Environment setup

python3 -m pip install -r requirements.txt

Model Training

To launch training you can just run hypermodel.py

python3 hypermodel.py
python3 hypermodel_graph.py

model logs of all our training attempts can be checked at https://app.neptune.ai/koritsky/DL2021-Bio

logs of hypermodel.py are at https://app.neptune.ai/koritsky/DL2021-Bio/e/DLBIO-165

logs of hypermodel_graph.py are at https://app.neptune.ai/koritsky/DL2021-Bio/e/DLBIO-178

Model Evaluation on the test set

Download datasets from https://drive.google.com/drive/folders/1wy0lwgR_zzb2GDVkwRELqITkSY1jt6v8 and put it into the dataset/ directory

To start model evaluation ...

python3 eval.py --model model_type --checkpoint PATH_TO_WEIGHTS --cuda 1 --plot 2

PATH_TO_WEIGHTS is hypermodel.pth or hypermodel-graph.pth

model_type is conv or graph

Hybrid model

The key idea of our model is incorporation of information about a DNA sequence to a low resolution image to obtain an image with higher resolution

VeHiCLE model

VEHiCLE, a DL algorithm for resolution enhancement of Hi-C contact data.

@article{VEHiCLE,
  title={VEHiCLE: a Variationally Encoded Hi-C Loss Enhancement algorithm for improving and generating Hi-C data},
  author={Highsmith, Max and Cheng, Jianlin},
  journal={Scientific Reports},
  volume={11},
  number={1},
  pages={1--13},
  year={2021},
  publisher={Nature Publishing Group}
}

Akita model

Sequence-to-image model that accurately predicts genome folding from DNA sequence.

@article{Akita,
  title={Predicting 3D genome folding from DNA sequence with Akita},
  author={Fudenberg, Geoff and Kelley, David R and Pollard, Katherine S},
  journal={Nature Methods},
  volume={17},
  number={11},
  pages={1111--1117},
  year={2020},
  publisher={Nature Publishing Group}
}