CYP2C19 enzyme drug metabolism phenotype prediction method and device based on graph convolutional network

Abstract

According to the CYP2C19 enzyme drug metabolism phenotype prediction method and device based on the graph convolutional network, protein three-dimensional structure features are introduced by constructing a CYP2C19 wild type amino acid sequence residue contact graph, database information is integrated to generate a diplotype source domain / target domain data set, a residue contact graph and a source domain data set are utilized to train a graph convolutional network regression model, and a CYP2C19 enzyme drug metabolism phenotype prediction result is obtained. Migrating the trained model parameters to a five-classification graph convolutional network model by using a transfer learning algorithm; splitting a diploid feature matrix of the data set into two single allele feature matrixes, and respectively inputting two image volume integral branches together with the edge index matrix to respectively extract features; variation information to be analyzed is input into the five-classification graph convolutional network model subjected to target domain data set optimization training for prediction, the predicted CYP2C19 enzyme drug metabolism phenotype is output, and the problems that in a traditional method, the influence of a protein three-dimensional structure is ignored, multi-site variation prediction is difficult, and the model generalization and automation degree are low are effectively solved.

Description

Technical Field

[0001] This invention relates to gene variation prediction processing technology, specifically a method and device for predicting the metabolic phenotype of CYP2C19 enzyme drugs based on graph convolutional networks. Background Technology

[0002] In the fields of precision medicine and pharmacogenomics research, the CYP2C19 enzyme, as a key member of the cytochrome P450 superfamily, plays a decisive role in the metabolism of more than 50 commonly used clinical drugs, including clopidogrel, omeprazole, and citalopram. Amino acid variations in its gene sequence directly alter the enzyme's spatial conformation and catalytic activity, leading to differences in drug metabolism efficiency and potentially causing insufficient efficacy or adverse side effects. For example, patients carrying the CYP2C19*2 / *3 compound heterozygous variant have a significantly reduced ability to metabolize clopidogrel, and their incidence of adverse cardiovascular events is more than three times higher than that of wild-type patients. Furthermore, in weak metabolizers, the blood concentration of proton pump inhibitors is more likely to exceed the safety threshold, increasing the risk of gastrointestinal adverse reactions. Therefore, the drug metabolism phenotype of the CYP2C19 enzyme (such as normal metabolizer, intermediate metabolizer, weak metabolizer, etc.) plays a crucial role in guiding the development of personalized clinical medication regimens and drug research.

[0003] Due to the complexity and diversity of CYP2C19 variants, current databases only include metabolic phenotypes of a subset of amino acid variants. Therefore, metabolic phenotype prediction is needed for amino acid variants with unknown metabolic phenotypes. However, current metabolic phenotype prediction technologies based on CYP2C19 amino acid variants have three key limitations that make it difficult to meet actual clinical needs: First, it neglects the impact of protein three-dimensional structure on metabolic function. Traditional methods (such as the allele matching method and AS scoring method recommended by the CPIC guidelines) rely solely on the linear sequence characteristics of the CYP2C19 gene and determine the metabolic type by manually comparing known variant sites, without considering the disruption of the three-dimensional spatial conformation of enzyme proteins by amino acid mutations. For example, although a mutation may be located at an inactive site in the sequence, it may indirectly lead to the complete loss of enzyme activity by changing the spatial distance between neighboring residues (such as disrupting the pocket structure of the active site). This kind of "spatial synergistic effect" cannot be captured by sequence-level analysis.

[0004] Secondly, the predictive ability of multi-site variants is weak. With the updates to databases such as PharmVar, more than 120 stellate alleles have been identified for CYP2C19, many of which contain two or more amino acid variations (e.g., allele 23 contains the double mutations R286C and R431C). Traditional methods rely on known "variant combination-phenotype" correspondences for matching. For unannotated compound heterozygous variants (e.g., 2 / *23, *17 / *3), only a rough inference can be made based on the function of a single variant, with a prediction accuracy of less than 60%.

[0005] Third, the models have low generalization and automation levels. Existing prediction models are mostly trained on small sample clinical data and do not incorporate artificial intelligence technology for feature learning. On the one hand, when faced with newly added DMS (multiple effects of variation) data in the MAVEDB database, the models have difficulty quickly adapting to the feature distribution of unknown variations. On the other hand, the manual interpretation process is time-consuming (2-4 hours for a single analysis), which cannot meet the efficiency requirements of batch clinical sample testing.

[0006] Although machine learning techniques have been applied to drug metabolism prediction in recent years (such as prediction models based on random forests and artificial neural networks), these models still have significant shortcomings: either they fail to incorporate the three-dimensional structural features of proteins into the modeling process, or the limited clinical sample size in the target domain (usually less than 500 cases) leads to model overfitting, making it difficult to achieve effective transfer from simulated data to real clinical data. Therefore, there is an urgent need to develop a prediction technique that can integrate multi-dimensional features of "sequence-physicochemical-spatial" and has strong generalization ability, to overcome the limitations of current methods and improve the accuracy and practicality of CYP2C19 enzyme metabolic phenotype prediction. Summary of the Invention

[0007] The technical problem to be solved by the present invention is the defect of low prediction accuracy caused by the influence of protein three-dimensional structure in existing prediction methods. The present invention provides a method and device for predicting the metabolic phenotype of CYP2C19 enzyme drugs based on graph convolutional networks.

[0008] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a CYP2C19 enzyme drug metabolism phenotype prediction method based on graph convolutional networks, comprising the following steps: (1) Construct a residue contact diagram of the CYP2C19 wild-type amino acid sequence.

[0009] (2) Obtain the amino acid variation information of CYP2C19 from the database, generate the source domain dataset of diploids based on the haplotype variation function data in the amino acid variation information, and generate the target domain dataset of diploids based on the amino acid variation site information of the known single star allele and the star allele pairs of the known metabolic phenotype.

[0010] (3) Construct a graph convolutional network regression model and a five-class graph convolutional network model respectively. Train the graph convolutional network regression model using the residue contact map and the source domain dataset. Then, use the transfer learning algorithm to transfer the parameters of the trained graph convolutional network regression model to the five-class graph convolutional network model. Finally, use the residue contact map and the target domain dataset to optimize and train the five-class graph convolutional network model.

[0011] The feature extraction modules of both the graph convolutional network regression model and the five-class graph convolutional network model have two graph convolutional branches with the same structure. When training the graph convolutional network regression model and the five-class graph convolutional network model using the source domain dataset and the target domain dataset respectively, the double-type feature matrix in the dataset is split into two single-allelic feature matrices. Then, the two single-allelic feature matrices, together with the edge index matrix obtained based on the residue contact graph, are input into the two graph convolutional branches respectively. The features obtained after processing by the two graph convolutional branches are concatenated and used for the prediction task.

[0012] (4) Input the CYP2C19 amino acid variation information to be analyzed into the optimized five-class graph convolutional network model for prediction, and output the predicted CYP2C19 enzyme drug metabolism phenotype.

[0013] In step (1), the method for constructing the residue contact map of the CYP2C19 wild-type amino acid sequence is as follows: obtain the amino acid residue data of the CYP2C19 wild-type amino acid sequence from the database, and based on the C in the amino acid residues... β Atomic coordinates or C α Multiple residue pairs are constructed using atomic coordinates, and the Euclidean distance between each residue pair is calculated to generate a residue distance matrix. Based on the residue distance matrix, a cutoff distance is set to determine the spatial proximity relationship between two residues, and a residue contact map is constructed based on the spatial proximity relationship.

[0014] In step (2), the haplotype variation data in the amino acid variation information is copied and merged to form homozygous diploid data for generating source domain dataset.

[0015] In step (2), the amino acid variation site information of a single star allele is associated and matched with star allele pairs of known metabolic phenotypes, so that each allele pair corresponds to the variation information of two single star alleles, forming a double type data with a correspondence between double type amino acid variation information and metabolic phenotype labels, which is used to generate the target domain dataset.

[0016] The diploid data is feature-encoded to form a source domain dataset or a target domain dataset. The method is as follows: using the CYP2C19 wild-type amino acid sequence as a reference, according to the mutation position of each variant haplotype in the diploid data, the corresponding amino acid at the corresponding site is replaced. Then, each amino acid in the sequence is mapped to a unique integer value. The two haplotypes in the diploid data are encoded by integers and then merged into an integer encoding matrix. The integer encoding matrix is ​​one-hot encoded to form a sequence feature encoding matrix. The physicochemical characteristics of each amino acid are obtained, and the integer encoding matrix is ​​encoded by physicochemical characteristics to form a physicochemical feature encoding matrix. Then, the sequence feature encoding matrix and the physicochemical feature encoding matrix are concatenated and merged along the feature dimension to obtain the source domain dataset or the target domain dataset.

[0017] Principal component analysis was used to reduce the dimensionality of the obtained amino acid physicochemical features to obtain physicochemical features for encoding.

[0018] The graph convolutional network regression model and the five-class graph convolutional network model both include a residue contact graph processing module, an edge index expansion module, a feature extraction module, and a prediction module. The prediction module of the graph convolutional network regression model is used to predict the enzyme metabolic activity score, and the prediction module of the five-class graph convolutional network model is used to predict the five metabolic phenotypes of the enzyme.

[0019] The residue contact graph processing module is used to load the residue contact graph, generate an adjacency matrix, and convert the adjacency matrix into edge indices corresponding to each individual sample residue in the residue contact graph. The edge index expansion module is used to calculate the node ID offset of each sample according to the training batch size of the dataset, and to add the node ID offset to the value of each edge index, and then concatenate them along the column dimension to obtain the expanded edge index matrix.

[0020] The feature extraction module includes a first graph convolution branch and a second graph convolution branch with identical structures, as well as a feature merging module for feature splicing and merging. Two single allele feature matrices split from the source domain dataset or the target domain dataset are input together with the extended edge index matrix into the corresponding first graph convolution branch and second graph convolution branch for neighbor residue feature aggregation processing. The outputs of the first graph convolution branch and the second graph convolution branch are sent to the feature merging module for splicing and merging, and then passed to the prediction module.

[0021] A CYP2C19 enzyme drug metabolism phenotype prediction device based on graph convolutional networks includes: Residue contact map building unit, used to construct residue contact maps of the CYP2C19 wild-type amino acid sequence.

[0022] The dataset generation unit is used to obtain CYP2C19 amino acid variation information for training from the database and generate source domain datasets and target domain datasets.

[0023] The model training unit is used to construct a graph convolutional network regression model and a five-class graph convolutional network model. The graph convolutional network regression model is trained using the residue contact map and the source domain dataset. The weights and features of the trained graph convolutional network regression model are transferred to the five-class graph convolutional network model using a transfer learning algorithm. Then, the transferred five-class graph convolutional network model is optimized and trained using the residue contact map and the target domain dataset.

[0024] The prediction unit is used to input the CYP2C19 amino acid variation information to be analyzed into the optimized five-class graph convolutional network model for prediction, and output the predicted CYP2C19 enzyme drug metabolism phenotype.

[0025] The beneficial effects of this invention are as follows: By constructing a residue contact map of the CYP2C19 wild-type amino acid sequence to introduce the three-dimensional structural features of the protein, integrating database information to generate source / target domain datasets, and then training a graph convolutional network regression model, constructing a five-classification model through transfer learning, and optimizing model parameters with target domain data, the metabolic phenotype prediction of CYP2C19 amino acid variations is finally achieved. This effectively solves the problems of traditional methods ignoring the influence of protein three-dimensional structure, difficulty in predicting multi-site variations, and low model generalization and automation. It captures the indirect influence of mutations on enzyme spatial conformation through residue contact maps and overcomes the limitation of target domain sample size by relying on transfer learning, which significantly improves the prediction accuracy of the model for unannotated compound heterozygous variations. At the same time, the automated prediction process greatly shortens the analysis time, providing efficient and reliable technical support for clinical development of precision medicine plans based on CYP2C19 variations. Attached Figure Description

[0026] Figure 1 A flowchart of the CYP2C19 enzyme drug metabolism phenotype prediction method provided in the embodiments of the present invention; Figure 2 Distance matrix diagram of wild-type enzyme protein CYP2C19 residues provided in embodiments of the present invention; Figure 3 The residue contact diagram of wild-type enzyme protein CYP2C19 provided in the embodiments of the present invention, constructed based on the cutoff distance; Figure 4 The interaction flowchart of each module of the graph convolutional network provided in this embodiment of the invention; Figure 5 This is a flowchart of the dual-branch graph convolution module provided in an embodiment of the present invention; Figure 6 The MSE variation with epoch of the model provided in this embodiment of the invention in the source domain dataset; Figure 7The Pearson r variation with epoch for the model provided in this embodiment of the invention in the source domain dataset; Figure 8 A graph showing the accuracy of the model provided in this embodiment of the invention as a function of epochs in the target domain dataset; Figure 9 The F1-Score of the model provided in this embodiment of the invention varies with epoch in the target domain dataset; Figure 10 This is a schematic diagram of the functional modules of the CYP2C19 enzyme drug metabolism phenotype prediction device provided in an embodiment of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. The specific contents listed in the following embodiments are not limited to the technical features necessary to solve the technical problem of the present invention. Furthermore, the listed embodiments are merely a part of the present invention, and not all embodiments.

[0028] like Figure 1 As shown, this invention provides a method for predicting the metabolic phenotype of CYP2C19 enzyme drugs based on graph convolutional networks, comprising the following steps: S1. Construct a residue contact diagram of the CYP2C19 wild-type amino acid sequence; S2. Obtain CYP2C19 amino acid variation information for training from the specified database, and generate source domain dataset and target domain dataset; S3. Construct a graph convolutional network regression model. Train the graph convolutional network regression model using the residue contact graph and the source domain dataset to obtain the parameters and features of the trained graph convolutional network regression model. S4. Construct a five-class graph convolutional network model, and use the transfer learning algorithm to transfer the weights and features of the trained graph convolutional network regression model to the five-class graph convolutional network model to obtain the transferred five-class graph convolutional network model. S5. Input the target domain dataset into the transferred five-class convolutional graph network model to adjust the model parameters and obtain the optimized five-class convolutional graph network model. S6. Input the CYP2C19 amino acid variation information to be analyzed into the optimized five-class graph convolutional network model for prediction, and output the predicted CYP2C19 enzyme drug metabolism phenotype.

[0029] The following describes the methods for each step in detail with reference to the embodiments: I. Constructing the residue contact diagram of the CYP2C19 wild-type amino acid sequence: The PDB file containing the wild-type amino acid sequence of CYP2C19 was obtained from the AlphaFold Protein Structure Database. Multiple amino acid residues were read from the PDB file, and C10 residues were extracted from each residue. β Atomic coordinates or C α Atomic coordinates, based on all extracted C β Atomic coordinates or C α Multiple residue pairs are constructed using atomic coordinates. The Euclidean distance between each residue pair is calculated. A residue distance matrix is ​​generated based on all the Euclidean distances. A cutoff distance is set based on the residue distance matrix. The spatial proximity relationship between two residues is determined based on the cutoff distance. A residue contact map is constructed based on the spatial proximity relationship.

[0030] Specifically, the formula for calculating the Euclidean distance between each residue pair is: , in, This represents the Euclidean distance between the i-th residue and the j-th residue, which is used to determine the spatial proximity of residue pairs in subsequent tests. Indicates C in the i-th amino acid residue β Atom (or C) α Atom, when the residue is glycine, use C. α The three-dimensional coordinates of the atom are extracted from the PDB file of the CYP2C19 wild-type amino acid sequence to represent the position of the residue in the three-dimensional space of the protein. This represents the atom (C) corresponding to the j-th amino acid residue. β Or C α The three-dimensional coordinates of ) and Similarly, they are used together to calculate the spatial distance between two residues.

[0031] By extracting the three-dimensional coordinates of residue atoms and calculating the distances using formulas, the spatial relationships of residues can be determined. This provides a structural basis for constructing residue contact maps and subsequent analysis of the impact of CYP2C19 enzyme variations on metabolic phenotypes based on graph convolutional networks. The constructed residue contact maps can accurately capture the spatial relationships between residues (such as the spatial distance between active site residues and neighboring residues), providing a crucial "structural framework" for subsequent graph convolutional network models. This allows the models to further learn the "indirect effects of mutations on the enzyme's three-dimensional conformation" (such as mutations at inactive sites disrupting enzyme function by altering the spatial relationships of neighboring residues). This lays a structural foundation for overcoming the weakness of traditional methods in predicting multi-site variations and unannotated variations, and also provides core technical support for improving the accuracy of CYP2C19 enzyme drug metabolic phenotype prediction.

[0032] II. Generating source domain datasets and target domain datasets: 1. Source domain dataset generation: CYP2C19 amino acid variation information for training was obtained from the MAVEDB database, including haplotype variation functional data. The MAVEDB database, through VAMP-seq experiments, measured the effects of missense and nonsense mutations in the CYP2C19 gene on protein function and provided standardized scores for enzyme activity after mutation. This database compiled each mutation and its corresponding enzyme activity score (0 for loss of function, 1 for normal function, and >1 for enhanced function) into DMS data. The DMS data recording CYP2C19 gene missense and nonsense mutations in the MAVEDB database were downloaded to obtain relevant haplotype variation functional data for source domain dataset generation.

[0033] The haplotype variation data in the amino acid variation information is copied and merged to form source domain homozygous diploid data. The method is as follows: First, the original data is copied as the basis for homozygous mutations. The original DataFrame is copied using `homozygous_df = df.copy()`, retaining all haplotype information; subsequent modifications are only made to homozygous characteristics. Second, the variant name (variant column) is modified using relevant code to reflect homozygous characteristics. For example, if the original haplotype variation is *2 (representing a specific variation of a certain allele), the variant name for the homozygous mutation will become *2 / *2; if the original variation is C185T (a base mutation), it will become C185T / C185T, i.e., converted to homozygous. Then, the mutation count column is adjusted. Homozygous mutations mean that each chromosome carries one identical mutation, therefore the total number of mutations is twice that of the haplotypes.

[0034] Based on the CYP2C19 wild-type amino acid sequence (490 units in length), according to the mutation position of each variant haplotype in the source domain diploid data, the corresponding amino acid is replaced. Then, each amino acid in the sequence is mapped to a unique integer value (e.g., alanine corresponds to 1, valine corresponds to 2). The two haplotypes in the diploid data are encoded into integers and then merged into a three-dimensional tensor, generating a shape of... The integer encoding matrix, where, denoted as the number of samples in the source domain dataset, M as the number of alleles, and P as the length of the amino acid sequence.

[0035] The integer encoding matrix is ​​sequence feature encoded using the one-hot encoding method to form a sequence feature encoding matrix, which represents the integer values ​​of amino acids using one-hot encoding. Taking 20 common amino acids as an example, with the addition of a terminator, the one-hot representation forms a binary vector of length 21, resulting in a total of 21 dimensions, thus generating a sequence feature encoding matrix of shape N1×M×P×21.

[0036] Based on the physicochemical characteristics of each amino acid obtained from the aa-index database, the integer coding matrix is ​​then encoded with physicochemical features, that is, each amino acid is assigned biochemical properties (such as hydrophobicity, isoelectric point, etc.). Originally, the aa-index database contained 114 different amino acid attribute indicators, each describing a certain characteristic of the amino acid. Principal component analysis (PCA) was used for dimensionality reduction, that is, the original features were transformed into a set of linearly uncorrelated principal components through orthogonal transformation, resulting in 19 attribute indicators retained, forming a binary sequence of length 19, generating a physicochemical feature coding matrix of shape N1×M×P×19.

[0037] The sequence feature encoding matrix and the physicochemical feature encoding matrix are merged along the feature dimension to obtain a shape of... The feature matrix is ​​given by F, where F is the sum of the features of the sequence feature code and the physicochemical feature code, i.e., 40. The source domain dataset is obtained based on the feature matrix.

[0038] 2. Target domain dataset generation: We retrieved all known single stellate alleles of CYP2C19 amino acids and their corresponding amino acid variation sites from the PharmVar database, and all known metabolic phenotypes of CYP2C19 amino acids from the CPIC database. The PharmVar database includes single CYP2C19 stellate alleles obtained from the literature, each labeled with specific amino acid variation information; we collected the amino acid variation information for each stellate allele. The CPIC (Clinical Pharmacogenomics Implementation Consortium) database categorized the metabolic phenotypes of CYP2C19 biallelic amino acid variation information into normal metabolizers, intermediate metabolizers, possibly intermediate metabolizers, weak metabolizers, possibly weak metabolizers, and indeterminate metabolic phenotypes based on activity scores; we collected the metabolic phenotype data for CYP2C19 biallelic variation information from these categories.

[0039] The amino acid variation information of a single stellate allele is associated and matched with stellate allele pairs of known metabolic phenotypes, so that each allele pair corresponds to the variation information of two single stellate alleles, forming a correspondence between diploid amino acid variation information and metabolic phenotype tags, thus obtaining diploid data of the target domain.

[0040] Specifically, firstly, the PharmVar database contains 39 CYP2C19 alleles. Since this design involves the residue contact map of the enzyme protein CYP2C19, only amino acid variation information was collected. Variation information of 19 star alleles was collected, resulting in monoallelic amino acid variation data.

[0041] Secondly, based on the amino acid variation information of the 19 stellate alleles collected, the biallelic information in the CPIC database was screened to select biallelic combinations that contain only these 19 stellate alleles, thus obtaining biallelic phenotypic data.

[0042] Then, monoallelic amino acid variation data and bialelic phenotype data are read, diploid types are split, and data association technology is used to integrate genotype-mutation information from different data sources to obtain a diploid dataset that includes not only amino acid variation information but also the corresponding metabolic phenotype.

[0043] Finally, the one-hot encoding method and the aa_index database were selected to convert the double-sized dataset into a data matrix, following the same method as the source dataset.

[0044] In this embodiment, CYP2C19 amino acid variation information is accurately obtained from databases such as MAVEDB, PharmVar, and CPIC. First, a source domain dataset is constructed using MAVEDB haplotype variation data. This dataset undergoes multi-dimensional encoding, including integer encoding, one-hot encoding, and physicochemical feature encoding, and features are merged to provide rich simulated data for model training. Then, variation information associated with known metabolic phenotypes is acquired to construct a target domain dataset, achieving coverage from simulation to real clinical data. This solves the problems of single sample features and poor clinical data adaptability in traditional methods. The multi-database integration and encoding method employed ensures that the dataset covers sufficient variation scenarios and aligns with clinical phenotypes, providing high-quality, multi-dimensional data support for subsequent model training. This helps overcome the limitations of traditional prediction methods in predicting multi-site variations and unknown variations, improving the accuracy and generalization of CYP2C19 enzyme drug metabolic phenotype prediction.

[0045] III. Constructing a Graph Convolutional Network Regression Model The graph convolutional network regression model includes a residue contact graph processing module, an edge index expansion module, a two-branch graph convolution module, a feature merging module, and a fully connected metabolic score prediction module. The residue contact graph processing module and the edge index expansion module are data preprocessing modules, the two-branch graph convolution module and the feature merging module are feature extraction modules, and the fully connected metabolic score prediction module is used to output the prediction results.

[0046] The residue contact map processing module is used to load residue contact maps and generate dimensions of [missing information]. The adjacency matrix is ​​converted into edge indices corresponding to each individual sample residue in the residue contact graph. These edge indices are represented as follows: Where 2 represents a binary structure, meaning that each edge needs to store two pieces of information: the source residue number and the target residue number. P represents the number of residues linked in a single sample, where P is the length of the CYP2C19 amino acid sequence. P is typically set to 490. Matrix elements "1" indicate that there is a spatial proximity between residues, and "0" indicates that there is no proximity.

[0047] The edge index expansion module calculates the node ID offset for each sample based on the training batch size of the source domain dataset, adds the node ID offset to the value of each edge index, and then concatenates them along the column dimension using the `torch.cat` function to obtain the expanded edge index matrix. The expanded edge index matrix is ​​represented as follows: ,in, This refers to the number of samples input at one time during training. This refers to the number of residue connections per sample, i.e., the total number of edges in a single sample (such as a CYP2C19 amino acid sequence); for example, when batch size = 16 and number of nodes per sample = 490, the offset is [0, 490, 980, ..., 16×490-490].

[0048] The dual-branch graph convolution module includes two identical first graph convolution branches and a second graph convolution branch. It splits the feature matrix of the source domain dataset into two single allele feature matrices. The two single allele feature matrices and the extended edge index matrix are respectively input into the corresponding first graph convolution branch and second graph convolution branch for neighbor residue feature aggregation processing. The aggregated features are then compressed, and finally the first graph convolution branch and the second graph convolution branch output the allele feature matrix respectively.

[0049] The feature merging module is used to concatenate two allele feature matrices along the node dimension to obtain a fused feature matrix.

[0050] The fully connected metabolic scoring prediction module includes a Flatten layer, two fully connected layers, and a linear activation layer. The Flatten layer is used to flatten the fused feature matrix into a two-dimensional feature vector. Corresponding neuron parameters are set for the two fully connected layers. The first fully connected layer performs nonlinear feature transformation on the two-dimensional feature vector based on the set neuron parameters and through the ReLU activation function. The second fully connected layer scores the enzyme metabolic activity of the source domain samples based on the set neuron parameters and through linear activation, thereby completing the training of the graph convolutional network regression model.

[0051] In one embodiment of the present invention, each graph convolution branch of the dual-branch graph convolution module includes two GCN convolutional layers, a batch normalization layer, and a ReLU activation layer. The two branches use independent batch normalization layers to avoid feature distribution confusion. The framework of the CYP2C19 enzyme drug metabolism phenotype prediction model in this embodiment is shown in Table 1: Combination Figure 4 and Figure 5 As shown, the dual-branch graph convolution module first splits the input biallelic data (batch×2×490×40) into a batch×490×40 shape. Then, the feature matrices and edge index matrices of the two single alleles are input together into the corresponding graph convolution branches. For example, the feature matrix and edge index matrix of allele 1 are input into the first graph convolution branch, and the feature matrix and edge index matrix of allele 2 are input into the second graph convolution branch. The two graph convolution branches extract the features of the two alleles respectively. In each branch graph convolution module, the first layer GCN maps the 40-dimensional node features to 64 dimensions and aggregates neighbor information. The second layer GCN further compresses the features to 32 dimensions. The aggregated neighbor information includes: first, traversing all neighbors j of node i (defined by the edge index edge_index) and obtaining their features h. j (l), then perform a linear transformation on each neighbor feature (multiply by the weight matrix W(l)), and then based on the degree (d) of node i and neighbor j. i and d j The transformed features are scaled to ensure fair participation in the computation for nodes with different connection densities. Then, all weighted neighbor features are summed, and a new feature for node i is generated using the activation function σ. .

[0052] The GCN neighbor aggregation formula, which inputs the two monoallelic feature matrices and the extended edge index matrix into the corresponding first and second graph convolution branches for neighbor residue feature aggregation processing, is as follows: in, This represents the feature vector of the i-th residue in the CYP2C19 amino acid sequence at the (l+1)-th layer of the graph convolutional network. Represents the ReLU activation function, used to perform nonlinear transformations on residue features. Let i be the set of spatially neighboring residues of the i-th residue. These represent the number of residues contained in residue i and residue j, respectively, used to normalize neighbor features. This represents the learnable weight matrix of the l-th layer of a graph convolutional network. This represents the feature vector of the j-th residue in layer l, which serves as neighbor information in the feature update of residue i. For example, the value of i ranges from 1 to 490, corresponding to the 490 amino acid residues of CYP2C19.

[0053] After feature extraction by the bi-branch graph convolution module, the CYP2C19 enzyme protein sequence data is converted into a matrix of size batch×490×32. Subsequently, the features of the two alleles are merged and concatenated, converting the data into a matrix of size batch×980×32.

[0054] Finally, the merged and spliced ​​feature map is input into the fully connected prediction module to complete the prediction task for the sequence data. This prediction module consists of one flattening layer and two fully connected layers: the flattening layer reduces the dimensionality of the three-dimensional tensor data to two-dimensional tensors so that it can be input into the subsequent fully connected layers. The output of the flattening layer is batch×(980×32); the first fully connected layer contains 100 neurons, and the second layer contains 1 neuron, which is used to output the metabolic score.

[0055] When training the graph convolutional network regression model using the source domain dataset, the batch size was set to 16 (i.e., training 16 double-batch datasets at a time), and the tensor shape was [16, 2, 490, 40]. The learning rate was set to 0.0001 as the initial learning rate, and the ReduceLROnPlateau method was used to dynamically adjust the learning rate. If the validation set loss did not improve for 10 consecutive epochs, the learning rate was automatically reduced by one order of magnitude. The mean squared error loss function was chosen. The Adam method was used as the optimizer. Furthermore, an early stopping mechanism was employed during training to prevent overfitting. The changes in the model's loss function MSE and Pearson r with epochs in the source domain dataset are shown below. Figure 6 and Figure 7 As shown.

[0056] In this embodiment, a graph convolutional network regression model is constructed, incorporating modules such as residue contact graph processing, edge index expansion, and bi-branch graph convolution. Trained using residue contact graphs and source domain datasets, this model addresses the limitations of traditional methods that fail to consider the three-dimensional structure of proteins and struggle to handle multi-site variations. By adapting the bi-branch graph convolution structure to the learning of two single allele features in diploid data, and by adapting edge index expansion to batch training, the collaborative work of multiple modules enables the model to accurately aggregate neighboring residue features and fuse allele information, outputting enzyme metabolic activity scores. This lays the foundation for subsequent metabolic phenotype prediction, improves predictive capabilities in complex variation scenarios, and overcomes the limitations of traditional models. By inputting the two monoallelic feature matrices and the extended edge index matrix into the convolutional branches of the first and second graphs respectively, and performing neighbor residue feature aggregation based on the GCN neighbor aggregation formula, nonlinear feature transformation is achieved using the ReLU activation function. Residue degree normalization is combined to avoid feature aggregation bias. The correlation between residue features and enzyme activity is captured through a learnable weight matrix. This effectively solves the problems of traditional methods that cannot take into account the features of CYP2C19 bicelestems and ignore the influence of residue spatial proximity on enzyme function. The model can accurately aggregate the neighbor residue features of each bicelestem and fully learn the mapping law of "residue spatial correlation - enzyme metabolic activity". This lays a key feature learning foundation for the subsequent construction of a five-classification model through transfer learning and improving the accuracy of metabolic phenotype prediction. It also breaks through the limitation of traditional models that rely solely on linear sequence features and have weak prediction ability for complex variations.

[0057] IV. Constructing a Five-Classification Graph Convolutional Network Model The five-class graph convolutional network model's module composition (residue contact graph processing module, edge index expansion module, bi-branch graph convolution module, and feature merging module) is completely consistent with the graph convolutional network regression model structure. Only the "fully connected prediction module" is adapted and adjusted to obtain the fully connected metabolic phenotype prediction module. The fully connected metabolic phenotype prediction module retains the Flatten layer (functionally consistent with the regression model) and the first fully connected layer (100 neurons + ReLU activation). The second fully connected layer has 5 neurons (corresponding to five metabolic phenotypes: normal metabolizer, intermediate metabolizer, possibly intermediate metabolizer, weak metabolizer, and possibly weak metabolizer). The output is converted into probability values ​​for each phenotype (the sum of probabilities is 1) through the Softmax activation function, achieving classification prediction.

[0058] Using the transfer learning algorithm, the weights and learned features of the trained graph convolutional network regression model are transferred to the five-class graph convolutional network model to obtain the transferred five-class graph convolutional network model.

[0059] V. Optimization and Training of the Five-Classification Graph Convolutional Network Model The target domain dataset is input into the transferred five-class graph convolutional network model for optimization training. The model parameters are adjusted, and the training strategy includes: The training process utilizes transfer learning, transferring the model weights and learned features obtained from the pre-trained model to the new model, and updating the model weights with an initial learning rate of 0.00001. Five-fold cross-validation is used to evaluate model performance on the validation set, which helps to select the best combination of hyperparameters (such as learning rate, batch size, etc.). By using an early stopping mechanism, the loss on the validation set is monitored, and the optimal model state is saved.

[0060] Data partitioning: Divide the target domain dataset into a "training set" and a "validation set" in a 7:3 (or 8:2) ratio. The training set is used for parameter updates, and the validation set is used to evaluate the model's generalization ability.

[0061] It should be understood that five-fold cross-validation is used to make full use of the limited samples. The dataset is divided into 5 parts, and 4 parts are used as the training set and 1 part as the validation set each time. This is repeated 5 times, and the average of the 5 validation results (such as average accuracy and average F1-Score) is taken as the final performance index of the model. This also helps to select the optimal hyperparameters (such as learning rate and batch size).

[0062] Training convergence judgment: When the "accuracy" and "F1-Score" of the training set and validation set are stable for several consecutive epochs (without significant increase or decrease), and the validation set loss is stable, the model is judged to have converged, training is stopped, and the "trained and optimized target domain five-class classification model" is saved (as shown in Figure 8). Figure 9 As shown, the model's accuracy and F1-score tend to stabilize after approximately 80 epochs.

[0063] The core of transfer learning lies in adapting the prior representations learned from the source domain to the target domain through parameter sharing and knowledge transfer. Specifically, the low-level weights of the source model's graph convolutional layers (such as conv1-1 and conv2-1) are reused to preserve general structural features, while some deep convolutional layers (such as conv2-1 and conv2-2) are partially unfrozen for fine-tuning. The weights in the fully connected layer CYP2C19 enzyme drug metabolism phenotype prediction module are changed with a small learning rate, ultimately completing the training task of transfer learning.

[0064] The target domain dataset is divided into a training set and a validation set according to a set ratio. The labels corresponding to the features in the training set and validation set are encoded, and the training set and validation set are loaded in batches using a data iterator.

[0065] The difference between the predicted probability and the true label is calculated using a multi-class cross-entropy loss function. Based on the difference, the training parameters of the transferred five-class graph convolutional network model are adjusted using the Adam optimizer. The adjusted five-class graph convolutional network model is iteratively trained based on the set parameter update range and the loaded training set. The loss value of each training round is recorded, and the average loss value of n consecutive rounds is calculated. If the average loss value is continuously and stably below a preset threshold, the optimized five-class graph convolutional network model is obtained.

[0066] When training a five-class graph convolutional network model using the target domain dataset, the batch size was set to 16 (i.e., training 16 double-class datasets at a time), and the tensor shape was [16, 2, 490, 40]. The learning rate was set to 0.0001 as the initial learning rate, and the ReduceLROnPlateau method was used to dynamically adjust the learning rate. If the validation set loss did not improve for 10 consecutive epochs, the learning rate was automatically reduced by one order of magnitude. The multi-class cross-entropy loss function was chosen. The Adam method was used as the optimizer. Furthermore, an early stopping mechanism was employed during training to prevent overfitting. The model's accuracy and F1 score on the target domain dataset as a function of epochs are shown below. Figure 8 and Figure 9 As shown.

[0067] In this embodiment, the target domain dataset is divided into training and validation sets according to a set ratio, labels are encoded, and data is loaded in batches. Then, a multi-class cross-entropy loss function is used to calculate the difference between the predicted and true labels, and the Adam optimizer is used to adjust the parameters of the transferred model. Simultaneously, the parameter update range is defined for iterative training, and the convergence criterion is that the average loss value stabilizes below a preset threshold. This effectively solves the problems of overfitting and poor generalization caused by limited clinical sample size and lack of targeted parameter adjustment in traditional models in the background technology. This process adapts to the characteristics of the target domain data through reasonable data partitioning and batch loading, and achieves targeted fine-tuning of model parameters through precise loss calculation and optimization strategies. This ensures that the transferred model can fully adapt to the feature distribution of real clinical data, significantly improving the stability and accuracy of CYP2C19 enzyme drug metabolism phenotype prediction, and overcoming the limitations of traditional methods in predicting insufficient performance in small sample clinical data scenarios.

[0068] VI. Metabolic Phenotypic Prediction of CYP2C19 Amino Acid Variation Information to be Analyzed Preprocessing of variant information to be analyzed (adapting to model input format): First, extract the amino acid variant information of the CYP2C19 bicelestem of the sample to be analyzed (e.g., allele 1 contains the R286C mutation and allele 2 contains the R431C mutation). Based on the CYP2C19 wild-type amino acid sequence (length 490), replace the corresponding mutation sites with variant amino acids. Then, map each amino acid to a unique integer value (e.g., alanine corresponds to 1 and valine corresponds to 2). Encode the two alleles separately with integers and merge them to generate an integer encoding matrix of shape 1×2×490 (1 is the number of samples to be analyzed, 2 is the number of alleles, and 490 is the sequence length).

[0069] Multi-dimensional feature encoding (matching model training feature dimensions): Based on the one-hot encoding method, sequence feature encoding is performed on the integer encoding matrix to generate a 1×2×490×21 one-hot sequence feature encoding matrix (including 20 common amino acids, plus a terminator); at the same time, the physicochemical characteristics (such as hydrophobicity, isoelectric point, etc.) of 20 amino acids are obtained from the aa-index database, and the integer encoding matrix is ​​encoded with physicochemical features. The physicochemical feature dimension is compressed to 19 through PCA dimensionality reduction, resulting in a 1×2×490×19 physicochemical feature matrix; finally, the two feature matrices are merged along the feature dimensions to generate a 1×2×490×40 feature matrix of the sample to be analyzed (40 is the total number of sequence and physicochemical features, consistent with the feature dimension during model training).

[0070] Structural input preparation (supplementing residue spatial association information): Load the CYP2C19 wild-type residue contact map consistent with that used during optimized model training. Generate a 490×490 adjacency matrix through the residue contact map processing module, and convert it into the edge index (shape) corresponding to a single sample. ).

[0071] Forward propagation of the model (feature extraction and prediction): The feature matrix of the sample to be analyzed and the edge index are simultaneously input into the optimized five-class graph convolutional network model: First, the feature matrix of the two alleles is processed by the two branches of the two-branch graph convolution module, and the features of neighboring residues are aggregated based on the edge index. After the GCNConv layer, BN layer and ReLU activation, the allele feature matrix is ​​output. Then, the two allele features are concatenated by the feature merging module to obtain a 1×980×32 fused feature matrix. Finally, it is flattened into a two-dimensional feature vector by the Flatten layer and input into the fully connected metabolic phenotype prediction module. The probability values ​​of five metabolic phenotypes (normal metabolizer, intermediate metabolizer, possible intermediate metabolizer, weak metabolizer, and possible weak metabolizer) are output by the Softmax activation function.

[0072] Metabolic phenotype determination (outputting the final prediction result): Select the category with the highest probability value output by the model as the predicted metabolic phenotype of the sample to be analyzed. For example, if the probability corresponding to "intermediate metabolizer" is 0.82 (higher than the probability of other categories), then the CYP2C19 enzyme drug metabolism phenotype of the sample will be output as "intermediate metabolizer". The specific probability values ​​of each category can also be output to provide a reference for clinical judgment.

[0073] The above embodiments provide a method for predicting the metabolic phenotype of CYP2C19 enzyme drugs based on graph convolutional networks, which can predict the metabolic phenotype of CYP2C19 based on multi-site amino acid variation information. To fully utilize training samples and address the problem of model overfitting, this invention employs five-fold cross-validation to train the five-class graph convolutional network model. Through a highly accurate five-class graph convolutional network model, it can predict the drug metabolic phenotype of CYP2C19 based on multi-site variation combinations in the protein spatial structure. This helps provide a scientific basis for clinical drug use at the genetic level, prevent adverse drug reactions, and achieve precision medicine.

[0074] The CYP2C19 enzyme drug metabolism phenotype prediction method of this invention can be widely applied to the metabolic function assessment of drugs that require CYP2C19 enzyme metabolism (such as the antidepressant clomipramine, the proton pump inhibitor omeprazole, and the antiplatelet drug clopidogrel). Since CYP2C19 gene polymorphisms (such as mutation sites 1, 2, and 3) significantly affect drug metabolism rates, leading to individual differences in metabolic phenotypes (such as normal metabolizers, intermediate metabolizers, and weak metabolizers), this method integrates mutant sequence characteristics and structural information through a graph convolutional network (GCN) to accurately predict the patient's metabolic phenotype. Finally, based on the metabolic phenotype label (such as normal / intermediate / weak metabolizer), personalized medication regimens and dosage adjustment strategies are recommended.

[0075] This invention also provides a CYP2C19 enzyme drug metabolism phenotype prediction device based on graph convolutional networks, such as... Figure 10 As shown, it includes: Residue contact map building unit, used to construct residue contact maps of the CYP2C19 wild-type amino acid sequence; The dataset generation unit is used to obtain CYP2C19 amino acid variation information for training from the database and generate source domain datasets and target domain datasets. The model training unit is used to construct a graph convolutional network regression model and a five-class graph convolutional network model. The graph convolutional network regression model is trained using the residue contact map and the source domain dataset. The weights and features of the trained graph convolutional network regression model are transferred to the five-class graph convolutional network model using a transfer learning algorithm. Then, the transferred five-class graph convolutional network model is optimized and trained using the residue contact map and the target domain dataset. The prediction unit is used to input the CYP2C19 amino acid variation information to be analyzed into the optimized five-class graph convolutional network model for prediction, and output the predicted CYP2C19 enzyme drug metabolism phenotype.

[0076] Based on the above technical content, the present invention achieves the following significant advantages in solving the problem of predicting the metabolic phenotype of CYP2C19 enzyme drugs: (1) Improve prediction accuracy By integrating multi-dimensional features and graph convolutional networks, this method overcomes the limitations of traditional methods that rely solely on sequence features for training, significantly improving the prediction accuracy of CYP2C19 enzyme drug metabolism phenotypes.

[0077] (2) Capturing spatial structural features Using graph convolutional networks to capture the spatial structural features of CYP2C19 mutants can more accurately reflect the impact of mutations on enzyme protein structure, thereby more accurately predicting metabolic phenotypes.

[0078] (3) Handling multisite variations The method of this invention can handle multi-site mutations on CYP2C19 bicelestes. It extracts the topological features of the two alleles through a bibranch graph convolution module, thereby enabling synergistic effect analysis of compound heterozygous variants.

[0079] (4) Optimize the training process We employ a five-fold cross-validation method and an early stopping mechanism to further optimize model parameters, avoid overfitting, and improve the model's generalization ability.

[0080] This invention establishes an efficient and accurate CYP2C1 enzyme drug metabolism phenotype prediction system through transfer learning, fusion of multi-dimensional features, application of graph convolutional networks, design of bi-branch graph convolutional modules, and optimization of the training process, providing strong support for precision medicine in clinical practice.

[0081] The above description of specific embodiments is only for the purpose of helping to understand the technical concept and core idea of ​​the present invention. Although specific preferred embodiments have been used to describe and illustrate the technical solutions, they should not be construed as limiting the present invention itself. Those skilled in the art can make various changes in form and detail without departing from the technical concept of the present invention. These easily conceived changes or substitutions should all be covered within the protection scope of the present invention.

Claims

1. A method for predicting CYP2C19 enzyme pharmacometabolic phenotype based on a graph convolution network, characterized in that: The method comprises the following steps: (1) constructing a residue contact map of a wild-type amino acid sequence of CYP2C19; (2) obtaining amino acid variation information of CYP2C19 from a database, generating source domain data set according to haplotype variation function data in the amino acid variation information, and generating target domain data set according to amino acid variation site information of a known single star-like allele and a known metabolic phenotype of a star-like allele pair; (3) constructing a graph convolution network regression model and a five-class graph convolution network model respectively, training the graph convolution network regression model by using the residue contact map and the source domain data set, and migrating the parameters of the trained graph convolution network regression model to the five-class graph convolution network model by using a transfer learning algorithm, and then optimizing and training the five-class graph convolution network model by using the residue contact map and the target domain data set; wherein the feature extraction modules of the graph convolution network regression model and the five-class graph convolution network model both have two graph convolution branches with the same structure, when training the graph convolution network regression model and the five-class graph convolution network model by using the source domain data set and the target domain data set respectively, the haplotype feature matrix in the data set is split into two single-allele feature matrices, then the two single-allele feature matrices are respectively input into the two graph convolution branches together with the edge index matrix obtained based on the residue contact map, and the features obtained after processing by the two graph convolution branches are spliced and combined for a prediction task; (4) inputting the CYP2C19 amino acid variation information to be analyzed into the five-class graph convolution network model optimized and trained to perform prediction, and outputting the predicted CYP2C19 enzyme drug metabolic phenotype. 2.The CYP2C19 enzyme metabolic phenotyping prediction method based on a graph convolution network according to claim 1, wherein: In step (1), the method for constructing the residue contact map of the CYP2C19 wild-type amino acid sequence is as follows: obtaining the amino acid residue data of the CYP2C19 wild-type amino acid sequence from a database, determining the C β atomic coordinates or C α atomic coordinates, constructing a plurality of residue pairs, calculating the Euclidean distance between each residue pair, generating a residue distance matrix, setting a cutoff distance based on the residue distance matrix, determining the spatial proximity relationship between two residues, and constructing a residue contact map according to the spatial proximity relationship. 3.The CYP2C19 enzyme metabolic phenotyping prediction method based on graph convolution network according to claim 1, wherein: In the step (2), the haplotype variation function data in the amino acid variation information is copied and combined to form homozygous haplotype data for generating the source domain data set. 4.The CYP2C19 enzyme metabolic phenotyping prediction method based on a graph convolution network according to claim 1, wherein: In the step (2), the amino acid variation site information of a single star-like allele is associated and matched with a known metabolic phenotype of a star-like allele pair, so that each allele pair corresponds to the variation information of two single star-like alleles, forming haplotype data having a corresponding relationship between haplotype amino acid variation information and metabolic phenotype labels, for generating the target domain data set. 5.The CYP2C19 enzyme metabolic phenotyping prediction method based on graph convolution network according to claim 3 or 4, characterized in that: The haplotype data is feature-encoded to form the source domain data set or the target domain data set, and the method is as follows: taking the wild-type amino acid sequence of CYP2C19 as a reference, replacing the amino acids at the corresponding positions according to the mutation positions of each variation haplotype in the haplotype data, then mapping each amino acid in the sequence to a unique integer value, and merging the integer encodings of the two haplotypes in the haplotype data into an integer encoding matrix; one-hot encoding the integer encoding matrix to form a sequence feature encoding matrix; obtaining the physicochemical characteristics of each amino acid, and physicochemical feature encoding the integer encoding matrix to form a physicochemical feature encoding matrix; and then splicing and combining the sequence feature encoding matrix and the physicochemical feature encoding matrix along the feature dimension to obtain the source domain data set or the target domain data set. 6.The CYP2C19 enzyme metabolic phenotyping prediction method based on a graph convolution network according to claim 5, wherein: The principal component analysis method is used to reduce the dimension of the obtained amino acid physicochemical characteristics, and the physicochemical characteristics for physicochemical characteristic coding are obtained. 7.The CYP2C19 enzyme metabolic phenotyping prediction method based on graph convolution network according to claim 1, wherein: The graph convolution network regression model and the five-classification graph convolution network model both comprise a residue contact map processing module, an edge index expansion module, a feature extraction module and a prediction module; the prediction module of the graph convolution network regression model is used for predicting the enzyme metabolic activity score, and the prediction module of the five-classification graph convolution network model is used for predicting the five metabolic phenotypes of the enzyme. 8.The CYP2C19 enzyme metabolic phenotyping prediction method based on graph convolution network according to claim 7, wherein: The residue contact map processing module is used for loading the residue contact map, generating an adjacency matrix, and converting the adjacency matrix into edge indexes corresponding to each single-sample residue in the residue contact map; the edge index expansion module is used for calculating the node ID offset of each sample according to the training batch size of the data set, superimposing the node ID offset on the value of each edge index, and then splicing along the column dimension to obtain an expanded edge index matrix. 9.The CYP2C19 enzyme metabolic phenotyping prediction method based on graph convolution network according to claim 8, wherein: The feature extraction module comprises a first graph convolution branch and a second graph convolution branch with the same structure, and a feature merging module for feature splicing and merging; two single-allele feature matrices split from the source domain data set or the target domain data set are input into the corresponding first graph convolution branch and the second graph convolution branch together with the expanded edge index matrix for neighbor residue feature aggregation processing; the outputs of the first graph convolution branch and the second graph convolution branch are sent to the feature merging module for splicing and merging, and then transmitted to the prediction module.

10. The device for predicting CYP2C19 enzyme pharmacometabolic phenotype based on graph convolution network, adopting the method of claim 1, characterized in that: It comprises: a residue contact map construction unit configured to construct a residue contact map of a CYP2C19 wild-type amino acid sequence; a data set generation unit configured to obtain CYP2C19 amino acid variation information for training from a database, and generate a source domain data set and a target domain data set; a model training unit configured to construct a graph convolution network regression model and a five-classification graph convolution network model, train the graph convolution network regression model through the residue contact map and the source domain data set, and migrate the trained graph convolution network regression model weight and feature into the five-classification graph convolution network model by using a transfer learning algorithm; then, the five-classification graph convolution network model after migration is optimized and trained through the residue contact map and the target domain data set; a prediction unit configured to input CYP2C19 amino acid variation information to be analyzed into the optimized five-classification graph convolution network model for prediction, and output the predicted CYP2C19 enzyme drug metabolism phenotype.

Images