sgRNA cleavage efficiency prediction and sgRNA design method
Patent Information
- Application Number
- CN202611236712.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-14
- Publication Date
- 2026-10-30
AI Technical Summary
[0004]预测精度有限：现有工具的预测准确性、精度和召回率等指标仍有较大提升空间，无法完全满足高精度基因编辑的需求
[0037]预测性能卓越：本发明采用基于ResNet50的深度神经网络架构，能够从sgRNA与targetDNA的配对序列中自动学习复杂的序列特征和相互作用模式。与现有工具相比，本方法在准确率(Accuracy)、精度(Precision)、召回率(Recall)、AUROC和AUPRC等多个关键评估指标上均实现了全面领先，各项指标逼近理论极限值1，显著提升了预测的可靠性。
Smart Images
Figure CN123017201A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of gene editing technology, and in particular to a method for predicting sgRNA (single guide RNA) cleavage efficiency and designing sgRNA. Background Technology
[0002] CRISPR-Cas9 gene editing technology is widely used in biomedical research, gene therapy, and synthetic biology due to its high efficiency, convenience, and low cost. The core of this technology is the use of sgRNA to guide the Cas9 protein to a specific location in the genome for DNA cleavage. Cas9 is an RNA-guided DNA endonuclease that performs site-specific cleavage of target DNA under the guidance of sgRNA. However, the cleavage efficiency of sgRNAs varies greatly, making the design of an efficient sgRNA a crucial prerequisite for successful gene editing.
[0003] Currently, sgRNA design primarily relies on empirical rules or existing computational prediction tools. These tools are mostly trained on large-scale screening data generated in mammalian cells. However, these existing tools have significant limitations:
[0004] Limited prediction accuracy: Existing tools still have considerable room for improvement in terms of prediction accuracy, precision, and recall, and cannot fully meet the needs of high-precision gene editing.
[0005] Poor cross-species applicability: Since the training data mainly comes from mammals, the predictive effect of these tools drops sharply when applied to microorganisms (such as Escherichia coli and yeast) whose genome structure and subcellular environment differ significantly from those of mammals. This makes it difficult to effectively guide sgRNA design in microbial gene editing and synthetic biology research.
[0006] The design threshold is high: researchers need to have certain bioinformatics knowledge to effectively use these tools, and it is difficult to achieve automated, high-throughput sgRNA screening and optimization of specific gene sequences. Summary of the Invention
[0007] To address the problems existing in the prior art, this invention provides a method for predicting sgRNA cleavage efficiency and designing sgRNA.
[0008] This invention provides a method for predicting sgRNA cleavage efficiency, comprising:
[0009] The sgRNA to be predicted is paired and aligned with the target DNA of the sgRNA to be predicted to form a paired sequence;
[0010] One-hot encoding is performed based on the base pairing combination of sgRNA and target DNA at each position of the paired sequence to obtain a two-dimensional matrix;
[0011] The two-dimensional matrix is input into a deep convolutional neural network to obtain the cutting efficiency score of the sgRNA to be predicted output by the deep convolutional neural network.
[0012] The deep convolutional neural network is trained using a training set, which includes the pairing sequence of the sgRNA sample and the target DNA of the sgRNA sample, as well as a tag indicating whether it has been cut.
[0013] According to the sgRNA cleavage efficiency prediction method provided by the present invention, the pairing sequence between the sgRNA sample and the target DNA of the sgRNA sample includes pairing sequences collected from open source literature and pairing sequences determined from microorganisms.
[0014] According to the present invention, the sgRNA cleavage efficiency prediction method includes Escherichia coli and lipophilic yeast.
[0015] According to the sgRNA cleavage efficiency prediction method provided by the present invention, the tag acquisition step includes:
[0016] The pairing sequences of the sgRNA sample and the target DNA of the sgRNA sample were detected by high-throughput sequencing to obtain the cleavage count;
[0017] If the cut count is greater than or equal to a preset threshold, the paired sequence is marked as a cut label and used as a positive sample;
[0018] If the cut count is less than the preset threshold, the paired sequence is marked as a non-cut label and used as a negative sample.
[0019] According to the sgRNA cleavage efficiency prediction method provided by the present invention, the acquisition of the tag further includes the following steps:
[0020] Traverse the gene sequence containing the target DNA of the sgRNA sample, and pair and align sequence fragments of a preset length that are not marked as cleavage tags with the sgRNA sample to form paired sequences, which are then marked as non-cleavage tags and used as negative samples.
[0021] According to the sgRNA cleavage efficiency prediction method provided by the present invention, a two-dimensional matrix is obtained by one-hot encoding based on the base pairing combination of sgRNA and target DNA at each position of the paired sequence, including:
[0022] Identify all possible base pairings between the bases in the sgRNA and the bases in the target DNA;
[0023] An encoding vector is constructed based on each position of the pairing sequence with the encoding length specified in the encoding vector. The element corresponding to the pairing combination at each position in the encoding vector is 1, and the other elements are 0. The elements in the encoding vector correspond one-to-one with all pairing combinations.
[0024] A two-dimensional matrix is constructed based on the encoding vectors of all positions in the paired sequence.
[0025] According to the sgRNA cleavage efficiency prediction method provided by the present invention, the deep convolutional neural network adopts the ResNet50 architecture.
[0026] According to the sgRNA cleavage efficiency prediction method provided by the present invention, the training steps of the deep convolutional neural network include:
[0027] The sgRNA sample cutting efficiency score and the corresponding label are input into the loss function to calculate the loss;
[0028] The weight parameters of the deep convolutional neural network are updated using the backpropagation algorithm and optimizer based on the loss.
[0029] According to the sgRNA cleavage efficiency prediction method provided by the present invention, after the deep convolutional neural network is trained, it further includes:
[0030] The performance of the deep convolutional neural network was evaluated using a separate test set.
[0031] This invention also provides an sgRNA design method, comprising:
[0032] The target sites on the gene sequence to be edited are retrieved using sequence alignment algorithms;
[0033] Multiple sgRNAs are generated for each target site, and each sgRNA is paired and aligned with the target DNA at the target site to form a paired sequence;
[0034] Based on any of the above-described sgRNA cleavage efficiency prediction methods, the cleavage efficiency score of each sgRNA is obtained according to the pairing sequence corresponding to each sgRNA.
[0035] The sgRNAs were sorted according to their cleavage efficiency scores, and the optimal sgRNA was selected based on the sorting results.
[0036] The sgRNA cleavage efficiency prediction and sgRNA design method provided by this invention have the following beneficial effects:
[0037] Superior Predictive Performance: This invention employs a ResNet50-based deep neural network architecture, capable of automatically learning complex sequence features and interaction patterns from the paired sequences of sgRNA and target DNA. Compared to existing tools, this method achieves comprehensive superiority across multiple key evaluation metrics, including accuracy, precision, recall, AUROC, and AUPRC, with each metric approaching the theoretical limit of 1, significantly improving the reliability of predictions.
[0038] Microbial Adaptability: This invention is the first to introduce sgRNA cleavage datasets obtained through experimental determination and sequencing in microorganisms (Escherichia coli and lipophilic yeast) for model training and optimization. This innovation solves the problem of existing tools being "unsuitable" in the field of microorganisms, enabling them to have excellent and reliable predictive capabilities in microbial gene editing, filling a technological gap in this field.
[0039] Achieving automated and intelligent design: This invention can be integrated into automated processes, enabling full-sequence scanning of target gene sequences, automatic retrieval of all potential targets, and iterative evaluation of all possible sgRNA designs. This eliminates the need for researchers to have any theoretical knowledge, significantly improving the efficiency and comprehensiveness of sgRNA design and accelerating research and industrial processes.
[0040] With broad application prospects, this method is not only applicable to basic research, but also provides a powerful core tool for industrial directions such as CRISPR drug development, gene editing therapy and industrial microbial modification, and has extremely high scientific research and commercial value. Attached Figure Description
[0041] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0042] Figure 1 This is a flowchart illustrating the sgRNA cleavage efficiency prediction method provided by the present invention;
[0043] Figure 2 This is a schematic diagram of the neural network model architecture for predicting sgRNA cleavage efficiency based on ResNet50 in the sgRNA cleavage efficiency prediction method provided by this invention;
[0044] Figure 3This is a schematic diagram comparing the performance of the model (referred to as MOP) in the sgRNA cleavage efficiency prediction method provided by this invention with existing mainstream tools.
[0045] Figure 4 This is a schematic diagram of the sgRNA design process provided by the present invention. Detailed Implementation
[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0047] The following is combined with Figure 1 A method for predicting sgRNA cleavage efficiency according to the present invention includes:
[0048] Step 101: Align the sgRNA to be predicted with the target DNA of the sgRNA to be predicted using standard pairing to form a paired sequence (e.g., 23 nt in length).
[0049] Step 102: Perform one-hot encoding based on the base pairing combination of sgRNA and target DNA at each position (e.g., 23 positions in total) of the paired sequence to obtain a two-dimensional matrix;
[0050] Step 103: Input the two-dimensional matrix into a deep convolutional neural network to obtain the cutting efficiency score of the sgRNA to be predicted output by the deep convolutional neural network.
[0051] The deep convolutional neural network is trained using a training set, which includes the pairing sequence of the sgRNA sample and the target DNA of the sgRNA sample, as well as a tag indicating whether it has been cut.
[0052] sgRNA is a single-stranded guide RNA and a key component of the CRISPR-Cas9 gene editing system, responsible for recognizing and binding to the target DNA sequence.
[0053] targetDNA refers to the specific DNA sequence that sgRNA recognizes and binds to on the genome.
[0054] One-hot encoding is a method of encoding categorical variables into binary vectors that are easy for machine learning algorithms to use.
[0055] This embodiment obtains matching features between the sgRNA to be predicted and its target DNA by performing one-hot encoding. Based on the matching features, a deep convolutional neural network is used to predict the sgRNA cleavage efficiency, thereby improving the accuracy and efficiency of the prediction.
[0056] Based on the above embodiments, in this embodiment, the pairing sequence of the sgRNA sample and the target DNA of the sgRNA sample includes pairing sequences collected from open-source literature and pairing sequences determined from microorganisms.
[0057] Data acquisition in this embodiment includes:
[0058] Public data collection: Existing sgRNA-target DNA pairing sequences and their cleavage counts detected by high-throughput sequencing were collected from open-source literature.
[0059] Own Data Generation: To fill the gaps in microbial applications, we conducted our own experiments, determining and sequencing data on sgRNA-target DNA pairing sequences and their cleavage capabilities in E. coli and lipophilic yeast.
[0060] By integrating all data (public and proprietary data), sgRNA cleavage efficiency prediction methods can be applied across species (especially adapted microorganisms).
[0061] Based on the above embodiments, the tag acquisition step in this embodiment includes:
[0062] The pairing sequences of the sgRNA sample and the target DNA of the sgRNA sample were detected by high-throughput sequencing to obtain the cleavage count;
[0063] If the cut count is greater than or equal to a preset threshold (e.g., 100), the paired sequence is marked as a cut label and used as a positive sample;
[0064] If the cut count is less than the preset threshold, the paired sequence is marked as a non-cut label and used as a negative sample.
[0065] Based on the above embodiments, the step of obtaining the tag in this embodiment further includes:
[0066] Traverse the gene sequence containing the target DNA of the sgRNA sample, and pair and align sequence fragments of a preset length that are not marked as cleavage tags with the sgRNA sample to form paired sequences, which are then marked as non-cleavage tags and used as negative samples to expand the negative sample set.
[0067] The preset length can be 23 bases (nt).
[0068] Based on the above embodiments, this embodiment performs one-hot encoding according to the base pairing combination of sgRNA and target DNA at each position of the paired sequence, resulting in a two-dimensional matrix, including:
[0069] Identify all possible base pairings between the bases in the sgRNA and the bases in the target DNA;
[0070] An encoding vector is constructed based on each position of the pairing sequence with the encoding length specified in the encoding vector. The element corresponding to the pairing combination at each position in the encoding vector is 1, and the other elements are 0. The elements in the encoding vector correspond one-to-one with all pairing combinations.
[0071] A two-dimensional matrix is constructed based on the encoding vectors of all positions in the paired sequence.
[0072] For example, there are 4x4=16 possible pairing combinations between the bases (A, U, C, G) of sgRNA and the bases (A, T, C, G) of targetDNA, and each pairing is represented by a 16-dimensional one-hot encoding.
[0073] After encoding the entire 23nt pairing sequence, a 16x23 two-dimensional matrix with 16 rows and 23 columns is obtained, which is then used as the input to a deep convolutional neural network.
[0074] Based on the above embodiments, the deep convolutional neural network in this embodiment adopts the ResNet50 architecture.
[0075] ResNet50 is a deep residual network with 50 layers. It is a classic convolutional neural network architecture in the field of computer vision. In this embodiment, it is used to extract features from the encoding matrix of sgRNA-targetDNA pairing sequences.
[0076] Model Construction: A deep convolutional neural network based on the ResNet50 architecture is built. This network receives a 16x23 input matrix, such as... Figure 2 As shown.
[0077] Model Output: After feature extraction and computation through multiple layers of convolutional, pooling, and fully connected layers, the network finally outputs a probability distribution vector containing two elements [cut score, non-cut score] through a Softmax layer. Both scores are between 0 and 1, and their sum is 1. The closer the cut score is to 1, the greater the probability that the predicted sgRNA can effectively cut the target DNA.
[0078] Based on the above embodiments, the training steps of the deep convolutional neural network in this embodiment include:
[0079] The sgRNA sample cutting efficiency score and the corresponding label are input into the loss function to calculate the loss;
[0080] The weight parameters of the deep convolutional neural network are updated using the backpropagation algorithm and optimizer based on the loss.
[0081] Model optimization: The predicted probability distribution and the true label distribution ("cut" to [1, 0], "non-cut" to [0, 1]) are input into a loss function (such as the cross-entropy loss function) to calculate the loss. The network weights are updated using a backpropagation algorithm and an optimizer (such as Adam).
[0082] Iterative training: Repeat the above process for multiple rounds of training until the model's performance on the validation set converges.
[0083] Based on the above embodiments, this embodiment further includes the following after the deep convolutional neural network is trained:
[0084] The performance of the deep convolutional neural network was evaluated using a separate test set.
[0085] Model Evaluation: After training, the model performance is evaluated using an independent test set. The model's predictions (taking the category corresponding to the larger value in the probability distribution as the predicted category) are compared with the true labels, and metrics such as accuracy, precision, recall, AUROC (Area Under the Receiver Operating Characteristic Curve), and AUPRC (Area Under the Precision-Recall Curve) are calculated.
[0086] AUROC is a comprehensive metric used to evaluate the performance of binary classification models. AUPRC is an important metric for evaluating model performance, especially when there is an imbalance between positive and negative samples.
[0087] The specific steps for model training and performance evaluation include:
[0088] Data preparation: Mammalian datasets from publicly available literature, along with experimentally determined datasets from *E. coli* and *Saccharomyces lipolyticus*, totaling approximately 150,000 sgRNA-target DNA pairings. Positive and negative samples were defined based on a cut count >100.
[0089] Data encoding: Encode all paired sequences into a 16*23 input matrix.
[0090] Model training: A ResNet50-based network architecture was used, and the dataset was divided into training, validation, and test sets in a 6:2:2 ratio. The batch size was set to 64, the learning rate to 0.001, and the training run was performed for 100 epochs.
[0091] Performance Results: On the independent test set, the evaluation results of this model are as follows: Accuracy 0.9819, Precision 0.9769, Recall 0.9768, AUROC 0.9984, AUPRC 0.9985. Figure 3 As shown in (a), (b) and (c) in the figure.
[0092] Compared with current mainstream sgRNA design tools, this model shows significant advantages in all metrics, and its predictive performance approaches the theoretical limit.
[0093] like Figure 4 As shown, the present invention also provides an sgRNA design method, comprising:
[0094] The target sites on the gene sequence to be edited are retrieved using sequence alignment algorithms;
[0095] Multiple sgRNAs are generated for each target site, and each sgRNA is paired and aligned with the target DNA at the target site to form a paired sequence;
[0096] Based on the sgRNA cleavage efficiency prediction method described in any of the above embodiments, the cleavage efficiency score of each sgRNA is obtained according to the pairing sequence corresponding to each sgRNA.
[0097] The sgRNAs were sorted according to their cleavage efficiency scores, and the optimal sgRNA was selected based on the sorting results.
[0098] The steps involved in the intelligent design and application of sgRNA include:
[0099] Target retrieval: Input the target gene sequence to be edited, and retrieve all possible target sites on the gene (usually the 20nt region upstream of the NGG PAM sequence) through sequence alignment algorithm.
[0100] Iterative generation: For each target site, all possible sgRNA sequences are generated iteratively and paired with the target DNA.
[0101] Efficiency evaluation: After encoding all sgRNA-targetDNA pairing sequences into a two-dimensional matrix, the matrix is input into a trained model to predict its cleavage probability (i.e., cleavage score).
[0102] Intelligent sorting and selection: All sgRNA designs are sorted according to the predicted cleavage probability, and the top-ranked sgRNAs are automatically recommended for subsequent gene editing experiments.
[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting sgRNA cleavage efficiency, characterized in that, include: The sgRNA to be predicted is paired and aligned with the target DNA of the sgRNA to be predicted to form a paired sequence; One-hot encoding is performed based on the base pairing combination of sgRNA and target DNA at each position of the paired sequence to obtain a two-dimensional matrix; The two-dimensional matrix is input into a deep convolutional neural network to obtain the cutting efficiency score of the sgRNA to be predicted output by the deep convolutional neural network. The deep convolutional neural network is trained using a training set, which includes the pairing sequence of the sgRNA sample and the target DNA of the sgRNA sample, as well as a tag indicating whether it has been cut.
2. The sgRNA cleavage efficiency prediction method according to claim 1, characterized in that, The pairing sequence between the sgRNA sample and the target DNA of the sgRNA sample includes pairing sequences collected from open-source literature and pairing sequences determined from microorganisms.
3. The method for predicting sgRNA cleavage efficiency according to claim 2, characterized in that, The microorganisms include Escherichia coli and lipophilic yeast.
4. The sgRNA cleavage efficiency prediction method according to claim 1, characterized in that, The steps for obtaining the tag include: The pairing sequences of the sgRNA sample and the target DNA of the sgRNA sample were detected by high-throughput sequencing to obtain the cleavage count; If the cut count is greater than or equal to a preset threshold, the paired sequence is marked as a cut label and used as a positive sample; If the cut count is less than the preset threshold, the paired sequence is marked as a non-cut label and used as a negative sample.
5. The sgRNA cleavage efficiency prediction method according to claim 4, characterized in that, The steps for obtaining the tag also include: Traverse the gene sequence containing the target DNA of the sgRNA sample, and pair and align sequence fragments of a preset length that are not marked as cleavage tags with the sgRNA sample to form paired sequences, which are then marked as non-cleavage tags and used as negative samples.
6. The method for predicting sgRNA cleavage efficiency according to claim 1, characterized in that, One-hot encoding is performed based on the base pairing combinations of sgRNA and target DNA at each position of the paired sequence to obtain a two-dimensional matrix, including: Identify all possible base pairings between the bases in the sgRNA and the bases in the target DNA; An encoding vector is constructed based on each position of the pairing sequence with the encoding length specified in the encoding vector. The element corresponding to the pairing combination at each position in the encoding vector is 1, and the other elements are 0. The elements in the encoding vector correspond one-to-one with all pairing combinations. A two-dimensional matrix is constructed based on the encoding vectors of all positions in the pairing sequence.
7. The method for predicting sgRNA cleavage efficiency according to claim 1, characterized in that, The deep convolutional neural network uses the ResNet50 architecture.
8. The method for predicting sgRNA cleavage efficiency according to claim 1, characterized in that, The training steps of the deep convolutional neural network include: The sgRNA sample cutting efficiency score is input into the corresponding label loss function to calculate the loss; The weight parameters of the deep convolutional neural network are updated using the backpropagation algorithm and optimizer based on the loss.
9. The method for predicting sgRNA cleavage efficiency according to claim 1, characterized in that, After training the deep convolutional neural network, the following is also included: The performance of the deep convolutional neural network was evaluated using a separate test set.
10. A method for designing sgRNA, characterized in that, include: The target sites on the gene sequence to be edited are retrieved using sequence alignment algorithms; Multiple sgRNAs are generated for each target site, and each sgRNA is paired and aligned with the target DNA at the target site to form a paired sequence; Based on the sgRNA cleavage efficiency prediction method according to any one of claims 1-9, the cleavage efficiency score of each sgRNA is obtained according to the pairing sequence corresponding to each sgRNA. The sgRNAs were sorted according to their cleavage efficiency scores, and the optimal sgRNA was selected based on the sorting results.