Deep Learning for Single-Cell RNA-seq Cell Type Classification

Benchmarking five deep learning architectures for automated cell type annotation from scRNA-seq data, evaluated on the TOSICA benchmark (Chen et al., 2023).

Key Results

Dataset	Best Model	Accuracy	Benchmark Rank
hPancreas (14 classes)	GNN	97.3%	3rd of 24 methods
mPancreas (21 classes)	Transformer	78.9%	4th of 19 methods
mAtlas (120 classes)	MLP	84.1%	1st of 21 methods

Architectures

Model	Key Mechanism	Params
MLP	Squeeze-and-excitation channel attention	2.6M
1D CNN	Residual convolutions + SE + dual-pool head	58K
GraphSAGE GNN	Cosine k-NN graph, transductive learning	2.9M
Transformer	Reactome pathway-masked attention with [CLS] token	144M
TOSICA	Published pathway-masked transformer (Chen et al.)	~144M

All models use cross-entropy loss, SGD with cosine annealing, and 10K highly variable genes.

Project Structure

.
├── 1_download.py                # Download and preprocess datasets
├── 2_visualize.py               # Dataset visualization (UMAP, PCA, etc.)
├── 3_MLP.py                     # Train MLP
├── 3_CNN.py                     # Train 1D CNN
├── 3_GNN.py                     # Train GraphSAGE GNN
├── 3_Transformer.py             # Train pathway-masked Transformer
├── 3_TOSICA.py                  # Train using original TOSICA library
├── 5_hPancreas.py               # Evaluate all models on hPancreas
├── 5_mPancreas.py               # Evaluate all models on mPancreas
├── 5_mAtlas.py                  # Evaluate all models on mAtlas
├── 6_figures.py                 # Generate paper figures
├── allen_brain/
│   ├── models/                  # Model definitions and training loop
│   │   ├── CellTypeMLP.py
│   │   ├── CellTypeCNN.py
│   │   ├── CellTypeGNN.py
│   │   ├── CellTypeAttention.py # Pathway-masked Transformer
│   │   ├── train.py             # Shared training infrastructure
│   │   ├── gnn_train.py         # GNN-specific training (transductive)
│   │   ├── losses.py            # Cross-entropy and focal loss
│   │   ├── blocks.py            # SE block
│   │   └── config.py            # Hyperparameter configs
│   ├── TOSICA/                  # Original TOSICA implementation
│   ├── cell_data/               # Data loading and preprocessing
│   └── data_sets/               # Per-dataset download scripts
├── hyperparametertuning/        # Optuna tuning scripts
├── tests/                       # Unit tests
├── figures/                     # Generated figures
└── data/                        # Downloaded datasets (not tracked)

Setup

pip install torch torchvision torch-geometric
pip install anndata scanpy optuna rich scikit-learn scipy

Usage

# 1. Download datasets
python 1_download.py

# 2. Train models (example: MLP on mPancreas)
python 3_MLP.py

# 3. Evaluate on all datasets
python 5_hPancreas.py
python 5_mPancreas.py
python 5_mAtlas.py

# 4. Generate figures
python 6_figures.py

Datasets

Three datasets from the TOSICA benchmark with condition-based train/test splits:

Dataset	Species	Train	Test	Classes	Split Criterion
hPancreas	Human	9,540	4,218	14	Study of origin
mPancreas	Mouse	22,918	10,886	21	Dev. day != 15.5
mAtlas	Mouse	30,624	76,797	120	Biological condition

Code Availability

The source code for this project is available at github.com/magana272/CellTypeClassification.

References

Chen, J., Xu, H., Tao, W. et al. Transformer for one stop interpretable cell type annotation. Nature Communications 14, 223 (2023). https://doi.org/10.1038/s41467-023-35923-4