Implementation of BPNet, a base-resolution deep neural network for functional genomics tasks. The offical implementation of BPNet can be found here.
$ pip install bpnet-pytorchfrom bpnet_pytorch import BPNet
model = BPNet()
x = torch.randn(1, 4, 1000)
out = model(x)
# out['x'] contains the output of the convolution layers.
# May not be useful, but left for debugging purpose for now.
### shape: (1, 64, 1000)
# out['profile'] contains the output of profile head.
### shape: (1, 1000, 2), 2 for +/- strands.
# out['total_count'] contains the output of total count head.
### shape: (1, 2), 2 for +/- strands.Training data used to train the original BPNet model can be found here (30.8GB).
- Confirm that the receptive field is +-1034bp.
-
Implement early stopping.
-
Implement multinomial negative log-likelihood loss.
-
Modify heads for multi-task (multi-TF) prediction.
-
Prepare training data.
-
Train the model and reproduce the performance.
@article{avsec2021base,
title={Base-resolution models of transcription-factor binding reveal soft motif syntax},
author={Avsec, {\v{Z}}iga and Weilert, Melanie and Shrikumar,
Avanti and Krueger, Sabrina and Alexandari, Amr and Dalal, Khyati and Fropf,
Robin and McAnany, Charles and Gagneur, Julien and Kundaje, Anshul and others},
journal={Nature Genetics},
volume={53},
number={3},
pages={354--366},
year={2021},
publisher={Nature Publishing Group US New York}
}