[!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 Pangolin did not write this model card for this model so this model card has been written by the MultiMolecule team.
Model Details
Pangolin is a deep convolutional neural network (CNN) that predicts splice site strength from primary pre-mRNA sequence.
It extends the dilated-residual SpliceAI architecture to predict tissue-specific splice site usage, and is trained on splicing measurements derived from RNA-seq data across multiple tissues.
The network processes a one-hot encoded nucleotide sequence and, for each position, predicts a splice-site score and a splice-site usage score per tissue.
Pangolin is typically used to estimate the effect of genetic variants on splicing by scoring reference and alternate sequences and taking the difference.
Please refer to the Training Details section for more information on the training process.
The probabilities tensor reproduces the original Pangolin output: for each of the four tissues, two splice-site score channels (softmax) and one splice-site usage channel (sigmoid).
Downstream Use
Token Prediction
You can fine-tune Pangolin for per-nucleotide splice site strength regression with [PangolinForTokenPrediction][multimolecule.models.PangolinForTokenPrediction], which adds a shared token prediction head on top of the backbone.
Interface
Input length: variable pre-mRNA sequence
Padding: flanking context padded with N near transcript ends
Output: per-position tissue-specific channels — for each of 4 tissues, 2 splice-site score channels + 1 splice-site usage channel
Training Details
Pangolin was trained to predict tissue-specific splice site usage from primary pre-mRNA sequence.
Training Data
Pangolin was trained on splice site usage derived from RNA-seq data in heart, liver, brain, and testis tissues from human and three other species, using gene annotations from GENCODE.
For each nucleotide whose splicing status was predicted, a sequence window centered on that nucleotide was used, with the flanking context padded with N (unknown nucleotide) when near transcript ends.
Training Procedure
Pre-training
The model was trained to minimize a combination of cross-entropy loss over splice-site classification and a regression loss over splice-site usage, comparing predictions against measurements derived from RNA-seq.
Optimizer: AdamW
Learning rate scheduler: Step decay
Citation
bibtex
1@article{zeng2022predicting,
2 author = {Zeng, Tony and Li, Yang I.},
3 title = {Predicting RNA splicing from DNA sequence using Pangolin},
4 journal = {Genome Biology},
5 volume = {23},
6 number = {1},
7 pages = {103},
8 year = {2022},
9 doi = {10.1186/s13059-022-02664-4},
10 publisher = {BioMed Central}
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 Pangolin paper for questions or comments on the paper/model.
