همایش ملی بیوانفورماتیک ایران

صفحه اصلی / 4th international edition and 13th Iranian Conference on Bioinformatics

A New Method in Analyzing Drug-Drug Interaction Using Artificial Intelligence: Combination of Channel and Spatial Maps in Deep Attention Networks

نویسندگان :

Seyed Vahab Shojaedini¹ AmirReza Moeini²

1- Biomedical engineering Department, Iranian Research Organization for Science & Technology, Tehran, Iran 2- Department of Computer Engineering, Faculty of Engineering, Islamic Azad University, Qazvin Branch, Qazvin, Iran

کلمات کلیدی :

Drug-Drug Interaction،Adverse Drug Reaction،Attention Deep Learning،Artificial Intelligence.

چکیده :

Drug-Drug Interaction (i.e., DDI) is mutual influence between a drug and another medical treatment which a patient takes. Although such influence may some positive effects but serious adverse effects may occur which are known as Adverse Drug Reactions (Liu et al, 2022). As ADRs may lead to high costs for healthcare system, their identification has been an important part of pharmaceutical research. For several decades laboratory research has been traditional approach for recognizing DDIs, however this scheme is costly and time-consuming. Furthermore in some cases performing experimental approaches are impossible (Safdari et al, 2016). Accordingly, Computer Aided Design (i.e., CAD) methods were gradually used to analyze the DDI, and among them, artificial intelligence methods have been much more widely used than other techniques due to the large volume of pharmaceutical data on the one hand and the presence of complex and contradictory statements in them on the other hand. Among artificial intelligence domain, deep learning methods have been developed during recent years in order to address several challenges of laboratory DDI recognition in parallel with increase the speed and accuracy of the process. In this article, the concept of attention-learning, is utilized in the process of constructing well fit deep neural network for DDI recognition purpose. In order to carry out this idea, firstly channel attention map may be estimated from input DDI patterns by exploiting the inter-channel relationship of features. On the other hand, spatial attention map is extracted from the same input by utilizing the inter-spatial relationships of features. The above attention maps are multiplied by the input feature map for adaptive feature refinement, which leads to training the deep neural network with DDI data by strengthening logical connections between its samples and eliminating redundancies. Thus, applying such enriched input, may improve the performance of deep neural networks in recognizing DDIs, in form of reducing the challenge of overfitting during their training process. The proposed architecture was examined on SemEval-2013 Task-9 dataset, a benchmark in which DDI types has been classified among five classes including: non-interaction, mechanism, effect, advice and int (Yi et al, 2017). The results of the above tests showed satisfactory performance for the proposed structure in DDI recognition in such way that weighted average values for the three parameters precision, recall, and F-score were 85%, 87%, and 86%, respectively, which led to an overall accuracy of 87 percent for correctly diagnosing five mentioned types of DDI. Based on the above investigations, it may be concluded that the combination of channel and spatial maps for deep attention learning may be considered as a high-potential candidate in artificial intelligence domain for computerized recognizing DDI types in parallel with decreasing the risk of ADRs.