Application of AI algorithms for off-target prediction in CRISPR-Cas9 gene editing

“Report on the seminar given by Jeremy Charlier on February 22^th 2022”

BIF7002

Written by: 

Wasmi Algasim, Salwa Haidar, and Zeinab Sherkatghanad

UQAM - Winter 2022

table of contents

1. Introduction

2. CRISPR/Cas9 Genome editing

3. Application of AI algorithms on target prediction

3.1 Sequence Encoding

3.2 Application of Machine learning

3.3 Application of Deep learning

4. Off-target prediction with novel sgRNA-DNA sequence encoding in CRISPR-Cas9 gene editing

5. Conclusion

6. Reference

1. Introduction

Discovery of genome editing technology has revolutionized molecular biology and genetics.

Genome editing or gene editing represents a set of approaches that permit the manipulation of the genome of living organisms. Gene editing allows to make specific changes in DNA sequences such as DNA insertion, deletion, modification of certain nucleotides or replacement of a specific DNA fragment [1].

Gene editing techniques rely on the use of particular enzymes called nucleases that can be directed to target a specific DNA sequence and enable gene editing at the cut sites. DNA nucleases induce double-strand breaks (DSB) at the DNA level that can be repaired by one of the 2 major DNA repair mechanisms in the cell, homology directed repair (HDR) and non-homologous end joining (NHEJ) repair, frequently accompanied by indels (insertion/deletion) at the cut site [2].

Zinc-finger nucleases (ZFNs) and transcription activator-like effector nucleases (TALENs) were the two main strategies, based on protein-DNA interaction, used for gene editing between 1985 and 2012 [3] [4] [5].

Genome editing technologies have evolved rapidly in the past few years leading to the discovery of a more powerful and efficient tool for DNA manipulation called CRISPR/Cas9. CRISPR/Cas9 system has been launched in 2012 and it is based on RNA-DNA complementarity, instead of protein-DNA interaction, for target DNA sequence recognition. It represents the most popular gene editing technique nowadays and it improved our understanding of several gene roles, biological functions and genetical diseases [5] [6].

Significant properties of CRISPR-associated protein 9 (Cas9) such as flexibility, cost-efficiency, simplicity and ability to remove more than one gene at a time have evoked great enthusiasm on this technique. CRISPR/Cas9 as a superior gene editing technology, has been rapidly developed in various research fields, ranging from basic research on genetic therapies at the cellular level to applied biomedical research [6] [7]. Beside clinical potential of CRISPR/Cas9 for treating human disease such as cancer, genetic disorders and beyond [8] [9] [10], it has efficient achievements in genetic engineering for plants [11] and in animal disease models [12]. A Schematic of CRISPR/Cas9 genome editing system and its applications has been provided in Figure 1.

Figure 1. CRISPR/Cas9 genome editing system and its application domains.

Adapted from [1].

2. CRISPR/Cas9 Genome editing

CRISPR-Cas9 is an adaptive immune system used in some bacteria and archaea to defend themselves against foreign DNA from invading viruses. CRISPR, stands for Clustered Interspaced Short Palindromic Repeats, consists of a succession of repeats separated by distinct sequences called “spacers” from viral genomes. Cas9 (CRISRP-associated protein 9) is an endonuclease that cuts DNA within 20 nucleotides located immediately before a protospacer adjacent motif (PAM) composed of 3 nucleotides (NGG) (Figure 1) [13] [14].

CRISPR/Cas9 system has two main components: a Cas9 protein responsible of cleaving the double strands of DNA sequences and a single guide RNA (sgRNA), composed of 20 nucleotides, directing the Cas9 to the target sequence in the genome and ensuring cutting takes place in the right place [1]. The Protospacer-Adjacent Motif (PAM) which is a 3 nucleotide motif located at the end of the DNA target site, is required for Cas9 protein to cleave at a specified site [15]. The CRISPR/Cas9 is quite accurate in gene editing. sgRNA can precisely edit the target site (i.e. on-target editing), though it may bind and wrongly edit at other additional sites leading to unintended off-target effects.

Therefore, the safety aspect of CRISPR/Cas9 on humans is still an open issue and there are concerns about the practical applications of this technique. It is thus highly desirable to design data-driven models which;

(1) Influence on-target efficiency

(2) Improve off-target specificity

(3) Simultaneously maximize on-target activity and minimize off-target effects

Over the last few years, the main objective of data-driven models has been to predict of on-target and off-target activity, and in all cases, investigating an informative encoding technique and designing effective models for learning features is a major challenge. These computational algorithms can improve our understanding of the mechanisms of CRISPR/Cas9 system and provoke more enthusiasm on the clinical application of this technique. A lot of tools and bioinformatic algorithms have been developed to help in the design of the sgRNA and to check possible off-target effects. A summary about some of these tools is presented in Table 1 [16].

Off-target effects are undesired and should be minimized.

Tool Name	Search by Gene Name	Alternate PAM Sequence	Predicts Off-targets	Ranks Output	All in One Tool	Link
Cas9-Design	×	×	✓	×	✓	http://cas9.cbi.pku.edu.cn/
CCTop	×	✓	✓	×	✓	http://crispr.cos.uni-heidelberg.de/
CGAT	✓	×	✓	✓	✓	http://cbc.gdcb.iastate.edu/cgat/
CHOPCHOP	✓	✓	✓	✓	✓	https://chopchop.rc.fas.harvard.edu/
COSMID	×	✓	✓	✓	✓	https://crispr.bme.gatech.edu/
CRISPR design	×	✓	✓	✓	✓	http://crispr.mit.edu/
CRISPRdirect	✓	✓	✓	✓	✓	http://crispr.dbcls.jp/
Crispr Finder	×	×	✓	×	×	http://crispr.u-psud.fr/Server/
CRISPR Multitargeter	×	✓	✓	×	×	http://www.multicrispr.net/
Crispr-P	✓	✓	✓	✓	✓	http://cbi.hzau.edu.cn/crispr/
CRISPRseek	×	✓	✓	×	✓	http://www.bioconductor.org/packages/release/bioc/html/CRISPRseek.html
CROP-IT	✓	✓	✓	✓	×	http://cheetah.bioch.virginia.edu/AdliLab/CROP-IT/homepage.html
E-crisp	×	✓	✓	×	×	http://www.e-crisp.org/E-CRISP/
flyCRISPR	×	✓	×	×	✓	http://flycrispr.molbio.wisc.edu/
GT-SCAN	×	✓	✓	×	✓	http://flycrispr.molbio.wisc.edu/
sgRNAcas9	×	✓	✓	×	×	http://www.biootools.com/col.jsp?id=140
SSFinder	×	×	×	×	×	https://code.google.com/p/ssfinder/

Table 1: Tools for sgRNA design and off-target effect prediction [16].

3. Application of AI algorithms on target prediction

In this section, we discuss the most important applications of Deep Learning and Machine learning Methods for Genome Editing. On-target and off-target predictions in the context of deep learning and machine learning can generally be divided into two main challengeable parts. First, those that obtain an encoding technique to convert sgRNA-DNA sequence pairs into an appropriate input for machine and deep learning models without loss of information. Second, those that design effective machine and deep learning models to learn features from vector or matrix representations and provides high accuracy predictions.

3.1 Sequence Encoding

Recently, several studies have been conducted to use the sequence pair information effectively. In fact, one of the main challenges in off-target prediction is converting sgRNA-DNA sequence pairs into an appropriate input for deep learning models without loss of information. In this case, several attractive encodings techniques have been explored based on one-hot encoding and k-mer embedding.

In one-hot encoding, each sgRNA-DNA sequence pair encoded in a one-hot matrix with 4 rows correspond to the size of nucleotide type, i.e. A, C, G and T, and L columns correspond to the length of the sequence. Therefore, each base in in the sgRNA and target DNA is encoded in the form of four one-hot vectors [1,0,0,0], [0,1,0,0], [0,0,1,0] and [0,0,0,1]. There are many outstanding one-hot encodings technique to encode sequence pairs. To cover this important encoding technique, we briefly summarize the more recent studies and their used encoding schemes.

Lin et al [17] introduced a novel one-hot sequence encoding technique to transfer each sgRNA-DNA sequence pair with length of 23 (3-bp PAM adjacent to the 20 bases) into a 4´23 matrix using the OR operation on two one-hot vectors of base-pairing. Lin et al. [18] proposed an encoding technique that each gRNA-target pair can be represented by 10´L binary matrix (L is the length of sequence). They encode each gRNA target sequence pair with indels and mismatches with five-bit one-hot encoding with channel-wise concatenation. Stortz et al. [19], investigated four feature encoding techniques called target-guide encoding, target-mismatch encoding, target-mismatch-type encoding and target-OR-guide encoding to implement physically informed features.

Chuai et al [20] presented an encoding method to encode a DNA region which contains both nucleotide sequence and epigenetic information. They considered DNA region as a one-row multi-channel picture includes A-channel, C-channel, G-channel, and T-channel, and also each epigenetic feature is considered as one channel. Zhang et al. [21] proposed a new encoding scheme to encode sequence pairs with bulges without information loss in the process of encoding. For different two bases in a base pair, an OR operation on the two vectors is performed to represent the base pair. For same two bases in a base pair, the encoding approach used OR operation and reverses it. Y. Zhang et al. investigated a new encoding that each sequence converted into 20´20 matrix. They used concatenation on the one-hot features for on-target sequence, one-hot features for putative off-target sequence, and the mismatch position. Since, the mismatch positions include 12 different mutation type, mutation position and type information can be represented by a 12´20 matrix.L. Xue et al. [22] used one-hot encoding on sgRNA on-target activity prediction.

Another important encoding technique is the k-mer embedding, which is inspired by the word2vec technique [23]. In this encoding approach, the input sequence is split into overlapped k-mers of length k using a sliding window with stride s, then each k-mer is mapped into a d-dimensional vector using word2vec method. Word2vec is an unsupervised learning algorithm which maps k-mers from the vocabulary to vectors of real numbers in a low-dimensional space.

Charlier et al. [24] represented a new one-hot encoding method based on One-hot encoding that maps sgRNA-DNA sequence pair into 8´23 matrix according to a bijective function. This encoding approach is based on the concatenation of the sgRNA and the DNA nucleobase sequences and preserves information throughout the encoding process (Figure 2).

Figure 2. 8x23 encoding of guide-RNA and target DNA.

3.2 Application of Machine learning

In CRISPR/Cas9, traditional machine learning algorithms has been studied extensively for on-target and off-target prediction. In this section, we provide a summary of the application of machine learning methods in genome editing and their impacts on target predictions. Before the data driven model was used as a common method in genome editing, some custom prediction methods based on scoring functions such as MIT CRISPR Design Tool2, CCTop algorithm, CRISPR Design, E-CRISP and CHOPCHOP were designed. These custom tools have their own limitations and there are concerns that they may miss extra information.

The initial attempt to use a data driven model was provided in 2014 by Wang et al. [25] that implement support-vector-machine (SVM) classifier. Then, and Doench et al. [26] proposed SVM and Logistic regression.

3.3 Application of Deep learning

Although CRISPR-Cas is a powerful genome editing technology, there are concerns about the safety aspect of its translational applications. Design of off-target prediction models to evaluate the off-target situation of various gRNAs accurately has a great impact on the selection of gRNAs with high specificity and targeting accuracy.

Lin et al. [17] proposed a deep neural network for off-target predictions. They investigated two deep convolutional neural network and deep feedforward neural network to predict off-target mutations. Engineer SpCas9 with higher specificities plays an important role to address the off-target problem. Wang et al [27], analyzed CNN and RNN framework for gRNA activity prediction for Cpf1, WT-SpCas9 and SpCas9-HF1. The RNN frame work have shown promising results compare to CNN and other algorithms. They developed an online tool based on RNN called DeepHF to improve on-target activity prediction. This tool, which combines RNN and important biological feature (biofeature), is able to predict the activities of all gRNAs with DNA sequence as input and select gRNAs that are suitable for gene knockout with eSpCas9(1.1) and SpCas9-HF1 data.

Chuai et al. [20] implemented a novel deep learning framework named DeepCRISPR to simultaneously predict sgRNA on-target knockout efficacy and the off-target cleavage. They presented a deep unsupervised representation learning approach to pre-train huge amounts of unlabeled sgRNAs. In this framework, first sgRNAs encoded with its sequence and epigenetic information and then a deep coevolutionary denoising neural network (DCDNN) is presented to learn meaningful representation of sgRNAs. Finally, the output is used as the input of a coevolutionary neural network. Thus, the model not only trained the wights of CNN but also fine-tuned the weights of DCDNN-based network with limited labeled sgRNAs. Also, they extend this model for sgRNA off-target site prediction by reusing the pre-trained DCDNN-based network.

Attention mechanism is also effective to achieve a satisfactory performance for sgRNA off-target specificity prediction and on-target efficiency prediction. Liu et al. [28] analyzed two transformer-based deep neural network model called AttnToMismatch_CNN and AttnToCrispr_CNN to consider cell-specific information of genes. These attention-based models, AttnToMismatch_CNN and AttnToCrispr_CNN have competitive performance for off-target sgRNA specificity prediction and on-target efficiency prediction, respectively. Also, they implemented seqCrispr model which compromise a LSTM component and CNN component in parallel for on-target efficiency predictions. AttnToMismatch_CNN architecture contains four components: an embedding layer (the base pairs at different positions are encoded into distinct vector representations), a transformer layer, a convolutional neural network layer and fully connected layer. AttnTo-Crispr_CNN frameworks consist of four components the same as AttnToMismatch_CNN except that AttnToCrispr_CNN has a linear regression layer for the final output.

4. Off-target prediction with novel sgRNA-DNA sequence encoding in CRISPR-Cas9 gene editing

In this section, we consider an interesting method for off-target prediction with novel sgRNA-DNA sequence encoding in CRISPR-Cas9 gene editing. Charlier et al. [24] propose the novel encoding method that has shown promising results for off-target prediction. They considered two experiments, one on each data set. The first is the CRISPOR dataset and the second is GUIDE-seq dataset. On both, many types of deep learning models were tested, such as Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) and Feedforward Neural Network (FNN) for example. Some of these tests were performed on different depth of layers, and the best of each (i.e., ones that performs best according to their F1-score) were picked and used to be compared against the best model of its type from the other dataset. For example, FNN models tested in Charlier et al. experiments were FNN 3 layers, FNN 5 layers and FNN 10 layers, meaning 3-5-10 connected dense layers respectively. Using their CRISPOR dataset, FNN 5 layers for the 4x23 encoding was the best model in terms of performance. As for the 8x23 encoding, the FNN 10 layers was the best performing one. For the first test, the model predicted 6 positive samples out of 43 positives, as for the latter the top 20 predictions were true positive samples [24]. Figure 3 could be looked at to see which deep learning models were used.

Figure 3. Roc curves (AUC) from Charlier et al 2021 experiment using the 4x23 encoding on the CRISPOR dataset [24].

Figure 4. Roc curves (AUC) from Charlier et al 2021 experiment using the 8x23 encoding on the CRISPOR dataset [24]

To present the performance of each model, the results in Charlier et al 2021 study were presented are the form of Receiver Operating Characteristic [29] curves. It is clear that this novel encoding method improves the accuracy of off-target prediction on every model tested, with the highest ROC (AUC) score of 99.5% (Figure 4). However, this method presented Charlier et al could be improved, for example, if it took into consideration the possibility of insertions, deletions (indels) in addition to mismatches to portray reality more reliably. It is reported by them that indels could be considered by adding two additional rows to the encoded matrix.

Figure 5. Roc curves (AUC) from Charlier et al 2021 experiment using the 8x23 encoding on the GUIDE-seq dataset [24].

Figure 6. Roc curves (AUC) from Lin et al 2020 experiment using the 7x24 encoding on the GUIDE-seq dataset [18]

A study performed by Lin et al 2020 use another method to consider indels. They have done that by encoding the sgRNA-DNA sequences in a 7 x 24 matrix. However, the deep learning model developed to score the mismatch and indels is a recurrent convolution network which is a combination of inception-based convolutional neural network and bidirectional LSTM (Lin 2020). When we compare the two highest scores from Lin et al 2020 and Charlier et al 2021 on the same dataset (Figure 5 and Figure 6), which are FNN3 for the first and CRIPSR-Net-cif for the second, we notice that the latter model outperforms the first one. This analysis still must be taken with precaution, because it’s difficult to compare two methods that share a few parameters.

5. Conclusion

Deep Learning is a powerful method for the target prediction and learning complex patterns at multiple layers that has been used in many research works. Compared to traditional machine learning methods, deep Learning have been applied to computational biology to manage growing amounts of generated data. In the experiment performed by Charlier et al 2021 [24], the novel encoding method has shown promising results for off-target prediction. They propose a novel encoding method for off-target prediction in CRISPR/Cas 9 technique. They consider CRISPOR and GUIDE-seq dataset for deep learning models such as Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) and Feedforward Neural Network (FNN) for example.

Overall, the prediction for off-target activities is still a successful step forward in research and would enable better outcomes in the clinical usage. It would be interesting explore a comparison of the model proposed by Lin et al. [18] and the model proposed by Charlier et al. [24] by adding 2 extra rows, as indicated earlier, to evaluate performance of prediction including indels. Also, it is interesting to address the insertions and deletions (indels) in the 8x23 encoding and use parameter optimization method for better performance of deep learning algorithms.

References

1. Redman, M., et al., What is CRISPR/Cas9? Arch Dis Child Educ Pract Ed, 2016. 101(4): p. 213-5.

2. Maeder, M.L. and C.A. Gersbach, Genome-editing Technologies for Gene and Cell Therapy. Mol Ther, 2016. 24(3): p. 430-46.

3. Chuang, C.K. and W.M. Lin, Points of View on the Tools for Genome/Gene Editing. Int J Mol Sci, 2021. 22(18).

4. Gaj, T., C.A. Gersbach, and C.F. Barbas, 3rd, ZFN, TALEN, and CRISPR/Cas-based methods for genome engineering. Trends Biotechnol, 2013. 31(7): p. 397-405.

5. Adli, M., The CRISPR tool kit for genome editing and beyond. Nat Commun, 2018. 9(1): p. 1911.

6. Jinek, M., et al., A programmable dual-RNA-guided DNA endonuclease in adaptive bacterial immunity. Science, 2012. 337(6096): p. 816-21.

7. Cho, S.W., et al., Targeted genome engineering in human cells with the Cas9 RNA-guided endonuclease. Nat Biotechnol, 2013. 31(3): p. 230-2.

8. Cong, L., et al., Multiplex genome engineering using CRISPR/Cas systems. Science, 2013. 339(6121): p. 819-23.

9. Mali, P., et al., CAS9 transcriptional activators for target specificity screening and paired nickases for cooperative genome engineering. Nat Biotechnol, 2013. 31(9): p. 833-8.

10. Ma, H., et al., Correction of a pathogenic gene mutation in human embryos. Nature, 2017. 548(7668): p. 413-419.

11. Liu, H., et al., CRISPR-P 2.0: An Improved CRISPR-Cas9 Tool for Genome Editing in Plants. Mol Plant, 2017. 10(3): p. 530-532.

12. Zarei, A., et al., Creating cell and animal models of human disease by genome editing using CRISPR/Cas9. J Gene Med, 2019. 21(4): p. e3082.

13. Zhang, F., Y. Wen, and X. Guo, CRISPR/Cas9 for genome editing: progress, implications and challenges. Hum Mol Genet, 2014. 23(R1): p. R40-6.

14. Ma, Y., L. Zhang, and X. Huang, Genome modification by CRISPR/Cas9. Febs j, 2014. 281(23): p. 5186-93.

15. Shah, S.A., et al., Protospacer recognition motifs: mixed identities and functional diversity. RNA Biol, 2013. 10(5): p. 891-9.

16. Brazelton, V.A., Jr., et al., A quick guide to CRISPR sgRNA design tools. GM Crops Food, 2015. 6(4): p. 266-76.

17. Lin, J. and K.C. Wong, Off-target predictions in CRISPR-Cas9 gene editing using deep learning. Bioinformatics, 2018. 34(17): p. i656-i663.

18. Lin, J.C., et al., CRISPR-Net: A Recurrent Convolutional Network Quantifies CRISPR Off-Target Activities with Mismatches and Indels. Advanced Science, 2020. 7(13).

19. Florian Störtz, J.M., Peter Minary, piCRISPR: Physically Informed Features Improve Deep Learning Models for CRISPR/Cas9 Off-Target Cleavage Prediction.

20. Chuai, G., et al., DeepCRISPR: optimized CRISPR guide RNA design by deep learning. Genome Biol, 2018. 19(1): p. 80.

21. Zhang, Z.R. and Z.R. Jiang, Effective use of sequence information to predict CRISPR-Cas9 off-target. Comput Struct Biotechnol J, 2022. 20: p. 650-661.

22. Xue, L., et al., Prediction of CRISPR sgRNA Activity Using a Deep Convolutional Neural Network. J Chem Inf Model, 2019. 59(1): p. 615-624.

23. Tomas Mikolov, I.S., Kai Chen, Greg Corrado, Jeffrey Dean, Distributed Representations of Words and Phrases and their Compositionality, 2013.

24. Charlier, J., R. Nadon, and V. Makarenkov, Accurate deep learning off-target prediction with novel sgRNA-DNA sequence encoding in CRISPR-Cas9 gene editing. Bioinformatics, 2021.

25. Wang, T., et al., Genetic screens in human cells using the CRISPR-Cas9 system. Science, 2014. 343(6166): p. 80-4.

26. Doench, J.G., et al., Rational design of highly active sgRNAs for CRISPR-Cas9-mediated gene inactivation. Nat Biotechnol, 2014. 32(12): p. 1262-7.

27. Wang, D., et al., Optimized CRISPR guide RNA design for two high-fidelity Cas9 variants by deep learning. Nat Commun, 2019. 10(1): p. 4284.

28. Liu, Q., D. He, and L. Xie, Prediction of off-target specificity and cell-specific fitness of CRISPR-Cas System using attention boosted deep learning and network-based gene feature. PLoS Comput Biol, 2019. 15(10): p. e1007480.

29. Elzamly, S., et al., Epithelial-Mesenchymal Transition Markers in Breast Cancer and Pathological Responseafter Neoadjuvant Chemotherapy. Breast Cancer (Auckl), 2018. 12: p. 1178223418788074.