[!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 OpenSpliceAI did not write this model card for this model so this model card has been written by the MultiMolecule team.
Model Details
OpenSpliceAI is a deep dilated residual convolutional neural network that reimplements the SpliceAI architecture in native PyTorch. It predicts, for each nucleotide of a pre-mRNA transcript, whether the position is a splice acceptor, a splice donor, or neither. The model stacks dilated residual units with increasing kernel size and atrous rate so that a wide genomic context window contributes to each per-nucleotide prediction, while skip connections aggregate multi-scale features. OpenSpliceAI reproduces the predictive behavior of SpliceAI while providing an efficient, modular training pipeline that can be retrained on non-human species.
Variants
OpenSpliceAI ships trained model families for human MANE and four non-human species. Each family provides four
flanking-context sizes.
Model size is determined by flanking context and is shared across species for the same context. FLOPs and MACs are
reported for a single 5,000-nucleotide output sequence.
OpenSpliceAI was trained to predict the location of splice donor and acceptor sites from nucleotide sequence, following the SpliceAI training methodology.
Training Data
The MANE variants were trained on transcripts from the GENCODE/MANE human reference annotation. The non-human variants use the species annotations released by OpenSpliceAI for mouse, zebrafish, honeybee, and Arabidopsis. For each predicted nucleotide, the model receives a flanking context of 80, 400, 2,000, or 10,000 nucleotides, split evenly across the two sides of the output sequence, with sequence ends padded with N. Annotated splice donor and acceptor sites serve as positive labels; all other positions are negative.
Training Procedure
Pre-training
The model was trained to minimize a cross-entropy loss between predicted splice-site probabilities and the reference annotation.
Optimizer: Adam
Loss: cross-entropy
Please refer to the OpenSpliceAI paper for the full training protocol and hardware details.
Citation
bibtex
1@article{chao2025openspliceai,
2 author = {Chao, Kuan-Hao and Mao, Alan and Liu, Anqi and Salzberg, Steven L and Pertea, Mihaela},
3 title = {OpenSpliceAI: An efficient, modular implementation of SpliceAI enabling easy retraining on non-human species},
4 journal = {eLife},
5 volume = 14,
6 pages = {RP107454},
7 year = 2025,
8 doi = {10.7554/eLife.107454.3},
9 publisher = {eLife Sciences Publications, Ltd}
10}
[!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 OpenSpliceAI paper for questions or comments on the paper/model.
