Graphs describe objects as nodes and their connections as edges. Many graph neural networks update a node's internal features by combining information from connected nodes and edges.
The learned features support predictions about nodes, connections or an entire graph. In molecular modeling, atoms can be nodes and chemical bonds edges, allowing structure to inform a prediction.
ELI5
Imagine learning about each person by looking at both their own information and the people connected to them.
For example, a model could examine atoms and their bonds to estimate a property of a molecule.
What does a graph neural network learn from?
It learns patterns in object features and connections represented by a graph.
Does it guarantee an accurate molecular prediction?
No. Predictions still depend on suitable training data and testing on relevant examples.
