The survey was limited to nine pharmaceutical companies within the AO focus group of the Centre for Applied Pharmacokinetic Research. To minimize potential bias due to intercompany communications, the survey was conducted prior to the focus group discussions. The response rate was 100%. Six out of nine companies (67%) claimed to have experience in developing AO/mixed AO and CYP substrates, although no compound had received marketing authorization. All nine companies used in vitro systems for the prediction of in vivo CL including human hepatocytes, human liver cytosols, and human liver S9. However, there was no consensus on which in vitro system is the best when predicting in vivo CL. Reaction phenotyping using chemical inhibitors (e.g., hydralazine, raloxifene, or menadione) was the most common in vitro method for predicting fmAO (Fig. 1). Only one company considered storage instability of human tissue samples or extrahepatic contribution of AO for total clearance predictions, although 22% of the companies had quantified AO abundance or variability in various human tissues. Over half of the companies had some experience in developing PBPK and/or static mechanistic models for translational prediction of human pharmacokinetic and victim DDIs of AO or mixed AO/CYP substrates. One-third of the companies applied modeling in limited cases or used only static models for worst-case DDI predictions.

Survey responses from pharmaceutical companies (n = 9). (A) Method(s) in use for the prediction of in vivo clearance for aldehyde oxidase/mixed aldehyde oxidase and CYP substrates. (B) In vitro system(s) in use by companies among human hepatocytes (HH), human liver cytosols (HLC), and human liver S9 (HLS9). (C) Method(s) used for predicting fraction metabolized by aldehyde oxidase versus fraction metabolized by CYPs. (D) Favored method for in vivo clearance predictions for aldehyde oxidase and mixed aldehyde oxidase and CYP substrates. Different color bars differentiate multiple-choice answers.

A systematic literature search of AO substrates yielded 37 compounds with diverse pharmacokinetic properties for which in vitro studies in human hepatocytes, human liver cytosols, and/or human liver S9 were available. Among these compounds, CLint,u,in vitro data could not be calculated for nine compounds (A7701, bafetinib, CL387785, INCB28060, lapatinib, LDN193189, ML347, SB525344, and VX509) (Dick, 2018). Three compounds (phthalazine, ripasudil, VU0409106) did not have available in vivo intravenous or oral pharmacokinetic data. The AO contribution to hepatocyte CLint was insufficient to allow a reliable calculation of fmAO for imatinib (Toselli et al., 2022), and AMG900 and favipravir had missing measured fup or B/P data. Only CLint data measured in the absence of NADPH were included in the human liver S9 dataset, to reduce potential confounding impact of CYP-mediated metabolism in this system. In-house data for human hepatocytes (Supplemental Table 1) were included in the dataset after correcting for nonspecific binding using the literature data. CLint measurements of compounds were sensitive to assay conditions especially in case of low turnover compounds such as XK469. After excluding data that were below limits of quantification, the final database consisted of 100 CLint,in vitro measurements for 22 AO substrates (Table 2). Among those, 10 compounds were also metabolized by CYPs to varying degrees. The largest dataset was for human hepatocytes (n = 19 compounds), followed by human liver cytosols (n = 16) and liver S9 (n = 11), with nine compounds overlapping across all in vitro systems. Table 3 lists the fuinc values, either corrected from literature (n = 16) or predicted from logP data (n = 6), that were used in the calculation of CLint,u,in vitro.

Predicted CLint,u obtained from scaling in vitro data measured in human hepatocytes, human liver cytosols, and human liver S9

References for all in vitro studies are available in Supplemental Table 2.

logP, fuinc, fup, and B/P data used in the IVIVE analysis

References for human plasma protein binding and B/P ratio data are available in Supplemental Table 2.

Mean fup and B/P values from literature and in-house data (Table 3) were used in the back-calculation of observed CLint,u. Despite the similar in vitro methodology between studies, fup values of DACA, fasudil, idelalisib, and ziprasidone showed %CV over 30%. B/P data were more consistent between different studies.

Plasma clearance data after intravenous administration, and apparent clearance data after oral drug administration, were collated for the investigated AO substrates from 25 reports (Table 4). The intravenous plasma clearance (n = 10) ranged from 0.04 to 73.2 mL/min/kg for XK469 and fasudil, respectively. Six compounds had CLH over 80% of average QH, of which BIBX1382, carbazeran, and fasudil had CLH exceeding QH suggesting non-AO metabolic pathways (Table 4) may be involved in extrahepatic metabolism, since AO clearance of extra-hepatic tissues (kidney, lung, vasculature, and intestine) predicted from S9 fractions were to be < 1% of the liver (Kozminski et al., 2021). For the compounds with CLH exceeding QH, the cut-off CLH of 90% of QH was applied (Cubitt et al., 2009) in the calculation of CLint,u via the well-stirred liver model. Oral plasma clearance (n = 12) ranged between 2.05 to 17,857 mL/min/kg for lenvatinib and LuAF09535, respectively. The correction for fa was applied for 6-deoxypenciclovir (Filer et al., 1994), capmatinib (Glaenzel et al., 2020), idelalisib (https://www.tga.gov.au/sites/default/files/auspar-idelalisib-150714.pdf), and lenvatinib (https://www.ema.europa.eu/en/documents/assessment-report/lenvima-epar-public-assessment-report_en.pdf) with respective fa values of 0.8, 0.67, 0.91, and 0.97.

Intravenous or oral (apparent) plasma clearance, CLint,u, fmAO, and observed CLint,u,AO

Weighted mean of plasma clearance for compounds for which more than one clinical study or multiple dose studies were available. Number of studies and between study coefficient of variation are given in parenthesis. References for fmAO data are available in Supplemental Table 3. When fmAO data from human mass balance studies were available, only those were used as the input in the CLint,u,AO calculation; otherwise, all fmAO values in Supplemental Table 3 were considered.

All available data on fmAO, from in vitro hepatocyte or S9 inhibition studies; PBPK model-based predictions in literature; absorption, distribution, metabolism, and excretion (ADME) studies; and previous literature were collated, with up to 10 different fmAO values reported for individual compounds (Supplemental Table 3). Mean fmAO ranged from 0.2 to 1, with compounds grouped as low (0.2–0.49; n = 6), medium (0.5–0.79; n = 6), and high (0.8 and over; n = 10) AO-metabolized compounds. However, large variability was noted for some compounds, including ziprasidone, capmatinib, and lenvatinib, with %CV in fmAO of 117%, 76%, and 39%, respectively, due to the discrepancy between in vitro and in vivo data. Considering this high uncertainty, human mass balance data were considered as a primary input for fmAO when reported (available for six compounds in the database). For the remaining compounds (n = 16), mean and range of fmAO from all reported values were used (Supplemental Table 3).

IVIVE using purely physiologic scaling factors was applied to predict CLint,u (human hepatocytes) and CLint,u,AO (human liver cytosols and S9) (Table 2). Mean predicted CLint,u,AO using human liver cytosols or S9 data were higher than the mean of predicted total CLint,u in human hepatocytes for six compounds having data from all in vitro systems. Mean predicted CLint,u,AO between human liver cytosol and S9 were correlated with R2 of 0.5 (n = 11 compounds), and most of these compounds (n = 7) had higher predicted mean CLint,u,AO using human liver S9 than human liver cytosols.

Large inter-assay variability of predicted CLint,u was observed for all three in vitro systems. For compounds with data from multiple studies, the %CV ranged from 26% to 91% for human hepatocytes, 22% to 73% for human liver cytosols, and 5% to 98% for human liver S9 (Table 2). Potential contributions of factors such as lot-to-lot variability, data analysis methodology, and publication date to the inter-assay variability were investigated. For most of the compounds, in vitro data were generated using hepatocytes or subcellular fractions from different commercial sources, with lot numbers and supplier information on AO activity or assay sensitivity typically not reported in literature studies. The predicted CLint,u for eight compounds from a lot of cryopreserved hepatocytes pooled from individual donors with relatively high AO activity were comparable with corresponding data in lots for which AO activity of donors was unknown. The highest CLint,in vitro for carbazeran in human liver S9 (twofold higher than mean value) came from measurement of the early fast rate of the metabolite formation using the modulated activity model (Abbasi et al., 2019). Data were also analyzed to explore whether in vitro systems/methods that had been refined over the past decade (e.g., due to improved tissue processing, modified assay formats, etc.) would result in higher predicted CLint,u. However, there were insufficient data for each substrate to assess potential trends between CLint,in vitro and publication date.

The observed CLint,u values back-calculated from intravenous clinical data ranged between 3.17 and 5012 mL/min/kg, for methotrexate and ziprasidone, respectively (Table 4). Our static approach did not consider differences in physiologic parameters (i.e., variability in liver blood flow) between study groups. The back-calculated CLint,u was highly sensitive to the variability in plasma and blood binding especially for high clearance compounds; therefore the mean fup and B/P values from various literature was taken when applicable. From oral data, the lowest and the highest CLint,u were observed for RS8359 (17.1 mL/min/kg) and LuAF09535 (68,681 mL/min/kg), respectively. Overall, observed CLint,u values covered a wide range, with compounds evenly distributed across low (< 100 mL/min/kg; n = 6), medium (101–1000 mL/min/kg; n = 10), and high (> 1000 mL/min/kg; n = 6) CLint,u groups. After applying mean fmAO values, the mean of observed CLint,u,AO in the complete dataset ranged from 0.63 to 59,753 mL/min/kg for methotrexate and LuAF09535, respectively (Table 4).

The predictive performance of IVIVE based solely on physiologic scalars was assessed with the bias and precision of the predictions for each in vitro system (Table 5). The overall trend of underprediction of in vivo CLint,u was prominent in all systems with the highest gmfe of 10.4 noted in human hepatocyte dataset (n = 19 substrates). There was no correlation (R2= 0.02) between observed and predicted CLint,u from human hepatocytes (Fig. 2). FK3453 and LuAF09535 were outlier compounds in the human hepatocyte dataset with predicted CLint,u below 1% of observed values, whereas 11% of the compounds were predicted within twofold. Outliers were not excluded from the dataset, and the distribution of gmfe values were assessed via the L-O-O approach. The arithmetic mean±SD (range) of L-O-O gmfe values of predictions in human hepatocytes was 10.4±0.8 (8.4–11.7) (Fig. 3).

Bias and precision of in vitro to in vivo human hepatic unbound intrinsic clearance (for human hepatocytes) and human unbound intrinsic clearance by aldehyde oxidase (for human liver cytosols and human liver S9) predictions

Mean predicted and observed unbound intrinsic clearance values were used in these predictions.

Comparison of predicted and observed CLint,u for 22 aldehyde oxidase substrates. IVIVE of in vitro data from human hepatocytes, cytosol, and S9 was performed using corresponding physiologic scaling factors for each in vitro system. (A) Comparison of observed and predicted CLint,u from in vitro data generated in human hepatocytes (n = 19, red square). (B) Comparison of observed and predicted CLint,u,AO from human liver cytosols (n = 16, green circle) and human liver S9 (n = 11, blue triangle) in vitro data. Dashed lines represent the twofold deviation from the line of unity. Vertical and horizontal error bars represent inter-assay variability and uncertainty in fmAO, respectively (min-max). Compounds are numbered in the following order: 1, 6-deoxypenciclovir; 2, BIBX1382; 3, capmatinib; 4, carbazeran; 5, DACA; 6, fasudil; 7, FK3453; 8, idelalisib; 9, JNJ38877605; 10, lenvatinib; 11, LuAF09535; 12, methotrexate; 13, O6-Benzylguanine; 14, PF4217903; 15, PF5190457; 16, PF6273340; 17, PF-945863; 18, RS8359; 19, XK469; 20, zaleplon; 21, ziprasidone; 22, zoniporide.

Box plot of gmfe values calculated after leaving one compound out from each dataset. In vitro data includes human hepatocytes (HH; n = 19), human liver cytosols (HLC; n = 16), and human liver S9 (n = 11). Within each box, horizontal black lines denote median values and + denote mean values; boxes extend from the 25th to the 75th percentile of each group's distribution of values; vertical extending lines denote the minimum and maximum values; dashed lines represent the geometric mean of gmfe calculated in each dataset (overall gmfe) as 10.4, 5.6, 7.7, and 5.0 for HH, HLC, HLC (fucyt = 1 approach), and HLS9, respectively.

The predictive performance of CLint,u,AO from human liver cytosol and S9 were comparable, with gmfe of 5.6 (n = 16) and 5.0 (n = 11), respectively. The observed and predicted CLint,u,AO in human liver cytosols and human liver S9 were weakly correlated with R2 of 0.28 and 0.37, respectively. A majority of compounds were underpredicted using data from human liver cytosols and human liver S9, with CLint,u,AO overpredicted for only one (different) compound in each dataset, by 1.6- and 2.4-fold, respectively. The corresponding underpredictions in these systems were up to 29- and 18-fold, respectively. The arithmetic mean±SD (range) of L-O-O gmfe values in human liver cytosol and S9 datasets was 5.6±0.4 (5.1–6.1) and 5.0±0.5 (4.4–5.7), respectively (Fig. 3). The predictive performance of CLint,u,AO using human liver cytosols and S9 was dependent on the fmAO estimates with associated uncertainty.

Considering the general lack of measured fucyt in the database, assumption of fucyt of 1 was explored as alternative approach to predicted fucyt from microsomal data and to evaluate sensitivity of fucyt toward IVIVE performance. For most of the compounds, the impact of assuming fucyt of 1 on predicted CLint,u,AO was minimal. However, for highly protein bound compounds BIBX1382, lenvatinib, ziprasidone, and methotrexate predicted CLint,u,AO was twofold lower when assuming fucyt of 1 compared with using predicted fucyt from available fumic or logP/D data (Supplemental Fig. 1), causing an increase in gmfe to 7.7 (Fig. 3).

The predictive performance of all in vitro systems showed high variability between compounds, even when considering groups with similar observed CLint,u or CLint,u,AO. In human liver cytosols and S9, there was no clear trend between the observed CLint,u and the fold underprediction. Conversely, in the human hepatocyte dataset, the IVIVE performance improved with decreasing observed CLint,u. The gmfe of predictions including low (n = 3), medium (n = 10), and high clearance (n = 6) compounds were 3.4, 7.6, and 30, respectively. For the high clearance group, predictive performance was still the worst [gmfe of 10.4 (n = 4)] even after excluding the two outliers with the highest fold underprediction (FK3453 and LuAF09535). The extent of underprediction of CLint,u in human hepatocytes increased with higher fmAO. The gmfe of predictions for low, medium, and high fmAO compounds were 6.1- (n = 5), 10.5- (n = 6), and 14.2-fold (n = 8), respectively. However, when the two outlier compounds (FK3453 and LuAF09535) were excluded, the gmfe was 5.4 (n = 6) for high fmAO compounds.