Emerging interdisciplinary directions
Reaction prediction, machine-designed synthesis
How data-driven models predict chemical transformations and search for practical multistep routes, while exposing uncertainty and chemical constraints.
IntuitionA reaction model proposes, not proves
A model learns patterns linking reactants and conditions to products. A plausible product is a hypothesis: chemists still check atom balance, selectivity, safety, and whether the transformation can actually be run.
SchoolRepresenting a reaction
Definition: Reaction record
A reaction record may encode reactant and product structures, reagents, solvent, temperature, catalyst, yield, and provenance. Missing or inconsistently recorded conditions can make identical chemistry appear contradictory to a model.
Atom-mapped representations assign atom indices across reactants and products. Comparing changed bonds can describe a transformation compactly, but mapping is bookkeeping—not evidence for a mechanism.
UndergraduatePrediction as conditional inference
The probability is conditional on the examples and representation used for training. It is not automatically a probability of experimental success; omitted variables, publication bias, and changed protocols all matter.
Definition: Top-k accuracy
A prediction is counted correct if the recorded product appears among the model’s first k suggestions. This metric rewards useful candidate ranking, but says nothing by itself about yield, purity, or generalization to new chemistry.
AdvancedFrom one-step models to synthesis trees
Retrosynthesis searches backward from a target to simpler precursors. Search algorithms expand candidate disconnections into a tree or graph, then rank routes using learned scores and practical constraints such as purchasability, step count, chemoselectivity, and hazardous reagents.
A route is not a recipe until conditions, work-up, purification, and scale are credible. A short route may be inferior if it relies on unstable intermediates or poorly selective steps.
ResearchResearch frontier
A useful workflow treats AI output as ranked proposals: inspect chemistry, identify uncertainty, select a discriminating experiment, record failures as well as successes, and update the plan. The model supports expert judgment rather than replacing it.
References
- Computer-Assisted Synthetic Planning · E. J. Corey, 1988
- Predicting reaction performance in C–N cross-coupling using machine learning · A. A. A. N. Ahneman et al., 2018
- Predicting reaction outcomes with deep learning · P. Schwaller et al., 2019