CNV detection method integrating learning-driven PSO algorithm and neural network

Abstract

The invention provides a CNV detection method integrating a learning-driven PSO algorithm and a neural network, and relates to the technical field of genome data processing and analysis, and the method comprises the steps: inputting a sequencing sample and a reference genome, carrying out the preprocessing, extracting the features of each genome window, carrying out the normalization processing of the features, and generating a normalized feature vector; constructing a three-layer BP neural network model based on the normalized feature vector; and encoding the weight and the threshold of the BP neural network model as particle positions, randomly generating a group of particles as an initial solution space, selecting a particle updating direction from three speed updating modes by adopting a roulette strategy, and updating the particle speed based on a self-adaptive inertia weight adjustment strategy. According to the method, CNV efficient and accurate detection is realized by applying ensemble learning, a PSO algorithm and a neural network.

Description

Technical Field

[0001] The present invention relates to the technical field of genome data processing and analysis, and in particular to a CNV detection method that integrates a learning-driven PSO algorithm and a neural network. Background Art

[0002] In the biomedical field, copy number variation (CNV) detection is crucial for genetic disease research and clinical diagnosis and treatment. With the advancement of high-throughput sequencing technology, CNV detection methods are constantly evolving. Early traditional methods based on read depth and SNP arrays suffered from low sensitivity and poor specificity, making them inadequate for complex sample testing. While recent advances have been made in machine learning and optimization algorithms, such as CNV_IWOABP, they still suffer from high computational complexity and premature convergence, making them inadequate for rapid large-scale data detection and clinical precision diagnosis. Furthermore, existing methods also lack feature utilization and algorithm optimization, hindering the full mining of data information and efficient optimization. Therefore, there is an urgent need for a CNV detection method that can integrate features from multiple data sources and possess efficient global search capabilities. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a CNV detection method that uses integrated learning to drive PSO algorithm and neural network, and to achieve efficient and accurate CNV detection using integrated learning, PSO algorithm and neural network.

[0004] In order to solve the above technical problems, the technical solutions of the present invention are as follows: In a first aspect, a CNV detection method using an integrated learning-driven PSO algorithm and a neural network is provided, the method comprising: S1: Input sequencing samples and reference genomes, perform preprocessing, extract features of each genome window, normalize the features, and generate normalized feature vectors; S2: Based on the normalized feature vector, a three-layer BP neural network model is constructed; S3: The weights and thresholds of the BP neural network model are encoded as particle positions, a set of particles are randomly generated as the initial solution space, a roulette wheel strategy is used to select the particle update direction from three velocity update modes, and the particle velocity is updated based on the adaptive inertia weight adjustment strategy; S4: In the particle swarm update process, the selection and mutation mechanism of evolutionary game theory is introduced. The strategy probability distribution is dynamically updated by the imitator, the strategy selection probability is dynamically adjusted according to the particle fitness, and the upper limit of the strategy probability is set to 0.5 and the adaptive mutation probability is set to 0.1 to obtain the updated particle swarm; S5: Based on the updated particle swarm, a reinforcement learning algorithm is used to perform hierarchical optimization to generate the optimal weights and thresholds; S6: Substitute the optimal weights and thresholds into the BP neural network and train it using simulated and real data sets until the convergence conditions are met. Use the trained model to classify the CNV status of the target genomic data and output the detection results.

[0005] Furthermore, S1: inputs the sequencing sample and the reference genome, performs preprocessing, extracts the features of each genome window, normalizes the features, and generates a normalized feature vector, including: Input sequencing samples and reference genomes for preprocessing; Based on the preprocessed sequencing samples and reference genome, five features are extracted for each genomic window, including read depth, GC content, base quality, correlation between adjacent genomic sites, and correlation between the number of duplicate reads within the window; Normalize the features to generate normalized feature vectors.

[0006] Furthermore, S2: Based on the normalized feature vector, a three-layer BP neural network model is constructed, including: Based on the normalized feature vector, a three-layer BP neural network consisting of input layer, hidden layer and output layer is constructed, where the number of nodes in the input layer is 5, the number of nodes in the hidden layer is 7, and the number of nodes in the output layer is 4. Configure the input layer feature map and map the five nodes of the input layer to the RD, GC content, base quality, adjacent site correlation, and repeat read number correlation in the normalized feature vector to achieve standardized input of the feature vector. Set the hidden layer activation function, use the Sigmoid activation function in the hidden layer nodes, and perform nonlinear transformation on the input features; Design the classification probability output of the output layer, use the Softmax function in the output layer to normalize the signal output by the hidden layer, and generate the classification probability distribution of the four categories of CNV status; A complete three-layer BP neural network model is constructed through hierarchical structure, feature mapping and function configuration.

[0007] Furthermore, S3: the weights and thresholds of the BP neural network model are encoded into particle positions, a set of particles are randomly generated as the initial solution space, a roulette wheel strategy is used to select the particle update direction from three velocity update modes, and the particle velocity is updated based on the adaptive inertia weight adjustment strategy, including: The weights and thresholds of the BP neural network model are encoded as the position vectors of the particles, and an initial population of 50 particles is randomly generated. The particle fitness value is determined by calculating the absolute error between the expected output and the predicted output. The particle update direction is selected from three velocity update modes through a roulette wheel strategy; Based on the adaptive inertia weight adjustment strategy, the inertia weight is dynamically adjusted by linearly decreasing from the maximum value to the minimum value according to the formula based on the maximum number of iterations, the current number of iterations, and the maximum and minimum values of the inertia weight coefficient.

[0008] Furthermore, S4: In the particle swarm update process, the selection and mutation mechanism of evolutionary game theory is introduced. The strategy probability distribution is dynamically updated by the imitator, the strategy selection probability is dynamically adjusted according to the particle fitness, and the upper limit of the strategy probability is set to 0.5 and the adaptive mutation probability is set to 0.1. The updated particle swarm is obtained, including: The imitator dynamically updates the probability distribution of strategies, calculates the probability of each strategy being selected in the next iteration based on the fitness value of the particle, and normalizes the calculation results; Set the upper limit of strategy probability to 0.5. When the probability of a strategy exceeds the upper limit, the excess will be forcibly allocated to other strategies according to the returns of other strategies; Randomly perturb the particle position with an adaptive mutation probability of 0.1 to generate a new particle position; Based on the obtained new strategy probability distribution and new particle position, the speed and position of each particle in the particle swarm are updated to obtain the updated particle swarm.

[0009] Furthermore, S5: Based on the updated particle swarm, a reinforcement learning algorithm is used to perform hierarchical optimization to generate optimal weights and thresholds, including: In the local search phase, based on the updated particle swarm, an improved Q-learning algorithm is used to define five local perturbation operators as the action space. The perturbation action is dynamically selected according to the fitness change state, and the Q value is updated according to the reward function. In the global search stage, based on the SARSA algorithm, four crossover operators are defined as the action space. The global search path is optimized through the state-action sequence, and the global optimal weights and thresholds are generated by combining the Q value update rule.

[0010] Furthermore, S6: the optimal weights and thresholds are substituted into the BP neural network, and training is performed using simulated and real data sets until convergence conditions are met. The trained model is used to classify the CNV status of the target genomic data and output the test results, including: The optimal weights and thresholds were substituted into the BP neural network. Simulated datasets with coverage depths of 4x and 6x and tumor purity of 0.2-0.4, as well as the real 1000 Genomes Project dataset, were used for training. The mean squared error loss function and gradient descent optimizer were set, and the training was performed until either the error was less than 0.01 or 200 iterations were achieved. After training is completed, the normalized feature vector of the target genome data is forward propagated and calculated to output the classification probabilities of the four types of CNV states, and the final detection results are output based on the maximum probability.

[0011] Secondly, a CNV detection system integrating a learning-driven PSO algorithm and a neural network includes: The feature extraction module is used to input sequencing samples and reference genomes, perform preprocessing, extract features of each genome window, normalize the features, and generate normalized feature vectors; A construction module is used to construct a three-layer BP neural network model based on the normalized feature vector; The optimization module encodes the weights and thresholds of the BP neural network model into particle positions, randomly generates a set of particles as the initial solution space, uses a roulette wheel strategy to select the particle update direction from three speed update modes, and updates the particle speed based on the adaptive inertia weight adjustment strategy. During the particle swarm update process, the selection and mutation mechanism of evolutionary game theory is introduced. The strategy probability distribution is dynamically updated through the imitator, the strategy selection probability is dynamically adjusted according to the particle fitness, and the strategy probability upper limit is set to 0.5 and the adaptive mutation probability is set to 0.1 to obtain the updated particle swarm. Based on the updated particle swarm, the reinforcement learning algorithm is used for hierarchical optimization to generate the optimal weights and thresholds. The output module is used to substitute the optimal weights and thresholds into the BP neural network, use simulated and real data sets for training until the convergence conditions are met, use the trained model to classify the CNV status of the target genome data, and output the detection results.

[0012] According to a third aspect, a computing device includes: one or more processors; The storage device is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0013] In a fourth aspect, a computer-readable storage medium stores a program, which implements the method when executed by a processor.

[0014] The above solution of the present invention includes at least the following beneficial effects: By considering the interaction between bins and repeat reads and integrating multi-feature analysis data, this method accurately captures CNVs, reduces errors in complex sample detection, and provides a reliable basis for disease diagnosis. An improved PSO algorithm optimizes the weights and thresholds of the BP neural network. Incorporating evolutionary game theory, the strategy probability is adjusted based on particle fitness, avoiding local optima and increasing the probability of obtaining a global optimal solution. Based on reinforcement learning, local search updates the Q value by selecting perturbation actions based on fitness, while global search optimizes the path through state-action sequences, enhancing overall optimization and accelerating algorithm convergence. Setting an upper limit on strategy probability and an adaptive mutation probability balances overcoming local optima with maintaining diversity. The method exhibits low relative deviation across different sample tests and exhibits superior stability compared to other methods. Using both simulated and real-world datasets for training and testing, the method demonstrates excellent performance in sensitivity, accuracy, and F1 scores across a variety of scenarios. Computing equipment and computer-readable storage media provide hardware and software support for the detection method, ensuring repeatable and stable algorithm operation and facilitating its widespread application. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of a CNV detection method that integrates a learning-driven PSO algorithm and a neural network, provided by an embodiment of the present invention.

[0016] Figure 2 Schematic diagram of a CNV detection system that integrates a learning-driven PSO algorithm and a neural network, provided in an embodiment of the present invention.

[0017] Figure 3 It is a workflow diagram of CNV_ILPSOBP of the present invention.

[0018] Figure 4 It is the BP neural network topology diagram of CNV_ILPSOBP of the present invention.

[0019] Figure 5 It is the flow chart of the Q-learning algorithm of the present invention.

[0020] Figure 6 It is an action relationship diagram in Q-learning of the present invention.

[0021] Figure 7 It is a detailed flow chart of the four global search operators of the present invention.

[0022] Figure 8 It is the algorithm optimization network flow chart of the present invention.

[0023] Figure 9 The performance comparison of the ILPSOBP of the present invention and six other methods on the simulated data set is shown in FIG.

[0024] Figure 10 This is the ROC curve analysis of the seven methods of the present invention.

[0025] Figure 11 is the relative deviation of MFCNV and ILPSOBP of the present invention.

[0026] Figure 12 This is the boundary error analysis of the five methods of the present invention.

[0027] Figure 13 It is a radar chart analysis of the number of CNVs detected by the six methods of the present invention.

[0028] Figure 14 This is a performance comparison of the seven methods of the present invention on real samples. DETAILED DESCRIPTION

[0029] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0030] like Figure 1 As shown, an embodiment of the present invention proposes a CNV detection method that integrates a learning-driven PSO algorithm and a neural network, and the method includes the following steps: Step S1: Input the sequencing sample and reference genome, perform preprocessing, extract the features of each genome window, normalize the features, and generate a normalized feature vector; Step S2: construct a three-layer BP neural network model based on the normalized feature vector; Step S3: Encode the weights and thresholds of the BP neural network model into particle positions, randomly generate a set of particles as the initial solution space, use the roulette strategy to select the particle update direction from three velocity update modes, and update the particle velocity based on the adaptive inertia weight adjustment strategy; Step S4: In the particle swarm update process, the selection and mutation mechanism of evolutionary game theory is introduced. The strategy probability distribution is dynamically updated by the imitator, the strategy selection probability is dynamically adjusted according to the particle fitness, and the upper limit of the strategy probability is set to 0.5 and the adaptive mutation probability is set to 0.1 to obtain the updated particle swarm; Step S5: Based on the updated particle swarm, a reinforcement learning algorithm is used to perform hierarchical optimization to generate optimal weights and thresholds; Step S6: Substitute the optimal weights and thresholds into the BP neural network and train it using simulated and real data sets until the convergence conditions are met. Use the trained model to classify the CNV status of the target genome data and output the detection results.

[0031] In an embodiment of the present invention, sequencing samples and reference genomes are preprocessed, and genomic window features are extracted and normalized to generate normalized feature vectors, thereby improving data quality and enhancing model performance. A three-layer BP neural network model is constructed based on the normalized feature vectors. With its nonlinear mapping capability, it can mine the complex patterns of genomic data and is suitable for classification tasks. The weights and thresholds of the BP neural network model are encoded as particle positions, and an initial solution space is randomly generated. A roulette wheel strategy is used to select the particle update direction, and the particle velocity is updated based on an adaptive inertia weight adjustment strategy. The roulette wheel and adaptive weight strategies expand the search range, accelerate convergence, and optimize weights and thresholds. The selection and mutation mechanism of evolutionary game theory is introduced to dynamically adjust the strategy probability, balance the search capability, avoid local optimality, and improve the stability of the algorithm. Based on the updated particle swarm, a reinforcement learning algorithm is used for hierarchical optimization, and better weights and thresholds are selected based on environmental feedback to enhance model performance. Through simulation and real data training, the model is adapted to actual data and provides accurate results for genomic research and disease diagnosis.

[0032] In a preferred embodiment of the present invention, the above step S1: inputting the sequencing sample and the reference genome, performing preprocessing, extracting the features of each genome window, normalizing the features, and generating a normalized feature vector may include: Step S11, input sequencing samples and reference genomes for preprocessing; Step S12: Extract five features of each genomic window based on the preprocessed sequencing samples and the reference genome, including read depth, GC content, base quality, correlation between adjacent genomic sites, and correlation of the number of duplicate reads within the window; Step S13: normalize the features to generate a normalized feature vector.

[0033] In this embodiment of the present invention, by preprocessing the sequencing samples and reference genome, low-quality sequences, adapter contamination, and other interfering information in the raw data are effectively removed, thereby improving data purity. Five key features, including read depth (RD), GC content, base quality, correlation between adjacent genomic sites, and correlation between the number of duplicate reads within the window, are extracted from each genomic window to reflect genomic characteristics from multiple dimensions. The extracted features are normalized to eliminate differences in dimension and numerical range between different features and assign equal importance to each feature. The standardized data preprocessing and feature extraction process supports automated and batch processing, significantly reducing manual intervention and the time cost of manual processing.

[0034] In an embodiment of the present invention, the specific steps include: Step S11: Preprocess the sequencing sample data, such as removing low-quality reads and adapter sequences. Preprocess the reference genome data, such as by creating an index to facilitate subsequent alignment with the sequencing sample. Tools such as BWA and Bowtie2 can be used to create the reference genome index. The preprocessed sequencing sample data is aligned with the reference genome data. Alignment tools (such as BWA and Bowtie2) are used to align the sequencing reads to the reference genome, generating an alignment result file (e.g., in SAM or BAM format). The generated alignment result file is sorted and deduplicated. Tools such as Samtools are used to sort the BAM file and mark duplicate reads, removing duplicate reads that may have occurred due to PCR amplification, etc.

[0035] Step S12: Divide the pre-processed alignment result file into multiple windows according to the genomic position. The window size can be set according to the actual situation, for example, it can be set to an appropriate length such as 1000bp. For each genomic window: count the number of sequencing reads aligned to the reference genome in each window, and use this as the read depth feature; count the number of G and C bases in the window, and then calculate their proportion to the total number of bases in the window to obtain the GC content feature; perform statistical analysis on the base quality of each sequencing read in the window, for example, calculate the average base quality, etc., as the base quality feature; calculate the correlation between adjacent genomic sites, which can be obtained by calculating the correlation of the sequencing depth, base composition, etc. of adjacent sites.

[0036] In step S13, normalization is performed on each extracted feature, and the five normalized features are combined into a feature vector as follows: ,in, represents the eigenvector of the bin, Indicates the RD value, Represents the base quality, Represents the correlation between adjacent sites in the genome, Indicates the The GC content of each bin, Indicates the The number of repeated reads within a bin range is the normalized feature vector of each genomic window.

[0037] In a preferred embodiment of the present invention, the above step S2: constructing a three-layer BP neural network model based on the normalized feature vector may include: Step S21: Based on the normalized feature vector, a three-layer BP neural network is constructed, which includes an input layer, a hidden layer, and an output layer, wherein the number of nodes in the input layer is 5, the number of nodes in the hidden layer is 7, and the number of nodes in the output layer is 4; Step S22: Configure the input layer feature map, and make a one-to-one correspondence between the five nodes of the input layer and the RD, GC content, base quality, adjacent site correlation, and repeat read number correlation in the normalized feature vector, to achieve standardized input of the feature vector; Step S23, setting the hidden layer activation function, using the Sigmoid activation function in the hidden layer nodes to perform nonlinear transformation on the input features; Step S24, designing the classification probability output of the output layer, using the Softmax function in the output layer to normalize the signal output by the hidden layer, and generating the classification probability distribution of the four categories of CNV states; Step S25: construct a complete three-layer BP neural network model through hierarchical structure, feature mapping and function configuration.

[0038] In this embodiment of the present invention, the five nodes in the input layer precisely correspond to the five features, avoiding information redundancy; the seven nodes in the hidden layer balance fitting capability and computational efficiency; and the four nodes in the output layer directly correspond to the four CNV states, ensuring that the network output is consistent with actual classification requirements and that the predictions have clear biological significance. The Sigmoid function in the hidden layer introduces nonlinear mapping, enhancing the ability to capture complex patterns; and the Softmax function in the output layer generates a probability distribution that intuitively reflects the classification confidence. The one-to-one correspondence between features and nodes enables the network to independently learn the association between key features and CNV states, improving detection sensitivity and specificity.

[0039] In an embodiment of the present invention, the specific steps include: Step S21 explicitly constructs a three-layer BP neural network architecture with an input layer, a hidden layer, and an output layer. This structure effectively handles nonlinear relationships and is suitable for the complex data characteristics of CNV detection. Based on the number of features required for detection and the complexity of the problem, the number of input layer nodes is set to 5, corresponding to the five key features of CNV detection; the number of hidden layer nodes is set to 7, determined through multiple experiments and experience to achieve a good balance between network complexity and performance; and the number of output layer nodes is set to 4, corresponding to the four CNV states: normal, amplification, hemizygous deletion, and homologous deletion.

[0040] In step S22, the five nodes of the input layer are mapped one-to-one with the RD (read depth), GC content, base quality, correlation between adjacent sites in the genome, and correlation between bins and the number of duplicate reads in the normalized feature vector; after the input features are normalized, they are mapped to the corresponding nodes.

[0041] In step S23, a Sigmoid activation function is used in the hidden layer nodes. This function can map the input signal to a range between 0 and 1, perform nonlinear transformation on the input features, and enable the neural network to learn complex nonlinear relationships. The signal transmitted from the input layer is input to each node of the hidden layer, and the Sigmoid function is used to calculate and output the signal after nonlinear transformation.

[0042] In step S24, the Softmax function is used in the output layer to normalize the signal output by the hidden layer, and the output value is converted into a probability distribution with a sum of 1, so that each output node corresponds to the probability of a CNV state, which facilitates intuitive judgment of the possibility of the sample belonging to each CNV state; through the calculation of the Softmax function, the output layer generates a classification probability distribution of four types of CNV states (normal, amplification, hemizygous deletion and homologous deletion).

[0043] In step S25, the determined hierarchical structure (input layer, hidden layer, output layer), feature mapping relationship, activation function and output function configuration are integrated to construct a complete three-layer BP neural network model.

[0044] In a preferred embodiment of the present invention, the above step S3: encoding the weights and thresholds of the BP neural network model into particle positions, randomly generating a set of particles as an initial solution space, selecting a particle update direction from three velocity update modes using a roulette wheel strategy, and updating the particle velocity based on an adaptive inertia weight adjustment strategy, may include: Step S31: Encode the weights and thresholds of the BP neural network model into the position vector of the particle, randomly generate an initial population of 50 particles, and determine the particle fitness value by calculating the absolute error between the expected output and the predicted output; Step S32, selecting a particle update direction from three speed update modes using a roulette wheel strategy; In step S33, based on the adaptive inertia weight adjustment strategy, the inertia weight is dynamically adjusted by linearly decreasing from the maximum value to the minimum value according to the formula based on the maximum number of iterations, the current number of iterations, and the maximum and minimum values of the inertia weight coefficient.

[0045] In an embodiment of the present invention, the neural network weights and thresholds are encoded as particle positions, and 50 particles are randomly generated. This not only ensures the diversity of search starting points and prevents the algorithm from being trapped in the local optimum, but also calculates the fitness through absolute error, provides a quantitative basis for parameter evaluation, and accelerates the screening of high-quality solutions. A roulette strategy is used to choose between three speed update modes, allowing particles to flexibly switch between local search and global exploration based on their own conditions. In conjunction with an adaptive inertia weight adjustment strategy, the inertia weight is linearly reduced with the iteration process, balancing the algorithm's global search capability and local development capability, and enhancing convergence performance.

[0046] In an embodiment of the present invention, the specific steps include: Step S31: Arrange the weights and thresholds in the BP neural network model in a certain order and encode them as the position vector of the particle; randomly generate an initial population of 50 particles. During the generation process, ensure that the parameter values in the position vector of each particle are within a reasonable range. For example, the weights are usually between [-1, 1], and the thresholds can be set in an appropriate range according to the actual situation. For each particle in the initial population, calculate the fitness value as follows: ,in, is the fitness value, is the expected output, is the predicted output, is a coefficient used to scale or adjust the fitness value to suit the needs of a specific problem or the settings of the algorithm. is a related quantity such as the number of samples or output dimension. The smaller the absolute error, the better the prediction effect of the BP neural network parameter combination corresponding to the particle on the current training data, and the higher the fitness value.

[0047] In step S32, three particle velocity update modes are used: self-learning-only update, global learning-only update, and a combination of self-learning and global learning. These modes achieve different update effects by adjusting the influence of the particle's own historical best position and the global best position on the particle's current velocity. A roulette wheel strategy is used to select the update mode. A random number between 0 and 1 is generated for each particle. The position of the random number on the roulette wheel determines the update mode used for the particle in this iteration. Each mode corresponds to a different mathematical formula as follows: ,in, No. The particle in The speed of the iteration, and is the individual learning factor, and is the group learning factor, and these four factors are all non-negative constants. is the inertia weight, arrive are four random numbers in the interval (0, 1), For the The particle in The current position at the iteration, For the The optimal position of a particle, is the optimal position of all particles, is the current iteration number.

[0048] Step S33: Determine the maximum number of iterations, the maximum and minimum values of the inertia weight coefficient. These parameters are set based on experimental requirements and experience. During the algorithm iteration process, the number of iterations currently in progress is recorded. Starting from the first iteration, the number of iterations is increased by 1 each time an iteration is completed until the maximum number of iterations is reached, at which point the algorithm stops. The formula for the adaptive adjustment process is as follows: ,in, is the inertia weight, is the maximum number of iterations, and are the maximum and minimum values of the inertia weight coefficient, is the current iteration number.

[0049] In a preferred embodiment of the present invention, the above step S4: in the particle swarm update process, the selection and mutation mechanism of evolutionary game theory is introduced, the strategy probability distribution is dynamically updated by the imitator, the strategy selection probability is dynamically adjusted according to the particle fitness, and the strategy probability upper limit is set to 0.5 and the adaptive mutation probability is set to 0.1, and the updated particle swarm is obtained, which may include: Step S41: Dynamically update the strategy probability distribution through the imitator, calculate the probability of each strategy being selected in the next iteration according to the fitness value of the particle, and normalize the calculation results to ensure that the sum of the probabilities is 1; Step S42: Set the upper limit of the strategy probability to 0.5. When the probability of a strategy exceeds the upper limit, the excess will be forcibly allocated to other strategies according to the benefits of other strategies; Step S43, randomly perturbing the particle position with an adaptive mutation probability of 0.1 to generate a new particle position; Step S44 : Based on the obtained new strategy probability distribution and the new particle position, the speed and position of each particle in the particle swarm are updated to obtain an updated particle swarm.

[0050] In an embodiment of the present invention, based on the dynamic update of strategy probability by the imitator, particles select strategies according to fitness, and strategies with high fitness are adopted more frequently, forming an evolutionary trend of "survival of the fittest", driving the population to quickly gather in the high-quality solution area, and updating and normalizing by combining the current strategy probability, relative benefit advantage and learning rate to ensure the stability and effectiveness of the probability system, achieve a balance between global exploration and local development, improve search efficiency, optimize BP neural network parameters, and improve CNV detection accuracy; set the strategy probability upper limit to 0.5 to limit the dominance of a single strategy, redistribute the excess part, retain the algorithm search potential, and ensure the reliability of CNV detection results; introduce a 0.1 adaptive mutation probability to perturb the particle position, break the local optimum, expand the search range, optimize the BP neural network performance, and improve the detection accuracy of various CNV samples.

[0051] In an embodiment of the present invention, the specific steps include: Step S41: After each iteration, the reward vector of each strategy is calculated based on the fitness value of each particle after adopting different strategies. At the same time, the average reward of all strategies in this iteration is calculated as a benchmark for measuring the advantage of each strategy. The specific process formula for updating the strategy distribution is as follows based on the current probability of each strategy, its advantage relative to the average reward, and the set learning rate: ,in, For the The strategy in the iteration The probability of being selected, For the The strategy in the iteration The probability of being selected, is the learning rate, used to control the evolution speed, For strategy In the The reward vector in the iteration, For the The average return of all strategies in iterations, is the number of strategies, which is set to 3 in this paper. For the The strategy in the iteration The probability of being selected, For strategy In the For the profit vector in the iteration, the selection probability of the strategy with high fitness and large profit advantage will be improved after the update; conversely, the selection probability of the strategy with poor performance will be reduced, thereby simulating the evolutionary process of "survival of the fittest"; the updated strategy selection probability is normalized to ensure that the sum of the probabilities of all strategies is 1.

[0052] Step S42: After each update of the strategy probability distribution, the probability value of each strategy is monitored in real time, and the probability of each strategy is compared with the set upper limit value of 0.5 to determine whether there is a situation of exceeding the upper limit; once it is found that the probability of a strategy exceeds 0.5, the forced allocation mechanism is immediately activated; the part exceeding 0.5 is redistributed to the remaining strategies according to the current benefits or probability ratios of other strategies.

[0053] In step S43, during each iteration, a random number between 0 and 1 is generated for each particle in the population, and the random number is compared with the set adaptive mutation probability of 0.1. If the random number is less than 0.1, the particle is selected as the mutation object; otherwise, no mutation operation is performed.

[0054] For particles selected for mutation, their position vectors are randomly perturbed. Specifically, a random value that obeys a specific distribution (such as uniform distribution or Gaussian distribution) is added to each dimension of the position vector to change the position of the particle in the solution space. This random perturbation gives the particle the opportunity to jump out of the current local optimal area and explore new areas of the solution space.

[0055] Step S44 : Based on the obtained new strategy probability distribution and the new particle position, the speed and position of each particle in the particle swarm are updated to obtain an updated particle swarm.

[0056] In a preferred embodiment of the present invention, the above step S5: performing hierarchical optimization based on the updated particle swarm using a reinforcement learning algorithm to generate optimal weights and thresholds may include: Step S51, local search phase, based on the updated particle swarm, adopts the improved Q-learning algorithm, defines five local perturbation operators as the action space, dynamically selects perturbation actions according to the fitness change state, and updates the Q value according to the reward function; Step S52, the global search phase, defines four crossover operators as the action space based on the SARSA algorithm, optimizes the global search path through the state-action sequence, and generates the global optimal weights and thresholds in combination with the Q value update rule.

[0057] In this embodiment of the present invention, by designing five perturbation operators, such as exchange and insertion, the limitations of a single operator are overcome and the diversity of local space search is enhanced. Perturbation actions are dynamically selected based on fitness, and the Q value is updated through a reward function to form an adaptive learning mechanism. The local search strategy is iteratively optimized to improve the convergence speed and solution quality. Four crossover operators are introduced to construct the action space, accelerating the discovery of global high-quality solutions and avoiding falling into local optimality. The search path is optimized through state-action sequences, and the strategy is dynamically adjusted in combination with Q value updates. Local and global searches work together to generate optimal weights and threshold parameters, improving neural network performance and enhancing the accuracy and reliability of CNV detection.

[0058] In an embodiment of the present invention, the specific steps include: Step S51, based on the updated particle swarm, determine the initial state of the improved Q-learning algorithm, set the Q table, where the Q value represents a certain action taken in a certain state (i.e., local perturbation operator); define five local perturbation operators as the action space, which are: the exchange operator is to randomly select two different positions in the solution and ,Will and The elements at the position are exchanged; the insertion operator randomly selects two different positions in a solution and ( ), and The element at the position is inserted into The inversion operator randomly selects two different positions in the solution. and , and and The block-based exchange operator first randomly selects three consecutive positions in a solution to form a subsequence, and then randomly selects two different positions in the solution to form a subsequence. and ( ) as the starting position of the two subsequences, and exchange the two subsequences as a whole; the block-based crossover operator is the same as the block-based exchange operator in that its subsequence selection method is the same as the block-based exchange operator, except that Insert the subsequence at the starting position into The subsequence preceding the starting position.

[0059] In ILPSOBP, a state is defined as the change of fitness value over time. To describe this state, the fitness value metric is defined as: ,in, is the fitness value at the previous moment, is the fitness value at the current moment, It is used to measure the change of fitness value over time. The smaller it is, the better the fitness value is. On the contrary, The larger it is, the worse the fitness value is. ; These states are based on Classified: Status 1: , State 2: , State 3: .

[0060] In addition, the agent will receive positive or negative rewards after performing an action, which can be defined as follows: ,in The reward value obtained by the agent after performing the action, which is a change in the fitness value To determine the discrete value, To measure the change of fitness value over time, >0 means the fitness value becomes better, ≤0 means that the fitness value has not improved (including deterioration and no change).

[0061] Traditional Q-learning generally uses a greedy strategy to select the action in each state, but this strategy is extreme and prone to fall into local optimality. - Greedy strategy is used for optimization, and a method is proposed to dynamically adjust the exploration factor according to the learning progress. This method can avoid the algorithm from falling into invalid local search and improve the convergence speed and search efficiency of the algorithm to a certain extent. The formula is as follows: ,in, is the preset exploration factor parameter, is the preset iteration threshold, is the number of iterations, that is, the number of iterations the algorithm has executed.

[0062] The update rules according to Q value are as follows: ,in, is the learning rate used to control the degree of learning, The two parameters are both between (0, 1) and are used to reduce the discount factor of future feedback. When the value is relatively large, the algorithm tends to converge faster. is the reward value obtained by the agent after performing the action, For the current state The following actions are taken, and is the current state and the next state, For Actions taken in In state Take action The Q value when For the execution strategy selected After the action, in the state The expected Q value under .

[0063] In step S52, based on the updated particle swarm, the initial state of the SARSA algorithm is determined, and a Q table is set, where the Q value represents the long-term cumulative reward that can be obtained by taking a certain action (i.e., the global search operator) in a certain state.

[0064] Four global search operators are defined as action spaces, namely: (Partially mapped crossover): Select a pair of chromosomes ( and ), randomly select two split points in the chromosome and (the start and end positions of several genes); exchange and between genes and produce two daughter chromosomes ( and ); Based on the mapping relationship established between the two groups of exchanged genes, conflict detection is performed on the unexchanged genes until and All genes are free of conflict.

[0065] (Sequential cross): In and Randomly select the starting and end position ,Will The genes in this region are copied At the same position, General The missing genes in the other offspring are filled in according to the order; Obtained in the same way.

[0066] (Cyclic Cross): In Randomly select a position on the The gene number at the corresponding position, then return to Find the position of the gene with the same number and repeat the previous work until a ring is formed. The positions of all genes in the ring are the last selected positions. Selected genes generated , and ensure the position corresponds, and finally The remaining genes were placed Another child Obtained in the same way.

[0067] (Position-based crossover): and Randomly select several locations in 、 、 ), the position can be discontinuous. The genes at these locations are copied At the same position of General The missing genes in the other offspring are filled in according to the order; Obtained in the same way.

[0068] During the interaction between the agent and the environment, the Q value is updated according to the sequence of state-action-reward-state-action. Specifically, in each iteration, the algorithm updates the Q value according to the formula ,in, is the learning rate used to control the degree of learning, is the discount factor used to reduce future feedback, is the reward value obtained by the agent after performing the action, For the current state The following actions are taken, and is the current state and the next state, For Actions taken in In state Take action The Q value when For the execution strategy selected After the action, in the state The expected Q value under generates the Q table corresponding to the state and action.

[0069] Call a random function to generate a random number between 0 and 1 and compare it with the exploration factor To compare; if , the agent will randomly choose an action; if , the intelligent experience obtains the action with the maximum reward value in this state according to the Q table; this process can be expressed as Description, where is the policy function, is a parameter between 0 and 1, is a random value between 0 and 1, In state Take action The maximum Q value when The process of randomly choosing an action during exploration.

[0070] The same reward function as Q-learning is used to evaluate the quality of actions. During the global search phase, the algorithm can quickly find high-quality solutions within a larger search range, accelerate the convergence of the algorithm, and ultimately generate the globally optimal weights and thresholds.

[0071] In a preferred embodiment of the present invention, step S6: substituting the optimal weights and thresholds into the BP neural network, training using simulated and real data sets until convergence conditions are met, using the trained model to classify the CNV status of the target genome data, and outputting the detection results, may include: Step S61: Substitute the optimal weights and thresholds into the BP neural network and train it using simulated datasets with coverage depths of 4x and 6x and tumor purity of 0.2-0.4, as well as the real dataset from the 1000 Genomes Project. Set the mean square error loss function and gradient descent optimizer, and iterate until either the error is less than 0.01 or 200 iterations are reached. Step S62: After the training is completed, the normalized feature vector of the target genome data is forward propagated and calculated, and the classification probabilities of the four types of CNV states are output, and the final detection result is output according to the maximum probability.

[0072] In an embodiment of the present invention, a BP neural network is trained by combining simulated and real data sets with different coverage depths and tumor purities, taking into account both model performance testing and actual application needs to improve generalization capabilities; using mean square error as the loss function and gradient descent as the optimizer, reasonable iteration termination conditions are set to ensure training accuracy and efficiency and avoid overfitting; the trained model forward propagates the normalized feature vector, determines the detection result based on the maximum value of the classification probability of the four types of CNV status, and utilizes the nonlinear mapping capability of the BP neural network to achieve accurate classification, providing a reliable basis for disease diagnosis and treatment.

[0073] In an embodiment of the present invention, the specific steps include: In step S61, the optimal weights and thresholds obtained using an improved particle swarm optimization (ILPSO) algorithm are substituted into a three-layer BP neural network. Simulated datasets with coverage depths of 4x and 6x, and tumor purities ranging from 0.2 to 0.4, are selected, along with a real dataset from the 1000 Genomes Project. The mean squared error (MSE) is set as the loss function to measure the difference between the network's predicted output and the true labels. A gradient descent optimizer is used to continuously adjust the weights and thresholds to minimize the loss function by calculating the gradient of the loss function with respect to the network parameters. Training is initiated, allowing the BP neural network to process the input data, calculate the error, and backpropagate parameter updates. Training continues iteratively until one of the following two termination criteria is met: First, the training error falls below 0.01, indicating that the network has achieved high prediction accuracy; second, a preset number of 200 iterations is reached to avoid overtraining or prolonged computation. During training, the error changes and iteration progress are monitored in real time.

[0074] In step S62, five features, including RD and GC content, are extracted from the target genomic data and normalized to form a normalized feature vector. This vector is then input into the trained BP neural network, ensuring that the data format matches the network input requirements. In the BP neural network, the normalized feature vector is passed from the input layer to the hidden layer, where it is calculated using the activation functions of the neurons in each layer, ultimately outputting the classification probabilities of the four CNV states at the output layer. This process, based on the optimal weights and thresholds obtained through training, achieves feature mapping and classification prediction for the input data. The four CNV state classification probabilities obtained at the output layer are compared, and the category with the highest probability value is selected as the final CNV state detection result.

[0075] like Figure 2As shown, an embodiment of the present invention further provides a CNV detection system 20 that integrates a learning-driven PSO algorithm and a neural network, comprising: The feature extraction module 21 is used to input the sequencing sample and the reference genome, perform preprocessing, extract the features of each genome window, normalize the features, and generate a normalized feature vector; A construction module 22 is used to construct a three-layer BP neural network model based on the normalized feature vector; The optimization module 23 is used to encode the weights and thresholds of the BP neural network model into particle positions, randomly generate a group of particles as the initial solution space, use a roulette wheel strategy to select the particle update direction from three speed update modes, and update the particle speed based on the adaptive inertia weight adjustment strategy; during the particle swarm update process, the selection and mutation mechanism of evolutionary game theory is introduced, the strategy probability distribution is dynamically updated through the imitator, the strategy selection probability is dynamically adjusted according to the particle fitness, and the strategy probability upper limit is set to 0.5 and the adaptive mutation probability is set to 0.1 to obtain the updated particle swarm; based on the updated particle swarm, a reinforcement learning algorithm is used to perform hierarchical optimization to generate the optimal weights and thresholds; The output module 24 is used to substitute the optimal weights and thresholds into the BP neural network, use simulated and real data sets for training until the convergence conditions are met, use the trained model to classify the CNV status of the target genome data, and output the detection results.

[0076] In a specific embodiment of the present invention, the initial parameter configuration of the BP neural network of the CNV-ILPSOBP method is consistent with that of the CNV-IWOABP method. The remaining parameters are as follows:

[0077] Prior to clinical trials, the reliability and validity of ILPSOBP were verified using a simulated dataset. To mitigate the effects of external factors such as cell contamination, samples with varying purity and coverage depths were used as training data. To analyze samples with low purity and low tumor coverage, simulated data were generated using the IntSIM software. During the simulations, coverage depths were set to 4x and 6x, and tumor purity was set to 0.2, 0.3, and 0.4, respectively. Fifty samples were generated for each configuration, for a total dataset of 300 samples. ILPSOBP was then compared with six existing detection methods (GROM_RD, CNVnator, FREEC, MFCNV, CNV_IFTV, and CNVABNN), using sensitivity, precision, and F1 score as reference metrics. Figure 9 The comparative results of these seven methods under different configurations are presented, from which the detection performance and advantages of ILPSOBP under low purity and low coverage conditions can be intuitively evaluated.

[0078] exist Figure 9 In the figure, the horizontal axis represents accuracy and the vertical axis represents sensitivity. The closer to the upper right corner of the image, the higher the F1 score and the better the algorithm performance. Figure 9 It can be seen that with the increase of coverage, the performance of all algorithms has improved to varying degrees, especially in terms of accuracy. For example, when the tumor purity is 0.3, the accuracy of the FREEC method is significantly improved from 0.62 at a coverage of 4x to 0.94 at a coverage of 6x. Figure 9 The results show that ILPSOBP's F1 score is not consistently the highest across all six sample types; ILPSOBP achieves the highest F1 score at 4x coverage, while CNVABNN performs best at 6x coverage. However, other methods also demonstrate advantages in certain scenarios. For example, at a tumor purity of 0.4 and a coverage of 6x, GROM_RD is slightly more accurate than ILPSOBP. However, in this setting, GROM_RD's sensitivity is only 0.32 and its F1 score is only 0.48, while ILPSOBP's F1 score reaches 0.92. Similarly, CNVnator achieves near 100% accuracy at a tumor purity of 0.2 and a coverage of 6x, but its sensitivity is less than 0.4. CNV_IFTV achieves the highest accuracy at a tumor purity of 0.3 and a coverage of 6x, but its sensitivity is only 0.6. In summary, ILPSOBP is an effective and reliable method for CNV detection.

[0079] In addition, the ROC curve can reflect the changing trends of the true positive rate (TPR) and the false positive rate (FPR). When the ROC curve is closer to the upper left corner, it means that the algorithm performs better in detection. Figure 10 The ROC curves of seven methods at three different tumor purity levels are shown. Each curve represents the average TPR and FPR of the corresponding method under different conditions. The horizontal axis represents FPR, and the smaller the value, the lower the probability of false detection of the algorithm; the vertical axis represents TPR, and the higher the value, the more accurately the algorithm can detect positive samples; under various configurations, ILPSOBP achieved the highest TPR and lower FPR, although in some cases CNV_IFTV's FPR was slightly better than ILPSOBP. In contrast, the TPR of the remaining detection methods is generally low, and as the tumor purity decreases, the TPR value shows a downward trend. Overall, the ROC curve of ILPSOBP is at Figure 10 The upper left side shows that ILPSOBP still has satisfactory detection capability under various test conditions.

[0080] We also analyzed the stability of ILPSOBP and a single detection method. For ILPSOBP and MFCNV, when the tumor purity was 0.2, 0.3, and 0.4 and the BP neural network structure was the same, we used the average F1-score of 50 samples as a reference and compared the deviation of the F1-score in 50 samples. The test results are shown in Figure 2. Figure 11 As shown, the relative deviation is calculated as follows: ,in, Indicates the The relative deviation of the samples, Indicates the The F1 score of samples, Indicates the average F1 score of 50 samples. In the 50 consecutive tests under three tumor purities, 50 samples were used sequentially to expand the scope of the test; at the same time, using different samples each time also ensures the randomness of the test, making the test results convincing; Figure 11 In the figure, the horizontal axis represents the number of tests (i.e., the number of samples), and the vertical axis represents the error range of the F1 score. The results show that when the tumor purity is 0.2, the relative deviation of ILPSOBP remains within the range of [-5%, 5%], with a relatively stable curve and minimal fluctuation. In contrast, the relative deviation curve of MFCNV varies dramatically, exhibiting large fluctuations, even with samples exhibiting deviations approaching -10%. As the tumor purity increases to 0.3, the relative deviation of ILPSOBP becomes concentrated in the range of [-10%, 5%], and the curve becomes smoother. In contrast, the deviation range of MFCNV expands to [-20%, 5%], with multiple obvious local extremes in the region below -10%. Furthermore, when the tumor purity is 0.4, the relative deviation of ILPSOBP is limited to [-8%, 1%], with almost no significant fluctuations. The deviation range of MFCNV is [-15%, 2%]. Although the curve fluctuation amplitude has decreased, significant extreme values still exist locally. As the sample coverage depth increases, the fluctuation amplitude of MFCNV gradually decreases, and the error curve of ILPSOBP remains below MFCNV. This shows that ILPSOBP has higher stability and robustness than MFCNV, thus ensuring the accuracy of overall CNV prediction. However, ILPSOBP also has some samples with relatively large deviations. For example, at a coverage depth of 0.4, the 9th and 39th tests produced deviations close to -8%. This may be due to the presence of insignificant CNVs in the sample, but under these conditions, ILPSOBP still exhibits good stability.

[0081] Detecting boundary deviations is also a measure of algorithm effectiveness. Accurately detecting CNV boundaries helps pinpoint the CNV mutation site, thus providing a basis for the selection and use of targeted drugs. Generally speaking, greater boundary deviations indicate poorer algorithm detection capabilities. Figure 12 Comparisons of the margin deviations of ILPSOBP and four other algorithms are presented. The results show that ILPSOBP achieves the lowest margin deviations in all four sample configurations, while CNVABNN achieves the lowest margin deviations in two sample configurations. This suggests that optimizing the network with ILPSO can effectively reduce the margin deviations of CNV detection. However, in some cases, ILPSOBP may also exhibit high outlier values; for example, in a scenario with a tumor purity of 0.3 and a coverage of 6x, the outlier value reaches 0.36. This phenomenon may be due to deviations in the CNV distribution in some simulated samples, where the other algorithms also exhibit high margin deviations.

[0082] To apply ILPSOBP to clinical diagnosis, this study used a real-world dataset to validate the method. Similar to the real-world sample experiments used to validate the reliability of CNV_IWAOBP, the data included three real samples (NA19238, NA19239, and NA19240) from the CEU family in the public database 1000 Genomes Project (http: / / www.internationalgenome.org / ). For each sample, ILPSOBP was used in conjunction with five other methods to detect CNVs on chromosome 21. The results are shown in Figure 2. Figure 13 As shown; in the radar chart visualization, ILPSOBP has the largest coverage area, indicating that it detects the largest number of CNVs compared with other algorithms, followed by ReadDepth; while the point of GROM_RD is located closest to the center of the radar chart, indicating that its detection effect is the worst.

[0083] Figure 14 The performance comparison results of seven detection methods using real-world datasets are shown. As can be seen from the figure, ILPSOBP achieves the position closest to the upper right corner in all three sets of real-world samples, indicating that it achieves the highest F1-score across all samples, followed by CNVABNN and MFCNV. Although MFCNV has the highest accuracy across the three samples, its sensitivity is relatively low. In contrast, ILPSOBP achieves the highest sensitivity across all samples and is second only to MFCNV in accuracy on the NA19238 sample. In summary, ILPSOBP strikes a good balance between accuracy and sensitivity, and also exhibits good detection performance on real-world samples.

[0084] 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 within the scope of protection of the present invention.

Claims

1. A CNV detection method that integrates learning-driven PSO algorithm and neural network, characterized in that: The method comprises: S1: Input sequencing samples and reference genomes, perform preprocessing, extract features of each genome window, normalize the features, and generate normalized feature vectors; S2: Based on the normalized feature vector, a three-layer BP neural network model is constructed; S3: The weights and thresholds of the BP neural network model are encoded as particle positions, a set of particles are randomly generated as the initial solution space, a roulette wheel strategy is used to select the particle update direction from three velocity update modes, and the particle velocity is updated based on the adaptive inertia weight adjustment strategy; S4: In the particle swarm update process, the selection and mutation mechanism of evolutionary game theory is introduced. The strategy probability distribution is dynamically updated by the imitator, the strategy selection probability is dynamically adjusted according to the particle fitness, and the upper limit of the strategy probability is set to 0.5 and the adaptive mutation probability is set to 0.1 to obtain the updated particle swarm; S5: Based on the updated particle swarm, a reinforcement learning algorithm is used to perform hierarchical optimization to generate the optimal weights and thresholds; S6: Substitute the optimal weights and thresholds into the BP neural network and train it using simulated and real data sets until the convergence conditions are met. Use the trained model to classify the CNV status of the target genomic data and output the detection results.

2. The CNV detection method of integrated learning driven PSO algorithm and neural network according to claim 1, characterized in that: S1: Input sequencing samples and reference genomes, perform preprocessing, extract features of each genomic window, normalize the features, and generate a normalized feature vector, including: Input sequencing samples and reference genomes for preprocessing; Based on the preprocessed sequencing samples and reference genome, five features are extracted for each genomic window, including read depth, GC content, base quality, correlation between adjacent genomic sites, and correlation between the number of duplicate reads within the window; Normalize the features to generate normalized feature vectors.

3. The CNV detection method of integrated learning driven PSO algorithm and neural network according to claim 2, characterized in that: S2: Based on the normalized feature vector, a three-layer BP neural network model is constructed, including: Based on the normalized feature vector, a three-layer BP neural network consisting of input layer, hidden layer and output layer is constructed, where the number of nodes in the input layer is 5, the number of nodes in the hidden layer is 7, and the number of nodes in the output layer is 4. Configure the input layer feature map and map the five nodes of the input layer to the RD, GC content, base quality, adjacent site correlation, and repeat read number correlation in the normalized feature vector to achieve standardized input of the feature vector. Set the hidden layer activation function, use the Sigmoid activation function in the hidden layer nodes, and perform nonlinear transformation on the input features; Design the classification probability output of the output layer, use the Softmax function in the output layer to normalize the signal output by the hidden layer, and generate the classification probability distribution of the four categories of CNV status; A complete three-layer BP neural network model is constructed through hierarchical structure, feature mapping and function configuration.

4. The CNV detection method of integrated learning driven PSO algorithm and neural network according to claim 3, characterized in that: S3: Encode the weights and thresholds of the BP neural network model into particle positions, randomly generate a set of particles as the initial solution space, use a roulette wheel strategy to select the particle update direction from three velocity update modes, and update the particle velocity based on the adaptive inertia weight adjustment strategy, including: The weights and thresholds of the BP neural network model are encoded as the position vectors of the particles, and an initial population of 50 particles is randomly generated. The particle fitness value is determined by calculating the absolute error between the expected output and the predicted output. The particle update direction is selected from three velocity update modes through a roulette wheel strategy; Based on the adaptive inertia weight adjustment strategy, the inertia weight is dynamically adjusted by linearly decreasing from the maximum value to the minimum value according to the formula based on the maximum number of iterations, the current number of iterations, and the maximum and minimum values of the inertia weight coefficient.

5. The CNV detection method of integrated learning driven PSO algorithm and neural network according to claim 4, characterized in that: S4: In the particle swarm update process, the selection and mutation mechanism of evolutionary game theory is introduced. The strategy probability distribution is dynamically updated by the imitator, the strategy selection probability is dynamically adjusted according to the particle fitness, and the upper limit of the strategy probability is set to 0.5 and the adaptive mutation probability is set to 0.

1. The updated particle swarm is obtained, including: The imitator dynamically updates the probability distribution of strategies, calculates the probability of each strategy being selected in the next iteration based on the fitness value of the particle, and normalizes the calculation results; Set the upper limit of strategy probability to 0.

5. When the probability of a strategy exceeds the upper limit, the excess will be forcibly allocated to other strategies according to the returns of other strategies; Randomly perturb the particle position with an adaptive mutation probability of 0.1 to generate a new particle position; Based on the obtained new strategy probability distribution and new particle position, the speed and position of each particle in the particle swarm are updated to obtain the updated particle swarm.

6. The CNV detection method of integrated learning driven PSO algorithm and neural network according to claim 5, characterized in that: S5: Based on the updated particle swarm, a reinforcement learning algorithm is used to perform hierarchical optimization to generate optimal weights and thresholds, including: In the local search phase, based on the updated particle swarm, an improved Q-learning algorithm is used to define five local perturbation operators as the action space. The perturbation action is dynamically selected according to the fitness change state, and the Q value is updated according to the reward function. In the global search stage, based on the SARSA algorithm, four crossover operators are defined as the action space. The global search path is optimized through the state-action sequence, and the global optimal weights and thresholds are generated by combining the Q value update rule.

7. The CNV detection method of integrated learning driven PSO algorithm and neural network according to claim 6, characterized in that: S6: Substitute the optimal weights and thresholds into the BP neural network and train it using simulated and real data sets until convergence conditions are met. Use the trained model to classify the CNV status of the target genomic data and output the test results, including: The optimal weights and thresholds were substituted into the BP neural network. Simulated datasets with coverage depths of 4x and 6x and tumor purity of 0.2-0.4, as well as the real 1000 Genomes Project dataset, were used for training. The mean squared error loss function and gradient descent optimizer were set, and the training was performed until either the error was less than 0.01 or 200 iterations were achieved. After training is completed, the normalized feature vector of the target genome data is forward propagated and calculated to output the classification probabilities of the four types of CNV states, and the final detection results are output based on the maximum probability.

8. A CNV detection system integrating a learning-driven PSO algorithm and a neural network, the system implementing the method according to any one of claims 1 to 7, characterized in that: include: The feature extraction module is used to input sequencing samples and reference genomes, perform preprocessing, extract features of each genome window, normalize the features, and generate normalized feature vectors; A construction module is used to construct a three-layer BP neural network model based on the normalized feature vector; The optimization module encodes the weights and thresholds of the BP neural network model into particle positions, randomly generates a set of particles as the initial solution space, uses a roulette wheel strategy to select the particle update direction from three speed update modes, and updates the particle speed based on the adaptive inertia weight adjustment strategy. During the particle swarm update process, the selection and mutation mechanism of evolutionary game theory is introduced. The strategy probability distribution is dynamically updated through the imitator, the strategy selection probability is dynamically adjusted according to the particle fitness, and the strategy probability upper limit is set to 0.5 and the adaptive mutation probability is set to 0.1 to obtain the updated particle swarm. Based on the updated particle swarm, the reinforcement learning algorithm is used for hierarchical optimization to generate the optimal weights and thresholds. The output module is used to substitute the optimal weights and thresholds into the BP neural network, use simulated and real data sets for training until the convergence conditions are met, use the trained model to classify the CNV status of the target genome data, and output the detection results.

9. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a program, which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images