ИСТИНА

Selectivity–the ratio of the desired product to the side products–is the central property of a chemical reaction, which largely determines its usefulness. Arguably, finding reactions with better selectivity and improving selectivity of the known reactions are what synthetic chemistry is about. Theoretical chemistry offers a simple model for computing the ratio between the chemical reaction products given that relative energies of the transition states leading to the products are known. This has stimulated use of quantum chemical modeling (mostly at density functional theory level) for discerning mechanisms of chemical reactions and their computer-aided optimization. However, the computed selectivity incorporates errors, propagated from inexact density functional, incomplete basis set, imperfect representation of the molecular surroundings and so on, raising the question of predictions fidelity. In this work we propose a universal approach for propagating errors in individual energies into the predicted selectivities and then invert it, finding out that current computational approaches, accounting for dispersion and solvation, have mean absolute errors (MAE) as small as 0.3 kcal/mol for relative energies of enantiomeric transition states. Our approach allows one to directly use this MAE to make predictions in terms of “There is X% probability that the ratio between products A and B is between Y and Z.” instead of usual “Calculations predict A:B ratio to be W.”; the MAE of 0.3 kcal/mol is small enough to make error bars within 10% in most cases. Interestingly, we find that switching off solvation leads to the error rise up to ~1 kcal/mol, rendering the results meaningless: in most cases, energy differences between the transition states leading to concurrent products are less than that. Our approach in its direct form can be used to quantify the computational error in any observable property given that MAE of the used approximations in this property was estimated using the inverted approach (which requires only several computed cases with known experimental answers). Besides, MAEs obtained using the inverted approach measure the random error in the level of theory and are good for their benchmarking.