## TL;DR
Graph Neural Networks extend deep learning to graph-structured data — molecules, social networks, knowledge graphs. Message passing enables nodes to learn from their local neighborhood, creating representations that capture both structure and features.

## Core Explanation
Message passing: each node receives messages from neighbors, aggregates them (mean, sum, max, attention), transforms via neural network, and updates. Stacking L layers gives each node information from its L-hop neighborhood. Key architectures: GCN (graph convolution), GAT (graph attention), GraphSAGE (sampling for large graphs).

## Detailed Analysis
Applications: molecular property prediction (GNN scores candidates for drug discovery), recommendation (PinSage powers Pinterest recommendations at billion-node scale), traffic prediction (road network as graph), and protein interaction prediction (AlphaFold uses GNN-like operations).

## Further Reading
- Stanford CS224W: Machine Learning with Graphs
- PyTorch Geometric Library
- OGB: Open Graph Benchmark