Article Preview
TopIntroduction
The majority of drugs are toxic (Robciuc et al., 2017; Erdem Kuruca et al., 2019). Consequently, the assessment of drug toxicity is an important task at the interface of chemistry, biology, medicine. Quantitative structure – property / activity relationships (QSPRs/QSARs) are useful tools to solve the task (Zhao & Li, 2018; Zhang et al., 2018; Kar & Roy, 2015; Toropov et al., 2014; Gissi et al., 2014; Raevsky et al., 2012). It is to be noted that modelling of both therapeutic efficacy and toxicity can be performed using the QSPR/QSAR technique. There are different categories of drug toxicity, e.g., cardiac toxicity (Frid & Matthews, 2010), liver-related adverse effects of drugs (Wong, 2006; Rodgers et al., 2010; Toropov et al., 2012), toxicity of psychotropic drugs (Esteki & Khayamian, 2008; Gissi et al., 2014; Gómez-Lumbreras et al., 2018).
One can find newer toxicity of existing drugs from further medicinal experiments as well as exchange of old drug forms by new drug forms. In addition, simultaneous application of different therapeutic agents can lead to detection of unexpected effects. Sometimes, the detection can be made during long term observations, e.g. years, or even tens of years (Ziuganov et al., 2005; Mizuno et al., 2006). In order to obtain a more stable and reliable system for registration of drugs, so-called ADMET conception (absorption, distribution, metabolism, excretion, and toxicity) has been developed and suggested (Gola et al., 2006; Dearden, 2007; Cheng et al., 2013; Vora et al., 2019; Zaki et al., 2019).
In the framework of the ADMET conception as well as in the frame of generalized pharmaceutical applications, several approaches are formulated to solve the tasks related to sphere of pharmaceutical sciences. These are: (i) linear regression analysis (Roy & Mitra, 2012); (ii) artificial neural networks (Fjodorova et al., 2010; Verma & Matthews, 2015); (iii) random forest (Schöning et al., 2018); (iv) super vector machine (Yu et al., 2017); (v) Monte Carlo technique (Veselinović et al., 2016; Bhargava et al., 2017).
Each of the above approaches is characterized by a list of advantages as well as by a list of disadvantages. The common practice in the QSPR/QSAR studies is to search for improvement of models. Naturally, some criteria to compare QSPR/QSAR models are necessary. Some widely used criteria contradict common sense and results of manifold computational experiments (Hartung & Hoffmann, 2009) but nonetheless they are often applied for estimation of the QSPR/QSAR models. For instance, well-known q2 being calculated with the training set (Golbraikh & Tropsha, 2002) factually does not correlate with R2 (determination coefficient) for external validation set (Kubinyi, 2004). This situation is known as “Kubinyi paradox” (Hartung & Hoffmann, 2009). The Kubinyi paradox indicated the necessity of external validation for any QSPR/QSAR models. In addition, the Kubinyi paradox questioned reliability of internal validation (i.e. validation via q2).