Graph Neural Networks in Microbiology

One-Sentence Definition

Graph neural networks (GNNs) learn on nodes and edges — natural for molecules, protein contact graphs, assembly graphs, and microbial interaction/transmission networks.

Simple Explanation

When your data is a network (atoms bonded, contigs linked, patients connected), GNNs pass messages along those links instead of flattening everything into a spreadsheet.

Detailed Scientific Explanation

Graph typeNodes / edgesMicro task
Molecular graphAtoms / bondsAntibiotic activity, toxicity (AI for Antibiotic Discovery)
Protein graphsResidues / contactsFunction, binding, stability
Assembly / DBGUnitigs / overlapsStrain resolution, plasmid links
Bipartite host–geneGenomes ↔ genesPlasmid Host Attribution with ML style problems
Contact / mobilityPatients/wards / transfersOutbreak risk scoring (careful with privacy)

How to use

  1. Define biologically meaningful edges (don’t invent dense noise graphs)
  2. Choose task: node label, graph label (molecule active?), link prediction
  3. Baselines: fingerprints + Tree Ensembles in Microbiology for molecules — GNN must beat them
  4. Wet-lab validate chemical hits; epi-validate network scores

Architectures you will see: GCN, GraphSAGE, GAT, MPNN; geometric networks for 3D conformers.

Mechanism

Iterative neighborhood aggregation updates node embeddings; readout pools them for graph-level prediction.

Clinical Importance

  • Discovery and research-facing more than routine bench diagnostics today

Research Importance

Diagnostic Relevance

  • Limited direct use; assembly-graph ML is still specialized bioinformatics

AMR Relevance

  • Molecule GNNs for new agents; bipartite models for resistome→host inference in metagenomes

Active Recall Questions

  1. Why are molecules naturally GNN problems?
  2. What baseline should a molecular GNN beat?
  3. Name one non-chemistry graph in microbiology.

Connections