CPA Catalyst Selectivity ML Prediction

What this research found

Can a model predict how selectively a chiral catalyst will produce one mirror-image product over the other, and does it still work for catalysts it has never seen? Using the 1,075-reaction benchmark of chiral phosphoric acid catalysed thiol additions from a 2019 Science study, K-Dense trained four regression families on two-dimensional molecular fingerprints alone. A random forest predicted the selectivity energy to within 0.132 kilocalories per mole on a random held-out split, and gradient boosting stayed at 0.189 when eight entire catalysts were withheld from training — both inside the conventional chemical-accuracy threshold of 0.2.

• On a random 80/20 split covering 215 held-out reactions, the random forest reached a mean absolute error of 0.132 kilocalories per mole with an R-squared of 0.915. The median error was 0.084, and 79.5% of reactions fell within 0.2 kilocalories per mole.
• When all 25 reactions of each of eight catalysts were removed from training, histogram gradient boosting reached 0.189 kilocalories per mole with an R-squared of 0.895 across the 200 unseen-catalyst reactions, narrowly ahead of the random forest at 0.195.
• Moving from interpolation to catalyst extrapolation raised the error by 43%, but R-squared fell only from 0.915 to 0.895, indicating the models learned transferable structure-selectivity relationships rather than memorising catalyst identities.
• Tree ensembles beat the linear and kernel baselines in both regimes and degraded far less under extrapolation. Ridge regression went from 0.183 to 0.273 kilocalories per mole and support vector regression from 0.146 to 0.324, while both tree models stayed below 0.2.
• Per-catalyst errors on the holdout varied widely: 0.088 for a bis(trifluoromethyl)phenyl catalyst and 0.107 for a para-methoxyphenyl one, against 0.369 for a catalyst carrying a sterically distinctive biphenyl substituent with no close analogue in the training set. Two-dimensional connectivity alone sufficed, without the conformer ensembles or quantum-chemical descriptors of the original study.

How it was done

The complete factorial matrix of five acylimines, five thiols, and 43 BINOL-derived chiral phosphoric acid catalysts was reconstructed, and every structure was parsed and validated. Measured enantiomeric ratios were converted to the differential activation free energy between the two competing transition states, a quantity that ranges from -0.419 to +3.135 kilocalories per mole across the dataset and is preferable to enantiomeric excess because it is additive and unbounded. Each reaction was featurized by concatenating, for all three components, a 1,024-bit Morgan fingerprint of radius 2 with 19 standardized physicochemical descriptors, giving a 3,129-dimensional design matrix. Ridge regression, support vector regression, random forests, and histogram gradient boosting were tuned by five-fold cross-validation on the training partition and evaluated on untouched test sets under both split regimes, with all random operations seeded.

Data sources

• Zahrt et al., Science 363:eaau5631 (2019) — 1,075 measured enantioselectivities across a 5 by 5 by 43 factorial design
• RDKit 2026.03 — Morgan fingerprints and two-dimensional physicochemical descriptors
• scikit-learn 1.9 — Ridge, support vector regression, random forest, and histogram gradient boosting implementations