Chapter 7 of Cheminformatics

Graph neural networks for molecules soon

Molecules as graphs: message passing, learned representations, and when GNNs beat fingerprints (and when they do not).

soon This chapter is planned and not yet written. The outline below is the current plan; it may change.

What it will cover

• Molecules as graphs: atoms as nodes, bonds as edges, featurisation
• Message passing and readout, step by step
• A minimal GNN in PyTorch (Geometric)
• Training on MoleculeNet tasks with proper splits
• GNN versus fingerprint baselines: a fair comparison
• Lab 7: train a GNN for solubility and inspect its learned embeddings

Planned learning outcomes

1. Featurise a molecule as a graph and explain message passing in your own words.
2. Train and evaluate a graph neural network on a molecular property task.
3. Decide, with evidence, whether a GNN is worth it over a fingerprint model for a given dataset size.

In the meantime

Everything here builds on Part I, in particular the evaluation habits of Chapter 2. Episode 2.6 already introduces the group splits that become scaffold splits here.

PreviousChapter 6 QSAR and property predictionNextChapter 8 Structure-based design

Edit this pageReport a problemLast updated 2026-11-02