CRISPR-Cas9 off-target prediction engine
Art by me btw ^__^
The CRISPR-Cas system is an RNA-guided nuclease commonly used to introduce a single-stranded break (ssDNA nick) or double-stranded break (dsDNA break). The guide RNA is typically 20 nucleotides (nts) long, and the genomic locus must often be adjacent to a PAM (Protospacer Adjacent Motif) site for DNA cleavage to occur.
The RNA sequence and PAM site determine the specificity of the nuclease. However, many Cas and Cas-like nucleases exhibit off-target activity, meaning they can occasionally bind to unintended sites resembling the target sequence. When the CRISPR-Cas system localizes to such unintended sites and performs its function, this is referred to as an off-target effect.
In in vitro methods, if off-target editing occurs in protein-coding regions of a gene, it can lead to significant unintended consequences. Thus, predicting the likelihood of off-target effects is critical.
This problem cannot be solved solely by algorithmic sequence alignment methods, as off-target effects depend not only on nucleotide complementarity between the RNA and DNA but also on more complex mechanisms.
Machine learning (ML) approaches are effective because, when trained on large datasets of experimentally observed off-target effects, models learn to capture hidden factors influencing targeting.
Input: Two sequences, sgRNA and targetDNA, represented as strings of nucleotides (A, T, G, C) approximately 20 characters in length (variable size).
Output: A binary prediction (0 or 1) indicating the presence (1) or absence (0) of an off-target effect.
Models in this project are evaluated via Accuracy, AUROC, Precision and Recall.
the main CRISPR-BERT model is expected to perform around following:
| Model | Val Accuracy | Val AUROC | Val Precision | Val Recall |
|---|---|---|---|---|
| Bidirectional LSTM with BERT embedding | 0.86 | 0.90 | 0.77 | 0.56 |
The main model is CRISPR-BERT based on architecture suggested in the paper Orhan Sari, Ziying Liu, Youlian Pan, Xiaojian Shao, Predicting CRISPR-Cas9 off-target effects in human primary cells using bidirectional LSTM with BERT embedding, Bioinformatics Advances, Volume 5, Issue 1, 2025, vbae184, https://doi.org/10.1093/bioadv/vbae184 (https://academic.oup.com/bioinformaticsadvances/article/5/1/vbae184/7934878)
Stack encoding was used to encode the CRISPR-Cas9 sgRNA–DNA sequence pairs, after which the data is passed into a BERT embedding layer. The output of BERT is fed into a BiLSTM layer. The output of the BiLSTM is then passed through dense layers with a sigmoid activation at the end to predict the presence of off-target effects.
Final project runs as an ML Flow inference server with interactive frontend. Please follow instructions below to run it locally on your machine.
To run the program, follow these steps:
-
Clone repository and enter project dir:
git clone https://github.com/tenebrissilvam/CRISPR-off-t-engine.git cd /CRISPR-off-t-engine -
Activate the Virtual Environment: First, you need to activate the virtual environment of
uv. This ensures that all the necessary dependencies are available.pip install uv uv venv --python $(cat .python-version) --prompt crispr-off-t-engine source .venv/uv/bin/activate uv sync
- You can download csv file of Change_seq dataset from the following link: https://1drv.ms/x/c/695ce78a94063f8c/EbdYXZVnrT5ElMKf6cKtlugBznQbWkVrqr7tNWPrDLpcmg?e=iHYCsy
and paste it in the project directory ./data/Change_seq.csv
-
Or you can download it automatically via script to data/Change_seq.py
uv run download_data.py
- Training After environment preparation navigate to conf folder, switch label field at conf/mode/default.yaml to train and modify other parameters if needed, then run following command from the root of the repository to train the model.
uv run src/run_scripts/train.pyPlots visualisation To view loss plits and AUROC, recall and precision metrics first navigate to conf/logging/default.yaml and specify your wandb username and project name.Then run
uv run src/utils/data_visuals.pyMlFlow server preparation You need to build the model first via command
uv run mlflow-serve/mlflow/mlflow_model_wrapper.pyONNX model conversion To convert model into onnx format using pretrained weights run following command to save model to inference_model_modifications/onnx/crispr_detector.onnx
uv run inference_model_modifications/onnx/onnx_model_conversion.pyTensorRT 0. First pull docker image via
docker pull nvcr.io/nvidia/tensorrt:23.04-py3- Fix onnx model to have int32 weights via
cd /inference_model_modifications/onnx
uv run model_surgeon.py
cd ..
cd ..- then run interactive container
docker run -it --rm --gpus device=0 -v $(pwd)/inference_model_modifications/onnx:/models nvcr.io/nvidia/tensorrt:23.04-py3- then inside the container run
trtexec \
--onnx=/models/crispr_detector_fixed.onnx \
--saveEngine=/models/crispr_detector_FP32.plan \
--minShapes=input:1x22 \
--optShapes=input:8x22 \
--maxShapes=input:16x22 \
--profilingVerbosity=detailed \
--builderOptimizationLevel=5 \
--hardwareCompatibilityLevel=ampere+Run the Script: for model inference on the provided dataset put csv dataset path in conf/data/default.yaml and switch label field at conf/mode/default.yaml to train. Also make sure you placed model weights correct path at conf/mode/default.yaml. Then run
uv run src/run_scripts/infer.pyMlFlow inference
-
Run mlflow server with the model via
mlflow models serve -m mlflow-serve/mlflow/crispr_off_t_model -p 8888 --host 0.0.0.0 --no-conda
-
Run mlflow model in a docker via
sudo docker build -t mlflow-app -f mlflow-serve/mlflow-serve/mlflow/Dockerfile . sudo docker run -d -p 8888:8888 --name mlflow-server mlflow-app
Then you can send requests using following format
curl -X POST http://localhost:8888/invocations -H "Content-Type: application/json" -d '{
"inputs": [
{"sequence":"GTCACCAATCCTGTCCCTAGTGG",
"Target sequence": "TAAAGCAATCCTGTCCCCAGAGT"
}
]
}'
- To run all services use
sudo docker-compose -f mlflow-serve/docker-compose.yml up --build
Then navigate to the http://localhost:9000 in your browser to use user friendly RNA and DNA site input to predict off-target effect
- 07.06.2025
- Added ONNX conversion
- Added autodownload script for the data
- 10.05.2025
- Added frontend to ML Flow server
- 20.04.2025
- Added Ml Flow inference
- Added Dockerised model
- Added support for curl requests for the model
-
31.03.2025
- Added Hydra configs.
- Created run script.
- Integrated DVC for version control.
-
30.03.2025
- Added project description.
- Implemented metrics for performance evaluation on the Change-seq dataset for all four methods.
-
23.03.2025
- Added R-Crispr method.
- Added Deep-CNN method.
-
17.03.2025
- Added Crispr DipOff method code.
-
15.03.2025
- Added bidirectional LSTM with BERT embeddings method code.
This code is based on methods presented in following papers:
- Lin J, Wong KC. Off-target predictions in CRISPR-Cas9 gene editing using deep learning. Bioinformatics. 2018 Sep 1;34(17):i656-i663. doi: 10.1093/bioinformatics/bty554. PMID: 30423072; PMCID: PMC6129261.
- Orhan Sari, Ziying Liu, Youlian Pan, Xiaojian Shao, Predicting CRISPR-Cas9 off-target effects in human primary cells using bidirectional LSTM with BERT embedding, Bioinformatics Advances, Volume 5, Issue 1, 2025, vbae184, https://doi.org/10.1093/bioadv/vbae184
- Md Toufi kuzzaman, Md Abul Hassan Samee, M Sohel Rahman, CRISPR-DIPOFF: an interpretable deep learning approach for CRISPR Cas-9 off-target prediction, Briefi ngs in Bioinformatics, Volume 25, Issue 2, March 2024, bbad530, https://doi.org/10.1093/bib/bbad530
- Niu R, Peng J, Zhang Z, Shang X. R-CRISPR: A Deep Learning Network to Predict Off-Target Activities with Mismatch, Insertion and Deletion in CRISPR-Cas9 System. Genes (Basel). 2021 Nov 25;12(12):1878. doi: 10.3390/genes12121878. PMID: 34946828; PMCID: PMC8702036.