Methods for selecting mismatched sgRNAs and their application in CRISPR systems at specific DNA sites

Abstract

This invention relates to a method for selecting mismatched sgRNAs and their application in a CRISPR system at specific DNA sites, relating to the field of gene editing. The method includes: building a model for predicting sgRNA off-target mismatch activity; predicting and selecting the activity of mismatched sgRNAs at a given DNA site; and predicting the off-target effects of the selected mismatched sgRNAs. This invention analyzes the changes in the activity of mismatched sgRNAs under different matching conditions using sgRNA-DNA sequences, establishes a deep learning-based model for predicting sgRNA mismatch activity, and performs activity prediction and off-target effect analysis on mismatched sgRNAs binding to a given DNA site. This invention includes sgRNAs with mismatches to the target DNA site as candidates, expanding the possibilities for sgRNA design and effectively reducing the off-target effects of the CRISPR system at that site. It is applicable to all situations requiring the reduction of off-target effects using the CRISPR / Cas9 system, such as gene knockout and gene therapy, and can overcome the limitations of existing sgRNA technology methods.

Description

Technical Field

[0001] This invention relates to the field of gene editing, specifically to a method for selecting mismatched sgRNAs and its application in a CRISPR system at specific sites on DNA. Background Technology

[0002] CRISPR / Cas9 (clustered regularly interspaced short palindromic repeat / CRISPR-associated protein 9) system-mediated genome editing is currently the most widely used site-specific genome editing technology. It uses a single guide RNA (sgRNA) to recognize the target site through base complementarity, guiding the Cas9 protein to perform enzymatic cleavage, creating double-strand breaks, thereby achieving gene editing at the DNA level. Compared to other previously invented gene editing technologies, CRISPR / Cas9 technology is widely applicable, inexpensive, and easy to operate, and has been rapidly applied to various aspects of gene research and therapy. In gene therapy, CRISPR / Cas9 technology has also demonstrated significant advantages and potential. However, because the sgRNA only requires the first 20 bases from its 5' end to be complementary to the target DNA sequence, and because Cas9 allows for some mismatch in the complementary structure between the sgRNA and the target DNA when performing its endonuclease function, this leads to the fundamental off-target problem inherent in the CRISPR system. Therefore, it is particularly important to evaluate sgRNA by using a more accurate predictive model of sgRNA targeting cleavage efficiency and off-target effects.

[0003] Currently, various methods exist for predicting sgRNA efficiency, and deep learning has also been increasingly applied to sgRNA efficiency prediction in recent years. Existing deep learning-based sgRNA efficiency prediction methods utilize convolutional neural networks, taking the sgRNA sequence as input. These models can capture the global features of the sgRNA sequence effectively, achieving higher prediction accuracy than other traditional machine learning models. However, all current sgRNA design methods are limited to predicting sgRNA efficiency and off-target effects at all possible binding sites within the target region of DNA. With limited site selection, it is often impossible to find usable sgRNAs that combine high targeting efficiency with low off-target effects. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method for selecting mismatched sgRNAs and their application in a CRISPR system at specific DNA sites. This method can predict the targeting cleavage efficiency and off-target effects of these mismatched sgRNAs at the DNA site, and then compare them with the targeting cleavage efficiency and off-target effects of perfectly matched sgRNAs. This allows for the identification of mismatched sgRNAs that combine high targeting cleavage efficiency with low off-target effects.

[0005] This invention is achieved through the following technical solution:

[0006] A method for selecting mismatched sgRNAs includes the following steps:

[0007] P1. A model for predicting the relative activity of sgRNA mismatch was built using the publicly available sgRNA-DNA relative activity dataset in HEK293T.

[0008] P2 generates a set of mismatched sgRNAs and a set of perfectly matched sgRNAs that bind to a given DNA site for any sgRNA. Then, using the sgRNA mismatch activity prediction model, it obtains the relative activity prediction value of each mismatched sgRNA in the mismatched sgRNA set compared with the perfectly matched sgRNA.

[0009] P3, select candidate mismatched sgRNAs from the mismatched sgRNA set based on the relative activity prediction value, and predict the off-target effect of each candidate mismatched sgRNA.

[0010] P4 compares the targeting activity and off-target effect of perfectly matched sgRNAs and candidate mismatched sgRNAs in P3. The candidate mismatched sgRNA with high targeting activity and low off-target effect is selected as the target mismatched sgRNA.

[0011] Preferably, in P1, the steps for building the sgRNA mismatch activity prediction model are as follows:

[0012] P11, using one-hot encoding to process the sgRNA-DNA sequences in the sgRNA-DNA relative activity dataset, obtained a multidimensional matrix encoded as 2×4×22;

[0013] P12 uses a multidimensional matrix as the input layer, extracts sequence features using a convolutional neural network, and then sequentially uses a global max pooling layer, a first fully connected layer, a second fully connected layer, and an output layer to obtain an sgRNA relative activity prediction model.

[0014] Preferably, in P11, the multidimensional matrix includes sequence information of DNA target and sgRNA, PAM sequence information, and mismatch site information.

[0015] Preferably, in P12, the convolutional layer of the convolutional neural network includes 32 convolutional kernels, and the convolutional layer includes a first layer and a second layer. The convolutional kernel of the first layer is 4×4, and the convolutional kernel of the second layer is 4×1.

[0016] Preferably, in P12, the number of neurons in the first fully connected layer is 256, and the number of neurons in the second fully connected layer is 128.

[0017] Preferably, in P2, the process of generating the mismatched sgRNA set includes the following steps:

[0018] P21 generates a perfectly matching sgRNA sequence for a given DNA site.

[0019] P22. Select n positions on the perfectly matched sgRNA sequence, change the type of bases at the n positions to form m combinations, and denote the changed sequence as mismatched sgRNA, until all m base combinations different from the perfectly matched sgRNA sequence have been exhausted.

[0020] Preferably, in P3, the following steps are specifically included:

[0021] P31, select mismatched sgRNAs with relative activity ≥0.8 and include them in the candidate mismatched sgRNA set;

[0022] P32, for each candidate mismatched sgRNA, uses the Blast alignment algorithm to identify all DNA off-target sites with ≤6 differential bases on the target genome; the types of DNA off-target sites include exon / / CDS / UTR regions, promoters, introns and other regions; predicts and calculates the off-target activity of each candidate mismatched sgRNA at its respective DNA off-target sites.

[0023] P33, calculate the off-target effects of each candidate misplaced sgRNA.

[0024] Preferably, in P32, the targeting activity of each sgRNA that perfectly matches the off-target site of DNA is predicted using a publicly available sgRNA targeting activity prediction tool.

[0025] The relative activity of each mismatched sgRNA at each off-target DNA site was predicted by the sgRNA mismatch activity prediction model compared with that of a perfectly matched sgRNA at each off-target DNA site.

[0026] Multiplying the targeting activity by the relative activity yields the off-target activity of the misplaced sgRNA at the off-target site.

[0027] Preferably, in P33, the calculation process is as follows:

[0028] S331 assigns values ​​to the types of DNA off-target sites: exons / CDS / UTR regions are assigned 10, promoters are assigned 5, intron regions are assigned 3, and other regions are assigned 1.

[0029] P332, sum and calculate the off-target activity × genome annotation weight of each mismatched sgRNA at each DNA off-target site, as the off-target effect of the mismatched sgRNA, i.e., off-target effect = ∑ off-target activity × genome annotation weight.

[0030] An application of a mismatched sgRNA selected by the aforementioned mismatched sgRNA selection method in a CRISPR system at a specific DNA site, wherein the mismatched sgRNA is used to reduce the off-target effects of the CRISPR system at a specific DNA site.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The mismatched sgRNA selection method employed in this invention differs from previous methods that only considered the efficiency of sgRNAs perfectly matching the target DNA site. Instead, it generates mismatched sgRNAs as candidates and uses a convolutional neural network to predict their targeted cleavage activity against the target DNA site, while simultaneously predicting off-target effects. This significantly broadens the selection range of sgRNAs when the target editing sites of the CRISPR / Cas9 system are limited, selecting sgRNAs that possess both high targeted cleavage activity and low off-target effects. This method is applicable to all scenarios using CRISPR / Cas9 technology, such as targeted gene knockout, targeted gene modification, and gene therapy.

[0033] Furthermore, in the process of producing mismatched sgRNA candidate libraries, selecting fewer than or equal to 3 positions on the sgRNA sequence for mutation is to reduce the number of mismatched sgRNAs produced, which helps to reduce computation time and the generation of false positive results in the prediction model.

[0034] Furthermore, when assessing the off-target effects of mismatched sgRNAs, the genome alignment uses no more than 6 bases as potential off-target sites because previous studies on the off-target mechanisms of the CRISPR / Cas9 system have statistically shown that the number of mismatched bases between sgRNAs and their off-target sites is no more than 6.

[0035] The model reported in this invention has been fully validated on the test dataset and further validated by experimental data. Compared with other existing sgRNA design methods, this invention can generate sgRNAs with both high-efficiency targeted cleavage activity and low off-target effects, which can help the subsequent research and application of CRISPR / Cas9 technology in gene editing and gene therapy. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the calculation process of a mismatched sgRNA selection method according to the present invention.

[0037] Figure 2 The relative activity prediction model based on convolutional neural networks used in this invention

[0038] Figure 3 This invention demonstrates the design results of mismatched sgRNAs to reduce off-target effects at a given example DNA site.

[0039] Figure 4 The relative activity obtained from the dual-luciferase test in the examples is shown. Detailed Implementation

[0040] The present invention will be further described in detail below with reference to specific embodiments. These descriptions are for explanation purposes only and are not intended to limit the scope of the invention.

[0041] This invention uses a publicly available sgRNA-DNA relative activity dataset, which contains a total of 26,248 mismatched sgRNAs and 1,978 corresponding DNA targets and matching sgRNAs. Each mismatched sgRNA has experimentally verified activity relative to the matching sgRNA.

[0042] This invention proposes a method for selecting mismatched sgRNAs. This method, in the process of finding sgRNAs with lower off-target effects, first requires constructing and training a neural network model to predict the relative activity of mismatched sgRNAs. The trained model is then used to predict the relative activity of generated mismatched sgRNAs. Afterwards, the off-target sites of mismatched sgRNAs with higher relative activity are identified, and their off-target effects are calculated. The main process is described in [reference needed]. Figure 1 This includes the following steps:

[0043] P1. Using the publicly available sgRNA-DNA relative activity dataset from HEK293T, the sgRNA was constructed. The specific steps are as follows:

[0044] P11 uses one-hot encoding to process the sgRNA-DNA sequences in the sgRNA-DNA relative activity dataset, obtaining a 2×4×22 multidimensional matrix. For example, for the bases A and T(U) in the sequence, one-hot encoding is used to encode them into four one-dimensional vectors: [1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0], and [0, 0, 0, 1]. For the 22bp DNA target sequence and the 19nt sgRNA sequence, a 4×22 and a 4×19 matrix are obtained, respectively. To ensure consistent model input, the sgRNA is also encoded with the PAM sequence, extending it to 22nt. Combining these two matrices yields a final 2×4×22 multidimensional matrix containing the sequence information of the DNA target and sgRNA, the PAM sequence information, and the mismatch site information.

[0045] P12, Reference Figure 2 A convolutional neural network was constructed, using a multidimensional matrix as the input layer. After extracting sequence features through two convolutional layers with 32 convolutional kernels (4×4 and 4×1 respectively), a global max pooling layer, a first fully connected layer with 256 neurons, a second fully connected layer with 128 neurons, and an output layer were used to obtain the output sgRNA relative activity prediction model.

[0046] The formula used in feature extraction is: When using the ReLU activation function, the horizontal size of each convolutional kernel is set to 4, and the scan stride is set to 1, which is used to extract feature information of adjacent base combinations.

[0047] The pooling window size is 1×2, and the pooled data is batch normalized. Then, the first and second connecting layers are connected, and the activation function is calculated using the formula g(x)=1 / (1+e^(-x)). Finally, the predicted value of the mismatched sgRNA activity is obtained by the linear activation function f(x)=ax of the output layer.

[0048] P2 generates a set of mismatched sgRNAs and a set of perfectly matched sgRNAs that bind to a given DNA site for any sgRNA. Then, using an sgRNA mismatch activity prediction model, it obtains a relative activity prediction value for each mismatched sgRNA in the mismatched sgRNA set compared to the perfectly matched sgRNA.

[0049] The process of generating a mismatched sgRNA set includes the following steps:

[0050] P21 generates a perfectly matching sgRNA sequence for a given DNA site; for example, simply replacing the T base with U in the DNA site sequence yields a perfectly matching sgRNA sequence.

[0051] P22. Select n positions on a perfectly matched sgRNA sequence and change the base types at these n positions to form m combinations. The altered sequences are denoted as mismatched sgRNAs. This process continues until all m base combinations different from the perfectly matched sgRNA sequence have been exhausted. For a perfectly matched sgRNA, there are 4 possible combinations at the selected n positions. n -1 mismatched sgRNA base combination.

[0052] In other words, n positions on a perfectly matched sgRNA sequence are selected, and the base types at these n positions are changed to form m combinations. The altered sequences are denoted as mismatched sgRNAs. This process continues until all m base combinations different from the perfectly matched sgRNA sequence have been exhausted. Then, new n positions are selected, and the process of exhaustively searching for m base type combinations to form mismatched sgRNAs is repeated until all combinations at the n positions on the sgRNA sequence have been exhausted. Here, n and m are constants, n = 1, 2, and 3. For a perfectly matched sgRNA, when n = 1, there are 20 combinations; when n = 2, there are 190 combinations; when n = 3, there are 1140 combinations; for a total of 1350 combinations.

[0053] P3. Select a set of mismatched sgRNAs with targets from the mismatched sgRNA set based on their relative activity prediction values, and predict their respective off-target probabilities. The specific steps are as follows:

[0054] P31, select mismatched sgRNAs with relative activity ≥0.8 from the mismatched sgRNA set and include them in the candidate mismatched sgRNA set;

[0055] P32, for each candidate mismatched sgRNA, uses the Blast alignment algorithm to identify all DNA off-target sites with ≤6 differential bases on the target genome; the types of DNA off-target sites include exon / / CDS / UTR regions, promoters, introns and other regions; predicts and calculates the off-target activity of each candidate mismatched sgRNA at its respective DNA off-target sites.

[0056] Among them, the targeting activity of each sgRNA that perfectly matches the off-target site of DNA was predicted using a publicly available sgRNA targeting activity prediction tool.

[0057] The relative activity of each mismatched sgRNA at each off-target DNA site was predicted by the sgRNA mismatch activity prediction model compared with that of a perfectly matched sgRNA at each off-target DNA site.

[0058] Multiplying the targeting activity by the relative activity yields the off-target activity of the misplaced sgRNA at the off-target site.

[0059] P33, calculate the off-target effects of each candidate misplaced sgRNA, the steps are as follows:

[0060] S331 assigns values ​​to the types of DNA off-target sites: exons / CDS / UTR regions are assigned 10, promoters are assigned 5, intron regions are assigned 3, and other regions are assigned 1.

[0061] P332, sum and calculate the off-target activity × genome annotation weight of each mismatched sgRNA at each DNA off-target site, as the off-target effect of the mismatched sgRNA, i.e., off-target effect = ∑ off-target activity × genome annotation weight.

[0062] P4, Reference Figure 3 By comparing the targeting activity and off-target probability of perfectly matched sgRNAs and mismatched sgRNAs in P3, candidate mismatched sgRNAs with high targeting activity and low off-target effect are selected as target mismatched sgRNAs.

[0063] Example

[0064] For the firefly fluorescence gene (Luciferase) on pGL3-promoter, the method provided in this invention was used to identify the corresponding mismatched sgRNAs at each of the 37 DNA sites. The DNA sites, mismatched sgRNA sequences, and predicted relative activities are shown in Table 1. In the HEK293T transfection experiment, more than 60% of the mismatched sgRNAs showed a targeting cleavage activity ≥0.8 relative to the perfectly matched sgRNAs. The relative activities obtained from the dual-luciferase assay are shown in Table 1. Figure 4 .

[0065] Table 1. DNA sites and mismatched sgRNA sequences in the firefly fluorescent gene on pGL3-promoter, and predicted relative activities.

[0066]

[0067]

[0068] The present invention also discloses the application of a mismatched sgRNA in a CRISPR system at a specific site on DNA, that is, the mismatched sgRNA is used to reduce the off-target effects of the CRISPR system at a specific site on DNA.

[0069] In summary, this invention establishes an sgRNA method to reduce the off-target effects of the CRISPR system at specific DNA sites using mismatched sgRNAs. The main focus is on the question of how much targeting activity a mismatched sgRNA with a given DNA site can have. A convolutional neural network-based model is introduced to predict the activity of mismatched sgRNAs, which has been fully validated on a test dataset and further validated by experimental data.

[0070] Compared to other existing sgRNA design methods that only consider the activity of sgRNAs that are perfectly matched with the target DNA site, this invention includes sgRNAs with mismatches with the target DNA site as candidates, expanding the possibilities of sgRNA design. This can effectively reduce the off-target effects of the CRISPR system at this site and can produce sgRNAs with both high-efficiency targeted cleavage activity and low off-target effects, thus providing assistance for subsequent research and application of CRISPR / Cas9 technology in gene editing and gene therapy.

[0071] This invention analyzes the changes in the activity of mismatched sgRNAs under different matching conditions using sgRNA-DNA sequence pairs, establishes a deep learning-based model for predicting sgRNA mismatch activity, and performs activity prediction and off-target effect analysis on mismatched sgRNAs binding to a given DNA site. This invention includes sgRNAs with mismatches to target DNA sites as candidates, expanding the possibilities for sgRNA design and effectively reducing off-target effects of the CRISPR system at that site. It is applicable to all situations requiring the reduction of off-target effects using the CRISPR / Cas9 system, such as gene knockout and gene therapy, and can overcome the limitations of existing sgRNA technologies.

[0072] The above descriptions are merely preferred embodiments of this application, and the present invention is not limited to the above embodiments. It can be understood that other improvements and variations directly derived or conceived by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included within the protection scope of the present invention.

Claims

1. A method for selecting mismatched sgRNAs, characterized in that, Includes the following steps: P1. A predictive model for sgRNA mismatch relative activity was built using the publicly available sgRNA-DNA relative activity dataset from HEK293T. P2 generates a set of mismatched sgRNAs and a set of perfectly matched sgRNAs that bind to the DNA site targeted by any sgRNA. Then, using the sgRNA mismatch activity prediction model, it obtains the relative activity prediction value of each mismatched sgRNA in the mismatched sgRNA set compared with the perfectly matched sgRNA. The process of generating a mismatched sgRNA set includes the following steps: P21 generates a perfectly matching sgRNA sequence for a given DNA site. P22, select n positions on the perfectly matched sgRNA sequence, change the type of bases at the n positions to form m combinations, and denote the changed sequence as mismatched sgRNA, until all m base combinations that are different from the perfectly matched sgRNA sequence have been exhausted; P3, select candidate mismatched sgRNAs from the mismatched sgRNA set based on the relative activity prediction value, and predict the off-target effect of each candidate mismatched sgRNA. P4 compares the targeting activity and off-target effect of perfectly matched sgRNAs and candidate mismatched sgRNAs in P3. The candidate mismatched sgRNA with high targeting activity and low off-target effect is selected as the target mismatched sgRNA.

2. The method for selecting mismatched sgRNAs according to claim 1, characterized in that, In P1, the steps for building the sgRNA mismatch activity prediction model are as follows: P11, using one-hot encoding to process the sgRNA-DNA sequences in the sgRNA-DNA relative activity dataset, obtained a multidimensional matrix encoded as 2×4×22; P12 uses a multidimensional matrix as the input layer, extracts sequence features using a convolutional neural network, and then sequentially uses a global max pooling layer, a first fully connected layer, a second fully connected layer, and an output layer to obtain an sgRNA relative activity prediction model.

3. The method for selecting mismatched sgRNAs according to claim 2, characterized in that, In P11, the multidimensional matrix includes sequence information of DNA targets and sgRNA, PAM sequence information, and mismatch site information.

4. The method for selecting mismatched sgRNAs according to claim 2, characterized in that, In P12, the convolutional layers of the convolutional neural network consist of 32 convolutional kernels. The convolutional layers include a first layer and a second layer. The first layer has 4×4 convolutional kernels, and the second layer has 4×1 convolutional kernels.

5. The method for selecting mismatched sgRNAs according to claim 4, characterized in that, In P12, the number of neurons in the first fully connected layer is 256, and the number of neurons in the second fully connected layer is 128.

6. The method for selecting mismatched sgRNAs according to claim 1, characterized in that, In P3, the specific steps include: P31, select mismatched sgRNAs with relative activity ≥0.8 and include them in the candidate mismatched sgRNA set; P32, for each candidate mismatched sgRNA, uses the Blast alignment algorithm to identify all DNA off-target sites with ≤6 differential bases on the target genome; the types of DNA off-target sites include exon / / CDS / UTR regions, promoters, introns and other regions; predicts and calculates the off-target activity of each candidate mismatched sgRNA at its respective DNA off-target sites. P33, calculate the off-target effects of each candidate misplaced sgRNA.

7. The method for selecting mismatched sgRNAs according to claim 6, characterized in that, In P32, the targeting activity of each sgRNA that perfectly matches the off-target site of DNA was predicted using a publicly available sgRNA targeting activity prediction tool. The relative activity of each mismatched sgRNA at each off-target DNA site was predicted by the sgRNA mismatch activity prediction model compared with that of a perfectly matched sgRNA at each off-target DNA site. Multiplying the targeting activity by the relative activity yields the off-target activity of the misplaced sgRNA at the off-target site.

8. The method for selecting mismatched sgRNAs according to claim 7, characterized in that, The calculation process is as follows on page 33: S331 assigns values ​​to the types of DNA off-target sites: exons / CDS / UTR regions are assigned 10, promoters are assigned 5, intron regions are assigned 3, and other regions are assigned 1. P332, summate the off-target activity of each mismatched sgRNA at each DNA off-target site × genome annotation weight, and use this as the off-target effect of the mismatched sgRNA, i.e. .

9. The application of a mismatched sgRNA selected by the mismatched sgRNA selection method according to any one of claims 1 to 8 in a CRISPR system at a specific site on DNA, characterized in that, This mismatched sgRNA is used to reduce the off-target effects of the CRISPR system at specific sites on DNA.

Images