doNet

DeepOMAPNet is a deep learning framework for multi-modal single-cell analysis of CITE-seq data. Given RNA expression profiles, it simultaneously predicts surface protein (ADT) levels, classifies cell types, and performs disease diagnosis (AML vs. Normal). It does this by combining Graph Attention Networks (GAT) with cross-modal Transformer Fusion.

---

What problem does it solve?

CITE-seq experiments measure both RNA and surface protein (ADT) levels in the same cell, but protein measurements are expensive and noisy. DeepOMAPNet learns to predict protein abundance directly from RNA expression while exploiting the cell-cell neighborhood graph structure. It also jointly classifies whether a cell belongs to an AML or Normal sample, sharing representations across both tasks.

---

How to use it

Start with Tutorials/Training.ipynb. It walks through the complete pipeline end-to-end: data loading, normalization, graph construction, model training, and evaluation. This is the recommended starting point for all users.

jupyter notebook Tutorials/Training.ipynb

---

Architecture

The model processes cells as nodes in a k-nearest-neighbor graph built in PCA space.

Raw CITE-seq AnnData (RNA + ADT)
    --> CLR normalization (ADT) + Z-score normalization (both modalities)
    --> Train / val / test split (70 / 15 / 15, stratified)
    --> PCA (50 components) + k-NN graph (k=15) + Leiden clustering
    --> PyTorch Geometric Data objects
    --> GATWithTransformerFusion
    --> ADT predictions  |  AML classification  |  Fused cell embeddings

Model components (scripts/model/doNET.py)

Component	Description
GATWithTransformerFusion	Top-level model: GAT encoder followed by TransformerFusion and task-specific output heads
GraphPositionalEncoding	Enriches node embeddings with graph topology signals (node degree, clustering coefficient)
SparseCrossAttentionLayer	Sparse multi-head cross-attention operating over edge lists; scales to large graphs (>100k cells)
CrossAttentionLayer	Dense cross-attention variant with layer norm; suitable for smaller graphs
AdapterLayer	Bottleneck adapter (dim -> dim/r -> dim) for parameter-efficient fine-tuning without retraining the full model
TransformerFusion	Stacks multiple cross-attention layers to fuse RNA and ADT modalities bidirectionally

The model produces three outputs:

• adt_pred — predicted protein levels per cell (regression)
• aml_pred — binary disease classification score per cell
• fused_embeddings — latent cell representations usable for downstream analysis (UMAP, clustering)

Training (scripts/trainer/gat_trainer.py)

Training is handled by train_gat_transformer_fusion(), which provides:

• Multi-task loss: MSE for ADT regression + BCE for AML classification, with an optional cross-entropy term for cell-type classification
• Stratified train/val/test splits
• Automatic mixed precision (AMP) and gradient accumulation
• Early stopping with best-model restoration

---

Installation

Conda (recommended)

git clone https://github.com/SreeSatyaGit/DeepOMAPNet.git
cd DeepOMAPNet
conda env create -f environment.yml
conda activate deepomapnet

pip

pip install -r requirements.txt

Core dependencies: Python 3.8, PyTorch >= 2.0, PyTorch Geometric >= 2.3, ScanPy >= 1.9, AnnData >= 0.9.

---

Benchmark results

Evaluated on a synthetic CITE-seq benchmark (500 cells, 250 Normal + 250 AML, 30 proteins, 500 genes).

Metric	Value
ADT prediction — mean Pearson r	0.785
ADT prediction — best single protein r	0.948
AML classification — AUC-ROC	0.836
AML classification — F1	0.719

To reproduce these results:

python run_experiment.py   # saves figures to results/

---

Repository structure

DeepOMAPNet/
├── scripts/
│   ├── model/
│   │   └── doNET.py                 # GATWithTransformerFusion and all model components
│   ├── data_provider/
│   │   ├── graph_data_builder.py    # k-NN graph construction, PCA, PyG Data objects
│   │   └── data_preprocessing.py    # CLR and Z-score normalization, train/val/test splitting
│   ├── trainer/
│   │   └── gat_trainer.py           # Multi-task training loop with AMP and early stopping
│   └── visualizations.py            # Plotting utilities
├── Tutorials/
│   └── Training.ipynb               # End-to-end tutorial
├── tests/                           # pytest test suite
├── R/                               # Supporting R scripts (WNN, preprocessing)
├── research/                        # Experiment scripts
├── run_experiment.py                # Main experiment execution script
├── environment.yml
└── requirements.txt

---

Tests

# Run the full test suite
pytest

# Run a specific test file
pytest tests/test_model_components.py -v

File	Tests	Coverage
test_model_components.py	36	Forward pass, gradients, sparse attention, adapters
test_data_pipeline.py	25	Normalization correctness, graph validity, split integrity
test_training.py	10	Loss convergence, gradient clipping, reproducibility
test_performance_benchmark.py	16	Pearson r vs. baselines, Wilcoxon test, AML AUC

---

License

MIT. See LICENSE.