OptMRL

Convolutional neural network for predicting the mean ribosome load (MRL) of an mRNA from the 50 nucleotides upstream of the coding sequence.

Disclaimer

This is an UNOFFICIAL implementation of Interpreting Deep Neural Networks for the Prediction of Translation Rates by Frederick Korbel, et al.

The OFFICIAL repository of OptMRL is at ohlerlab/mlcis.

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

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

Model Details

OptMRL is a small 1D convolutional neural network trained to predict the mean ribosome load (MRL), a polysome-profiling-derived translation efficiency proxy, from the 50 nucleotides of 5' untranslated region (5'UTR) sequence immediately upstream of the coding sequence. The model was first pre-trained on roughly 260,000 random 5'UTR reporters and then fine-tuned on roughly 20,000 endogenous human 5'UTRs. Please refer to the Training Details section for more information on the training process.

The architecture is a stack of three Conv1D layers (120 filters, kernel size 8, same padding, ReLU activation) followed by a Flatten, a 40-unit Dense bottleneck with ReLU activation and dropout, and a final scalar Dense regression head.

Model Specification

Num Layers	Hidden Size	Num Parameters (M)	FLOPs (M)	MACs (M)	Max Num Tokens
5	40	0.476	24.04	12.00	50

Links

• Code: multimolecule.optmrl
• Data: 260,000 random 5'UTR reporters (pre-training) + 20,000 human 5'UTR reporters (fine-tuning)
• Paper: Interpreting Deep Neural Networks for the Prediction of Translation Rates
• Developed by: Frederick Korbel, Ekaterina Eroshok, Uwe Ohler
• Model type: 1D CNN for mean-ribosome-load regression from 5'UTR sequence
• Original Repository: ohlerlab/mlcis

Usage

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

pip install multimolecule

Direct Use

Mean Ribosome Load Prediction

You can use this model directly to predict the mean ribosome load of a 50-nucleotide 5'UTR window:

>>> from multimolecule import RnaTokenizer, OptMrlForSequencePrediction

>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/optmrl")
>>> model = OptMrlForSequencePrediction.from_pretrained("multimolecule/optmrl")
>>> sequence = "ACGU" * 12 + "AC"  # 50 nt
>>> input = tokenizer(sequence, add_special_tokens=False, return_tensors="pt")
>>> output = model(**input)

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

Interface

• Input length: 50 nt fixed 5'UTR window taken immediately upstream of the coding sequence
• Padding: shorter sequences are right-padded with zeros to 50 nt; longer sequences are truncated to the first 50 nt
• Alphabet: ACGU only; unknown / N tokens contribute zero one-hot signal
• Special tokens: do not add (add_special_tokens=False)
• Output: single scalar mean-ribosome-load (MRL) score per window

Training Details

OptMRL was first pre-trained on a large random-5'UTR reporter library and then fine-tuned on a smaller library of endogenous human 5'UTRs.

Training Data

• Pre-training: ~260,000 random 5'UTR reporters paired with polysome-profiling MRL measurements.
• Fine-tuning: ~20,000 endogenous human 5'UTR reporters paired with polysome-profiling MRL measurements.

Each reporter contributes a 50-nucleotide 5'UTR window immediately upstream of the coding sequence and a scalar MRL label.

Note [RnaTokenizer][multimolecule.RnaTokenizer] will convert "T"s to "U"s for you, you may disable this behaviour by passing replace_T_with_U=False.

Training Procedure

Pre-training

The model was first pre-trained as a regression task to predict the measured MRL of each random 5'UTR reporter, then fine-tuned end-to-end on the human-5'UTR reporters using the same regression objective. The published model is the fine-tuned model.

Citation

@article{korbel2023interpreting,
  author    = {Korbel, Frederick and Eroshok, Ekaterina and Ohler, Uwe},
  title     = {Interpreting Deep Neural Networks for the Prediction of Translation Rates},
  journal   = {bioRxiv},
  publisher = {Cold Spring Harbor Laboratory},
  year      = {2023},
  doi       = {10.1101/2023.06.02.543405}
}

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 OptMRL 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