[!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 a2z-chromatin did not write this model card for this model so this model card has been written by the MultiMolecule team.
Model Details
a2z-chromatin is a recurrent convolutional neural network (CNN+BLSTM, DanQ topology) trained to predict chromatin state from a fixed-length 600 bp one-hot encoded angiosperm DNA sequence. The single convolutional layer applies 320 filters with a kernel size of 26, followed by dropout and a max-pool over 13 positions; the resulting feature sequence is fed to a bidirectional LSTM (320 units per direction) whose final forward and backward hidden states are concatenated, projected through a 925-unit dense layer, and read out as a single per-window probability.
Please refer to the Training Details section for more information on the training process.
The model file depends on the multimolecule library. You can install it using pip:
pip install multimolecule
Direct Use
Chromatin State Prediction
You can use this model directly to predict the chromatin accessibility (or lack of DNA methylation, for the methylation variant) of a 600 bp angiosperm DNA sequence:
Alphabet: DNA IUPAC tokens; ambiguous bases use upstream fractional A/C/G/T mixtures, and non-IUPAC tokens map to zero
Output: single per-window logit (binary chromatin accessibility for a2z-accessibility, lack of DNA methylation for a2z-methylation)
Training Details
a2z-chromatin was trained to predict per-window chromatin state across angiosperms using a single shared cross-species DanQ topology.
Training Data
a2z-chromatin was trained on two cross-species data resources:
Chromatin accessibility: leaf ATAC-seq peaks from 12 angiosperm species, with each 600 bp genomic interval labelled as accessible or inaccessible.
DNA methylation: unmethylated-region (UMR) calls from 10 angiosperm species, with each 600 bp genomic interval labelled as unmethylated or methylated. Unmethylated regions overlap significantly with accessible chromatin in plants, so the two tasks share the same architecture.
Each training example is a 600 bp one-hot encoded DNA sequence with a single binary label.
Training Procedure
Pre-training
Each variant was trained to minimize a binary cross-entropy loss between its sigmoid-activated per-window prediction and the observed accessibility / unmethylation label, sweeping cross-species splits to evaluate generalization.
Optimizer: Adam
Loss: Binary cross-entropy
Regularization: Dropout (0.2 after the convolution, 0.5 after the bidirectional LSTM)
Citation
bibtex
1@article{wrightsman2022a2z,
2 author = {Wrightsman, Travis and Marand, Alexandre P. and Crisp, Peter A. and Springer, Nathan M. and Buckler, Edward S.},
3 title = {Modeling chromatin state from sequence across angiosperms using recurrent convolutional neural networks},
4 journal = {The Plant Genome},
5 volume = 15,
6 number = 3,
7 pages = {e20249},
8 year = 2022,
9 publisher = {Wiley},
10 doi = {10.1002/tpg2.20249}
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 a2z-chromatin paper for questions or comments on the paper/model.
