DeepCRISPR: optimized CRISPR guide RNA design by deep learning.
Chuai G, Ma H, Yan J, Yan J, Chen M, Hong N, Xue D, Zhou C, Zhu C, Chen K, Duan B, Gu F, Qu S, Huang D, Wei J, Liu Q. · 2018
Read DeepCRISPR: optimized CRISPR guide RNA design by deep learning. by Chuai G, Ma H, Yan J, Yan J, Chen M, Hong N, Xue D, Zhou C, Zhu C, Chen K, Duan B, Gu F, Qu S, Huang D, Wei J, Liu Q. (2018) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc6020378.txt.
Introduction
Chuai and co-authors present DeepCRISPR, a design platform for gene editing. Their article appeared in Genome Biology. CRISPR uses a guide RNA to direct cutting. A good guide hits its target strongly. It should also avoid cutting at wrong sites.
DeepCRISPR predicts both effects in one framework. It uses deep learning trained on guide data. The system also finds sequence traits that shape success. Epigenetic features, tied to how DNA is packaged, matter too. The authors report it beat existing design tools. Automated feature finding removes much manual guesswork. Readers get a guide to choosing safer, stronger guides. The platform is free online for research use.
Catalog record
| CREATOR | Chuai G, Ma H, Yan J, Yan J, Chen M, Hong N, Xue D, Zhou C, Zhu C, Chen K, Duan B, Gu F, Qu S, Huang D, Wei J, Liu Q. |
|---|---|
| YEAR | 2018 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 72 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
DeepCRISPR: optimized CRISPR guide RNA design by deep learning By Guohui Chuai, Hanhui Ma, Jifang Yan, Ming Chen, Nanfang Hong, Dongyu Xue, Chi Zhou, Chenyu Zhu, Ke Chen, Bin Duan, Feng Gu, Sheng Qu Published in: Genome Biology DOI: 10.1186/s13059-018-1459-4 ABSTRACT A major challenge for effective application of CRISPR systems is to accurately predict the single guide RNA (sgRNA) on-target knockout efficacy and off-target profile, which would facilitate the optimized design of sgRNAs with high sensitivity and specificity. Here we present DeepCRISPR, a comprehensive computational platform to unify sgRNA on-target and off-target site prediction into one framework with deep learning, surpassing available state-of-the-art in silico tools. In addition, DeepCRISPR fully automates the identification of sequence and epigenetic features that may affect sgRNA knockout efficacy in a data-driven manner. DeepCRISPR is available at http://www.deepcrispr.net/. Electronic supplementary material The online version of this article (10.1186/s13059-018-1459-4) contains supplementary material, which is available to authorized users. BACKGROUND CRISPR-based gene knockout is widely implemented in various cell types and organisms. In this system, a single-guide RNA (sgRNA) guides Cas9 proteins to specific genomic targets. Recognition and cleavage occur via complementarity of a 20-nucleotide (nt) sequence within the sgRNA to the genomic target, i.e., the on-target, upstream of a protospacer adjacent motif (PAM) at its 3′ end [1]. Extensive studies have also demonstrated that multiple mismatches as well as DNA or RNA bulges can be tolerated, resulting in cleavage of unintended genomic sites, termed off-targets [2]. Such a CRISPR-Cas9 endonuclease system permits genome editing at nucleotide resolution [3, 4], while a major challenge for its effective application is to accurately predict the sgRNA on-target knockout efficacy and off-target (OT) profile beforehand. Accurate prediction would facilitate the optimized design of sgRNAs by maximizing their on-target efficacy (high sensitivity) and minimizing their off-target effects (high specificity) [1, 2, 5–7]. Various sgRNA design rules and tools have been developed for sgRNA on-target identification and efficacy prediction. These methods are categorized into three types: (1) alignment-based, where the sgRNAs are aligned from the given genome purely by locating the PAM (CasFinder [8], etc.); (2) hypothesis-driven, where the sgRNA knockout efficacies are scored empirically by considering the impact of genome context factors (E-CRISP [9], CRISPR [6], CHOPCHOP [10], GuideScan [11], etc.); and (3) learning-based, where the sgRNA knockout efficacies are predicted from a training model by considering different features (sgRNA Designer [2], SSC [5], sgRNA Scorer [12, 13], CRISPRscan [14], etc.). A benchmark study indicated that the latter two types of tools generally perform better than the alignment-based tool, although the predictions do not scale well in different cell types [15, 16]. Novel computational methods as well as comprehensive exploration of DNA sequence and epigenetic features affecting sgRNA knockout efficacy are required [7, 16]. Off-targets are proven to occur in CRISPR system [2, 6, 17]. Although sgRNA-guided Cas9 cutting at a particular site does not necessarily lead to functional consequences (such as an in-frame shift mutation) [18], how to accurately and quantitatively detect or predict off-target cleavage sites is still an important issue and remains challenging [19]. Most existing tools use simple sequence alignment with different nucleotide mismatches to exhaustively search for off-target sites [20]. A few tools predict cleavage efficacy at the mismatched locus by designing off-target scores (for example, CFD score [2], MIT score [6, 16, 21], etc.). Their prediction results were compared with data generated by whole sgRNA off-target cleavage detection techniques like GUIDE-seq [22], Digenome-seq [16, 23, 24], High-throughput genome-wide translocation sequencing (HTGTS) [25], direct in situ breaks labeling sequencing (BLESS) [26], and integration-deficient lentiviral vector capture (IDLV) [27]. These are essentially hypothesis-based methods that use empirically defined off-target criteria to identify off-target sites. Effective learning-based prediction of the whole genome off-target profile is needed. Currently, building a learning model for sgRNA efficacy prediction faces several obstacles: (1) data heterogeneity issues where effective integration is required for data from different cell types and experimental platforms. (2) data sparsity issues where the labeled sample size, i.e., the amount of sgRNAs with known efficacies, is relatively small and experimentally expensive to collect—insufficiently labeled data makes the current learning models inefficient; (3) data imbalance issues in off-target site prediction—the number of true off-target cleavage sites recognized by whole-genome off-target detection techniques is small among all the possible nucleotide mismatch loci; (4) the leading sequence and epigenetic features affecting sgRNA efficacy remain unclear and await further exploration [5]. Recently, several studies have tried sophisticated learning models for on-target or off-target prediction [28–30], but none of them have addressed these issues thoroughly. In our study, we present a novel and powerful deep learning framework [31–33] to simultaneously predict sgRNA on-target knockout efficacy and a whole-genome off-target cleavage profile that competes favorably with the available state-of-the-art tools. Our approach, called DeepCRISPR, is based on a carefully designed hybrid deep neural network for model training and prediction. To the best of our knowledge, this is the most comprehensive computational platform available to unify on-target and off-target site prediction into one framework with deep learning. We applied a deep unsupervised representation learning technique [32, 34] to automatically learn the underlying representation of sgRNAs using a complete set of genome-wide unlabeled sgRNAs. The learned model was further tuned by a supervised deep neural network using the existing labeled sgRNAs. We point out that pre-training of huge amounts of unlabeled sgRNAs can be used to boost the model prediction, which has never been studied in traditional sgRNA efficacy prediction. Together, DeepCRISPR addresses the above challenges with the following advantages: (1) by considering the epigenetic information in different cell types, it represents different DNA regions from different cell types in a unified feature space and integrates the data from different experiments and cell types. Although DeepCRISPR is trained on limited cell type data, we validated that it has a generally good prediction ability when adapting to new cell types. (2) It learns from billions of genome-wide unlabeled sgRNAs to automatically derive a “parent network”, thus generating a high-level feature representation simultaneously for sgRNA on-target and off-target design. In this way, DeepCRISPR optimizes sgRNA design for both coding and non-coding regions by considering the unsupervised pre-training of genome-wide sgRNA sequences from these regions. The models with and without pre-training were compared and the superiority of unsupervised pre-training was validated. (3) It applies a specific data augmentation technique to generate novel sgRNAs with biologically meaningful labels, thus increasing the labeled training size in sgRNA on-target site prediction. We further validated that such data augmentation indeed improves the prediction performance and makes the training model robust. (4) It further fine-tunes the parent network using the labeled sgRNA data, which helps to boost the prediction performance on limited labeled samples. (5) It integrates an efficient bootstrapping sampling algorithm with the training procedure, dramatically alleviating the data imbalance issue in off-target site prediction. (6) Finally, it fully automates the identification of sequence and epigenetic features. The model learns which features are important in optimized sgRNA design using limited training samples. The identified features can be used for optimized sgRNA design. This helps to decipher the CRISPR on-target and off-target mechanisms in a much more efficient data-driven manner. DeepCRISPR is available at http://www.deepcrispr.net/. The command line code is also available at https://github.com/bm2-lab/DeepCRISPR and https://zenodo.org/record/1246320.