PrimeNet is a deep learning model designed to predict Prime editing efficiency with high accuracy by integrating epigenetic factors such as chromatin accessibility and DNA methylation. Unlike existing models, PrimeNet leverages multi-scale convolution and attention mechanisms to extract diverse sequence features, achieving superior generalization across different conditions. Our results show that PrimeNet achieves a Spearman correlation coefficient of 0.94, outperforming existing models.
- Sequence: Wild Sequence and the Edited Sequence.
- Epigenetic information: chromatin accessibility and DNA methylation status. For chromatin accessibility features, a site is labeled "Y" if it is chromatin accessible and "N" otherwise. For methylation features, a site is labeled as "Y" if it is methylated and "N" if it is unmethylated.
- Functional area information: The positions of Protospacer, PAM, PBS, and RT in the sequence.
We encode the input data as an image for deep learning processing:
- valid editing efficiency: The proportion of sequences that are successfully edited as intended, meaning the desired genetic modification is correctly introduced.
- unedited efficiency: The proportion of sequences that remain unchanged, indicating that Prime editing did not occur.
- erroneous editing efficiency: The proportion of sequences that were edited incorrectly, meaning unintended modifications occurred instead of the expected edit.
# Clone the repository
git clone https://github.com/bm2-lab/PrimeNet.git
cd PrimeNet# Ensure you have Python 3.8+ installed, then run:
pip install numpy pandas torch scikit-learn scipy- PrimeNet.pth: The trained model weights file.
- model.py: Defines the PrimeNet architecture and forward pass.
- train_val.py: Training and validation script for PrimeNet.
- test.py: Inference script for loading a trained model and evaluating on held-out data.
- data/: Raw and processed datasets used for training and testing.
- DeepPE/: Retrained DeepPE model and data.
- PRIDICT/: Retrained PRIDICT model and data.
- deepprime/: Retrained deepprime model and data.
In addition to the full‐size PrimeNet teacher model, we are implementing a model distillation workflow to produce lightweight “student” networks with varying parameter counts. These distilled models are optimized to run efficiently on CPU while retaining high accuracy.
This project is licensed under the Apache License. See LICENSE for details.
For questions or collaborations, please contact [email protected] or open an issue on GitHub.