Graph Neural Networks (GNN)
Graphs are a general language for describing and analyzing entities with relationships/interactions.
Graph Neural Networks (GNN) are a new form of artificial neural network based on graphs. A graph consists of several points (nodes or vertices) that are connected to each other (by edges) and form pairs. Many data can be represented in graphs.
Graphs allow formalizing complex research problems. This allows to take into account the complexity of interactions between heterogeneous data.
For example, interactions between different drugs, between different proteins or between drugs and proteins. Taking this complexity into account should allow researchers to make better predictions.
GNN : Encoding then predictions
Graph neural networks (GNN) consist in encoding the network in the form of vectors and then using this encoding to make predictions.
1/ Encoding is based on the following principle: encode nodes with a similar vector if the neighborhood of the nodes in the graph is similar.
2/ GNN is making four types of prediction : predict the property of a node, predict the link between two nodes, predict the whole property of a graph, predict the similarity between two nodes or two graphs.
Optimizing the performance of GNNs
Graph neural networks (GNN) are as “hype” as Transformers. Researchers are trying to optimize the training and performance of these models. They divide graphs into sub-graphs (Sub-GNN), train networks with ill-defined links between nodes (G-Meta), measure the impact of small changes in the graph on predictions and model performance (GNNGuard).
GNN and biomedical research
Biomedical research aims at a better understanding of diseases: to better detect them (tests, diagnostics) and to better cure them (drugs, medical devices). Most biomedical data can be represented as networks.
For Marinka Zitnik, Professor of Biomedical Informatics at Harvard Medical School, convolutional neural networks and recurrent neural networks (RNNs) have enabled advances in computer vision, automatic language processing (NLP), speech recognition and robotics.
However, the modern deep learning toolbox is designed for simple sequences and grids, not for analyzing complex interactions represented by a graph.
At the end of December 2020, it was GNNs that allowed DeepMind to solve with its AlphaFold 2 algorithm, one of the most important problems in biology, 50 years old: protein folding.
GNN to prevent drug interactions
Graphical neural networks are used by researchers to discover optimal drug combinations, to make drugs more effective, to evaluate drug use in new pathologies. The search for interactions between drugs is crucial because multiple drug use can be more harmful than beneficial for many patients.
Marinka Zitnik, Monica Agrawal, and Jure Leskovec developed Decagon, an approach for modeling the side effects of polypharmacy.
The approach is based on a multimodal graph of protein-protein, drug-target protein interactions and polypharmacy side effects (represented as drug-drug interactions).
The nodes are drugs or proteins and the edges are interactions. The GNN must be able to predict the side effects on muscle of the combination of two drugs : Simvastatin and Ciprofloxacin.
Decagon achieves excellent accuracy in the task of predicting the side effects of polypharmacy. For example out of the ten side effects predicted bellow by Decagon between the drugs below; 5 have been verified by medical literature.
The researchers used the t-SNE package (Maaten and Hinton, 2008) to map the side effects predicted by Decagon. The three main side effects that often occur with uterine polyp are: uterine bleeding, breast dysplasia, and postmenopausal bleeding. Decagon is able to deduce clusters of side effects by itself (different colors on this graph).
GNN to discover new antibiotics
A team of MIT researchers has developed a model that can predict molecular properties directly from the graphical structure of the molecule, where atoms are represented as nodes and bonds between atoms as edges.
These researchers performed predictions on several chemical libraries and discovered a molecule — halicin — that structurally diverges from conventional antibiotics and exhibits activity against a broad spectrum of pathogens.
Appendix
Learn more about Graph Neural Networks in the field of drug development
Neural-relational Learning and some Biomedical Applications
This video gives an overview of the research on graph-based machine learning conducted at NEC Labs Europe. Biomedical applications include cancer vaccine development, variant calling and drug side effect prediction.
Learn more about GNN
