Backgrounds: The CORAL software has been developed as a tool to build up quantitative structure–
activity relationships (QSAR) for various endpoints.
Objective: The task of the present work was to estimate and to compare QSAR models for biochemical activity of
various therapeutic agents, which are built up by the CORAL software.
Method: The Monte Carlo technique gives possibility to build up predictive model of an endpoint by means of selection
of so-called correlation weights of various molecular features extracted from simplified molecular input-line entry system
(SMILES). Descriptors calculated with these weights are basis for building up correlations "structure – endpoint".
Results: Optimal descriptors, which are aimed to predict values of endpoints with apparent influence upon metabolism
are crytically compared in aspect of their robustness and heuristic potential. Arguments which are confirming the necessity
of reformulation of basics of QSARs are listed: (i) each QSAR model is stochastic experiment. The result of this
experiment is defined by distribution into the training set and validation set; (ii) predictive potential of a model should be
checked up with a group of different splits; and (iii) only model stochastically stable for a group of splits can be estimated as
a reliable tool for the prediction. Examples of the improvement of the models previously suggested are demonstrated.
Conclusion: The current version of the CORAL software remains a convenient tool to build up predictive models.
The Monte Carlo technique involved for the software confirms the principle “QSAR is a random event” is important
paradigm for the QSPR/QSAR analyses.