Multi-Step Retrosynthesis: Predict a full route to purchasable materials

Unlike the single-step page, this tool proposes a complete synthetic route: the target is disconnected recursively until every leaf is (ideally) a purchasable compound, checked against a 23-million-compound stock (eMolecules + buyables). Input your target molecule by drawing its chemical structure using the drawing tool (Ketcher). Then click on Get SMILES from drawing. The computed routes will be rendered by RDKit.js directly from the model's raw SMILES, shown beneath each structure.

checking server…
Model: Pistachio dataset fine-tuned base-model Explorer XL
See base model by the Batista group (Yale): Shee, Morgunov, Li, Batista, J. Chem. Inf. Model. 2025, 65(8), 3903–3914, doi:10.1021/acs.jcim.4c01982. MIT-licensed. This web interface (Ketcher input, RDKit.js rendering) built independently.
The served checkpoint is a fine-tune of Explorer XL, trained here on 2.2 million multi-step routes extracted from the Pistachio patent reaction set. On 200 held-out 2026 patent targets it solves 68.5% to purchasable materials within its 5 suggested routes, versus 32.5% for the published checkpoint — the largest gains are on routes of five or more steps.