- Ottawa, Ontario
Plotting solutions to real‑world problems on graphs unlocks doors to new frontiers
If you're tracking your family tree on Ancestry.ca, your ancestors appear on nodes in graphs with the edges showing connections among them. Looking behind the screens of Facebook or LinkedIn, you'd see similar graphs that plot your networks and interactions.
Graph neural networks (GNNs) such as these are ideal for organizing and connecting vast amounts of data on social media. When deep learning methods are applied to graph edges, they extract knowledge and make predictions, track your community activity and generate recommendations. For example, Pinterest improved the performance of its recommendation system by 150% by using GNNs.
They can also serve as powerful probes into uncharted territory, opening doors to new material discovery, advanced circuit design and novel drug invention among others. By developing new graph designs, structures and properties, and using artificial intelligence (AI) and machine learning (ML) to mine information from large data sets, researchers have found immense potential for solving real‑world problems. And in the race to find a COVID-19 vaccine, this has enabled them to discover new molecules for drug development.
"Using AI graph illustrations can improve our ability to quickly predict which molecular properties among billions can be used to develop solutions for specific challenges," says Harry Guo, Research Officer at the National Research Council of Canada's (NRC) Digital Technologies Research Centre. "The GNN project, which is part of the NRC's AI for Design Challenge program, is creating methods to rapidly search and identify graph designs with the most potential for science and engineering design."
The project's state‑of‑the‑art predictive performance was recognized recently in an international MIT‑hosted competition related to AI for Design and COVID-19. The NRC's paper on the topic, prepared with collaborator Montréal, Quebec‑based Mila research institute in artificial intelligence, was also accepted at the 2020 International Conference on Machine Learning. This success has garnered attention that will soon lead to other projects in this cutting‑edge field.
Channelling the power of AI
In pandemics such as COVID-19 that blindside the world and come with no treatments or cures as of yet, AI is particularly important in speeding up the drug‑discovery process since most of that data is graph‑structured. Traditionally a long, expensive activity, drug development includes not only complex research and testing, but also lengthy approval procedures by authorities such as Health Canada and the U.S. Food and Drug Administration.
According to Jian Tang, Assistant Professor at Mila and HEC Montréal, developing a new drug takes more than 10 years on average and costs about US$2.5 billion. "AI is a huge opportunity to accelerate the process," he says, estimating that within the next 5 years, harnessing the power of AI could reduce the time from 10 years to 1.
A collaboration between the NRC and Mila to develop general ML techniques for GNNs has ensured their place as major players in a global arena with renowned competitors such as MIT and Stanford University. Tang points out that the "natural collaboration" between the NRC and Mila blends expertise in ML and AI for design, drug discovery and deep learning that will push the envelope for years to come.
The methods developed in the GNN project can also be applied in the short term to find COVID-19 solutions by searching datasets of existing drugs with a view to repurposing them. This will identify medications that show high anti‑viral potential and provide material for laboratory testing of the most promising molecules. "These methods would also be invaluable during future pandemics to get therapies and vaccines into place faster, saving lives and mitigating the impacts on our health and economy," adds Tang.
Beyond the current pandemic
According to Kevin Thomson, Program Director for the NRC's AI for Design program, the GNN project is part of a broader, longer‑term NRC thrust to develop AI tools that "accelerate and advance Canada's capacity to design scientific and engineering innovations." This can be anything from better communications networks that bring high‑speed internet access to all corners of the country, to accelerated discovery of new materials that combat global warming.
"AI for Design is one in a series of challenge programs that aim to address research objectives of national importance through collaborative projects with leading academic and small and mediu‑sized organizations," he says. "And such collaborations can only strengthen the technology ecosystem for the benefit of all Canadians."
Media Relations, National Research Council of Canada
1-855-282-1637 (toll-free in Canada only)
1-613-991-1431 (elsewhere in North America)
Follow us on Twitter: @NRC_CNRC