4 GPUs (or A100 for faster embedding tasks) with Google Colab Pro.
• If using the free version (T4 GPU), follow the instructions in the testing_examples notebooks.

Python Version

• The notebooks are designed for Python 3.10.12.
• IMPORTANT: In Colab, set your runtime to "fallback" to Python 3.10 if you see Python 3.11 was loaded.

Dependencies

• The notebooks install and upgrade packages such as scgpt, wandb, faiss-gpu, scanpy, torch, among others.
• For the full list for reproducibility, see the last cell of each notebook (i.e., pip freeze).

---

Setup & Usage

1. Open the Notebooks in Google Colab

2. Set Up the Runtime

   • Connect to a GPU runtime (L4 is recommended).
   • Use the Command Palette to select "Use fallback runtime version" so that the Python version is set to 3.10.12.

3. Mount Your Google Drive

   • Both notebooks require you to mount your Google Drive to read/write data and save generated files (e.g., embedded AnnData objects, UMAP coordinates).

4. Run Notebook Cells Sequentially

   • Start with the embedding notebook to generate the reference and query embeddings.
   • Then run the UMAP post-processing notebook to visualize the embeddings and evaluate the performance metrics (e.g., ARI).

5. Data Access

   • The notebooks download data from Google Drive. Make sure you have proper access or adjust the URLs if you use your own data.

---

External Data Sources

GBmap Datasets

• GBmap data was downloaded via CELLxGENE.
• Included in this analysis is the "Core" GBmap atlas.
• Read more: GBmap Pre-print.

Neftel 2019 Datasets

• Neftel 2019 data is available via Broad Single Cell Portal.
• Data for this analysis was filtered from the original GBmap data.
• Read more: Neftel 2019 Cell Publication.

Richards 2021 Dataset

• Richards 2019 data was downloaded and available via CReSCENT and Broad Single Cell Portal.
• Richards tumour data for this analysis was filtered from the original GBmap data and GSC data (glioblastoma stem cells) from CReSCENT.
• Read more: Richards 2021 Nature Cancer Publication.

---

Additional Notes

FAISS Installation

• The first notebook checks for faiss installation for fast similarity search. If it is not installed, you can follow the FAISS installation guide.

Customization

• You can modify cell type keys (annotation_level_3 or cell_type) and other parameters (like the number of nearest neighbors k) to suit your analysis.

scGPT

• This repository is based on the scGPT GitHub tutorial. Read more: scGPT 2024 Nature Methods Publication.

---

Contact

If you run into any problems, please contact Suluxan Mohanraj or leave a GitHub issue in this repository!