Method for predicting target activity of sgRNA

Abstract

The invention discloses a target activity prediction method of sgRNA, which comprises the following steps: step 1, preparing a data set, and obtaining an sgRNA activity sequence data set; 2, performing sequence feature extraction on the sgRNA active sequence data set to obtain multiple pieces of feature information; 3, fusing the multiple pieces of feature information to obtain a feature set; 4, constructing an unbalanced data set processing algorithm based on the graph weighted adversarial network; 5, introducing ensemble learning, and constructing a deep learning prediction model based on a voting algorithm; and 6, carrying out conservative motif analysis on the sgRNA high-activity sequence. According to the method, the prediction effect and robustness of the model can be ensured, different characteristics of data can be captured through diversified base learners, the over-fitting risk is reduced, sequence characteristics can be efficiently extracted, and the prediction precision and efficiency are remarkably improved; in addition, the biological significance of the model is explored, and an interpretable analysis attempt is carried out on the model.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method for predicting the target activity of sgRNA. Background Art

[0002] Single-guide RNA (sgRNA) is a guide RNA composed of two RNAs (tracRNA and crRNA) fused together. These RNAs can bind to the Cas9 protein and guide the Cas9 enzyme to the target region of genomic DNA to cut the DNA.

[0003] sgRNA is a crucial component of the CRISPR / Cas9 gene knockout system and plays a crucial role in gene editing and disease treatment. Highly active sgRNAs exhibit higher mutation efficiency against their targets, improving the efficiency of subsequent screening and identification protocols.

[0004] Therefore, how to determine the target activity of sgRNA before gene editing is an urgent problem to be solved. Summary of the Invention

[0005] In view of this, the present invention provides a method for predicting the target activity of sgRNA.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0007] A method for predicting target activity of sgRNA comprises the following steps:

[0008] Step 1: Dataset preparation, obtain sgRNA activity sequence dataset;

[0009] Step 2: Extract sequence features from the sgRNA activity sequence dataset to obtain multiple feature information;

[0010] Step 3: Fusing the multiple feature information to obtain a feature set;

[0011] Step 4: Construct an imbalanced dataset processing algorithm based on graph weighted adversarial network;

[0012] Step 5: Introduce ensemble learning and build a deep learning prediction model based on the voting algorithm;

[0013] Step 6: Conserved motif analysis of sgRNA high activity sequences.

[0014] Preferably, in step 1, the sgRNA activity sequence dataset includes a positive example dataset and a negative example dataset, the positive example dataset is a high-activity sgRNA sequence, and the negative example dataset is a low-activity sgRNA sequence.

[0015] Preferably, in step 1, the sgRNA activity sequence data has a total of 8 sample set pairs, including: G17, Gr, Gnr and Gm.

[0016] Preferably, in step 2, the feature extraction algorithm includes a nucleotide composition method, a sequence autocorrelation method, a nucleotide composition method, and a sequence structural feature method.

[0017] Preferably, in step 2, the nucleotide composition method includes a k-mer extraction algorithm and a subsequence extraction algorithm; the self-organizing correlation feature extraction algorithm includes a feature extraction algorithm based on automatic covariance DAC, an extraction algorithm based on cross covariance DCC, an extraction algorithm based on automatic cross covariance DACC, a Geary autocorrelation GAC algorithm, and a normalized Moreau-Broto autocorrelation NMBAC algorithm; the pseudo nucleic acid composition feature extraction algorithm includes an algorithm based on parallel correlation pseudo dinucleotide composition and a method based on continuous correlation pseudo dinucleotide composition; the structural feature extraction algorithm includes a local structure sequence triplet feature Triplet extraction algorithm.

[0018] Preferably, in step 4, first, a graph convolution layer is used to aggregate the features of each node and its neighbors, a fully connected layer is used to find a potential representation for each node, and finally a probabilistic classifier is used to select the class with the highest probability as the probability distribution of the label prediction class to predict the class label of the input node;

[0019] Then, a separate weighted network learning parameter function is used to weight the samples in each class to help the classifier fit better between classes, and the classifier parameters are optimized through adversarial training methods to prevent the classifier from being biased towards any one class.

[0020] Preferably, in step 4, the optimization formula is:

[0021]

[0022] in, and They are y i and q i The cth element of is the set of nodes c with labels. c are the weights of the weighted cross entropy.

[0023] Preferably, in step 4, weighted cross entropy is used as the classifier loss function, so that the weighted network dynamically provides sample weights for the classifier, and a two-person minimum maximum game is used to formulate the confrontation process, wherein both the classifier and the weighted network set attempt to optimize the following objectives:

[0024]

[0025] Among them, q i and w i It is the network D and W c s in sample x i Output.

[0026] Preferably, in step 5, after obtaining the reweighted adversarial graph convolutional network, it is used as the base classifier. Then, the voting algorithm of ensemble learning is used to directly count the predicted categories of each model, select the category with the most votes, and use a ten-fold cross-validation method to verify the model results.

[0027] Preferably, in step 6, the frequently occurring motifs with a length of 5-10 bp are identified by MEME, parameters are set, the distribution of the motifs in the sgRNA is checked, a sequence logo map WebLogo is drawn, and the positional distribution characteristics of the nucleotides in the conserved motifs are analyzed.

[0028] Compared with the prior art, the present invention has achieved the following technical effects:

[0029] (1) This paper proposes a graph weighted adversarial network, which can be used to solve the classification bias problem under imbalanced samples and ensure the prediction effect and robustness of the model;

[0030] (2) The present invention introduces an ensemble learning algorithm, uses a deep learning architecture as a base learner, uses different feature subsets to train the base learner, and integrates the voting algorithm output results. The diverse base learners can capture different characteristics of the data and reduce the risk of overfitting;

[0031] (3) The present invention explores the biological significance of the model by analyzing the conserved motifs of the sgRNA high-activity sequences and attempts to analyze the model in an interpretable manner;

[0032] (4) The present invention adopts the GNN model and takes advantage of its local perception to efficiently extract sequence features, significantly improving prediction accuracy and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Flowchart of the present invention.

[0034] Figure 2 This is a schematic diagram of the ACC effect of CS-Smote under an unbalanced dataset provided in an embodiment of the present application.

[0035] Figure 3 Schematic diagram of the G-mean effect of CS-Smote under an unbalanced dataset provided in an embodiment of the present application.

[0036] Figure 4 Schematic diagram of the results under different classifiers provided in the embodiments of this application.

[0037] Figure 5 A schematic diagram comparing the recognition effects of the embodiment of the present application with those of the existing recognition algorithms. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] The present invention discloses a method for predicting target activity of sgRNA, comprising the following steps:

[0040] Step 1: Obtain sgRNA sequence dataset

[0041] The sgRNA activity sequence dataset includes a positive dataset and a negative dataset. The positive dataset is a high-activity sgRNA sequence, and the negative dataset is a low-activity sgRNA sequence. The sgRNA activity sequence data has a total of 8 sample sets, including: G17, Gr, Gnr, and Gm;

[0042] Step 2: Extract sequence features from the sgRNA sequence dataset to obtain multiple feature information

[0043] Performing sequence feature extraction on the sgRNA sequence dataset based on a plurality of different feature extraction algorithms to obtain a plurality of feature information, wherein in some optional embodiments, the feature extraction algorithms include a nucleotide composition method, a sequence autocorrelation method, a nucleotide composition method, and a sequence structural feature method;

[0044] Nucleotide composition methods include k-mer extraction algorithm and Subsequence extraction algorithm;

[0045] The self-organizing correlation feature extraction algorithm includes the feature extraction algorithm based on automatic covariance DAC, the extraction algorithm based on cross covariance DCC, the extraction algorithm based on automatic cross covariance DACC, the algorithm of Geary autocorrelation GAC and the algorithm of normalized Moreau-Broto autocorrelation NMBAC;

[0046] The pseudo nucleic acid composition feature extraction algorithm includes an algorithm based on parallel correlation pseudo dinucleotide composition and a method based on continuous correlation pseudo dinucleotide composition;

[0047] The structural feature extraction algorithm includes the local structure sequence triplet feature extraction algorithm;

[0048] Step 3: Fuse the multiple feature information to obtain the feature set

[0049] In some optional implementations, the 11 feature files are spliced ​​and fused using a pre-fusion method;

[0050] Of course, in some other optional implementations, a late fusion method may be selected for feature fusion;

[0051] Step 4: Build an algorithm for processing imbalanced datasets based on graph weighted adversarial networks

[0052] First, we use a spectral spatial convolutional network (GCN) to address the class imbalance in disease prediction. We use its own features and its neighbor information to construct a neighborhood feature map for each node between samples. We use a graph convolutional layer to first aggregate the features of each node and its neighbors (based on the graph structure). We use a fully connected layer to find a potential representation for each node. Finally, we use a probabilistic classifier to select the class with the highest probability as the probability distribution of the label prediction class, thus predicting the class label of the input node.

[0053] Then, a separate weighted network learning parameter function is used to weight the samples in each class to help the classifier better fit between classes, and the classifier parameters are optimized through adversarial training methods to prevent the classifier from being biased towards any one class;

[0054] The optimization formula is as follows:

[0055]

[0056] in, and They are y i and q i The cth element of is the set of labeled nodes c, β c is the weight of the weighted cross entropy;

[0057] In order to emphasize the misclassified samples and the samples that are correctly classified but with low confidence, the classifier network and the weighted network We design an adversarial training process, using weighted cross entropy as the classifier loss function, so that the weighted network dynamically provides sample weights for the classifier. We use a two-player minimax game to formulate the adversarial process, in which both the classifier and the weighted network set try to optimize the following objectives:

[0058]

[0059] Among them, qi and w i It is the network D and W c s in sample x i During training, the classifier tries to correctly classify the training samples, especially those with high weights.

[0060] Step 5: Introduce ensemble learning and build a deep learning prediction model based on voting algorithm

[0061] After obtaining the reweighted adversarial graph convolutional network, it is used as the base classifier;

[0062] Then, using the voting algorithm of ensemble learning, we directly count the predicted categories of each model, select the category with the most votes, and use ten-fold cross-validation to verify the model results;

[0063] Step 6: Conserved motif analysis of sgRNA high activity sequences

[0064] Use MEME to identify frequently occurring motifs with a length of 5-10 bp. Set parameters such as -maxw 10 -nmotifs 5 and check the distribution of the motif in the sgRNA, such as whether it is enriched near the PAM, in the middle, or far away. Finally, draw a sequence logo map WebLogo to analyze the positional distribution characteristics of the nucleotides in the conserved motif.

[0065] Example 1:

[0066] Based on the G17 dataset, ACC and G-mean are used to evaluate the performance of the graph weighted adversarial network's imbalanced dataset processing algorithms CS-Smote, original data, smote, and bsmote. The Smote algorithm is a classic oversampling algorithm. The Bsmote algorithm mainly selects a small number of sample sets on the edge of the minority class and the majority class samples, and is an improved algorithm of the Smote algorithm. The results are shown in Figure 2. Figure 2 and Figure 3 shown.

[0067] from Figure 2 As can be seen, the ACC in the original dataset was generally low. The classification accuracy for the MED12 gene was the highest, approximately 91%, while the NF2 gene had the lowest, approximately 73%, a difference of approximately 18%. The classic smoothing algorithm performed best and achieved the highest classification accuracy for the 12 genes. Among these genes, the MED12 gene achieved a classification accuracy of at least 95%, while the THY1 gene had the lowest accuracy, approximately 80%. The bsmote and CS-Smote methods were then tested, and their prediction accuracy improved compared to the original data.

[0068] In addition, the algorithm used in this analysis achieved the highest classification accuracy for CD28, CUL3, and TAD2B genes, which to some extent demonstrates the effectiveness of this study's algorithm. G-mean was then used to analyze the performance of the four cases.

[0069] from Figure 3 As can be seen, CS-Smote achieves the best g-mean, followed by Smote and the original data. Seventeen genes have the highest CS-Smote values, with most genes having CS-Smote values ​​around 0.8. This result demonstrates the effectiveness of the CS-Smote method. The combined ACC and G-mean values ​​indicate that the CS-Smote algorithm improves predictive performance for imbalanced datasets to a certain extent.

[0070] Then, we obtained a better dataset by processing the imbalanced data and compared the performance of the four classifiers (RF, SVM, NB and J48) generated by ACC, SN, SP and MCC. Figure 4 shown.

[0071] from Figure 4 As can be seen from the broken line position of the algorithm sgRNA-RF of the present invention, sgRNA-RF achieved good results on all datasets. In the fifth dataset, all four genes achieved good results overall. For the Gr, Gnr, and Gm datasets, the recognition effect was generally poor, but in comparison, sgRNA-RF achieved the best recognition effect, demonstrating its effectiveness. However, a classifier needs to be constructed to adapt to additional datasets.

[0072] Finally, the present invention is compared with the research results of the existing excellent recognition algorithms. When comparing, the consistent data sets used are ensured and the consistent evaluation index ACC is used, such as Figure 5 shown.

[0073] To demonstrate the effectiveness of the proposed algorithm, sgRNA-RF, we compared it with other previously studied predictors in the G17 dataset and performed 10-fold cross-validation. Ge-CRISPR, Az-imuth, CRISPRpred, sgRNA-psm, and sgRNA-expsm are six of the most advanced predictors for sgRNA target activity in previous studies. The 10-fold cross-validation results show that all sgRNAs outperform previous studies. sgRNA-RF demonstrates reasonable effectiveness in predicting sgRNA activity and can provide a new approach to sgRNA research.

[0074] The above description is merely a preferred embodiment of the present invention and does not limit the technical scope of the present invention. Therefore, any minor modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A method for predicting target activity of sgRNA, characterized in that: The steps include: Step 1: Dataset preparation, obtain sgRNA activity sequence dataset; Step 2: Extract sequence features from the sgRNA activity sequence dataset to obtain multiple feature information; Step 3: Fusing the multiple feature information to obtain a feature set; Step 4: Construct an imbalanced dataset processing algorithm based on graph weighted adversarial network; Step 5: Introduce ensemble learning and build a deep learning prediction model based on the voting algorithm; Step 6: Conserved motif analysis of sgRNA high activity sequences.

2. The method for predicting target activity of sgRNA according to claim 1, wherein: In step 1, the sgRNA activity sequence dataset includes a positive example dataset and a negative example dataset, wherein the positive example dataset is a high-activity sgRNA sequence, and the negative example dataset is a low-activity sgRNA sequence.

3. The method for predicting target activity of sgRNA according to claim 1, wherein: In step 1, there are a total of 8 sample set pairs for sgRNA activity sequence data, including: G17, Gr, Gnr, Gm, hela, hct116, hek293t and h16.

4. The method for predicting target activity of sgRNA according to claim 1, wherein: In step 2, the feature extraction algorithm includes a nucleotide composition method, a sequence autocorrelation method, a nucleotide composition method, and a sequence structural feature method.

5. The method for predicting target activity of sgRNA according to claim 1, wherein: In step 2, the nucleotide composition method includes a k-mer extraction algorithm and a subsequence extraction algorithm; the self-organizing correlation feature extraction algorithm includes a feature extraction algorithm based on automatic covariance DAC, an extraction algorithm based on cross covariance DCC, an extraction algorithm based on automatic cross covariance DACC, a Geary autocorrelation GAC algorithm, and a normalized Moreau-Broto autocorrelation NMBAC algorithm; the pseudo nucleic acid composition feature extraction algorithm includes an algorithm based on parallel correlation pseudo dinucleotide composition and a method based on continuous correlation pseudo dinucleotide composition; the structural feature extraction algorithm includes a local structure sequence triplet feature Triplet extraction algorithm.

6. The method for predicting target activity of sgRNA according to claim 1, wherein: In step 4, first, the features of each node and its neighbors are aggregated using the graph convolution layer, and the potential representation of each node is found using the fully connected layer. Finally, the probability classifier is used to select the class with the highest probability as the probability distribution of the label prediction class to predict the class label of the input node. Then, a separate weighted network learning parameter function is used to weight the samples in each class to help the classifier fit better between classes, and the classifier parameters are optimized through adversarial training methods to prevent the classifier from being biased towards any one class.

7. The method for predicting target activity of sgRNA according to claim 6, wherein: In step 4, the optimization formula is: in, and They are y i and q i The cth element of is the set of labeled nodes c, β c are the weights of the weighted cross entropy.

8. The method for predicting target activity of sgRNA according to claim 1, wherein: In step 4, weighted cross entropy is used as the classifier loss function, so that the weighted network dynamically provides sample weights for the classifier, and a two-person minimum-maximum game is used to formulate the confrontation process, in which both the classifier and the weighted network set attempt to optimize the following objectives: Among them, q i and w i It is the network D and W c s in sample x i Output.

9. The method for predicting target activity of sgRNA according to claim 1, wherein: In step 5, after obtaining the reweighted adversarial graph convolutional network, it is used as the base classifier. Then, the voting algorithm of ensemble learning is used to directly count the predicted categories of each model, select the category with the most votes, and use the ten-fold cross-validation method to verify the model results.

10. The method for predicting target activity of sgRNA according to claim 1, wherein: In step 6, MEME is used to identify frequently occurring motifs with a length of 5-10 bp, parameters are set, the distribution of the motifs in the sgRNA is checked, a sequence logo map WebLogo is drawn, and the positional distribution characteristics of the nucleotides in the conserved motifs are analyzed.

Images