engleski

The outcome of reasoning based on models greatly depends on the procedure used for their validation

In modeling process, three main validation procedures can be used: fitting (the obligatory one), cross-validation, and prediction. The first two are internal validations, and are based on the training data set used also for estimation and optimization of model coefficients. The last one is called external validation, and is performed on external data set containing additional samples (compounds) never used in training. According to literature, severity of these validation methods increases in this order: fitting, cross- validation, and prediction. For each validation procedure corresponding statistical parameters are calculated and reported, like standard errors or correlation coefficients. Conclusions on quality of models or on quality of novel molecular descriptors involved in the models can be different or even opposite if one draws conclusions based on statistical parameters calculated in fitting or in prediction. It will be illustrated on data sets from literature related to structure-property modeling of organic compounds. Also, the lack of pure external validation in modeling of secondary structure of proteins will be addressed. It comes out that more rigorous validation of models should be completed if one wants to draw more definite conclusions based on models.

structure-property modeling ; validation procedure ; fitting ; cross-validation ; prediction ; statistical parameter ; protein secondary structure

nije evidentirano