Accelerate Drug Discovery with Multitask Graph Neural Networks
, Genentech
, Amazon Web Services
We'll present a successful use case of applying graph neural networks (GNNs) to molecular property prediction. GNNs generalize molecular fingerprints with task-specific, data-driven representations. For this project, we utilized DGL-LifeSci, an open source package for GNN-based modeling that doesn't require background knowledge. Experiments on real-world datasets from Genentech show that multitask learning using DGL-LifeSci consistently outperforms traditional single-task approaches. The GNN-based approaches also show better ability to extrapolate to uncharted chemical space and allow for efficient transfer learning in real-world settings. We assume that you have a basic background in machine learning and cheminformatics.