Method and system for predicting drug interactions based on atomic and chemical bond structures

Abstract

The application provides a drug interaction prediction method and system based on atomic and chemical bond structure, and relates to the technical field of biological information analysis, and comprises the following steps: acquiring atomic features, chemical bond features, molecular fingerprint features, two-dimensional structure information and three-dimensional spatial conformation information of each drug molecule; calculating a two-dimensional distance matrix of atoms and chemical bonds based on the two-dimensional structure information; calculating a three-dimensional distance matrix of atoms and chemical bonds based on the three-dimensional spatial conformation information; using a graph Transformer network to introduce the two-dimensional distance matrix and the three-dimensional distance matrix as spatial prior bias into a self-attention mechanism of the network to process the atomic features and the chemical bond features, so as to obtain structure features of each drug molecule; and outputting a prediction score of drug molecule interaction based on the structure features and the molecular fingerprint features of the drug molecule. The application realizes multi-classification prediction of drug interaction types and risk levels, and provides more comprehensive decision support for clinical rational drug use.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics analysis technology, and in particular to a method and system for predicting drug interactions based on atomic and chemical bond structures. Background Technology

[0002] Drug interactions (DDIs) refer to the phenomenon where the simultaneous or sequential use of two or more drugs leads to altered drug efficacy or increased toxicity due to pharmacokinetic or pharmacodynamic interactions. Adverse clinical consequences resulting from DDIs are termed drug interaction events (DDIEs). Multiple studies have shown that DDIs are one of the leading causes of adverse drug reactions in hospitalized patients, seriously threatening patient medication safety. With the increasing complexity of clinical combination drug regimens, accurate prediction of DDIs has become a crucial step in ensuring rational drug use in clinical practice.

[0003] Currently, AI-based drug interaction prediction methods include similarity-based methods and matrix factorization-based methods. Similarity-based methods follow the assumption that similar chemical structures lead to similar biological properties, inferring potential interactions by calculating the similarity of drugs in chemical structure, target proteins, or pathways of action. Matrix factorization-based methods, on the other hand, decompose known drug interaction matrices to learn low-dimensional potential characterizations of drugs, thereby reconstructing unknown interactions.

[0004] In recent years, deep learning technology has been widely applied in the field of drug discovery defect (DDI) prediction. Research shows that DDI is often determined by specific chemical substructures within drug molecules, thus molecular graph-based deep learning methods have become a research hotspot. These methods extract substructure features of drugs for DDI prediction using graph neural networks (GNNs). For example, the DSN-DDI model considers both global and local structural representations of the drug through a dual-path learning mechanism.

[0005] However, existing technologies still have the following limitations: First, existing molecular graph-based DDI prediction methods generally suffer from the problem of decoupling between atomic and chemical bond features. These methods typically treat atoms as the main nodes for feature learning, while considering chemical bonds merely as connectivity relationships, failing to fully consider the synergistic relationship between atoms and chemical bonds, resulting in insufficient characterization of the local chemical environment.

[0006] Secondly, most studies are limited to two-dimensional molecular diagram characterization, neglecting the three-dimensional spatial conformation information of drug molecules. In fact, the three-dimensional spatial structure of drug molecules directly affects their binding ability and mode with the target, and modeling without three-dimensional spatial information will lose key stereochemical features.

[0007] Furthermore, traditional graph neural network methods are limited by local neighborhood message passing mechanisms, making it difficult to effectively capture long-range synergistic interactions between non-adjacent atoms / chemical bonds within drug molecules. This limitation prevents the models from accurately identifying interactions between spatially distant but functionally related chemical substructures.

[0008] Finally, most current mainstream DDI prediction models can only predict whether two drugs will interact, lacking the ability to finely classify specific interaction types and clinical risk levels, which is of vital value in guiding clinical medication decisions. Summary of the Invention

[0009] To address at least one of the technical problems mentioned above, this invention proposes a method and system for predicting drug interactions based on atomic and chemical bond structures.

[0010] To achieve the above objectives, a first aspect of the present invention provides a method for predicting drug interactions based on atomic and chemical bond structures, comprising: Acquire atomic features, chemical bond features, molecular fingerprint features, two-dimensional structural information, and three-dimensional spatial conformation information for each drug molecule; Based on the two-dimensional structural information, calculate the two-dimensional distance matrix between atoms and chemical bonds; based on the three-dimensional spatial conformation information, calculate the three-dimensional distance matrix between atoms and chemical bonds. Using a graph Transformer network, the two-dimensional distance matrix and the three-dimensional distance matrix are introduced as spatial prior biases into the self-attention mechanism of the Transformer to process the atomic features and chemical bond features, thereby obtaining the structural features of each drug molecule. Based on the structural features and molecular fingerprint features of the two drug molecules, a prediction score for drug molecule interaction is output.

[0011] Furthermore, the calculation of the two-dimensional distance matrix between atoms and chemical bonds includes: Calculate the atom-to-atom distance matrix, where for any two atom nodes in the molecular graph, the shortest path distance is calculated as the atom-to-atom distance. If the two atom nodes are not connected, their distance is set to a specific value. The chemical bond-chemical bond distance matrix is ​​calculated by constructing a dual graph with chemical bonds as nodes and atoms as edges. The shortest path distance between any two chemical bonds is calculated on the dual graph. If two chemical bonds are not connected, their distance is set to a specific value. The atomic-chemical bond distance matrix is ​​calculated by constructing an extended graph, the node set of which includes the original atomic nodes and edge nodes representing chemical bonds. The shortest path distance from the atomic nodes to the edge nodes in the extended graph is calculated to form the atomic-chemical bond spatial position matrix.

[0012] Furthermore, the calculation of the three-dimensional distance matrix between atoms and chemical bonds includes: The three-dimensional conformation of drug molecules is obtained through Merck molecular force field, and the conformation with the lowest energy is selected. Calculate the atom-to-atom Euclidean distance matrix based on the three-dimensional coordinates of atoms; Based on the three-dimensional coordinates of chemical bonds, calculate the Euclidean distance matrix between chemical bonds, where the coordinates of the chemical bonds are calculated by the arithmetic mean of the coordinates of the atoms at both ends of the chemical bond; Calculate the Euclidean distance matrix between atoms and chemical bonds based on atomic coordinates and chemical bond coordinates.

[0013] Furthermore, the method also includes adding virtual nodes connected to all atoms and all chemical bonds in the two-dimensional distance matrix and the three-dimensional distance matrix; and exchanging information with atoms and chemical bonds through the virtual nodes to aggregate global structural information.

[0014] Furthermore, the step of introducing the two-dimensional distance matrix and the three-dimensional distance matrix as spatial prior biases into the network's self-attention mechanism specifically includes: In the self-attention layer of the graph Transformer network, the attention score is calculated according to the following formula: ; in, , K, V These are query, key, and value matrices, respectively. d For feature dimension, and These are the two-dimensional distance matrix and the three-dimensional distance matrix, respectively.

[0015] Furthermore, the molecular fingerprint feature is obtained through the following steps: The Morgan fingerprint, RDKit fingerprint, and MACCS fingerprint of the drug molecule are calculated; the three fingerprint vectors are concatenated; the concatenated vector is input into a multilayer perceptron (MLP) for fusion and dimensionality reduction to obtain the final molecular fingerprint feature representation.

[0016] A second aspect of the present invention provides a drug interaction prediction system based on atomic and chemical bond structures, comprising: The acquisition module is used to acquire the atomic characteristics, chemical bond characteristics, molecular fingerprint characteristics, two-dimensional structural information, and three-dimensional spatial conformation information of each drug molecule; The calculation module is used to calculate the two-dimensional distance matrix of atoms and chemical bonds based on the two-dimensional structural information; and to calculate the three-dimensional distance matrix of atoms and chemical bonds based on the three-dimensional spatial conformation information. The Graph Transformer module is used to introduce the two-dimensional and three-dimensional distance matrices as spatial prior biases into the network's self-attention mechanism using a graph Transformer network to process the atomic features and chemical bond features, thereby obtaining the structural features of each drug molecule. The prediction module is used to output a prediction score of the interaction between the two drug molecules based on their respective structural features and molecular fingerprint features.

[0017] A third aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps in the drug interaction prediction method based on atomic and chemical bond structures as described in the first aspect of the present invention.

[0018] A fourth aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps in the drug interaction prediction method based on atomic and chemical bond structures as described in the first aspect of the present invention.

[0019] A fifth aspect of the present invention provides a computer program product comprising software code, wherein a program in the software code performs steps in the drug interaction prediction method based on atomic and chemical bond structures as described in the first aspect of the present invention.

[0020] Compared with existing technologies, the drug interaction prediction method and system based on atomic and chemical bond structures provided by this invention have the following advantages: (1) To address the decoupling of atom-chemical bond characteristics, this invention calculates the two-dimensional and three-dimensional distance matrices of atoms and chemical bonds respectively. By constructing a complete distance matrix system including atom-to-atom, chemical bond-to-chemical bond, and atom-to-chemical bond distance matrices, it achieves effective modeling of the synergistic effects of atoms and chemical bonds. This method breaks through the limitation of traditional methods that only treat chemical bonds as a connection relationship, enabling the model to simultaneously perceive the interrelationships of atoms and chemical bonds in two-dimensional topological structures and three-dimensional spatial conformations, thereby significantly improving the accuracy and completeness of characterizing the local chemical environment of drugs.

[0021] (2) This invention calculates the three-dimensional distance matrix of atoms and chemical bonds based on three-dimensional spatial conformation information, obtains the three-dimensional conformation with the lowest energy through the Merck molecular force field, and calculates the Euclidean distance matrix based on the spatial coordinates of atoms and chemical bonds, thus introducing the stereochemical characteristics of drugs into the prediction model. This design enables the model to fully consider the influence of the spatial conformation of drug molecules on their interactions, making up for the shortcomings of traditional two-dimensional molecular diagram characterization methods and enhancing the ability to capture the sensitivity of drug spatial structure.

[0022] (3) To address the technical problem of effectively capturing the synergistic effects of non-adjacent atoms / chemical bonds within drug molecules in existing technologies, this invention introduces the distance matrix as a spatial prior bias into the self-attention mechanism of the graph Transformer. By integrating the two-dimensional and three-dimensional distance matrices into the multi-head self-attention calculation as bias terms, the model can overcome the limitations of local neighborhoods and directly model the spatial relationships between any two atoms or chemical bonds. This mechanism effectively expands the model's receptive field, enabling it to identify synergistic effects between spatially distant but functionally related chemical substructures.

[0023] (4) This invention is based on multi-class prediction using fused molecular fingerprint features. By combining the rich chemical functional information provided by three molecular fingerprints—Morgan, RDKit, and MACCS—and effectively fusing them with deep molecular characterization, a multi-class prediction model capable of distinguishing different interaction types and risk levels is constructed. This method can not only predict the existence of interactions but also provide specific interaction mechanism information and clinical risk level assessment, providing a more comprehensive reference for clinical medication decisions. Attached Figure Description

[0024] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0025] Figure 1 This is a flowchart of the drug interaction prediction method based on atomic and chemical bond structures provided in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the ABT-DDI model framework provided in Embodiment 1 of the present invention; Figure 3 This is a radar chart comparing the performance of the ABT-DDI model and its variants provided in Embodiment 1 of the present invention on the DDInter dataset; Figure 4 This is a bar chart comparing the performance of the ABT-DDI model and its variants provided in Embodiment 1 of the present invention on the DDInter dataset; Figure 5 This is a radar chart comparing the performance of the ABT-DDI model and its variants provided in Embodiment 1 of the present invention on the DeepDDI dataset; Figure 6 This is a bar chart comparing the performance of the ABT-DDI model and its variants provided in Embodiment 1 of the present invention on the DeepDDI dataset; Figure 7This is a visualization of the drug attention to key substructures of acetylsalicylic acid and nebivolol provided in Embodiment 1 of the present invention. Figure 8 This is a visualization of the drug attention to key substructures of bromfenac and ziprasidone provided in Embodiment 1 of the present invention. Figure 9 This is a visualization of the feature space distribution of the ABT-DDI model before training on the DDInter dataset using UMAP dimensionality reduction, as provided in Embodiment 1 of the present invention. Figure 10 This is a visualization of the feature space distribution of the trained ABT-DDI model on the DDInter dataset using UMAP dimensionality reduction, as provided in Embodiment 1 of this invention. Figure 11 This is a visualization of the feature space distribution of the ABT-DDI model before training on the DDInter dataset after t-SNE dimensionality reduction, as provided in Embodiment 1 of the present invention. Figure 12 This is a visualization of the feature space distribution of the trained ABT-DDI model on the DDInter dataset after t-SNE dimensionality reduction, as provided in Embodiment 1 of the present invention. Figure 13 This is a schematic diagram of a drug interaction prediction system based on atomic and chemical bond structures provided in Embodiment 2 of the present invention. Detailed Implementation

[0026] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0028] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0029] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0030] Before describing the embodiments of the present invention in detail, the technical terms involved in the present invention are explained and defined as follows: 1. Molecular Graph: A mathematical model used to represent the topological structure of drug molecules, where atoms are represented as nodes and chemical bonds as edges connecting the nodes. In this invention, a molecular graph is defined as... ,in It is a collection of atomic nodes. It is a collection of chemical bond edges.

[0031] 2. Atom Features: These are feature vectors used to characterize the chemical properties of atoms. In this invention, the atom features include, but are not limited to: atom type (e.g., C, N, O), chiral label, connectivity (number of atoms bonded to the atom), explicit valence state, formal charge, hybridization state (e.g., sp, sp2, sp3), whether it is an aromatic atom, atomic mass, total number of bonded hydrogen atoms, and whether the atom is within a ring. These features collectively constitute a digital description of the local chemical environment of an atom.

[0032] 3. Bond Features: These are feature vectors used to characterize the properties of chemical bonds. In this invention, the bond features include, but are not limited to: bond orientation, bond type (e.g., single, double, triple bonds), whether the bond is within a ring, whether the bond is conjugated, and the stereochemical type of the bond (e.g., tetrahedral, Z / E, etc.). Through these features, the model can distinguish the physical and chemical properties of different chemical bonds.

[0033] 4. Expanded Graph: This is a special graph structure constructed in this invention for calculating the distances between atoms and chemical bonds. Based on the original molecular graph, it creates a new "edge node" for each chemical bond, thus expanding the node set of the expanded graph. ,in This represents the set of edge nodes that symbolize chemical bonds. This structure allows for the direct calculation of the shortest path distance between atoms and chemical bonds within the same graph framework.

[0034] 5. Dual Graph: This invention presents a graph structure for calculating the distances between chemical bonds. In this graph, the chemical bonds in the original molecular graph are transformed into nodes in the dual graph, while the atoms in the original molecular graph become the edges connecting these new nodes. Using the dual graph, the shortest path distance between any two chemical bonds in the topological structure can be calculated.

[0035] 6. Spatial Prior Bias: In the graph Transformer network of this invention, the calculated two-dimensional and three-dimensional distance matrices are introduced into the self-attention mechanism as bias terms. Specifically, when calculating the attention score, the distance matrix is ​​added to the query-key dot product, as shown in the formula: This mechanism infuses the model with prior knowledge about the spatial relative positions of atoms and chemical bonds, guiding the model to focus on spatially neighboring elements.

[0036] 7. Virtual Node: This is a special type of node introduced in this invention, used to aggregate global information of the molecular graph. Specifically, it includes an "atomic virtual node" connected to all atomic nodes and a "chemical bond virtual node" connected to all chemical bond edges. As a learnable parameter, the virtual node learns and characterizes the global features of the entire molecule through information interaction with the whole graph, ultimately used to generate a molecular-level representation.

[0037] 8. Merck Molecular Force Field (MMFF): A force field model widely used in computational chemistry to predict and optimize the three-dimensional conformation of organic molecules. This invention uses MMFF for conformational search and energy optimization, selecting the lowest-energy conformation as the most stable three-dimensional structure for subsequent calculation of the three-dimensional distance matrix.

[0038] 9. Molecular Fingerprint: A computational method that transforms molecular structure into a fixed-length bit vector for efficient characterization and comparison of molecules. This invention integrates three fingerprinting methods: Morgan fingerprint (based on circular substructures), RDKit fingerprint (based on linear atomic paths), and MACCS fingerprint (based on predefined key substructures) to comprehensively capture local functional groups, topological frameworks, and key chemical features of molecules.

[0039] 10. Cross-Entropy Loss: This is a commonly used loss function in training multi-class classification models in deep learning. It guides the optimization of model parameters by measuring the difference between the probability distribution predicted by the model and the probability distribution of the true labels. This invention uses this loss function to train a multi-class prediction model, aiming to minimize prediction error.

[0040] Example 1 like Figure 1 This embodiment provides a method for predicting drug interactions based on atomic and chemical bond structures, including: Acquire atomic features, chemical bond features, molecular fingerprint features, two-dimensional structural information, and three-dimensional spatial conformation information for each drug molecule; Based on the two-dimensional structural information, calculate the two-dimensional distance matrix between atoms and chemical bonds; based on the three-dimensional spatial conformation information, calculate the three-dimensional distance matrix between atoms and chemical bonds. Using a graph Transformer network, the two-dimensional distance matrix and the three-dimensional distance matrix are introduced as spatial prior biases into the network's self-attention mechanism to process the atomic features and chemical bond features, thereby obtaining the structural features of each drug molecule. Based on the structural features and molecular fingerprint features of drug molecules, a prediction score for drug molecule interactions is output.

[0041] To address the technical challenges of existing DDI prediction methods, such as decoupling of atomic-chemical bond features, lack of 3D spatial information, and difficulty in capturing long-range dependencies, this invention provides an innovative multimodal feature learning framework. First, a comprehensive drug characterization foundation is constructed by acquiring atomic features, chemical bond features, molecular fingerprint features, and 2D and 3D structural information of drug molecules. Second, distance matrices based on 2D molecular graphs and 3D spatial conformations are calculated separately, each characterizing the spatial relationships between atoms and chemical bonds from different dimensions. Then, a graph Transformer network is used to introduce the distance matrices as spatial prior biases into a self-attention mechanism. This design allows the model to fully consider spatial proximity when calculating attention weights, maintaining the graph neural network's ability to perceive local structures while overcoming the layer limitations of traditional message passing mechanisms. Finally, by fusing deep structural features and traditional molecular fingerprint features, multi-classification prediction of drug interactions is achieved.

[0042] At the feature representation level, richer and more accurate drug characterization is constructed through collaborative modeling of atoms and chemical bonds and complementary fusion of two-dimensional and three-dimensional information. At the model architecture level, long-range dependencies within drug molecules are effectively captured through a distance matrix-guided attention mechanism. At the application level, more clinically valuable and refined prediction results are provided through multi-feature fusion and multi-classification output.

[0043] Specifically, the calculation of the two-dimensional distance matrix between atoms and chemical bonds includes: Calculate the atom-to-atom distance matrix, where for any two atom nodes in the molecular graph, the shortest path distance is calculated as the atom-to-atom distance. If the two atom nodes are not connected, their distance is set to a specific value. The chemical bond-chemical bond distance matrix is ​​calculated by constructing a dual graph with chemical bonds as nodes and atoms as edges. The shortest path distance between any two chemical bonds is calculated on the dual graph. If two chemical bonds are not connected, their distance is set to a specific value. The atomic-chemical bond distance matrix is ​​calculated by constructing an extended graph, the node set of which includes the original atomic nodes and edge nodes representing chemical bonds. The shortest path distance from the atomic nodes to the edge nodes in the extended graph is calculated to form the atomic-chemical bond spatial position matrix.

[0044] Based on multi-view distance calculation in graph theory, the shortest path distance is used to characterize the relative positions of atoms in the molecular topology in the calculation of the atom-to-atom distance matrix. Specific values ​​are assigned to unconnected atom pairs, ensuring the integrity of the distance matrix. In the calculation of the chemical bond-to-chemical bond distance matrix, a dual graph structure is innovatively constructed, converting chemical bonds in the original molecular graph into nodes in the dual graph. The topological relationships between chemical bonds are characterized by calculating the shortest path on the dual graph. In the calculation of the atom-to-chemical bond distance matrix, an extended graph is constructed, unifying atomic nodes and edge nodes representing chemical bonds in the same graph structure, thus enabling direct calculation of the spatial relationships between atoms and chemical bonds.

[0045] The two-dimensional distance matrix calculation method provided by this invention first breaks the feature isolation between atoms and chemical bonds in traditional methods, establishing a complete atom-chemical bond correlation system. Second, by constructing dual and extended graphs, it provides an independent expression space for chemical bond feature learning, avoiding the marginalization of chemical bond information. Finally, the combined effect of the three distance matrices forms a comprehensive description of the two-dimensional topological structure of molecules, providing rich structural prior information for subsequent graph Transformer modeling. This method is particularly suitable for capturing the topological relationships between functional groups in drug molecules, providing important support for understanding the chemical structural basis of drug molecule dihydrogen ionomers (DDIs).

[0046] Specifically, the calculation of the three-dimensional distance matrix between atoms and chemical bonds includes: The three-dimensional conformation of drug molecules is obtained through Merck molecular force field, and the conformation with the lowest energy is selected. Calculate the atom-to-atom Euclidean distance matrix based on the three-dimensional coordinates of atoms; Based on the three-dimensional coordinates of chemical bonds, calculate the Euclidean distance matrix between chemical bonds, where the coordinates of the chemical bonds are calculated by the arithmetic mean of the coordinates of the atoms at both ends of the chemical bond; Calculate the Euclidean distance matrix between atoms and chemical bonds based on atomic coordinates and chemical bond coordinates.

[0047] This invention is based on the theories of computational chemistry and structural biology. First, it uses the Merck Molecular Force Field (MMFF) to perform conformational search and energy optimization, selecting the stable conformation with the lowest energy as the three-dimensional structural benchmark. This step ensures the physical rationality and biological relevance of the conformation used. In the atom-to-atom distance calculation, the Euclidean distance is calculated directly based on the three-dimensional Cartesian coordinates of the atoms, accurately reflecting the relative positions of the atoms in real space. In the chemical bond-to-chemical bond distance calculation, it innovatively uses the arithmetic mean of the coordinates of the atoms at both ends of the chemical bond as the spatial position of the chemical bond, and then calculates the Euclidean distance between the bond centers. This method considers the spatial orientation of the chemical bond while maintaining computational efficiency. In the atom-to-chemical bond distance calculation, it is also calculated directly based on spatial coordinates, forming a complete three-dimensional spatial relationship description.

[0048] The technical advantages of this three-dimensional distance matrix calculation method are mainly reflected in the following aspects: First, it introduces the stereochemical characteristics of drug molecules into the DDI prediction model, fully considering the influence of three-dimensional structural factors such as steric hindrance and chirality on drug interactions; Second, by selecting the lowest energy conformation, it ensures the representativeness of the conformation used under physiological conditions, thereby improving the biological reliability of the prediction results; Third, the three-dimensional distance matrix and the two-dimensional distance matrix effectively complement each other, with the former capturing spatial proximity and the latter reflecting topological connectivity, together forming a comprehensive description of the spatial structure of the drug.

[0049] Specifically, the method further includes adding virtual nodes connected to all atoms and all chemical bonds in the two-dimensional distance matrix and the three-dimensional distance matrix; and exchanging information with atoms and chemical bonds through the virtual nodes to aggregate global structural information.

[0050] Two special virtual nodes are introduced into the distance matrix system: atomic virtual nodes are connected to all atomic nodes, and chemical bond virtual nodes are connected to all chemical bond edges, forming a global information aggregation channel. These virtual nodes serve as learnable parameter embeddings, automatically learning how to integrate the feature information of atoms and chemical bonds from a global perspective during model training. Through information interaction with all real nodes, virtual nodes gradually accumulate the global context information of the entire molecular graph, while simultaneously feeding back global information to each local node through parameter updates, achieving a bidirectional flow of global and local information.

[0051] This virtual node mechanism effectively addresses the problem in graph neural networks where information transmission requires multiple layers of stacking to reach distant nodes, achieving rapid perception of global information through the "shortcut connections" of virtual nodes. Secondly, the separate setup of atomic and chemical bond virtual nodes maintains differentiated treatment of atomic and chemical bond features, avoiding feature confusion. Finally, the global representation learned by virtual nodes can serve as a molecular-level feature representation, directly used in subsequent prediction tasks, improving the model's computational efficiency and representational capabilities. In DDI prediction tasks, this mechanism is particularly helpful in capturing chemical properties that require identification at the entire molecular level, such as the overall polarity and hydrophilicity / hydrophobicity of molecules.

[0052] Specifically, the step of introducing the two-dimensional distance matrix and the three-dimensional distance matrix as spatial prior biases into the self-attention mechanism of the graph Transformer network includes: In the self-attention layer of the graph Transformer network, the attention score is calculated according to the following formula: ; in, , K, V These are query, key, and value matrices, respectively. d For feature dimension, and These are the two-dimensional distance matrix and the three-dimensional distance matrix, respectively.

[0053] Based on this spatial prior bias mechanism, the graph Transformer is endowed with explicit spatial awareness, enabling it to allocate attention resources more rationally and improving the model's interpretability and generalization ability. Secondly, through the synergistic effect of two-dimensional and three-dimensional distance matrices, the model can simultaneously consider topological proximity and spatial proximity, resulting in a more comprehensive understanding of spatial relationships. Finally, the introduction of this bias term does not increase the model's complexity, maintaining computational efficiency while significantly improving performance. In the application of DDI prediction, this mechanism is particularly helpful in identifying interactions between active centers that are spatially close but topologically distant, providing a new technical means for understanding complex drug interaction mechanisms.

[0054] Specifically, the molecular fingerprint feature is obtained through the following steps: The Morgan fingerprint, RDKit fingerprint, and MACCS fingerprint of the drug molecule are calculated; the three fingerprint vectors are concatenated; the concatenated vector is input into an MLP for fusion and dimensionality reduction to obtain the final molecular fingerprint feature representation.

[0055] The inventors discovered that single-molecule characterization methods suffer from incomplete information coverage. Therefore, they improved the model's chemical information understanding capability through multimodal fingerprint fusion. The technical principle is based on the ensemble learning approach of feature engineering: First, Morgan fingerprint, RDKit fingerprint, and MACCS fingerprint are calculated separately. These three fingerprints describe the chemical properties of a molecule from different perspectives—Morgan fingerprint describes the local functional group environment of the molecule through circular substructures, RDKit fingerprint captures the topological skeleton features of the molecule based on a path method, and MACCS fingerprint focuses on the presence of specific key chemical substructures. Then, these three complementary fingerprint vectors are concatenated to form a rich hybrid feature representation. Finally, feature fusion and dimensionality reduction are performed using MLP, automatically learning the optimal combination of different fingerprint features, eliminating redundant information, and retaining the best feature combination.

[0056] By leveraging the complementary advantages of multiple fingerprints, a more comprehensive and robust molecular characterization is constructed, reducing reliance on single feature extraction methods. Furthermore, the nonlinear transformation capability of MLP can learn complex interactions between different fingerprint features, generating a synergistic effect. Simultaneously, dimensionality reduction not only reduces computational complexity but also serves to denoise and select features, improving the model's generalization ability. In the DDI prediction task, this rich molecular characterization is particularly helpful in distinguishing the impact of subtle chemical structural differences on interactions, providing an important feature foundation for accurately predicting interaction types and risk levels.

[0057] Specifically, the predicted score for drug interactions includes: The structural features of the first drug molecule and the structural features of the second drug molecule are fused with the molecular fingerprint features respectively, and then concatenated to obtain a joint feature vector of the drug pair; The joint feature vector is input into the classification prediction model, and the multi-class probability of drug interaction type (such as enzyme inhibition, affecting drug serum concentration, etc.) and corresponding risk level (such as severe, moderate, mild) is output according to the specific prediction task.

[0058] In one specific embodiment, the present invention provides a drug interaction prediction method ABT-DDI based on atomic and chemical bond structures, which specifically includes the following steps: Step S101: Construct a drug molecule map and generate atomic and chemical bond features.

[0059] Reference Figure 2 This embodiment describes the data processing flow in the drug interaction prediction stage. First, the drug molecule is represented as a molecular graph structure, and this is modeled as a graph. ,in This represents the set of all nodes (i.e., drug atoms) in the graph. This represents the set of edges (i.e., chemical bonds between atoms) between nodes. To characterize the properties of nodes and edges in a graph, the node feature matrix and edge feature matrix are defined as follows: Node Feature Matrix Store the feature vectors of all atoms, where This represents the eigenvector of the i-th atom. It refers to the number of nodes (atoms) in a drug molecule. The atomic feature dimension. Edge feature matrix. Store the feature vectors of all chemical bonds, where It refers to the number of edges (chemical bonds) in a drug molecule. The dimensions of chemical bond characteristics include: atomic characteristics such as atom type, chiral label, connectivity, explicit valence state, formal charge, hybridization state, whether it is an aromatic atom, atomic mass, total number of connected hydrogen atoms, and whether the atom is inside a ring. The chemical bond characteristics include bond orientation, bond type, whether the bond is inside a ring, whether the bond is conjugated, and the stereochemical type of the bond.

[0060] Step S102: Calculate the two-dimensional spatial distance matrix of atomic-chemical bonds of the drug molecule.

[0061] Based on the two-dimensional graph structure of the molecule, three distance matrices are calculated: (1) The shortest path distance between any two nodes in the drug molecule graph is calculated as the atom-to-atom distance matrix. For the molecular graph Any two atomic nodes and If a path exists between two nodes, the distance between them is defined as the shortest path length in the graph; if the two nodes are not connected, their distance is set to a specific value -1 as a marker. This calculation process is represented as a mapping. Finally, a two-dimensional atom-to-atom distance matrix is ​​obtained. (2) Calculate the shortest path distance between any two chemical bonds in the drug molecular diagram, and use it as the chemical bond-chemical bond distance matrix. Then, use the molecular diagram... A transformation is performed to construct a dual graph with chemical bonds as nodes and atoms as edges. The shortest path between any two chemical bonds is calculated on this dual graph to obtain a two-dimensional chemical bond-to-chemical bond distance matrix. If the two chemical bonds are not connected, their distance is also set to a specific value of -1. (3) Calculate the shortest path distance between any atom and chemical bond in the drug molecule diagram, and use it as the atom-chemical bond distance matrix. Construct an extended diagram The node set of this extended graph From the original atomic nodes And the "edge node" representing each edge. Together they form a whole, where each edge node Corresponding to a primitive chemical bond in the molecular diagram ,Right now

[0062] The total number of nodes in the extended graph is .

[0063] The edge set of the extended graph All edges in the original graph are preserved. If the original graph... The edge in Connected atoms and So in Add two new edges: and The final extended graph is obtained. The adjacency matrix is ​​a The matrix is ​​represented as:

[0064] in: yes The original graph adjacency matrix, It is The incidence matrix defines the connection relationships between atoms and edges. Its elements... for:

[0065] yes of Transpose matrix. It is A zero matrix represents a matrix where there is no direct connection between edge nodes.

[0066] Calculated in the same way as the calculation of interatomic distances. The distance matrix is ​​obtained by finding the shortest path distance between all pairs of nodes. Finally, from the distance matrix Extract the sub-matrix corresponding to the distance from the atomic node to the "edge node" to form the atomic-chemical bond spatial position matrix. .matrix elements equal matrix From the atom To the "edge node" The distance. In the matrix In the middle, atoms Corresponding to the Yes, but "edge nodes" Corresponding to the Therefore, The definition is as follows:

[0067] in and .

[0068] make This yields the 2D-based distance matrix. :

[0069] in, yes The transpose of the matrix. Since the two-dimensional distances are all integers, a learnable scalar needs to be assigned to the values ​​in this matrix, which will serve as the bias term in the self-attention module.

[0070] Step S103: Calculate the three-dimensional spatial conformation matrix of the atomic-chemical bonds of the drug molecule.

[0071] The three-dimensional conformation of the drug molecule is obtained through Merck molecular force field (MMFF), and the conformation with the lowest energy is selected from the conformations, which represents the most stable state of the molecule under given conditions. Therefore, the distance matrix within the molecule is calculated using the three-dimensional conformation of the drug molecule: (1) The atom-to-atom distance matrix is ​​calculated based on the three-dimensional coordinates of the atoms, and the Euclidean distance is calculated for any two atoms. For a molecule containing A drug molecule of 1 atom can be obtained as a 3D matrix (2) Calculate the three-dimensional coordinates of the chemical bonds based on the three-dimensional coordinates of the atoms to obtain the chemical bond-chemical bond matrix. Specifically, the geometric center of the chemical bond is calculated from the arithmetic mean of the coordinates of the atoms at both ends of the chemical bond, and then the Euclidean distance of all chemical bond coordinates is calculated. For bonds containing A molecule with one chemical bond can obtain one 3D matrix (3) The three-dimensional distance between atoms and chemical bonds can be obtained from the atomic coordinates and chemical bond coordinates, and the atom-chemical bond matrix can be calculated. For those containing One atom and A molecule with 1 chemical bond can be obtained as a whole. 3D matrix .

[0072] Based on the above calculation process, the three-dimensional distance matrix can be obtained. :

[0073] in, yes The transpose of the matrix represents the Euclidean distances between atoms and chemical bonds within the molecule. The Euclidean distances are then encoded: for each atom pair... Their Euclidean distance is processed using the Gaussian kernel function:

[0074] in It refers to the number of Gorsky nuclei.

[0075] For each atom pair Calculate three-dimensional distance encoding :

[0076] in and , It is a learnable weight matrix, a vector By the learnable core center and learnable scaling factor Common definition. Furthermore... and It is an atomic type index pair Learnable scalar. Represented as a matrix form of three-dimensional distance encoding, its shape is as follows: .

[0077] Step S104: Integrate atomic and chemical bond features using virtual nodes.

[0078] To extract a representation of the entire molecule from atomic and chemical bond information, the model sets up virtual nodes for atoms. ] and virtual nodes for chemical bonds [ [, used to characterize atomic and chemical bond information within a molecule.] ] is considered to be connected to all atoms, while [ They are connected to all chemical bonds. That is, they can interact with atoms and chemical bonds respectively. Therefore, the two virtual nodes contain information about the atoms and chemical bonds within the molecule and can serve as a layer-level molecular representation. For drug molecules... Its molecular characterization It can be represented as:

[0079] in, and yes[ ]and[ The eigenvectors of ].

[0080] Step S105: Use the Graph Transformer to process the drug molecule structure features.

[0081] The ABT-DDI model uses a graph Transformer architecture to learn the edge and node relationships in drug molecules. The graph Transformer consists of stacked Transformer layers. Each Transformer layer comprises a normalized self-attention layer and a feedforward layer, which learns from node features. Sum of edge features For input, it is represented as .Will Represented as the first The layer's input, and definition . No. Attention matrix of layer It can be represented as:

[0082]

[0083] in , and It is a trainable weight matrix. H is the number of attention heads. It is the dimension of each head. and This refers to the two-dimensional and three-dimensional distance encoding matrices calculated earlier.

[0084] Step S106: Calculate the molecular fingerprint characteristics of the drug.

[0085] Molecular fingerprints represent drug structures as sequence bit vectors, providing a compressed, abstract view that can aid in molecular learning tasks. Morgan fingerprints finely encode local functional groups and structural features of a molecule by iteratively hashing the ever-expanding circular environment surrounding each atom. MACCS fingerprints resemble a list of 166 chemicals, each representing a predefined substructure indicating its presence in the molecule. RDKit fingerprints describe the overall topology and backbone connectivity of a molecule by exhaustively enumerating and hashing all linear atomic paths of a specific length within the molecule. The model uses these three types of molecular fingerprints to guide the structural characterization of molecules:

[0086]

[0087]

[0088]

[0089] in , These represent the two drug molecules input into the model. This represents the molecular fingerprint information.

[0090] Step S107: Output the predicted scores for drug multi-classification.

[0091] After passing through multiple Transformer layers, the model outputs the characterizations of the two drug molecules. and Combined with molecular fingerprint characterization Calculate the probability score of the interaction between the two drugs:

[0092]

[0093]

[0094]

[0095] in, and For a trainable weight matrix, and For the corresponding bias term, This represents the number of classification categories, and the activation function used is PReLU. The training objective is to minimize... Cross-entropy loss across categories:

[0096] in, For real labels, The probability prediction value output by the model. For the sample size, This represents the total number of categories.

[0097] Step S108: Compare the experimental performance of the ABT-DDI model with other baselines.

[0098] Since DDI prediction is a multi-classification task, four metrics were used for evaluation: accuracy (ACC), AUROC, AP, and F1-score. Because the overall performance of the model on the data was given more attention, all metrics were calculated using a micro-average method. ABT-DDI was evaluated against state-of-the-art baseline algorithms, including substructure-based and multimodal fusion-based methods: (1) GCN: Basic graph convolutional network model, learning drug characterization through neighborhood information aggregation. (2) GAT: Graph attention network, dynamically learning the importance weights between nodes using an attention mechanism. (3) MR-GNN: Employing a multi-resolution GNN architecture, integrating drug pair interaction features through LSTM. (4) SSI-DDI: Pioneering direct modeling of the original drug molecular graph, resolving pairwise interactions between substructures. (5) DeepDDS: Fusion of molecular graph structure and genomic features to identify key substructures. (6) DSN-DDI: Dual-view learning architecture, collaboratively optimizing the local and global representations of drug molecules. (7) MOTOR: Topology optimization representation learning method, integrating multi-granularity primitive information. (8) GraphSAGE: It pioneered a neighborhood sampling strategy to solve the generalization problem of new nodes through local feature aggregation. (9) MDF-SA-DDI: It combines multi-source fusion networks and Transformer self-attention to integrate heterogeneous features. (10) MathEagle: It combines heterogeneous graph convolution and multi-head attention to jointly model drug behavior and chemical structure.

[0099] To evaluate the effectiveness of the ABT-DDI model, experiments were conducted on two benchmark datasets, DDInter and DeepDDI. To ensure statistical significance, 5-fold cross-validation was performed on both datasets. The comparison results of ABT-DDI with the baseline methods are shown in Tables 1 and 2, with the best results highlighted in bold. Experimental results show that ABT-DDI significantly outperforms all comparison methods on both datasets in terms of ACC, AUROC, AP, and F1. Specifically, on DDInter, compared to the second-best method, ABT-DDI improves accuracy and AUROC by 11.37% and 2.59%, respectively. On DeepDDI, ABT-DDI improves by 3.18% and 5.77%, respectively. Notably, the model achieves AUROCs of 99.52% and 99.95% on both datasets, indicating that the model exhibits near-perfect DDI prediction capabilities. Furthermore, compared to SSI-DDI and DSN-DDI, two substructure-based methods, the ABT-DDI model demonstrates significant superiority. This indicates that the ABT-DDI model captures the interaction information of atoms and chemical bonds through a multi-scale attention mechanism, enhancing its ability to perceive drug substructures and thus exhibiting higher performance. These results demonstrate that ABT-DDI provides an advanced solution for DDI prediction tasks and has significant practical application value.

[0100] Table 1. Results of the ABT-DDI model and other baselines on the DDInter dataset.

[0101] Table 2. Results of the ABT-DDI model and other baselines on the DeepDDI dataset.

[0102] Step S109: Compare the ablation experimental performance of the ABT-DDI model and its variants.

[0103] To systematically evaluate the contributions of each key component in the ABT-DDI model, a comprehensive ablation experiment was conducted, and the following five variant models were designed for comparative analysis: (1) w / o 3D-bias: The intramolecular distance bias based on the three-dimensional conformation of the drug molecule (3D-bias) was removed, and only the distance bias based on the two-dimensional structure (2D-bias) was retained to verify the importance of the three-dimensional spatial structure information of the drug. (2) w / o 2D-bias: 2D-bias was disabled and only 3D-bias was used to evaluate the importance of the two-dimensional structure bias in characterizing the molecular topological relationship. (3) w / o bias: The 2D and 3D bias terms in each layer were completely removed to explore the role of spatial structure bias in the identification of key substructures during the hierarchical feature extraction process. (4) w / o virtual node: Virtual nodes were not used, and instead, global average pooling of drug node features was performed to generate drug representations, thereby evaluating the effectiveness of the virtual node mechanism for feature aggregation in the graph transformer architecture. (5) w / ofingerprint: Molecular fingerprint features are not used to verify the contribution of molecular fingerprint features to the model classification performance.

[0104] The ablation experimental results on two benchmark datasets are shown in Table 3 and Figures 3 to 6 As shown. It can be seen that the performance of all variant models decreased to varying degrees, which fully verifies the effectiveness of the design of each component in the ABT-DDI model. The following conclusions can be drawn: (1) The performance decrease caused by removing both 2D and 3D biases (w / o bias) is more significant, which is significantly higher than the impact of removing either bias alone (w / o 2D-bias or w / o 3D-bias). This shows that the two-dimensional topological structure and three-dimensional spatial information are complementary in characterizing drug molecules, and their combined effect helps the model to more comprehensively identify the interaction patterns between key atoms and chemical bonds. (2) On the DDInter dataset, removing the molecular fingerprint (w / o fingerprint) caused a significant decrease in model performance, far exceeding the impact of other components. This shows that the chemical functional groups and pharmacophore information encoded by the molecular fingerprint have indispensable discriminative value for the DDI prediction task. (3) After removing the virtual node, the model performance showed a stable decrease, indicating that the virtual node mechanism can more effectively integrate and refine the long-range dependencies and global context in the molecular graph, thereby generating a more discriminative drug representation.

[0105] Table 3. Results of the ABT-DDI model and its variants on the DDInter and DeepDDI datasets.

[0106] Step S110: Conduct a visualization experiment of the substructure.

[0107] To enhance the interpretability of model predictions, a hierarchical attention mechanism was integrated into the model. This mechanism, through dynamic allocation of attention weights, can identify substructure information in drug molecules closely related to DDI. To verify the model's effectiveness in practical applications, two pairs of drugs from the DDInter dataset were randomly selected: Acetylsalicylic acid and Nebivolol, and Bromfenac and Ziprasidone. Visual analysis was used to reveal the association between substructures and the DDI mechanism. The results are as follows: Figure 7 and Figure 8 As shown, when the model is untrained, the attention score distribution for drug molecule structures is relatively uniform, failing to identify key substructures that may lead to drug dissociation defects (DDI); however, the trained model can accurately locate important drug substructures. For example, for the drug pairs Acetylsalicylic acid and Nebivolol, from... Figure 7 The model successfully localized the carboxylic acid group (-COOH) and ester bond (-OCOCH3) of Acetylsalicylicacid, as well as the calcium ion binding site (Ca2+) in the Nebivolol molecule. This finding is highly consistent with known interaction mechanisms; the carboxylic acid group (-COOH), as a strong metal chelating site, can bind to the carboxylate group (-COO3) after its ionization. - ) and divalent cations (such as Ca) 2+ This chelation reaction forms insoluble salt complexes, which significantly reduces the gastrointestinal solubility of drugs, leading to impaired absorption. For the drugs Bromfenac and Ziprasidone, from... Figure 8 As can be seen from the training analysis, the bromophenylacetic acid structure of Bromfenac and the polyhydroxy structure of Ziprasidone are highlighted by the model. The carboxyl group of Bromfenac may form a hydrogen bond network with the hydroxyl group. This binding may hinder the active secretion of both in the renal tubules, thus potentially affecting the renal tubular excretion of the drug. Both sets of visualization analyses demonstrate that the model can accurately identify key substructures in drug molecules and reveal their potential interaction mechanisms, exhibiting good interpretability.

[0108] Step S111: Conduct a feature dimensionality reduction visualization experiment.

[0109] To evaluate the feature learning capability of the ABT-DDI model, the embedding space of drug pairs in the model is visualized. Specifically, high-dimensional feature representations of drug pairs generated by the model before and after training are extracted, and then projected into a two-dimensional space using t-SNE and UMAP dimensionality reduction methods for visualization and comparison. Figures 9 to 12As shown. In Figures 9 to 12 In the diagram, each data point represents a drug pair, and its category [Major, Moderate, Minor] is distinguished by color. Higher discriminative power among points of different categories indicates stronger feature discrimination ability of the model. Visualization results show that after dimensionality reduction using UMAP and t-SNE, the drug pair representations generated by the untrained model exhibit a disordered, mixed state in two-dimensional space, with no clear boundaries between categories, indicating a lack of meaningful semantic structure in the initial features. In contrast, the feature distribution of the trained model demonstrates significant intra-class clustering and inter-class separation, with clear boundaries between the three categories. This comparison intuitively proves that the model successfully learns discriminative feature representations through training, enabling it to extract more discriminative feature patterns from the original data.

[0110] Example 2 like Figure 13 As shown, this embodiment provides a drug interaction prediction system based on atomic and chemical bond structures, including: The acquisition module is used to acquire the atomic characteristics, chemical bond characteristics, molecular fingerprint characteristics, two-dimensional structural information, and three-dimensional spatial conformation information of each drug molecule; The calculation module is used to calculate the two-dimensional distance matrix of atoms and chemical bonds based on the two-dimensional structural information; and to calculate the three-dimensional distance matrix of atoms and chemical bonds based on the three-dimensional spatial conformation information. The Graph Transformer module is used to introduce the two-dimensional and three-dimensional distance matrices as spatial prior biases into the network's self-attention mechanism using a graph Transformer network to process the atomic features and chemical bond features, thereby obtaining the structural features of each drug molecule. The prediction module is used to output a prediction score of drug molecule interactions based on the structural features and molecular fingerprint features of drug molecules.

[0111] Example 3 Embodiment 3 of the present invention provides an electronic device.

[0112] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps in the drug interaction prediction method based on atomic and chemical bond structures as described in Embodiment 1 of the present invention.

[0113] The detailed steps are the same as those of the drug interaction prediction method based on atomic and chemical bond structures provided in Example 1, and will not be repeated here.

[0114] Example 4 Embodiment 4 of the present invention provides a computer-readable storage medium.

[0115] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the drug interaction prediction method based on atomic and chemical bond structures as described in Embodiment 1 of the present invention.

[0116] The detailed steps are the same as those of the drug interaction prediction method based on atomic and chemical bond structures provided in Example 1, and will not be repeated here.

[0117] Example 5 Embodiment 5 of the present invention provides a computer program product.

[0118] A computer program product includes software code, wherein the program in the software code performs the steps in the drug interaction prediction method based on atomic and chemical bond structures as described in Embodiment 1 of the present invention.

[0119] The detailed steps are the same as those of the drug interaction prediction method based on atomic and chemical bond structures provided in Example 1, and will not be repeated here.

[0120] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0121] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0122] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0123] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0124] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0125] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0126] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A method for predicting drug interactions based on atomic and chemical bond structures, characterized in that, include: Acquire atomic features, chemical bond features, molecular fingerprint features, two-dimensional structural information, and three-dimensional spatial conformation information for each drug molecule; Based on the aforementioned two-dimensional structural information, the two-dimensional distance matrix between atoms and chemical bonds is calculated; Based on the aforementioned three-dimensional spatial conformation information, the three-dimensional distance matrix between atoms and chemical bonds is calculated; The three-dimensional distance matrix for calculating atoms and chemical bonds includes: The three-dimensional conformation of drug molecules is obtained through Merck molecular force field, and the conformation with the lowest energy is selected. Calculate the atom-to-atom Euclidean distance matrix based on the three-dimensional coordinates of atoms; Based on the three-dimensional coordinates of chemical bonds, calculate the Euclidean distance matrix between chemical bonds, where the coordinates of the chemical bonds are calculated by the arithmetic mean of the coordinates of the atoms at both ends of the chemical bond; Calculate the Euclidean distance matrix between atoms and chemical bonds based on atomic coordinates and chemical bond coordinates; Encode the obtained Euclidean distance for each atom pair. Their Euclidean distance is processed using the Gaussian kernel function: ； in, It is the number of Gorsky nuclei; For each atom pair Calculate the three-dimensional distance code : ； in and , It is a learnable weight matrix, a vector By the learnable core center and learnable scaling factor Common definition; in addition, and It is an atomic type index pair Learnable scalars; Virtual atomic nodes connected to all atoms and virtual chemical bond nodes connected to all chemical bonds are added to the two-dimensional distance matrix and the three-dimensional distance matrix. The virtual atomic nodes interact with all atomic nodes to aggregate the global structural information of atoms, and the virtual chemical bond nodes interact with all chemical bond nodes to aggregate the global structural information of chemical bonds. Using a graph Transformer network, the two-dimensional distance matrix and the three-dimensional distance matrix are introduced as spatial prior biases into the network's self-attention mechanism to process the atomic features and chemical bond features, thereby obtaining the structural features of each drug molecule. Based on the structural features and molecular fingerprint features of drug molecules, a prediction score for drug molecule interactions is output.

2. The method as described in claim 1, characterized in that, The two-dimensional distance matrix for calculating atoms and chemical bonds includes: Calculate the atom-to-atom distance matrix, where for any two atom nodes in the molecular graph, the shortest path distance is calculated as the atom-to-atom distance. If the two atom nodes are not connected, their distance is set to a specific value. The chemical bond-chemical bond distance matrix is ​​calculated by constructing a dual graph with chemical bonds as nodes and atoms as edges. The shortest path distance between any two chemical bonds is calculated on the dual graph. If two chemical bonds are not connected, their distance is set to a specific value. The atomic-chemical bond distance matrix is ​​calculated by constructing an extended graph, the node set of which includes the original atomic nodes and edge nodes representing chemical bonds. The shortest path distance from the atomic nodes to the edge nodes in the extended graph is calculated to form the atomic-chemical bond spatial position matrix.

3. The method as described in claim 1, characterized in that, The step of introducing the two-dimensional distance matrix and the three-dimensional distance matrix as spatial prior biases into the network's self-attention mechanism specifically includes: In the self-attention layer of the graph Transformer network, the attention score is calculated according to the following formula: ； in, , K, V These are query, key, and value matrices, respectively. d For feature dimension, and These are the two-dimensional distance matrix and the three-dimensional distance matrix, respectively.

4. The method as described in claim 1, characterized in that, The molecular fingerprint feature is obtained through the following steps: The Morgan fingerprint, RDKit fingerprint, and MACCS fingerprint of the drug molecule are calculated; the three fingerprint vectors are concatenated; the concatenated vector is input into a multilayer perceptron for fusion and dimensionality reduction to obtain the final molecular fingerprint feature representation.

5. A drug interaction prediction system based on atomic and chemical bond structure, characterized in that, include: The acquisition module is used to acquire the atomic characteristics, chemical bond characteristics, molecular fingerprint characteristics, two-dimensional structural information, and three-dimensional spatial conformation information of each drug molecule; The calculation module is used to calculate the two-dimensional distance matrix of atoms and chemical bonds based on the two-dimensional structure information; Based on the aforementioned three-dimensional spatial conformation information, the three-dimensional distance matrix between atoms and chemical bonds is calculated; The three-dimensional distance matrix for calculating atoms and chemical bonds includes: The three-dimensional conformation of drug molecules is obtained through Merck molecular force field, and the conformation with the lowest energy is selected. Calculate the atom-to-atom Euclidean distance matrix based on the three-dimensional coordinates of atoms; Based on the three-dimensional coordinates of chemical bonds, calculate the Euclidean distance matrix between chemical bonds, where the coordinates of the chemical bonds are calculated by the arithmetic mean of the coordinates of the atoms at both ends of the chemical bond; Calculate the Euclidean distance matrix between atoms and chemical bonds based on atomic coordinates and chemical bond coordinates; Encode the obtained Euclidean distance for each atom pair. Their Euclidean distance is processed using the Gaussian kernel function: ； in, It is the number of Gorsky nuclei; For each atom pair Calculate the three-dimensional distance code : ； in and , It is a learnable weight matrix, a vector By the learnable core center and learnable scaling factor Common definition; in addition, and It is an atomic type index pair Learnable scalars; Virtual atomic nodes connected to all atoms and virtual chemical bond nodes connected to all chemical bonds are added to the two-dimensional distance matrix and the three-dimensional distance matrix. The virtual atomic nodes interact with all atomic nodes to aggregate the global structural information of atoms, and the virtual chemical bond nodes interact with all chemical bond nodes to aggregate the global structural information of chemical bonds. The Graph Transformer module is used to introduce the two-dimensional and three-dimensional distance matrices as spatial prior biases into the network's self-attention mechanism using a graph Transformer network to process the atomic features and chemical bond features, thereby obtaining the structural features of each drug molecule. The prediction module is used to output a prediction score of drug molecule interactions based on the structural features and molecular fingerprint features of drug molecules.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the drug interaction prediction method based on atomic and chemical bond structures as described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the steps of the drug interaction prediction method based on atomic and chemical bond structures as described in any one of claims 1 to 4.

8. A computer program product, comprising software code, characterized in that, The program in the software code performs the steps of the drug interaction prediction method based on atomic and chemical bond structures as described in any one of claims 1 to 4.

Images