Drug interaction event prediction method based on multi-modal diagram diffusion static subgraph

Abstract

The invention discloses a drug-drug interaction prediction method fusing multi-modal drug information and a static subgraph. The method mainly comprises the following steps: extracting biochemical similarity characteristics of a drug chemical substructure, a target spot and an enzyme; obtaining pharmacological features of the medicine based on a natural language processing model; using graph diffusion convolution to obtain learned node representation from each static sparse sub-graph; embedding of multi-modal feature fusion different drug features; a multi-layer perceptron is used to predict the category of action. According to the drug interaction event prediction method based on the diffusion static subgraph of the multi-modal graph, DDI multi-category prediction is achieved, and compared with an existing algorithm, the method has the following advantages that richer features are provided in combination with multi-modal features such as pharmacology and the like; (2) dividing into a plurality of static sparse sub-graphs by using random edges, learning node representation from each sub-graph, and providing global structure features and local structure features of the DDI network; and (3) richer potential representation is obtained by using graph diffusion convolution.

Description

Technical Field

[0001] The present invention relates to the field of multimodal feature learning and deep learning, and is based on a drug-drug interaction prediction method that integrates multimodal drug information and static subgraphs to predict drug-drug interactions. Background Art

[0002] Multidrug combination therapy has become an effective strategy for managing complex or coexisting diseases. However, the simultaneous use of multiple drugs may lead to drug-drug interactions (DDIs), which may cause adverse reactions and affect patient safety. DDIs contribute significantly to adverse drug events (ADEs), with a considerable proportion of ADEs attributed to DDIs. Given the potential impact on patient health and medical costs, accurate identification and prediction of DDIs are critical for clinicians and pharmaceutical companies.

[0003] In the existing field of DDI prediction, a variety of computational methods have been proposed, ranging from natural language processing (NLP) technology to machine learning frameworks. These methods mainly focus on binary classification tasks to determine whether a DDI event occurs. However, these methods usually ignore the in-depth analysis of specific pharmacological effects. In recent years, predicting multi-type DDIs has become a research hotspot, aiming to more comprehensively analyze the pharmacological effects of drug interactions. For example, computational frameworks such as Deep-DDI use the name and structural information of drug pairs as input to generate 86 important DDI types, presented in the form of human-readable sentences. Lee et al. predicted the pharmacological effects of DDIs through autoencoders and deep feedforward networks trained with drug pair structural similarity, gene ontology similarity, and target gene similarity data. However, these methods have limitations in the following aspects: 1) The drug features are single and rarely pay attention to other multimodal features; 2) Secondly, most methods only use the DDI network from a holistic perspective and ignore the details inside the network, losing some features. 3) The deep network of traditional graph convolution is prone to over-smoothing problems, resulting in the homogeneity of node features. Summary of the invention

[0004] In view of the shortcomings in the field of predicting drug-drug interactions (DDI), the purpose of the present invention is to provide a multimodal feature extraction strategy and perform multimodal feature fusion to perform DDI multi-category prediction tasks.

[0005] The present invention provides a drug-drug interaction prediction method that integrates multimodal drug information and static subgraphs, and the method steps are as follows.

[0006] Step 1: Perform multi-feature fusion representation learning on drugs, convert the enzyme features, target features, and chemical substructures of drugs into binary (0 or 1) vector representations, and calculate the Jaccard similarity between drugs to obtain the enzyme feature similarity matrix, target feature similarity matrix, and chemical substructure similarity matrix of drugs respectively. The three similarity matrices of enzyme features, target features, and chemical substructures are learned and concatenated to obtain the drug feature similarity matrix. The pharmacological text features of drugs are encoded using a pre-trained natural language processing (NLP) model to obtain the pharmacological text feature encoding matrix of drugs. The drug feature similarity matrix and the pharmacological text feature encoding matrix are combined. The encoding matrices are merged, and the initial drug representation matrix is ​​obtained through a multi-head attention mechanism. The autoencoder is used to perform representation learning on the initial drug matrix. The hidden layer representation of the autoencoder is used as the biological feature representation of the drug to obtain the drug biological feature representation matrix. The random edge segmentation method is used to segment the DDI network into multiple static sparse subgraphs with non-overlapping edges. The graph encoder model is used to perform representation learning on the drug biological feature representation matrix of (1.3). The hidden layer representation is used as the structural feature representation of the drug to obtain the drug structural feature representation matrix. The drug biological feature representation matrix and the drug structural feature representation matrix are merged to obtain the drug representation matrix.

[0007] Step 2: Use the multimodal feature fusion module to learn the drugs and obtain the drug pair representation vector.

[0008] Step 3: Use a multi-layer perceptron to learn the drug pair representation vector and predict the type of drug interaction event. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a model diagram of the present invention.

[0010] Figure 2 The performance of the model of this method under various parameters on the dataset (Deng et al). Implementation

[0011] In order to further understand the content of the present invention, the prediction model and the training method thereof of the present invention are described in detail in combination with the accompanying drawings and specific implementation methods, and the specific technology is explained and further illustrated with the help of legends.

[0012] Figure 1 This is a model diagram of a drug interaction prediction method that integrates multimodal drug information and static subgraphs of the present invention. Figure 1As shown, the method includes learning multimodal drug features based on three feature extraction modules, fusion of learned feature embedding, and prediction of DDIs based on a predictor. The three feature extraction modules learn low-dimensional embedding vectors of different drug features, including a feature extraction module based on biosimilarity, a module for extracting drug pharmacological properties based on Bio-BERT, and a static sparse subgraph feature extraction module based on a drug interaction network. Then, the multimodal feature fusion module fuses the three aspects of drug feature embedding. Finally, the fused feature vector is fed into the predictor to predict the category of DDIs. The specific steps are as follows.

[0013] Step 1: Constructing drug feature matrix, a drug feature matrix (drug-enzyme, drug-target and drug-transporter) was constructed to represent these drug features. Each feature corresponds to a set of descriptors, so that the drug can be represented as a binary feature vector, whose value (1 or 0) indicates the presence or absence of the corresponding element. Calculate the drug feature similarity matrix, we use the similarity profile to represent the drug feature vector, which is generated by the similarity between drug A and other drugs (i.e., drug B) in the dataset, and calculate the Jaccard similarity of each drug feature to obtain the similarity matrix of the three features of each drug. Among them, A and B are the original bit vectors of the two drugs; |A∩B| is the number of elements in the intersection of A and B; |A∪B| is the number of elements in the union of A and B. Through the Jaccard similarity, we correspondingly obtain the drug feature similarity matrix. The pre-trained BIO-BERT model obtains the pharmacological characteristics of the drug. The pre-trained BIO-BERT model is used to encode the pharmacological text attributes of the drug, and the rich semantic information related to pharmacology is captured to obtain the pharmacological attribute feature matrix of the drug. We use the pre-trained Bio-BERT model to process the pharmacological text attributes of the drug and obtain vector features. The pre-training goal of Bio-BERT is to simultaneously minimize the masked language modeling (MLM) and next sentence prediction (NSP) losses. The MLM loss aims to predict the masked tokens in the input text, while the NSP loss aims to predict whether two input sentences are continuous. Formally, the training objective of Bio-BERT can be expressed as: Where LMLM is the masked language modeling loss and LNSP is the next sentence prediction loss.

[0014] Step 2: Use an autoencoder to obtain the learned node representation from each static sparse subgraph, concatenate the feature matrices obtained in steps 2 and 3 into the feature matrix of the DDI network, use the random edge segmentation method to segment the DDI network into multiple static sparse subgraphs with non-overlapping edges, use an autoencoder to obtain the learned node representation in each subgraph and aggregate them together to obtain the joint representation of the graph. We define an undirected, unweighted large graph as G = (V,E), where V represents the vertex set, i.e., the drug set, and E represents the edge set, i.e., the DDI between drugs. We denote its adjacency matrix as A∈R N*N The node set is represented as X∈R N*F We assume that the diagonal elements of the adjacency matrix are set to 1, i.e., including self-loops. For the autoencoder setting, we define the latent vector Z∈R N*d , where N = |V| represents the total number of nodes in the graph, i.e., the total number of drugs, and f and d are the feature dimensions of nodes in X and Z, respectively. We define k subgraphs, their node representations, adjacency matrices, edges, and potential representations as , , , and{ }, we formally define the kth subgraph as , where v k represents a vertex set, e k Represents the edge set. Adjacency matrix a k can be considered as a masked version of A while preserving the self-loops of all nodes: a k = M k ⊙ A. In the data preparation step, the training set is further divided into static subgraphs, each subgraph undergoes graph diffusion convolution to obtain its corresponding latent representation, and finally we concatenate the latent features of K subgraphs.

[0015] Step 3: The embedding of different drug features is fused through the multimodal feature fusion module to obtain the drug pair information in DDI. The feature matrices of steps 2, 3 and 4 are concatenated. The two drugs enter the drug feature encoder composed of multi-head attention and linear layers respectively. The two drug features are concatenated and reduced to obtain the drug pair features. A drug pair contains two drugs (i.e., drug A and drug B). Through the drug feature encoder module containing a multi-head attention module and three linear layers, we can obtain three potential features of drug A: DA1, DA2 and DA3. Similarly, we can obtain three potential features of drug B: DB1, DB2 and DB3, where the multi-head attention is expressed as: , , ; , , We concatenate DA1 and DB1, DA2 and DB2, and DA3 and DB3 respectively to represent a drug pair. Then, we input the concatenated features into the fully connected layer to obtain the fused drug pair features FD1, FD2, and FD3.

[0016] Step 4: Predict the DDI category. We concatenate the drug pair features FD1, FD2, and FD3 to obtain the total drug pair features and feed them into two fully connected layers to obtain the category of the corresponding DDI.

[0017] Figure 2 It is the performance index of the algorithm of the present invention (a drug interaction prediction method integrating multimodal drug information and static subgraph) in dataset 1. In this study, six multi-classification task indicators were selected to evaluate the drug-drug interaction (DDI) event prediction performance of the model. These indicators are precision, recall, F1 score, accuracy, area under the ROC curve (AUC), and area under the precision-recall curve (AUPR). Our model achieved an F1 score of 0.9070 and a Precision score of 0.9106 under the best parameter setting, which are 2.8 and 2.94 percentage points higher than the SOTA model MDDI-SCL, respectively, showing excellent classification performance and effective ranking of prediction probabilities of different DDI categories. In addition, the model has a high AUPR of 0.9830, AUC of 0.9983, and accuracy (ACC) of 0.9515. It further supports its superior performance relative to other models.

Claims

1. A drug interaction event prediction method based on multimodal graph diffusion static subgraph, characterized in that: The following steps are involved: Step 1.1: Calculate the Jaccard similarity between the drug's enzyme features, target features, and chemical substructures to obtain the drug's enzyme feature similarity matrix, target feature similarity matrix, and chemical substructure similarity matrix, respectively. Study and concatenate the three similarity matrices of enzyme features, target features, and chemical substructures to obtain the drug feature similarity matrix; Step 1.2: Use the pre-trained NLP model to encode the pharmacological text features of the drug and obtain the pharmacological text feature encoding matrix of the drug; Step 1.3: Combine the drug feature similarity matrix and the pharmacology text feature encoding matrix, obtain the drug initial representation matrix through the multi-head attention mechanism, use the autoencoder to learn the representation of the drug initial matrix, and use the hidden layer representation of the autoencoder as the drug biological feature representation to obtain the drug biological feature representation matrix; Step 1.4: Use the random edge segmentation method to segment the DDI network into multiple static sparse subgraphs with non-overlapping edges, and use graph diffusion convolution to perform representation learning on the drug biological feature representation matrix of (1.3) to obtain the drug structure feature representation matrix; Step 2: Use the multimodal feature fusion module to learn the drugs and obtain the drug pair representation vector.

2. A drug interaction event prediction method based on multimodal graph diffusion static subgraph according to claim 1, characterized in that: The pharmacological text feature encoding algorithm of the drug using the pre-trained NLP model in step 1.2 is: The pre-trained Bio-BERT model is used to process the pharmacological text attributes of drugs and obtain vector features. For example, for the pharmacology text "Aceclofenac is a NSAID that inhibits both isoforms of COX enzyme....", the text sequence is represented as a tag sequence: X = {[CLS], Aceclofenac is a NSAID that inhibits both isoforms of COX enzyme. [SEP]}, input representation: The text sequence X is converted into the input representation of Bio-BERT, including the following three embeddings: Token Embeddings: Convert each token into a corresponding vector representation, Position Embeddings: Add the position information of the token in the sequence, Segment Embeddings: Indicates the sentence to which the token belongs (usually zero in single sentence tasks), Combining the above three embeddings, the input representation is: Input_Embedding(x i )=E token (x i )+E position (i)+E segment (i), Input to the Bio-BERT model: The input is embedded into the Bio-BERT model. The core of Bio-BERT is a multi-layer Transformer structure. Each layer uses a multi-head self-attention mechanism to capture contextual information. For example, Bio-BERT has L layers, and the input is processed by each layer of Transformer. H 0 =Input_Embedding(X) H (l) =Trans former (l) (H (l-l) ), for l=1,2,...,L, the output embedding of the final layer is: H (L) = BioBERT(X), extract feature embeddings: Extract feature embeddings from the output of Bio-BERT, typically using the embedding of the first token [CLS] as the feature representation of the entire sequence:

3. The method for predicting drug interaction events based on multimodal graph diffusion static subgraphs according to claim 1, characterized in that: The algorithm for obtaining node representation from each static sparse subgraph by graph diffusion convolution in step 1.4 is: Define an undirected, unweighted large graph G = (V, E), where V represents the vertex set, i.e., the drug set, and E represents the edge set, i.e., the DDI between drugs. We denote its adjacency matrix as A∈R N*N The node set is represented as X∈R N*F We assume that the diagonal elements of the adjacency matrix are set to 1, i.e., including self-loops,} We formally define the kth subgraph as g k =(v k , e k ) where v k represents a vertex set, e k represents the edge set, the adjacency matrix a k can be considered as a masked version of A while preserving the self-loops of all nodes: a k =M k ⊙A, in the data preparation step, the training set is further divided into static subgraphs, and each subgraph undergoes graph diffusion convolution to obtain its corresponding latent representation.

Images