Convolutional Networks on Graphs for Learning Molecular Fingerprints

This paper introduces a convolutional neural network that operates directly on graphs, allowing end-to-end learning of prediction pipelines whose inputs are graphs of arbitrary size and shape. The architecture is designed to generalize standard molecular feature extraction methods based on circular fingerprints, replacing the fixed, hand-engineered fingerprinting procedure with a differentiable, learnable network that can be trained jointly with a downstream predictor on molecular data.

The authors demonstrate that the resulting data-driven features are more interpretable than conventional circular fingerprints and deliver better predictive performance across a variety of tasks. This mattered because it showed how learned, graph-based representations could improve on established cheminformatics feature extraction, an early and influential step toward graph neural networks for molecular property prediction.