Key Findings
The OrgAIcat framework demonstrates high predictive accuracy for enantioselectivity in two major reaction classes. Using an ExtraTrees regression model, the team achieved mean absolute errors in free energy difference (ΔΔG) of 0.282 kcal/mol for Aldol reactions and 0.265 kcal/mol for Michael additions. These figures correspond to mean absolute errors in enantiomeric excess (ee) of approximately 7.30% and 6.78%, respectively. Additionally, a four-category classification model for selectivity intervals reached AUC values of 0.908 and 0.922.
Technical Details
The model is powered by iSynth, a dataset extracted from 844 papers published between 2000 and 2020. After rigorous cleaning, the final set included 13,360 Aldol reactions and 9,194 Michael addition reactions. To represent these reactions, the team developed the R-SPOC descriptor, which integrates three data streams:
- Catalyst structure and physicochemical characteristics.
- Structural changes (the difference between product and reactant descriptors).
- Reaction conditions, including temperature, solvent, and additives.
R-SPOC eliminates the need for quantum chemical calculations or explicit transition state searches, reducing computational costs.
Background and Context
The researchers identified a significant “selection bias” in existing chemistry literature. While the potential reaction space for Aldol and Michael additions is estimated at 1016 and 1017 combinations respectively, published literature only records approximately 104, heavily skewed toward high-performing reactions. To ensure the model learned structure-selectivity relationships rather than mere popularity, the team performed low-similarity tests and downsampled high-frequency components.
Significance and Outlook
OrgAIcat’s utility was validated through external tests on papers published between 2021 and 2024 and laboratory experiments. In a test involving 24 newly synthesized chiral primary amine catalysts, the model tracked selectivity differences with a Pearson correlation coefficient of 0.736. This framework transforms imperfect historical literature into a predictive tool for the production of fine chemicals and chiral drugs.
Get our weekly technology intelligence — free
Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.
Subscribe Free — Weekly Tech Intelligence
By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.
- Your email and selected fields are used only to deliver the newsletter.
- We never share your information with third parties.
- You can unsubscribe anytime via the link in each email.
See our Privacy Policy for details.
Takes about a minute · Unsubscribe anytime
