Applying deep learning to the pervasive graph data is significant because of the unique characteristics of graphs. Recently, substantial amounts of research efforts have been keen on this area, greatly advancing graph-analyzing techniques. In this study, the authors comprehensively review different kinds of deep learning methods applied to graphs. They discuss with existing literature into sub-components of two: graph convolutional networks, graph autoencoders, and recent trends including chemoinformatics research area including molecular fingerprints and drug discovery. They further experiment with variational autoencoder (VAE) analyze how these apply in drug target interaction (DTI) and applications with ephemeral outline on how they assist the drug discovery pipeline and discuss potential research directions.
TopIntroduction
Graph neural networks (GNNs) are deep learning-centered methods that function in the graph region. Due to its substantial performance and high interpretability, GNN has been a widely applied graph analysis method recently. In the following paragraphs, we will illustrate the fundamental motivations of graph neural networks. Graphs are a kind of data structure, which models a set of objects (nodes) and their relationships (edges). Newly, researches of analyzing graphs with machine learning have been getting more and more attention because of the great communicative power of graphs, i.e. graphs can use as signification of a large number of systems across various areas including science (biomedical networks).
The motivation of GNNs roots in convolutional neural networks (CNNs) (LeCun et. al, 1998). CNN's have the ability to extract multi-scale localized spatial features and compose them to construct highly expressive representations, which led to breakthroughs in almost all machine-learning areas and started the new era of deep learning (LeCun, Bengio, and Hinton, 2015). However, CNN's can only operate on regular Euclidean data like images (2D grid) and text (1D sequence) while the data structures can regarded as instances of graphs. As we are going deeper into CNNs and graphs, we found the keys of CNNs: local connection shared weights and the use of multi-layer (LeCun, Bengio, and Hinton, 2015). However, as shown in Figure 1, it is hard to define localized convolutional filters and pooling operators, which delays the transformation of CNN from Euclidean domain to non-Euclidean domain.
Figure 1(a) Euclidean Space, 1(b) non-Euclidean space
In this chapter, we try to fill this gap by comprehensive reviewing of deep learning methods on graphs. Specifically, as shown in Figure 1, we divide the existing methods into three main categories: semi-supervised methods, unsupervised methods. Concretely speaking, semi-supervised methods include Graph Neural Networks (GNNs) and Graph Convolutional Networks (GCNs), unsupervised methods are mainly composed of Graph Autoencoders (GAEs). We summarize some main distinctions of these categories in Table 1. Approximately, GNNs and GCNs are semi-supervised as they apply node attributes and node labels to train model parameters end-to-end for a specific task, while GAEs mainly focus on learning representation using unsupervised methods. Recently advanced methods use other unique algorithms that do not fall in previous categories. Besides these high-level distinctions, the model architectures also differ significantly. We will provide a comprehensive overview of these graph methods in detail. We also analyze the differences between these models and how to composite different architectures. In the end, we briefly outline the applications of these methods applied in chemoinformatics and discuss potential future directions.
The tasks for learning a deep model on graphs have considered, and that will be the contributions of this chapter:
- •
To apply and provide a fundamental knowledge associated with graph deep learning models for prediction, and node recommendation.
- •
To provide a basis for the use of graphs and nodes for chemoinformatics, a summary of autoencoders for graphs and nodes has presented.
- •
Applications of graph neural networks, graph convolutional networks (GCN), graph autoencoders (GAE) models, which uses in chemoinformatics, has described in detail.
The rest of the chapter has organized as follows. In the next section, we give the formal graph neural networks with benchmarks. In Section 3, recent Graph Convolutional networks (GCNs). Followed by applications in chemoinformatics explaining Graph Autoencoders (GAE) in molecule design have presented in Section 4, and 5. We worked on the application areas of these in experimental analysis with drug target interaction (DTI) using VAE in section 6. The chapter has concluded with an overview in Section 7.