CHEMOMETRIC STUDY AND VALIDATION STRATEGIES IN THE STRUCTURE-ACTIVITY-RELATIONSHIPS OF NEW CARDIOTONIC AGENTS

Forty-two molecules, thirty-eight milrinone analogues, two lead compou nds, amrinone and milrinone, and two commercial products have been stu died using chemometrical techniques applied to thirty theoretical desc riptors and two biological activities (each one at three different con centrations). PLS Regression was applied both in the usual form PLS-1, with one response variable, and as PLS-2, with the contemporary study of more activities in the block of response variables. Regression mod els (both with the original activities and with log and arctan transfo rms) were refined by progressive elimination of conformers and of non- relevant predictors, one-at-a-time, on the basis of the relevance in t he regression equation. Different sorts of model refinement gave origi n to four chemometrical strategies. Special attention was deserved to the development of validation procedures for the regression analysis, in order to evaluate the true predictive ability of the refined models . The predictive optimization was based on cross-validation. Complete validation using three sets (training, optimization, external) was app lied in one of the strategies. Both optimization and validation were p erformed in different conditions in order to eliminate the possibility of chance correlation. The severe validation procedures applied preve nt underestimate of prediction error, frequently encountered when part ial validation procedures are applied. Only one biological activity at the highest concentration can be predicted from the theoretical descr iptors with a reasonable prediction error, measured by cross-validated explained variance. Only volume descriptors have a sure importance in the final regression model.