Hugging Face's logo

multimolecule
/
deepmel

Other
MultiMolecule
PyTorch
Safetensors
Upper Grand Valley Dani
deepmel
Biology
DNA
Model card Files
xet
 Community
deepmel / README.md
ZhiyuanChen's picture
ZhiyuanChen
Upload folder using huggingface_hub
010e2db verified
preview code
|
raw
history blame contribute delete
11.9 kB
metadata 
language: dna
library_name: multimolecule
license: agpl-3.0
pipeline: regulatory-activity
pipeline_tag: other
tags:
  - Biology
  - DNA
widget:
  - example_title: tumor protein p53
    pipeline_tag: regulatory-activity
    sequence_type: DNA
    task: regulatory-activity
    text: >-
      ACTCCCCTGCCCTCAACAAGATGTTTTGCCAACTGGCCAAGACCTGCCCTGTGCAGCTGTGGGTTGATTCCACACCCCCGCCCGGCACCCGCGTCCGCGCCATGGCCATCTACAAGCAGTCACAGCACATGACGGAGGTTGTGAGGCGCTGCCCCCACCATGAGCGCTGCTCAGATAGCGATGG
  - example_title: BRCA1 DNA repair associated
    pipeline_tag: regulatory-activity
    sequence_type: DNA
    task: regulatory-activity
    text: >-
      TCATTGGAACAGAAAGAAATGGATTTATCTGCTCTTCGCGTTGAAGAAGTACAAAATGTCATTAATGCTATGCAGAAAATCTTAGAGTGTCCCATCTGG
  - example_title: hemoglobin subunit beta
    pipeline_tag: regulatory-activity
    sequence_type: DNA
    task: regulatory-activity
    text: >-
      CATTTGCTTCTGACACAACTGTGTTCACTAGCAACCTCAAACAGACACCATGGTGCATCTGACTCCTGAGGAGAAGTCTGCCGTTACTGCCCTGTGGGGCAAGGTGAACGTGGATGAAGTTGGTGGTGAGGCCCTGGGCAGG
  - example_title: CF transmembrane conductance regulator
    pipeline_tag: regulatory-activity
    sequence_type: DNA
    task: regulatory-activity
    text: >-
      ACTTCACTTCTAATGGTGATTATGGGAGAACTGGAGCCTTCAGAGGGTAAAATTAAGCACAGTGGAAGAATTTCATTCTGTTCTCAGTTTTCCTGGATTATGCCTGGCACCATTAAAGAAAATATCATCTTTGGTGTTTCCTATGATGAATATAGATACAGAAGCGTCATCAAAGCATGCCAACTAGAAGAG
  - example_title: telomerase reverse transcriptase
    pipeline_tag: regulatory-activity
    sequence_type: DNA
    task: regulatory-activity
    text: >-
      CGCGGGGGTGGCCGGGGCCAGGGCTTCCCACGTGCGCAGCAGGACGCAGCGCTGCCTGAAACTCGCGCCGCGAGGAGAGGGCGGGGCCGCGGAAAGGAAGGGGAGGGGCTGGGAGGGCCCGGAGGGGGCTGGGCCGGGGACCCGGGAGGGGTCGGGACGGGGCGGGGTCCGCGCGGAGGAGGCGGAGCTGGAAGGTGAAGGGGCAGGACGGGTGCCCGGGTCCCCAGTCCCTCCGCCACGTGGGAAGCGCGGTCCTGGGCGTCTGTGCCCGCGAATCCACTGGGAGCCCGGCCTGGCCCCGACAGCGCAGCTGCTCCGGGCGGACCCGGGG
  - example_title: KRAS proto-oncogene
    pipeline_tag: regulatory-activity
    sequence_type: DNA
    task: regulatory-activity
    text: >-
      GCCTGCTGAAAATGACTGAATATAAACTTGTGGTAGTTGGAGCTGGTGGCGTAGGCAAGAGTGCCTTGACGATACAGCTAATTCAGAATCATTTTGTGGACGAATATGATCCAACAATAGAG
  - example_title: prion protein (Kanno blood group)
    pipeline_tag: regulatory-activity
    sequence_type: cDNA
    task: regulatory-activity
    text: ATGGCGAACCTTGGCTGCTGGATGCTGGTTCTCTTTGTGGCCACATGGAGTGACCTGGGCCTCTGC
  - example_title: interleukin 10
    pipeline_tag: regulatory-activity
    sequence_type: cDNA
    task: regulatory-activity
    text: ATGCACAGCTCAGCACTGCTCTGTTGCCTGGTCCTCCTGACTGGGGTGAGGGCC
  - example_title: Zaire ebolavirus
    pipeline_tag: regulatory-activity
    sequence_type: cDNA
    task: regulatory-activity
    text: >-
      AATGTTCAAACACTTTGTGAAGCTCTGTTAGCTGATGGTCTTGCTAAAGCATTTCCTAGCAATATGATGGTAGTCACAGAGCGTGAGCAAAAAGAAAGCTTATTGCATCAAGCATCATGGCACCACACAAGTGATGATTTTGGTGAGCATGCCACAGTTAGAGGGAGTAGCTTTGTAACTGATTTAGAGAAATACAATCTTGCATTTAGATATGAGTTTACAGCACCTTTTATAGAATATTGTAACCGTTGCTATGGTGTTAAGAATGTTTTTAATTGGATGCATTATACAATCCCACAGTGTTAT
  - example_title: SARS coronavirus
    pipeline_tag: regulatory-activity
    sequence_type: cDNA
    task: regulatory-activity
    text: >-
      ATGTTTATTTTCTTATTATTTCTTACTCTCACTAGTGGTAGTGACCTTGACCGGTGCACCACTTTTGATGATGTTCAAGCTCCTAATTACACTCAACATACTTCATCTATGAGGGGGGTTTACTATCCTGATGAAATTTTTAGATCAGACACTCTTTATTTAACTCAGGATTTATTTCTTCCATTTTATTCTAATGTTACAGGGTTTCATACTATTAATCATACGTTTGACAACCCTGTCATACCTTTTAAGGATGGTATTTATTTTGCTGCCACAGAGAAATCAAATGTTGTCCGTGGTTGGGTTTTTGGTTCTACCATGAACAACAAGTCACAGTCGGTGATTATTATTAACAATTCTACTAATGTTGTTATACGAGCATGTAACTTTGAATTGTGTGACAACCCTTTCTTTGCTGTTTCTAAACCCATGGGTACACAGACACATACTATGATATTCGATAATGCATTTAAATGCACTTTCGAGTACATATCT
  - example_title: insulin
    pipeline_tag: regulatory-activity
    sequence_type: cDNA
    task: regulatory-activity
    text: >-
      ATGGCCCTGTGGATGCGCCTCCTGCCCCTGCTGGCGCTGCTGGCCCTCTGGGGACCTGACCCAGCCGCAGCCTTTGTGAACCAACACCTGTGCGGCTCACACCTGGTGGAAGCTCTCTACCTAGTGTGCGGGGAACGAGGCTTCTTCTACACACCCAAGACCCGCCGGGAGGCAGAGGACCTGCAGGTGGGGCAGGTGGAGCTGGGCGGGGGCCCTGGTGCAGGCAGCCTGCAGCCCTTGGCCCTGGAGGGGTCCCTGCAGAAGCGTGGCATTGTGGAACAATGCTGTACCAGCATCTGCTCCCTCTACCAGCTGGAGAACTACTGCAACTAG
  - example_title: cyclin dependent kinase inhibitor 2A
    pipeline_tag: regulatory-activity
    sequence_type: cDNA
    task: regulatory-activity
    text: >-
      ATGGAGCCGGCGGCGGGGAGCAGCATGGAGCCTTCGGCTGACTGGCTGGCCACGGCCGCGGCCCGGGGTCGGGTAGAGGAGGTGCGGGCGCTGCTGGAGGCGGGGGCGCTGCCCAACGCACCGAATAGTTACGGTCGGAGGCCGATCCAGGTCATGATGATGGGCAGCGCCCGAGTGGCGGAGCTGCTGCTGCTCCACGGCGCGGAGCCCAACTGCGCCGACCCCGCCACTCTCACCCGACCCGTGCACGACGCTGCCCGGGAGGGCTTCCTGGACACGCTGGTGGTGCTGCACCGGGCCGGGGCGCGGCTGGACGTGCGCGATGCCTGGGGCCGTCTGCCCGTGGACCTGGCTGAGGAGCTGGGCCATCGCGATGTCGCACGGTACCTGCGCGCGGCTGCGGGGGGCACCAGAGGCAGTAACCATGCCCGCATAGATGCCGCGGAAGGTCCCTCAGACATCCCCGATTGA
  - example_title: human papillomavirus type 16 E6
    pipeline_tag: regulatory-activity
    sequence_type: cDNA
    task: regulatory-activity
    text: >-
      ATGCACCAAAAGAGAACTGCAATGTTTCAGGACCCACAGGAGCGACCCAGAAAGTTACCACAGTTATGCACAGAGCTGCAAACAACTATACATGATATAATATTAGAATGTGTGTACTGCAAGCAACAGTTACTGCGACGTGAGGTATATGACTTTGCTTTTCGGGATTTATGCATAGTATATAGAGATGGGAATCCATATGCTGTATGTGATAAATGTTTAAAGTTTTATTCTAAAATTAGTGAGTATAGACATTATTGTTATAGTTTGTATGGAACAACATTAGAACAGCAATACAACAAACCGTTGTGTGATTTGTTAATTAGGTGTATTAACTGTCAAAAGCCACTGTGTCCTGAAGAAAAGCAAAGACATCTGGACAAAAAGCAAAGATTCCATAATATAAGGGGTCGGTGGACCGGTCGATGTATGTCTTGTTGCAGATCATCAAGAACACGTAGAGAAACCCAGCTGTAA

DeepMEL

Convolutional and recurrent neural network for predicting melanoma-specific accessible chromatin regions and chromatin topics directly from DNA sequence.

Disclaimer

This is an UNOFFICIAL implementation of Cross-species analysis of enhancer logic using deep learning by Liesbeth Minnoye, Ibrahim Ihsan Taskiran, et al.

The OFFICIAL repository of DeepMEL is at aertslab/DeepMEL.

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

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

Model Details

DeepMEL is a hybrid convolutional / recurrent neural network trained to predict 24 melanoma chromatin topics (a 4-MEL melanocytic, a 7-MES mesenchymal-like, and additional accessibility programs) directly from 500 bp DNA sequence. Each input sequence is processed by a shared encoder consisting of a 1D convolution, max pooling, a time-distributed dense projection, and a bidirectional LSTM, followed by a fully-connected layer. The same encoder is applied independently to the forward DNA strand and to its reverse complement; a final 24-way decoder produces a sigmoid probability per topic in each branch, and the two branches' probabilities are averaged into the model's prediction. Please refer to the Training Details section for more information on the training process.

Model Specification

Conv Filters Conv Kernel BiLSTM Hidden FC Hidden Num Topics Num Parameters (M) FLOPs (M) MACs (M) Max Num Tokens
128 20 128 256 24 3.44 40.76 20.19 500

Links

  • Code: multimolecule.deepmel
  • Data: Melanoma cell-line single-cell ATAC-seq topic models
  • Paper: Cross-species analysis of enhancer logic using deep learning
  • Developed by: Liesbeth Minnoye, Ibrahim Ihsan Taskiran, David Mauduit, Maurizio Fazio, Linde Van Aerschot, Gert Hulselmans, Valerie Christiaens, Samira Makhzami, Monika Seltenhammer, Panagiotis Karras, Aline Primot, Edouard Cadieu, Ellen van Rooijen, Jean-Christophe Marine, Giorgia Egidy, Ghanem-Elias Ghanem, Leonard Zon, Jasper Wouters, Stein Aerts
  • Model type: 1D CNN + BiLSTM over 500 bp DNA with reverse-complement averaging for multi-task chromatin-topic prediction
  • Original Repository: aertslab/DeepMEL

Usage

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

pip install multimolecule

Direct Use

Chromatin Topic Prediction

You can use this model directly to predict the 24 melanoma chromatin-topic activities of a 500 bp DNA sequence:

>>> import torch
>>> from multimolecule import DnaTokenizer, DeepMelForSequencePrediction

>>> tokenizer = DnaTokenizer.from_pretrained("multimolecule/deepmel")
>>> model = DeepMelForSequencePrediction.from_pretrained("multimolecule/deepmel")
>>> sequence = "ACGT" * 125
>>> output = model(**tokenizer(sequence, return_tensors="pt"))

>>> output.logits.shape
torch.Size([1, 24])

Interface

  • Input length: fixed 500 bp DNA window
  • Alphabet: ACGT (one-hot encoded); the reverse complement is computed internally
  • Output: 24 chromatin-topic logits (multi-label binary); postprocess returns the branch-averaged sigmoid probability per topic

Training Details

DeepMEL was trained to predict cell-type-specific accessible chromatin topics derived from single-cell ATAC-seq of melanoma cell lines.

Training Data

DeepMEL was trained on accessible genomic intervals derived from melanoma single-cell ATAC-seq experiments and modeled as 24 chromatin topics (including the 4-MEL melanocytic-like and 7-MES mesenchymal-like programs). Each training example is a 500 bp genomic interval labelled with a binary vector indicating which topics are active. Chromosome 2 was held out for validation and testing.

Training Procedure

Pre-training

The model was trained to minimize a multi-label binary cross-entropy loss between the branch-averaged sigmoid probabilities and the observed topic-activity labels.

  • Optimizer: Adam
  • Loss: Multi-label binary cross-entropy
  • Regularization: Dropout (0.2 after pooling, 0.1 LSTM input and recurrent dropout, 0.2 after the BiLSTM, 0.4 before the prediction head)

Citation 

@article{minnoye2020deepmel,
  author    = {Minnoye, Liesbeth and Taskiran, Ibrahim Ihsan and Mauduit, David and Fazio, Maurizio and Van Aerschot, Linde and Hulselmans, Gert and Christiaens, Valerie and Makhzami, Samira and Seltenhammer, Monika and Karras, Panagiotis and Primot, Aline and Cadieu, Edouard and van Rooijen, Ellen and Marine, Jean-Christophe and Egidy, Giorgia and Ghanem, Ghanem-Elias and Zon, Leonard and Wouters, Jasper and Aerts, Stein},
  title     = {Cross-species analysis of enhancer logic using deep learning},
  journal   = {Genome Research},
  volume    = 30,
  number    = 12,
  pages     = {1815--1834},
  year      = 2020,
  publisher = {Cold Spring Harbor Laboratory Press},
  doi       = {10.1101/gr.260844.120}
}

The artifacts distributed in this repository are part of the MultiMolecule project. If MultiMolecule supports your research, please 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 DeepMEL 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