Hypertension medication gene polymorphism melting curve detection decision system

Abstract

This invention relates to the field of medical electro-digital data processing, and discloses a decision-making system for detecting melting curves of gene polymorphisms in hypertension medication. It collects fluorescence melting curve data of multiple gene targets; maps the melting temperature-fluorescence intensity derivative curve as graph structure nodes, and maps the melting feature associations between different gene loci as edges, constructing a multidimensional melting curve topology graph; extracts the co-fluctuation features of node peak temperature offset and edge weights and inputs them into a graph convolutional network, outputting a polymorphic combination feature vector; maps the polymorphic combination feature vector to a hypertension drug metabolism network graph model, calculates the joint decay coefficient of drug metabolic enzyme activity and receptor binding rate, and determines the type and dosage adjustment parameters for hypertension medication based on the joint decay coefficient. This invention overcomes the deficiency of single-peak independent alignment in separating overlapping peaks, achieves quantitative analysis of co-fluctuation associations of multi-gene polymorphisms, and eliminates medication bias caused by isolated qualitative matching.

Description

Technical Field

[0001] This invention relates to the field of medical electronic digital data processing, and discloses a decision system for detecting melting curves of gene polymorphisms in hypertension medications. Background Technology

[0002] In existing technologies, decision-making systems for detecting gene polymorphisms in hypertension medication typically process the fluorescence melting curves of multiple gene targets independently, extracting the melting temperature peak or derivative extreme value of a single gene locus. These values ​​are then compared with preset reference thresholds to qualitatively determine the wild-type or mutant status of each gene. Subsequently, based on the independent qualitative results of each gene, the corresponding hypertension drug type and dosage adjustment rules are matched one by one in a preset medication rule table. However, existing methods of independently extracting single-gene melting curve features for qualitative interpretation cannot analyze the synergistic association of melting features between multiple gene loci, leading to inaccurate medication decisions. Summary of the Invention

[0003] The purpose of this invention is to provide a decision system for detecting melting curves of gene polymorphisms in hypertension medications, which can solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A decision system for detecting and analyzing melting curves of hypertension drug gene polymorphisms includes: a data acquisition terminal that acquires fluorescence melting curve data of multiple gene targets related to hypertension; a graph construction processor that maps the melting temperature-fluorescence intensity derivative curve of each gene locus to nodes in a graph structure, and maps the melting feature associations between different gene loci to edges, constructing a multidimensional melting curve topology graph; a feature extraction processor that extracts the co-fluctuation features of peak temperature offset and edge weights of nodes in the multidimensional melting curve topology graph, inputs the co-fluctuation features into a graph convolutional network, and outputs a polymorphic combination feature vector; and a decision processor that maps the polymorphic combination feature vector to a hypertension drug metabolism network graph model, calculates the joint decay coefficient of drug metabolic enzyme activity and receptor binding rate, and determines the type and dosage adjustment parameters of hypertension drugs based on the joint decay coefficient, thus constituting the decision system for detecting and analyzing melting curves of hypertension drug gene polymorphisms.

[0005] Preferably, the graph construction processor maps the melting temperature-fluorescence intensity derivative curve of each gene locus to nodes of a graph structure, and maps the melting feature associations between different gene loci to edges, constructing a multidimensional melting curve topology graph, including: setting a sliding window along the temperature axis to extract local waveform segments of the melting temperature-fluorescence intensity derivative curve, calculating the curvature extrema and area integral of the local waveform segments as attribute features of the nodes of the graph structure; calculating the mutual information value between the local waveform segments corresponding to different gene loci, using the mutual information value as the weight of the edges between different nodes, and generating the multidimensional melting curve topology graph based on the attribute features of the nodes and the weights of the edges.

[0006] Preferably, the feature extraction processor extracts the co-fluctuation features of peak temperature offset and edge weights of nodes in the multidimensional melting curve topology graph, inputs the co-fluctuation features into a graph convolutional network, and outputs a polymorphic combination feature vector, including: in each layer of graph convolution operation of the graph convolutional network, aggregating the peak temperature offset and edge weights of the neighboring nodes of the target node in the multidimensional melting curve topology graph, calculating the Hadamard product of the feature difference between the target node and the neighboring nodes and the edge weights, concatenating the Hadamard product with the features of the previous hidden layer of the target node and processing it through a nonlinear activation function, and extracting high-order topological features as the polymorphic combination feature vector after propagation layer by layer.

[0007] Preferably, the decision processor maps the polymorphic combination feature vector to a hypertension drug metabolism network graph model and calculates the joint decay coefficient of drug-metabolizing enzyme activity and receptor binding rate, including: constructing a hypertension drug metabolism network graph model with drug-metabolizing enzymes and receptors as nodes and metabolic response pathways as edges; projecting the polymorphic combination feature vector to the node space of the hypertension drug metabolism network graph model to obtain the feature representation of each node; calculating the inhibitory factor of the metabolic response pathway based on the feature representation of each node; and accumulating the inhibitory factor through backpropagation along the edges of the hypertension drug metabolism network graph model to generate the joint decay coefficient characterizing the overall metabolic capacity.

[0008] Preferably, the decision processor determines the type of antihypertensive medication and dosage adjustment parameters based on the joint attenuation coefficient, including: constructing an objective function of drug concentration and blood pressure reduction amplitude in a preset antihypertensive drug efficacy response surface space, using the joint attenuation coefficient as a constraint; performing non-convex optimization on the objective function to obtain the extreme point that makes the blood pressure reduction amplitude reach the target threshold and the drug concentration within a safe window; extracting the drug identifier corresponding to the extreme point as the type of antihypertensive medication, and extracting the concentration ratio corresponding to the extreme point as the dosage adjustment parameter.

[0009] Preferably, the data acquisition terminal acquires fluorescence melting curve data of hypertension-related multi-gene targets, including: acquiring time-series data of the original fluorescence signal intensity changing with temperature; performing wavelet transform decomposition on the time-series data to separate low-frequency baseline drift components and high-frequency fluorescence noise components; reconstructing denoised fluorescence melting curve data from the decomposed data; obtaining the first and second derivatives of the denoised fluorescence melting curve data along the temperature axis; and inputting the joint data stream of the first and second derivatives as the fluorescence melting curve data of the hypertension-related multi-gene targets into the graph construction processor.

[0010] Preferably, calculating the mutual information value between the local waveform segments corresponding to different gene loci, and using the mutual information value as the weight of the edge between different nodes, includes: performing kernel density estimation on the local waveform segments to obtain the edge probability density distribution and joint probability density distribution of the local waveform segments; calculating the conditional entropy between the local waveform segments based on the edge probability density distribution and the joint probability density distribution; solving for the nonlinear dependency metric based on the difference between the conditional entropy and the edge entropy; setting the nonlinear dependency metric as the mutual information value; and normalizing the mutual information value and assigning it to the corresponding edge in the multidimensional melting curve topology graph to form a weighted adjacency matrix.

[0011] Preferably, the Hadamard product is concatenated with the features of the previous hidden layer of the target node and processed by a nonlinear activation function. After propagation layer by layer, higher-order topological features are extracted as the polymorphic combination feature vector. This includes: introducing a multi-head attention mechanism to calculate the attention weight coefficients corresponding to each neighbor node of the target node; using the attention weight coefficients to perform a weighted summation of the Hadamard product; performing cross-channel feature fusion with the weighted summation result and the features of the previous hidden layer of the target node; inputting the fused features into a linear unit with leakage correction for nonlinear mapping; iteratively executing the cross-channel feature fusion and nonlinear mapping operations until a preset number of propagation layers is reached; and outputting the polymorphic combination feature vector of fixed dimensions.

[0012] Preferably, the inhibitory factor of the metabolic response pathway is calculated based on the characteristic representation of each node, and the inhibitory factor is accumulated by backpropagation along the edges of the hypertension drug metabolism network graph model to generate the joint decay coefficient characterizing the overall metabolic capacity. This includes: correcting the enzyme reaction rate on the edges of the hypertension drug metabolism network graph model based on the Michaelis-Menten kinetic equation; determining the Michaelis constant correction term and the maximum reaction rate correction term based on the characteristic representation of each node; calculating the inhibitory factor of the metabolic response pathway based on the Michaelis constant correction term and the maximum reaction rate correction term; performing message passing and cumulative multiplication operations on the inhibitory factor along the topological path of the hypertension drug metabolism network graph model, and converting the cumulative multiplication result into a scalar value as the joint decay coefficient.

[0013] Preferably, the objective function is solved using non-convex optimization to obtain the extreme point where the blood pressure reduction reaches the target threshold and the drug concentration is within a safe window. This includes: transforming the safe window constraint of the drug concentration into a dual problem using the Lagrange multiplier method; constructing a Lagrange augmented function using the joint decay coefficient; iteratively optimizing the Lagrange augmented function using a quasi-Newton method; calculating the approximate value of the Hessian matrix and the gradient direction of the Lagrange augmented function; updating the drug concentration variable along the gradient direction until a preset convergence criterion is met; outputting the combination of drug concentration variables corresponding to the convergence point as the extreme point; and analyzing each drug component in the extreme point to obtain the drug identifier and the dose adjustment parameters.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The melting temperature-fluorescence intensity derivative curve of each gene locus is mapped to nodes in a graph structure. The melting feature associations between different gene loci are mapped to edges to construct a multidimensional melting curve topology. The peak temperature offset and edge weight of the nodes are extracted through a graph convolutional network to output a polymorphic combination feature vector. This overcomes the defect that single-peak independent alignment cannot separate overlapping peaks and realizes the quantitative analysis of the polymorphic co-fluctuation associations of multiple genes. The polymorphic combination feature vector is mapped to a graph model of hypertension drug metabolism to calculate the joint decay coefficient of drug metabolic enzyme activity and receptor binding rate. Based on the joint decay coefficient, the type of hypertension drug and dosage adjustment parameters are determined. The synergistic effect between genes is mapped to the pharmacokinetic network, eliminating the medication bias caused by isolated qualitative matching.

[0015] 2. By setting a sliding window along the temperature axis to extract local waveform segments and calculating the curvature extrema and area integral as node attributes, and calculating the mutual information value between local waveform segments of different gene loci as edge weights, the expressive power of local features in overlapping regions and the measurement accuracy of nonlinear associations between nodes are improved. In the hypertension drug metabolism network graph model, the inhibitory factor of the metabolic response pathway is calculated based on the node characteristics and the joint attenuation coefficient is generated by backpropagation along the edges. Combined with the objective function of drug concentration and blood pressure reduction, non-convex optimization is performed to achieve closed-loop quantitative calculation from thermodynamic signals to pharmacokinetic parameters, and the interference of misjudgment of a single site on the overall decision is suppressed. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall workflow of the hypertension medication gene polymorphism melting curve detection and decision-making system of the present invention. Figure 2 This is a flowchart of the fluorescence melting curve data acquisition and preprocessing process of the present invention; Figure 3 This is a flowchart illustrating the construction of the multidimensional melting curve topology diagram of the present invention. Figure 4 This is a flowchart of the polymorphic combination feature vector extraction process of the present invention; Figure 5 This is a flowchart of the calculation of the joint attenuation coefficient of the drug metabolism network in this invention; Figure 6 This is a flowchart illustrating the decision-making process for adjusting the types and dosages of hypertension medications according to the present invention. Detailed Implementation

[0017] refer to Figure 1In one embodiment, the data acquisition terminal acquires fluorescence melting curve data of hypertension-related multi-gene targets. Hypertension-related multi-gene targets include multiple polymorphic sites of genes such as CYP2D6, CYP2C9, CYP3A4, ADRB1, AGT, and ACE. The data acquisition terminal receives raw data output from a real-time quantitative PCR instrument, which contains time pairs of temperature values ​​and corresponding fluorescence intensity values. The graph construction processor receives the fluorescence melting curve data output from the data acquisition terminal, maps the melting temperature-fluorescence intensity derivative curve of each gene site to nodes of a graph structure, and maps the melting feature associations between different gene sites to edges, constructing a multidimensional melting curve topology graph. Each node in the graph structure corresponds to a set of melting curve features for a gene site, and each node contains multiple attribute features used to characterize the local morphology and global features of the melting curve of that gene site. Each edge in the graph structure corresponds to the degree of melting feature association between two gene sites, and the edge weight is used to quantify the strength of this association. The feature extraction processor receives the multidimensional melting curve topology graph output by the graph construction processor and extracts the co-variation features of peak temperature offset and edge weights of nodes in the multidimensional melting curve topology graph. Peak temperature offset refers to the difference between the peak melting temperature of the test sample and the peak melting temperature of the wild-type reference sample. Co-variation features refer to the joint change pattern between the peak temperature offsets of multiple nodes and the edge weights connecting these nodes. The feature extraction processor inputs the co-variation features into a graph convolutional network and outputs a polymorphic combination feature vector. The polymorphic combination feature vector is a fixed-dimensional vector containing polymorphic information of all gene loci and the co-variation information between these polymorphisms. The decision processor receives the polymorphic combination feature vector output by the feature extraction processor and maps it to a hypertension drug metabolism network graph model. The hypertension drug metabolism network graph model is pre-constructed and includes the main metabolic pathways of hypertension drugs in the human body, the enzymes involved in metabolism, and the receptors for drug action. The decision processor calculates the combined attenuation coefficient of drug-metabolizing enzyme activity and receptor binding rate. This combined attenuation coefficient is a scalar value used to characterize the comprehensive impact of gene polymorphism on the overall metabolism and efficacy of the drug. Based on the combined attenuation coefficient, the decision processor determines the type and dosage adjustment parameters for hypertension medication, thus constructing a decision system for detecting melting curves of gene polymorphisms in hypertension medication.

[0018] In this embodiment, the time-series data of the original fluorescence signal intensity changing with temperature acquired by the data acquisition terminal is represented as follows: ,in This represents temperature values ​​in degrees Celsius. The sampling interval for the raw time-series data is 0.1 degrees Celsius, and the temperature range is from 50 degrees Celsius to 95 degrees Celsius. The graph construction processor first preprocesses the raw time-series data to generate a melting temperature-fluorescence intensity derivative curve. The melting temperature-fluorescence intensity derivative curve is represented as follows: ,in This represents the change in fluorescence intensity. This represents the amount of temperature change. The graph construction processor will assign each gene locus to... A curve is mapped to a node in a graph structure. ,in An index representing gene loci. , This represents the total number of gene loci. The graph construction processor calculates the total number of any two nodes. and The degree of correlation of melting characteristics between them is mapped to the degree of correlation of connection nodes. and edge weight The multidimensional melting curve topology diagram is represented as follows: ,in Represents a set of nodes. Denotes the set of edges. This represents the set of edge weights.

[0019] The feature extraction processor first calculates each node Peak temperature shift Peak temperature offset The calculation formula is:

[0020] in, Indicates the first in the sample to be tested Peak melting temperature of each gene locus Indicating the first wild-type reference sample The peak melting temperature of each gene locus. A feature extraction processor extracts a multidimensional melting curve topology map. The set of peak temperature offsets of all nodes in the middle and the set of weights for all edges Construct a cooperative fluctuation feature matrix Cooperative fluctuation characteristic matrix The dimension is , of which Before the journey Each element represents the edge weight. , No. The first line Each element represents the peak temperature offset. The feature extraction processor will coordinate the fluctuation feature matrix. Input a graph convolutional network. A graph convolutional network contains... Graph convolutional layers, each with an output dimension of . ,in The input layer dimension of a graph convolutional network is... The output layer dimension is The output layer outputs a polymorphic combination feature vector. .

[0021] Decision processors pre-build graph models of hypertension drug metabolism networks. ,in This represents a set of drug-metabolizing enzymes and receptor nodes. This indicates the total number of drug-metabolizing enzymes and receptors. This represents the set of edges in a metabolic response pathway. The decision processor combines polymorphic feature vectors. Mapping to a graph model of hypertension drug metabolism The node space is obtained to get each node. Feature representation ,in This represents the feature dimensions of each node in the drug metabolism network graph model. The decision processor bases its decisions on the feature representations of each node. Calculate the attenuation coefficient of drug-metabolizing enzyme activity Attenuation coefficient of receptor binding rate Drug metabolism enzyme activity attenuation coefficient The range of values ​​is ,when When the enzyme's activity is unaffected by gene polymorphism, it indicates that the enzyme's activity is not affected by gene polymorphism; when... At this point, it indicates that the enzyme's activity is completely lost. Receptor binding rate attenuation coefficient. The range of values ​​is ,when When the receptor binding rate is 1, it indicates that the binding rate of the drug is not affected by gene polymorphism; when When the threshold is reached, it indicates that the receptor cannot bind to the drug at all. The decision processor calculates the combined attenuation coefficient of drug-metabolizing enzyme activity and receptor binding rate. Joint attenuation coefficient The calculation formula is:

[0022] in, Indicates the first The weighting coefficient of each drug-metabolizing enzyme in the overall metabolic process. Indicates the first The weighting coefficient of each receptor in the drug action process , .

[0023] The decision processor pre-constructs a preset drug efficacy response surface space for hypertension drugs. This drug efficacy response surface space for hypertension drugs includes... The pharmacodynamic response surfaces of several commonly used antihypertensive drugs are shown below. The pharmacodynamic response surface for each drug is represented as follows: ,in Indicates drug concentration. Indicates the joint attenuation coefficient. This indicates the voltage drop magnitude. The decision processor uses the joint attenuation coefficient. As constraints, construct an objective function relating drug concentration and blood pressure reduction. Objective function The calculation formula is:

[0024] in, A penalty coefficient representing drug concentration, used to balance the magnitude of blood pressure reduction with drug side effects. The decision processor operates on the objective function. We perform non-convex optimization to find the extreme point where the blood pressure reduction reaches the target threshold and the drug concentration is within a safe window. The safety window refers to the range between the lowest effective concentration and the highest tolerated concentration for each drug. The decision processor extracts the drug identifiers corresponding to drugs with concentrations greater than zero at extreme points as the types of hypertension medications, and extracts the ratio of the concentration of each drug at the extreme points to the concentration corresponding to the standard dose of that drug as the dose adjustment parameter.

[0025] In this embodiment, the hypertension-related gene loci and their corresponding melting temperature reference values ​​are shown in Table 1.

[0026] Table 1. Hypertension-related gene loci and corresponding melting temperature reference values. ; Table 1 shows the reference melting temperatures for wild-type, heterozygous mutant, and homozygous mutant types of the six hypertension-related gene loci involved in this embodiment. These reference values ​​were obtained through fluorescence melting curve experiments on a large number of wild-type and mutant standard samples. In the actual detection process, the peak melting temperature of the sample to be tested was compared with the reference values ​​in Table 1. Combined with the cooperative fluctuation characteristics of the multidimensional melting curve topology, the polymorphism status of each gene locus can be accurately determined. For example, when the peak melting temperature of the CYP2D6*10 locus is 83.5℃, the locus is a homozygous mutant; when the peak melting temperature shows two peaks at 83.5℃ and 86.2℃, the locus is a heterozygous mutant; and when the peak melting temperature is 86.2℃, the locus is wild-type.

[0027] In this embodiment, the multidimensional melting curve topology graph constructed by the graph construction processor contains 6 nodes, corresponding to the 6 gene loci in Table 1. Each node includes attribute features such as peak temperature offset, curve curvature extrema, and curve area integral. An edge exists between any two nodes, and the weight of the edge represents the degree of correlation between the melting features of the two gene loci. For example, the melting curves of the CYP2D610 and CYP2C93 gene loci overlap in the 85℃ to 87℃ range, indicating a strong correlation between their melting features; therefore, the edge connecting these two nodes has a larger weight. Conversely, the melting curves of the CYP3A4*1G and ACEI / D gene loci overlap less, indicating a weaker correlation between their melting features; therefore, the edge connecting these two nodes has a smaller weight. The feature extraction processor extracts the co-variation features of these node attributes and edge weights through a graph convolutional network, effectively distinguishing overlapping peaks and accurately identifying the polymorphic state of each gene locus.

[0028] In this embodiment, the hypertension drug metabolism network model constructed by the decision processor contains 12 nodes, of which 8 nodes correspond to drug-metabolizing enzymes and 4 nodes correspond to drug receptors. The drug-metabolizing enzyme nodes include CYP2D6, CYP2C9, CYP3A4, CYP1A2, CYP2B6, CYP2C19, UGT1A1, and SULT1A1. The drug receptor nodes include ADRB1, AGTR1, ACE, and AT1R. Metabolic pathway edges represent the metabolic reactions of drugs catalyzed by enzymes and the binding reactions between drugs and receptors. For example, metoprolol is mainly metabolized by CYP2D6, therefore, there is an edge connecting metoprolol and CYP2D6 in the hypertension drug metabolism network model. Valsartan is mainly metabolized by CYP2C9 and acts on the AT1R receptor; therefore, there are edges connecting valsartan and CYP2C9, and edges connecting valsartan and AT1R in the hypertension drug metabolism network model. The decision processor maps the polymorphism combination feature vectors to the model, calculates the activity decay coefficient of each enzyme and receptor, and then obtains the joint decay coefficient. The joint decay coefficient can comprehensively reflect the influence of all gene polymorphisms on drug metabolism and efficacy. Based on the joint decay coefficient, the decision processor performs optimization in a preset pharmacodynamic response surface space to obtain the optimal drug type and dosage adjustment parameters.

[0029] refer to Figure 2 and Figure 3In a preferred embodiment, the data acquisition end acquires time-series data of the original fluorescence signal intensity changing with temperature. Wavelet transform decomposition is performed on the time-series data to separate the low-frequency baseline drift component and the high-frequency fluorescence noise component. The denoised fluorescence melting curve data is reconstructed from the decomposed data. The first and second derivatives of the denoised fluorescence melting curve data are calculated along the temperature axis, and the joint data stream of the first and second derivatives is input as the fluorescence melting curve data of hypertension-related multi-gene targets to the graph construction processor. The graph construction processor sets a sliding window along the temperature axis to extract local waveform segments of the melting temperature-fluorescence intensity derivative curve, and calculates the curvature extrema and area integral of the local waveform segments as attribute features of the nodes in the graph structure. The mutual information values ​​between local waveform segments corresponding to different gene loci are calculated, and the mutual information values ​​are used as the weights of the edges between different nodes. A multidimensional melting curve topology graph is generated based on the attribute features of the nodes and the weights of the edges.

[0030] In this embodiment, the raw fluorescence signal intensity time series data acquired by the data acquisition terminal The data includes baseline drift and random noise. Baseline drift is a low-frequency signal component caused by factors such as temperature variations in the PCR instrument and non-specific binding of fluorescent dyes. Random noise is a high-frequency signal component caused by factors such as thermal noise from the photodetector and noise from electronic circuits. Discrete wavelet transform is used at the data acquisition end to process the raw time-series data. Decomposition is performed. The discrete wavelet transform decomposition has 5 levels, and the db4 wavelet is selected as the wavelet basis function. Original time series data. After discrete wavelet transform decomposition, a low-frequency approximate component is obtained. and 5 high-frequency detail components Among them, the low-frequency approximation component Corresponding to baseline drift components, high-frequency detail components and Corresponding to high-frequency fluorescence noise components, high-frequency detail components This corresponds to the effective fluorescence signal components. The data acquisition end will use the low-frequency approximation components. and high-frequency detail components Set to zero to preserve high-frequency detail components. The denoised fluorescence melting curve data was reconstructed using inverse discrete wavelet transform. The formula for calculating the inverse discrete wavelet transform is:

[0031] in, Indicates the first Layer Wavelet coefficients, Indicates the first Layer Wavelet basis functions.

[0032] Data acquisition terminal for denoised fluorescence melting curve data Taking the first derivative along the temperature axis and second derivative First derivative Used to characterize the rate of change of fluorescence intensity with temperature, its peak value corresponds to the melting temperature. Second derivative. The rate of change used to characterize the rate of change in fluorescence intensity has a zero-crossing point that corresponds to the precise location of the melting temperature. The data acquisition end will use the first derivative... and second derivative The joint data stream is input as fluorescence melting curve data to the graph construction processor. The joint data stream is represented as... ,in This indicates the temperature value.

[0033] The graph-constructing processor sets a sliding window along the temperature axis to capture the combined data stream. A partial waveform segment. The width of the sliding window is... The sliding step size is In this embodiment, the width of the sliding window... 2 degrees Celsius, sliding step size It is 0.5 degrees Celsius. For the first... A gene locus, captured by the graph construction processor. There are several local waveform segments, each represented as... ,in The graph-building processor calculates each local waveform segment. curvature extrema and area integral Curvature extrema The calculation formula is:

[0034] in, Indicates the first The initial temperature of each sliding window. Area integral. The calculation formula is:

[0035] The graph-constructing processor will extract the curvature extrema of all local waveform segments. and area integral As the first Nodes The attributes and characteristics of a node. The attribute feature vector is represented as: .

[0036] The graph construction processor calculates the mutual information values ​​between local waveform segments corresponding to different gene loci, and uses these mutual information values ​​as the weights of edges between different nodes. For the first... The gene locus and the first For each gene locus, the graph construction processor first performs kernel density estimation on its local waveform segments to obtain the marginal probability density distribution and joint probability density distribution of the local waveform segments. The kernel density estimation uses a Gaussian kernel function, and the bandwidth of the Gaussian kernel function is determined using cross-validation. The first gene locus The edge probability density distribution of a local waveform segment is represented as follows: , No. The first gene locus The edge probability density distribution of a local waveform segment is represented as follows: Their joint probability density distribution is expressed as ,in Indicates the first The first gene locus Eigenvalues ​​of a local waveform segment Indicates the first The first gene locus The characteristic values ​​of a local waveform segment.

[0037] The graph-based processor calculates the conditional entropy between local waveform segments based on the edge probability density distribution and the joint probability density distribution. The first gene locus The local waveform segment relative to the first The first gene locus Conditional entropy of a local waveform segment The calculation formula is:

[0038] The graph construction processor calculates the nonlinear dependency metric based on the difference between conditional entropy and edge entropy, and sets the nonlinear dependency metric as a mutual information value. The first gene locus The local waveform segment and the first The first gene locus Mutual information value between local waveform segments The calculation formula is:

[0039] in, Indicates the first The first gene locus The marginal entropy of a local waveform segment is calculated using the following formula:

[0040] Graph construction processor computes the first The gene locus and the first The average of the mutual information values ​​of all local waveform fragment pairs between two gene loci is used as the mutual information value between those two gene loci. Mutual information value The calculation formula is:

[0041] The graph construction processor will use mutual information values Normalization is performed, and the normalized mutual information value is obtained. The calculation formula is:

[0042] in, This represents the maximum mutual information value among all gene locus pairs. The normalized mutual information value. The range of values ​​is The graph construction processor will Assigning connection nodes to the multidimensional melting curve topology graph and edge This forms a weighted adjacency matrix. Weighted adjacency matrix It is A symmetric matrix, where , .

[0043] In this embodiment, the normalized mutual information values ​​between different gene locus pairs are shown in Table 2.

[0044] Table 2 Normalized mutual information values ​​between different gene locus pairs ; Table 2 shows the normalized mutual information values ​​between each pair of the six gene loci in this embodiment. These values ​​were calculated using fluorescence melting curve data from 1000 clinical samples. As can be seen from Table 2, the normalized mutual information value between CYP2D610 and CYP2C93 is the highest, at 0.87, indicating the strongest association between the melting characteristics of these two gene loci. This is because the melting curves of these two gene loci overlap significantly in the 85℃ to 87℃ range, and their melting characteristic changes influence each other. The normalized mutual information value between ADRB1C1165G and AGTM235T is the second highest, at 0.79. The normalized mutual information value between CYP2D6*10 and ACEI / D is the lowest, at 0.31, indicating the weakest association between the melting characteristics of these two gene loci. This is because the overlap range of the melting curves of these two gene loci is small, and their melting characteristic changes have little mutual influence.

[0045] In this embodiment, the graph construction processor uses a sliding window to extract local waveform segments and calculates curvature extrema and area integrals as node attribute features, effectively capturing the local morphological features of the melting curve. Especially for overlapping peaks, the curvature extrema and area integrals of the local waveform segments can reflect subtle changes in the overlapping region, thereby improving the distinguishability of overlapping peaks. The graph construction processor uses mutual information values ​​as edge weights, accurately measuring the degree of nonlinear correlation between melting features of different gene loci. Compared with linear correlation coefficients, mutual information values ​​can capture nonlinear dependencies, making them more suitable for characterizing complex correlations between melting features of gene loci. By constructing a multidimensional melting curve topology graph that includes node local attributes and edge nonlinear correlations, the feature information of melting curves at multiple gene loci can be comprehensively and accurately represented, laying the foundation for subsequent feature extraction and polymorphism recognition.

[0046] In this embodiment, wavelet transform is used at the data acquisition end to denoise the original fluorescence signal, effectively separating baseline drift and high-frequency noise, and improving the signal-to-noise ratio of the melting curve. Compared with traditional moving average filtering and polynomial fitting denoising methods, wavelet transform has the characteristics of multi-resolution analysis, enabling signal decomposition and reconstruction at different scales, thereby better preserving the local features of the signal. In particular, for sharp peaks in the melting curve, wavelet transform can remove noise while maintaining the position and shape of the peaks, improving the accuracy of melting temperature measurement. The data acquisition end simultaneously outputs a joint data stream of the first and second derivatives, providing richer melting curve feature information than a single derivative. The first derivative reflects the rate of the melting process, and the second derivative reflects the acceleration of the melting process; the combination of the two can more accurately determine the location of the melting temperature and the morphological characteristics of the melting curve.

[0047] refer to Figure 4In a preferred embodiment, in each layer of the graph convolutional network, the feature extraction processor aggregates the peak temperature offsets and edge weights of the target node's neighboring nodes in the multidimensional melting curve topology graph. It calculates the Hadamard product of the feature differences between the target node and its neighboring nodes and the edge weights. The Hadamard product is concatenated with the features of the target node's previous hidden layer and processed through a nonlinear activation function. After propagation layer by layer, higher-order topological features are extracted as a polymorphic combined feature vector. The feature extraction processor introduces a multi-head attention mechanism to calculate the attention weight coefficients corresponding to each neighboring node of the target node, and uses these attention weight coefficients to perform a weighted summation of the Hadamard product. The weighted summation result is then fused with the features of the target node's previous hidden layer across channels. The fused features are input into a linear unit with leakage correction for nonlinear mapping. This cross-channel feature fusion and nonlinear mapping operation is iteratively executed until a preset number of propagation layers is reached, outputting a fixed-dimensional polymorphic combined feature vector.

[0048] refer to Figure 5 The decision processor constructs a hypertension drug metabolism network graph model with drug-metabolizing enzymes and receptors as nodes and metabolic reaction pathways as edges. It projects polymorphic combination feature vectors onto the node space of the hypertension drug metabolism network graph model to obtain the feature representation of each node. Based on the feature representation of each node, it calculates the inhibitory factor of the metabolic reaction pathway and accumulates the inhibitory factor through backpropagation along the edges of the hypertension drug metabolism network graph model, generating a joint decay coefficient characterizing the overall metabolic capacity. The decision processor corrects the enzyme-catalyzed reaction rates on the edges of the hypertension drug metabolism network graph model based on the Michaelis-Menten kinetic equation, determining the Michaelis constant correction term and the maximum reaction rate correction term based on the feature representation of each node. It calculates the inhibitory factor of the metabolic reaction pathway based on the Michaelis constant correction term and the maximum reaction rate correction term. It performs graph message passing and cumulative multiplication operations on the inhibitory factor along the topological path of the hypertension drug metabolism network graph model, converting the cumulative multiplication result into a scalar value as the joint decay coefficient.

[0049] refer to Figure 6The decision processor constructs an objective function relating drug concentration and blood pressure reduction within a predefined antihypertensive drug efficacy response surface space, constrained by a joint decay coefficient. Non-convex optimization is performed on the objective function to obtain the extreme points where the blood pressure reduction reaches the target threshold while the drug concentration remains within a safe window. The drug identifier corresponding to each extreme point is extracted as the type of antihypertensive drug, and the concentration ratio corresponding to each extreme point is extracted as the dose adjustment parameter. The decision processor uses the Lagrange multiplier method to transform the safe window constraint of drug concentration into a dual problem, constructing a Lagrange augmented function based on the joint decay coefficient. The quasi-Newton method is used to iteratively optimize the Lagrange augmented function, calculating the approximate value of the Hessian matrix and the gradient direction. The drug concentration variable is updated along the gradient direction until a predefined convergence criterion is met. The combination of drug concentration variables corresponding to the convergence point is output as the extreme point, and the drug identifier and dose adjustment parameter are obtained by analyzing each drug component in the extreme point.

[0050] In this embodiment, the graph convolutional network used by the feature extraction processor comprises three graph convolutional layers and one fully connected output layer. Each graph convolutional layer has an output dimension of 64, and the fully connected output layer has an output dimension of 128. The input to the graph convolutional network is a multidimensional melting curve topology graph. and node attribute feature matrix ,in This represents the attribute feature dimension of each node. In the... In layer graph convolution operations, the target node The hidden layer features are represented as ,in Indicates the first Output dimension of the layer. Target node. The set of neighbor nodes is represented as .

[0051] The feature extraction processor first calculates the target node With each neighbor node Feature difference Feature difference The calculation formula is:

[0052] in, and Representing the target node respectively and neighboring nodes In the Hidden layer features of the layer. The feature extraction processor calculates the feature difference. With edge weight Hadama Hadama accumulation The calculation formula is:

[0053] in, This represents the Hadamard product operation, which is the element-wise multiplication of two vectors.

[0054] The feature extraction processor introduces a multi-head attention mechanism to compute the target node. Each neighboring node Corresponding attention weight coefficient Multi-head attention mechanisms include There are *n* attention heads, and each attention head calculates a set of attention weight coefficients. In the *n*th... In each attention head, the attention weight coefficient The calculation formula is:

[0055] in, Indicates the first The learnable weight matrix for each attention head. Indicates the first A learnable attention vector for each attention head. This represents a vector concatenation operation. This represents a linear unit activation function with leakage correction. The feature extraction processor will... The final attention weight coefficient is obtained by averaging the attention weight coefficients calculated from each attention head. .

[0056] Feature extraction processors utilize attention weight coefficients For Hadama Perform a weighted summation to obtain the feature aggregation result of neighbor nodes. Neighbor node feature aggregation results The calculation formula is:

[0057] The feature extraction processor aggregates the features of neighboring nodes. Features of the previous hidden layer of the target node Cross-channel feature fusion is performed. Cross-channel feature fusion employs a gating mechanism to fuse features. The calculation formula is:

[0058] in, The gating coefficient is represented by the following formula:

[0059] in, and This represents the learnable weight matrix and bias vector. This represents the sigmoid activation function.

[0060] The feature extraction processor will fuse features The input is fed into a linear unit with leakage correction for nonlinear mapping to obtain the target node. In the Hidden layer features of the layer Hidden layer features The calculation formula is:

[0061] in, and Indicates the first The graph convolutional layer has a learnable weight matrix and bias vector. The feature extraction processor iteratively executes the above graph convolution operation until the preset number of propagation layers is reached. After the final graph convolution operation is completed, the feature extraction processor performs global average pooling on the hidden layer features of all nodes to obtain a polymorphic combined feature vector. The formula for calculating global average pooling is:

[0062] A graphical model of hypertension drug metabolism constructed using a decision processor It is a directed graph where the direction of the edges represents the direction of metabolic reactions or the direction of drug action. The decision processor uses a graph attention network to combine polymorphic feature vectors. Projecting the model onto the node space of the hypertension drug metabolism network diagram, we obtain each node. Feature representation The structure of the graph attention network is similar to that of the graph convolutional network described above, containing two graph attention layers, each with an output dimension of 64.

[0063] The decision processor modifies the enzyme reaction rates on the edges of the hypertension drug metabolism network model based on the Michaelis-Menten kinetic equation. The Michaelis-Menten kinetic equation describes the relationship between enzyme reaction rate and substrate concentration, and its basic form is:

[0064] in, Indicates the rate of enzyme-catalyzed reaction. Indicates the maximum reaction rate. Indicates substrate concentration. This represents the Mie constant. The decision processor is based on the node... Feature representation Determine the correction term for the Michaelis constant With the maximum reaction rate correction term Michaelis constant correction term and maximum reaction rate correction term The calculation formula is:

[0065]

[0066] in, This represents the learnable weight matrix and bias vector. The corrected Mie constant. and maximum reaction rate The calculation formula is:

[0067]

[0068] in, and denoted as Michaelis constant and maximum reaction rate in the wild-type state, respectively.

[0069] The decision processor calculates the inhibitory factor of the metabolic response pathway based on the corrected Michaelis constant and the maximum reaction rate. Inhibitory factors Represents gene polymorphism in connection nodes and metabolic response pathways The degree of inhibition is calculated using the following formula:

[0070] Inhibitors The range of values ​​is ,when When the time is right, it indicates that the metabolic pathway is not affected by gene polymorphism; when When the signal is 0, it indicates that the metabolic pathway is completely inhibited.

[0071] The decision processor performs message passing and cumulative multiplication operations on the inhibitory factors along the topological path of the hypertension drug metabolism network model. For each drug, the decision processor starts from the receptor node of drug action and propagates backward along the metabolic response pathway, accumulating the inhibitory factors of all relevant metabolic response pathways. Its joint attenuation coefficient The calculation formula is:

[0072] in, Indicates drug The set of edges along all topological paths from the receptor node to the drug intake node. The decision processor calculates a weighted average of the joint decay coefficients of all drugs to obtain the overall joint decay coefficient. Overall joint attenuation coefficient The calculation formula is:

[0073] in, Indicates drug The weighting coefficients, .

[0074] The pre-constructed pharmacodynamic response surface space of hypertension drugs by the decision processor includes The pharmacodynamic response surfaces of several commonly used antihypertensive drugs are presented. The pharmacodynamic response surface for each drug was obtained by fitting a large amount of clinical data, and is represented as follows: ,in This represents the fitted parameters. The decision processor uses the joint decay coefficient. As constraints, construct an objective function relating drug concentration and blood pressure reduction. Objective function The calculation formula is:

[0075] The decision processor introduces a safety window constraint on drug concentration: ,in Indicates drug The minimum effective concentration, Indicates drug The highest tolerable concentration. The decision processor also introduces a target threshold constraint on the magnitude of the pressure drop: ,in This indicates the target reduction in blood pressure.

[0076] The decision processor uses the Lagrange multiplier method to transform the above constraints into a dual problem and constructs the Lagrange augmented function. The formula for calculating the Lagrange augmented function is:

[0077] in, Denotes Lagrange multipliers, and The decision processor uses a quasi-Newton method to iteratively optimize the Lagrange augmented function. The quasi-Newton method uses the BFGS formula to update the approximate value of the Hessian matrix. The formula for calculating BFGS is as follows:

[0078] in, Indicates the iteration step size. Represents the gradient difference. Indicates the first The gradient of the next iteration.

[0079] The decision processor calculates the gradient direction of the Lagrange augmented function. Update drug concentration variables along the gradient direction ,in The step size is determined through line search. The decision processor iteratively executes the above update operation until the preset convergence criterion is met: ,in This represents the convergence threshold. After convergence, the decision processor outputs the combination of drug concentration variables corresponding to the convergence point. As extreme points, the decision processor extracts the drug identifiers corresponding to drugs with concentrations greater than zero at the extreme points as the types of hypertension medications, and extracts the ratio of the concentration of each drug at the extreme points to the concentration corresponding to the standard dose of that drug as the dose adjustment parameter.

[0080] In this example, the optimal dosing regimens corresponding to different combined attenuation coefficients are shown in Table 3.

[0081] Table 3 Optimal Dosing Regimens for Different Combined Attenuation Coefficients ; Table 3 shows the optimal dosing regimens for different combined attenuation coefficients in this embodiment. These regimens were obtained through non-convex optimization of the Lagrange augmented function. As can be seen from Table 3, as the combined attenuation coefficient decreases, indicating an increased influence of gene polymorphism on drug metabolism and efficacy, a wider variety of drugs is required for combination therapy, and the dosage adjustment parameters for each drug also decrease accordingly. For example, when the combined attenuation coefficient is 0.95, only metoprolol monotherapy is needed, with a dosage adjustment parameter of 0.9, i.e., 90% of the standard dose. When the combined attenuation coefficient is 0.55, a combination of amlodipine, hydrochlorothiazide, and metoprolol is required, with a dosage adjustment parameter of 0.5 for metoprolol, i.e., 50% of the standard dose. This is because gene polymorphism leads to reduced activity of drug-metabolizing enzymes and decreased receptor binding rates, necessitating a reduction in drug dosage to avoid drug accumulation and increased side effects, while simultaneously increasing the variety of drugs to achieve the target blood pressure reduction.

[0082] In this embodiment, the graph convolutional network employed by the feature extraction processor incorporates a multi-head attention mechanism and a cross-channel feature fusion gating mechanism. This adaptively aggregates feature information from neighboring nodes, improving the accuracy of feature extraction. The multi-head attention mechanism learns the varying importance of different neighboring nodes to the target node, thus more effectively aggregating valuable feature information. The cross-channel feature fusion gating mechanism adaptively balances the contributions of the target node's own features and the aggregated features from neighboring nodes, enhancing the robustness of feature representation. By extracting high-order topological features through layer-by-layer propagation, complex collaborative relationships between multiple gene loci can be captured, thereby more accurately representing polymorphic combination features.

[0083] In this embodiment, the decision processor modifies the enzyme reaction rate of the drug metabolism network model based on the Michaelis-Menten kinetic equation. This quantifies the impact of gene polymorphism on enzyme activity into a Michaelis constant and a correction term for the maximum reaction rate, allowing for more accurate calculation of inhibitory factors in metabolic pathways. The decision processor performs backpropagation accumulation of inhibitory factors along the topological path of the drug metabolism network model, comprehensively considering the inhibitory effects of all relevant metabolic pathways to obtain a more accurate joint decay coefficient. This joint decay coefficient comprehensively reflects the combined impact of gene polymorphism on the overall metabolism and efficacy of the drug, providing a reliable basis for medication decisions.

[0084] In this embodiment, the decision processor uses the Lagrange multiplier method to transform the constraints into a dual problem, and employs the quasi-Newton method to perform non-convex optimization of the Lagrange augmented function, efficiently finding the optimal drug concentration combination that satisfies all constraints. The quasi-Newton method has advantages such as fast convergence speed and low computational complexity, making it suitable for solving high-dimensional non-convex optimization problems. The optimal medication plan obtained through optimization can control the drug concentration within a safe window while ensuring that the blood pressure reduction reaches the target threshold, thereby achieving personalized precision medication.

[0085] In this embodiment, mapping the polymorphic combination feature vector to a hypertension drug metabolism network model to calculate the joint decay coefficient allows for direct mapping of inter-gene synergistic effects to the pharmacokinetic network, thereby eliminating medication bias caused by isolated qualitative matching. Compared to existing methods that match medication rules one by one based on the independent qualitative results of each gene, the method in this embodiment comprehensively considers the synergistic effects of all gene polymorphisms, resulting in more accurate medication decisions. Especially for complex cases with multiple gene mutations, the method in this embodiment can more accurately assess drug metabolism and efficacy, thus enabling the development of more rational individualized medication regimens.

Claims

1. A decision system for the detection of genetic polymorphisms in antihypertensive drugs by melting curve analysis, characterized in that, include: The data acquisition terminal acquires fluorescence melting curve data of hypertension-related multi-gene targets; The graph construction processor maps the melting temperature-fluorescence intensity derivative curve of each gene locus to nodes of a graph structure, and maps the melting feature associations between different gene loci to edges, thus constructing a multidimensional melting curve topology graph. The feature extraction processor extracts the co-fluctuation features of peak temperature offset and edge weight of nodes in the multidimensional melting curve topology graph, inputs the co-fluctuation features into the graph convolutional network, and outputs a polymorphic combination feature vector. The decision processor maps the polymorphic combination feature vector to a hypertension drug metabolism network graph model, calculates the joint decay coefficient of drug metabolic enzyme activity and receptor binding rate, and determines the type of hypertension drug and dosage adjustment parameters based on the joint decay coefficient, thus constituting the hypertension drug gene polymorphism melting curve detection decision system.

2. The hypertension pharmacogenetic polymorphism melting curve detection decision system according to claim 1, wherein, The graph construction processor maps the melting temperature-fluorescence intensity derivative curve of each gene locus to nodes of the graph structure, maps the melting feature associations between different gene loci to edges, and constructs a multidimensional melting curve topology graph, including: setting a sliding window along the temperature axis to extract local waveform segments of the melting temperature-fluorescence intensity derivative curve, and calculating the curvature extrema and area integral of the local waveform segments as attribute features of the nodes of the graph structure; The mutual information value between the local waveform segments corresponding to different gene loci is calculated, and the mutual information value is used as the weight of the edge between different nodes. The multidimensional melting curve topology graph is generated based on the attribute characteristics of the node and the weight of the edge.

3. The hypertension pharmacogenetic polymorphism melting curve detection decision system according to claim 1, wherein, The feature extraction processor extracts the co-fluctuation features of peak temperature offset and edge weights of nodes in the multidimensional melting curve topology graph, inputs the co-fluctuation features into a graph convolutional network, and outputs a polymorphic combination feature vector. This includes: in each layer of graph convolution operation of the graph convolutional network, aggregating the peak temperature offset and edge weights of the target node's neighboring nodes in the multidimensional melting curve topology graph, calculating the Hadamard product of the feature difference between the target node and its neighboring nodes and the edge weights, concatenating the Hadamard product with the features of the previous hidden layer of the target node and processing it through a nonlinear activation function, and then extracting high-order topological features as the polymorphic combination feature vector after propagation layer by layer.

4. The hypertension drug gene polymorphism melting curve detection and decision system according to claim 1, characterized in that, The decision processor maps the polymorphic combination feature vector to a hypertension drug metabolism network graph model and calculates the joint decay coefficient of drug-metabolizing enzyme activity and receptor binding rate, including: constructing a hypertension drug metabolism network graph model with drug-metabolizing enzymes and receptors as nodes and metabolic reaction pathways as edges, projecting the polymorphic combination feature vector to the node space of the hypertension drug metabolism network graph model, and obtaining the feature representation of each node. Based on the characteristic representation of each node, the inhibitory factor of the metabolic response pathway is calculated, and the inhibitory factor is accumulated by backpropagation along the edges of the hypertension drug metabolism network model to generate the joint attenuation coefficient characterizing the overall metabolic capacity.

5. The hypertension drug gene polymorphism melting curve detection and decision system according to claim 1, characterized in that, The decision processor determines the type and dosage adjustment parameters of antihypertensive drugs based on the joint attenuation coefficient, including: constructing an objective function of drug concentration and blood pressure reduction amplitude in a preset antihypertensive drug efficacy response surface space with the joint attenuation coefficient as a constraint. The objective function is solved by non-convex optimization to obtain the extreme point that makes the blood pressure reduction reach the target threshold and the drug concentration within the safe window; The drug identifier corresponding to the extreme point is extracted as the type of hypertension medication, and the concentration ratio corresponding to the extreme point is extracted as the dosage adjustment parameter.

6. The hypertension drug gene polymorphism melting curve detection and decision system according to claim 1, characterized in that, The data acquisition terminal acquires fluorescence melting curve data of hypertension-related multi-gene targets, including: acquiring time-series data of the original fluorescence signal intensity changing with temperature, performing wavelet transform decomposition on the time-series data, and separating the low-frequency baseline drift component and the high-frequency fluorescence noise component. Reconstruct the denoised fluorescence melting curve data from the decomposed data; The first and second derivatives of the denoised fluorescence melting curve data are calculated along the temperature axis, and the combined data stream of the first and second derivatives is input into the graph construction processor as the fluorescence melting curve data of the hypertension-related multi-gene target.

7. The hypertension drug gene polymorphism melting curve detection and decision system according to claim 2, characterized in that, Calculating the mutual information value between the local waveform segments corresponding to different gene loci, and using the mutual information value as the weight of the edge between different nodes, includes: performing kernel density estimation on the local waveform segments to obtain the edge probability density distribution and joint probability density distribution of the local waveform segments; Calculate the conditional entropy between the local waveform segments based on the marginal probability density distribution and the joint probability density distribution; The nonlinear dependency metric is calculated based on the difference between the conditional entropy and the edge entropy. The nonlinear dependency metric is set as the mutual information value. After normalization, the mutual information value is assigned to the corresponding edge in the multidimensional melting curve topology graph to form a weighted adjacency matrix.

8. The hypertension drug gene polymorphism melting curve detection and decision system according to claim 3, characterized in that, The Hadamard product is concatenated with the features of the previous hidden layer of the target node and processed by a nonlinear activation function. After propagation layer by layer, higher-order topological features are extracted as the polymorphic combination feature vector. This includes: introducing a multi-head attention mechanism to calculate the attention weight coefficients corresponding to each neighbor node of the target node, and using the attention weight coefficients to perform a weighted summation of the Hadamard product. The weighted summation result is fused with the features of the previous hidden layer of the target node through cross-channel feature fusion. The fused features are then input into a linear unit with leakage correction for nonlinear mapping. The cross-channel feature fusion and nonlinear mapping operations are iteratively executed until a preset number of propagation layers is reached, and the polymorphic combination feature vector with a fixed dimension is output.

9. The hypertension drug gene polymorphism melting curve detection and decision system according to claim 4, characterized in that, Based on the characteristic representation of each node, the inhibitory factor of the metabolic reaction pathway is calculated, and the inhibitory factor is accumulated by backpropagation along the edge of the hypertension drug metabolism network diagram model to generate the joint decay coefficient characterizing the overall metabolic capacity. This includes: correcting the enzyme reaction rate on the edge of the hypertension drug metabolism network diagram model based on the Michaelis kinetic equation, and determining the Michaelis constant correction term and the maximum reaction rate correction term based on the characteristic representation of each node. The inhibitory factor of the metabolic response pathway is calculated based on the Michaelis constant correction term and the maximum reaction rate correction term; The inhibitory factor is subjected to message passing and cumulative multiplication operations along the topological path of the hypertension drug metabolism network graph model, and the cumulative multiplication result is converted into a scalar value as the joint decay coefficient.

10. The hypertension drug gene polymorphism melting curve detection and decision system according to claim 5, characterized in that, The objective function is solved by non-convex optimization to obtain the extreme point where the blood pressure reduction reaches the target threshold and the drug concentration is within the safe window. This includes: using the Lagrange multiplier method to transform the safe window constraint of the drug concentration into a dual problem, and constructing a Lagrange augmented function in combination with the joint decay coefficient. The Lagrange augmented function is iteratively optimized using the quasi-Newton method to calculate the approximate value of the Hessian matrix and the gradient direction of the Lagrange augmented function; The drug concentration variable is updated along the gradient direction until the preset convergence criterion is met. The combination of drug concentration variables corresponding to the convergence point is output as the extreme point. The drug identifier and the dose adjustment parameter are obtained by parsing each drug component in the extreme point.

Images