BPNet

Base-resolution convolutional neural network for predicting transcription-factor binding profiles from DNA sequence.

Disclaimer

This is an UNOFFICIAL implementation of Base-resolution models of transcription-factor binding reveal soft motif syntax by Žiga Avsec, Melanie Weilert et al.

The OFFICIAL repository of BPNet is at kundajelab/bpnet.

The MultiMolecule team has confirmed that the provided model and checkpoints are producing the same intermediate representations as the original implementation.

The team releasing BPNet did not write this model card for this model so this model card has been written by the MultiMolecule team.

Model Details

BPNet is a convolutional neural network (CNN) trained to predict base-resolution transcription-factor binding signal (ChIP-nexus) from primary DNA sequence. It uses a convolutional motif stem followed by a stack of dilated residual convolutions that aggregate ~1 kb of genomic context.

BPNet predicts a single base-resolution signal task whose output is factorized into two terminal branches that share the dilated-convolution backbone:

• a profile branch predicting the shape of the binding signal as per-position multinomial logits, trained with a multinomial negative log-likelihood;
• a count branch predicting the total magnitude of the signal as a scalar per task and strand (in log space), trained with mean-squared error.

The usable base-resolution prediction recombines the two branches as softmax(profile_logits, positions) * exp(count_logits), exposed via BPNetForProfilePrediction.postprocess. Please refer to the Training Details section for more information on the training process.

Model Specification

Num Layers	Hidden Size	Num Parameters (M)	FLOPs (G)	MACs (G)
10	64	0.13	0.24	0.12

Links

• Code: multimolecule.bpnet
• Weights: multimolecule/bpnet
• Data: BPNet manuscript data
• Paper: Base-resolution models of transcription-factor binding reveal soft motif syntax
• Developed by: Žiga Avsec, Melanie Weilert, Avanti Shrikumar, Sabrina Krueger, Amr Alexandari, Khyati Dalal, Robin Fropf, Charles McAnany, Julien Gagneur, Anshul Kundaje, Julia Zeitlinger
• Original Repository: kundajelab/bpnet

Usage

The model file depends on the multimolecule library. You can install it using pip:

pip install multimolecule

Direct Use

You can use this model directly to predict transcription-factor binding profiles of a DNA sequence:

>>> from multimolecule import DnaTokenizer, BPNetForProfilePrediction

>>> tokenizer = DnaTokenizer.from_pretrained("multimolecule/bpnet")
>>> model = BPNetForProfilePrediction.from_pretrained("multimolecule/bpnet")
>>> output = model(**tokenizer("ACGTNACGTN", return_tensors="pt"))

>>> output.keys()
odict_keys(['profile_logits', 'count_logits'])

>>> output["profile_logits"].shape
torch.Size([1, 10, 8])

>>> output["count_logits"].shape
torch.Size([1, 8])

>>> track = model.postprocess(output)
>>> track.shape
torch.Size([1, 10, 8])

The recombined track is the usable base-resolution prediction. The last dimension stacks num_tasks (Oct4, Sox2, Nanog, Klf4) by num_strands (forward, reverse).

Training Details

BPNet was trained to predict the base-resolution ChIP-nexus binding profiles of the pluripotency transcription factors Oct4, Sox2, Nanog and Klf4 in mouse embryonic stem cells.

Training Data

The published BPNet-OSKN model was trained on ChIP-nexus profiles for Oct4, Sox2, Nanog and Klf4, using 1 kb genomic windows centered on detected binding peaks. The training regions and trained Keras checkpoint are distributed as part of the BPNet manuscript data.

Training Procedure

Training

The model was trained with a composite loss: a multinomial negative log-likelihood on the per-position profile shape plus a mean-squared-error regression on the log total counts.

• Backbone: 1 motif convolution (64 filters, kernel 25, ReLU) + 9 dilated residual convolutions (64 filters, kernel 3, dilations 2, 4, 8, …, 512, ReLU)
• Profile head: per-task transposed convolution (kernel 25) producing per-position logits
• Count head: per-task global average pooling + linear layer producing log-count scalars
• Optimizer: Adam

Citation

@article{avsec2021baseresolution,
  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 Zeitlinger, Julia},
  title     = {Base-resolution models of transcription-factor binding reveal soft motif syntax},
  journal   = {Nature Genetics},
  volume    = 53,
  number    = 3,
  pages     = {354--366},
  year      = 2021,
  publisher = {Nature Publishing Group},
  doi       = {10.1038/s41588-021-00782-6}
}

The artifacts distributed in this repository are part of the MultiMolecule project. If you use MultiMolecule in your research, you must cite the MultiMolecule project as follows:

@software{chen_2024_12638419,
  author    = {Chen, Zhiyuan and Zhu, Sophia Y.},
  title     = {MultiMolecule},
  doi       = {10.5281/zenodo.12638419},
  publisher = {Zenodo},
  url       = {https://doi.org/10.5281/zenodo.12638419},
  year      = 2024,
  month     = may,
  day       = 4
}

Contact

Please use GitHub issues of MultiMolecule for any questions or comments on the model card.

Please contact the authors of the BPNet paper for questions or comments on the paper/model.

License

This model implementation is licensed under the GNU Affero General Public License.

For additional terms and clarifications, please refer to our License FAQ.

SPDX-License-Identifier: AGPL-3.0-or-later