Toward heterogeneous information fusion: bipartite graph convolutional networks for in silico drug repurposing

Abstract

Motivation Mining drug–disease association and related interactions are essential for developing in silico drug repurposing (DR) methods and understanding underlying biological mechanisms. Recently, large-scale biological databases are increasingly available for pharmaceutical research, allowing for deep characterization for molecular informatics and drug discovery. However, DR is challenging due to the molecular heterogeneity of disease and diverse drug–disease associations. Importantly, the complexity of molecular target interactions, such as protein–protein interaction (PPI), remains to be elucidated. DR thus requires deep exploration of a multimodal biological network in an integrative context. Results In this study, we propose BiFusion, a bipartite graph convolution network model for DR through heterogeneous information fusion. Our approach combines insights of multiscale pharmaceutical information by constructing a multirelational graph of drug–protein, disease–protein and PPIs. Especially, our model introduces protein nodes as a bridge for message passing among diverse biological domains, which provides insights into utilizing PPI for improved DR assessment. Unlike conventional graph convolution networks always assuming the same node attributes in a global graph, our approach models interdomain information fusion with bipartite graph convolution operation. We offered an exploratory analysis for finding novel drug–disease associations. Extensive experiments showed that our approach achieved improved performance than multiple baselines for DR analysis. Availability and implementation Source code and preprocessed datasets are at: https://github.com/zcwang0702/BiFusion.

Integration strategies of multi-omics data for machine learning analysis

“…Additionally, Graph Neural Networks (GNN) [119] and its upgrade Graph Convolutional Neural Networks (GCN), are specifically designed to receive graph data as input. GCNs have also been developed for DTIs prediction [120] or node classification [121] . Going further than simple link prediction, Zitnik et al (2018) [122] have used a GCN called Decagon to predict the presence of side effects between two drugs as well as the side effects’ type.…”

Predictions, Pivots, and a Pandemic: a Review of 2020's Top Translational Bioinformatics Publications

“…The representations are mainly extracted by deep learning [ 33 34 35 36 37 38 ] or latent probabilistic [ 35 ] methods. These distributed representations, i.e., embedding, are used to encode various modalities of data, including gene expressions [ 39 40 ], events [ 36 ], images [ 33 ], and other relational graph data [ 37 41 ]. The embedding methods are data-driven representations that can capture semantic and contextual information and incorporate them into a numerical representation.…”

“…Use of GCN in DeepCDR [ 43 ] and use of directed-message passing deep neural network model [ 44 ] for antibiotic drug discovery [ 37 ] are among these practices. In multimodal studies [ 41 45 46 ] the information fusion is designed in a graph-based form according to a domain-driven information flow. Wang et al proposed a bipartite GCN for drug re-purposing prediction, which accounts for the central role of proteins in drug-disease association [ 41 ].…”

Predicting activatory and inhibitory drug–target interactions based on mol2vec and genetically perturbed transcriptomes

“…However, their prediction is often unreliable when the 3D structure of a protein or target is unavailable or when an insufficient number of ligands is known for the target (6). Recently, chemical genomic approaches have emerged as an alternative enabling large-scale predictions by leveraging recent advances in network-based approaches or machine learning techniques (7)(8)(9)(10)(11). For example, Yunan et al…”