Graph-based Deep Learning Research Articles

Uncertain Graph Neural Networks for Facial Action Unit Detection

Capturing the dependencies among different facial action units (AU) is extremely important for the AU detection task. Many studies have employed graph-based deep learning methods to exploit the dependencies among AUs. However, the dependencies among AUs in real world data are often noisy and the uncertainty is essential to be taken into consideration. Rather than employing a deterministic mode, we propose an uncertain graph neural network (UGN) to learn the probabilistic mask that simultaneously captures both the individual dependencies among AUs and the uncertainties. Further, we propose an adaptive weighted loss function based on the epistemic uncertainties to adaptively vary the weights of the training samples during the training process to account for unbalanced data distributions among AUs. We also provide an insightful analysis on how the uncertainties are related to the performance of AU detection. Extensive experiments, conducted on two benchmark datasets, i.e., BP4D and DISFA, demonstrate our method achieves the state-of-the-art performance.

An effective self-supervised framework for learning expressive molecular global representations to drug discovery.

How to produce expressive molecular representations is a fundamental challenge in artificial intelligence-driven drug discovery. Graph neural network (GNN) has emerged as a powerful technique for modeling molecular data. However, previous supervised approaches usually suffer from the scarcity of labeled data and poor generalization capability. Here, we propose a novel molecular pre-training graph-based deep learning framework, named MPG, that learns molecular representations from large-scale unlabeled molecules. In MPG, we proposed a powerful GNN for modelling molecular graph named MolGNet, and designed an effective self-supervised strategy for pre-training the model at both the node and graph-level. After pre-training on 11 million unlabeled molecules, we revealed that MolGNet can capture valuable chemical insights to produce interpretable representation. The pre-trained MolGNet can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of drug discovery tasks, including molecular properties prediction, drug-drug interaction and drug-target interaction, on 14 benchmark datasets. The pre-trained MolGNet in MPG has the potential to become an advanced molecular encoder in the drug discovery pipeline.

Structure motif-centric learning framework for inorganic crystalline systems.

Incorporation of physical principles in a machine learning (ML) architecture is a fundamental step toward the continued development of artificial intelligence for inorganic materials. As inspired by the Pauling's rule, we propose that structure motifs in inorganic crystals can serve as a central input to a machine learning framework. We demonstrated that the presence of structure motifs and their connections in a large set of crystalline compounds can be converted into unique vector representations using an unsupervised learning algorithm. To demonstrate the use of structure motif information, a motif-centric learning framework is created by combining motif information with the atom-based graph neural networks to form an atom-motif dual graph network (AMDNet), which is more accurate in predicting the electronic structures of metal oxides such as bandgaps. The work illustrates the route toward fundamental design of graph neural network learning architecture for complex materials by incorporating beyond-atom physical principles.

Graph-based deep learning frameworks for molecules and solid-state materials

Recent years have witnessed the rapid increase in the application of deep learning in atomistic systems including both molecules and solid-state materials. The use of graphs and associated design of message passing strategies have enabled multiple deep learning frameworks to achieve reliable and efficient predictions of materials properties with a much smaller cost compared with first-principles atomistic simulations. In this review, we will focus on recent development of graph-based deep learning frameworks and their applications for both molecules and solid-state material systems. The history of the development of graph-based representations for molecules and crystals will be introduced. Essential learning processes defined by the so-called message passing will be reviewed, based on which the performance of different models will be compared. Furthermore, recent development of graph learning frameworks that incorporate material information beyond atom level will be introduced. Current challenges and future perspectives on this emerging field at the crossroad of material science, physics, chemistry, and computer science will be given, with an emphasize on how multiple tiers of material information can be organized and combined in a graph neural network setup.

Hyperspectral Image Classification with Localized Graph Convolutional Filtering

The nascent graph representation learning has shown superiority for resolving graph data. Compared to conventional convolutional neural networks, graph-based deep learning has the advantages of illustrating class boundaries and modeling feature relationships. Faced with hyperspectral image (HSI) classification, the priority problem might be how to convert hyperspectral data into irregular domains from regular grids. In this regard, we present a novel method that performs the localized graph convolutional filtering on HSIs based on spectral graph theory. First, we conducted principal component analysis (PCA) preprocessing to create localized hyperspectral data cubes with unsupervised feature reduction. These feature cubes combined with localized adjacent matrices were fed into the popular graph convolution network in a standard supervised learning paradigm. Finally, we succeeded in analyzing diversified land covers by considering local graph structure with graph convolutional filtering. Experiments on real hyperspectral datasets demonstrated that the presented method offers promising classification performance compared with other popular competitors.

Auto-Metric Graph Neural Network Based on a Meta-Learning Strategy for the Diagnosis of Alzheimer's Disease.

Alzheimer's disease (AD) is the most common cognitive disorder. In recent years, many computer-aided diagnosis techniques have been proposed for AD diagnosis and progression predictions. Among them, graph neural networks (GNNs) have received extensive attention owing to their ability to effectively fuse multimodal features and model the correlation between samples. However, many GNNs for node classification use an entire dataset to construct a large fixed-graph structure, which cannot be used for independent testing. To overcome this limitation while maintaining the advantages of the GNN, we propose an auto-metric GNN (AMGNN) model for AD diagnosis. First, a metric-based meta-learning strategy is introduced to realize inductive learning for independent testing through multiple node classification tasks. In the meta-tasks, the small graphs help make the model insensitive to the sample size, thus improving the performance under small sample size conditions. Furthermore, an AMGNN layer with a probability constraint is designed to realize node similarity metric learning and effectively fuse multimodal data. We verified the model on two tasks based on the TADPOLE dataset: early AD diagnosis and mild cognitive impairment (MCI) conversion prediction. Our model provides excellent performance on both tasks with accuracies of 94.44% and 87.50% and median accuracies of 94.19% and 86.25%, respectively. These results show that our model improves flexibility while ensuring a good classification performance, thus promoting the development of graph-based deep learning algorithms for disease diagnosis.

Deep Graph Generators: A Survey

Deep generative models have achieved great success in areas such as image, speech, and natural language processing in the past few years. Thanks to the advances in graph-based deep learning, and in particular graph representation learning, deep graph generation methods have recently emerged with new applications ranging from discovering novel molecular structures to modeling social networks. This paper conducts a comprehensive survey on deep learning-based graph generation approaches and classifies them into five broad categories, namely, autoregressive, autoencoder-based, reinforcement learning-based, adversarial, and flow-based graph generators, providing the readers a detailed description of the methods in each class. We also present publicly available source codes, commonly used datasets, and the most widely utilized evaluation metrics. Finally, we review current trends and suggest future research directions based on the existing challenges.

Deep Graph Learning with Property Augmentation for Predicting Drug-Induced Liver Injury.

Drug-induced liver injury (DILI) is a crucial factor in determining the qualification of potential drugs. However, the DILI property is excessively difficult to obtain due to the complex testing process. Consequently, an in silico screening in the early stage of drug discovery would help to reduce the total development cost by filtering those drug candidates with a high risk to cause DILI. To serve the screening goal, we apply several computational techniques to predict the DILI property, including traditional machine learning methods and graph-based deep learning techniques. While deep learning models require large training data to tune huge model parameters, the DILI data set only contains a few hundred annotated molecules. To alleviate the data scarcity problem, we propose a property augmentation strategy to include massive training data with other property information. Extensive experiments demonstrate that our proposed method significantly outperforms all existing baselines on the DILI data set by obtaining a 81.4% accuracy using cross-validation with random splitting, 78.7% using leave-one-out cross-validation, and 76.5% using cross-validation with scaffold splitting.

Multiprincipal Component P2-Na0.6(Ti0.2Mn0.2Co0.2Ni0.2Ru0.2)O2 as a High-Rate Cathode for Sodium-Ion Batteries.

Mixing transition metal cations in nearly equiatomic proportions in layered oxide cathode materials is a new strategy for improving the performances of Na-ion batteries. The mixing of cations not only offers entropic stabilization of the crystal structure but also benefits the diffusion of Na ions with tuned diffusion activation energy barriers. In light of this strategy, a high-rate Na0.6(Ti0.2Mn0.2Co0.2Ni0.2Ru0.2)O2 cathode was designed, synthesized, and investigated, combining graph-based deep learning calculations and complementary experimental characterizations. This new cathode material delivers high discharge capacities of 164 mA g–1 at 0.1 C and 68 mAh g–1 at a very high rate of 86 C, demonstrating an outstanding high rate capability. Ex situ and operando synchrotron X-ray diffraction were used to reveal the detailed structural evolution of the cathode upon cycling. Using the climbing-image nudged elastic-band calculation and Ab initio molecular dynamics simulations, we show that the optimal transition metal composition enables a percolating network of low barrier pathways for fast, macroscopic Na diffusion, resulting in the observed high rate performance.

Identifying influential neighbors in social networks and venue affiliations among young MSM: a data science approach to predict HIV infection.

Young MSM (YMSM) bear a disproportionate burden of HIV infection in the United States and their risks of acquiring HIV may be shaped by complex multilayer social networks. These networks are formed through not only direct contact with social/sex partners but also indirect anonymous contacts encountered when attending social venues. We introduced a new application of a state-of-the-art graph-based deep learning method to predict HIV infection that can identify influential neighbors within these multiple network contexts. We used empirical network data among YMSM aged 16-29 years old collected from Houston and Chicago in the United States between 2014 and 2016. A computational framework GAT-HIV (Graph Attention Networks for HIV) was proposed to predict HIV infections by identifying influential neighbors within social networks. These networks were formed by multiple relations constituted of social/sex partners and shared venue attendances, and using individual-level variables. Further, GAT-HIV was extended to combine multiple social networks using multigraph GAT methods. A visualization tool was also developed to highlight influential network members for each individual within the multiple social networks. The multigraph GAT-HIV models obtained average AUC values of 0.776 and 0.824 for Chicago and Houston, respectively, performing better than empirical predictive models (e.g. AUCs of random forest: 0.758 and 0.798). GAT-HIV on single networks also delivered promising prediction performances. The proposed methods provide a comprehensive and interpretable framework for graph-based modeling that may inform effective HIV prevention intervention strategies among populations most vulnerable to HIV.

Spectral-Based Directed Graph Network for Malware Detection

As a kind of behavioral-feature based malware detection approach, spectral graph-based deep learning has attracted considerable research efforts with the fast growth of threats of malicious programs. However, previous spectral based graph neural networks can hardly be applied to directed graphs due to the asymmetrical nature of the graph adjacency matrix. In order to address the issues of existing techniques, we propose a Spectral-based Directed Graph Network (SDGNet) architecture to classify directed graphs. In SDGNet, the weighted graph matrix normalization methods transform the graph adjacency matrix into three symmetrical graph matrices that describe different aspects of node information interaction. Then, the SDGNet extracts graph representations with different layers of multi-aspect directed GCN. On each layer, three node embeddings learned from the symmetrical graph matrices are fused together for a graph representation. The multi-layer graph representations are further concatenated together to form a comprehensive representation for classification with a combined loss function. We evaluate the proposed algorithm on a public benchmark data, and the experimental results show that it outperforms state-of-the-art algorithms.

Pseudo2GO: A Graph-Based Deep Learning Method for Pseudogene Function Prediction by Borrowing Information From Coding Genes.

Pseudogenes are indicating more and more functional potentials recently, though historically were regarded as relics of evolution. Computational methods for predicting pseudogene functions on Gene Ontology is important for directing experimental discovery. However, no pseudogene-specific computational methods have been proposed to directly predict their Gene Ontology (GO) terms. The biggest challenge for pseudogene function prediction is the lack of enough features and functional annotations, making training a predictive model difficult. Considering the close functional similarity between pseudogenes and their parent coding genes that share great amount of DNA sequence, as well as that coding genes have rich annotations, we aim to predict pseudogene functions by borrowing information from coding genes in a graph-based way. Here we propose Pseudo2GO, a graph-based deep learning semi-supervised model for pseudogene function prediction. A sequence similarity graph is first constructed to connect pseudogenes and coding genes. Multiple features are incorporated into the model as the node attributes to enable the graph an attributed graph, including expression profiles, interactions with microRNAs, protein-protein interactions (PPIs), and genetic interactions. Graph convolutional networks are used to propagate node attributes across the graph to make classifications on pseudogenes. Comparing Pseudo2GO with other frameworks adapted from popular protein function prediction methods, we demonstrated that our method has achieved state-of-the-art performance, significantly outperforming other methods in terms of the M-AUPR metric.

Graph-based prediction of Protein-protein interactions with attributed signed graph embedding

BackgroundProtein-protein interactions (PPIs) are central to many biological processes. Considering that the experimental methods for identifying PPIs are time-consuming and expensive, it is important to develop automated computational methods to better predict PPIs. Various machine learning methods have been proposed, including a deep learning technique which is sequence-based that has achieved promising results. However, it only focuses on sequence information while ignoring the structural information of PPI networks. Structural information of PPI networks such as their degree, position, and neighboring nodes in a graph has been proved to be informative in PPI prediction.ResultsFacing the challenge of representing graph information, we introduce an improved graph representation learning method. Our model can study PPI prediction based on both sequence information and graph structure. Moreover, our study takes advantage of a representation learning model and employs a graph-based deep learning method for PPI prediction, which shows superiority over existing sequence-based methods. Statistically, Our method achieves state-of-the-art accuracy of 99.15% on Human protein reference database (HPRD) dataset and also obtains best results on Database of Interacting Protein (DIP) Human, Drosophila, Escherichia coli (E. coli), and Caenorhabditis elegans (C. elegan) datasets.ConclusionHere, we introduce signed variational graph auto-encoder (S-VGAE), an improved graph representation learning method, to automatically learn to encode graph structure into low-dimensional embeddings. Experimental results demonstrate that our method outperforms other existing sequence-based methods on several datasets. We also prove the robustness of our model for very sparse networks and the generalization for a new dataset that consists of four datasets: HPRD, E.coli, C.elegan, and Drosophila.

How to Build a Graph-Based Deep Learning Architecture in Traffic Domain: A Survey

In recent years, various deep learning architectures have been proposed to solve complex challenges (e.g. spatial dependency, temporal dependency) in traffic domain, which have achieved satisfactory performance. These architectures are composed of multiple deep learning techniques in order to tackle various challenges in traffic tasks. Traditionally, convolution neural networks (CNNs) are utilized to model spatial dependency by decomposing the traffic network as grids. However, many traffic networks are graph-structured in nature. In order to utilize such spatial information fully, it's more appropriate to formulate traffic networks as graphs mathematically. Recently, various novel deep learning techniques have been developed to process graph data, called graph neural networks (GNNs). More and more works combine GNNs with other deep learning techniques to construct an architecture dealing with various challenges in a complex traffic task, where GNNs are responsible for extracting spatial correlations in traffic network. These graph-based architectures have achieved state-of-the-art performance. To provide a comprehensive and clear picture of such emerging trend, this survey carefully examines various graph-based deep learning architectures in many traffic applications. We first give guidelines to formulate a traffic problem based on graph and construct graphs from various kinds of traffic datasets. Then we decompose these graph-based architectures to discuss their shared deep learning techniques, clarifying the utilization of each technique in traffic tasks. What's more, we summarize some common traffic challenges and the corresponding graph-based deep learning solutions to each challenge. Finally, we provide benchmark datasets, open source codes and future research directions in this rapidly growing field.

Graph weeds net: A graph-based deep learning method for weed recognition

Robotic weed control through weed detection has become increasingly important due to mounting pressure on herbicides from resistance and the large impact of weeds on agricultural productivity. One of the major challenges is accurate classification of weed species for selective targeting in crop situations, whilst the existing studies are often conducted in well-controlled settings with consistent lighting, species and backgrounds. Therefore, in this study, we propose a novel graph-based deep learning architecture, namely Graph Weeds Net (GWN), which aims to recognize multiple types of weeds from conventional RGB images collected from complex rangelands. GWN collects regional patterns in line with set image scopes and formulates multi-scale graph representations for weed classification. Additionally, GWN provides suggestions for key regions, creating opportunities for further within-image actions for robotic in-field systems. The architecture was evaluated on a recently published benchmark dataset, achieving the state-of-the-art performance with a top-1 accuracy 98.1%.

DLPNet: A deep manifold network for feature extraction of hyperspectral imagery

Deep learning has received increasing attention in recent years and it has been successfully applied for feature extraction (FE) of hyperspectral images. However, most deep learning methods fail to explore the manifold structure in hyperspectral image (HSI). To tackle this issue, a novel graph-based deep learning model, termed deep locality preserving neural network (DLPNet), was proposed in this paper. Traditional deep learning methods use random initialization to initialize network parameters. Different from that, DLPNet initializes each layer of the network by exploring the manifold structure in hyperspectral data. In the stage of network optimization, it designed a deep-manifold learning joint loss function to exploit graph embedding process while measuring the difference between the predictive value and the actual value, then the proposed model can take into account the extraction of deep features and explore the manifold structure of data simultaneously. Experimental results on real-world HSI datasets indicate that the proposed DLPNet performs significantly better than some state-of-the-art methods.

KGCN: a graph-based deep learning framework for chemical structures

Deep learning is developing as an important technology to perform various tasks in cheminformatics. In particular, graph convolutional neural networks (GCNs) have been reported to perform well in many types of prediction tasks related to molecules. Although GCN exhibits considerable potential in various applications, appropriate utilization of this resource for obtaining reasonable and reliable prediction results requires thorough understanding of GCN and programming. To leverage the power of GCN to benefit various users from chemists to cheminformaticians, an open-source GCN tool, kGCN, is introduced. To support the users with various levels of programming skills, kGCN includes three interfaces: a graphical user interface (GUI) employing KNIME for users with limited programming skills such as chemists, as well as command-line and Python library interfaces for users with advanced programming skills such as cheminformaticians. To support the three steps required for building a prediction model, i.e., pre-processing, model tuning, and interpretation of results, kGCN includes functions of typical pre-processing, Bayesian optimization for automatic model tuning, and visualization of the atomic contribution to prediction for interpretation of results. kGCN supports three types of approaches, single-task, multi-task, and multi-modal predictions. The prediction of compound-protein interaction for four matrixmetalloproteases, MMP-3, -9, -12 and -13, in the inhibition assays is performed as a representative case study using kGCN. Additionally, kGCN provides the visualization of atomic contributions to the prediction. Such visualization is useful for the validation of the prediction models and the design of molecules based on the prediction model, realizing “explainable AI” for understanding the factors affecting AI prediction. kGCN is available at https://github.com/clinfo.

Semi-supervised graph-based deep learning for multi-modal prediction of knee osteoarthritis incidence

Purpose: Osteoarthritis (OA) is a multifactorial disorder accompanied by biochemical and morphological changes in the articular cartilage, modulated by skeletal biomechanics and gait. While we can acquire detailed information about OA patients, we are not yet able to leverage the multifactorial factors for diagnosis and disease management of knee OA. Recent advancements in machine learning can help in achieving this goal; however, obstacles in collection of large enough samples to train complex models and lack of established methods for multi-modal data fusion make the prediction of future OA onset still an unmet challenge.

Assessing Graph-based Deep Learning Models for Predicting Flash Point.

Flash points of organic molecules play an important role in preventing flammability hazards and large databases of measured values exist, although millions of compounds remain unmeasured. To rapidly extend existing data to new compounds many researchers have used quantitative structure-property relationship (QSPR) analysis to effectively predict flash points. In recent years graph-based deep learning (GBDL) has emerged as a powerful alternative method to traditional QSPR. In this paper, GBDL models were implemented in predicting flash point for the first time. We assessed the performance of two GBDL models, message-passing neural network (MPNN) and graph convolutional neural network (GCNN), by comparing against 12 previous QSPR studies using more traditional methods. Our result shows that MPNN both outperforms GCNN and yields slightly worse but comparable performance with previous QSPR studies. The average and Mean Absolute Error (MAE) scores of MPNN are, respectively, 2.3 % lower and 2.0 K higher than previous comparable studies. To further explore GBDL models, we collected the largest flash point dataset to date, which contains 10575 unique molecules. The optimized MPNN gives a test data of 0.803 and MAE of 17.8 K on the complete dataset. We also extracted 5 datasets from our integrated dataset based on molecular types (acids, organometallics, organogermaniums, organosilicons, and organotins) and explore the quality of the model in these classes.

Synchrophasor Recovery and Prediction: A Graph-Based Deep Learning Approach

Data integrity of power system states is critical to modern power grid operation and control due to communication latency, state measurements are not immediately available at the control center, rendering slow responses of time-sensitive applications. In this paper, a new graph-based deep learning approach is proposed to recover and predict the states ahead of time utilizing the power network topology and existing measurements. A graph-convolutional recurrent adversarial network is devised to process available information and extract graphical and temporal data correlations. This approach overcomes drawbacks of the existing synchrophasor recovery and prediction implementation to improve the overall system performance. Additionally, the approach offers an adaptive data processing method to handle power grids of various sizes. Case studies demonstrate the outstanding recovery and prediction accuracy of the proposed approach, and investigations are conducted to illustrate its robustness against bad communication conditions, measurement noise, and system topology changes.