Deep convolutional neural network that predicts noncoding chromatin features (DNase I hypersensitivity, transcription-factor binding, and histone marks) from DNA sequence, used to score the regulatory impact of noncoding variants.
[!TIP]
The MultiMolecule team has confirmed that the provided model and checkpoints are producing the same intermediate representations as the original implementation.
The team releasing DeepSEA did not write this model card for this model so this model card has been written by the MultiMolecule team.
Model Details
DeepSEA is a convolutional neural network (CNN) trained to predict 919 chromatin features—DNase I hypersensitivity peaks, transcription-factor binding peaks, and histone-mark peaks—across multiple human cell types from a fixed-length 1000 bp DNA sequence. The model applies three convolutional blocks (convolution, ReLU, max pooling, and dropout) followed by a single fully-connected layer and a multi-label sigmoid output. The sequence-prediction model averages forward and reverse-complement probabilities. The trained model is then used to score the regulatory impact of noncoding single-nucleotide variants by computing the difference between reference- and alternate-allele predictions. Please refer to the Training Details section for more information on the training process.
Output: 919 chromatin-feature logits (multi-label binary), covering DNase I hypersensitivity, transcription-factor binding, and histone-mark peaks across multiple cell types
Training Details
DeepSEA was trained to predict the chromatin features of DNA sequences across a panel of human cell types and then used to score the regulatory impact of noncoding variants.
Training Data
DeepSEA was trained on chromatin profiling data from ENCODE and the Roadmap Epigenomics project, comprising 690 transcription-factor ChIP-seq profiles, 125 DNase I hypersensitivity profiles, and 104 histone-mark ChIP-seq profiles for a total of 919 chromatin features. Each 1000 bp genomic interval centered on a 200 bp bin is labeled with a binary vector indicating which of the 919 chromatin features have a peak overlapping the central bin.
Training Procedure
Pre-training
The model was trained to minimize a multi-label binary cross-entropy loss, comparing its predicted per-feature probabilities against the observed chromatin-feature labels.
Optimizer: Stochastic gradient descent with momentum
Loss: Multi-label binary cross-entropy
Regularization: Dropout (0.2 after the first two convolutions, 0.5 after the third convolution) and L2 weight decay
Citation
bibtex
1@article{zhou2015deepsea,
2 author = {Zhou, Jian and Troyanskaya, Olga G.},
3 title = {Predicting effects of noncoding variants with deep learning-based sequence model},
4 journal = {Nature Methods},
5 volume = 12,
6 number = 10,
7 pages = {931--934},
8 year = 2015,
9 publisher = {Nature Publishing Group},
10 doi = {10.1038/nmeth.3547}
11}
[!NOTE]
The artifacts distributed in this repository are part of the MultiMolecule project.
If MultiMolecule supports your research, please cite the MultiMolecule project as follows:
bibtex
1@software{chen_2024_12638419,
2 author = {Chen, Zhiyuan and Zhu, Sophia Y.},
3 title = {MultiMolecule},
4 doi = {10.5281/zenodo.12638419},
5 publisher = {Zenodo},
6 url = {https://doi.org/10.5281/zenodo.12638419},
7 year = 2024,
8 month = may,
9 day = 4
10}
Contact
Please use GitHub issues of MultiMolecule for any questions or comments on the model card.
Please contact the authors of the DeepSEA paper for questions or comments on the paper/model.
