Single-Cell RNA-seq Foundation Model Project

This project implements and compares various deep learning foundation models for single-cell RNA sequencing (scRNA-seq) analysis, focusing on trajectory inference and cell state generation across three biological processes: epithelial-mesenchymal transition (EMT), hematopoiesis, and thymocyte development.

Project Structure

fm-project/
├── data/                   # Datasets and generated samples
├── experiments/            # Jupyter notebooks for model experiments
├── models/                 # Saved model checkpoints
├── src/                    # Source code implementations
├── utils/                  # Utility functions

Directory Contents

🧪 experiments/ - Model Training and Evaluation Notebooks

Jupyter notebooks for preprocessing data and running different models:

Data Preprocessing:

• emt_preprocess.ipynb - EMT dataset preprocessing
• hematopoiesis_preprocess.ipynb - Hematopoiesis dataset preprocessing
• thymocyte_preprocess.ipynb - Thymocyte dataset preprocessing

Model Experiments:

Each model has dedicated notebooks for each dataset:

• scNODE: scnode_[dataset].ipynb - Neural ODE-based generative model
• scDiffusion: scdiff_[dataset].ipynb - Diffusion model for cell generation
• scGPT: scgpt_[dataset].ipynb - GPT-based foundation model
• scVI: scvi_[dataset].ipynb - Variational inference model

💻 src/ - Source Code Implementations

Contains the implementation of two main models:

src/scNODE/ - Neural ODE Model Implementation

Cloned from https://github.com/rsinghlab/scNODE

src/scDiffusion/ - Diffusion Model Implementation

Cloned from https://github.com/EperLuo/scDiffusion

🔧 utils/ - Utility Functions

Helper functions for data processing and evaluation:

• __init__.py - Package initialization
• adata.py - AnnData object utilities
• evaluation.py - Evaluation metrics including marker gene monotonicity
• latent.py - Latent space analysis utilities
• plot.py - Plotting utilities