Skip to content

Single-Step Models

Syntheseus currently supports 11 single-step models: RetroChimera (which we recommend as default), its two submodels, and 8 additional external models.

For convenience, for each model we include a default checkpoint (trained on Pistachio in the case of RetroChimera, and on USPTO-50K otherwise). If no checkpoint directory is provided during model loading, syntheseus will automatically download a default checkpoint and cache it on disk for future use. The default path for the cache is $HOME/.cache/torch/syntheseus, but it can be overriden by setting the SYNTHESEUS_CACHE_DIR environment variable. See tables below for the links to the default checkpoints.

Model checkpoint link Source
RetroChimera trained by us
RetroChimeraEdit trained by us
RetroChimeraDeNovo trained by us

Other models (trained on USPTO-50K)

Model checkpoint link Source
Chemformer finetuned by us starting from checkpoint released by authors
GLN released by authors
Graph2Edits released by authors
LocalRetro trained by us
MEGAN trained by us
MHNreact trained by us
RetroKNN trained by us
RootAligned released by authors
Choice of dataset

The USPTO-50K dataset is well-established but relatively small. Advanced users may prefer to either use our Pistachio-trained models, or retrain any model class of interest on their own data. To do that, please follow the instructions in the original model repositories.

In reaction_prediction/cli/eval.py a forward model can be used for computing back-translation (round-trip) accuracy. See here for a Chemformer checkpoint finetuned for forward prediction on USPTO-50K. As for the backward direction, pretrained weights released by original authors were used as a starting point.

Licenses

All checkpoints were produced in a way that involved external model repositories, hence may be affected by the exact license each model was released with. For more details about a particular model see the top of the corresponding model wrapper file in reaction_prediction/inference/.