Reaction Models learns a property of a reaction rather than of a molecule: a yield, an enantiomeric or diastereomeric ratio, an activation barrier or a rate constant. Each reaction is read as a condensed graph of reaction, which combines the reactants and products atom by atom, together with the molecules that change between experiments, such as the ligand, base, additive or solvent.
Yield models work best on consistent in-house data, such as a high-throughput experimentation plate run under one protocol. Yields pooled from the literature mix different scales, purities and work-ups, and give weak models.
Inputs
A table with one atom-mapped reaction per row and its measured value. Map the reactions first with the Reaction Atom Mapper: its mapped table keeps your other columns (yield, ligand, plate position), so it can be used for training as it is. Rows that are not atom-mapped are skipped, rows the mapper rejected are dropped, and low-confidence mappings are excluded unless you keep them. Up to 10,000 reactions.
| Column | Role |
|---|---|
| Reaction | Atom-mapped reaction SMILES (mapped_reaction from the mapper) |
| Target | The measured value: yield, ee, a barrier, log k (or a class) |
| Conditions | Up to six columns of molecules that vary between reactions (SMILES; an empty cell means none) |
| Group (optional) | Keeps every reaction with the same value in one fold, for a leave-one-ligand-out test |
Molecules on the left of the reaction that contribute no atoms to the product (reagents written into the reaction) are left out of the graph; give them as condition columns instead.
How the model is checked
The model is trained like MPNN Models: a validation set picks the saved checkpoint, and a held-out test set is scored against a mean and a nearest-neighbour baseline. Splits can be random, by group (the honest test for a plate: hold out a ligand or an additive the model never saw), by reaction template or by product scaffold. Ensembles add an uncertainty to every prediction, regression models carry a 90% prediction interval, and each prediction reports how similar the reaction and its conditions are to the training data.
Set target bounds (for example 0 and 100 for a yield in percent) to keep predictions and intervals inside the possible range.
Predictions
Predict up to 1,000 reactions at a time on the tool page or through the API, giving the same condition columns the model was trained with. Score larger sets, or reactions enumerated from a template, in a workflow with Saved Model Scoring: the workflow templates Train a reaction yield model from your reactions and Enumerate products and keep the reactions your model predicts best chain the mapper, training and scoring.
Outputs
| File | Contents |
|---|---|
model/ | The saved model with its model card: test metrics, baselines, the split, the conditions it reads and the mapping counts (no pickled code) |
training_summary.json | Reactions used and skipped |
The saved model appears in My Models.
Related tools
Map and check reactions with the Reaction Atom Mapper, enumerate products with Reaction Enumeration, or model molecules instead with MPNN Models and QSAR Models.