piCRISPR

piCRISPR is a Python-based prediction tool for CRISPR off-target effects using physically informed features. It is developed by Florian Störtz at the group of Dr. Minary, University of Oxford.

Usage

piCRISPR can be run as

python picrispr.py test_input.csv

Positional arguments are:

• input file: path to input csv file
• model number: integer between 0 and 7 to choose which model is used, order corresponds to the one within the publication
• path to models: path to folder where models are saved, default is models
• regression: True or False

Example input file (comma-separated):

target_sequence,grna_target_sequence
AAGGCGCATAAAGATGAGGCGCTGG,GACGCATAAAGATGAGACGCTGG
AAACGCATAAAGATGAGACGCTGGG,GACGCATAAAGATGAGACGCTGG
GGTGATAAGTGGAATGACCATGTGG,GTGATAAGTGGAATGCCATGTGG
GTGATAAGCTGGAATGCCATTGTGG,GTGATAAGTGGAATGCCATGTGG

Further features can be supplied when named in the top line:

experiment_id,target_sequence,grna_target_sequence,epigen_ctcf,epigen_dnase,epigen_rrbs,epigen_h3k4me3,epigen_drip,energy_1,energy_2,energy_3,energy_4,energy_5
0,AAGGCGCATAAAGATGAGGCGCTGG,GACGCATAAAGATGAGACGCTGG,0.0,0.0,0.0,0.0,0.0,20.12,0.9369158778408675,0.9369158778408675,8.767159085783554,20.12
0,AAACGCATAAAGATGAGACGCTGGG,GACGCATAAAGATGAGACGCTGG,0.0,0.0,0.0,0.0,0.0,0.75,-32.84039169058475,-32.84039169058475,-25.639584889882755,0.75
1,GGTGATAAGTGGAATGACCATGTGG,GTGATAAGTGGAATGCCATGTGG,0.0,0.0,0.0,0.0,0.0,-2.83,-27.117216094014633,-27.117216094014633,-21.500510272206174,-0.33000000000000007
1,GTGATAAGCTGGAATGCCATTGTGG,GTGATAAGTGGAATGCCATGTGG,0.0,0.0,0.0,0.0,0.0,-13.614999999999998,-49.92028390034596,-49.92028390034596,-44.50227987453346,-11.114999999999998

The maximum set of features is given in test_input.csv.

Running piCRISPR like this results in an output file output.csv with columns

• piCRISPR prediction: binary label or score between 0 and 1, depending on whether regression mode was chosen
• ground truth: true label, if given in the input file as column cleavage_freq
• ground truth_transformed: true label transformed to lie between 0 and 1, if a true label has been given in the input file as column cleavage_freq

Requirements

python==3.8.3, jupyter-notebook==6.0.3, torch==1.7.0, tensorflow==2.3.1, sklearn==0.23.1, scipy==1.5.0, numpy==1.18.5, pandas==1.0.5, xgboost==1.4.2, matplotlib==3.2.2, pickle==4.0

Installation

Unzip models/models_torch.zip.

Data preparation

In order to predict on custom off-target cleavage data, it must be annotated with epigenetic markers and physically informed features as detailed in the publication. We provide a readily annotated dataset in the file offtarget_260520_nuc.csv.zip which contains a zip-compressed pandas dataframe and can be loaded using pd.read_csv('offtarget_260520_nuc.csv.zip').