CRISPR · Software
DeepCRISPR
Predicts SpCas9 sgRNA on-target knockout efficacy and off-target cleavage profiles in human genomic sequences.
What DeepCRISPR is
DeepCRISPR is a deep learning tool designed for guide RNA design in CRISPR experiments. Developed by Guohui Chuai, Qi Liu, and colleagues, it predicts both the on-target knockout efficacy of single guide RNAs (sgRNAs) and their genome-wide off-target cleavage profiles.
Selecting sgRNAs with high cleavage activity and minimal unintended genomic cuts is a central challenge in genome editing. DeepCRISPR addresses this by modelling conventional NGG-based target sites for SpCas9 in human species, evaluating target efficacy through sequence composition and epigenetic context.
How it works
DeepCRISPR uses a hybrid deep neural network framework based on convolutional neural networks (CNNs), trained using data augmentation and pre-training strategies. Target regions are represented as 23-nucleotide sequences encoded into tensors with dimensions of batch size, channels, 1, and 23. For on-target predictions, inputs can be formatted using 4 channels for sequence-only data or 8 channels incorporating selected epigenetic features sourced from ENCODE.
The software provides both classification and regression models implemented through dedicated classes for on-target prediction (DCModelOntar) and off-target evaluation (DCModelOfftar). Off-target prediction operates exclusively on full-featured 8-channel inputs, scoring cleavage likelihood across guide and off-target sequence pairs.
Key facts
- Target enzyme
- SpCas9 (NGG PAM)
- Target species
- Human
- Input tensor shape
- [batch_size, channels, 1, 23]
- Input channels
- 4 channels (sequence only) or 8 channels (sequence and epigenetic features)
- Core architecture
- CNN with pre-training and data augmentation
- Software dependencies
- Python 3.6, TensorFlow 1.3.0, Sonnet 1.9
What you can use it for
- On-target knockout efficacy regression
- Predicting continuous knockout efficacy scores for SpCas9 candidate guides using either nucleotide sequence alone or combined sequence and epigenetic features.
- On-target guide classification
- Classifying potential sgRNAs into efficacy categories using pre-trained convolutional neural network models.
- Genome-wide off-target cleavage profiling
- Evaluating the likelihood of cleavage at off-target genomic sites by comparing sgRNA targets against putative off-target loci using 8-channel epigenetic and sequence encodings.
Run it here
The research agent sets DeepCRISPR up on its machine the first time a task needs it, then runs it and reads the output back to you. Nothing to install on your side.
Try asking
“Design and rank SpCas9 guides for this region with DeepCRISPR.”
Strengths and limits
Good at
- Supports both classification and regression models for predicting on-target knockout efficacy and off-target cleavage.
- Integrates cell-type-specific epigenetic markers alongside sequence data to inform cleavage predictions.
- Offers a sequence-only model alternative for on-target scoring when experimental epigenetic data is unavailable.
Watch out for
- Limited to human SpCas9 targets containing conventional NGG protospacer adjacent motifs.
- Off-target cleavage prediction requires the full 8-channel input and cannot run on sequence data alone.
- Depends on legacy software frameworks, specifically Python 3.6, TensorFlow 1.3.0, and Sonnet 1.9.
Questions
Can I run DeepCRISPR online without installing it?
The research agent sets DeepCRISPR up on its machine the first time a task needs it, then runs it and reads the output back to you. Nothing to install on your side.
What licence is DeepCRISPR under?
As published by the authors — see the bm2-lab/DeepCRISPR repository. The terms are published at https://github.com/bm2-lab/DeepCRISPR.
Which PAM sequences and Cas variants are supported?
DeepCRISPR specifically focuses on SpCas9 guide RNA design using conventional NGG protospacer adjacent motifs in human genomic DNA.
Can predictions be run without epigenetic datasets?
Yes, but only for on-target predictions using the dedicated 4-channel sequence-only regression model. Off-target prediction models strictly require the full 8-channel representation containing epigenetic features.
What is the expected input shape for guide sequences?
Input arrays must follow the shape [batch_size, channels, 1, 23], corresponding to the 23-nucleotide target site and PAM across 4 or 8 feature channels.
Where are the required epigenetic features obtained?
Epigenetic annotations used for the 8-channel encoding schema can be retrieved from the ENCODE project.
How to cite it
Guohui Chuai, Qi Liu et al. DeepCRISPR: optimized CRISPR guide RNA design by deep learning. 2018
Written from DeepCRISPR's own README and project page. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.