Prediction Of Drug-drug Interaction Events Based On Multimodal Neural Network Associations

In recent years,combination of drugs has been a rational strategy for the treatment of disease because they improve therapeutic efficacy and reduce drug resistance,but they can cause adverse reactions in patients,which can have serious consequences and lead to adverse drug events.Therefore,the prediction of potential drug-drug interactions plays a very critical role in patient treatment.However,most of the currently available prediction methods can only predict whether there is an interaction between two drugs,while few methods can predict the interaction events between two drugs.Predicting drug interaction events is more useful for studying the hidden mechanisms behind a combination of drugs or adverse reactions.Although current methods for predicting drug-drug interaction events have performed well,they still suffer from some limitations.On the one hand,most of them do not take into account the complementary nature of drug multimodal data;on the other hand,most deep learning-based models splice or sum two drug feature vectors together and only consider training classifiers on the fused features to predict drug interaction events or averaging/weighting the classification scores under each modality,with few attempts at other fusion approaches.In order to address the above limitations and improve the accuracy of DDI event prediction,a multimodal neural network model is proposed to predict drug interaction events,and the main research work in this paper is shown below:(1)A prediction channel based on heterogeneous features.In this paper,the three features of a drug are combined to calculate the similarity between drugs,and the three feature vectors of the drug pair are stitched together and fed into two autoencoders with self-attention mechanisms to output potential feature vectors,which are used as input to the Transformer module,and then passed through a neural network classifier to obtain an initial vector of interaction event prediction results.(2)A prediction channel based on drug atomic map structures.For each drug in the drug interaction matrix,the corresponding SMILES sequence was collected in DrugBank.The RDKit library in Python was called to convert the SMILES sequences of each drug into the form of an undirected graph,and we used MPAN to generate feature vectors for each drug’s atomic graph.The generated feature vectors for the drug pairs were stitched together and fed into a neural network classifier to obtain an initial vector of interaction event prediction results.(3)A multimodal fusion layer VCDN.The initial prediction result vector of interaction events for drug pairs in both modalities is obtained based on the two channels,respectively.The view correlation discovery network(VCDN)is used to further obtain the final prediction results.The multimodal fusion layer effectively combines biological property information and chemical structure information of drugs,exploring the cross-modal complementary and category uniqueness of multimodal data.Through ablation experiments and performance comparisons with other methods,the results show that the model in this paper outperforms current better model algorithms for the DDI event prediction task.The algorithm is validated to be effective in predicting new drug interactions through case analysis.