Deep learning-based CRISPR / Cas9 mediated DNA double-strand breakage repair result prediction method
Through the deep learning-based ECnnIndel model, using k-mers digital sequence and Electra and CNN modules to extract features, the accuracy and efficiency of prediction of CRISPR/Cas9-mediated DNA double-strand break repair results in the prior art was solved, and more efficient prediction results were achieved.
Patent Information
- Application Number
- CN202411882532.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-19
- Publication Date
- 2025-04-11
AI Technical Summary
There is still room for improvement in the prediction accuracy and efficiency of CRISPR/Cas9-mediated DNA double-strand break repair results in the prior art, especially affected by factors such as sgRNA design, Cas9 protein type and PAM sequence recognition.
Using a deep learning-based method, the ECnnIndel model is constructed, the k-mers digital sequence is used as input, and features are extracted in combination with Electra and CNN modules are extracted to predict the DNA double-strand break repair results. This method improves the accuracy and efficiency of prediction by automatically extracting features.
It effectively improves the prediction accuracy and efficiency of the results of DNA double-strand break repair mediated by CRISPR/Cas9, and promotes the research progress of the double-strand DNA break repair mechanism.
Smart Images
Figure CN119811482A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of genome editing technology, and in particular to a method for predicting the results of CRISPR / Cas9-mediated DNA double-strand break repair based on deep learning. Background Art
[0002] The CRISPR / Cas9 system is a widely used genome editing technology that specifically recognizes and cuts the target DNA sequence through the combination of a specifically designed single guide RNA and the Cas9 nuclease. The Cas9 protein relies on the PAM sequence for precise targeted cutting, resulting in double-stranded DNA breaks. Cells repair this break mainly through two pathways: non-homologous end joining (NHEJ) and homologous recombination (HDR). Several studies have now gained in-depth insights into the mechanisms that affect Cas9-induced DNA editing. However, the existing technology has limited understanding of how Cas9 interacts with sgRNA and how cells respond to Cas9-mediated DNA damage.
[0003] At present, studies have shown that the repair results of Cas-induced double-strand DNA breaks are predictable and non-random. The use of computer-assisted prediction of CRISPR / Cas9-mediated DNA double-strand break repair results can increase the understanding of the mechanism of double-strand DNA break repair, but its accuracy and efficiency are still affected by many factors, including the design of sgRNA, the type of Cas9 protein, and the recognition of PAM sequences. In recent years, the use of computer-assisted prediction of CRISPR / Cas9-mediated DNA double-strand break repair results has become a research hotspot, mainly including traditional machine learning methods, deep learning methods and their hybrid applications. Traditional machine learning methods rely on manual feature construction and may have redundancy and feature selection problems. The use of deep learning methods significantly improves the accuracy and efficiency of prediction by automatically extracting features, and has become the mainstream direction of current research. Summary of the invention
[0004] The technical problem to be solved by the embodiments of the present invention is to provide a CRISPR / Cas9-mediated DNA double-strand break repair result prediction method based on deep learning, which can accurately predict the double-strand DNA break repair result.
[0005] In order to solve the above technical problems, an embodiment of the present invention provides a method for predicting the results of CRISPR / Cas9-mediated DNA double-strand break repair based on deep learning, comprising the following steps: S1: Construct a dataset for editing result prediction, obtain each dataset for editing result prediction, and divide it into a training dataset and an independent test dataset; S2: Build a token dictionary for sequence encoding; S3: constructing a first k-mers digital sequence for model input; using a sliding window method to obtain a k-mers sequence of length 118 for the training data set, and converting it into a k-mers sequence of length 120, using the token dictionary to match each position of the k-mers sequence and converting it into the first k-mers digital sequence; S4: ECnnIndel model was trained and evaluated using a five-fold cross validation algorithm; S5: Convert the independent test data set into a second k-mers digital sequence and input it into the trained model to make predictions, and analyze the generalization ability and prediction performance of the ECnnIndel model.
[0006] A further solution that can be adopted is that each data in the data set in S1 contains a biological sequence with a length of no more than 60bp, and corresponding six editing result labels, namely deletion frequency, 1 bp insertion frequency, 1 bp deletion frequency, 1 bp frameshift frequency, 2 bp frameshift frequency and frameshift frequency.
[0007] A further solution that can be adopted is that S2 includes the steps of: [CLS]: 0, [SEP]: 1, [PAD]: 2, 'A': 3, 'C': 4, 'G': 5, 'T': 6, 'AAA': 7, 'AAC': 8, 'AAG': 9, 'AAT': 10, 'ACA': 11, 'ACC': 12, 'ACG': 13, 'ACT': 14, 'AGA': 15, 'AGC': 16, 'AGG': 17. 'AGT': 18, 'ATA': 19, 'ATC': 20, 'ATG': 21, 'ATT': 22, 'CAA': 23, 'CAC': 24, 'CAG': 25, 'CAT': 2 6. 'CCA': 27, 'CCC': 28, 'CCG': 29, 'CCT': 30, 'CGA': 31, 'CGC': 32, 'CGG': 33, 'CGT': 34, 'CTA': 35 , 'CTC': 36, 'CTG': 37, 'CTT': 38, 'GAA': 39, 'GAC': 40, 'GAG': 41, 'GAT': 42, 'GCA': 43, 'GCC': 44 , 'GCG': 45, 'GCT': 46, 'GGA': 47, 'GGC': 48, 'GGG': 49, 'GGT': 50, 'GTA': 51, 'GTC': 52, 'GTG': 53, 'GTT':54, 'TAA':55, 'TAC':56, 'TAG':57, 'TAT':58, 'TCA':59, 'TCC':60, 'TCG':61, 'TCT':62, 'TGA':63, 'TGC':64, 'TGG':65, 'TGT':66, 'TTA':67, 'TTC':68, 'TTG':69, 'TTT':70 are added to the token dictionary.
[0008] A further possible solution is that S3 comprises the steps of: S31: Scan each sequence of 60 bp in length using the method of sliding window size of 1 and 3 and step size of 1 in the training data set, respectively, to obtain a 1-mer subsequence of 60 and a 3-mer subsequence of 58, respectively, and concatenate the 1-mer subsequence and the 3-mer subsequence to obtain the k-mers sequence of 118 in length; S32: Add a start identifier [CLS] and an end identifier [SEP] to the beginning and the end of the k-mers sequence respectively to obtain a k-mers sequence with a length of 120; if the length of the k-mers sequence is less than 120, fill it with a padding identifier [PAD] before the end identifier of the k-mers sequence.
[0009] A further possible solution is that S4 comprises the steps of: S41: randomly dividing the first k-mers digital sequence into 5 subsets, selecting one subset each time as a k-mers digital sequence test set, and the remaining 4 subsets as k-mers digital sequence training sets, and the ratio of the k-mers digital sequence training set to the k-mers digital sequence test set is 0.8:0.2; S42: For each fold, the model is trained using the k-mers digital sequence training set, and the model performance is evaluated using the k-mers digital sequence test set; S43: Calculate the average performance and standard deviation of the five-fold cross validation to evaluate the generalization ability of the model.
[0010] A further possible solution is that the S4 further comprises the steps of: S44: The input of the ECnnIndel model is a k-mers digital sequence training set; S45: Apply the Electra module on the input data to capture the global dependencies of the sequence; S46: Input the output of Electra to the CNN module to extract local features; the CNN module includes two one-dimensional convolution layers, the number of convolution kernels is 64, and the sizes of the convolution kernels are set to 1x1 and 3x3 respectively; S47: Flatten the output of the CNN module, convert it into a one-dimensional vector, input it into the fully connected layer for linear transformation, map the extracted features to a new space, and the output dimension is 192; S48: The output layer outputs six prediction results, corresponding to six editing results, including deletion frequency, 1 bp insertion frequency, 1 bp deletion frequency, 1 bp frameshift frequency, 2 bp frameshift frequency and frameshift frequency.
[0011] A further possible solution is that the S4 further comprises the steps of: S49: Use the Sigmoid activation function to process each output of the final layer so that each output value is between [0,1].
[0012] A further solution that can be adopted is that the training process of the ECnnIndel model also includes an adaptive learning rate optimization algorithm Adam with a learning rate of 0.0001, and the training process includes the steps of: S411: input training data and pass it through the model forward propagation to generate prediction values; S412: Calculate a loss function using the predicted value and the true label to measure the prediction error; S413: Calculate the partial derivative of the loss function with respect to each model parameter through back propagation as a basis for parameter update; S414: After the gradient calculation is completed, the Adam algorithm processes and updates the gradient through the following steps: First and second moment estimates: Calculate the first-order moment momentum (mean) and second-order moment momentum (variance) of the gradient as follows
[0013]
[0014] in, and denote the mean and variance of the gradient, respectively. and denote the decay rates of the first-order moment and the second-order moment, respectively. is the current gradient; Bias correction: The first-order moment and the second-order moment are corrected, and the formula is as follows:
[0015] in, and represents the bias-corrected moment estimate; Parameter update: Use the corrected gradient mean and variance to dynamically adjust the learning rate of each parameter and update the model parameters. The formula is as follows:
[0016] in, is the learning rate, is a small constant to prevent zero errors, represents the updated parameters; Optimize iterations. The Adam algorithm repeats the above steps in each iteration, continuously adjusts the learning rate based on historical gradient information, and optimizes the model parameters until the loss function converges or the preset number of training times is reached.
[0017] A further possible solution is that S5 further comprises the steps of: The prediction performance of the ECnnIndel model was evaluated by calculating one of the method evaluation indicators: Pearson correlation coefficient, Spearman rank correlation coefficient, and Kendall correlation coefficient.
[0018] The implementation of the embodiments of the present invention has the following beneficial effects: the present invention uses the k-mers method to convert the DNA sequence into a digital sequence, inputs the deep learning model for prediction, and obtains six results, which effectively solves the shortcomings of insufficient equipment resources and low precision. ECnnIndel promotes the research progress of computer-assisted double-strand DNA break repair mechanism. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 This is a flow chart of the ECnnIndel prediction method in the present invention.
[0020] Figure 2 It is the token dictionary construction diagram in the present invention.
[0021] Figure 3 This is a sample diagram of sequence coding in the present invention. DETAILED DESCRIPTION
[0022] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings.
[0023] Combination Figures 1 to 3 Schematic diagram shown.
[0024] A method for predicting the results of CRISPR / Cas9-mediated DNA double-strand break repair based on deep learning in an embodiment of the present invention can accurately predict the results of double-strand DNA break repair. The method disclosed in the present invention requires training ECnnIndel before using it to predict the results of CRISPR / Cas9-mediated DNA double-strand break repair. Therefore, the present invention is divided into two parts. The first part is the training model, and the second part is to test the prediction accuracy of CRISPR / Cas9-mediated DNA double-strand break repair results. Main process reference Figure 1 , token dictionary reference Figure 2 , sequence coding reference Figure 3 , and implement it through the following steps.
[0025] Step 1: Create a dataset for editing result prediction.
[0026] First, each dataset for editing result prediction is obtained from a public database. Each data in the dataset contains a biological sequence with a length of no more than 60 bp and six corresponding editing result labels, namely deletion frequency, 1bp insertion frequency, 1 bp deletion frequency, 1 bp frameshift frequency, 2 bp frameshift frequency and frameshift frequency. Among them, the biological sequence will be converted in step 3 as the input of the model, and the six editing result labels will be used as the target output of the model. Secondly, among the obtained datasets, the dataset with large data volume is selected for training the model, and the dataset with small data volume is used as an independent test dataset.
[0027] Step 2: Build a token dictionary for sequence encoding.
[0028] Set [CLS]: 0, [SEP]: 1, [PAD]: 2, 'A': 3, 'C': 4, 'G': 5, 'T': 6, 'AAA': 7, 'AAC': 8, 'AAG': 9, 'AAT': 10. 'ACA': 11, 'ACC': 12, 'ACG': 13, 'ACT': 14, 'AGA': 15, 'AGC': 16, 'AGG': 17, 'AGT': 18, 'ATA ': 19, 'ATC': 20, 'ATG': 21, 'ATT': 22, 'CAA': 23, 'CAC': 24, 'CAG': 25, 'CAT': 26, 'CCA': 27, 'C CC': 28, 'CCG': 29, 'CCT': 30, 'CGA': 31, 'CGC': 32, 'CGG': 33, 'CGT': 34, 'CTA': 35, 'CTC': 36, 'CTG': 37, 'CTT': 38, 'GAA': 39, 'GAC': 40, 'GAG': 41, 'GAT': 42, 'GCA': 43, 'GCC': 44, 'GCG': 4 5. 'GCT': 46, 'GGA': 47, 'GGC': 48, 'GGG': 49, 'GGT': 50, 'GTA': 51, 'GTC': 52, 'GTG': 53, 'GTT' :54、'TAA':55、'TAC':56、'TAG':57、'TAT':58、'TCA':59、'TCC':60、'TCG':61、'TCT':62、'TGA':63、'TGC':64、'TGG':65、'TGT':66、'TTA':67、'TTC':68、'TTG':69、'TTT':70 are added to the token dictionary.
[0029] Step 3: Construct the first k-mers digital sequence for model input.
[0030] (1) For the training data set in step 1, we scan each 60 bp sequence using a sliding window size of 1 and 3 and a step size of 1, respectively, to obtain a 1-mer subsequence of 60 and a 3-mer subsequence of 58. Then we concatenate the 1-mer subsequence and the 3-mer subsequence to obtain a k-mers sequence of 118 in length.
[0031] (2) Add the start identifier [CLS] and the end identifier [SEP] to the beginning and end of the k-mers sequence, respectively, to obtain a k-mers sequence with a length of 120. If the length of the k-mers sequence is less than 120 (that is, the original sequence length is less than 60 bp), use the padding identifier [PAD] to fill in the end identifier of the k-mers sequence.
[0032] (3) Using the token dictionary from step 2, match each position of the k-mers sequence and convert it into the first k-mers numeric sequence (for example, through the token dictionary, 'A' is mapped to 3 and 'AAA' is mapped to 7) as the input of the model.
[0033] Step 4: Training and prediction methods.
[0034] (1) The ECnnIndel model was trained and evaluated using a five-fold cross-validation algorithm.
[0035] First, the first k-mers digital sequence in step 3 is randomly divided into 5 subsets, one subset is selected each time as the k-mers digital sequence test set, and the remaining 4 subsets are used as the k-mers digital sequence training set. The ratio of the k-mers digital sequence training set to the k-mers digital sequence test set is 0.8:0.2.
[0036] Secondly, for each fold, the model is trained using the k-mers digital sequence training set, and the model performance is evaluated using the k-mers digital sequence test set.
[0037] Finally, the average performance and standard deviation of the five-fold cross validation were calculated to evaluate the generalization ability of the model.
[0038] (2) ECnnIndel model structure.
[0039] First, the input of the ECnnIndel model is a k-mers digital sequence training set.
[0040] Second, the Electra module is applied on the input data to capture the global dependencies of the sequence.
[0041] Next, the output of Electra is input into the CNN module (containing two one-dimensional convolutional layers, the number of convolution kernels is 64, and the convolution kernel sizes are set to 1x1 and 3x3 respectively) to extract local features.
[0042] Then, the output of the CNN module is flattened, converted into a one-dimensional vector, and input into the fully connected layer for linear transformation to map the extracted features to a new space with an output dimension of 192.
[0043] Finally, the output layer outputs six prediction results, corresponding to six editing results, including deletion frequency, 1 bp insertion frequency, 1 bp deletion frequency, 1 bp frameshift frequency, 2 bp frameshift frequency and frameshift frequency.
[0044] In the above steps, the "Sigmoid" activation function is used to process each output of the final layer so that each output value is between [0, 1]. The formula of the "Sigmoid" activation function is
[0045] in, It is the output of the "Sigmoid" function, and its value range is between [0, 1]. is the input value, is the base of natural logarithms (approximately 2.71828).
[0046] (3) During the entire training process, the binary cross-entropy loss function (BCE Loss) is used for multi-label regression tasks.
[0047] The formula for the binary cross entropy loss function is
[0048] in, Indicates the number of samples. is the true value of the i-th sample, ranging from [0, 1]. is the predicted probability of the ith sample, and its value is between [0, 1]. It represents the natural logarithm with base e, which is used to measure the difference between the predicted value and the true value. and the true value The closer it is, the smaller the binary cross entropy loss value is, which means the better the prediction accuracy of the model is.
[0049] (4) In the above model training process, in order to improve the training efficiency, the present invention also uses the adaptive learning rate optimization algorithm Adam with a learning rate of 0.0001.
[0050] The training process is handled as follows: First, the input training data is forward propagated through the model to generate predicted values.
[0051] Secondly, the loss function is calculated by comparing the predicted value with the true label to measure the prediction error.
[0052] Next, the partial derivative of the loss function with respect to each model parameter is calculated through back propagation as the basis for parameter updating.
[0053] Then, after the gradient calculation is completed, the Adam algorithm processes and updates the gradient through the following steps: (i) First-order and second-order moment estimates.
[0054] Calculate the first-order moment momentum (mean) and second-order moment momentum (variance) of the gradient as follows
[0055]
[0056] in, and denote the mean and variance of the gradient respectively. and represent the decay rates of the first-order moment and the second-order moment, respectively. is the current gradient.
[0057] (ii) Bias correction.
[0058] In order to avoid the deviation of momentum estimation at the initial moment, the first-order moment and the second-order moment are corrected as follows:
[0059] in, and represents the moment estimate after bias correction. (iii) Parameter update. Use the corrected gradient mean and variance to dynamically adjust the learning rate of each parameter and update the model parameters. The formula is as follows:
[0060] in, is the learning rate, is a small constant to prevent zero errors. Represents the updated parameters. Finally, the optimization iteration, the Adam algorithm repeats the above steps in each iteration, continuously adjusts the learning rate based on historical gradient information, and optimizes the model parameters until the loss function converges or reaches the preset number of training times.
[0061] (5) For the ECnnIndel model of the present invention, the implementation process is as follows: (i) Input feature processing.
[0062] The model directly uses DNA sequences as input without the need for additional manual feature engineering, thus avoiding the complex process of manually constructing features.
[0063] (ii) Design and application of Electra module and CNN module.
[0064] In the feature extraction stage of the model, first, the Electra module is introduced to capture long-distance dependencies in the input sequence. Electra generates feature representations rich in global information by modeling the contextual information of the sequence. At the same time, the number of parameters of Electra is relatively small, which helps to reduce the computational overhead of the model. Secondly, the output features of the Electra module are processed by the CNN module. This module contains two one-dimensional convolutional layers with 64 convolution kernels, and the convolution kernel sizes are set to 1x1 and 3x3 respectively, which are used to extract features in the local area of the sequence.
[0065] (iii) Implementation of dual feature extraction mechanism. By combining the Electra module with the CNN module, the model simultaneously captures the long-distance dependencies and local feature patterns in the sequence during the feature extraction process, thereby generating more expressive feature representations to support subsequent predictions.
[0066] (iv) Prediction stage: The extracted features are flattened and then input into the fully connected layer, and finally the output layer generates the predicted values of the six editing results.
[0067] Step 5: Convert the independent test data set into a second k-mers digital sequence and input it into the trained model for prediction, and analyze the generalization ability and prediction performance of the ECnnIndel model.
[0068] Evaluate the trained model using an independent test set: First, use the token dictionary constructed in step 2 and the method proposed in step 3 to convert each sample sequence in the independent test set into a digital sequence.
[0069] Secondly, the digital sequence is input into the trained ECnnIndel model for prediction, and six frequency values are obtained, corresponding to the six editing results described above.
[0070] Finally, the binary cross entropy loss of the training set and the independent test set, as well as the variance calculated during the cross validation process, are compared. If a large variance or a significantly higher loss value on the test set occurs, it indicates that the model has poor generalization ability. At the same time, the prediction performance of the ECnnIndel model is evaluated by calculating evaluation indicators such as the Pearson correlation coefficient, the Spearman rank correlation coefficient, and the Kendall correlation coefficient.
[0071] For the evaluation indicators, the Pearson Correlation coefficient measures the linear relationship between the predicted value and the actual value. The value range is [-1, 1]. When the value is closer to 1, it means that there is a strong positive correlation between the model's prediction and the true value, and the performance of the model is better; and the closer the value is to -1 or 0, it means that the correlation is weak or there is no correlation.
[0072] The formula for Pearson Correlation is
[0073] in, represents the model prediction value of the i-th sample, Represents the true value of the i-th sample. represents the mean of the model prediction values of all samples, Represents the mean of the true values of all samples. Indicates the number of samples.
[0074] The Spearman rank correlation coefficient measures the rank correlation between the predicted value and the actual value. The value range is [-1, 1]. The closer it is to 1, the more consistent the ranking relationship between the model's predicted results and the actual values, and the better the model performance. The formula for Spearman Correlation is:
[0075] in, represents the ranking of the i-th sample in the predicted value, Represents the ranking of the i-th sample in the true value. Indicates the number of samples.
[0076] Kendall Tau measures the rank relationship between the predicted value and the actual value. The larger the value, the better the performance of the model. Its formula is:
[0077] Among them, is the number of positively correlated ranked pairs, is the number of negatively correlated ranked pairs, is the sum of all sorted pairs, is the number of possible remaining permutations. The above three evaluation indicators can be used as independent evaluation indicators to evaluate the prediction effect of the model. In addition, using these indicators at the same time and integrating the evaluation results of different dimensions makes the model performance evaluation more comprehensive.
[0078] The above disclosure is only a preferred embodiment of the present invention, which certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.
Claims
1. A method for predicting the results of CRISPR / Cas9-mediated DNA double-strand break repair based on deep learning, characterized in that: The following steps are involved: S1: Construct a dataset for editing result prediction, obtain each dataset for editing result prediction, and divide it into a training dataset and an independent test dataset; S2: Build a token dictionary for sequence encoding; S3: constructing a first k-mers digital sequence for model input; using a sliding window method to obtain a k-mers sequence of length 118 for the training data set, and converting it into a k-mers sequence of length 120, using the token dictionary to match each position of the k-mers sequence and converting it into the first k-mers digital sequence; S4: ECnnIndel model was trained and evaluated using a five-fold cross validation algorithm; S5: Convert the independent test data set into a second k-mers digital sequence and input it into the trained model to perform prediction, and analyze the generalization ability and prediction performance of the ECnnIndel model.
2. The method according to claim 1, characterized in that Each data item in the data set in S1 contains a biological sequence with a length of no more than 60 bp, and six corresponding editing result labels, namely, deletion frequency, 1 bp insertion frequency, 1 bp deletion frequency, 1 bp frameshift frequency, 2 bp frameshift frequency and frameshift frequency.
3. The method according to claim 1, characterized in that The S2 comprises the steps of: [CLS]: 0, [SEP]: 1, [PAD]: 2, 'A': 3, 'C': 4, 'G': 5, 'T': 6, 'AAA': 7, 'AAC': 8, 'AAG': 9, 'AAT': 10, 'ACA': 11, 'ACC': 12, 'ACG': 13, 'ACT': 14, 'AGA': 15, 'AGC': 16, 'AGG': 17, 'AGT':
18. 'ATA': 19, 'ATC': 20, 'ATG': 21, 'ATT': 22, 'CAA': 23, 'CAC': 24, 'CAG': 25, 'CAT': 26, 'CCA' :27, 'CCC': 28, 'CCG': 29, 'CCT': 30, 'CGA': 31, 'CGC': 32, 'CGG': 33, 'CGT': 34, 'CTA': 35, 'CTC' :36, 'CTG': 37, 'CTT': 38, 'GAA': 39, 'GAC': 40, 'GAG': 41, 'GAT': 42, 'GCA': 43, 'GCC': 44, 'GCG ': 45, 'GCT': 46, 'GGA': 47, 'GGC': 48, 'GGG': 49, 'GGT': 50, 'GTA': 51, 'GTC': 52, 'GTG': 53, 'GT 'T': 54, 'TAA': 55, 'TAC': 56, 'TAG': 57, 'TAT': 58, 'TCA': 59, 'TCC': 60, 'TCG': 61, 'TCT': 62, 'TGA': 63, 'TGC': 64, 'TGG': 65, 'TGT': 66, 'TTA': 67, 'TTC': 68, 'TTG': 69, 'TTT': 70 are added to the token dictionary.
4. The method according to claim 1, characterized in that: The S3 comprises the steps of: S31: Scan each sequence of 60 bp in length using the method of sliding window size of 1 and 3 and step size of 1 in the training data set, respectively, to obtain a 1-mer subsequence of 60 and a 3-mer subsequence of 58, and concatenate the 1-mer subsequence and the 3-mer subsequence to obtain the k-mers sequence of 118 in length; S32: Add a start identifier [CLS] and an end identifier [SEP] to the beginning and the end of the k-mers sequence respectively to obtain a k-mers sequence with a length of 120; if the length of the k-mers sequence is less than 120, fill it with a padding identifier [PAD] before the end identifier of the k-mers sequence.
5. The method according to claim 1, characterized in that The S4 comprises the steps of: S41: randomly dividing the first k-mers digital sequence into 5 subsets, selecting one subset each time as a k-mers digital sequence test set, and the remaining 4 subsets as k-mers digital sequence training sets, and the ratio of the k-mers digital sequence training set to the k-mers digital sequence test set is 0.8:0.2; S42: For each fold, the model is trained using the k-mers digital sequence training set, and the model performance is evaluated using the k-mers digital sequence test set; S43: Calculate the average performance and standard deviation of the five-fold cross validation to evaluate the generalization ability of the model.
6. The method according to claim 5, characterized in that The S4 further comprises the steps of: S44: The input of the ECnnIndel model is a k-mers digital sequence training set; S45: Apply the Electra module on the input data to capture the global dependencies of the sequence; S46: Input the output of Electra to the CNN module to extract local features; the CNN module includes two one-dimensional convolution layers, the number of convolution kernels is 64, and the sizes of the convolution kernels are set to 1x1 and 3x3 respectively; S47: Flatten the output of the CNN module, convert it into a one-dimensional vector, input it into the fully connected layer for linear transformation, map the extracted features to a new space, and the output dimension is 192; S48: The output layer outputs six prediction results, corresponding to six editing results, including deletion frequency, 1 bp insertion frequency, 1 bp deletion frequency, 1 bp frameshift frequency, 2 bp frameshift frequency and frameshift frequency.
7. The method according to claim 6, characterized in that The S4 further comprises the steps of: S49: Use the Sigmoid activation function to process each output of the final layer so that each output value is between [0, 1].
8. The method according to claim 1, characterized in that The training process of the ECnnIndel model also includes an adaptive learning rate optimization algorithm Adam with a learning rate of 0.0001. The training process includes the following steps: S411: input training data and pass it through the model forward propagation to generate prediction values; S412: Calculate a loss function using the predicted value and the true label to measure the prediction error; S413: Calculate the partial derivative of the loss function with respect to each model parameter through back propagation as a basis for parameter update; S414: After the gradient calculation is completed, the Adam algorithm processes and updates the gradient through the following steps: First and second moment estimates: Calculate the first-order moment momentum (mean) and second-order moment momentum (variance) of the gradient as follows in, and denote the mean and variance of the gradient, respectively. and denote the decay rates of the first-order moment and the second-order moment, respectively. is the current gradient; Bias correction: The first-order moment and the second-order moment are corrected, and the formula is as follows: in, and represents the bias-corrected moment estimate; Parameter update: Use the corrected gradient mean and variance to dynamically adjust the learning rate of each parameter and update the model parameters. The formula is as follows: in, is the learning rate, is a small constant to prevent zero errors, represents the updated parameters; Optimize iterations. The Adam algorithm repeats the above steps in each iteration, continuously adjusts the learning rate based on historical gradient information, and optimizes the model parameters until the loss function converges or the preset number of training times is reached.
9. The method according to claim 1, characterized in that: The S5 further comprises the steps of: The prediction performance of the ECnnIndel model was evaluated by calculating one of the method evaluation indicators: Pearson correlation coefficient, Spearman rank correlation coefficient, and Kendall correlation coefficient.