Graph Neural Networks for Retrosynthesis

How we're using message-passing neural networks to predict synthetic routes for complex molecules.

Retrosynthesis — working backwards from a target molecule to find viable starting materials — is one of the hardest problems in computational chemistry.

The Approach

We model chemical reactions as graph transformations. Atoms are nodes, bonds are edges. A reaction is a change in that graph.

Results

On the USPTO-50k benchmark, our model achieves top-1 accuracy of 54.6%, competitive with state-of-the-art approaches while being 3x faster at inference.