A method for detecting the integrity of adeno-associated virus using hidden Markov model

Abstract

The present invention belongs to the second generation sequencing data analysis technology in the field of bioinformatics, and in particular, it relates to a method for detecting the integrity of adeno-associated virus using a hidden Markov model; the present invention applies the HMM method to the detection of AAV genome integrity for the first time, obtains the optimal emission matrix through parameter search, optimizes the emission state, and constructs the best HMM model, thereby significantly improving the sensitivity and accuracy of detecting the integrity of the AAV genome. To solve the above problems, the present invention is based on the second generation sequencing and hidden Markov model (HMM) method, and optimizes the emission matrix state through a grid search strategy, which effectively improves the accuracy and detection efficiency of AAV genome integrity detection. This method can not only quickly test the integrity of the AAV genome at low cost, but also significantly improve the prediction accuracy, providing a new and efficient tool for quality control of gene therapy products.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics and relates to a method for detecting the integrity of adeno-associated virus using a hidden Markov model. Background Art

[0002] Hidden Markov Model: Hidden Markov Model, referred to as HMM, is a statistical model used to describe a system with hidden state Markov process. In this model, the state of the system is not directly visible (they are "hidden"), but the change of state can be inferred by a series of observed events or signals. HMM is widely used in speech recognition, natural language processing, bioinformatics and many other fields.

[0003] Emission Matrix: describes the probability of observing a particular observation in a given state. The emission matrix is ​​a core component of the HMM, which defines the probability distribution of observations that may be generated (or "emitted") by each hidden state in the model. In other words, it tells us how likely it is to observe a particular output in any given hidden state. For each state, the emission matrix provides the probabilities of all possible observations that constitute the visible output of the model.

[0004] Grid Search: Grid Search is a method for model hyperparameter optimization, which is widely used in machine learning and data science. The core idea of ​​this method is to systematically explore multiple parameter combinations within a specified parameter range to find the best model performance.

[0005] EM algorithm: Expectation-Maximization Algorithm. The EM algorithm is an iterative optimization algorithm for parameter estimation problems in probability models with hidden variables. The algorithm alternates between two steps: the E step (expectation step), which calculates the expectation of the hidden variables given the current parameter estimates; and the M step (maximization step), which updates the parameter estimates based on these expectations to maximize the likelihood function. The EM algorithm is often used in statistical modeling scenarios such as clustering and hidden Markov models.

[0006] ACC: Accuracy, referred to as ACC, is an important indicator to measure the performance of a classification model. It indicates the proportion of samples correctly predicted by the model to the total number of samples. In binary or multi-classification problems, accuracy provides an intuitive performance measure that reflects the overall predictive ability of the model for a given data set. bwa alignment: Burrows-WheelerAligner, referred to as bwa, is a widely used software tool specifically for aligning or mapping nucleic acid sequences (such as DNA or RNA) to larger reference genomes. It uses the Burrows-Wheeler transform (BWT) to efficiently process large-scale sequence data and is one of the key components in genomic research and high-throughput sequencing data analysis.

[0007] Adeno-associated virus (AAV) is a commonly used vector in clinical gene therapy. During the AAV packaging process, a three-plasmid system is required for packaging, and the AAV genome may be broken during this process. Recent studies have identified three types of incompletely packaged vector genomes: 1) double-stranded DNA containing 3' or 5' ITRs, called "self-complementary genomes" (SBGs), which may be completely complementary (symmetrical) or lack single-stranded complementary sequences (asymmetric); 2) incomplete single-stranded genomes containing only 3' or 5' ITRs; 3) genomes containing both ITRs but lacking the internal part of the AAV genome. Any vector that does not contain a complete AAV genome can only play a neutral or negative role, which will increase the number of AAV vectors delivered during gene therapy, but will not increase the expression of functional therapeutic products, and will lead to an increase in the total amount of DNA delivered to patients, which is not beneficial to gene therapy. Therefore, in therapeutic products using AAV as a vector, it is crucial to verify the integrity of the AAV vector and its integrity in the production process.

[0008] At present, the commonly used methods for detecting the integrity of AAV mainly include the following: 1) PCR-based methods, such as digital titration PCR (ddPCR) and traditional PCR. Although these methods are low-cost and easy to operate, they have a limited detection range and can usually only detect the integrity of the AAV genome in a specific primer region, and cannot fully detect breakpoint information; 2) First-generation sequencing-based methods, such as Sanger sequencing, can provide 1-2 times sequencing depth, but it is difficult to identify rare AAV breakpoints; 3) Second-generation sequencing-based methods, such as fast-seq, SSV-Seq, etc., can identify the reads coverage of the AAV genome, but cannot comprehensively evaluate the integrity of the AAV genome from a statistical perspective; 4) Third-generation sequencing-based methods, such as single-molecule real-time sequencing (SMRT), can detect the integrity of the AAV genome, but are expensive and require the construction of a self-complementary AAV genome. These methods have their own limitations. Although the cost of detecting the integrity of the AAV genome based on PCR or first-generation sequencing such as Sanger sequencing is low, different primer probes need to be designed to detect the integrity of different regions of AAV, which is not comprehensive. Although the methods based on second-generation sequencing such as fast-seq and SSV-Seq sequencing are relatively low in cost, they do not detect the integrity of the AAV genome from a statistical perspective such as the HMM method, and do not use grid search to find the best HMM model parameters and the optimal emission state to predict the integrity of the AAV genome, so the detection accuracy is relatively low. Although the methods based on third-generation sequencing such as SMRT theoretically have the ability to detect the integrity of the AAV genome, they are relatively low in accuracy in detecting the integrity of the AAV genome due to their high cost and slow speed, and they do not detect the integrity of the AAV genome from a statistical perspective. In summary, there is currently no method that can quickly and low-costly test the integrity of the AAV genome, and has a relatively high detection accuracy. The present invention is based on second-generation sequencing and HMM methods, and through grid search to traverse the probability of the emission matrix, optimize the emission matrix state, and can quickly and low-costly test the integrity of the AAV genome, and detect the integrity of the AAV genome with a higher accuracy.

[0009] In sequence analysis, HMM is widely used in many fields such as DNA copy number variation detection, gene prediction, protein family classification, and genome annotation. In detecting the integrity of the AAV genome, the HMM method can set the hidden state of the genome, such as defining the DNA molecule as complete or missing, which is based on an assumption that consecutive genomic bases are more likely to have the same state, that is, there is a local correlation between genomic bases. The state transition probability in the HMM can describe the conversion effect of missing or normal states between consecutive bases. In addition, the HMM can also predict the state of each base by the standardized value of the sequencing depth, which can be described by the emission matrix. The currently known HMM method detects CNVs. The method proposed by Seyed Amir Malekpour et.al to detect CNV integrity is to solve the parameters of the HMM model through the EM algorithm, without detecting it through the grid search method, and without optimizing the optimal emission matrix state. The accuracy of detecting the integrity of the CNV genome is not high enough. In theory, detecting the CNV features of the genome and detecting the integrity features of the AAV genome are similar methods, and both can determine whether the genome is missing based on the sequencing coverage of the reads of each base of the DNA molecule. Therefore, the application of the HMM method can be transferred to the detection of the integrity of the AAV genome. Therefore, there is currently a lack of a method for quickly detecting the integrity of the AAV genome at a low cost based on the second-generation sequencing data image through the HMM method. In addition, due to the limited number of our samples, in the process of constructing the HMM model, the model parameters may be overfitted. The data that performs well in the training set may not meet the expectations in the prediction results of the validation set. Therefore, we use the grid search method in machine learning, that is, traversing the emission matrix, to count the accuracy of the AAV genome integrity prediction under different emission matrix parameters, and select the optimal emission matrix. In addition, by setting different emission matrix states, such as 2 or 3 states, the optimal emission matrix state is selected. The above two optimization methods can significantly improve the accuracy of predicting the integrity of the AAV genome.

[0010] Therefore, the present invention proposes for the first time to detect the integrity of the AAV genome using the HMM method through the second-generation sequencing method. In order to improve the accuracy of detection and avoid overfitting of model parameters, the optimal emission matrix is ​​obtained through parameter search (grid search), the state of the emission matrix is ​​optimized (2 emission matrix states or 3 emission matrix states), and the optimal HMM model is constructed to predict the integrity state of the AAV genome. It can provide higher accuracy from a statistical point of view with a low-cost and rapid detection method, and improve the sensitivity of detecting whether the AAV genome state is complete. Summary of the invention

[0011] The present invention firstly provides a method for detecting the integrity of adeno-associated virus using a hidden Markov model. The method comprises the steps of using a hidden Markov model HMM in combination with a grid search method to find the optimal emission matrix parameters.

[0012] In some embodiments, the optimal transmission matrix parameters are two states or three states.

[0013] In some embodiments, the optimal transmission matrix parameters are in three states.

[0014] In certain embodiments, the adeno-associated virus is an adeno-associated virus based on next-generation sequencing data.

[0015] In some embodiments, the steps include:

[0016] 1) Select the standard set;

[0017] 2) Select the training validation set;

[0018] 3) Select the test set;

[0019] 4) According to the training validation set in step 2), the training validation set is divided into a training set and a validation set by three-fold cross validation, an HMM model is constructed, and the initial probability and the transfer matrix in the HMM model are calculated;

[0020] 5) Based on the training and validation sets, two types of emission matrices are constructed;

[0021] 6) For the emission probabilities in 2 states or 3 states of the emission matrix, grid search parameters are searched respectively, and the emission probabilities are traversed from 0.01 to 0.99 with an interval of 0.02, and the states of the AAV genomes of the samples in the validation set are predicted under different emission probabilities, that is, 2 states or 3 states respectively, and the accuracy under different emission probabilities of all validation sets is calculated, overfitting states are excluded, and the optimal emission parameters are selected;

[0022] 7) The optimal HMM model is obtained based on the optimal emission matrix probability obtained by grid search as a parameter, 3 emission matrix states are selected, and prediction is performed on all training validation set data in step 4), and the final prediction state of each base of the AAV genome in all training validation sets is counted, and the accuracy in all training validation sets is calculated.

[0023] In certain embodiments, the step of selecting a standard set includes: aligning N1 sets of original complete AAV virus liquid second-generation sequencing data to the AAV genome, selecting the optimal alignment, and obtaining a standardized sequencing depth for each AAV base position; calculating the mean and standard deviation of the standardized sequencing depth for each base of the AAV genome as the standard set; optionally, N1 is a natural number greater than or equal to 3.

[0024] In certain embodiments, the step of selecting a training validation set includes: aligning the second-generation sequencing data of AAV virus liquid with known different breakpoints, a total of N2 sets of breakpoint data, and N3 repeats in each group to the AAV genome, selecting the optimal alignment, and obtaining a standardized sequencing depth for each AAV base position as the training validation set of the model; optionally, N2 can be a natural number greater than or equal to 3, and N3 can be a natural number greater than or equal to 3.

[0025] In certain embodiments, the step of selecting a test set includes: aligning the second-generation sequencing data of AAV virus liquid with known different breakpoints, a total of N4 sets of breakpoint data, and N5 repeats in each group to the AAV genome, selecting the optimal alignment, and obtaining a standardized sequencing depth for each AAV base position as a test set, wherein N4 can be a natural number greater than or equal to 3, and N5 can be a natural number greater than or equal to 3.

[0026] In certain embodiments, the step 5) includes the following steps: 5.1) using the threshold 1 calculated in step 1) as the threshold, setting the emission matrix to two states, normal or missing, that is, judging whether each base of the AAV genome in the training set is greater than or equal to the threshold 1, if yes, judging that the base position in the AAV genome in the validation set is in a complete state; if not, judging that the base position in the AAV genome in the training set is in a missing state;

[0027] 5.2) According to the threshold 1 and threshold 2 calculated in step 1), respectively, the emission matrix is ​​set to 3 states: normal, missing or deeply missing. The specific steps are as follows: for each base of the AAV genome in the training set, it is determined whether it is greater than or equal to threshold 1. If so, it is determined that the base position in the AAV genome in the training set is in a complete state; if not, threshold 2 is used as a judgment. If it is less than threshold 2, it is determined that the base position in the AAV genome in the training set is in a deeply missing state, otherwise it is in a missing state.

[0028] Compared with the prior art, the present invention has at least the following beneficial effects:

[0029] 1) The current traditional methods for detecting the integrity of AAV virus have high technical costs and are slow, which limits the detection of AAV integrity. The present invention proposes a method for detecting the integrity of AAV virus at low cost and quickly using second-generation sequencing technology and HMM method, which can more efficiently identify and verify the integrity of AAV virus. This new method helps to quickly ensure the quality of AAV vectors in virus preparation and gene therapy research.

[0030] 2) The accuracy of detecting AAV virus integrity by conventional methods is not high. In order to improve the accuracy of detecting AAV virus integrity, the present invention adopts a grid search strategy to determine the optimal parameters of the HMM model emission matrix, which can effectively avoid the problem of overfitting of model parameters. In addition, compared with the selection of 2 emission matrix states in the conventional HMM method, the present invention selects 3 emission matrix states as the optimal emission matrix states, which significantly enhances the accuracy of detecting AAV genome integrity. By combining these two improvement strategies, not only the problem of overfitting of model parameters is avoided, but also a significant improvement is achieved in the accuracy of AAV genome integrity detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is the flowchart of AAVseq. For the AAV virus liquid second-generation sequencing data with missing fragments, the standardized sequencing depth of the AAV genome is obtained by bwa alignment and calculation, and the grid search method is used to traverse the emission matrix parameters to obtain the best emission matrix. When there are three emission states, it is selected as the optimal HMM model to predict the integrity status of the AAV genome;

[0032] Figure 2 The ACC value is calculated by grid search for the emission matrix parameters in the validation set under two emission states. The optimal emission matrix is ​​shown in Table 1, the accuracy of the data in the training validation set is shown in Table 3, and the accuracy of the data in the test set is shown in Table 4;

[0033] Figure 3 is the ACC value calculated by grid search parameter search in the validation set under three emission states. The optimal emission matrix finally obtained is shown in Table 2, the accuracy of the data in the training validation set is shown in Table 3, and the accuracy of the data in the test set is shown in Table 4.

[0034] Figure 4 Density map of the normalized sequencing depth of the AAV genome obtained by ContaVect analysis of the four test sets. The ContaVect software cannot statistically determine whether the AAV genome is complete or provide possible breakpoint information. DETAILED DESCRIPTION

[0035] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0036] Example 1 A method for detecting the integrity of adeno-associated virus using a hidden Markov model, the method comprising the following steps:

[0037] 1) First, 10 sets of original complete AAV virus liquid second-generation sequencing data were aligned to the AAV genome, and the optimal alignment was selected to obtain the standardized sequencing depth of each AAV base position; the mean and standard deviation of the standardized sequencing depth of each base in the AAV genome were calculated; as the standard set. Two thresholds were set according to the standardized sequencing depth of each base in the AAV genome in the standard set, threshold 1: mean, threshold 2: mean minus 2 times the standard deviation.

[0038] 2) The second-generation sequencing data of AAV virus liquid with known different breakpoints (a total of 4 sets of breakpoint data, 5 replicates in each group) were aligned to the AAV genome, and the optimal alignment was selected to obtain the standardized sequencing depth of each AAV base position; this was used as the training validation set for the model.

[0039] 3) The next-generation sequencing data of AAV virus fluid with known different breakpoints (a total of 3 sets of breakpoint data, 5 replicates in each group) were aligned to the AAV genome, and the optimal alignment was selected to obtain the standardized sequencing depth of each AAV base position; this was used as the test set.

[0040] 4) The training validation set in step 2) is divided into a training set and a validation set by 3-fold cross validation, and an HMM model is constructed. The initial probability in the HMM model and the transfer matrix are calculated according to the sequencing depth ratio of the AAV genome with different fragments in the training set.

[0041] 5) According to the training and validation sets, two types of emission matrices are constructed. The specific steps are as follows:

[0042] 5.1) Using the threshold 1 calculated in step 1) as the threshold, the emission matrix is ​​set to two states, normal or missing, that is, for each base of the AAV genome in the training set, it is determined whether it is greater than or equal to the threshold 1. If so, it is determined that the base position in the AAV genome in the training set is in a complete state; if not, it is determined that the base position in the AAV genome in the training set is in a missing state;

[0043] 5.2) According to the threshold 1 and threshold 2 calculated in step 1), respectively, the emission matrix is ​​set to 3 states: normal, missing or deeply missing. The specific steps are as follows: for each base of the AAV genome in the training set, it is determined whether it is greater than or equal to threshold 1. If so, it is determined that the base position in the AAV genome in the training set is in a complete state; if not, threshold 2 is used as a judgment. If it is less than threshold 2, it is determined that the base position in the AAV genome in the training set is in a deeply missing state, otherwise it is in a missing state.

[0044] 6) For the emission probabilities in the two states or three states of the emission matrix, grid search parameters are searched respectively, and the emission probabilities are traversed from 0.01 to 0.99 with an interval of 0.02, and the states of the AAV genomes of the samples in the validation set are predicted under different emission probabilities, that is, two states or three states respectively. The accuracy under different emission probabilities in all validation sets is calculated, overfitting states are excluded, and the optimal emission parameters are selected, such as Figure 2 , Figure 3 shown.

[0045] 7) The optimal HMM model is obtained based on the optimal emission matrix probability obtained by grid search as a parameter, and three emission matrix states are selected. The data of all training validation sets in step 4) are used as predictions. The final prediction state of each base of the AAV genome in all training validation sets is counted, and the accuracy in all training validation sets is calculated.

[0046] Example 2

[0047] This embodiment uses two sets of data sets: one set is a standard set consisting of 10 samples, which have complete AAV genomes and a sequencing depth of 2000x; the other set is a training validation set, which contains 20 samples divided into 4 groups, each with a different AAV breakpoint position, and an average sequencing depth of about 1500x. The training validation set is divided into a training set and a validation set by a 3-fold cross-validation method; the truncated AAV genome data of another 15 samples are additionally selected and used as a test set to verify the ability of the present invention in detecting the integrity of the AAV genome, that is, detecting whether the AAV genome has breakpoints, the location information of the breakpoints, etc.

[0048] In this example, we used the HMM method to detect the integrity of the AAV genome on the second-generation sequencing data.

[0049] 1. According to steps 1)-3) of Example 1, the second-generation sequencing data of the AAV virus liquid of the samples in the standard set, training set, validation set, and test set are first aligned to the AAV genome to obtain the standardized sequencing depth of the AAV genome base of each sample in the standard set, training validation set, and test set. And two thresholds are set according to the standardized sequencing depth of each base of the AAV genome in the standard set, threshold 1: mean, threshold 2: mean minus 2 times the standard deviation. The training set and validation set are divided by a 3-fold cross-validation method, and it is determined whether the standardized sequencing depth of each base in the AAV genome in the training set is significantly lower than the threshold.

[0050] 2. According to step 4), the parameters of the HMM model are initialized, and the initial probability values ​​are set according to the data in the training set, such as the transfer matrix, and the average length ratio of the breakpoints of the four groups of data in the training set is used as the transfer matrix, the emission matrix, etc. First, two emission matrix states are set, and the two parameters of the emission matrix are searched by grid search, that is, the probability of detecting that the mean of the AAV genome is less than that in the standard set in the case of a breakpoint, or the probability of detecting that the mean of the AAV genome is less than that in the standard set in the normal case. These two parameters are traversed from 0.01 to 0.99 with a gradient of 0.02, and the state of each base of the AAV genome of each sample in the validation set is judged. Finally, the model accuracy of the four groups of data in the validation set under all emission matrix parameters is statistically analyzed to obtain the best emission matrix, as shown in Table 1, that is, the state of the AAV base in each sample of each group of data is predicted by the HMM model, whether it is a breakpoint or a normal point.

[0051] Table 1 The best transition probability matrix obtained by grid search method under two transition states

[0052] Emission probability under different hidden states The probability that the site is in a normal state The probability that a site is in a deep missing state normal 0.69 0.31 Missing 0.49 0.51

[0053] Similarly, this embodiment tests three emission matrix states, namely, a normal state, a missing state, and a deep missing state, and searches for the best emission matrix under the three emission states by a grid search method, as shown in Table 2.

[0054] Table 2 The best transition probability matrix obtained by grid search method under three transition states

[0055]

[0056] For the above two emission matrix states, the accuracy of the HMM model under different emission matrix parameters is statistically analyzed (the ACC value is statistically analyzed), and the accuracy (acc) of all samples in all training validation sets and test sets is sorted out, as shown in Tables 3 and 4.

[0057] Table 3 ACC values ​​in the training and validation sets when the emission matrix is ​​in 2 or 3 states, compared with the ACC not obtained by the grid search method

[0058]

[0059] When the grid search method is used to optimize the emission matrix parameters, the accuracy (ACC) of the training validation set is significantly improved compared to the results without this method, whether in the case of two states or three states. This shows that the optimal emission matrix obtained by grid search can effectively increase the ACC value. In addition, compared with the emission matrix of two states, when the emission matrix is ​​three states, the ACC value in the training validation set is further significantly improved.

[0060] Table 4. Comparison of the ACC obtained by the test set with the emission matrix in 2 or 3 states and without the grid search method.

[0061]

[0062] When the emission matrix is ​​in 2 or 3 states, the accuracy (ACC) of the test set calculated according to the optimal parameters is significantly improved compared with the ACC obtained without the grid search method, indicating that the optimal emission matrix searched by the grid search method can significantly improve the ACC value in the test set; in addition, compared with selecting the emission matrix as 2 states, when the emission matrix is ​​in 3 states, the ACC value in the test set is also significantly improved.

[0063] It can be seen that the grid search method avoids the overfitting problem of the model due to the small number of samples. When the emission matrix is ​​in two states, the best state transfer matrix is ​​selected, and the average accuracy in the training and validation sets is about 83.89% (as shown in Table 3, mean), and the average accuracy in the test set is about 79.93% (as shown in Table 4, mean). When the emission matrix is ​​selected as three states at the same time, the HMM model constructed by the optimal parameters obtained by the grid search method is selected, and the best state transfer matrix is ​​selected. Finally, the average accuracy in the training and validation sets is about 90.02% (as shown in Table 3, mean), and the average accuracy in the test set is about 82.29% (as shown in Table 4, mean). Without the grid-search method, regardless of whether the emission matrix state is 2 or 3, the average accuracy in the training validation set is 73.90% and 77.73% (as shown in Table 3, mean), and the average accuracy in the test set is about 71.21% and 75.57% (as shown in Table 4, mean), indicating that the grid search method significantly improves the accuracy of HMM in detecting AAV genome breakpoints, making the HMM model have a higher generalization error. Therefore, the optimal parameters obtained by grid search and the state of the emission matrix can better detect the integrity of the AAV genome. Moreover, compared with the traditional PCR method, it has higher accuracy in detecting the integrity of the AAV genome, and can quickly detect the integrity of the AAV genome of multiple samples at one time, while the traditional PCR method requires the design of different primers and multiple tests to detect the integrity of different regions of the AAV genome.

[0064] In summary, compared with other detection methods, the present invention can quickly detect the integrity of multiple AAV sample genomes at low cost. The present invention avoids the overfitting problem of the model by running a statistical HMM method, traversing and searching the parameters of the emission matrix, and selecting the optimal model parameters. An HMM model with a relatively small generalization error is obtained through training, and predictions are made in the test set, and the accuracy in the test set is also relatively high. Compared with traditional PCR, this low-cost and rapid detection method can better detect the integrity of the AAV genome.

[0065] Example 3

[0066] In this example, the method of Example 2 is compared with the ContaVect software used in the SSV-seq method to illustrate that the method proposed by the present invention can better predict the integrity of the AAV genome from a statistical perspective. In this example, four groups of second-generation sequencing data of AAV virus liquid are first selected as the test data, and the average sequencing depth is 6000x. According to the optimal HMM model obtained in Example 1, that is, there are three optimized emission matrix states, and the optimal emission matrix parameters obtained by gridsearch search are used as the breakpoint information of the AAV genome predicted by the optimal HMM model, and finally the breakpoint information of the AAV genome is obtained. However, the ContaVect software can only give the average sequencing depth information of the AAV genome. As shown in Table 5.

[0067] Table 5 Comparison of the optimized HMM model (using three emission states and the optimal emission matrix) and ContaVect software in calculating the standardized sequencing depth of the AAV genome

[0068]

[0069] The normalized average sequencing depth values ​​of the non-deleted and deleted regions of the AAV genome calculated by the two methods are close, indicating that the two methods show a high degree of consistency in calculating the average sequencing depth of the AAV genome.

[0070] By comparing the calculation results of the average sequencing depth of non-deleted and deleted regions in the AAV genome by these two methods, the present invention found that their values ​​are very close, which confirms that the two methods have considerable uniformity and reliability in determining the average sequencing depth of the AAV genome. However, ContaVect cannot give the integrity of the AAV genome from a statistical point of view, as shown in Table 6. By selecting the optimal HMM method, the prediction of the breakpoint information of the AAV genome can be given. The average ACC in the four groups of test data is about 85.01%, which shows that compared with the currently known ContaVect software, the optimized HMM method can quickly detect the integrity of the AAV genome with higher accuracy and low cost.

[0071] Table 6 Comparison of the optimized HMM model (using three emission states and optimal emission matrix parameters) and ContaVect software in predicting AAV genome breakpoints in the training set data

[0072]

[0073] It can be seen that the HMM model can predict the breakpoint information of the AAV genome, while the ContaVect software cannot provide the breakpoint information of the AAV genome from a statistical point of view, so it is NA.

[0074] Therefore, the method for detecting the integrity of the AAV genome in the AAV virus liquid based on the optimized HMM method proposed in the present invention can quickly detect the integrity of the AAV genome at a low cost compared to other second-generation sequencing analysis methods such as ContaVect software.

[0075] In summary, the present invention uses a method based on a hidden Markov model (HMM) to find the optimal emission matrix parameters through a grid search technique. By comparing the model accuracy when the emission state is set to two and three situations, the present invention successfully determines the optimal emission matrix parameters and state configuration. This method significantly improves the accuracy of detecting the integrity of the adeno-associated virus (AAV) genome, proving its application value in the field of high-precision genome analysis.

[0076] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for detecting the integrity of adeno-associated virus using a hidden Markov model, characterized in that: The method comprises the following steps: 1) Select the standard set; 2) Select the training validation set; 3) Select the test set; 4) According to the training validation set in step 2), the training validation set is divided into a training set and a validation set by three-fold cross validation, an HMM model is constructed, and the initial probability and the transfer matrix in the HMM model are calculated; 5) Based on the training and validation sets, two types of emission matrices are constructed; 6) For the emission probabilities in 2 states or 3 states of the emission matrix, grid search parameters are searched respectively, and the emission probabilities are traversed from 0.01 to 0.99 with an interval of 0.02, and the states of the AAV genomes of the samples in the validation set are predicted under different emission probabilities, that is, 2 states or 3 states respectively, and the accuracy under different emission probabilities of all validation sets is calculated, overfitting states are excluded, and the optimal emission parameters are selected; 7) The optimal HMM model is obtained based on the optimal emission matrix probability obtained by grid search as a parameter, 3 emission matrix states are selected, and the data of all training and validation sets in step 4) are predicted. The final predicted state of each base of the AAV genome in all training and validation sets is counted, and the accuracy in all training and validation sets is calculated.

2. The method according to claim 1, characterized in that The step of selecting the standard set includes: aligning N1 sets of original complete AAV virus liquid second-generation sequencing data to the AAV genome, selecting the optimal alignment, and obtaining the standardized sequencing depth of each AAV base position; calculating the mean and standard deviation of the standardized sequencing depth of each base of the AAV genome as the standard set; optionally, N1 is a natural number greater than or equal to 3.

3. The method according to claim 1, characterized in that The step of selecting a training validation set includes: aligning the second-generation sequencing data of AAV virus liquid with known different breakpoints, a total of N2 sets of breakpoint data, and N3 repeats in each group to the AAV genome, selecting the optimal alignment, and obtaining a standardized sequencing depth for each AAV base position as a training validation set for the model; wherein N2 is a natural number greater than or equal to 3, and N3 is a natural number greater than or equal to 3.

4. The method according to claim 1, characterized in that: The step of selecting a test set includes: aligning the second-generation sequencing data of AAV virus liquid with known different breakpoints, a total of N4 sets of breakpoint data, and N5 repeats in each group to the AAV genome, selecting the optimal alignment, and obtaining a standardized sequencing depth for each AAV base position as a test set, wherein N4 is a natural number greater than or equal to 3, and N5 is a natural number greater than or equal to 3.

5. The method according to claim 1, characterized in that The step 5) comprises the following steps: 5.1) Using the threshold 1 calculated in step 1) as the threshold, the emission matrix is ​​set to two states, normal or missing, that is, for each base of the AAV genome in the training set, it is determined whether it is greater than or equal to the threshold 1. If so, it is determined that the base position in the AAV genome in the validation set is in a complete state; if not, it is determined that the base position in the AAV genome in the training set is in a missing state; 5.2) According to the threshold 1 and threshold 2 calculated in step 1), respectively, the emission matrix is ​​set to 3 states: normal, missing or deeply missing. The specific steps are as follows: for each base of the AAV genome in the training set, it is determined whether it is greater than or equal to threshold 1. If so, it is determined that the base position in the AAV genome in the training set is in a complete state; if not, threshold 2 is used as a judgment. If it is less than threshold 2, it is determined that the base position in the AAV genome in the training set is in a deeply missing state, otherwise it is in a missing state.

Images