A sliding window-based method for evaluating primer pool dimerization in targeted PCR.

Abstract

This application discloses a sliding window-based method for evaluating primer pool dimerization in targeted PCR, relating to the field of bioinformatics. The method includes: establishing a standardized primer pool sequence database; extracting all primer sequences to build a primer pair combination library; performing base alignment within a sliding window to form a complete primer pool sliding window alignment database; simulating the introduction of single-base vacancies in overlapping regions based on error-tolerant rules to construct an optimal alignment database after error-tolerant processing of the complete primer pool; calculating the final dimerization risk score to construct a complete primer pool dimerization risk score database; constructing a database of graded results for complete primer pool dimerization risk assessment and a list of high-risk primers; inputting the high-risk primer list into a DQN reinforcement learning model, iteratively obtaining a preliminary optimized primer pool, and performing a secondary assessment to obtain the final optimized primer pool. This application overcomes the limitations of manual rule sets and achieves a dynamic balance between dimerization risk and primer pool function.

Description

Technical Field

[0001] This application relates to the field of bioinformatics, and in particular to a method for evaluating the dimerization of targeted PCR primer pools based on a sliding window. Background Technology

[0002] Targeted PCR technology is a core method in molecular biology for achieving precise gene amplification and detection. Ultra-multiplex targeted PCR, with its advantage of simultaneously amplifying tens to tens of thousands of gene targets in a single experiment, has been widely applied in various scenarios such as early cancer screening, high-throughput detection of pathogenic microorganisms, genomics research, and genetic disease diagnosis. As a core component of ultra-multiplex targeted PCR, the rationality of its design directly determines the efficiency, specificity, and accuracy of PCR amplification results. The non-specific formation of primer dimers is a core challenge that urgently needs to be addressed in primer dimer design. The double-stranded nucleic acid structure formed by complementary base pairing between primers in the primer dimer system competitively consumes Taq polymerase, dNTPs, and Mg²⁺ in the PCR reaction system. 2 The presence of key reaction materials such as α and β can lead to a significant decrease in the amplification efficiency of target genes. It may also cause non-specific amplification and even false negative or false positive experimental results. Therefore, accurate and efficient dimer risk assessment of the primer pool for ultra-multiplex targeted PCR is a necessary prerequisite for ensuring the success of ultra-multiplex targeted PCR experiments.

[0003] Several patents have disclosed technologies related to the assessment of dimerization in targeted PCR primer pools (CN120126567A, CN119296644A), providing basic technical ideas and implementation methods for primer pool dimer assessment. These technologies are mostly based on base pairing comparison and have initially achieved quantitative assessment of dimer risk. However, in the practical application of ultra-multiple primer pools, the following shortcomings still exist. First, the dimerization risk determination process relies excessively on manually constructed multidimensional rule sets. Existing patent technologies all formulate judgment rules by manually sorting out typical characteristics of dimer formation to classify the severity of dimerization risk. These artificial rules can only capture basic dimer formation characteristics such as continuous base complementarity and simple 3' end matching. They cannot accurately determine complex dimer formation modes such as 3' end extension-induced dimerization, dimerization formed by multi-regional dispersed complementarity, and dimerization mediated by GC-enriched regions. Second, the primer pool optimization strategy is too simplistic and crude. Existing patented technologies all directly remove high-risk primers based on scores, without considering the overall functional requirements of the primer pool. The target coverage of a multi-multiplex primer pool is the core indicator of the experiment. Directly removing high-risk primers can easily lead to some target sites having no corresponding amplification primers, resulting in a decrease in target coverage. At the same time, it will also disrupt the uniformity of amplification efficiency in the primer pool. Excessive removal of some primers may also cause non-specific pairing of the remaining primers, which will increase the risk of dimer formation. It is impossible to achieve a balance between dimerization risk control and the targeted amplification function of the primer pool.

[0004] The two defects mentioned above in the existing patented technology lead to problems such as inaccurate judgment in complex scenarios and functional imbalance after optimization in the dimerization evaluation of primer pools for ultra-multiplex targeted PCR, making it difficult to meet the core requirements of ultra-multiplex targeted PCR experiments for high specificity, high coverage and high amplification efficiency of primer pools. Summary of the Invention

[0005] This application provides a sliding window-based method for evaluating the dimerization of targeted PCR primer pools, in order to solve the above-mentioned problems in the prior art.

[0006] This application provides a sliding window-based method for evaluating primer pool dimerization in targeted PCR, including: The original sequences of primer pools designed for targeted PCR experiments were collected and preprocessed to form a standardized primer pool sequence database. All primer sequences are extracted from the standardized primer pool sequence database to establish a primer pair combination library. For any primer combination in the primer pair combination library, base alignment is performed in a sliding window to form a full primer pool sliding window alignment of the original database. The overlapping region information, original base pairing information and basic statistical data of each sliding position are extracted from the original database of the full primer pool sliding window alignment. Based on the fault tolerance processing rules, a single base vacancy is simulated to be introduced in the overlapping region to generate a fault tolerance alignment scheme under the current sliding position. The optimal result of the fault tolerance alignment scheme under each sliding position is selected to construct the optimal alignment database after fault tolerance processing of the full primer pool. The final dimerization risk score of each primer pair is calculated based on the optimal alignment database after full primer pool fault tolerance processing, and a full primer pool dimerization risk score database is constructed. The data from the full primer pool dimerization risk score database is input into the CNN+BiLSTM model to obtain the dimerization risk level and risk type results for each primer pair, and to construct a full primer pool dimerization risk determination and classification result library and a list of high-risk primers. The initial environment of the DQN reinforcement learning model is established using the full primer pool dimerization risk assessment and classification result library. The high-risk primer list and the pre-set candidate primer library are input into the DQN reinforcement learning model. Iteration is performed in the initial environment to obtain a preliminary optimized primer pool. The preliminary primer pool is then re-performed with sliding window alignment, fault tolerance processing, risk score calculation and risk assessment. If the primer pool obtained again meets the optimization requirements, the primer pool obtained again is used as the final optimized primer pool.

[0007] The sliding window-based targeted PCR primer pool dimerization evaluation method in this application has the following advantages: 1. Deep Learning Drives Intelligent Judgment, Overcoming the Limitations of Manual Rule Sets. This application employs a CNN+BiLSTM deep learning model to replace the existing patented technology's manual multidimensional rule set, achieving intelligent classification of dimer risk and accurate identification of complex risk types. By learning the correlation between the sliding window alignment features, fault-tolerant processing features, quantitative scoring features, and wet experiment dimer formation results of massive primer pairs, the model can accurately capture complex dimer formation patterns that cannot be identified by manual rules. It can not only output four precise risk levels (none / low / medium / high) but also clearly label core risk factors and risk types. Simultaneously, the deep learning model possesses extremely strong generalization ability, significantly reducing operating costs and solving the core defects of existing patented technologies, such as lagging rule set updates and poor generalization.

[0008] 2. Reinforcement learning-mediated global optimization achieves a dynamic balance between dimerization risk and primer pool function. This application introduces a Deep Q-Network (DQN) reinforcement learning model to replace the single, coarse-grained optimization strategy of directly eliminating high-risk primers in existing patented technologies. The model incorporates dimerization risk, target coverage, and amplification efficiency uniformity into the optimization objectives. High-risk primers are not directly eliminated; instead, the model first attempts to fine-tune the 5' end with 1-2 nonsense bases. Only when the risk is not reduced after fine-tuning are high-risk primers eliminated, and low-risk primers with matching target sites are automatically selected from a pre-set candidate primer library to supplement the model, ensuring that the target coverage of the primer pool does not decrease. This optimization strategy controls the risk of dimer formation at its source while ensuring the core amplification function of the ultra-multiple primer pool. It solves the core defects of existing patented technologies, such as decreased target coverage and imbalanced amplification efficiency after optimization, achieving dual protection of dimerization risk control and primer pool function.

[0009] 3. Refined sliding window alignment design, balancing evaluation accuracy and computational efficiency. This application retains the core alignment logic of a 1bp sliding window step size to ensure the precision of base pair alignment. At the same time, reasonable preprocessing is performed on primer pair combinations, and pairwise alignment is performed only on any two primers in the primer pool. Self-primer combinations only include the forward sequence and its own reverse complementary sequence, avoiding invalid alignment combinations and reducing computational load.

[0010] 4. Error tolerance processing aligns with actual PCR reactions, improving the accuracy of sliding window alignment. This application employs a single-base insertion / deletion error tolerance processing method. Combined with the actual rules of primer dimer formation, the error tolerance processing is finely designed, allowing only one vacancy to be introduced into the overlapping area at each sliding position. It clearly distinguishes between original mismatches and mismatches introduced by the error tolerance vacancy. This not only restores the actual base pairing scenario during random collisions and hybridization of primers in solution but also avoids the decrease in alignment accuracy caused by indiscriminate error tolerance processing. This makes the alignment results of the sliding window more closely resemble actual PCR experiments, providing more reliable basic data for subsequent quantitative scoring and risk assessment. Compared to the simple alignment or single error tolerance of existing patent technologies, the reference value of the alignment results in this application is significantly improved. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1This is a flowchart of a targeted PCR primer pool dimerization evaluation method based on a sliding window, provided for embodiments of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] Figure 1 This application provides a flowchart of a sliding window-based targeted PCR primer pool dimerization evaluation method. The method includes: S100 collects the original sequences of primer pools designed for targeted PCR experiments, and after preprocessing, forms a standardized primer pool sequence database.

[0015] For example, a complete set of original primer sequences for targeted PCR experimental design is collected. The sequence format is standardized to the FASTA standard, ensuring that each primer sequence contains core information such as a unique primer number, target site, and 5'-3' end base sequence. Then, invalid primer sequences with duplicate numbers, ambiguous target sites, or abnormal base types are removed. Sequences with a length <15bp or >30bp, duplicate numbering ≥2 times, or missing target sites are also removed to ensure the validity of the original sequences. Abnormal base types refer to bases labeled with degenerate base markers (general degenerate base marker N, dibase degenerate marker, or tribase degenerate marker) or unknown base markers (X, ?, ...). ) or incorrect base identifiers (space, hyphen, underscore, slash).

[0016] Furthermore, the preprocessing of the original primer pool sequences includes sequence normalization and verification, reverse complementary sequence generation, and structured database construction.

[0017] Specifically, sequence standardization verification involves standardizing the selected valid primer sequences, unifying the base writing format to uppercase English letters (A / T / C / G), eliminating sequences with a non-A / T / C / G base ratio of ≥1%, retaining the original length information of each primer without manual truncation or padding, and ensuring the originality and authenticity of the primer sequences.

[0018] Reverse complementary sequence generation involves generating a complete reverse complementary sequence for each standardized primer sequence, strictly following the base pairing principle (i.e., A pairs with T, C pairs with G). The sequence is generated in a 1:1 correspondence to ensure no omissions or errors. The reverse complementary sequence retains the primer number and target site information of the corresponding forward primer, forming a one-to-one sequence pair with the primer sequence, and clearly marking the sequence attributes, i.e., forward sequence / reverse complementary sequence.

[0019] The structured database construction involves integrating all primer sequences with their reverse complementary sequences to build a standardized primer pool sequence database. Each record in the database contains five core fields: primer number, target site, sequence attributes, 5'-3' end base sequence, and sequence length. The primer number and sequence attributes enable rapid indexing, retrieval, and association of data, ensuring efficient data retrieval in subsequent steps.

[0020] S110: Extract all primer sequences from the standardized primer pool sequence database, establish a primer pair combination library, and perform base alignment in a sliding window for any primer pair combination in the library to form a full primer pool sliding window alignment of the original database.

[0021] For example, the primer pair combination library contains primer combinations of sequence A and sequence B. The reverse complementary sequence of sequence B is compared with sequence A through a sliding window. During the sliding window comparison, sequence A is kept stationary, while the reverse complementary sequence of sequence B is controlled to move in 1bp increments along its 5' to 3' direction. During the sliding process, the sliding position number, overlapping region information, pairing results of each base position, total number of matching bases, number of GC matching bases, and number of mismatched bases are recorded to form the original database of sliding window comparison of the entire primer pool.

[0022] Specifically, all primer sequences are extracted from the standardized primer pool sequence database to generate a primer pair combination library. This library contains pairwise combinations of any two different primers from the primer pool. Self-combinations of a single primer are also included in this library, where the forward sequence of a primer is paired with its own reverse complementary sequence. The total number of combinations is C(n,2)+n, where C(·) represents the combination and n is the total number of primers. A unique alignment number is assigned to each primer pair combination, and a primer pair combination index table is constructed. The index table records information such as the alignment number, primer 1 number, primer 2 number, and target site combination, thus establishing a link between primer pair combinations and alignment data.

[0023] Any primer pair in the primer pair library is designated as sequence A and sequence B, both being forward sequences from the 5' to 3' ends. A uniform sliding window alignment rule is set as follows: ① Keep sequence A fixed, and control the reverse complementary sequence of sequence B to translate along its own 5' to 3' end in steps of 1 bp to ensure the precision of base alignment; ② The sliding start position is when the 3' end of sequence B just touches the 3' end of sequence A, ensuring no base overlap, and the sliding end position is when sequence B completely slides past sequence A, i.e., no base overlap. The total number of sliding positions = the length of sequence A + the length of sequence B - 1, covering all possible collision and pairing positions between primers; ③ At each sliding position, the base overlap region of the two primers is precisely defined, with the overlap region ranging from 0 bp to 25 bp. Only bases within the overlap region are paired and aligned; non-overlapping regions are not included in the calculation.

[0024] According to the above rules, a full sliding alignment process is performed for each primer pair. At each sliding position, the following operations are performed: ① Record the sliding position number, the start and end base sites of the overlapping region, and the total number of bases in the overlapping region; ② Pair each base position in the overlapping region one by one and determine the base pairing type, which is divided into matching (i.e., AT / CG standard complementary pairing) and mismatch (non-AT / CG base pairing); ③ Calculate the total number of matching bases, GC matching bases, and mismatch bases at the sliding position to provide data for the calculation of the basic alignment score.

[0025] Furthermore, after recording the sliding position number, overlapping region information, pairing results of each base position, total number of matching bases, number of GC matching bases and number of mismatched bases, the data is structured and integrated and associated with the corresponding alignment numbers to form the original database for full primer pool sliding window alignment.

[0026] Specifically, for each primer pair, the sliding positions are structurally integrated with the sliding position number, overlapping region information, base pairing results, total number of matching bases, number of GC-matched bases, and number of mismatched bases, and associated with the corresponding alignment number to form the raw sliding window alignment data for each primer pair. The raw alignment data for all primer pairs are then compiled to construct a full primer pool sliding window alignment raw database. Each record contains seven core data items, enabling rapid data retrieval and access via alignment number and sliding position number.

[0027] S120: Extract the overlapping region information, original base pairing information and basic statistical data of each sliding position from the original database of the full primer pool sliding window alignment. Based on the fault tolerance processing rules, simulate the introduction of single base vacancies in the overlapping region to generate a fault tolerance alignment scheme under the current sliding position. Perform optimal result selection on the fault tolerance alignment scheme under each sliding position to construct the optimal alignment database after full primer pool fault tolerance processing.

[0028] For example, the fault tolerance processing rules established in this application embodiment are as follows: ① In the base overlap region of each sliding position, only one gap is allowed to be introduced at any single base position of sequence A or sequence B, that is, only single base insertion or single base deletion is supported as the fault tolerance type, and the introduction of consecutive gaps or multiple gaps is not allowed; ② After the gap is introduced, the position is recorded as a mismatch, and the remaining base positions are matched one by one according to the original base pairing rules; ③ Fault tolerance processing is only for the current sliding position, and the fault tolerance processing of different sliding positions is independent of each other and does not generate correlation; ④ When the length of the overlap region is <5bp, fault tolerance processing is not performed, and the original comparison result is directly retained.

[0029] For each sliding position of each primer pair, the overlapping region information, original base pairing results, and basic statistical data for that position are first extracted from the original alignment database. Based on the error tolerance rules, single-base vacancies are sequentially simulated at all base positions in the overlapping region to generate all error tolerance alignment schemes for that sliding position. A maximum of 30 error tolerance schemes are generated for a single sliding position. Each scheme records in detail the primer sequence (Sequence A / Sequence B) introduced by the vacancies, the specific base position of the vacancies, the base pairing results after error tolerance, the total number of matched bases after error tolerance, the number of GC matched bases, and the number of mismatched bases.

[0030] Furthermore, based on local alignment fit, the optimal result of the fault-tolerant alignment scheme at each sliding position is selected. The criterion for judging local alignment fit is the product of the proportion of the number of matched bases after fault tolerance to the total number of bases in the overlapping region and the number of matched bases. The fault-tolerant alignment scheme with the highest local alignment fit is selected as the optimal alignment result after fault tolerance at the current sliding position. The optimal alignment result, overlapping region information, total number of matched bases after fault tolerance, number of GC matched bases, and number of mismatched bases at each sliding position are structured and integrated, and associated with the corresponding alignment number and sliding position number to form the optimal alignment database after fault tolerance processing of the entire primer pool.

[0031] Specifically, for all fault-tolerant alignment schemes at each sliding position, the optimal result is selected based on local alignment fit. The core criterion for fit is the product of the proportion of the number of matched bases after fault tolerance to the total number of bases in the overlapping region and the number of matched bases. The fault-tolerant alignment scheme with the highest fit and a score ≥60% is selected as the optimal alignment result after fault tolerance at that sliding position. If the fit of all fault-tolerant alignment schemes is lower than 60% or lower than the original alignment result, the original alignment result is retained as the optimal alignment result at that sliding position.

[0032] Next, for all sliding positions of each pair of primer combinations, the optimal alignment results (including original alignment / fault-tolerant alignment markers, gap introduction information), overlapping region information, total number of matching bases after fault tolerance, number of GC matching bases, and number of mismatched bases at each position are structurally integrated, and the corresponding alignment numbers and sliding position numbers are associated. At the same time, the mismatched sites caused by gap introduction are uniformly marked, with 1 marking gap mismatches and 0 marking original mismatches, clearly distinguishing them from the original mismatched sites to avoid confusion during subsequent scoring. Finally, an optimal alignment database after fault tolerance processing for the entire primer pool is constructed, and each record in the database contains 8 core fields.

[0033] S130. Calculate the final dimer risk score for each pair of primer combinations based on the optimal alignment database after fault tolerance processing for the entire primer pool, and construct a dimer risk score database for the entire primer pool.

[0034] Exemplarily, calculate the dimer score for each primer combination according to the following formula:

[0035] where Dimer score is the dimer score for a sliding position, N matched is the total number of correctly paired bases in the optimal alignment result after fault tolerance at the current sliding position, is the number of C-G paired bases in the optimal alignment result after fault tolerance at the current sliding position, α is the GC weighting coefficient, N mismatched is the number of mismatched bases in the optimal alignment result after fault tolerance at the current sliding position; For each pair of primer combinations, extract the dimer scores of all sliding positions, and select the maximum value as the final dimer risk score for the primer pair.

[0036] In the embodiments of the present application, the GC weighting coefficient α is adaptively assigned according to the overall GC content of the primer pair, taking four values of 0.5, 1.0, 1.5, and 2.0, to emphasize the characteristic that dimers are difficult to denature due to the higher thermal stability of GC pairing. Specifically, for each pair of primer combinations, calculate the average value of the overall GC content of the two primers, and perform α adaptive assignment of values according to the average value to ensure that the weighting coefficient is highly adapted to the physical and chemical characteristics of the primer sequence. The specific assignment rules are as follows: ① When the average value of the GC content ≤ 40%, α take 0.5; ② When 40% < average value of the GC content ≤ 50%, α take 1.0; ③ When 50% < average value of the GC content ≤ 60%, α take 1.5; ④ When the average value of the GC content > 60%, α take 2.0.

[0037] Next, for each sliding position of each primer pair, the optimal alignment database after fault tolerance processing was precisely extracted. N matched , , N mismatched The specific values, combined with the adaptive properties of this primer pair α The value is calculated using a quantitative scoring formula to determine the dimer score for the sliding position. The scoring result is recorded and associated with the comparison number and the sliding position number.

[0038] Finally, for each primer pair, the dimerization score of all its sliding positions is extracted, and the maximum value is selected as the final dimerization risk score of the primer pair. This score reflects the maximum risk and severity of dimer formation of the primer pair under all possible collision, pairing and fault-tolerant scenarios. The higher the score, the greater the probability of dimer formation and the stronger the harm.

[0039] The alignment number, primer number combination, and adaptive sequence of each primer pair are used. α The scores, scores at each sliding position, and final dimerization risk scores are structurally integrated and sorted from high to low according to the final dimerization risk scores, with a score accuracy of 0.1. A dimerization risk score database for the entire primer pool is constructed, and the scores are labeled with intervals: 0-5.0 is low risk, 5.1-10.0 is medium risk, and above 10.0 is high risk. This provides clear and orderly quantitative data for subsequent deep learning risk assessment.

[0040] S140. Input the data from the full primer pool dimerization risk score database into the CNN+BiLSTM model to obtain the dimerization risk level and risk type results for each primer pair, and construct a full primer pool dimerization risk determination and classification result library and a list of high-risk primers.

[0041] For example, the CNN+BiLSTM model includes an input layer, a convolutional layer, a pooling layer, a bidirectional LSTM layer, a fully connected layer, and an output layer.

[0042] Specifically, the CNN+BiLSTM model in this embodiment is an end-to-end multi-classification and multi-label joint decision-making model, capable of simultaneously achieving dichotomous risk level classification and complex risk type identification. This model is built on Python and the TensorFlow 2.10 framework, can run stably on conventional servers without requiring high-performance computing power, and is suitable for typical laboratory use cases. The model consists of six layers: an input layer, a convolutional layer, a pooling layer, a bidirectional LSTM layer, a fully connected layer, and an output layer. The structure and core parameters of each layer are as follows: Input layer: The input to the model is a primer pair feature matrix with a dimension of 32×32. This matrix is ​​constructed by standardizing and vectorizing 12 core features of the primer pair. The core features include the final dimerization risk score, GC weighting coefficient, the length of the overlapping region at the optimal alignment position, the number of 3' end matching bases, the number of GC matching bases, the vacancy introduction situation, the total number of matching bases, the number of mismatched bases, the average primer length, and the average GC content. The batch size of the input layer is set to 32, and the ReLU activation function is used to realize the non-linear transformation of features and improve the feature extraction capability of the model.

[0043] Convolutional Layers: This layer contains two one-dimensional convolutional layers. Its core function is to extract local spatial features from the feature matrix and capture the local correlation between primer alignment features and dichotomy risk. The first convolutional layer has 32 kernels, a kernel size of 3, a stride of 1, and uses the same padding method. It employs the ReLU activation function and a dropout rate of 0.2. The second convolutional layer has 64 kernels, a kernel size of 3, a stride of 1, and also uses the same padding method. It employs the ReLU activation function and a dropout rate of 0.2. The dropout rate is set to prevent overfitting and improve the model's generalization ability.

[0044] Pooling layer: It contains two max pooling layers, which correspond one-to-one with the convolutional layers. Its core function is to perform dimensionality reduction on the local features extracted by the convolutional layers, retain the core feature information, reduce the computational load of the model, and improve the running efficiency. The pooling kernel size is 2, the stride is 2, the padding method is same, and there is no activation function to ensure that the core features are not lost.

[0045] Bidirectional LSTM layer: This layer contains one bidirectional LSTM layer. Its core function is to extract temporal features from primer alignment data, adapting to the characteristics of one-dimensional temporal data of primer sequences. It accurately captures temporal information such as the positional distribution of base matches and the matching patterns from the 3' to the 5' ends. The number of LSTM units is set to 128, with 64 units each in the forward and reverse LSTM directions. The activation function is tanh, and the recurrent dropout rate is 0.2 to ensure the model's accurate extraction of temporal features.

[0046] Fully Connected Layers: This layer consists of two fully connected layers. Its core function is to perform deep fusion and nonlinear transformation on the local spatial features extracted by the convolutional layers and the temporal features extracted by the bidirectional LSTM layer, thereby uncovering deep correlations between the features. The first fully connected layer has 256 neurons, uses the ReLU activation function, and has a dropout rate of 0.3. The second fully connected layer has 128 neurons, uses the ReLU activation function, and has a dropout rate of 0.3.

[0047] Output Layer: This layer contains two independent output branches, enabling joint output of multi-classification and multi-label methods to meet the dual requirements of dimeric risk classification and risk type identification. The two branches are independent yet collaborative in their determination. The risk level branch is a multi-classification output, classifying dimeric risk into four levels (Level 0: no risk, Level 1: low risk, Level 2: medium risk, Level 3: high risk). The activation function is Softmax, the output dimension is 4, and the output result is the probability value for each risk level. The maximum probability is taken as the final risk level of the primer pair. The risk type branch is a multi-label binary classification output, identifying four core complex dimeric risk types: 3' extension risk, self-dimeric risk, GC-enriched dimeric risk, and fault-tolerant vacancy-induced dimeric risk. The activation function is Sigmoid, the output dimension is 4, and the output result is the confidence level for each risk type. A confidence level ≥ 0.8 indicates that the primer pair possesses that type of risk.

[0048] During model training, the Adam optimizer was used with an initial learning rate of 0.001. A learning rate decay strategy was employed, decreasing the learning rate by 10% every 10 training epochs to ensure rapid model convergence. The number of training epochs was set to 50 to balance training effectiveness and efficiency. An early stopping strategy with a patience of 5 was implemented, meaning the model automatically stopped training when the loss function on the validation set did not decrease for five consecutive epochs, effectively preventing overfitting.

[0049] The dataset construction process for model training is as follows: Primer pool data and wet-lab validation data from historical targeted PCR experiments in the laboratory are collected to construct the model's training, validation, and test sets. The dataset samples are primer pair combinations, with each sample containing two parts: feature data and label data. ① Feature data: The 32×32 feature matrix of the primer pair (consistent with the model input layer); ② Label data: Annotated based on a large amount of tNGS sequencing data, PCR gel electrophoresis, and quantitative fluorescence detection results, including risk level labels (0 / 1 / 2 / 3) and risk type labels (0 / 1, 0 indicating no risk, 1 indicating the presence of risk), ensuring the authenticity and accuracy of the label data. The dataset is divided into a 7:2:1 ratio for the training set (70%), validation set (20%), and test set (10%), with a total sample size of no less than 10,000 samples to ensure the model's generalization ability and accuracy.

[0050] After the data is collected, the following preprocessing is required: ① Perform Z-score standardization on the feature matrix to transform all feature values ​​into standard values ​​with a mean of 0 and a variance of 1, eliminating the impact of differences in the dimensions of different features on model training; ② Perform one-hot encoding on the risk level labels and binarization encoding on the risk type labels to adapt to the output requirements of the model; ③ Perform data augmentation on the training set by randomly rotating the feature matrix and perturbing local features to improve the robustness and generalization of the model.

[0051] Input the preprocessed training set into the CNN+BiLSTM model, set the parameters as described above, and begin model training. During training, a joint loss function is used to calculate the model's training loss, enabling collaborative training for multi-classification and multi-label tasks. The formula for the joint loss function is: Loss =0.6× Loss cls +0.4× Loss multi in: Loss cls The cross-entropy loss for risk level multi-classification reflects the model's error in determining the risk level of the dimerization. Loss multi The binary crossentropy loss, applied to multiple risk types, reflects the model's error in identifying complex dichotomous risk types. During training, the loss values ​​on the training and validation sets, the accuracy of risk level determination, and the F1 score for risk type identification are monitored in real time to ensure the model gradually converges.

[0052] After model training is complete, the preprocessed test set is input into the model for performance verification. The core verification metrics are: ① Risk level determination accuracy ≥ 90%; ② Risk type identification F1 score ≥ 85%. If the model performance meets the above metrics, the training is considered successful, and the model is saved as a dimeric risk determination and grading model file, which can be directly used for new primer pool evaluation. If the metrics are not met, the model is retrained by adjusting parameters such as the number of convolutional kernels, the number of LSTM units, and the learning rate, or by expanding the training set sample size, until the performance meets the standards.

[0053] After the model training is complete, the dimerization risk assessment can be performed. This process includes the following: Model input feature construction: For each primer pair in the full primer pool dimerization risk score database, a 32×32 primer pair feature matrix is ​​constructed by combining the core features in the optimal alignment database after error tolerance processing. After Z-score standardization, the standard input data of the model is formed.

[0054] Model invocation and prediction: The standard input data is batch-inputted into the trained CNN+BiLSTM model. The model automatically outputs the dimerization risk level (0 / 1 / 2 / 3) and risk type results for each primer pair (marking the existing risk type and corresponding confidence level).

[0055] Integration of judgment results: The alignment number, primer number combination, final dimerization risk score, risk level, risk type and confidence level of each primer pair are structured and integrated. Core risk factors (such as 6bp continuous matching at the 3' end, complementary GC enrichment region, etc.) are marked for high-risk primer pairs (level 3) to construct a full primer pool dimerization risk judgment classification result library.

[0056] High-risk primer screening: Based on the primer pair determination results, risk statistics are performed on individual primers in the primer pool. The number of high-risk primer pairs formed by each primer with other primers is calculated and recorded as the high-risk pair count. If the high-risk pair count of a primer is greater than or equal to 5% of the total number of primers in the primer pool, it is marked as a high-risk primer. A high-risk primer list is constructed, recording the primer number, target site, high-risk pair count, and core risk type, providing a clear target for subsequent primer pool optimization.

[0057] S150: The initial environment state of the DQN reinforcement learning model is established using the full primer pool dimerization risk assessment and classification result library. The high-risk primer list and the preset candidate primer library are input into the DQN reinforcement learning model. Iteration is performed in the initial environment state to obtain the preliminary optimized primer pool. The preliminary primer pool is then re-performed with sliding window alignment, fault tolerance processing, risk score calculation and risk assessment. If the primer pool obtained again meets the optimization requirements, the primer pool obtained again is used as the final optimized primer pool.

[0058] For example, the DQN reinforcement learning model in this application embodiment is a discrete reinforcement learning model based on value functions. Its core consists of modules such as Agent, Environment, State, Action, and Reward. The model is built on the Python language and the PyTorch 1.13 framework, and its operating environment is consistent with that of the CNN+BiLSTM model. It can achieve autonomous decision-making and global optimization for primer pool optimization. The structure and core parameters of each module are as follows: The agent, the decision-making body of the model, is responsible for perceiving the environmental state and selecting the optimal action. It consists of an evaluation network (Q-network) and a target network, both with identical structures—fully connected networks—used to fit a value function and output the Q-value (the reward value of each action). The network structure has 16 neurons in the input layer, 64 neurons in hidden layer 1, 32 neurons in hidden layer 2, and 4 neurons in the output layer. The activation function for both is ReLU, and the dropout rate is 0.2. The optimizer uses the Adam optimizer with a learning rate of 0.0001. Parameter updates are performed with a soft update coefficient of 0.01 between the evaluation and target networks, updating the target network parameters every 100 steps to ensure model training stability.

[0059] Environment: Simulates the actual experimental scenario of primer pool optimization. The current composition and core features of the primer pool are the environment state. After the agent performs an action, the environment state changes (the composition of the primer pool changes), and the agent is fed back a reward value and the new environment state. The environment termination conditions are: ① There are no high-risk primer pairs in the primer pool (the risk level of all primer pairs is ≤1); ② The target coverage of the primer pool is 100% (original target coverage); ③ The number of iterations reaches the maximum limit (1000 times). If any of the conditions are met, the environment stops interacting with the agent.

[0060] State: This is a feature vector of the environment, with a dimension of 16. It contains 16 core global features of the primer pool, and all feature values ​​are standardized to values ​​between 0 and 1, comprehensively reflecting the overall state of the primer pool. Features include: proportion of high-risk primer pairs, proportion of medium-risk primer pairs, number of high-risk primers, average GC content of primers, average 3' end matching length, predicted mean of amplification efficiency, and coefficient of variation of amplification efficiency. The calculation process for the predicted mean amplification efficiency is as follows: Primer length, primer GC content, melting temperature (Tm), and 3' end base stability are extracted from the standardized primer pool sequence database in step S100 to obtain physicochemical characteristics. Dimerization risk characteristics are extracted from the full primer pool dimerization risk assessment and grading result database in step S140. The physicochemical characteristics are standardized and scored to map to the 0-1 interval. The physicochemical characteristics are linearly weighted and summed using the scores as weights to obtain a basic physicochemical score. A risk correction coefficient is determined based on the highest associated risk level among the dimerization risk characteristics. The product of this coefficient and the basic physicochemical score is used as the predicted primer amplification efficiency value. The arithmetic mean of all predicted amplification efficiency values ​​is calculated to obtain the predicted mean amplification efficiency.

[0061] Actions: These are discrete operations performed by the agent on the primer pool. Four core actions are defined, with an action space of {0, 1, 2, 3}. All actions do not alter the primer's targeted amplification function. Specifically: ① Action 0 (Retain): No operation is performed on the current primer; it remains in the primer pool. ② Action 1 (Fine-tuning): 1-2 nonsense bases are fine-tuned at the 5' end of high-risk primers to reduce their complementary pairing risk with other primers. ③ Action 2 (Removal): High-risk primers whose risk remains unreduced after fine-tuning are directly removed from the primer pool. ④ Action 3 (Supplement): Low-risk primers with matching target sites are selected from the pre-defined candidate primer library and added to the primer pool to compensate for the target coverage loss after primer removal.

[0062] Reward: This is the feedback value from the environment to the agent's actions, reflecting the quality of the actions. It is calculated using a multi-objective weighted reward function. The reward value is a continuous numerical value, taking into account dimeric risk control, target coverage maintenance, amplification efficiency uniformity, and primer pool adjustment. The formula is: Reward =0.4× R 1+0.3× R 2+0.2× R 3 0.1× R 4. Among them: R 1 represents the degree of improvement in dimerization risk, with a value ranging from 0 to 1, indicating the proportion by which the risk level of the primer pool is reduced after the action is performed; R 2 represents the target coverage retention rate, with a value range of 0-1, which is the coverage after the action is executed / the original coverage. R 3 represents the uniformity of amplification efficiency, with a value range of 0-1, and 1 represents the coefficient of variation of amplification efficiency; R 4 represents the primer pool adjustment rate, ranging from 0 to 1. It represents the number of primers that are fine-tuned, removed, or added after the action is executed, as well as the original number of primers. The higher the reward value, the better the action performed by the agent.

[0063] During model training, the experience replay pool size is set to 10,000 to store interaction experience samples of the agent; an ε-greedy strategy is adopted, with an initial value of 1.0, which decays linearly to 0.1 with the number of iterations, with a decay rate of 0.001, to balance the model's exploration and utilization capabilities; the maximum number of iterations per round is set to 1,000 to balance optimization effect and efficiency.

[0064] This model is an unsupervised reinforcement learning model. It accumulates experience samples through continuous interaction between the agent and the environment, and continuously optimizes and evaluates the network parameters, allowing the agent to learn the optimal strategy for primer pool optimization. The training process is divided into four stages: experience replay pool construction, agent-environment interaction exploration, network parameter update, and model validation. The specific operations of each stage are as follows: Experience replay pool construction: Initialize an empty experience replay pool to store experience samples generated by the agent's interaction with the environment. Each sample contains (current state S, action A, reward R, new state S', and termination flag Done). The maximum storage capacity of the experience replay pool is 10,000 samples. When the sample size reaches the limit, the oldest sample is deleted to ensure the validity of the experience samples.

[0065] Agent-Environment Interaction Exploration: Initialize the primer pool environment to the state feature vector of the original primer pool, set the exploration rate ε=1.0, and begin the agent's interaction with the environment. In each iteration, the agent selects an action based on the current environment state S using an ε-greedy strategy: ① If the random number ≤ ε, randomly select an action (explore new optimization strategies); ② If the random number > ε, select the action with the highest Q-value output by the evaluation network (utilize existing optimal strategies). After the agent executes an action, the environment state is updated to S', and the environment calculates and feeds back the reward value R according to the reward function, while simultaneously determining whether the termination condition (Done=True / False) is met. Experience samples (S, A, R, S', Done) are stored in the experience replay pool until the sample size reaches 5000 (batch update threshold).

[0066] Network parameter update: Once the sample size in the experience replay pool reaches a threshold, batch updates of the evaluation network parameters begin. The update process is as follows: ① Each time, 32 experience samples are randomly selected from the experience replay pool, and the current Q-value and target Q-value of the evaluation network are calculated. Q target The formula is: Q target = R + γ ×max( Q ( S' , A )),in γ The discount factor is set to 0.99, reflecting the weight of future rewards. Q ( S' , A The environmental state is S' Time action A ① Calculate the Q value; ② Use the mean squared error (MSE) as the loss function to calculate the loss between the current Q value and the target Q value. Loss The formula is: Loss =MSE( Q ( S , A ), Q target ),in, Q ( S , A The environmental state is S Time actionA The Q value; ③ The parameters of the evaluation network are updated using the Adam optimizer through the backpropagation algorithm, and the parameters of the evaluation network are synchronized to the target network every 100 steps by soft updating the coefficients.

[0067] Model Validation: After the model has been trained iteratively 10,000 times, training is stopped, and the test primer pool is input into the model for optimization and validation. The core validation metrics are: ① No high-risk primer pairs in the optimized primer pool; ② 100% target coverage; ③ Amplification efficiency coefficient of variation ≤ 10%; ④ Primer adjustment rate ≤ 20%. If all metrics meet the standards, the model training is considered successful, and the model is saved as a globally optimized primer pool model file. If the standards are not met, training continues to 20,000 times, or the reward function weights and action space are adjusted and retrained until the performance meets the standards.

[0068] After model training is complete, global primer pool optimization can be performed, including the following steps: Optimize the environment and initialize parameters: Input the original primer pool (i.e., the standardized primer pool sequence database in step S100), the high-risk primer list, and the preset candidate primer library (low-risk target primers) into the trained DQN model, initialize the environment state to the feature vector of the original primer pool, and set the exploration rate ε=0.1.

[0069] Autonomous optimization and iteration of the agent: The model begins to iterate and optimize. The agent selects the optimal action to execute based on the current environmental state (retain / fine-tune / remove / supplement). The environment provides real-time feedback of reward values ​​and updates the state until the environmental termination condition is met, at which point the iteration stops.

[0070] Preliminary optimization results output: After the iteration stops, the model outputs a preliminary optimized primer pool, including the optimized primer list, primer pair dimerization risk assessment results, target coverage, amplification efficiency prediction values, and other information.

[0071] Secondary validation of optimization results: The initially optimized primer pool is re-entered into the entire process of steps S110-S140 for secondary sliding window comparison, fault tolerance processing, quantitative scoring and deep learning risk assessment. If there are no high-risk primer pairs in the secondary assessment results and the target coverage is 100%, the optimization results are deemed valid; if there are still high-risk primer pairs, the secondary assessment results are fed back to the model and the optimization iteration is re-executed.

[0072] The final optimization result was determined as follows: the optimized primer pool that has been verified to be effective was selected as the final optimized primer pool.

[0073] Furthermore, after the iteration is completed, an intermediate primer pool optimization report is output along with the preliminary optimized primer pool. The intermediate primer pool optimization report records the primer numbers that were fine-tuned / removed / added, the reasons for the adjustments, and the risk improvement after the adjustments. Moreover, after obtaining the final optimized primer pool, the sequence information, target site information, dimerization risk assessment results, and amplification efficiency prediction values ​​of the final optimized primer pool are integrated to construct the final optimized primer pool database, and a final primer pool dimerization evaluation and optimization report is generated.

[0074] Experimental process 1. Experimental materials and equipment.

[0075] Pathogen Samples and Primer Pool: 300 clinically common infectious pathogens were selected as target species, covering bacteria, fungi, viruses, and parasites commonly causing respiratory and bloodstream infections. A super-multiplex primer pool containing 1000 pairs of specific primers was designed using Primer 3.0, targeting their conserved genomic regions. Each pathogen covers an average of 3-4 target regions. Simultaneously, a candidate primer library containing 5000 pairs of alternative primers was constructed. This candidate primer library was designed for different conserved regions of the same pathogen, and all primers underwent screening for amplification specificity ≥99%.

[0076] Experimental Sample (Template): Mock microbial community standards: a mixed nucleic acid sample containing 20 representative pathogens (such as Staphylococcus aureus, Escherichia coli, influenza virus, Candida albicans, etc.) with known copy numbers (10^3~10^5 copies / μL), with a total concentration of 50 ng / μL.

[0077] Human background DNA: Human genomic DNA (10 ng / μL) extracted from the blood of healthy individuals, with a single reaction addition of 5 μL, is used to simulate host background interference in clinical samples.

[0078] Experimental reagents: 2× high-fidelity multiplex PCR enzyme system (Multiplex PCR MasterMix), 200U / μL active reverse transcriptase (for RNA viruses), library construction kit containing 8bp Index adapter (including Index adapter), 0.8× magnetic bead purification reagent, and Qubit quantitative reagent with a detection range of 1~1000ng / μL.

[0079] Experimental equipment: Illumina NovaSeq PE150 gene sequencer or MGI DNBSEQ-G400 sequencing platform, Agilent 2100 bioanalyzer with a fragment range of 50~500bp, ABI 9700 PCR instrument, and a high-performance computing server equipped with an Intel Xeon Platinum 8358 CPU + 256GB memory (equipped with the evaluation and optimization algorithm of this application).

[0080] 2. Experimental steps.

[0081] This experiment is divided into two stages: dry experiment (algorithm evaluation and optimization) and wet experiment (tNGS library construction and sequencing verification). Each experiment has 3 biological replicates.

[0082] Dry experiment evaluation and optimization: The original sequences of 1000 pairs (2000 lines) of primer pools are entered into the server, and steps S100-S150 are executed sequentially.

[0083] Full pool alignment: Pairwise sliding window alignment and self-alignment were performed on 2000 primers, with a total of 1,999,000 alignments.

[0084] Deep learning judgment: Use a trained CNN+BiLSTM model to identify the risk of complex dimers, especially 3' end extension and GC content ≥60% enriched dimers.

[0085] Reinforcement learning optimization: The optimization objective is set as minimizing the dimerization risk, with a risk threshold ≤ 0.5. The DQN model automatically performs primer fine-tuning (modification of 1-2 bases at the 5' end), removes high-risk primers, and intelligently supplements alternative primers from the candidate library.

[0086] Output: Generates a final optimized primer pool sequence list and optimization report.

[0087] Wet experimental verification (tNGS process): Experiments were conducted using both the original primer pool and the final optimized primer pool, with three replicates for each group.

[0088] Multiplex PCR amplification: Add simulated microbial community standards and human background DNA to a 50 μL reaction system and perform one round of ultra-multiplex PCR amplification (25 cycles, annealing temperature 60℃).

[0089] Library construction: The amplification products were purified twice with magnetic beads, ligated with dual-index adapters, and labeled with indexes to construct sequencing libraries.

[0090] Quality control: The size distribution of library fragments was detected using an Agilent 2100 with a sample loading volume of 1 μL. The presence of a large number of dimer peaks of approximately 80-120 bp was observed.

[0091] Sequencing: After normalization to 2nM, the sample was sequenced. Each sample yielded approximately 5M reads, with a read length of 150bp.

[0092] Data analysis metrics: Dimer Rate: The proportion of reads in sequencing data that are <100bp in length after removing adapter sequences and cannot be matched to the pathogen genome.

[0093] On-target rate: The percentage of reads that match 300 target pathogens out of the total reads.

[0094] Uniformity: Coefficient of variation (CV) of the abundance detected in 20 representative pathogens.

[0095] Coverage: The percentage of successfully detected pathogen targets out of all target regions of 300 pathogens.

[0096] 3. Experimental results.

[0097] Dry experiment evaluation and optimization results: The original pool of 1000 primer pairs was evaluated using this method, and 185 high-risk dimer primer pairs were detected, accounting for 18.5%. These were mainly concentrated between bacterial primers with high GC content ≥60% and some viral primers.

[0098] After global optimization using the DQN model: Forty-two primers were fine-tuned at the 5' end by 1-2 bases; 25 primers whose risk could not be reduced through fine-tuning were removed; Twenty-five low-risk alternative primers were successfully added from the candidate library.

[0099] Final results: The final optimized primer pool still maintains 1000 primer pairs, the theoretical target coverage remains at 100%, and the number of high-risk primer pairs is reduced to 0.

[0100] The results of the wet assay (tNGS sequencing) are shown in Table 1: Table 1 Results of wet test

[0101] Experimental results demonstrate that the evaluation and optimization method provided in this application can effectively identify and eliminate complex dimer risks when processing pathogen tNGS primer pools with up to 1000 pairs of primers. Compared to the original design, the optimized primer pool reduced the proportion of dimer reads by 14-fold, significantly improving the utilization rate of sequencing data and the detection rate of pathogens, without sacrificing target coverage. This proves the practicality and advancement of this method in complex scenarios of ultramultiplex PCR.

[0102] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0103] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for evaluating the dimerization of targeted PCR primer pools based on a sliding window, characterized in that, include: The original sequences of primer pools designed for targeted PCR experiments were collected and preprocessed to form a standardized primer pool sequence database. All primer sequences are extracted from the standardized primer pool sequence database to establish a primer pair combination library. For any primer combination in the primer pair combination library, base alignment is performed in a sliding window to form the original database of full primer pool sliding window alignment. The overlapping region information, original base pairing information and basic statistical data of each sliding position are extracted from the original database of the sliding window of the full primer pool. Based on the fault tolerance processing rules, a single base vacancy is simulated to be introduced in the overlapping region to generate a fault tolerance alignment scheme under the current sliding position. The optimal result of the fault tolerance alignment scheme under each sliding position is selected to construct the optimal alignment database after fault tolerance processing of the full primer pool. Based on the optimal alignment database after the full primer pool fault tolerance processing, the final dimerization risk score of each primer pair is calculated, and a full primer pool dimerization risk score database is constructed. The data from the full primer pool dimerization risk score database is input into the CNN+BiLSTM model to obtain the dimerization risk level and risk type results for each primer pair, and to construct a full primer pool dimerization risk determination and classification result library and a list of high-risk primers. The initial environment state of the DQN reinforcement learning model is established using the full primer pool dimerization risk assessment and classification result library. The high-risk primer list and the preset candidate primer library are input into the DQN reinforcement learning model. The model is iterated in the initial environment state to obtain a preliminary optimized primer pool. The preliminary primer pool is then re-performed with sliding window alignment, fault tolerance processing, risk score calculation, and risk assessment. If the primer pool obtained again meets the optimization requirements, it is used as the final optimized primer pool.

2. The method for evaluating the dimerization of targeted PCR primer pools based on a sliding window according to claim 1, characterized in that, The preprocessing of the original sequences in the primer pool includes sequence normalization and verification, reverse complementary sequence generation, and structured database construction.

3. The method for evaluating the dimerization of targeted PCR primer pools based on a sliding window according to claim 1, characterized in that, The primer pair combination library contains primer combinations of sequence A and sequence B. The reverse complementary sequence of sequence B is compared with sequence A through a sliding window. During the sliding window comparison, sequence A is kept stationary while sequence B is controlled to move in 1bp increments along its 5' to 3' direction. During the sliding process, the sliding position number, overlapping region information, pairing results of each base position, total number of matching bases, number of GC matching bases, and number of mismatched bases are recorded to form the original database of the full primer pool sliding window comparison.

4. The method for evaluating the dimerization of targeted PCR primer pools based on a sliding window according to claim 3, characterized in that, After recording the sliding position number, overlapping region information, pairing results of each base position, total number of matching bases, number of GC matching bases and number of mismatched bases, the data is structured and integrated and associated with the corresponding alignment numbers to form the original database for the sliding window alignment of the full primer pool.

5. The method for evaluating the dimerization of targeted PCR primer pools based on a sliding window according to claim 1, characterized in that, Based on local alignment fit, the optimal result of the fault-tolerant alignment scheme at each sliding position is selected. The local alignment fit is determined by the product of the ratio of the number of matched bases after fault tolerance to the total number of bases in the overlapping region and the number of matched bases. The fault-tolerant alignment scheme with the highest local alignment fit is selected as the optimal alignment result after fault tolerance at the current sliding position. The optimal alignment result, overlapping region information, total number of matched bases after fault tolerance, number of GC matched bases, and number of mismatched bases at each sliding position are structurally integrated and associated with the corresponding alignment number and sliding position number to form the optimal alignment database after fault tolerance processing of the full primer pool.

6. The method for evaluating the dimerization of targeted PCR primer pools based on a sliding window according to claim 1, characterized in that, The dimerization score for each primer combination is calculated using the following formula: in, Dimerscore For a sliding position, the dimer score is... N matched This represents the total number of correctly paired bases in the optimal alignment result after tolerating the current sliding position. This represents the number of CG bases paired in the optimal alignment result after fault tolerance at the current sliding position. α For GC weighting coefficients, N mismatched This represents the number of mismatched bases in the optimal alignment result after fault tolerance at the current sliding position. For each primer pair, the dimerization score of all sliding positions is extracted, and the maximum value is selected as the final dimerization risk score of that primer pair.

7. The method for evaluating the dimerization of targeted PCR primer pools based on a sliding window according to claim 1, characterized in that, The CNN+BiLSTM model includes an input layer, a convolutional layer, a pooling layer, a bidirectional LSTM layer, a fully connected layer, and an output layer.

8. The method for evaluating the dimerization of targeted PCR primer pools based on a sliding window according to claim 1, characterized in that, After the iteration is completed, the preliminary optimized primer pool is output along with the intermediate primer pool optimization report. The intermediate primer pool optimization report records the primer numbers that were fine-tuned / removed / added, the reasons for the adjustment, and the risk improvement after the adjustment.

9. The method for evaluating the dimerization of targeted PCR primer pools based on a sliding window according to claim 1, characterized in that, After obtaining the final optimized primer pool, the sequence information, target site information, dimerization risk judgment results, and amplification efficiency prediction values ​​of the final optimized primer pool are integrated to construct the final optimized primer pool database and generate a final report on primer pool dimerization evaluation and optimization.

Images