PRIDICT2.0: PRIme editing guide RNA preDICTion 2.0

Overview

PRIDICT2.0 is an advanced version of the original PRIDICT model designed for predicting the efficiency of prime editing guide RNAs. This repository allows you to run the model locally. For details on advancements over the original model, refer to our published study (Mathis et al., Nature Biotechnology, 2024) and the initial BioRxiv preprint.

Complementary Model

• ePRIDICT: This model focuses on the influence of local chromatin context (K562) on prime editing efficiencies and is designed to complement PRIDICT2.0. Access GitHub Repository

Resources

• Supplementary Files: Access Here
• Web Application: For an online version of PRIDICT2.0, visit our webapp.

Contact

For questions or suggestions, please either:

• Email us at [email protected]
• Open a GitHub issue

Citation

If find our work useful for your research please cite:

• Mathis et al., Nature Biotechnology, 2024 (PRIDICT2.0):

@article{mathis2024machine,
  title={Machine learning prediction of prime editing efficiency across diverse chromatin contexts},
  author={Mathis, Nicolas and Allam, Ahmed and Tálas, András, and Kissling, Lucas and Benvenuto, Elena and Schmidheini, Lukas and Schep, Ruben and Damodharan, Tanav and Balázs, Zsolt and Janjuha, Sharan and others},
  journal={Nature Biotechnology},
  year={2024},
  publisher={Nature Publishing Group},
  doi={10.1038/s41587-024-02268-2}
}

• Mathis & Allam et al., Nature Biotechnology, 2023 (PRIDICT):

@article{mathis2023predicting,
  title={Predicting prime editing efficiency and product purity by deep learning},
  author={Mathis, Nicolas and Allam, Ahmed and Kissling, Lucas and Marquart, Kim Fabiano and Schmidheini, Lukas and Solari, Cristina and Balázs, Zsolt and Krauthammer, Michael and Schwank, Gerald},
  journal={Nature Biotechnology},
  volume={41},
  number={8},
  pages={1151--1159},
  year={2023},
  publisher={Nature Publishing Group},
  doi={10.1038/s41587-022-01613-7}
}

Getting Started

Installation using Anaconda (Linux, Mac OS or WSL) 🐍

Windows is only supported via Windows Subsystem for Linux (WSL).

The easiest way to install and manage Python packages on various OS platforms is through Anaconda. Once installed, any package (even if not available on Anaconda channel) could be installed using pip.

• Install Anaconda or miniconda. (conda 22.11 or newer)

• Start a terminal and run:

  # clone PRIDICT2.0 repository
  git clone https://github.com/uzh-dqbm-cmi/PRIDICT2.git
  # navigate into repository
  cd PRIDICT2
  # create conda environment and install dependencies for PRIDICT2 (only has to be done before first run/install)
  conda env create -f pridict2_repo.yml
      
  # activate the created environment
  conda activate pridict2
  
  # after activating environment, pytorch has to be installed separately here:
  pip install torch==2.0.1 --index-url https://download.pytorch.org/whl/cpu
  
  # run desired PRIDICT2.0 command (single or batch mode, described below)
  python pridict2_pegRNA_design.py single --sequence-name seq1 --sequence "GCCTGGAGGTGTCTGGGTCCCTCCCCCACCCGACTACTTCACTCTCTGTCCTCTCTGCCCAGGAGCCCAGGATGTGCGAGTTCAAGTGGCTACGGCCGA(G/C)GTGCGAGGCCAGCTCGGGGGCACCGTGGAGCTGCCGTGCCACCTGCTGCCACCTGTTCCTGGACTGTACATCTCCCTGGTGACCTGGCAGCGCCCAGATGCACCTGCGAACCACCAGAATGTGGCCGC"
  # results are stored in 'predictions' folder

• PRIDICT2.0 environment only has to be installed once. When already installed, follow the following commands to use PRIDICT2.0 again:

  # open Terminal/Command Line
  # navigate into repository
  # activate the created environment
  conda activate pridict2
  # run desired PRIDICT2.0 command (single or batch mode, described below)
  python pridict2_pegRNA_design.py single --sequence-name seq1 --sequence "GCCTGGAGGTGTCTGGGTCCCTCCCCCACCCGACTACTTCACTCTCTGTCCTCTCTGCCCAGGAGCCCAGGATGTGCGAGTTCAAGTGGCTACGGCCGA(G/C)GTGCGAGGCCAGCTCGGGGGCACCGTGGAGCTGCCGTGCCACCTGCTGCCACCTGTTCCTGGACTGTACATCTCCCTGGTGACCTGGCAGCGCCCAGATGCACCTGCGAACCACCAGAATGTGGCCGC"
  # results are stored in 'predictions' folder

---

Running PRIDICT2.0 in 'single' mode:

Required:

• --sequence-name: name of the sequene (i.e. unique id for the sequence)
• --sequence: target sequence to edit in quotes (format: "xxxxxxxxx(a/g)xxxxxxxxxx"; minimum of 100 bases up and downstream of brackets are needed; put unchanged edit-flanking bases outside of brackets (e.g. xxxT(a/g)Cxxx instead of xxx(TAC/TGC)xxx)

Optional:

• --output-dir: output directory where results are dumped on disk (default: ./predictions; directory must already exist before running)
• --use_5folds: Use all 5-folds trained models. Default is to use fold-1 model
• --nicking: Additionally, design nicking guides for edit (PE3) with DeepSpCas9 prediction (Kim et al. 2019).
• --ngsprimer: Additionally, design high-throughput sequencing (NGS) primers for edit based on Primer3 design.

Example command:

python pridict2_pegRNA_design.py single --sequence-name seq1 --sequence "GCCTGGAGGTGTCTGGGTCCCTCCCCCACCCGACTACTTCACTCTCTGTCCTCTCTGCCCAGGAGCCCAGGATGTGCGAGTTCAAGTGGCTACGGCCGA(G/C)GTGCGAGGCCAGCTCGGGGGCACCGTGGAGCTGCCGTGCCACCTGCTGCCACCTGTTCCTGGACTGTACATCTCCCTGGTGACCTGGCAGCGCCCAGATGCACCTGCGAACCACCAGAATGTGGCCGC"

---

Running in batch mode:

For instructions on running PRIDICT2.0 in batch mode on an HPC cluster with SLURM (e.g. with >1000 inputs), see our HPC Batch Guide.

Required:

• --input-fname: input file name - name of csv file that has two columns [editseq, sequence_name]. See batch_template.csv in the ./input folder

Optional:

• --input-dir : directory where the input csv file is found on disk
• --output-dir: directory on disk where to dump results (default: ./predictions)
• --use_5folds: Use all 5-folds trained models. Default is to use fold-1 model
• --cores: Number of batch input sequences that can be predicted simultaneously. Maximum 3 cores due to memory limitations. Default value 0 uses 3 cores if available.
• --nicking: Design nicking guides for edit (PE3) with DeepSpCas9 prediction (Kim et al. 2019).
• --ngsprimer: Additionally, design high-throughput sequencing (NGS) primers for edit based on Primer3 design.
• --summarize: Summarize the best scoring pegRNA(s) of each batch input in a summary file (saved in output folder in the format date +time + HEK or K562 + batch_summary.csv). Choose either HEK or K562 to get highest scoring pegRNAs based on desired PRIDICT2.0 score. (e.g., --summarize K562)
• --summarize_number: Define the number of top scoring pegRNAs to be included in summary file. Default is 3.

Example command (including only required arguments):

 python pridict2_pegRNA_design.py batch --input-fname batch_template.csv

Notes:

• A log file will be saved in log file folder (date +time + batch filename (without file extension) + _batch_logfile.csv). Check the column log to catch any inputs which contained errors. If everything worked well, it should read Prediction successful! in each row.
• To run PRIDICT2.0 from outside the repository folder, add path to script file (e.g. /mnt/c/Users/pridictuser/github/PRIDICT2/pridict2_pegRNA_design.py) and use --input-dir to define input folder (e.g. /mnt/c/Users/pridictuser/github/PRIDICT2/input)

---

Prediction of pegRNAs with silent bystander edits

PRIDICT2.0's enhanced ability to predict multi-bp edits allows researchers to improve their outcomes by incorporating silent bystander edits into their targets. These edits can potentially boost editing efficiencies, particularly in MMR proficient contexts.

To facilitate this process, we provide a Jupyter Notebook notebook_silent_bystander_input.ipynb in the silentbystander_addon folder. This notebook generates all possible PRIDICT2.0 input sequences with silent bystanders up to 5bp up- and downstream of the edit.

Requirements:

• PRIDICT input sequence with 150 bp context on both sides of the edit
• Input sequence must be in-frame (ORF_start = 0) if the edit and its context are within an exon
• If not in an exon, still assume it is "in frame" and also keep the in frame value as yes if you do batch input file --> to get all possible mutations for an intron/non-genic region, enter your input sequence in all ORF variants (3 for the fw strand; 3 for the rv strand) and run the batch mode.
• PRIDICT2.0 conda environment

Usage overview:

1. Use the notebook or corresponding command line script to create an input batch file for PRIDICT2.0 prediction with silent bystanders.
2. Supports 1bp replacements and multi-bp replacements (not insertions/deletions).
3. Input format: PRIDICT2.0 format with 150bp flanking bases on both sides.
4. Choose between single (single mutation/edit) or batch (multiple inputs) functions.
5. Run PRIDICT2.0 in batch mode in a terminal outside of notebook with the generated input sequences to obtain efficiency predictions. (Optional: run PRIDICT2.0 from within notebook, see commented out example)
6. Summarize individual bystander predictions to one prediction file (best prediction of each bystander variant)