Dialysis drug and hemoglobin affinity prediction method based on graph neural network

Abstract

The application provides a dialysis drug and hemoglobin affinity prediction method based on a graph neural network. A molecular structure graph representation of the dialysis drug is established, and the drug molecular formula is converted into a graph data structure, wherein each atom is represented by a node and each chemical bond is represented by an edge. An adaptive attention graph convolutional neural network is used to model the dialysis drug, learn and update the features of the atoms and edges. A convolutional neural network is used to model the amino acid sequence of hemoglobin to extract the feature vector of the protein. A bidirectional attention module is used to fuse the dialysis drug features and hemoglobin features obtained from the two models, calculate the attention coefficients of the drug to the protein and the protein to the drug, and perform feature weighting fusion. A regression analysis type neural network is used to predict the affinity between the dialysis drug and hemoglobin. The method can effectively improve the accuracy of affinity prediction, reduce research and development costs, and has a wide application prospect.

Description

Technical fields:

[0001] This invention relates primarily to drug discovery, and more particularly to a method for predicting the affinity between dialysis drugs and hemoglobin based on graph neural networks. Background technology:

[0002] In the medical field, dialysis is an important method of blood purification used to remove metabolic waste and excess water from a patient's body. The choice of dialysis drugs directly affects the efficacy of dialysis treatment; therefore, accurately predicting the affinity of dialysis drugs for hemoglobin is of significant clinical importance. As hemoglobin is an oxygen transporter, its interaction with drugs can affect oxygen transport efficiency, thereby impacting efficacy and safety.

[0003] Traditional methods for predicting drug affinity for target proteins are typically based on biophysical experiments, which are time-consuming, labor-intensive, and costly. In recent years, with the rapid development of computer technology and bioinformatics, computational model-based prediction methods have emerged. These methods can provide guidance before experiments, reduce the number of experiments, and accelerate the drug development process.

[0004] Graph Neural Networks (GNNs), as efficient deep learning models, have been successfully applied to feature extraction of drug molecule structures and prediction of bioactivity. GNNs effectively capture complex intermolecular interactions by utilizing topological information within the graphical data structure of drug molecules. However, existing methods still have limitations in capturing the complex interactions and stereochemical properties of dialysis drug-protein complexes.

[0005] Therefore, a novel approach is needed to improve the accuracy and efficiency of predicting drug affinity for hemoglobin. This invention proposes a graph neural network-based prediction method that combines adaptive attention mechanisms and convolutional neural networks. This method can more precisely simulate the interaction between drugs and proteins and is further learned through a deep learning model to improve the accuracy and reliability of predictions. This method is expected to significantly optimize the screening process for dialysis drugs and has significant value for personalized medicine and drug development in dialysis treatment. Summary of the Invention:

[0006] To address the shortcomings of existing technologies, this invention proposes a method for predicting the affinity between dialysis drugs and hemoglobin based on graph neural networks. This method includes the following steps:

[0007] S1: Establish a molecular structure diagram representation of dialysis drugs, converting the drug's molecular formula into a graphical data structure, where each atom is represented by a node and each chemical bond is represented by an edge;

[0008] S2: Modeling dialysis drugs using an adaptive attention map convolutional neural network, which includes:

[0009] S2-1: Node-edge network layer, which uses graph attention mechanism to fuse the attribute information of atoms at both ends of the edge into the edge attribute;

[0010] S2-2: Edge-to-edge network layer, which uses adaptive attention to convert edge attributes and inter-atomic connection features obtained from the node-to-edge network layer into edge features;

[0011] S2-3: Edge-node network layer, which uses an adaptive attention mechanism to learn and update the features of atoms from edge features;

[0012] S3: A convolutional neural network is used to model the amino acid sequence of hemoglobin, and the feature vector of the protein is extracted through a convolutional layer with a fixed window size.

[0013] S4: Use a bidirectional attention module to fuse the dialysis drug features and hemoglobin features obtained from steps S2 and S3, where this module calculates:

[0014] S4-1: Attention coefficients from drug to protein and from protein to drug;

[0015] S4-2: Use these attention coefficients to weight and fuse the corresponding feature vectors to form a fused feature representation;

[0016] S5: Using a regression analysis-based neural network, based on the fusion features obtained in step S4, predict the affinity value between the dialysis drug and hemoglobin, and output a predicted affinity value.

[0017] As a preferred embodiment of the present invention, the specific structure of the graphic data structure in step S1 includes:

[0018] Molecular diagram representation: The molecular formula of a dialysis drug is converted into a graphical data structure, in which each atom is represented as a node in the graph, and each type of atom has a unique node type;

[0019] Chemical bond representation: Each chemical bond in a molecule is represented as an edge between two atomic nodes. The edge not only represents the connection between atoms, but also contains information about the type of bond.

[0020] Node characteristics: Each node contains multiple characteristics describing its chemical properties, including the atom's charge, electron affinity, atomic radius, and its connectivity in the molecule; the node is the atom.

[0021] Edge characteristics: Each edge also has multiple characteristics, including the bond length, bond strength, and the type of atoms connected by the bond. These edges are chemical bonds.

[0022] Topological structure: The graph data structure must also accurately reflect the three-dimensional spatial topology of the molecule to ensure that the stereochemical properties of the molecule can be captured during graph convolution.

[0023] As a preferred embodiment of the present invention, the amino acid sequence in step S3 is obtained in the following manner: each amino acid is converted into a numerical vector through an encoding scheme, which specifically uses the physicochemical properties of the amino acid, including hydrophobicity, charge, and size.

[0024] As a preferred embodiment of the present invention, the adaptive attention map convolutional neural network in step S2 further includes:

[0025] T1: Dynamically adjust the adjacency weights in the graph convolution process using a learning-based strategy, where the adaptive attention module learns the interaction strength between each atom node and its neighboring atoms.

[0026] T2: The node-edge network layer uses a multi-head attention mechanism, where each head captures different types of inter-atomic interaction features;

[0027] T3: The edge-to-edge network layer includes a serialization process that refines the feature representation of each inter-atomic connection. The feature transformation process is optimized through reinforcement learning algorithms to adapt to complex molecular structure changes.

[0028] T4: The edge-node network layer adopts a deep residual learning framework to output highly representative drug feature vectors.

[0029] As a preferred embodiment of the present invention, the adaptive attention module in step S2 specifically includes: node-edge network layer 1, edge-edge network layer 1, edge-node network layer 1, node-edge network layer 2, edge-edge network layer 2, and edge-node network layer 2, with the specific connection method as follows:

[0030] Node-edge network layer 1, edge-edge network layer 1, and edge-node network layer 1 are connected sequentially.

[0031] Node-edge network layer 2, edge-edge network layer 2, and edge-node network layer 2 are connected sequentially;

[0032] Edge-node network layer 1 connects both node-edge network layer 2 and edge-node network layer 2.

[0033] As a preferred embodiment of the present invention, the step of fusing the dialysis drug characteristics and hemoglobin characteristics in step S4 is as follows:

[0034] P1: Define the drug feature matrix and protein feature matrix as follows:

[0035]

[0036] in, This represents the drug feature matrix after processing by the third layer of the graph neural network, where each... represents the eigenvector of the i-th atom node, and N is the total number of atoms in the drug molecule. This represents the processed protein feature matrix, each... The feature vector represents the j-th amino acid residue, and M is the total number of amino acid residues in the protein sequence;

[0037] P2: Define the bidirectional attention matrix I as follows:

[0038]

[0039] in, It is the feature matrix of dialysis drugs; is the feature matrix of hemoglobin, used to transform drug features and protein features; q is a learnable weight vector used to calculate attention scores; σ represents the activation function; 1 is a vector of all 1s with the same dimension as q, which expands the dimension of q through a dot product;

[0040] P3: The learnable weight matrix for drugs and proteins, calculated using the following formula:

[0041]

[0042] in, U represents a matrix over the real number field, and K represents the dimension of the hidden layer or the transformed feature dimension. V represents a matrix over the real number field;

[0043] P4: The formula for calculating the combined characteristics of proteins and drugs is:

[0044]

[0045] Among them, f k ' represents the k-th element in the joint feature vector of the protein and drug, and σ is the activation function;

[0046] P5: Obtaining the final fusion characteristics through reddening operations:

[0047] f = SumPool(f′, S)

[0048] Where S is the window size for the pooling operation.

[0049] Compared with the relevant prior art, the present application has the following main technical advantages, and the beneficial effects of the present invention are:

[0050] Improved prediction accuracy: This invention utilizes an adaptive attention map convolutional neural network, which can more accurately capture the complex interactions between drug molecules and hemoglobin. Compared to traditional methods, this deep learning approach can provide more detailed and accurate molecular feature representations when predicting drug-protein interactions, thereby improving the accuracy of affinity prediction.

[0051] Reduced experimental requirements: By providing highly reliable predictive results before experiments, this invention reduces the need for biophysical experiments, which not only accelerates the drug screening process but also significantly reduces research and development costs.

[0052] Accelerating the drug development process: This invention enables rapid affinity assessment of a large number of drug molecules, which is of great value for screening candidate drugs in the early stages of drug development and helps to shorten the time from laboratory to market.

[0053] Broad application potential: Although this invention focuses on predicting the affinity between dialysis drugs and hemoglobin, the proposed method is also applicable to other types of drug-protein interaction prediction, demonstrating broad application potential.

[0054] Lowering the technical threshold: The method provided by this invention reduces the reliance on professional knowledge. Users do not need a deep background in biochemistry to use this system for prediction, which allows more non-professionals to participate in the drug screening process.

[0055] Supporting Personalized Medicine: Accurate affinity prediction supports personalized drug design, helps develop customized treatment plans for specific patients or disease states, and provides a powerful tool for achieving precision medicine. Attached image description:

[0056] Figure 1 This is a flowchart of the prediction method described in this invention. Detailed implementation method:

[0057] The present invention will be further described below with reference to the accompanying drawings and embodiments. However, the present invention can be implemented in many different ways and should not be construed as limited to the embodiments shown; rather, these embodiments provide those skilled in the art with implementation methods that meet applicable legal requirements.

[0058] Example 1: According to Figure 1 As shown in the figure, this embodiment provides a specific application process for predicting the affinity between dialysis drugs and hemoglobin based on graph neural networks.

[0059] Data preparation: First, structural data of dialysis drugs and hemoglobin were obtained from public databases. The molecular structures of dialysis drugs were represented by SMILES strings, and the amino acid sequences of hemoglobin were obtained from protein databases.

[0060] Graph Data Construction: The SMILES strings of dialysis drugs are converted into a graph data structure. Each atom is represented as a node in the graph, and each chemical bond is represented as an edge connecting the corresponding node. These transformations are handled using open-source cheminformatics toolkits such as RDKit.

[0061] Feature extraction: An adaptive attention map convolutional neural network is used to process the transformed drug image data and extract feature vectors of drug molecules. A one-dimensional convolutional neural network is used to process the amino acid sequence of hemoglobin and extract protein feature vectors.

[0062] Feature fusion: The attention coefficients between drug features and protein features are calculated using a bidirectional attention module, and then these features are weighted and fused to form a joint feature representation.

[0063] Affinity prediction: The fused features are input into a regression analysis neural network to predict the affinity between dialysis drugs and hemoglobin. The neural network architecture can be a multilayer perceptron with several hidden layers and activation functions.

[0064] Model Training and Validation: The model was trained and validated using real-world drug-protein affinity data. Cross-validation was employed to evaluate model performance, and the accuracy of predictions was verified by comparing experimental data.

[0065] Results analysis: The prediction results show that the method proposed in this embodiment is highly consistent with the experimental data, proving the effectiveness of the method in predicting the affinity between dialysis drugs and hemoglobin.

[0066] Example 2: Verification of the effectiveness of the method of the present invention in predicting the affinity of a specific dialysis drug—lidocaine—for the human hemoglobin α chain. The lidocaine molecule was obtained from the PubChem database, with the SMILES string CCN(CC)C(=O)C1=CN(C)C2=CC=CC=C21. The amino acid sequence of the human hemoglobin α chain was obtained from the UniProt database, with sequence ID P69905.

[0067] Molecular structure and sequence preparation: The first 5 atoms of the lidocaine molecular diagram, each atom possessing 3 distinct chemical properties (e.g., charge, atomic radius, and electronegativity). The eigenvector matrix is ​​shown below:

[0068]

[0069] The relevant fragment of the hemoglobin α chain contains a 150-amino acid sequence, and the feature vector matrix of the protein sequence is shown below:

[0070]

[0071] Feature extraction and fusion: Features are extracted using a defined adaptive attention map convolutional network and a one-dimensional convolutional network. Bidirectional attention scores are calculated, and feature vectors from the drug and protein are fused.

[0072] Affinity prediction: The joint feature vector f is a weighted fusion result of drug and protein features, specifically the matrix:

[0073]

[0074] Applying summation and pooling operations (with a window size of 3) to reduce the dimensionality of the joint feature vector, the final feature vector is: [2.83.0 2.1]

[0076] Example 3: To predict the affinity of dialysis drugs for hemoglobin, this invention requires acquiring a large amount of high-quality drug and protein data. Data sources include multiple public databases and resources to ensure accuracy and completeness. Drug data sources include: PubChem database: PubChem is an open database containing millions of compounds. Dialysis drug data is primarily sourced from PubChem, obtaining molecular structures and related property information for these drugs. The molecular structure of each drug is represented by the SMILES (Simplified Molecular Input Line Entry System) string. DrugBank database: DrugBank provides detailed drug information, including chemical structures, pharmacological information, and target data. Drug data in DrugBank is also used to supplement and validate data obtained from PubChem. Protein data: UniProt database: UniProt is a comprehensive database of protein sequence and functional information. Hemoglobin amino acid sequences are primarily sourced from UniProt, obtaining detailed information on these sequences, including sequence ID, sequence length, and related functional annotations.

[0077] Drug data format: SMILES string: The molecular structure of each drug is represented by a SMILES string. For example, the SMILES string for lidocaine is:

[0078] CCN(CC)C(=O)C1=CN(C)C2=CC=CC=C21.

[0079] Molecular characteristics include molecular weight, hydrophobicity (logP value), polar surface area (PSA), and the number of hydrogen bond donors and acceptors. These molecular characteristics are calculated from the SMILES string using the RDKit toolkit.

[0080] Protein data format: Amino acid sequence: The amino acid sequence of each hemoglobin is represented in FASTA format. For example, the sequence ID of the human hemoglobin α chain is P69905, and its amino acid sequence is as follows:

[0081] MVLSPADKTNVKAAWGKVGAHAGEYGAEALERMFSFPTTKTYFPHF DLSHGSAQVKGHGKKVADALTNAVAHVDDMPNALSALSDLHAHKLRVDP VNFKLLSHCLLVTLAAHLPAEFTPAVHASLDKFLASVSTVLTSKYR.

[0082] Amino acid characteristics: Each amino acid is encoded by its physicochemical properties, such as hydrophobicity, charge, and size. This property information can be obtained from the AAindex database and converted into a numerical vector.

[0083] The specific experimental steps are as follows:

[0084] 1. Data Collection: Download the SMILES strings and related characteristic data of the drugs from PubChem and DrugBank. Download the amino acid sequence of hemoglobin from UniProt and obtain the physicochemical properties of each amino acid using the AAindex database.

[0085] 2. Data Preprocessing: Drug Data: The SMILES string was converted into a molecular graph structure using the RDKit toolkit, where each atom is represented as a node in the graph and each chemical bond is represented as an edge connecting the corresponding node. Features of the nodes and edges were extracted, including atomic charge, electron affinity, atomic radius, and bond length and strength. Protein Data: Amino acid sequences were converted into numerical vectors, with the physicochemical properties of each amino acid used as features. A sliding window technique was used to segment the amino acid sequences to extract fixed-size local features.

[0086] 3. Data Augmentation: Molecular graph data augmentation: Techniques such as random rotation, flipping, and adding noise are used to enhance molecular graphs to increase the robustness of the model. Protein sequence data augmentation: Amino acid sequences are enhanced using techniques such as random cropping and adding noise to simulate different biological environments.

[0087] The amount of data is as follows:

[0088] Drug dataset: Contains 2000 dialysis drugs, each with its molecular structure represented by SMILES strings, averaging 30 features per drug. Protein dataset: Contains 1000 hemoglobins, each with an amino acid sequence length of 140 to 160 amino acids, averaging 20 features per protein. Total data volume: Approximately 2 million drug-hemoglobin pairs, used for model training, validation, and testing.

[0089] The experimental environment is as follows:

[0090]

[0091] The evaluation indicators are as follows:

[0092] Mean Squared Error (MSE): Measures the average squared error between the predicted and actual values. Coefficient of Determination (R²) 2 Predicted values: 1) Measure how well the model fits the data. 2) Mean Absolute Error (MAE): Measure the average absolute error between predicted and actual values. 3) Pearson Correlation Coefficient: Measure the linear correlation between predicted and actual values.

[0093] The evaluation method is as follows:

[0094] Data Splitting: The dataset is split into a training set (70%), a validation set (15%), and a test set (15%). Model Training: The model is trained using the training set, and hyperparameters are tuned using the validation set. Model Evaluation: The model performance is evaluated using the test set, and evaluation metrics are calculated.

[0095] The experimental results are as follows:

[0096]

[0097]

[0098] Below are 12 real-world experimental data examples for dialysis drugs and hemoglobin. Each example includes the name of the dialysis drug, the SMILES string, the name and amino acid sequence of the hemoglobin, as well as the predicted affinity and evaluation metrics.

[0099]

[0100] Experimental results show that the method provided in this invention performs excellently on the training set, validation set, and test set. Specific evaluation metrics are as follows: Training set: Mean Squared Error (MSE) is 0.0021, Coefficient of Determination (R²) is 0.0021, and Coefficient of Determination (R²) is 0.0021. 2 The mean absolute error (MAE) was 0.036, and the Pearson correlation coefficient was 0.97. The validation set had a mean squared error (MSE) of 0.0035 and a coefficient of determination (R²) of 0.95.2 The mean absolute error (MAE) was 0.045, and the Pearson correlation coefficient was 0.95. For the test set: the mean squared error (MSE) was 0.0038, and the coefficient of determination (R²) was 0.92. 2 The mean absolute error (MAE) was 0.91, the mean absolute error (MAE) was 0.048, and the Pearson correlation coefficient was 0.94.

[0101] Analysis of the experimental results shows that the model exhibits high accuracy and good generalization ability across various datasets. The model effectively captures the complex interactions between dialysis drugs and hemoglobin, providing accurate affinity predictions. Compared to traditional biophysical experimental methods, the method of this invention significantly reduces experimental requirements and accelerates the drug screening and development process.

[0102] In summary, the graph neural network-based method for predicting the affinity between dialysis drugs and hemoglobin proposed in this invention not only demonstrates excellent prediction accuracy but also has significant advantages in reducing experimental costs and accelerating drug development, showcasing broad application prospects and potential.

[0103] This embodiment details the workflow and method of the invention, and how to use graph neural networks to extract information from the characteristics of drug molecules and proteins, thereby predicting their affinity.

[0104] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for predicting the affinity between dialysis drugs and hemoglobin based on graph neural networks, comprising the following steps: S1: Establish a molecular structure diagram representation of dialysis drugs, converting the drug's molecular formula into a graphical data structure, where each atom is represented by a node and each chemical bond is represented by an edge; S2: Modeling dialysis drugs using an adaptive attention map convolutional neural network, which includes: S2-1: Node-edge network layer, which uses graph attention mechanism to fuse the attribute information of atoms at both ends of the edge into the edge attribute; S2-2: Edge-to-edge network layer, which uses adaptive attention to convert edge attributes and inter-atomic connection features obtained from the node-to-edge network layer into edge features; S2-3: Edge-node network layer, which uses an adaptive attention mechanism to learn and update the features of atoms from edge features; S3: A convolutional neural network is used to model the amino acid sequence of hemoglobin, and the feature vector of the protein is extracted through a convolutional layer with a fixed window size. S4: Use a bidirectional attention module to fuse the dialysis drug features and hemoglobin features obtained from steps S2 and S3, where this module calculates: S4-1: Attention coefficients from drug to protein and from protein to drug; S4-2: Use these attention coefficients to weight and fuse the corresponding feature vectors to form a fused feature representation; The steps for fusing dialysis drug characteristics and hemoglobin characteristics are as follows: P1: Define the drug feature matrix and protein feature matrix as follows: in, This represents the drug feature matrix after processing by the third layer of the graph neural network, where each... Representing the The eigenvectors of atomic nodes, and It is the total number of atoms in a drug molecule. This represents the processed protein feature matrix, each... Indicates the first Feature vectors of amino acid residues It is the total number of amino acid residues in the protein sequence; P2: Define the bidirectional attention matrix I as follows: in, It is the feature matrix of dialysis drugs; It is the feature matrix of hemoglobin, used to transform drug features and protein features; It is a learnable weight vector used to calculate the attention score; Indicates the activation function; It is a with A vector of all ones of the same dimension can be expanded using a dot product. The dimension; P3: The learnable weight matrix for drugs and proteins, calculated using the following formula: in, U represents a matrix over the real number field, and K represents the dimension of the hidden layer or the transformed feature dimension. V represents a matrix over the real number field; P4: The formula for calculating the combined characteristics of proteins and drugs is: in, This represents the k-th element in the joint feature vector of the protein and the drug, where σ is the activation function; P5: Obtain the final fused features through pooling operations: Where S is the window size for the pooling operation; S5: Using a regression analysis-based neural network, based on the fusion features obtained in step S4, predict the affinity value between the dialysis drug and hemoglobin, and output a predicted affinity value.

2. The method for predicting the affinity between dialysis drugs and hemoglobin based on graph neural networks according to claim 1, characterized in that, The specific structure of the graphics data structure mentioned in step S1 includes: Molecular diagram representation: The molecular formula of a dialysis drug is converted into a graphical data structure, in which each atom is represented as a node in the graph, and each type of atom has a unique node type; Chemical bond representation: Each chemical bond in a molecule is represented as an edge between two atomic nodes. The edge not only represents the connection between atoms, but also contains information about the type of bond. Node characteristics: Each node contains multiple characteristics describing its chemical properties, including the atom's charge, electron affinity, atomic radius, and its connectivity in the molecule; the node is the atom. Edge characteristics: Each edge also has multiple characteristics, including the bond length, bond strength, and the type of atoms connected by the bond. These edges are chemical bonds. Topological structure: The graph data structure must also accurately reflect the three-dimensional spatial topology of the molecule to ensure that the stereochemical properties of the molecule can be captured during graph convolution.

3. The method for predicting the affinity between dialysis drugs and hemoglobin based on graph neural networks according to claim 1, characterized in that, The amino acid sequence in step S3 is obtained in the following way: each amino acid is converted into a numerical vector through an encoding scheme, which specifically uses the physicochemical properties of the amino acid, including hydrophobicity, charge, and size.

4. The method for predicting the affinity between dialysis drugs and hemoglobin based on graph neural networks according to claim 1, characterized in that, The adaptive attention map convolutional neural network described in step S2 further includes: T1: Dynamically adjust the adjacency weights in the graph convolution process using a learning-based strategy, where the adaptive attention module learns the interaction strength between each atom node and its neighboring atoms. T2: The node-edge network layer uses a multi-head attention mechanism, where each head captures different types of inter-atomic interaction features; T3: The edge-to-edge network layer includes a serialization process that refines the feature representation of each inter-atomic connection. The feature transformation process is optimized through reinforcement learning algorithms to adapt to complex molecular structure changes. T4: The edge-node network layer adopts a deep residual learning framework to output highly representative drug feature vectors.

5. The method for predicting the affinity between dialysis drugs and hemoglobin based on graph neural networks according to claim 1, characterized in that, The adaptive attention module mentioned in step S2 specifically includes: node-edge network layer 1, edge-edge network layer 1, edge-node network layer 1, node-edge network layer 2, edge-edge network layer 2, and edge-node network layer 2, with the following connection method: Node-edge network layer 1, edge-edge network layer 1, and edge-node network layer 1 are connected sequentially. Node-edge network layer 2, edge-edge network layer 2, and edge-node network layer 2 are connected sequentially; Edge-node network layer 1 connects both node-edge network layer 2 and edge-node network layer 2.

Images