Alex Mattinson, David Lindley, Helena Engman, Manthena V.S. Varma, Anh-Thu Nguyen-Trung, Binfeng Xia, Xavier Pepin, Linette Ruston, Wei Zhu, Bill van Osdol, Celine Ollier, David Good, Jonathan Brown, Laurent Boulu, Talia Flanagan, Olivier Nicolas, Staffan Berg, Shanoo Budhdeo, Aarti Patel, Jan Bevernage, Handan He, Sweta Modi, Richard Barker, Claire Jackson, Christer Tannergren, Xiojun Ren, Stephane Beilles, J. Matthew Wood, Johanna Laru, Pierre Daublain, Sara Carlert, Louis Henrion, Jean-Flaubert Nguefack, Gunilla Hanisch, Eva Karlsson, Wen Lin, Shruthi Vaidhyanathan, Robert Carr, Fan Wu, Jin Zhang, Amais Ahmad, James M. Mullin, Sari Pappinen, Andrea Moir, Ke Szeto, Christine Xu, Claire Patterson, David B. Turner, Guillaume Louit, Yuya Wang, Tycho Heimbach, Richard Lloyd, Frans Franek, Masoud Jamei, Christophe Tistaert, Bertil Abrahamsson, Kerstin Julia Schäfer, Leon Aarons, Shriram M. Pathak, Adam S. Darwich, Suet Wong, Dónal Murphy, Amin Rostami-Hodjegan, Timo Korjamo, Helena Thörn, Kartrin Schmid, Michael B. Bolger, Johanna Tuunainen, and Mai Anh Nguyen
Oral drug absorption is a complex process depending on many factors, including the physicochemical properties of the drug, formulation characteristics and their interplay with gastrointestinal physiology and biology. Physiological-based pharmacokinetic (PBPK) models integrate all available information on gastro-intestinal system with drug and formulation data to predict oral drug absorption. The latter together with in vitro-in vivo extrapolation and other preclinical data on drug disposition can be used to predict plasma concentration-time profiles in silico. Despite recent successes of PBPK in many areas of drug development, an improvement in their utility for evaluating oral absorption is much needed. Current status of predictive performance, within the confinement of commonly available in vitro data on drugs and formulations alongside systems information, were tested using 3 PBPK software packages (GI-Sim (ver.4.1), Simcyp® Simulator (ver.15.0.86.0), and GastroPlusTM (ver.9.0.00xx)). This was part of the Innovative Medicines Initiative (IMI) Oral Biopharmaceutics Tools (OrBiTo) project. Fifty eight active pharmaceutical ingredients (APIs) were qualified from the OrBiTo database to be part of the investigation based on a priori set criteria on availability of minimum necessary information to allow modelling exercise. The set entailed over 200 human clinical studies with over 700 study arms. These were simulated using input parameters which had been harmonised by a panel of experts across different software packages prior to conduct of any simulation. Overall prediction performance and software packages comparison were evaluated based on performance indicators (Fold error (FE), Average fold error (AFE) and absolute average fold error (AAFE)) of pharmacokinetic (PK) parameters. On average, PK parameters (Area Under the Concentration-time curve (AUC0-tlast), Maximal concentration (Cmax), half-life (t1/2)) were predicted with AFE values between 1.11 and 1.97. Variability in FEs of these PK parameters was relatively high with AAFE values ranging from 2.08 to 2.74. Around half of the simulations were within the 2-fold error for AUC0-tlast and around 90% of the simulations were within 10-fold error for AUC0-tlast. Oral bioavailability (Foral) predictions, which were limited to 19 APIs having intravenous (i.v.) human data, showed AFE and AAFE of values 1.37 and 1.75 respectively. Across different APIs, AFE of AUC0-tlast predictions were between 0.22 and 22.76 with 70% of the APIs showing an AFE > 1. When compared across different formulations and routes of administration, AUC0-tlast for oral controlled release and i.v. administration were better predicted than that for oral immediate release formulations. Average predictive performance did not clearly differ between software packages but some APIs showed a high level of variability in predictive performance across different software packages. This variability could be related to several factors such as compound specific properties, the quality and availability of information, and errors in scaling from in vitro and preclinical in vivo data to human in vivo behaviour which will be explored further. Results were compared with previous similar exercise when the input data selection was carried by the modeller rather than a panel of experts on each in vitro test. Overall, average predictive performance was increased as reflected in smaller AAFE value of 2.8 as compared to AAFE value of 3.8 in case of previous exercise. QC 20200930