Enhanced drug-drug interaction prediction method based on multi-modal drug characteristics

Abstract

The invention provides an enhanced drug-drug interaction prediction method based on multi-modal drug characteristics. The method comprises the following steps: step 1, carrying out data processing on two medicines to obtain characteristic data of the two medicines, wherein the characteristic data comprises a three-dimensional molecular map, a molecular fingerprint and a one-dimensional SMIELS sequence; 2, based on the feature data of the two drugs, feature learning is conducted through a heterogeneous feature extraction network, and respective multi-modal feature embedding expressions of the two drugs are obtained; 3, performing feature fusion on the multi-modal feature embedding expressions of the two drugs to obtain respective fused features of the two drugs; and 4, performing fusion interaction on the fused features of the two drugs and the interaction category features, and predicting a fusion interaction result to obtain a prediction result. The method provided by the invention can integrate the multi-modal information of the drug molecules, does not depend on multivariate information and knowledge maps, and has better performance and wider applicability.

Description

Technical Field

[0001] The present invention relates to the interdisciplinary field of bioinformatics, computational biology and artificial intelligence medicine, and in particular to an enhanced drug-drug interaction prediction method based on multimodal drug features. Background Art

[0002] Drug-drug interaction refers to the effect of two or more drugs on each other when they act together in an organism. Drugs may enhance or inhibit each other's efficacy or induce adverse reactions through interaction. The intensity and effect of drug-drug interactions depend on their chemical structure, metabolic pathways, pharmacokinetic properties, etc. The structure of drug molecules and their metabolic pathways in the body determine the interaction pattern between them. High interactions between drugs may lead to significant changes in efficacy or increase the risk of side effects, so accurate prediction of such interactions is crucial for the safety and efficacy evaluation of drugs.

[0003] In the prediction of drug-drug interaction (DDI) side effects, the existing technologies are mainly divided into two categories: machine learning-based and deep learning-based.

[0004] Machine learning-based models usually rely on feature engineering and traditional supervised learning algorithms to achieve their goals. Common features include structural information, biological activity, and metabolic pathways of drugs. Classifiers or regression models are constructed to predict the possible types of interactions between drugs and their possible side effects. Although these methods are highly interpretable in specific fields and can be analyzed based on manually selected features, they have limitations when dealing with complex nonlinear relationships and large-scale data, and require more domain knowledge and feature engineering support. Therefore, the importance of deep learning methods has become increasingly prominent.

[0005] Deep learning-based methods have shown great potential in predicting DDI side effects. Most deep learning methods predict side effects between drugs by modeling drugs as two-dimensional molecular graphs and then using graph neural networks for feature learning. Another type of method builds a more complex deep learning model by introducing multiple types of information, including omics information such as enzymes, genes, and proteins. Another type of method models the problem as an edge prediction problem based on knowledge graphs. However, these methods all have certain shortcomings. First, the two-dimensional molecular graph-based methods only use the two-dimensional molecular graph information and ignore the more detailed three-dimensional molecular structure information, which is crucial for learning the conformational structure of drugs. Second, current research has failed to effectively integrate the one-dimensional, two-dimensional, and three-dimensional multimodal information of drugs, such as SMILES (Simplified Molecular Input Line Entry System) sequences, molecular fingerprints, and three-dimensional molecular graphs. This information contains two-dimensional information and atomic distance information in three-dimensional space. It is necessary to integrate multimodal information. Although methods based on omics information can utilize multivariate information, this information is difficult to obtain, the applicability of the method is low, and it is difficult to build a powerful network. Finally, the performance of knowledge graph-based methods depends on the quality of the constructed knowledge graph. It is difficult to obtain reliable embedding representations for unknown drugs, which further limits the applicability of the method. Summary of the invention

[0006] The present invention aims to provide a new DDI side effect prediction method based on deep learning, which can not only integrate the multimodal information of drug molecules and does not rely on difficult-to-obtain multivariate information and knowledge graphs, but also learn the interaction patterns from drug pairs through deep learning methods, thereby having better performance and wider applicability.

[0007] The present invention provides an enhanced drug-drug interaction prediction method based on multimodal drug features, comprising:

[0008] Step 1: performing data processing on the two drugs to obtain characteristic data of the two drugs, wherein the characteristic data includes a three-dimensional molecular graph, a molecular fingerprint and a one-dimensional SMIELS sequence;

[0009] Step 2: Based on the feature data of the two drugs, a heterogeneous feature extraction network is used to perform feature learning to obtain a multimodal feature embedding representation of each of the two drugs;

[0010] Step 3: Perform feature fusion on the multimodal feature embedding representations of the two drugs to obtain the fused features of the two drugs;

[0011] Step 4: Fuse and interact the fused features of the two drugs with the interaction category features, predict the fused interaction results, and obtain the predicted results.

[0012] Existing DDI prediction methods mainly rely on single modal information of drugs, such as two-dimensional molecular graphs or SMILES sequences. However, these single modal information cannot fully capture the structural and functional characteristics of drugs, resulting in limited accuracy of the prediction model when dealing with complex drug interactions. In order to solve this problem, the present invention designs a new multimodal feature extraction mechanism by integrating multiple modal information of drugs, including molecular fingerprints, one-dimensional SMILES sequences, and three-dimensional molecular structures. Using deep learning technology, the present invention can extract more comprehensive and accurate feature representations from these multimodal data, thereby significantly improving the performance and reliability of the DDI prediction model.

[0013] Furthermore, in step 1, data processing is performed on the two drugs to obtain a three-dimensional molecular map, specifically including:

[0014] A two-dimensional molecular graph of the drug is generated using the SMILES string of the drug, and self-loop information is added to the two-dimensional molecular graph of the drug; and the spatial distances between adjacent atoms of the drug molecules in the three-dimensional space are extracted, and the spatial distances are added to the corresponding node features in the two-dimensional molecular graph of the drug. The molecular graph generated at this time is the three-dimensional molecular graph of the drug.

[0015] Furthermore, in step 1, the molecular fingerprint of the drug is an ECFP molecular fingerprint.

[0016] Furthermore, in step 1, data processing is performed on the two drugs to obtain a one-dimensional SMIELS sequence of the drugs, specifically including:

[0017] According to the SMIELS string of the drug, the ID of each character is searched in the vocabulary to generate an ID list of the drug, and a mask list of the drug is generated according to the ID list of the drug, and the ID list and the mask list of the drug are used together as a one-dimensional SMIELS sequence of the drug; wherein the vocabulary refers to a pre-constructed table with a single character as a token, each token being assigned a unique ID.

[0018] Furthermore, step 2 specifically includes:

[0019] Based on the three-dimensional molecular graph of the drug, a graph neural network is used for feature learning to obtain the three-dimensional structural feature embedding representation of the drug;

[0020] Based on the molecular fingerprint of the drug, a feedforward neural network is used for feature learning to obtain the molecular fingerprint feature embedding representation of the drug;

[0021] Based on the one-dimensional SMIELS sequence of drugs, the Transformer encoder is used for feature learning to obtain the one-dimensional sequence feature embedding representation of the drugs.

[0022] Furthermore, the graph neural network includes a plurality of graph convolutional layers that are sequentially residually connected and a global pooling layer located after the last graph convolutional layer.

[0023] When using traditional graph neural networks to learn the features of drug molecule graphs, they can often only capture graph information of a specific scale and ignore subgraph information of different scales. This limitation prevents the model from fully utilizing the detailed information in the drug molecules, affecting the prediction effect. In order to overcome this problem, the present invention adds residual connections between the graph convolutional layers of the graph neural network, so that the model can learn multi-scale molecular graph level information. This improvement enables the model to not only capture the global graph structure, but also pay attention to more detailed subgraph information, thereby more accurately predicting drug interactions and the side effects they may cause.

[0024] Furthermore, in step 3, the attention mechanism is used to fuse the multimodal feature embedding representations of the two drugs to obtain the fused features of the two drugs.

[0025] Existing DDI prediction methods often use simple splicing or weighted averaging when fusing different modal features. This simple fusion mechanism cannot fully utilize the complementary information between different modal features, limiting the performance of the model. In order to solve this problem, the present invention introduces an attention mechanism to fuse molecular fingerprint features, one-dimensional SMILES sequence features, and three-dimensional graph structure features. In this way, the model can adaptively weigh and integrate the features of each modality, and finally obtain a molecular-level feature representation. This fusion mechanism not only improves the prediction accuracy of the model, but also enhances its generalization ability in different application scenarios.

[0026] Furthermore, step 4 specifically includes:

[0027] Convert interaction category features into embedding representations of given dimensions;

[0028] A multi-layer feed-forward neural network is used to project the embedded representation of the interaction category features into the same dimensional space as the fused features of the drugs;

[0029] The fused features of the two drugs are concatenated to form a drug-drug feature vector;

[0030] Performing an element-wise multiplication operation on the embedding representation of the projected interaction category features and the drug-drug feature vector to generate an embedding representation of the drug-drug interaction;

[0031] The drug-drug interaction embedding representation is input into a multi-layer feed-forward neural network to obtain a prediction result.

[0032] Beneficial effects of the present invention:

[0033] (1) The present invention fully utilizes the multimodal characteristics of drug molecules by integrating multiple modal information such as three-dimensional molecular structure, molecular fingerprint and one-dimensional SMILES sequence. Compared with the traditional method of single modal information, the comprehensive feature extraction and fusion method of the present invention can capture more comprehensive and accurate characteristics of drug molecules and comprehensively describe the structural and functional information of drugs, thereby significantly improving the accuracy and comprehensiveness of the prediction of drug-drug interaction side effects.

[0034] (2) The present invention introduces residual connections between the multi-graph convolutional layers of the graph neural network, so that the model can learn multi-scale molecular graph-level information. This multi-scale information fusion method enables the model to not only capture the global structure of drug molecules, but also pay attention to more detailed sub-graph information, thereby more accurately predicting drug interactions and their possible side effects. This improvement significantly improves the performance of the model when dealing with complex drug interactions.

[0035] (3) The present invention uses the attention mechanism to adaptively fuse molecular fingerprint features, one-dimensional SMILES sequence features, and three-dimensional graph structure features. This mechanism enables the model to adaptively weigh and integrate the features of each modality to generate a comprehensive molecular-level feature representation, thereby improving the generalization ability of the prediction model and its reliability in practical applications. In different application scenarios, the model of the present invention exhibits higher reliability and stability, providing powerful tool support for drug safety assessment and new drug development. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 One of the flow diagrams of an enhanced drug-drug interaction prediction method based on multimodal drug features provided by an embodiment of the present invention;

[0037] Figure 2 A schematic diagram of extracting characteristic data of each individual drug molecule provided by an embodiment of the present invention;

[0038] Figure 3 The second flowchart of an enhanced drug-drug interaction prediction method based on multimodal drug features provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] Example 1

[0041] Combination Figures 1 to 3 As shown, an embodiment of the present invention provides an enhanced drug-drug interaction prediction method based on multimodal drug features, comprising the following steps:

[0042] S101, data processing: Data processing is performed on the two drugs to be predicted to obtain characteristic data of the two drugs, wherein the characteristic data includes a three-dimensional molecular graph, a molecular fingerprint and a one-dimensional SMIELS sequence, such as Figure 2 shown.

[0043] Specifically, the process of obtaining the three-dimensional molecular graph of the drug includes: using the SMILES string of the drug to generate a two-dimensional molecular graph of the drug, adding self-loop information to the two-dimensional molecular graph of the drug; and extracting the spatial distance between adjacent atoms of the drug molecules in the three-dimensional space, adding the spatial distance to the corresponding node features in the two-dimensional molecular graph of the drug. The molecular graph generated at this time is the three-dimensional molecular graph of the drug.

[0044] It can be understood that the SMILES string is a text format for describing molecular structures, which can be easily converted into a two-dimensional graph representation of the molecule. By adding self-loop information to the two-dimensional molecular graph, the graph structure information can be increased to construct a more complete adjacency matrix. Self-loop information refers to adding self-connected edges to nodes (atoms), which helps capture the characteristics of the nodes themselves. The Rdkit tool can be used to extract the spatial distances between adjacent atoms. The generated three-dimensional molecular graph not only contains the two-dimensional structural information, but also integrates the three-dimensional spatial characteristics, making the graph structure more three-dimensional and comprehensive.

[0045] The molecular fingerprint of the drug is the ECFP molecular fingerprint. In this embodiment, the ECFP (Extended Connectivity Fingerprints) molecular fingerprint information of the drug of the Rdkit tool can be used. ECFP is a molecular fingerprint based on the molecular topology and is used to represent the characteristics of a molecule. For example, when calculating ECFP, the radius of the fingerprint is set to 2, which means that the atomic information of each atom and its two adjacent bonds in the molecular structure is taken into account. The fingerprint length is set to 1024 to obtain a fixed-length molecular fingerprint representation. This high-dimensional molecular fingerprint can capture detailed information about the molecular structure and contribute to more accurate molecular characterization in the feature learning process.

[0046] The process of obtaining the one-dimensional SMIELS sequence of a drug specifically includes: according to the SMIELS string of the drug, searching the ID of each character in the vocabulary to generate an ID list of the drug, and generating a mask list of the drug according to the ID list of the drug, and using the ID list and the mask list of the drug together as the one-dimensional SMIELS sequence of the drug; wherein the vocabulary refers to a pre-constructed table with a single character as a token, and each token is assigned a unique ID, as shown in Table 1.

[0047] Table 1 Vocabulary of SMILES sequences of drugs

[0048]

[0049] It can be understood that since the length of the SMILES string of different drugs is different, in order to unify the input length, a maximum length (max-length) needs to be set, which is set to 100 in this embodiment. If the length of the ID list is less than max-length, a padding ID (usually 0) is added to the end of the ID list to make up the length. In this way, the length of all ID lists will be 100. The mask list of the drug represents the valid bits and padding bits in the ID list; the valid bits are represented by 1 and the padding bits are represented by 0. For example, for the ID list [1,1,2], its mask list is [1,1,1,0,0,...,0] (the number of 0s added later makes the total length 100).

[0050] S102. Feature learning: Based on the feature data of the two drugs, a heterogeneous feature extraction network is used to perform feature learning to obtain a multimodal feature embedding representation of the two drugs.

[0051] It can be understood that for different feature data of each drug, feature extraction networks with different structures are used for feature learning, thereby obtaining embedded representations of different modal features. The feature extraction networks of the two drugs share weights.

[0052] S103. Fusion of multimodal features: performing feature fusion on the multimodal feature embedding representations of the two drugs to obtain the fused features of the two drugs.

[0053] Specifically, the previous step can be used to obtain feature representations of three different modalities, including the three-dimensional information embedding representation of the drug, the molecular fingerprint information embedding representation, and the one-dimensional sequence feature embedding representation. The three-dimensional information embedding representation captures the spatial structural characteristics of the drug molecules, the molecular fingerprint information embedding representation reflects the chemical and structural characteristics of the drug, and the one-dimensional sequence feature embedding representation provides contextual information of the drug molecular sequence.

[0054] In order to effectively fuse the above three features, this embodiment adopts the attention mechanism (also called Attention mechanism). This mechanism performs linear transformation on the features of each modality, calculates its weight coefficient, and weights and sums these features. Specifically, the three-dimensional information embedding representation, the molecular fingerprint information embedding representation, and the one-dimensional sequence feature embedding representation are mapped to the same space through independent linear transformations. Then, these mapped features are spliced ​​together, the weight coefficients are calculated by the softmax function, and finally a fused feature representation is obtained. This fused feature representation integrates the information of the three modalities to form a more comprehensive feature vector, which provides a rich feature basis for subsequent tasks (such as side effect analysis of drug interactions). This fusion method effectively improves the model's ability to understand drugs and enhances the accuracy and reliability of predictions.

[0055] S104, fusing and interacting the fused features of the two drugs with the interaction category features, predicting the fused interaction results, and obtaining a prediction result.

[0056] It should be noted that the interaction category features need to be collected in advance. In this embodiment, drug-drug interaction side effect information is extracted from three public data sets: DrugBank, Twosides and Dataset1. Among them, the DrugBank data set comes from the paper "Ryu JY, Kim HU, Lee SY. Deep learning improvesprediction of drug-drug and drug-food interactions[J]. 2018". The Twosides data set comes from the paper "Marinka Zitnik, Monica Agrawal, Jure Leskovec, Modeling polypharmacyside effects with graph convolutional networks[J]. 2018". The Dataset1 data set comes from the paper "Yifan Deng, Xinran Xu, Yang Qiu, Jingbo Xia, Wen Zhang, Shichao Liu, Amultimodal deep learning framework for predicting drug-drug interactionevents[J]. 2020". The above three public data sets are used as training and test data for the model. These data sets contain rich drug-drug interaction information and corresponding side effect categories, providing comprehensive data support for model training.

[0057] Specifically, the fusion interaction process is as follows:

[0058] Side effect category feature embedding: First, the category ID of the side effect is converted into an embedded representation of fixed dimension. Each side effect category ID is mapped into a fixed-size vector space through an embedding layer. Next, a multi-layer feedforward neural network is used to project the embedded representation of the side effect category into the same dimensional space as the multimodal features of the drug. This process involves transforming the embedding vector through one or more fully connected layers to match the dimension of the multimodal features of the drug. In this way, it is ensured that the side effect category features are fused with the multimodal features of the drug in the same feature space.

[0059] Feature fusion and interaction representation generation: The multimodal feature representations of the two drugs are concatenated to form a comprehensive drug-drug feature vector. Then, the embedding representation of the side effect category is element-wise multiplied with the concatenated drug feature vector. This multiplication operation combines the semantic information of the side effect category with the multimodal features of the drug to generate a fused drug-drug interaction embedding representation. This embedding representation combines the detailed features of the drug and the impact of the relevant side effect category, and can more accurately describe the potential interactions between drugs.

[0060] Predicting interactions: The generated drug-drug interaction embedding representation is input into a multi-layer feedforward neural network. Finally, after being processed by the feedforward neural network, the model outputs the prediction results of drug-drug interactions. These results provide detailed information on the interactions between drugs, including possible side effects and potential risks of drug combinations. These prediction results provide important reference data for drug development and clinical application, helping to identify and avoid potential drug interaction problems.

[0061] Example 2

[0062] Based on the above embodiment 1, Figure 2 As shown, this embodiment uses a graph neural network to perform feature learning based on the three-dimensional molecular graph of the drug to obtain an embedded representation of the three-dimensional structural features of the drug.

[0063] It can be understood that the initial node feature matrix and adjacency matrix can be obtained based on the three-dimensional molecular graph of the drug. In the node feature matrix, the feature of each node (i.e., atom) in the molecule is represented as a high-dimensional vector, and each node feature may include the physical and chemical properties and three-dimensional coordinates of the atom. The adjacency matrix represents the connection relationship between nodes in the molecular graph. The matrix describes which atoms are connected to each other and the strength or type of the connection.

[0064] In this embodiment, the graph neural network includes a plurality of graph convolution layers that are sequentially residually connected and a global pooling layer located after the last graph convolution layer. Correspondingly, the processing process of the graph neural network is as follows:

[0065] Graph convolution layer: The initial node feature matrix and adjacency matrix are input to the first graph convolution layer to generate updated node feature representations. In order to retain the information of the input features and help the model better learn complex graph structures, residual connections are introduced between the input and output of the first graph convolution layer. This means that the input features of the first graph convolution layer are not only used to generate the output of the first graph convolution layer, but are also directly passed to the subsequent graph convolution layers to help the model capture long-distance dependencies. In the subsequent graph convolution layers of the model, the residual connections continue to play a role. The output of the first graph convolution layer is passed to the third graph convolution layer, so that even after multiple layers of graph convolution, the model can retain feature information from earlier layers. This approach ensures that the model can capture different levels of graph information, from global structure to local details. The output of each layer is not only based on the features of the previous layer, but also retains information from earlier layers, thereby improving the ability to learn multi-scale information.

[0066] Global pooling layer: After multiple layers of graph convolution, a global pooling layer is used to convert node-level features into graph-level embedding representations. The global pooling operation aggregates the features of all nodes to generate a unified feature vector that fully reflects the overall information of the molecular graph.

[0067] After multi-layer graph convolution and global pooling, the obtained fixed-dimensional vector is embedded as the three-dimensional structural information of the drug molecule. This embedded representation combines the characteristics of all nodes in the molecular graph and their connection relationships, and can accurately capture the global structure and detailed local information of the drug. By introducing residual connections, the model of the present invention can not only capture the overall structure of the drug molecule, but also pay attention to the multi-scale molecular graph information at the detail level. This multi-scale information learning ability significantly improves the accuracy and reliability of the model in drug interaction and side effect prediction.

[0068] For example, if a three-layer graph convolutional neural network (one layer of GCN and two layers of GAT) is used to extract features from a three-dimensional molecular graph, the process is as follows: First, the drug molecules are feature extracted through three layers of graph convolutional layers, and each layer of graph convolutional layer is followed by a ReLU activation function to increase the nonlinear representation ability of the network. Subsequently, the node features of the drug molecules are aggregated and dimensionally compressed through a global maximum pooling layer. The network also introduces a residual connection mechanism to enhance the feature fusion capability. After the second and third layers of graph convolution, residual connections are implemented by calculating the weighted sum of features, making the transfer of features between layers smoother and facilitating the transfer of gradients. The aggregated features are further dimensionally transformed through two fully connected layers, and each fully connected layer is followed by a Dropout layer to prevent overfitting. The activation function of the entire network uses ReLU to ensure the nonlinear features of each layer output. The hyperparameters of the feature extraction module of the drug graph are shown in Table 2:

[0069] Table 2 Parameters of graph convolutional neural network

[0070]

[0071] Example 3

[0072] Based on the above embodiments, Figure 2 As shown, this embodiment uses a feedforward neural network to perform feature learning based on the molecular fingerprint of the drug to obtain an embedded representation of the molecular fingerprint feature of the drug.

[0073] It can be understood that in order to deeply understand the molecular characteristics of the drug, the molecular fingerprint information is input into the feedforward neural network (FNN) to generate an embedded representation of the molecular fingerprint information of the drug. The feedforward neural network contains an input layer and multiple hidden layers, each of which is composed of multiple neurons.

[0074] For example, the molecular fingerprint information with a length of 1024 dimensions is directly used as the feature of the input layer. Each input node corresponds to a dimension in the molecular fingerprint. The number of nodes in the input layer is 1024, and each feature value in the molecular fingerprint is received. Each hidden layer linearly transforms the output features of the previous layer through the weight matrix, and performs nonlinear transformation through the activation function (such as ReLU, Sigmoid, etc.). These hidden layers gradually extract and learn high-level abstract features of the input features, and enhance the network's ability to capture complex feature patterns. The vector output by the feedforward neural network is used as an embedded representation of the molecular fingerprint information of the drug. This embedded representation integrates all the important features of the molecular fingerprint, can effectively describe the chemical and structural information of the drug, and provide strong support for subsequent analysis and prediction.

[0075] For example, the input features can be processed by a sequence feature extraction module containing three layers of feedforward neural network (also called linear layer). Each linear layer is followed by a ReLU activation function to increase the nonlinear representation ability of the network. In addition, each linear layer is followed by a Dropout layer to prevent overfitting. The hyperparameters of the drug molecular fingerprint information feature learning module are shown in Table 3:

[0076] Table 3 Hyperparameters for learning molecular fingerprint information features of drugs

[0077]

[0078] Example 4

[0079] Based on the above embodiments, Figure 2 As shown, this embodiment uses a Transformer encoder to perform feature learning based on a one-dimensional SMIELS sequence of drugs to obtain a one-dimensional sequence feature embedding representation of the drugs.

[0080] It can be understood that the feature learning process of the Transformer encoder is as follows:

[0081] Position encoding: In order to make the model understand the information of each position in the sequence, the Transformer encoder adds position encoding to the embedding representation of the ID list. The position encoding is a vector with the same length as the input ID list, which is used to represent the relative or absolute position of each token in the sequence. This encoding method ensures that the model can capture the order relationship of tokens in the sequence.

[0082] Word embedding layer: The ID list is converted into a dense vector representation through the word embedding layer. The word embedding layer maps each ID into a high-dimensional space and obtains the embedding vector of each token. These vectors contain the semantic information of the token and serve as the input of the Transformer encoder layer.

[0083] Transformer encoder layer: It consists of multiple layers of self-attention mechanism and feedforward neural network. The self-attention mechanism generates a weighted representation of each token by calculating the relationship between each token in the input sequence (the sequence after adding position encoding). The encoder gradually extracts and fuses the contextual information in the sequence by performing multiple layers of processing on the input data. The processing of each layer includes two parts: the self-attention mechanism and the feedforward neural network. The first is to calculate the correlation between each token and other tokens and generate a weighted contextual representation. Through this mechanism, the model can pay attention to other tokens in the sequence that are related to the current token, thereby capturing the global information of the sequence. The second is to perform a nonlinear transformation on the output of the self-attention mechanism to further extract features.

[0084] Sequence feature embedding: After being processed by the Transformer encoder, an embedding representation containing drug sequence information is obtained. This embedding representation integrates the context information of each token in the drug molecule and converts it into a vector of fixed dimension. This vector not only retains the basic information of the drug sequence, but also incorporates the relationship and context information of each token in the sequence.

[0085] From the above content, we can see that by inputting the ID list and mask list into the Transformer encoder, the Transformer can make full use of the sequence information of the drug, so that the model can learn the complex patterns and contextual relationships in the sequence, thereby generating a one-dimensional sequence feature embedding representation of the drug.

[0086] The parameters of the Transformer encoder layer include model dimension, number of attention heads, dimension of feedforward neural network, etc. The design of the entire network structure ensures the full capture and feature extraction of SMILES sequence information. The hyperparameters of the one-dimensional sequence feature learning module of drugs are shown in Table 4:

[0087] Table 4 Hyperparameters for learning one-dimensional sequence features of drugs

[0088]

[0089] Through the above settings, the network can effectively extract and encode the features of SMILES sequences, ensuring the full expression and effective use of sequence information. The introduction of position encoding and self-attention mechanism enhances the network's ability to capture position information and long-range dependencies in the sequence.

[0090] Example 5

[0091] Based on the above embodiments, the attention mechanism in this embodiment includes three parallel attention mechanism modules. Correspondingly, the feature fusion process of the attention mechanism is as follows:

[0092] First, the three-dimensional information embedding representation, molecular fingerprint information embedding representation and one-dimensional sequence feature embedding representation of the drug are respectively input into three parallel attention mechanism modules. In these parallel modules, the three-dimensional information, molecular fingerprint information and one-dimensional sequence features are independently weighted to highlight the importance of different modal features. Subsequently, the output weights of all modalities are connected and normalized by the softmax function to obtain the final multimodal feature fusion representation. This design ensures the effectiveness and importance of different modal features in the fusion process, making the comprehensive feature representation of the drug more accurate and comprehensive.

[0093] In this embodiment, each attention mechanism module consists of two linear layers, the first linear layer is followed by a Tanh activation function, and the second linear layer outputs a weight value. The hyperparameters of this module are shown in Table 5:

[0094] Table 5 Hyperparameters of the multimodal feature fusion module of drugs

[0095]

[0096] Through the above settings, the network can effectively fuse the multimodal features of drugs, ensuring the effectiveness and importance of different features in the final comprehensive representation. The introduction of the attention mechanism enhances the network's ability to selectively focus on different modal features, thereby improving the accuracy and overall performance of feature fusion.

[0097] Example 6

[0098] On the basis of the above embodiments, based on the characteristics of the two drugs and the characteristics of the interaction categories, this embodiment uses a fusion interaction module to comprehensively represent the drug information, and introduces relationship type embedding to further enhance the feature representation capability.

[0099] First, the two drug embedding representations are connected together to form a preliminary drug pair feature representation. Next, an embedding layer is used to embed the relationship type into a vector representation of fixed dimension. Then, the drug pair feature representation is element-wise multiplied with the relationship type embedding representation to achieve the fusion of drug features and relationship type features. This design ensures the effectiveness and diversity of drug features in the fusion process, while introducing relationship type features to enhance the accuracy and comprehensiveness of feature representation. The hyperparameters of this module are shown in Table 6:

[0100] Table 6 Hyper parameters of fusion interaction module

[0101]

[0102] Through the above settings, the network can effectively fuse the multimodal features of drugs and combine the relationship type embedding to enhance the accuracy and comprehensiveness of feature representation. The embedding layer and element-by-element multiplication operation ensure the effectiveness of different modal features in the final comprehensive representation.

[0103] Example 7

[0104] On the basis of the above embodiments, this embodiment adopts a feedforward neural network as a prediction module, which performs feature processing and prediction through a series of linear layers, activation functions and Dropout layers.

[0105] First, the input drug-drug pair embedding representation is converted to 512 dimensions through the first linear layer. Then, the network uses the ReLU activation function after the first linear layer to increase the nonlinear representation ability of the network, and is followed by the Dropout layer to prevent overfitting. Next, the feature dimension is further reduced to 256 dimensions through the second linear layer, and the ReLU activation function and Dropout layer are also used for processing. Finally, the network compresses the feature dimension to 1 dimension through the third linear layer and outputs the final prediction result. The hyperparameters of this module are shown in Table 7.

[0106] Table 7 Hyperparameters of the prediction module

[0107]

[0108] Example 8

[0109] Based on the above embodiments, during model training, binary cross entropy is used as the loss function, and the Adam optimizer is used to update the model weights. The model training rounds are 500. The data set segmentation strategy of the model is that 80% of the data set is used as the training set and 20% of the data set is used as the test set. The hyperparameters of model training are shown in Table 8.

[0110] Table 8 Hyperparameters of the model

[0111]

[0112] The use of the method of the present invention to predict DDI side effects can improve the accuracy and practicality of drug safety assessment and new drug development.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An enhanced drug-drug interaction prediction method based on multimodal drug features, characterized in that: include: Step 1: performing data processing on the two drugs to obtain characteristic data of the two drugs, wherein the characteristic data includes a three-dimensional molecular graph, a molecular fingerprint and a one-dimensional SMIELS sequence; Step 2: Based on the feature data of the two drugs, a heterogeneous feature extraction network is used to perform feature learning to obtain a multimodal feature embedding representation of each of the two drugs; Step 3: Perform feature fusion on the multimodal feature embedding representations of the two drugs to obtain the fused features of the two drugs; Step 4: Fuse and interact the fused features of the two drugs with the interaction category features, predict the fused interaction results, and obtain the predicted results.

2. The enhanced drug-drug interaction prediction method based on multimodal drug features according to claim 1, characterized in that: In step 1, data processing is performed on the two drugs to obtain a three-dimensional molecular map, including: A two-dimensional molecular graph of the drug is generated using the SMILES string of the drug, and self-loop information is added to the two-dimensional molecular graph of the drug; and the spatial distances between adjacent atoms of the drug molecules in the three-dimensional space are extracted, and the spatial distances are added to the corresponding node features in the two-dimensional molecular graph of the drug. The molecular graph generated at this time is the three-dimensional molecular graph of the drug.

3. The enhanced drug-drug interaction prediction method based on multimodal drug features according to claim 1, characterized in that: In step 1, the molecular fingerprint of the drug is an ECFP molecular fingerprint.

4. The enhanced drug-drug interaction prediction method based on multimodal drug features according to claim 1, characterized in that: In step 1, data processing is performed on the two drugs to obtain the one-dimensional SMIELS sequences of the drugs, including: According to the SMIELS string of the drug, the ID of each character is searched in the vocabulary to generate an ID list of the drug, and a mask list of the drug is generated according to the ID list of the drug, and the ID list and the mask list of the drug are used together as a one-dimensional SMIELS sequence of the drug; wherein the vocabulary refers to a pre-constructed table with a single character as a token, each token being assigned a unique ID.

5. The enhanced drug-drug interaction prediction method based on multimodal drug features according to any one of claims 1 to 4, characterized in that: Step 2 specifically includes: Based on the three-dimensional molecular graph of the drug, a graph neural network is used for feature learning to obtain the three-dimensional structural feature embedding representation of the drug; Based on the molecular fingerprint of the drug, a feedforward neural network is used for feature learning to obtain the molecular fingerprint feature embedding representation of the drug; Based on the one-dimensional SMIELS sequence of drugs, the Transformer encoder is used for feature learning to obtain the one-dimensional sequence feature embedding representation of the drugs.

6. The enhanced drug-drug interaction prediction method based on multimodal drug features according to claim 5, characterized in that: The graph neural network includes a plurality of graph convolutional layers which are sequentially residually connected and a global pooling layer located after the last graph convolutional layer.

7. The enhanced drug-drug interaction prediction method based on multimodal drug features according to claim 1, characterized in that: In step 3, the attention mechanism is used to fuse the multimodal feature embedding representations of the two drugs to obtain the fused features of the two drugs.

8. The enhanced drug-drug interaction prediction method based on multimodal drug features according to claim 1, characterized in that: Step 4 specifically includes: Convert interaction category features into embedding representations of given dimensions; A multi-layer feed-forward neural network is used to project the embedded representation of the interaction category features into the same dimensional space as the fused features of the drugs; The fused features of the two drugs are concatenated to form a drug-drug feature vector; Performing an element-wise multiplication operation on the embedding representation of the projected interaction category features and the drug-drug feature vector to generate an embedding representation of the drug-drug interaction; The drug-drug interaction embedding representation is input into a multi-layer feed-forward neural network to obtain a prediction result.

Images