Graph Neural Networks

Recording

Learning Objectives

• Represent a molecule as a graph with atom and bond features and build it in PyTorch Geometric
• Implement one step of message passing (aggregate → update) by hand
• Distinguish GCN, GAT, and MPNN architectures and identify their trade-offs
• Explain why GNNs are permutation-invariant and why this matters for molecular inputs

Key Takeaways

• Takeaway 1. Message passing is the unifying abstraction: each atom aggregates information from its neighbours, updates its own state, and repeats for K layers — effectively encoding a K-hop chemical neighbourhood.

• Takeaway 2. Graph Attention Networks (GAT) learn which neighbours matter most, making them more expressive than GCN for heterogeneous molecules where not all bonds are equally informative.

• Takeaway 3. GNNs are not automatically better than fingerprints on small datasets. With fewer than ~1000 molecules, ECFP + gradient boosting frequently outperforms an MPNN due to overfitting risk.

• Takeaway 4. The readout function (sum, mean, or a learned aggregation over atom embeddings) is as important as the message-passing architecture; it determines what global properties the model can capture.

Further Reading & Resources