In “Neural Message Passing for Quantum Chemistry,” submitted to arXiv on April 4, 2017, Justin Gilmer, Samuel Schoenholz, Patrick Riley, Oriol Vinyals, and George Dahl proposed a common framework, Message Passing Neural Networks (MPNNs), that subsumes many previously separate graph neural network designs. They showed that graph convolutions, gated graph networks, and several other models could all be described as variations of the same two-step process.
In the MPNN view, each round of computation has a message phase and an update phase. During the message phase, every node gathers messages from its neighbors, computed by a learned message function that can also depend on edge features. During the update phase, each node combines the incoming messages with its current state to produce a new state. After several rounds, a readout function aggregates all node states into a prediction for the whole graph.
The authors applied this to predicting quantum-mechanical molecular properties on a standard chemistry benchmark, where a molecule is naturally a graph of atoms and bonds. Their best MPNN variant achieved state-of-the-art accuracy, demonstrating that the framework was not just a tidy abstraction but a strong practical model.
The lasting value of the paper is conceptual: by showing that so many graph models share one message-passing skeleton, it gave researchers a single vocabulary and a clear design space. Much of modern graph learning, including its use in drug discovery and materials science, is built on this framing.