Comparison of Software for Prediction of Fraction Absorbed and Unbound in Humans

The main objective of this study was to evaluate and compare the performance of 5 PK software, ANDROMEDA by Prosilico 2.0, and 4 free, web-based prediction tools, PkCSM, Swiss ADME, ADMETlab 3.0 and DruMAP 2.0, for predictions of fraction absorbed (fa) and unbound (fu) in humans. Sets with compounds with available and undisclosed estimates were selected (n=140). The risk that test compounds have been used in training sets for model building (and thereby, influenced and exaggerated the predictive performances) was minimized. At least, these were not included in training sets for ANDROMEDA. One set consisted of compounds that have not been marketed and for which pharmacokinetic information has not been publically disclosed. Quantitative and qualitative evaluations and comparisons were done. For both fa and fu, ANDROMEDA was clearly more accurate and balanced than the others, with higher Q2 (0.69 vs 0.35 for fa; 0.94 vs 0.62-0.76 for fu), lower mean errors (15 % vs 28 %; 2.3- vs 4.1- to 36-fold), lower maximum errors (54 % vs 92 %; 10- vs 30- to 524-fold), more correct predicted classes (70-77 % vs 13-54 %), no failed, inconclusive or poor predictions (as found for 3 of the other software), wider application range, and minimal skewness at low values. It had intercepts on fa- and fu-prediction axes that were ca 1/3 and 1/175 to 1/23 compared to those found for the other software, which is of particular importance. Two software, PkCSM and Swiss ADME, were considered inappropriate, whereas ADMETlab took an intermediate performance position. Apparently, DruMAP was second best performing software. Overall, there was poor performance overlap between the software (7-24 %), with many contradictory predictions. Advantages with ANDROMEDA suggest that this is the software of choice for those that desire adequate predictions of fa and fu in humans and estimates of certainty. The findings are of particular interest for the 3R-process.