Artificial Neural Network Modeling for Drug Dialyzability Prediction

Abstract

Purpose. The purpose of this study was to develop an artificial neural network (ANN) model to predict drug removal during dialysis based on drug properties and dialysis conditions. Nine antihypertensive drugs were chosen as model for this study. Methods. Drugs were dissolved in a physiologic buffer and dialysed in vitro in different dialysis conditions (UFRmin/UFRmax, with/without BSA). Samples were taken at regular intervals and frozen at -20ºC until analysis. Extraction methods were developed for drugs that were dialysed with BSA in the buffer.  Drug concentrations were quantified by high performance liquid chromatography (HPLC) or mass spectrometry (LC/MS/MS). Dialysis clearances (CLDs) were calculated using the obtained drug concentrations.  An ANOVA with Scheffe’s pairwise adjustments was performed on the collected data in order to investigate the impact of drug plasma protein binding and ultrafiltration rate (UFR) on CLD. The software Neurosolutions® was used to build ANNs that would be able to predict drug CLD (output). The inputs consisted of dialysis UFR and the herein drug properties: molecular weight (MW), logD and plasma protein binding. Results. Observed CLDs were very high for the majority of the drugs studied. The addition of BSA in the physiologic buffer statistically significantly decreased CLD for carvedilol (p= 0.002) and labetalol (p<0.001), but made no significant difference for atenolol (p= 0.100). In contrast, varying UFR does not significantly affect CLD (p>0.025). Multiple ANNs were built and compared, the best model was a Jordan and Elman network which showed learning stability and good predictive results (MSEtesting = 129). Conclusion. In this study, we have developed an ANN-model which is able to predict drug removal during dialysis. Since experimental determination of all existing drug CLDs is not realistic, ANNs represent a promising tool for the prediction of drug CLD using drug properties and dialysis conditions. This article is open to POST-PUBLICATION REVIEW. Registered readers (see “For Readers”) may comment by clicking on ABSTRACT on the issue’s contents page.