A Comprehensive Survey on Graph Neural Networks

Wu, Zonghan; Pan, Shirui; Chen, Fengwen; Long, Guodong; Zhang, Chengqi; Yu, Philip S.

doi:10.1109/tnnls.2020.2978386

Cited by 6,235 publications

(3,451 citation statements)

References 114 publications

Supporting

Mentioning

2,977

Contrasting

Unclassified

Order By: Relevance

“…From a machine learning perspective, KPNNs complement graph neural networks (GNNs) in interesting ways. GNNs build on domain network knowledge to define relationships between input nodes, which are iteratively updated based on their neighbors prior to predicting a given output 88,89 . Typical applications of GNNs include predictions of node labels (given labels of some nodes) or graph labels (given a set of graphs).…”

Section: Discussionmentioning

confidence: 99%

Knowledge-primed neural networks enable biologically interpretable deep learning on single-cell sequencing data

Fortelny

Bock

2019

Preprint

View full text Add to dashboard Cite

Deep learning has emerged as a powerful methodology for predicting a variety of complex biological phenomena. However, its utility for biological discovery has so far been limited, given that generic deep neural networks provide little insight into the biological mechanisms that underlie a successful prediction. Here we demonstrate deep learning on biological networks, where every node has a molecular equivalent (such as a protein or gene) and every edge has a mechanistic interpretation (e.g., a regulatory interaction along a signaling pathway).With knowledge-primed neural networks (KPNNs), we exploit the ability of deep learning algorithms to assign meaningful weights to multi-layered networks for interpretable deep learning. We introduce three methodological advances in the learning algorithm that enhance interpretability of the learnt KPNNs: Stabilizing node weights in the presence of redundancy, enhancing the quantitative interpretability of node weights, and controlling for the uneven connectivity inherent to biological networks. We demonstrate the power of our approach on two single-cell RNA-seq datasets, predicting T cell receptor stimulation in a standardized in vitro model and inferring cell type in Human Cell Atlas reference data comprising 483,084 immune cells.In summary, we introduce KPNNs as a method that combines the predictive power of deep learning with the interpretability of biological networks. While demonstrated here on single-cell sequencing data, this method is broadly relevant to other research areas where prior domain knowledge can be represented as networks.To overcome lack of interpretability as a key limitation of deep learning in biology and medicine, here we explore the concept -and demonstrate the feasibility -of deep learning on application-specific biological networks rather than on generic ANNs. By applying deep learning directly to biological networks (such as signaling pathways and gene-regulatory networks), this approach coerces deep learning algorithms to stick closely to the regulatory mechanisms that are relevant in cells, thereby making the learned models interpretable.We developed knowledge-primed neural networks (KPNNs) as a framework for interpretable deep learning on biological networks (applied here to signaling pathways and gene-regulatory networks), and we introduce three modifications to generic deep learning that enhance interpretability: (i) Repeated network training with random deletion of hidden nodes (a technique known as dropout 40 ), which yields robust results in the presence of network redundancy; (ii) dropout on input data in order to enhance quantitative interpretability of node weights; (iii) training on control inputs to normalize for the uneven connectivity of biological networks. To validate our method for interpretable deep learning on biological networks, we applied KPNNs to single-cell RNA-seq data for T cell receptor stimulation 41 and to a reference catalog of immune cells from the Human Cell Atlas 42 . RESULTS KPNNs enable deep learning on biolo...

show abstract

Section: Discussionmentioning

confidence: 99%

Knowledge-primed neural networks enable biologically interpretable deep learning on single-cell sequencing data

Fortelny

Bock

2019

Preprint

View full text Add to dashboard Cite

show abstract

“…Specifically, we consider the spatial organization of mRNAs inside tissues as a spatial functional network where different mRNA types interact based on their spatial proximity [ Figure 1], and where subcellular domains can be identified as clusters of local gene constellations that are shared or cell-type specific. In order to investigate the spatial mRNA network for recurrent gene constellations, we adopted a powerful graph representation learning technique [27] based on graph neural networks (GNN) [28], that has recently emerged as state-of-the-art machine learning technique for leveraging information from graph local neighborhoods. Therefore, each mRNA location is encoded in a graph as a node with a single feature representing the gene it belongs to and it is connected to all the other nodes representing the other mRNAs located in its neighborhood [ Figure 1a].…”

Section: Introductionmentioning

confidence: 99%

Spage2vec: Unsupervised detection of spatial gene expression constellations

Partel

Wählby

2020

Preprint

View full text Add to dashboard Cite

Investigation of spatial cellular composition of tissue architectures revealed by multiplexed in situ RNA detection often rely on inaccurate cell segmentation or prior biological knowledge from complementary single cell sequencing experiments. Here we present spage2vec, an unsupervised segmentation free approach for decrypting the spatial transcriptomic heterogeneity of complex tissues at subcellular resolution. Spage2vec represents the spatial transcriptomic landscape of tissue samples as a spatial functional network and leverages a powerful machine learning graph representation technique to create a lower dimensional representation of local spatial gene expression. We apply spage2vec to mouse brain data from three different in situ transcriptomic assays, showing that learned representations encode meaningful biological spatial information of re-occuring gene constellations involved in cellular and subcellular processes.

show abstract

“…GCNs have been introduced in the machine learning literature a few years ago [2]. Their main advantage is that they can utilize the power of convolutional NN even for cases where spatial relationships are not complete [29,31]. Specifically, rather than encoding the data using a 2D matrix (or a 1D vector) GCNs use the graph structure to encode relationships between samples.…”

Section: Introductionmentioning

confidence: 99%

“…The core structure of GCN is its graph convolutional layer, which enables it to combine graph structure (cell location) and node information (gene expression in specific cell) as inputs to a neural network. Since the graph structure encodes spatially related cells, GCNs can utilize convolutional layers that underly much of the recent success of neural networks, without directly using image data [29,31].…”

Section: Introductionmentioning

confidence: 99%

GCNG: Graph convolutional networks for inferring cell-cell interactions

Yuan

Bar‐Joseph

2019

Preprint

View full text Add to dashboard Cite

Several methods have been developed for inferring gene-gene interactions from expression data. To date, these methods mainly focused on intra-cellular interactions. The availability of high throughput spatial expression data opens the door to methods that can infer such interactions both within and between cells. However, the spatial data also raises several new challenges. These include issues related to the sparse, noisy expression vectors for each cell, the fact that several different cell types are often profiled, the definition of a neighborhood of cell and the relatively small number of extracellular interactions. To enable the identification of gene interactions between cells we extended a Graph Convolutional Neural network approach for Genes (GCNG). We encode the spatial information as a graph and use the network to combine it with the expression data using supervised training. Testing GCNG on spatial transcriptomics data we show that it improves upon prior methods suggested for this task and can propose novel pairs of extracellular interacting genes. Finally, we show that the output of GCNG can also be used for down-stream analysis including functional assignment.

show abstract

A Comprehensive Survey on Graph Neural Networks

Cited by 6,235 publications

References 114 publications

Knowledge-primed neural networks enable biologically interpretable deep learning on single-cell sequencing data

Knowledge-primed neural networks enable biologically interpretable deep learning on single-cell sequencing data

Spage2vec: Unsupervised detection of spatial gene expression constellations

GCNG: Graph convolutional networks for inferring cell-cell interactions

Contact Info

Product

Resources

About