[!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 DeepCpG-DNA did not write this model card for this model so this model card has been written by the MultiMolecule team.
Model Details
DeepCpG-DNA is the DNA submodule of the DeepCpG joint model. It is a 1D convolutional neural network that predicts the per-cell methylation state of a CpG site from a fixed-length 1001 bp DNA window centered on the site. The model consumes a one-hot encoded sequence and applies valid-padded convolutional blocks (Conv1D + ReLU + MaxPool) followed by a dense bottleneck and one binary classification head per single cell in the training dataset. Please refer to the Training Details section for more information on the training process.
The full DeepCpG model combines this DNA submodule with a recurrent CpG-context submodule and a joint head; this model card covers the DNA submodule only.
Variants
The DeepCpG-DNA module is trained per single-cell dataset, so each variant predicts a different number of output cells.
Each logit is a per-cell methylation score for one of the single cells in the chosen training dataset; apply a sigmoid to obtain methylation probabilities.
Interface
Input length: fixed 1001 bp DNA window centered on a CpG site
Padding: not supported; pad or crop genomic windows so they match sequence_length exactly
Alphabet: DNA (A, C, G, T); N is encoded as an all-zero channel
Output: per-cell methylation logits; the number of cells is dataset-specific (see Variants table)
Training Details
DeepCpG-DNA was trained to predict the per-cell methylation state of CpG sites from their flanking DNA context.
Training Data
DeepCpG-DNA was trained on single-cell bisulfite sequencing datasets:
Smallwood 2014: scBS-seq profiles of mouse embryonic stem cells, with 18 serum and 12 2i mESCs (excluding two serum cells whose methylation pattern deviated strongly from the remainder).
Hou 2016: scRRBS-seq profiles of 25 human hepatocellular carcinoma (HCC) cells, 6 human heptoplastoma-derived (HepG2) cells, and 6 mESCs, restricted to CpG sites covered by at least four reads.
Each training example is a 1001 bp DNA window centered on a CpG site, with a per-cell binary methylation label (methylated, unmethylated, or missing). Chromosomes were split into training, validation, and test sets to avoid sequence leakage.
Training Procedure
Pre-training
The model was trained to minimize a per-cell binary cross-entropy loss, comparing its predicted per-cell methylation probabilities (sigmoid of the per-cell logits) against the observed single-cell bisulfite labels. Missing labels are masked out during training.
Optimizer: Adam
Loss: Per-cell binary cross-entropy
Regularization: Dropout and L2 weight decay
Citation
bibtex
1@article{angermueller2017deepcpg,
2 author = {Angermueller, Christof and Lee, Heather J. and Reik, Wolf and Stegle, Oliver},
3 title = {{DeepCpG}: accurate prediction of single-cell {DNA} methylation states using deep learning},
4 journal = {Genome Biology},
5 volume = 18,
6 number = 1,
7 pages = {67},
8 year = 2017,
9 publisher = {BioMed Central},
10 doi = {10.1186/s13059-017-1189-z}
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 DeepCpG paper for questions or comments on the paper/model.
