Graph Neural Networks

Jure Leskovec's graduate course on graph ML. Slides and recorded lectures freely available. Covers GCN, GAT, GraphSAGE, knowledge graphs, and scalable GNNs.

Interactive article with animated visualisations of message passing, graph-level tasks, and GNN expressiveness. Essential before touching code.

Companion article explaining spectral vs. spatial graph convolutions and the relationship to classical signal processing.

Official Colab notebooks for node classification, link prediction, and graph classification using built-in datasets including QM9 and ZINC.

Reference implementations for molecular property prediction on ogbg-molhiv, ogbg-molpcba, and ogbg-moltox21.

Introduces the MPNN framework unifying prior GNN architectures. Applied to QM9 property prediction. The foundational paper for molecular GNNs.

Introduces edge-level attention coefficients, allowing the network to learn which neighbours matter most.

Foundational GCN paper. Simplifies spectral graph convolution to a localised first-order approximation.

Systematic study of node-level and graph-level GNN pre-training strategies for molecular and biological graphs.

Standard library for graph deep learning in PyTorch. Includes built-in molecular datasets (QM9, ZINC, OGB) and 60+ GNN implementations.

Standardised large-scale graph benchmarks for molecular property prediction with leaderboards.

Alternative to PyG with a focus on flexibility and large-scale graphs. Has its own molecular graph dataset collection.

134k small organic molecules with 19 DFT-computed properties. The standard benchmark for molecular GNNs — available directly via PyTorch Geometric.

Standardised molecular property prediction benchmarks (ogbg-molhiv, ogbg-molpcba) with scaffold splits and public leaderboards.

250k drug-like molecules from ZINC, widely used for graph generation benchmarks. Available via the chemical VAE repository.