🧬 SCALPEL: S͎pecific CRISPR-C͎as13d gRNA͎ design through deep L͎earning P͎rediction using endogenous E͎xperimentaL͎ RNA structure and binding information
Deciphering cellular context for efficient and cell type-specific CRISPR-Cas13d gRNA design using endogenous RNA structure and deep learning*
The efficacy and tissue specificity of RNA therapeutics are critical for clinical translation. By large-scale profiling of the dynamic RNA structurome across four cell lines, we systematically characterized the impact of endogenous target RNA structure and RNA-protein interactions on CRISPR/Cas13d gRNA activity. We identified the structural patterns of high-efficacy gRNA targets and observed that structural differences can lead to variations in efficacy across different cellular contexts. By stabilizing single-stranded structure, RNA-binding proteins also enhanced gRNA efficacy.
Leveraging this cell context information, along with approximately 290K RfxCas13d screening data, we developed SCALPEL, a deep learning model that predicts gRNA performance across various cellular environments. SCALPEL integrates:
| Feature | Description |
|---|---|
| 🔤 Sequence | Both target and gRNA sequence |
| 🧬 Structure | Endogenous icSHAPE data across different cell lines |
| 🔗 Binding | Cell type-specific RBP binding profiles |
SCALPEL significantly outperforms existing models and, most importantly, enables cell type-specific prediction of gRNA activity. Validation screens across multiple cell lines demonstrate that cellular context significantly influences gRNA performance, underscoring the feasibility of cell type-specific knockdown by targeting structural dynamic regions. SCALPEL also facilitates designing highly efficient virus-targeting gRNAs and gRNAs that robustly knockdown maternal transcripts essential for early zebrafish development.
Official Online Predicted Website: SCALPEL
- 🧬 Integrates endogenous RNA secondary structure data and cell type-specific RBP binding profiles
- 🤖 Transformer-based architecture for deep context modeling
- 🎯 Accurately predicts on-target effects of gRNAs
- 🔬 Assists in designing high-specificity gRNAs for different cellular contexts
- 🐟 Facilitates the design of gRNAs for animal models and patient-derived organoids
SCALPEL/
├── dataset/ # PyTorch dataloader and pre-processing
├── ckpt/ # PyTorch training model weights
├── models/ # SCALPEL model architecture
├── other_model_data/ # Training data path
├── loss_functions/ # Optional loss functions
├── Figs/ # Figures and diagrams
└── logicArchi.py # Main entry point
Download the following pre-trained weights and place them in the specified directories:
- BERT model → Google Drive — place files in the root directory
- Trained SCALPEL checkpoints (
m9_m1) → Google Drive — place files underckpt/
Prepare your input data following the format in other_model_data/ours/demo_data.csv. Each entry should include:
- RNA sequences
- icSHAPE reactivity profiles
- RBP-binding tracks or matrices from PrismNet
- Other features
The icSHAPE sequencing data for all cell lines have been deposited in NCBI GEO under accession GSE301234. The validation screening data are available under GSE30081, and in NGDC under PRJCA042228.
pip install -r requirements.txt # Python 3.9The SCALPEL model class is implemented as m9_m1 in models/m9.py. Use --model='m9_m1' to select it.
Hardware Requirements: Due to fine-tuning of the BERT model, training is computationally intensive. Our experiments used 96 GB RAM and NVIDIA A100 (48 GB) / L40S (48 GB) GPUs. A full 10-fold cross-validation training run takes over 140 hours. We recommend using hardware at or above this specification.
python3 -u logicArchi.py control \
--gpu_ids=[0] \
--model_des='seq_bert_fold_mfe1_mfe2_icshape_binding_relatelen_utrrate' \
--data_des='random' \
--model='m9_m1' \
--dataset='BertOnehotLoader30' \
--lr=1e-3 \
--data_path='other_model_data/ours/demo_data.csv'Note:
model_desconfigures which features to include.data_desacceptscell_line:{},match[i],random, ortarget_gene— refer totrain.shfor details. Remember to update theoriginpath in thecontrolfunction inlogicArchi.pyto point to your data files.
python3 -u logicArchi.py val \
--gpu_ids=[0] \
--model_des='seq_bert_fold_mfe1_mfe2_icshape_binding_relatelen_utrrate' \
--data_des='random' \
--model='m9_m1' \
--dataset='BertOnehotLoader30' \
--data_path={Your validation file path}Refer to val.sh for more usage examples.
Fig. Compared model performance.
We sincerely thank the following contributors and institutions for their support:
Collaborating Labs and Institutes We thank all members of Sunlab at Shandong University for their insightful discussions.
Funding Support This work was supported by the National Natural Science Foundation of China (No. 32300521, No. 32422013, and No. 82341086); the Open Grant from the Pingyuan Laboratory (No. 2023PY-OP-0104); the State Key Laboratory of Microbial Technology Open Projects Fund (No. M2023-20); the Intramural Joint Program Fund of the State Key Laboratory of Microbial Technology (No. SKLMTIJP-2024-02); the Double-First Class Initiative of Shandong University School of Life Sciences; the Young Innovation Team of Shandong Higher Education Institutions; the Taishan Scholars Youth Expert Program of Shandong Province; and the Program of Shandong University Qilu Young Scholars.
Open-source Tools This project builds upon many open-source tools and libraries, including PyTorch, Scikit-learn, and Biopython.
Special thanks to all community members and beta testers who provided feedback during model development and validation in Sunlab.
-
Cheng, Xiaolong, et al. "Modeling CRISPR-Cas13d on-target and off-target effects using machine learning approaches." Nature Communications 14.1 (2023): 752.
-
Wei, Jingyi, et al. "Deep learning and CRISPR-Cas13d ortholog discovery for optimized RNA targeting." Cell Systems 14.12 (2023): 1087–1102.
-
Wessels, Hans-Hermann, et al. "Prediction of on-target and off-target activity of CRISPR–Cas13d guide RNAs using deep learning." Nature Biotechnology 42.4 (2024): 628–637.
-
Zhu, Haoran, et al. "Dynamic characterization and interpretation for protein-RNA interactions across diverse cellular conditions using HDRNet." Nature Communications 14.1 (2023): 6824.
-
Yin, Weijie, et al. "Ernie-RNA: An RNA language model with structure-enhanced representations." bioRxiv (2024): 2024-03.
If you use SCALPEL in your research, please cite:
***This project is released under the MIT License.
Thank you for using SCALPEL! Questions, suggestions, and feedback are always welcome.