During drug development, several thousand candidate compounds must be screened and evaluated for their biopharmaceutical properties to determine their suitability as drug candidates. In assessing their drug-likeness, particular importance is attributed to the processes of absorption, distribution, metabolism, excretion, and toxicity, together referred to as biopharmaceutical drug properties. Although experimental in vitro determination of these properties represents the gold standard, it is often time-consuming and costly, while the obtained results do not always translate into successful in vivo outcomes. Consequently, most drug candidates fail during development. To overcome these limitations, computational methods have been developed as faster and more efficient option. These platforms use machine learning and deep learning algorithms to predict biopharmaceutical profile and support compound optimization. The predicted parameters can further be used for the development of physiologically based pharmacokinetic (PBPK) models to simulate systemic drug distribution and a more accurate evaluation of pharmacokinetic profiles.
The aim of this master’s thesis was to predict the biopharmaceutical properties of bisphenol A and several bisphenol analogues using 13 free web-based applications. The predictive performance, concordance and applicability of these tools were evaluated. For this purpose, simplified physiologically based pharmacokinetic models for bisphenol A and bisphenol S were developed for intravenous administration in rats. Physiologically based pharmacokinetic models were developed using two different approaches and the PK-Sim software. The developed models were validated against data from selected in vivo studies in rats reported in the literature. The influence of different pharmacokinetic parameters on model prediction performance was investigated, and the developed models were evaluated in terms of accuracy, precision of predicted pharmacokinetic profiles. Applicability of the selected approach of using web-based applications in model development was assessed. .
ADMETlab version 3.0 was found to provide the broadest prediction coverage and the highest overall reliability. Accurate predictions were also generated by DeepPK and PreADMET. The results confirmed that multiple prediction platforms should be used for compound evaluation. The highest agreement among predictions was observed for lipophilicity, whereas the greatest variability was observed for plasma unbound fraction predictions. Simplified models for intravenous administration were successfully developed and validated; however, their predictive accuracy remained limited due to model simplifications and the exclusion of enterohepatic recirculation in rats. Nevertheless, all developed models predicted pharmacokinetic parameters within a two-fold range of experimental values. The highest precision in PK-Sim was achieved when median biopharmaceutical values obtained from web-based platforms were applied.
|
