How to rank and discriminate artificial neural networks? Case study: prediction of anticancer activity of 17-picolyl and 17-picolinylidene androstane derivatives

Model discrimination is still not a resolved task. The classical statistical approaches lead to different results (for the same models) and at the same time a lot of models seem to be statistically equivalent. The authors deliberately select such conditions when their algorithm is superior. Hence, it is better to apply different approaches to compare and rank the models fairly. This paper presents the application of methodology called sum of ranking differences (SRD) to rank the artificial neural network models [quantitative structure–activity relationship (QSAR) models] designed for prediction of anticancer activity of 17-picolyl and 17-picolinylidene androstane derivatives toward androgen receptor negative prostate cancer cells (AR-, PC-3). The SRD method suggests the consistent models, in terms of compounds order and proximity to the golden standard, which should preferably be used in the prediction of anticancer activity of studied androstane derivatives.