Molecular Graph

A molecular graph represents atoms as nodes and bonds as edges. It is the natural input for graph neural networks and many molecular property models.

A molecule can be represented as:

G = (V, E, X_{V}, X_{E})

where $V$ is the set of atoms, $E$ is the set of bonds, $X_{V}$ are atom features, and $X_{E}$ are bond features.

Message passing then updates atom states:

h_{i}^{(t + 1)} = ϕ_{u} h_{i}^{(t)}, j \in N (i) \sum ϕ_{m} (h_{i}^{(t)}, h_{j}^{(t)}, e_{ij})

Checks

Are hydrogens explicit or implicit?
Are bond order, aromaticity, chirality, formal charge, and ring membership encoded?
Does graph batching isolate molecules correctly?
Does the graph lose conformer or stereochemical information needed by the task?
Are graph features generated from the standardized molecule definition?

Common atom features include:

Common bond features include:

The feature set is part of the model contract:

ϕ (M) = (X_{V}, X_{E}, A)

where $A$ is the adjacency or edge index. If $ϕ$ changes, cached features and splits may need regeneration.

For graph mini-batches, node tensors from many molecules are concatenated. A molecule-level readout must respect the batch index:

h_{G}^{(b)} = pool {h_{i} : batch (i) = b}

Pooling across all nodes without the batch boundary leaks information between examples.

Chiral tags or bond stereo are dropped while the label depends on stereochemistry.
Formal charge and hydrogens are inconsistent with the protonation protocol.
Global attention or normalization mixes nodes across molecules in a batch.
A graph-only model is evaluated on a task requiring 3D conformer information.
Features are generated from raw input while labels and splits use standardized identity.