Cross-validation strategies in QSPR modelling of chemical reactions

In this article, we consider cross-validation of the quantitative structure-property relationship models for reactions and show that conventional k-fold (CV) procedure gives an ‘optimistically’ biased assessment prediction performance. To address issue, suggest two strategies model cross-validation, ‘transformation-out’ CV, ‘solvent-out’ CV. Unlike approach does not nature objects, proposed procedures provide unbiased estimation predictive performance novel types structural transformations in chemical going under new conditions. Both suggested have been applied to predict rate constants bimolecular elimination nucleophilic substitution reactions, Diels-Alder cycloaddition. All methodologies tutorial are implemented open-source software package CIMtools (https://github.com/cimm-kzn/CIMtools).