Molecular Property Prediction: Predicting Chemical Properties from Molecular Graphs

Molecular property prediction uses graph neural networks to predict chemical and pharmacological properties from molecular structure. Atoms become nodes, chemical bonds become edges, and the GNN learns to map molecular topology (and optionally 3D geometry) to properties like toxicity, solubility, binding affinity, and quantum-mechanical energies. This is one of the highest-impact applications of graph ML, accelerating drug discovery from years-long experimental cycles to weeks of computational screening.

Molecular graph representation

Each molecule is converted to a graph with rich features:

from torch_geometric.data import Data
from rdkit import Chem

def mol_to_graph(smiles):
    """Convert SMILES string to PyG graph."""
    mol = Chem.MolFromSmiles(smiles)

    # Node features (per atom)
    node_features = []
    for atom in mol.GetAtoms():
        node_features.append([
            atom.GetAtomicNum(),        # element (C=6, N=7, O=8)
            atom.GetFormalCharge(),      # charge
            atom.GetNumImplicitHs(),    # hydrogen count
            atom.GetIsAromatic(),       # aromaticity
            atom.GetHybridization(),    # sp, sp2, sp3
        ])

    # Edge features (per bond)
    edges, edge_features = [], []
    for bond in mol.GetBonds():
        i, j = bond.GetBeginAtomIdx(), bond.GetEndAtomIdx()
        edges.extend([[i, j], [j, i]])  # undirected
        feat = [bond.GetBondTypeAsDouble(), bond.GetIsAromatic()]
        edge_features.extend([feat, feat])

    return Data(x=tensor(node_features), edge_index=tensor(edges).T,
                edge_attr=tensor(edge_features))

2D vs 3D molecular properties

2D graph properties

Many pharmacological properties depend primarily on molecular topology (which atoms are connected to which) rather than 3D geometry:

• Toxicity: Presence of toxic substructures (functional groups)
• Solubility: Balance of hydrophilic and hydrophobic groups
• Drug-likeness: Molecular weight, ring count, hydrogen bond donors/acceptors

Standard GNNs (GCN, GAT, GIN) work well for these properties. GIN is particularly popular because its sum aggregation is maximally expressive for graph-level classification.

3D geometric properties

Quantum-mechanical properties depend on exact 3D atom positions:

• Energy: Total electronic energy of the molecule
• Forces: Forces on each atom (for molecular dynamics)
• Electron density: Spatial distribution of electrons

These require equivariant GNNs (SchNet, PaiNN, MACE) that use 3D coordinates and respect rotational symmetry.

Graph-level prediction

Molecular property prediction is a graph-level task: one prediction per molecule, not per atom. This requires a readout (pooling) step that aggregates all node representations into a single molecular representation:

• Sum pooling: Preserves molecular size information (larger molecules have larger sums)
• Mean pooling: Size-invariant, suitable for intensive properties
• Set2Set / attention pooling: Learned, weighted aggregation that identifies the most important atoms

Key benchmarks

• QM9: 134K molecules, 12 quantum properties. The standard benchmark for 3D molecular GNNs.
• MoleculeNet: Suite of datasets for drug discovery properties (HIV inhibition, BBBP, Tox21). Standard for 2D molecular GNNs.
• PCQM4Mv2: 3.4M molecules, HOMO-LUMO gap prediction. Large-scale benchmark from OGB-LSC.
• MD17: Molecular dynamics trajectories for energy and force prediction. Standard for equivariant models.