Contact

Graph-Convolutional Neural Network (GNN)

A Graph-Convolutional Neural Network (GCNN) treats a molecule as what it structurally is: a graph, where atoms are nodes and chemical bonds are edges. Instead of relying on a fixed set of hand-picked descriptors, the network learns directly from this graph which structural patterns are relevant to the property being predicted, aggregating information from each atom’s neighbours layer by layer.

This lets a GCNN capture subtle structural effects — the kind that simpler, descriptor-based models can miss — without anyone having to anticipate and encode those effects in advance. The trade-off is data: graph-based models typically need a larger, well-curated training dataset to learn reliable patterns, which is why PAULY’s GCNN models are trained on datasets ranging from several hundred to tens of thousands of curated chemicals, depending on the property.

Models using this approach