Graph Neural Networks

Graph neural networks operate on nodes, edges, and message passing. In molecular modeling, atoms, residues, bonds, contacts, and interactions can all be represented as graph structure.

A common message-passing update is:

$$m_i^{(t)}=\sum _{j\in \mathcal{N}(i)}{\varphi }_m(h_i^{(t)},h_j^{(t)},e_{ij})$$ $$h_i^{(t+1)}={\varphi }_u(h_i^{(t)},m_i^{(t)})$$

Here $h_i$ is a node state, $e_{ij}$ is an edge feature, and $\mathcal{N}(i)$ is the neighborhood of node $i$.

For a graph-level prediction, a readout should be permutation invariant:

$$h_G=\rho (\{h_i^{(T)}:i\in V\})$$

For node-level outputs, a GNN is usually expected to be permutation equivariant. If $P$ permutes node order, then:

$$f(PX,PA{P}^{\top })=Pf(X,A)$$

Message Passing Contract

Component	Required Detail
node unit	atom, residue, token, pocket point, conformer, protein, ligand, or abstract entity
edge unit	bond, contact, distance edge, sequence adjacency, interaction edge, learned relation
edge attributes	bond type, distance, direction, residue separation, pair embedding, interaction label
update target	node state, edge state, graph state, coordinate, vector field, score, or ranking
aggregation	sum, mean, max, attention, normalized message, or degree-scaled message
readout	node-level, edge-level, graph-level, pair-level, or complex-level output
depth	number of message passing layers and effective receptive field

The output type changes what the GNN claim means:

Output	Example Task	Required Symmetry
node label	residue site, atom type, node mask	permutation equivariance
edge label	contact, bond, interaction	permutation equivariance over node pairs
graph scalar	molecular property, affinity, class	permutation invariance
pair score	molecule-target score, protein-ligand interaction	invariance within each entity and pair identity
coordinates or vectors	pose, force, velocity, update	geometric equivariance in addition to permutation behavior

Key Ideas

• Nodes hold entity features; edges hold relationships such as bonds, distances, contacts, or interaction types.
• Message passing repeatedly updates node or edge states from local neighborhoods.
• Readout functions pool graph information for graph-level tasks such as property prediction.
• Graph construction is a modeling choice: chemical bonds, k-nearest neighbors, radius graphs, contact maps, or learned edges change the problem.
• Weisfeiler-Lehman Test gives a reference point for the expressive power of standard message-passing GNNs.
• Geometry-aware variants connect message passing to equivariance and coordinate updates.

Common Failure Modes

Failure	Symptom	Check
oversmoothing	node embeddings become too similar after many layers	layer depth, residuals, normalization, pairwise embedding variance
oversquashing	long-range information is compressed through narrow neighborhoods	graph diameter, bottleneck edges, attention/global edges
graph leakage	edges use labels, future data, full benchmark, known pose, or test statistics	graph construction provenance
readout mismatch	graph-level pooling hides rare active sites or interaction regions	task-specific pooling and attention audit
split mismatch	train/test share scaffold, homolog, complex pair, or graph-derived artifact	split unit and duplicate controls
geometry mismatch	scalar GNN is used for coordinate/vector target without equivariance	target transformation rule

Paper Reading Pattern

Claim	First Evidence
better GNN architecture	same graph construction, features, split, optimizer budget, and readout
better graph representation	ablation over nodes, edges, pair features, and construction rule
better long-range reasoning	tasks requiring nonlocal signal, not only more parameters
better molecular performance	scaffold/source split and feature baseline
better protein/complex performance	protein-family or complex split plus leakage audit

Practical Checks

• Check what the graph nodes represent: atoms, residues, ligands, pockets, conformers, or abstract states.
• Check whether edge features include distances, directions, bond orders, residue separation, or learned relation types.
• Watch oversmoothing when many message-passing layers make node states too similar.
• For molecular tasks, inspect split strategy for leakage across scaffolds, protein families, or near-duplicate complexes.
• Verify that graph construction is part of the method or held fixed across baselines.
• Match readout to the output unit before comparing metrics.

Related

• Graph construction
• Weisfeiler-Lehman Test
• Graph Transformer
• Equivariant GNN
• Pooling and readout
• Invariant feature
• Equivariant feature
• Scaffold split
• Protein family split
• How Powerful are Graph Neural Networks?
• Ligand
• Protein-ligand complex