Computers in Biology and Medicine 149 (2022) 105984 Available online 18 August 2022 0010-4825/© 2022 Elsevier Ltd. All rights reserved. DeepPSE: Prediction of polypharmacy side effects by fusing deep representation of drug pairs and attention mechanism Shenggeng Lin a , Guangwei Zhang b , Dong-Qing Wei a, c, d, * , Yi Xiong a, e, ** a State Key Laboratory of Microbial Metabolism, School of Life Sciences and Biotechnology, Shanghai Jiao Tong University, Shanghai, 200240, China b School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, Guangdong, 510275, China c Zhongjing Research and Industrialization Institute of Chinese Medicine, Zhongguancun Scientifc Park, Meixi, Nayang, Henan, 473006, China d Peng Cheng National Laboratory, Vanke Cloud City Phase I Building 8, Xili Street, Nanshan District, Shenzhen, Guangdong, 518055, China e Shanghai Artifcial Intelligence Laboratory, Shanghai, China A R T I C L E INFO Keywords: Polypharmacy side effect prediction Drug-drug interactions Feature fusion Self-attention mechanism ABSTRACT Polypharmacy (multiple use of drugs) is an effective strategy for combating complex or co-existing diseases. However, a major consequence of polypharmacy is a higher risk of adverse side effects due to drug-drug in- teractions, which are rare and observed in relatively small clinical testing. Thus, identifcation of polypharmacy side effects remains challenging. Here, we propose a deep learning-based method, DeepPSE, to predict poly- pharmacy side effects in an end-to-end way. DeepPSE is composed of two main modules. First, multiple types of neural networks are constructed and fused to learn the deep representation of a drug pair. Second, the encoder block of transformer that includes self-attention mechanism is built to get latent features, which are further fed into the fully connected layer to predict polypharmacy side effects of drug pairs. Further, DeepPSE is compared with fve baseline or state-of-the-art methods on a benchmark dataset of 964 types of polypharmacy side effects across 63473 drug pairs. Experimental results demonstrate that DeepPSE achieves better performance than that of all fve methods. The source codes and data are available at https://github.com/ShenggengLin/DeepPSE 1. Introduction Polypharmacy (i.e., multiple drugs are jointly used) is an effective strategy for combating complex or co-existing diseases [1–4]. A major consequence of polypharmacy to a patient is a much higher risk of side effects due to adverse drug-drug interactions [5–8]. Reliable identif- cation of polypharmacy side effects is challenging because they are rare and observed in relatively small clinical testing. It is practically impos- sible to experimentally identify the polypharmacy side effects of all possible pairs of drugs. Therefore, it is desirable and urgent to develop computational methods to predict polypharmacy side effects, which is vital to drug discovery and development [9–15]. In recent years, side effect data of single drugs or drug combinations are collected from relevant literature, clinical trials, laboratory studies and electronic medical records to construct databases, which facilitate the development of computational methods for predicting polypharmacy side effects. Since 2018, there are several studies to develop data-driven or/and knowledge-driven approaches to predict polypharmacy side effects by deep neural network (DNN) [16], graph convolutional network (GCN) [17] and knowledge graph (KG) repre- sentation learning methods [18–21]. The pioneering study by Zitnik et al. [17] constructed the benchmark dataset of 964 commonly occur- ring types of polypharmacy side effects across 63473 drug combinations. Then, they formulated polypharmacy side effect modeling as a multi- relational link prediction problem on a multimodal graph consisting of drug, protein and side effect relationships. They proposed Decagon to predict what will the exact type of the side effect be for a given pair of drugs by using GCN in an end-to-end way, based on a multimodal graph of protein-protein interactions, drug-target interactions and DDIs, where each side effect is an edge of a different type. This architecture becomes a baseline for several other state-of-the-art methods for polypharmacy side effect prediction. * Corresponding author. State Key Laboratory of Microbial Metabolism, School of Life Sciences and Biotechnology, Shanghai Jiao Tong University, Shanghai, 200240, China. ** Corresponding author. State Key Laboratory of Microbial Metabolism, School of Life Sciences and Biotechnology, Shanghai Jiao Tong University, Shanghai, 200240, China. E-mail addresses: dqwei@sjtu.edu.cn (D.-Q. Wei), xiongyi@sjtu.edu.cn (Y. Xiong). Contents lists available at ScienceDirect Computers in Biology and Medicine journal homepage: www.elsevier.com/locate/compbiomed https://doi.org/10.1016/j.compbiomed.2022.105984 Received 22 June 2022; Received in revised form 17 July 2022; Accepted 14 August 2022