Multi-Omics Integration via Knowledge-Enhanced Graph Attention Networks for Alzheimer's Disease Progression Prediction
DeepOmicsNet is a deep learning framework that integrates multi-omics data (RNA-seq, proteomics, DNA methylation, miRNA) through knowledge-enhanced graph attention networks to predict Alzheimer's disease progression with high accuracy and interpretability.
- Knowledge-Enhanced GAT: Integrates biological knowledge graphs (PPIs, gene regulatory networks, pathways)
- Modality-Specific VAEs: Learns compact latent representations from each omics layer
- Multi-Objective Learning: Pathway coherence regularization ensures biological plausibility
- High Performance: AUROC 0.891 on ROSMAP dataset, outperforming state-of-the-art methods
- Interpretability: SHAP-based feature importance and attention weight analysis
Important: This repository does not include the ROSMAP dataset due to data use agreements.
To use DeepOmicsNet with ROSMAP data:
- Request access through AMP-AD Knowledge Portal
- Apply for dbGaP approval (accession phs000424.v1.p1)
- Follow Rush Alzheimer's Disease Center data use requirements
The repository includes pre-trained models and can be used with your own multi-omics data.
- Python 3.8+
- PyTorch 2.0+
- CUDA 11.0+ (optional, for GPU support)
git clone https://github.com/kushalrajroy/DeepOmicsNet.git
cd DeepOmicsNet
pip install -r requirements.txt
python setup.py installpip install -r requirements.txtfrom deepomicsnet import DeepOmicsNet
from deepomicsnet.data import MultiOmicsDataset
# Load your data
dataset = MultiOmicsDataset(
rna_path='rna_expression.csv',
protein_path='protein_abundance.csv',
methylation_path='methylation.csv',
mirna_path='mirna_expression.csv',
labels_path='labels.csv'
)
# Initialize model
model = DeepOmicsNet(
input_dims={'rna': 5000, 'protein': 5000, 'methylation': 10000, 'mirna': 1000},
latent_dim=128,
num_classes=2
)
# Train model
model.train(dataset, epochs=200)
# Make predictions
predictions = model.predict(test_data)DeepOmicsNet/
├── deepomicsnet/ # Core package
│ ├── models/ # Model architectures (VAE, GAT, DeepOmicsNet)
│ ├── data/ # Data loading and preprocessing
│ └── utils/ # Utilities (metrics, visualization)
├── notebooks/ # Jupyter notebooks with examples
├── scripts/ # Training and evaluation scripts
├── pretrained_models/ # Pre-trained model weights
└── docs/ # Documentation
See the notebooks/ directory for detailed examples:
01_data_preprocessing.ipynb: Data loading and preprocessing02_model_training.ipynb: Model training and hyperparameter tuning03_results_visualization.ipynb: Results analysis and visualization
DeepOmicsNet consists of three main components:
- Modality-Specific VAEs: Learn compact latent representations from each omics modality
- Knowledge-Guided GAT: Integrates multi-omics features through biological knowledge graph
- Classification Module: Attention-based aggregation and prediction
Evaluated on ROSMAP dataset (1,337 samples):
- AUROC: 0.891 ± 0.018
- AUPRC: 0.876 ± 0.021
- Sensitivity: 83.6%
- Specificity: 87.3%
Significantly outperforms:
- Standard DNN: 0.824 AUROC
- Elastic Net: 0.782 AUROC
- Single modality: 0.801 AUROC (best)
If you use DeepOmicsNet in your research, please cite:
@article{roy2026deepomicsnet,
title={DeepOmicsNet: Multi-Omics Integration via Knowledge-Enhanced Graph Attention Networks for Alzheimer's Disease Progression Prediction},
author={Roy, Kushal Raj},
journal={xxx},
year={2026}
}This project is licensed under the MIT License - see the LICENSE file for details.
Kushal Raj Roy
Department of Biology and Biochemistry
University of Houston
Email: kroy3@uh.edu
Data used in this study were obtained from:
- Religious Orders Study and Memory and Aging Project (ROSMAP)
- AMP-AD Knowledge Portal
- Rush Alzheimer's Disease Center
Knowledge graph sources:
- STRING (protein-protein interactions)
- KEGG, Reactome (pathways)
- miRTarBase (miRNA interactions)
- TRRUST, hTFtarget (gene regulatory networks)