Multimodal deep learning enhances genomic risk prediction for cardiometabolic diseases in UK Biobank

Leveraging Bidirectional Mamba's capacity to capture long-range dependencies across whole-genome SNP data, we propose DeepGP, a deep learning approach to improve genetic prediction for cardiometabolic diseases while providing interpretable insights into risk factors.

Papar Information

• Authors: Taiyu Zhu, Upamanyu Ghose, Héctor Climente-González, Joanna M. M. Howson, Sile Hu, Alejo Nevado-Holgado
• Affiliations: University of Oxford, Novo Nordisk Research Centre Oxford
• Preprint: TBA

Dataset

• UK Biobank Resource under Application Number 53639

Use

• To train and test the model, run:

   bash scripts/t2d.sh
  

Directory Hierarchy

|—— .gitignore
|—— README.md
|—— args_generator.py
|—— layers
|    |—— Embed.py
|    |—— SelfAttention_Family.py
|    |—— Transformer_EncDec.py
|—— main_genome.py
|—— models
|    |—— BaseModel.py
|    |—— DeepGP.py
|—— scripts
|    |—— t2d.sh
|—— utils.py

Code Details

Tested Platform

• software

  Python: 3.10.13
  PyTorch 2.1.1
  PyTorch Lightning 2.0.8
  

• hardware

  CPU: AMD EPYC 7R13 Processor
  GPU: NVIDIA A10 Tensor Core GPU
  

References

We would like to express our gratitude to the following GitHub repositories for their valuable code and contributions:

• Mamba
• Vision Mamba

License

BSD 3-Clause License

Copyright (c) 2025, University of Oxford. & Novo Nordisk. All rights reserved.

Citing

Please use the following BibTeX entry.

TBA