Drug interaction prediction method and device for ADME process and medium

Abstract

The invention discloses a drug interaction prediction method and device for an ADME process and a medium, and relates to the technical field of medicine. The method is applied to a drug interaction prediction model obtained by training a TRACE model, and comprises the processes of enhanced molecular map construction, feature extraction and prediction. According to the method, external chemical knowledge (chemical element knowledge graph) is introduced in a molecular graph construction process, so that the modeling capability of the model on a molecular complex structure and a long-range dependency relationship is improved, and the adaptability to a traditional Chinese medicine complex component system is remarkably enhanced; the TRACE model adopted by the invention can automatically identify high-risk substructures closely related to the ADME process, and a scientific and efficient solution is provided for safety evaluation and druggability prediction of drug combination of traditional Chinese medicine components and chemical drugs thereof. The risk of molecular interaction in the new drug research and development process and the traditional Chinese medicine modernization process can be effectively reduced, and the improvement of the drug safety management level is promoted.

Description

Technical Field

[0001] This application relates to the field of pharmaceutical technology, and in particular to a method, device and medium for predicting drug interactions in the ADME process. Background Technology

[0002] With the global aging population and the high incidence of complex diseases, combination therapy and integrated traditional Chinese and Western medicine treatments are becoming increasingly common. Whether it's chemical drugs or traditional Chinese medicine ingredients, the risks of molecular interactions arising from multi-drug combinations are constantly increasing, especially in the absorption-distribution-metabolism-excretion (ADME) stages, which have become critical issues that urgently need to be addressed in clinical safety and drug development. Therefore, technologies for predicting and assessing molecular interactions are continuously developing and are gradually becoming an important support for modern drug research and clinical practice.

[0003] In recent years, Graph Neural Networks (GNNs) have made groundbreaking progress in the fields of medicinal chemistry feature extraction and molecular interaction prediction due to their ability to effectively model molecular structural information. Deep learning methods based on molecular graphs, utilizing graphical representations of molecular structures, have significantly improved the accuracy and automation of molecular interaction prediction. However, most existing methods rely primarily on GNNs and their variants. While these models can extract information from atoms and their neighborhoods, they have limitations in capturing complex long-range dependencies and global structural features within molecules, resulting in insufficient predictive power for structurally complex or novel molecules (such as active ingredients in traditional Chinese medicine). Furthermore, most current mainstream models are data-driven, lacking deep integration with chemical knowledge, focusing only on the topological relationships between atoms in the molecular graph while neglecting the richer chemical semantics within the molecule, leading to insufficient model interpretability and generalization ability.

[0004] Traditional Chinese medicine (TCM) is characterized by its complex composition and diverse structures, with multiple potential molecular interactions among its active components. Many of its components, similar to modern drugs, contain key high-risk substructures involved in the absorption, distribution, metabolism, and excretion (ADME) processes. Therefore, traditional methods are insufficient for evaluating the safety and drug-likeness of TCM components. There is an urgent need for intelligent methods capable of automatically identifying and evaluating high-risk ADME structures in TCM to improve the scientific validity and safety of TCM formulations and combined drug use. Summary of the Invention

[0005] The purpose of this application is to provide a method, device, and medium for predicting drug interactions in the ADME process, so as to achieve accurate prediction of the interaction relationship between drug molecules and reduce the risk of molecular interaction in the process of new drug development and modernization of traditional Chinese medicine.

[0006] To achieve the above objectives, this application provides the following solution.

[0007] Firstly, this application provides a method for predicting drug interactions during the ADME process. This method utilizes a drug interaction prediction model, which is obtained by training a TRACE model. The drug interaction prediction model includes a molecular map generation module, a molecular feature extraction module, and a prediction module. The method for predicting drug interactions during the ADME process includes the following steps: Obtain chemical species structure data for the first and second drugs; Based on the chemical species structure data and chemical element knowledge graph of the first drug, an enhanced molecular graph of the first drug is constructed using the molecular graph generation module; Based on the chemical species structure data of the second drug and the chemical element knowledge graph, an enhanced molecular graph of the second drug is constructed using the molecular graph generation module; The enhanced molecular map of the first drug and the enhanced molecular map of the second drug are processed using the molecular feature extraction module to obtain the feature representations of the first drug and the second drug. Based on the feature representations of the first drug and the second drug, the prediction module is used to predict the probability of interaction between the first drug and the second drug.

[0008] Optionally, the molecular graph generation module includes an RDKit unit, a DGL unit, and an embedding unit. Based on the chemical species structure data and chemical element knowledge graph of the first drug, the molecular graph generation module is used to construct an enhanced molecular graph of the first drug, specifically including: The RDKit unit is used to perform feature calculations on the chemical species structure data of the first drug to obtain the molecular characteristics of the first drug; Based on the molecular characteristics of the first drug, a molecular map of the first drug is generated using the DGL unit; Based on the chemical element knowledge graph, the embedding unit is used to embed the chemical element features contained in the first drug into the molecular graph of the first drug to obtain an enhanced molecular graph of the first drug.

[0009] Optionally, the molecular feature extraction module applies a graph converter, which is used to dynamically aggregate local and global information in the enhanced molecular graph of the target drug using a self-attention mechanism to obtain a feature representation of the target drug; the target drug is a first drug or a second drug.

[0010] Optionally, local and global information from the enhanced molecular map of the target drug can be dynamically aggregated using a self-attention mechanism to obtain a feature representation of the target drug, specifically including: The initial node hiding representation and the initial edge hiding representation are obtained by linearly projecting the node features and edge features in the enhanced molecular graph of the target drug. The initial node hidden representation and the initial edge hidden representation are updated in multiple layers using a self-attention mechanism to obtain the final updated node hidden representation and edge hidden representation, which together form the feature representation of the target drug.

[0011] Optionally, a self-attention mechanism is used to perform multi-layer feature updates on the initial node hidden representation and the initial edge hidden representation to obtain the final updated node hidden representation and edge hidden representation, specifically including: Using batch normalization layer for the first The updated node and edge hidden representations are normalized, and the hidden representations of queries, keys, values, and edges are calculated in the multi-head self-attention mechanism. Calculate the attention score based on the hidden representations of queries, keys, and edges; Based on the value and attention score, the first Update the node hiding representation and edge hiding representation after the layer update to obtain the th layer. The intermediate representation of nodes and the intermediate representation of edges in a layer; The first The intermediate representations of nodes and edges in the first layer are passed to the feedforward network for residual connections and normalization to obtain the intermediate representations of the second layer. The updated node hiding representation and edge hiding representation.

[0012] Optionally, the batch normalization layer is used for the first... The updated node and edge hidden representations are normalized, and in the multi-head self-following mechanism, the formulas for calculating the hidden representations of queries, keys, values, and edges are as follows: ; ; ; ; in, k This indicates the first step in the multi-head self-following mechanism. k One thing to note: Indicates the first Layer self-attention mechanism, and Both represent nodes. , , , These are the hidden representations of queries, keys, values, and edges, respectively. , , and These are the projection matrices for the query, key, value, and edge, respectively. and The first The updated node hiding representation and edge hiding representation after layer update. Indicates normalization; The formula for calculating attention score is: ; in, This represents the attention score. Indicates the first k One dimension to pay attention to. for function; Based on the value and attention score, the first Update the node hiding representation and edge hiding representation after the layer update to obtain the th layer. The formulas for the intermediate representation of nodes and the intermediate representation of edges in a layer are as follows: ; ; in, and They represent the first The intermediate representation of nodes and the intermediate representation of edges in a layer. For the first The nodes of the layer output projection matrix. This is a random deactivation function used to prevent overfitting. For the first The edge output projection matrix of the layer, Indicates a connection. This indicates that attention should be paid to the number of heads. Represents a node i The number of adjacent nodes; The first The intermediate representations of nodes and edges in the first layer are passed to the feedforward network for residual connections and normalization to obtain the intermediate representations of the second layer. The formulas for the updated node hiding representation and edge hiding representation are as follows: ; ; in, and The first The updated node hiding representation and edge hiding representation after layer update. and For the first The edge-forward network weight matrix of the layer, where, Used for dimensionality increase or feature transformation Used to map back edge representation space For activation function, and For the first The node feedforward network weight matrix of the layer, where, Used for dimensionality increase or feature transformation Used to map back to the node representation space.

[0013] Optionally, the prediction module is used to project the feature representations of the first drug and the feature representations of the second drug onto a multilayer perceptron to obtain the interaction probability, as shown in the formula: ; in, Represents the probability of interaction. This is a characteristic representation of the first drug. This is a characteristic representation of the second drug. This represents a multilayer perceptron.

[0014] Optionally, the loss function used when training the TRACE model is the cross-entropy loss function.

[0015] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for predicting drug interactions in the ADME process.

[0016] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for predicting drug interactions in the ADME process.

[0017] According to the specific embodiments provided in this application, this application has the following technical effects.

[0018] This application provides a method, device, and medium for predicting drug interactions during the ADME process. The method utilizes a drug interaction prediction model obtained by training a TRACE model, including enhanced molecular graph construction, feature extraction, and prediction processes. This application introduces external chemical knowledge (a chemical element knowledge graph) during molecular graph construction, which not only improves the model's ability to model complex molecular structures and long-range dependencies but also significantly enhances its adaptability to complex traditional Chinese medicine (TCM) component systems. The TRACE model used in this application can automatically identify high-risk substructures closely related to the ADME process, providing a scientific and efficient solution for the safety evaluation and drug development prediction of TCM components and their combination with chemical drugs. This method can effectively reduce the risks of molecular interactions in new drug development and the modernization of TCM, promoting the improvement of drug safety management. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a drug interaction prediction method for the ADME process provided in one embodiment of this application; Figure 2 A schematic diagram illustrating the working principle of a TRACE model provided in an embodiment of this application; Figure 3 A high-risk substructure diagram related to metabolism found by the TRACE model provided in one embodiment of this application; Figure 4 A high-risk substructure diagram related to absorption, distribution, and excretion, as suggested by the TRACE model provided in an embodiment of this application; Figure 5 A diagram showing the results of predicting molecular interactions and docking of tanshinone IIA using the TRACE model, provided as an embodiment of this application. Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] In an exemplary embodiment, a method for predicting drug interactions during the ADME process is provided. This method applies a drug interaction prediction model, which is obtained by training a TRACE model. The drug interaction prediction model includes a molecular map generation module, a molecular feature extraction module, and a prediction module. Figure 1 As shown, the drug interaction prediction method for the ADME process includes the following steps 101-105.

[0024] Step 101: Obtain chemical species structure data of the first drug and the second drug.

[0025] Step 102: Based on the chemical species structure data and chemical element knowledge graph of the first drug, construct an enhanced molecular graph of the first drug using the molecular graph generation module.

[0026] Step 103: Based on the chemical species structure data of the second drug and the chemical element knowledge graph, construct an enhanced molecular graph of the second drug using the molecular graph generation module.

[0027] Step 104: The enhanced molecular map of the first drug and the enhanced molecular map of the second drug are processed by the molecular feature extraction module to obtain the feature representations of the first drug and the second drug.

[0028] Step 105: Based on the feature representations of the first drug and the second drug, the prediction module is used to predict the interaction probability of the first drug and the second drug.

[0029] The TRACE model in the above embodiments has the following advantages over the prior art: (1) Transformer-based Graph Representation LeArning with Chemical Embedding (TRACE) integrates elemental knowledge graphs to enhance knowledge representation. TRACE incorporates elemental knowledge graphs into the model, providing a structured and systematic knowledge foundation for a deeper understanding of drug properties and molecular interaction mechanisms.

[0030] (2) TRACE uses Graph Transformer to improve structural modeling capabilities. Unlike traditional methods that mainly rely on GNN and its variants, TRACE utilizes the global self-attention mechanism of Graph Transformer to efficiently capture complex structural features and long-range dependencies of molecular graphs, significantly improving the global learning capability of molecular structures.

[0031] (3) TRACE performs well in multiple tasks and has wide applicability. Whether in multi-class or binary classification tasks of molecular interactions, TRACE demonstrates better predictive performance than existing mainstream methods.

[0032] (4) TRACE intelligently identifies high-risk substructures to assist in the safety and drugability evaluation of traditional Chinese medicine. TRACE can automatically identify high-risk molecular substructures related to the absorption, distribution, metabolism, and excretion (ADME) processes, and is equally applicable to traditional Chinese medicine components. It reveals potential interaction risks and drugability barriers in the complex system of traditional Chinese medicine, and provides intelligent scientific support for the safety evaluation and rational compatibility of traditional Chinese medicine.

[0033] Currently, most existing molecular interaction prediction models are based on the topological structure of molecules, transforming molecular structures into graphs for modeling and training. While this method is convenient for representing local relationships between molecules, it has significant shortcomings in terms of knowledge-level representation and modeling of global molecular structural relationships. The introduction of elemental knowledge graphs can supplement domain knowledge beyond molecular graphs, effectively improving the model's reasoning ability and interpretability. Simultaneously, the global information aggregation characteristic of GraphTransformer overcomes the limitation of traditional GNNs, which can only process local neighborhood information, enabling the model to learn the overall structural features of molecules more comprehensively, thereby improving the accuracy and generalization ability of molecular interaction prediction. TRACE not only performs excellently in the field of chemical drugs but can also be extended to systems with complex components and diverse structures in traditional Chinese medicine (TCM), enabling intelligent identification and drug-likeness evaluation of high-risk structures related to ADME in TCM, providing strong support for the modernization and safe use of TCM.

[0034] The methods described in the above embodiments are not only applicable to chemical drug molecules, but also fully consider the molecular interaction mechanisms in processes such as absorption, distribution, metabolism, and excretion (ADME), and can intelligently identify high-risk substructures, thereby supporting the drug-likeness evaluation and safety prediction of traditional Chinese medicine molecules.

[0035] The TRACE framework consists of three main modules: a molecular map generation module, a molecular feature extraction module, and a prediction module (such as...). Figure 2 (As shown in a, b, and c). During the molecular graph construction process, elemental knowledge from ElementKG (Element Knowledge Graph) is embedded into the original drug molecule graph to generate an enhanced molecular graph that combines molecular structural features with chemical domain knowledge. This knowledge-enhanced molecular representation not only expands the breadth and depth of molecular description but also provides a solid information foundation for subsequent identification of ADME-related high-risk substructures and intelligent evaluation of drug drugability.

[0036] After obtaining the enhanced molecular graph, the model further inputs it into the Graph Transformer module. This module dynamically aggregates local and global information from the molecular graph through a self-attention mechanism, effectively uncovering complex structures and chemical patterns closely related to the ADME process and molecular interactions. Finally, the model concatenates the high-level characterizations of the two drug molecules and inputs them into a downstream multilayer perceptron (MLP) for molecular interaction prediction and druggability risk indication.

[0037] Based on this innovative framework, TRACE not only demonstrates excellent molecular interaction prediction capabilities in the field of chemical drugs, but also adapts to the complex and diverse traditional Chinese medicine system. It provides an efficient and reliable technical solution for the intelligent assessment of the compatibility safety of traditional Chinese medicine and ADME-related drugability barriers, thus promoting the modernization of traditional Chinese medicine and the scientific process of safe drug management.

[0038] In another exemplary embodiment, in step 101 above, the raw data presented in the format of the Simplified Molecular Input Line Entry System (SMILES), namely chemical species structure data, wherein SMILES is a line notation system that uses ASCII strings to describe the structure of chemical species.

[0039] In another exemplary embodiment, in steps 102 and 103 above, feature calculations are performed using RDKit (a cheminformatics toolkit), and a molecular graph is generated using a Deep Graph Library (DGL). Specifically, the molecular features mainly include the properties of atoms and chemical bonds. Each atom is represented by a 63-dimensional feature vector, including chemical properties such as atom type, chirality, and hybridization state. The features of each chemical bond are represented by a 6-dimensional vector, including the chemical bond type, conjugation state, and ring state.

[0040] Furthermore, chemical element knowledge is embedded into atomic feature vectors to create chemically enhanced molecular graphs. Specifically, embodiments of this application utilize element embeddings from ElementKG (a knowledge graph constructed from chemical elements) to add these embeddings to the atomic feature vectors. In TRACE, atomic features interact dynamically through an attention mechanism, updating atomic features by aggregating neighborhood information, thereby enhancing molecular representation. Therefore, the input for each atom is initialized as a 196-dimensional vector, thus optimizing the integration of structural and chemical knowledge in the representation.

[0041] Table 1 Feature extraction of atoms and bonds

[0042] In another exemplary embodiment, the molecular graph generation module includes an RDKit unit, a DGL unit, and an embedding unit. Based on the chemical species structure data and chemical element knowledge graph of the first drug, the molecular graph generation module is used to construct an enhanced molecular graph of the first drug. Step 102 above can be replaced by steps 201-203 as follows.

[0043] Step 201: Use the RDKit unit to perform feature calculations on the chemical species structure data of the first drug to obtain the molecular characteristics of the first drug.

[0044] Step 202: Based on the molecular characteristics of the first drug, generate a molecular map of the first drug using the DGL unit.

[0045] Step 203: Based on the chemical element knowledge graph, the embedding unit is used to embed the chemical element features contained in the first drug into the molecular graph of the first drug to obtain an enhanced molecular graph of the first drug.

[0046] The specific implementation method of step 103 is the same as that of step 102, and will not be repeated here.

[0047] In another exemplary embodiment, the molecular feature extraction module (which applies a GraphTransformer) is used to dynamically aggregate local and global information in the enhanced molecular graph of the target drug using a self-attention mechanism to obtain a feature representation of the target drug; the target drug is a first drug or a second drug.

[0048] In another exemplary embodiment, local and global information in the enhanced molecular graph of the target drug are dynamically aggregated using a self-attention mechanism to obtain a feature representation of the target drug, specifically including: The initial node hiding representation and the initial edge hiding representation are obtained by linearly projecting the node features and edge features in the enhanced molecular graph of the target drug.

[0049] The initial node hidden representation and the initial edge hidden representation are updated in multiple layers using a self-attention mechanism to obtain the final updated node hidden representation and edge hidden representation, which together form the feature representation of the target drug.

[0050] In another exemplary embodiment, a self-attention mechanism is used to perform multi-layer feature updates on the initial node hidden representation and the initial edge hidden representation to obtain the last layer of updated node hidden representation and edge hidden representation, specifically including the following steps 301-304.

[0051] Step 301, use the batch normalization layer to process the first... The updated node and edge hidden representations are normalized, and the hidden representations of queries, keys, values, and edges are calculated in the multi-head self-attention mechanism.

[0052] Step 302: Calculate the attention score based on the hidden representations of the query, key, and edge.

[0053] Step 303, based on the value and attention score, for the first... Update the node hiding representation and edge hiding representation after the layer update to obtain the th layer. The intermediate representation of nodes and the intermediate representation of edges in a layer.

[0054] Step 304, will the first The intermediate representations of nodes and edges in the first layer are passed to the feedforward network for residual connections and normalization to obtain the intermediate representations of the second layer. The updated node hiding representation and edge hiding representation.

[0055] In another exemplary embodiment, the molecular feature extraction module described above is specifically implemented as follows: First, the graph transformation process is initialized by preparing input nodes and edge embeddings. For Figure 2 Each node in The node features are represented as follows: For nodes and Each edge between them has the following edge features: Then, these nodes and edge features are transformed by two independent linear projections. Dimensional hidden representation, node is , side is : ; ; in, These are the weights and biases of the linear projection layer. The overall architecture of the graph converter layer is as follows: Figure 2 As shown in c, the improved multi-head self-attention (MHA) mechanism is mainly relied upon to update the characteristics of each node and edge in the graph. Taking the first... Taking layers as an example, the hidden representation of nodes and edges and First, normalization is performed using a batch normalization layer. Then, in the multi-head self-focus (MHA) mechanism, the corresponding query is calculated. ,key ,value Hidden representation of edges : ; ; ; ; in, All , This indicates that attention should be paid to the number of heads. The dimension of each attention head is equal to Divide by , Represents a node The adjacent nodes. The self-attention mechanism calculates the attention weight between the query and the key press, and then multiplies it by that value. The output attention score is expressed as: ; To ensure numerical stability, the output after taking the exponent values ​​of the softmax inner terms is clamped between -5 and +5. Then, the intermediate representations of nodes and edges are calculated: ; ; in Indicates a connection.

[0056] Then, output and The results are passed to the feedforward network, which is then replaced by residual connections and a normalization layer. ; ; in, ， This represents the intermediate representation. The node representation obtained from the last layer of the graph converter will be passed to downstream tasks that predict molecular interactions.

[0057] In another exemplary embodiment, the prediction module described above is implemented as follows: Given a set of all drugs Consider a pair of drugs ,in The enhanced molecular map is processed using a molecular feature extraction module. and To extract its feature representation and : ; ; in, and These represent the chemically enhanced atom and chemical bond features, respectively. The final interaction probability is calculated using multilayer perceptron (MLP) projection: ; In another exemplary embodiment, the TRACE model uses a cross-entropy loss function, specifically calculating the cross-entropy loss of all molecule pairs in the drug pairing set using the following formula. : ; in, Represents a drug pairing set. This is a basic, authentic label.

[0058] In another exemplary embodiment, the above-described drug interaction prediction model can be used for high-risk substructure identification, as follows: To identify key substructures in drug molecules that pose a high risk of molecular interactions, this application employs an automated analysis pipeline based on the attention scores output by the model. Specifically, each drug molecule is first decomposed into several substructures using a molecular decomposition algorithm, and then the model's atomic-level average attention score is extracted. For each substructure, the sum and average of the attention scores of its constituent atoms are calculated to obtain the total and average scores for each substructure. By sorting the substructures according to their average scores from highest to lowest, the high-risk substructures most likely to attract the model's attention are identified. Substructures with higher scores are considered to contribute more to the model's assessment of molecular interaction risk. The substructure decomposition algorithm used in this application is inspired by retrosynthetic analysis, systematically decomposing complex molecules into representative chemical substructures based on a series of typical chemical reaction rules.

[0059] To more scientifically and comprehensively evaluate the performance of TRACE, this application's embodiments designed a multi-level experimental scheme. Specifically, the experiments covered four datasets of different sizes, testing two task types: binary classification and multi-class classification. Furthermore, various scenarios, such as in-distribution and out-of-distribution, were set up to systematically verify the model's generalization ability and stability under different application conditions.

[0060] In another exemplary embodiment, an experiment is set up as follows: In the binary classification task, this application's embodiments evaluated the performance of TRACE on three widely used datasets: ZhangDDI, ChChMiner, and DeepDDI. The ZhangDDI dataset includes 543 drug molecules and 95,245 drug interaction (DDI) pairs; the ChChMiner dataset includes 997 drug molecules and 83,918 DDI pairs; and the DeepDDI dataset includes 1,704 drug molecules and 265,771 DDI pairs. Experiments were conducted in two data partitioning settings: conduction-based (in-distribution) and conduction-based (out-of-distribution).

[0061] In the transitive setup, each dataset is randomly divided into a training set, a validation set, and a test set in a 6:2:2 ratio, ensuring that drugs appearing in the test set also appear in the training set. For each dataset, this embodiment repeats the experiment five times using different random seeds, and reports the mean and standard deviation of the experimental results. The optimal test performance is defined as the model achieving its highest performance on the validation set.

[0062] In the inductive setting, embodiments of this application evaluated the model's ability to generalize to novel drugs not included in the training set.

[0063] In the inductive setting, this embodiment focuses on the DeepDDI dataset to evaluate the model's generalization ability to novel drugs not included in the training set. Specifically, it lets... This represents the collection of all drugs in DeepDDI. The embodiments of this application will... The subsets are randomly divided into five mutually exclusive subsets of approximately equal size: For all For the first A fold, as specified in the embodiments of this application. For testing drug sets, To train the drug set, for each fold, construct the following drug pairing dataset: Training set: drug pairs ,in .

[0064] S1 set (single unseen set): drug pairs ,in or However, it cannot be both.

[0065] S2 set (double hidden set): drug pair ,in .

[0066] During each fold, the model is trained only on drug pairs where both drugs are in the training set. It is then evaluated on drug pairs where at least one or both drugs are in the test set to measure the model's generalization ability to unseen drugs. This process is repeated five times, with each subset serving as the test set once. The final performance is reported as the mean and standard deviation over the five folds.

[0067] In multi-class classification tasks, to predict interaction types, embodiments of this application utilize the widely accepted Drugbank dataset, which contains 191,808 DDI records involving 1,706 drugs and 86 interaction types. For each dataset, experiments were repeated five times using different random seeds, and the mean and standard deviation of the experimental results were reported. The optimal test performance is defined as the model achieving its highest performance on the validation set.

[0068] In another exemplary embodiment, another experiment is set up as follows: During training, initial atomic and bond features are first mapped to a 128-dimensional hidden space using linear projection, and then input into the graph transformer module for training. The graph transformer consists of eight layers; the first seven layers utilize a multi-head self-attention mechanism to simultaneously update node and edge features, while the last layer focuses on refining node features. Each layer consists of the following key components: first, parallel multi-head attention computation, where four heads work in parallel, allowing the model to capture different aspects of the data simultaneously; second, BatchNorm, which normalizes the attention output; and third, non-linear projection (input and output dimensions are 128, inner dimension is 256), using SiLU activation with a 0.1 dropout rate and quadratic normalization.

[0069] The model is trained on the training set, and an early stopping mechanism is implemented based on performance on the validation set. The model is trained using the Adam optimizer with a batch size of 1024 and a learning rate of 0.001. TRACE is implemented in PyTorch and runs on a Linux server equipped with an NVIDIA A10 GPU.

[0070] In another exemplary embodiment, three specific sets of test methods are provided, along with test results.

[0071] Case Study 1: To systematically reveal the specific role of molecular structure in molecular interactions, the TRACE model was used to assign attention-based weights to different nodes in drug molecules, and a substructure decomposition algorithm was combined to automatically identify high-risk molecular fragments that may easily trigger molecular interactions during absorption, metabolism, distribution, and excretion.

[0072] Taking metabolism as an example, Figure 3 In the diagram, 'a' represents a flowchart illustrating the process of identifying high-interest substructures using a pre-trained model and a substructure decomposition algorithm. Figure 3 In the diagram, 'b' represents an example of attention visualization at the substructure level for four different molecules, along with an analysis of their functional mechanisms. Darker colors indicate higher attention weights received by that substructure. Figure 3As shown, TRACE highlights key functional groups in representative molecules by assigning higher attention weights to substructures known to readily induce metabolic-related molecular interactions. (1) The first example is the drug quinine, where the quinoline functional group was given a relatively high attention weight. Studies have shown that quinoline derivatives can coordinate with heme iron in CYP enzymes, acting as reversible inhibitors and thus reducing the metabolism of co-administered drugs, which is significant. (2) The second example is ziloutong, where the model gives more attention to the thiophene substructure. Studies have shown that P450 enzymes can oxidize thiophene to generate highly active metabolites, which readily interact with other cellular components and affect metabolic processes. (3) The third example is erythromycin, where the amine substructure receives the most attention. Studies have shown that under the catalysis of P450 enzymes, amines can be converted into nitrosoferric complexes, leading to enzyme inactivation and thus affecting the metabolism of other drugs. (4) The fourth example is cistanol, where the methylene dioxyphenyl substructure receives the most attention. It has been reported that this substructure can form a metabolic intermediate complex (MIC) with P450 enzymes through its metabolites, leading to quasi-irreversible inactivation of P450, which in turn affects the metabolism of other drugs, thereby triggering drug-drug interactions.

[0073] Besides metabolism, there are also absorption, distribution, and excretion processes. TRACE indicates potentially high-risk substructures in all of these. Specifically, for example... Figure 4 As shown, the TRACE algorithm can automatically identify high-risk substructures in molecular structures based on ADME processes (i.e., absorption, distribution, metabolism, and excretion). These substructures are often closely related to DDI (discharge-discharge-inducing) mechanisms, such as transporter binding, enzyme metabolic sensitivity, and protein binding capacity. It is worth emphasizing that although traditional Chinese medicine (TCM) components are diverse and structurally complex, many active TCM components also contain pharmacokineticly sensitive groups or skeletons similar to those in chemical drugs, such as carboxylic acids, phenolic hydroxyl groups, aromatic amines, and heterocyclic structures. These structures can also trigger competition or inhibition in ADME processes, leading to the risk of interactions between TCM and other drugs.

[0074] Therefore, TRACE can be used not only for screening high-risk substructures and predicting DDI in chemical drugs, but also effectively extended to the field of traditional Chinese medicine components. By performing structural analysis on complex TCM component libraries and combining it with the TRACE algorithm for accurate identification of high-risk ADME substructures, innovative tools and scientific support can be provided for evaluating the safety of TCM compatibility and predicting the risks of combined use of TCM and Western medicine.

[0075] Case Study 2: To validate the applicability of TRACE to traditional Chinese medicine (TCM) data, a case study was conducted on tanshinone IIA (a representative bioactive TCM molecule extracted from the traditional medicinal plant Salvia miltiorrhiza). By evaluating TRACE's performance in predicting interactions between tanshinone IIA and several clinically relevant drugs, and using molecular docking to interpret the results, the aim was to explore whether TRACE can be used to predict interactions involving structurally novel compounds and provide mechanistic insights at the molecular level. Published literature was systematically reviewed, and a dataset was constructed including seven drugs known to interact with tanshinone IIA: warfarin, dabigatran, rivaroxaban, apixaban, edoxaban, betrixaban, and losartan. For each pair of tanshinone IIA and drugs, a pre-trained TRACE model was used to predict the presence of an interaction. Notably, TRACE achieved 100% prediction accuracy on this dataset.

[0076] To further elucidate the molecular basis of these interactions, molecular docking analyses were performed. First, detailed docking studies were conducted using tanshinone IIA and warfarin as clinically relevant examples. Both compounds docked with human serum albumin (HSA), a major drug-binding protein in the blood and a common mediator of drug-drug interactions. Figure 5 As shown, Figure 5 In the figure, 'a' represents an overview of the prediction of tanshinone IIA drug interactions based on TRACE. Figure 5 In the figure, b represents the docking simulation of tanshinone IIA and warfarin on human serum albumin (HSA, PDB ID: 1HA2). Figure 5 In the diagram, 'c' represents the docking simulation of tanshinone IIA with P-glycoprotein (P-gp), and 'd' represents the docking simulation of tanshinone IIA with midazolam on CYP3A4. Yellow represents tanshinone IIA, and pink represents midazolam. The docking results show that both tanshinone IIA and warfarin bind to the same pocket on the HSA, with comparable binding affinities. Further structural analysis revealed that both molecules form hydrogen bonds with the same amino acid residue, ARG257, within the binding pocket. This suggests a possible competitive binding relationship between the two compounds. These findings provide a structural explanation for the clinically observed interaction between tanshinone IIA and warfarin, indicating that concomitant administration of these two drugs may lead to competitive substitution and pharmacokinetic alterations.

[0077] For direct oral anticoagulants (DOACs), including dabigatran, rivaroxaban, apixaban, edoxaban, and betrixaban, the potential mechanism was further investigated by docking tanshinone IIA with P glycoprotein (Pgp). The results showed that tanshinone IIA can occupy the substrate-binding pocket of Pgp and form a hydrogen bond with Y949, potentially inhibiting Pgp activity, thereby increasing the plasma concentration of DOACs and enhancing drug-drug interactions.

[0078] For losartan, molecular docking was performed between tanshinone IIA and the major metabolic enzyme CYP3A4. The results showed that tanshinone IIA and midazolam occupy the same binding pocket, suggesting that tanshinone IIA may have similar effects to midazolam. It can act as a substrate for CYP3A4, thereby slowing the metabolism of losartan and increasing the risk of drug interactions and adverse reactions.

[0079] The results of this case study fully demonstrate the broad applicability and superior performance of the TRACE model on complex molecules in traditional Chinese medicine (TCM). TRACE can not only accurately capture the interaction risks of TCM molecules such as tanshinone IIA, but also, by combining techniques such as molecular docking, provide a reasonable explanation of molecular interaction results at the molecular mechanism level. This result indicates that the TRACE algorithm has good generalization ability and can be applied to TCM components with high structural diversity, providing a powerful intelligent tool and scientific basis for the safety evaluation of TCM compatibility and the risk prediction of combined TCM and Western medicine use.

[0080] Case Study 3: To investigate the sources of TRACE's performance improvement, a series of ablation studies were conducted on three datasets to assess the contributions of its key components. Specifically, embodiments of this application focus on two modules: ElementKG (Chemical Element Knowledge Graph Embedding) and TRACE (Chemical Element Knowledge Graph Embedding).

[0081] First, the TRACE variant without ElementKG embeddings (denoted as w / oKG) was evaluated on all three datasets. As shown in Table 2, removing ElementKG leads to a decrease in performance on all metrics, and the full TRACE model improves ACC by 1.46% on the small-scale ZhangDDI dataset compared to the variant. These results highlight the effectiveness of incorporating chemical element knowledge into the model.

[0082] Secondly, to verify the role of the graph transformer, it was replaced with GCN and GAT architectures, creating variants with / o GCN and with / o GAT. Evaluation results on all three datasets show that TRACE using the graph transformer consistently outperforms other alternatives, especially on the DeepDDI dataset, where AUROC and ACC are improved by more than 7.9% and 12%, respectively. This clearly demonstrates the significant contribution of the molecular feature extraction module. In conclusion, each component of the TRACE model makes a significant contribution to the performance of molecular interaction prediction, further validating the rationality and scientific validity of the model design.

[0083] Table 2 Ablation experiments of TRACE

[0084] In another exemplary embodiment, prior art is used as a comparative example to illustrate how the solution of this application solves the problems of the prior art.

[0085] Regarding data: To adapt to various application scenarios, TRACE was tested on four commonly used datasets: DrugBank, ZhangDDI, ChChMiner, and DeepDDI. The data for each dataset is as follows: DrugBank contains 2456 drug molecules with 191,808 interaction records; ZhangDDI contains 543 drug molecules with 95,245 interaction records; ChChMiner contains 997 drug molecules with 83,918 interaction records; and DeepDDI contains 1704 drug molecules with 265,771 interaction records.

[0086] Evaluation: To ensure the scientific rigor and accuracy of the research, two experimental tasks were set up: binary classification and multi-class classification. The binary classification task aimed to confirm the existence of interactions; the multi-class classification task aimed to determine the type of intermolecular interactions. In the more common binary classification training, this embodiment of the application set up two data partitions: transduction (in-distribution) and transduction (out-of-distribution). In the transduction setting, each dataset was randomly divided into a training set, a validation set, and a test set in a 6:2:2 ratio to ensure that drugs appearing in the test set also appeared in the training set. For each dataset, the experiment was repeated five times using different random seeds, and the mean and standard deviation of the experimental results were reported. The best test performance was defined as the model achieving its highest performance on the validation set. In the inductive setting, the model's generalization ability to novel drugs not included in the training set was evaluated. In each fold, the model was trained only on drug pairs where both drugs belonged to the training set. Then, it was evaluated on drug pairs where at least one or both drugs came from the test set, thus measuring the model's generalization ability to unseen drugs. This process was repeated five times, with each subset serving as the test set once. Final performance is reported as the average and standard deviation of five folds.

[0087] Eight baseline models were selected for binary classification and five for multi-class classification tasks, respectively, for comparison. For binary classification, methods such as GCN, GAT, MPNN, MIRACLE, DSN-DDI, SSI-DDI, CMRL, and CGIB were selected for comparison. For multi-class classification, DeepDDI, DeepWalk, KGDDI, KGNN, and MUFFINE were selected as comparison models in this embodiment. Finally, TRACE was benchmarked against other molecular interaction prediction models to evaluate its performance in different scenarios.

[0088] Comparative Case 1: Table 3. Performance test of TRACE in a binary classification task transduction environment.

[0089] To evaluate the model's performance in binary classification tasks, specifically predicting the existence of intermolecular interactions, this application selected three datasets (ZhangDDI, ChChMiner, and DeepDDI) for experiments. First, TRACE's performance was evaluated in a widely used transduction setting, where drugs in the test set also appear in the training set. AUROC and ACC were used for evaluation; higher values ​​indicate better performance. The results (Table 3) show that TRACE outperforms other models in both AUROC and ACC across different datasets. Although state-of-the-art algorithms have achieved high accuracy in molecular interaction prediction, TRACE consistently outperforms the second-ranked method in key metrics such as ACC and AUROC across all three datasets, especially on the large-scale DeepDDI dataset, where TRACE improved ACC by 1.09%. Furthermore, TRACE's AUC values ​​exceeded 95% on all three datasets, reaching 98.62% on the ChChMiner dataset and 99.69% on the DeepDDI dataset. These results demonstrate that TRACE provides near-perfect molecular interaction predictions for existing drugs in a transduction setting.

[0090] Comparative Case 2: Table 4. Performance test of TRACE in a binary classification task under inductive environment.

[0091] To further evaluate TRACE's generalization ability to unknown drugs, experiments were conducted in a more challenging inductive setting. Specifically, TRACE was tested on the largest dataset, DeepDDI, in two scenarios: S1, where each sample in the test set contains one unknown drug and one known drug (simulating the DDI prediction when a new drug combines with an existing drug); and S2, where both drugs in the test samples are unknown (simulating the DDI prediction when two new drugs combine). The results shown in Table 4 demonstrate that TRACE again outperforms all baseline models across all metrics, proving its strong generalization ability. Comparing Tables 3 and 4 reveals that performance in the inductive setting is significantly lower than in the regressive setting, reflecting the greater difficulty in predicting DDIs for new drugs. Notably, in the more practical S1 scenario, TRACE's ACC and AUROC exceed those of the second-ranked CGIB by 2.62% and 1.81%, respectively. Despite the increased challenge in the S2 scenario, TRACE still achieves performance comparable to the current state-of-the-art methods. To further evaluate TRACE comprehensively, this application also followed previous research and performed DDI type prediction on the DrugBank dataset. The model in this application consistently outperformed all baseline methods in this multi-class classification task; detailed results are available in the supplementary information. In summary, TRACE outperformed other models in all benchmark evaluations, highlighting the effectiveness of incorporating chemical element information into molecular interaction prediction.

[0092] Comparative Case 3: Table 5. TRACE performance tests under multi-classification tasks.

[0093] For the multi-class tasks, the most representative DrugBank dataset was selected. Table 5 lists the experimental results. TRACE performed excellently on all evaluation metrics, with a significant improvement in accuracy of 2.26%. Specifically, the improved precision reduced false positive predictions, while the improved recall indicates that the model more effectively identified true DDI pairs, thus reducing false negatives. The improved F1 score reflects a balanced improvement in accuracy and comprehensiveness, demonstrating TRACE's competitive ability to better distinguish drug interaction types.

[0094] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 6As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a drug interaction prediction method for the ADME process.

[0095] Those skilled in the art will understand that Figure 6 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0096] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0097] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0098] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0099] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0100] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0101] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting drug interactions during the ADME process, characterized in that, The drug interaction prediction method for ADME process employs a drug interaction prediction model, which is obtained by training the TRACE model. The drug interaction prediction model includes a molecular map generation module, a molecular feature extraction module, and a prediction module. The drug interaction prediction method for ADME process includes the following steps: Obtain chemical species structure data for the first and second drugs; Based on the chemical species structure data and chemical element knowledge graph of the first drug, an enhanced molecular graph of the first drug is constructed using the molecular graph generation module; Based on the chemical species structure data of the second drug and the chemical element knowledge graph, an enhanced molecular graph of the second drug is constructed using the molecular graph generation module; The enhanced molecular map of the first drug and the enhanced molecular map of the second drug are processed using the molecular feature extraction module to obtain the feature representations of the first drug and the second drug. Based on the feature representations of the first drug and the second drug, the prediction module is used to predict the probability of interaction between the first drug and the second drug.

2. The method for predicting drug interactions in the ADME process according to claim 1, characterized in that, The molecular graph generation module includes an RDKit unit, a DGL unit, and an embedding unit. Based on the chemical species structure data and chemical element knowledge graph of the first drug, the molecular graph generation module constructs an enhanced molecular graph of the first drug, specifically including: The RDKit unit is used to perform feature calculations on the chemical species structure data of the first drug to obtain the molecular characteristics of the first drug; Based on the molecular characteristics of the first drug, a molecular map of the first drug is generated using the DGL unit; Based on the chemical element knowledge graph, the embedding unit is used to embed the chemical element features contained in the first drug into the molecular graph of the first drug to obtain an enhanced molecular graph of the first drug.

3. The method for predicting drug interactions during the ADME process according to claim 1, characterized in that, The molecular feature extraction module uses a graph converter to dynamically aggregate local and global information in the enhanced molecular graph of the target drug using a self-attention mechanism to obtain a feature representation of the target drug; the target drug is a first drug or a second drug.

4. The method for predicting drug interactions in the ADME process according to claim 3, characterized in that, By dynamically aggregating local and global information from the enhanced molecular map of the target drug using a self-attention mechanism, a feature representation of the target drug is obtained, specifically including: The initial node hiding representation and the initial edge hiding representation are obtained by linearly projecting the node features and edge features in the enhanced molecular graph of the target drug. The initial node hidden representation and the initial edge hidden representation are updated in multiple layers using a self-attention mechanism to obtain the final updated node hidden representation and edge hidden representation, which together form the feature representation of the target drug.

5. The method for predicting drug interactions during the ADME process according to claim 4, characterized in that, A self-attention mechanism is used to perform multi-layer feature updates on the initial node hidden representation and the initial edge hidden representation to obtain the final updated node hidden representation and edge hidden representation, specifically including: Using batch normalization layer for the first The updated node and edge hidden representations are normalized, and the hidden representations of queries, keys, values, and edges are calculated in the multi-head self-attention mechanism. Calculate the attention score based on the hidden representations of queries, keys, and edges; Based on the value and attention score, the first Update the node hiding representation and edge hiding representation after the layer update to obtain the th layer. Layer node intermediate representation and edge intermediate representation; The first The intermediate representations of nodes and edges in the first layer are passed to the feedforward network for residual connections and normalization to obtain the intermediate representations of the second layer. The updated node hiding representation and edge hiding representation.

6. The method for predicting drug interactions during the ADME process according to claim 5, characterized in that, Using batch normalization layer for the first The updated node and edge hidden representations are normalized, and in the multi-head self-following mechanism, the formulas for calculating the hidden representations of queries, keys, values, and edges are as follows: ； ； ； ； in, k This represents the k-th attention head in the multi-head self-attention mechanism. Indicates the first Layer self-attention mechanism, and Both represent nodes. , , , These are the hidden representations of queries, keys, values, and edges, respectively. , , and These are the projection matrices for the query, key, value, and edge, respectively. and The first The updated node hiding representation and edge hiding representation after layer update. Indicates normalization; The formula for calculating attention score is: ； in, This represents the attention score. This represents the dimension of the k-th attention head. for function; Based on the value and attention score, the first Update the node hiding representation and edge hiding representation after the layer update to obtain the th layer. The formulas for the intermediate representation of nodes and the intermediate representation of edges in a layer are as follows: ； ； in, and They represent the first The intermediate representation of nodes and the intermediate representation of edges in a layer. Output the projection matrix for the node. This is a random deactivation function used to prevent overfitting. Output the projection matrix for the edge. Indicates a connection. This indicates that attention should be paid to the number of heads. Represents a node i The number of adjacent nodes; The first The intermediate representations of nodes and edges in the first layer are passed to the feedforward network for residual connections and normalization to obtain the intermediate representations of the second layer. The formulas for the updated node hiding representation and edge hiding representation are as follows: ； ； in, and The first The updated node hiding representation and edge hiding representation after layer update. and For the first The edge-forward network weight matrix of the layer, where, Used for dimensionality increase or feature transformation Used to map back edge representation space For activation function, and For the first The node feedforward network weight matrix of the layer, where Used for dimensionality increase or feature transformation Used to map back to the node representation space.

7. The method for predicting drug interactions during the ADME process according to claim 1, characterized in that, The prediction module is used to project the feature representations of the first drug and the feature representations of the second drug onto a multilayer perceptron to calculate the interaction probability, as shown in the formula: ； in, Represents the probability of interaction. This is a characteristic representation of the first drug. This is a characteristic representation of the second drug. This represents a multilayer perceptron.

8. The method for predicting drug interactions during the ADME process according to claim 1, characterized in that, The loss function used when training the TRACE model is the cross-entropy loss function.

9. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the drug interaction prediction method for the ADME process according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the drug interaction prediction method for the ADME process as described in any one of claims 1-8.

Images