Deep residual neural network for predicting human 3' UTR Alternative Polyadenylation (APA) and cleavage magnitude at nucleotide resolution, and for deciphering the impact of genetic variants on polyadenylation.
[!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 APARENT2 did not write this model card for this model so this model card has been written by the MultiMolecule team.
Model Details
APARENT2 is a residual convolutional neural network (a ResNet successor to the original APARENT) trained on a 3' UTR massively parallel reporter assay (MPRA). Given a fixed 205 nt polyadenylation signal (PAS) sequence, it predicts a nucleotide-resolution cleavage probability distribution as well as the overall isoform abundance. It is primarily used to score the effect of genetic variants on polyadenylation by comparing the predictions for a reference and an alternate sequence.
The model file depends on the multimolecule library. You can install it using pip:
pip install multimolecule
Direct Use
Polyadenylation Cleavage Prediction
You can use this model directly to predict the cleavage distribution of a 205 nt polyadenylation signal sequence (core hexamer starting at position 70):
Hexamer position: core hexamer (e.g., AAUAAA) at position 70 (0-indexed) of the 205 nt window
Output: 206-dim cleavage distribution (one score per input position + trailing "no cleavage in window" bucket)
Variant Effect
Score reference and alternate sequences separately and compare their cleavage / isoform predictions
There is no separate ref/alt output dataclass
Training Details
APARENT2 was trained to predict nucleotide-resolution cleavage and isoform abundance from 3' UTR MPRA measurements.
Training Data
The model was trained on the 3' UTR MPRA library used by the original APARENT, re-processed with additional improvements (exact cleavage positions for the Alien1 Random sublibrary and a 20 nt random barcode upstream of the USE in the Alien1 sublibrary). The measured variant data and processed data repository are available at the original APARENT GitHub.
Training Procedure
Pre-training
The model minimizes a combination of a sigmoid KL-divergence isoform loss and a KL-divergence cleavage loss, weighted equally. The released inference model corresponds to the residual-network model trained for 5 epochs on all sublibraries (excluding ClinVar wild-type sequences), with dropout disabled for inference.
Citation
bibtex
1@article{linder2022deciphering,
2 author = {Linder, Johannes and Koplik, Samantha E. and Kundaje, Anshul and Seelig, Georg},
3 title = {Deciphering the impact of genetic variation on human polyadenylation using APARENT2},
4 journal = {Genome Biology},
5 volume = {23},
6 number = {1},
7 pages = {232},
8 year = {2022},
9 doi = {10.1186/s13059-022-02799-4},
10 publisher = {Springer Science and Business Media LLC}
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 APARENT2 paper for questions or comments on the paper/model.
