Graph Representations

molfp.graph

Molecular graph representations for deep learning.

Converts molecules to atom/bond feature matrices and adjacency structures suitable for graph neural networks (GCN, GAT, MPNN, SchNet, etc.).

Framework-agnostic core (NumPy only), with optional converters for PyTorch Geometric and DGL.

GraphTransformer

Bases: BaseEstimator, TransformerMixin

sklearn-compatible transformer for molecular graph features.

Parameters:

Name	Type	Description	Default
enriched		if True, use enriched atom features with Morgan hashes, functional group membership, and neighbor counts.	False
type		graph type — "standard", "distance", "line", "fragment".	'standard'

For feature selection, use standalone functions after transforming: X = molfp.GraphTransformer(enriched=True).fit_transform(smiles) result = molfp.select_features(X, y, k=500, threshold=0.01)

mol_graph(input_data, *, column=None, enriched=False, type='standard')

Convert molecule(s) to graph representation(s) for deep learning.

Parameters:

Name	Type	Description	Default
input_data		molecule input (str, Mol, list, Series, DataFrame).	required
column		column name for DataFrame input.	None
enriched		if True, add Morgan hashes, functional group membership, and neighbor counts per atom (432 dims vs 20 base dims). Only applies to type="standard" and type="distance".	False
type		graph type: "standard" — atoms=nodes, bonds=edges (GCN, GAT, GIN) "distance" — atoms=nodes, all pairs=edges with 3D distance (SchNet, DimeNet) "line" — bonds=nodes, bond-bond connections (D-MPNN, Chemprop) "fragment" — functional groups=nodes (hierarchical GNNs)	'standard'

Returns:

Type	Description
	Single molecule: dict with atom_features, edge_index, edge_features.
	Multiple molecules: list of dicts.

to_pyg(graph_dict, y=None)

Convert a graph dict to a PyTorch Geometric Data object.

Requires: pip install torch torch_geometric

Parameters:

Name	Type	Description	Default
graph_dict		output of mol_graph() for a single molecule.	required
y		optional label (int, float, or tensor).	None

Returns:

Type	Description
	torch_geometric.data.Data

graphs_to_pyg(graph_list, y=None, batch_size=32, shuffle=False)

Convert a list of graph dicts to a PyTorch Geometric DataLoader.

Parameters:

Name	Type	Description	Default
graph_list		list of graph dicts from mol_graph().	required
y		optional list/array of labels (same length as graph_list).	None
batch_size		batch size for DataLoader.	32
shuffle		whether to shuffle the DataLoader.	False

Returns:

Type	Description
	torch_geometric.loader.DataLoader

to_dgl(graph_dict)

Convert a graph dict to a DGL graph.

Requires: pip install dgl torch

Parameters:

Name	Type	Description	Default
graph_dict		output of mol_graph() for a single molecule.	required

Returns:

Type	Description
	dgl.DGLGraph with ndata['feat'] and edata['feat']

atom_feature_names()

Return ordered list of atom feature names.

edge_feature_names()

Return ordered list of edge feature names.

graph_to_vector(graph_dict)

Convert a graph dict to a fixed-length feature vector.

Uses chemistry-aware aggregation to preserve more structural information than simple mean/max pooling. Designed for sklearn classifiers. Adapts to both base (20-dim) and enriched (432-dim) atom features.

graph_vectors(input_data, *, column=None, enriched=False, type='standard')

Convert molecule(s) to fixed-length graph feature vectors for sklearn.

Parameters:

Name	Type	Description	Default
input_data		SMILES string, list, Series, or DataFrame.	required
column		column name for DataFrame input.	None
enriched		if True, use enriched atom features.	False
type		graph type — "standard", "distance", "line", "fragment".	'standard'

Returns:

Type	Description
	Single molecule: numpy array
	Multiple molecules: numpy array (n_samples, n_features)