license: mit
library_name: pytorch
tags:
- deep-learning
- genomics
- dna-sequence
- dna-shape
- transcription-factor
- histone-marks
- DNase
- regulatory-elements
model-index:
- name: DeepShape
results:
- task:
name: TF/DNase/Histone Binding Prediction
type: binary-classification
dataset:
name: ENCODE + Roadmap Epigenomics (919 targets)
type: genomics
split: 'held-out chromosomes (validation: chr6/7, test: chr8/9)'
metrics:
- name: AUROC (TF)
type: auc
value: 0.948
- name: AUPRC (TF)
type: average precision
value: 0.353
- name: AUROC (DNase)
type: auc
value: 0.907
- name: AUPRC (DNase)
type: average precision
value: 0.472
- name: AUROC (Histone)
type: auc
value: 0.85
- name: AUPRC (Histone)
type: average precision
value: 0.367
DeepShape
DeepShape is a deep convolutional neural network designed to predict molecular phenotypes from DNA sequences. Unlike traditional models that rely solely on one-hot encoded DNA sequences, DeepShape integrates DNA structural attributes indicative of local shape: minor groove width (MGW), helical twist (HelT), propeller twist (ProT), roll, and electrostatic potential (EP). This combination enhances the interpretability of the model and helps identify regulatory patterns that are not apparent from sequence information alone.
DeepShape is built upon DeeperDeepSEA, a PyTorch-based deep learning model originally designed to predict chromatin features from DNA sequence alone as implemented in Selene.
https://github.com/ni-lab/DeepShape
License
DeepShape is licensed under the MIT License. Portions of this software are derived from Selene, which is licensed under the Clear BSD License.
See the LICENSE file for full details.