MMSplice

Modular modeling of the effects of genetic variants on splicing.

Disclaimer

This is an UNOFFICIAL implementation of the MMSplice: modular modeling improves the predictions of genetic variant effects on splicing by Jun Cheng, Thi Yen Duong Nguyen, Kamil J. Cygan, Muhammed Hasan Çelik, William G. Fairbrother, Žiga Avsec and Julien Gagneur.

The OFFICIAL repository of MMSplice is at gagneurlab/MMSplice_MTSplice.

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

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

Model Details

MMSplice is a modular neural network for predicting the effect of genetic variants on pre-mRNA splicing. Instead of one monolithic network, MMSplice decomposes an exon together with its flanking introns into five regions and scores each region with an independent small convolutional sub-network:

• acceptor_intron: the intron stub upstream of the 3' splice site.
• acceptor: the 3' splice site (acceptor) region with a short exon flank.
• exon: the exon body.
• donor: the 5' splice site (donor) region with a short exon flank.
• donor_intron: the intron stub downstream of the 5' splice site.

Each sub-network consumes a one-hot encoded DNA sequence (a stack of convolution blocks followed by a small dense head) and emits a single scalar score. The five scalar scores form the module score vector. For variant-effect estimation, the model is run on both the reference and the alternative sequence and the per-module score deltas are combined by the fixed upstream linear model into a delta-logit-PSI splicing-effect score. Please refer to the Training Details section for more information on the training process.

Variant Effect Interface

MMSplice exposes variant effects as an input-schema concern, not a separate output type:

• Reference-only call (input_ids / inputs_embeds): returns the per-module score vector logits of shape (batch_size, 5).
• Reference + alternative call (also pass alternative_input_ids / alternative_inputs_embeds): additionally returns alternative_logits and the per-module deltas delta_logits (alternative_logits - logits).
• MMSpliceForSequencePrediction requires a reference and alternative sequence and returns the upstream scalar delta-logit-PSI score with shape (batch_size, 1).

MMSplice inputs are exon sequences with 100 nt of upstream intronic context and 100 nt of downstream intronic context.

Model Specification

Num Modules	Num Parameters	FLOPs (M)	MACs (M)
5	56,677	5.71	2.79

(FLOPs and MACs measured on a 220 bp exon-with-flanks input.)

Links

• Code: multimolecule.mmsplice
• Paper: MMSplice: modular modeling improves the predictions of genetic variant effects on splicing
• Developed by: Jun Cheng, Thi Yen Duong Nguyen, Kamil J. Cygan, Muhammed Hasan Çelik, William G. Fairbrother, Žiga Avsec, Julien Gagneur
• Original Repository: gagneurlab/MMSplice_MTSplice

Usage

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

pip install multimolecule

Direct Use

Module Scores

>>> import torch
>>> from multimolecule import DnaTokenizer, MMSpliceForSequencePrediction

>>> tokenizer = DnaTokenizer.from_pretrained("multimolecule/mmsplice")
>>> model = MMSpliceForSequencePrediction.from_pretrained("multimolecule/mmsplice")
>>> left_intron = "A" * 100
>>> exon = "C" * 20
>>> right_intron = "G" * 100
>>> reference = tokenizer(left_intron + exon + right_intron, return_tensors="pt")
>>> output = model.model(**reference)
>>> output["logits"].shape
torch.Size([1, 5])

Variant Effect

>>> import torch
>>> from multimolecule import DnaTokenizer, MMSpliceForSequencePrediction

>>> tokenizer = DnaTokenizer.from_pretrained("multimolecule/mmsplice")
>>> model = MMSpliceForSequencePrediction.from_pretrained("multimolecule/mmsplice")
>>> left_intron = "A" * 100
>>> exon = "C" * 20
>>> right_intron = "G" * 100
>>> reference = tokenizer(left_intron + exon + right_intron, return_tensors="pt")
>>> alternative_exon = exon[:10] + "T" + exon[11:]
>>> alternative = tokenizer(left_intron + alternative_exon + right_intron, return_tensors="pt")
>>> output = model(
...     reference["input_ids"],
...     alternative_input_ids=alternative["input_ids"],
... )
>>> output["logits"].shape
torch.Size([1, 1])

Training Details

MMSplice was trained as five independent modules on splicing data and the modules were combined with a linear model to predict variant effects on percent-spliced-in (PSI).

Training Data

The acceptor, donor, exon, and intron modules were trained on splice-site and exon data derived from human reference transcripts. The combining linear model was fit against a massively parallel reporter assay (MPRA) of exon-skipping variants.

Training Procedure

Pre-training

Each module was trained with a sequence-to-scalar objective scoring its region. The module scores (and their reference/alternative deltas) were then combined by a fixed linear model into a delta-logit-PSI splicing-effect score.

Citation

@article{cheng2019mmsplice,
  title     = {MMSplice: modular modeling improves the predictions of genetic variant effects on splicing},
  author    = {Cheng, Jun and Nguyen, Thi Yen Duong and Cygan, Kamil J and {\c{C}}elik, Muhammed Hasan and Fairbrother, William G and Avsec, {\v{Z}}iga and Gagneur, Julien},
  journal   = {Genome Biology},
  volume    = 20,
  number    = 1,
  pages     = {48},
  year      = 2019,
  publisher = {Springer},
  doi       = {10.1186/s13059-019-1653-z}
}

The artifacts distributed in this repository are part of the MultiMolecule project. If you use MultiMolecule in your research, you must 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 MMSplice 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