Purpose: Animal models are commonly used to predict human pharmacokinetics including first-in-human doses. Unfortunately, for drugs that are highly metabolized, animal models are often poor predictors of the overall extent of drug metabolism in humans. In this study, we ask if animal models are better at predicting human pharmacokinetics when looking only at drugs that are poorly metabolized, specifically Biopharmaceutics Drug Disposition Classification system (BDDCS) Class 3 and 4 drugs. BDDCS Class 3 and 4 drugs are those eliminated by a route other than metabolism for which it is confirmed that ≥30% of the drug is excreted unchanged in either the urine or the bile. The purpose of this project was to conduct a literature review of fraction excreted unchanged in the urine (fe) values, as a marker of a lack of metabolism in rats, dogs, and monkeys for all BDDCS Class 3 and 4 drugs identified to date including the amended list published by Hosey et al in 2015. The developed dataset of fe values was analyzed to determine the ability of animals to predict a lack of metabolism in humans.
Methods: A literature review of previously published pharmacokinetic studies in the three animal species and humans was conducted for all BDDCS Class 3 (n=288) and Class 4 (n=64) drugs. The pharmacokinetic studies identified were reviewed for fe values and only included in the analysis if the Class 3 or 4 drug was administered intravenously and unchanged drug in the urine was measured. The literature review focused only on urinary excretion rather than biliary elimination due to the difficulty in obtaining biliary values in humans. The final dataset contained 98 BDDCS Class 3 and 4 drugs. As fe data for many drugs were collected in more than one species, 159 total fe values were included in the dataset (80 from rats, 51 from dogs, and 28 from monkeys). The fe values for each animal species were compared to the human fe values through unweighted linear regression and two-fold error analysis. Spearman and Pearson correlation coefficients were calculated using GraphPad Prism. Additionally, binary classifications with corresponding sensitivity and specificity analyses were used to classify collected animal fe data with ability to predict a human fe ≥30% or <30%. When analyzing the data, two different thresholds were utilized for animal fe data. The first threshold was animal fe ≥30% or <30% to directly match the cutoff in humans. The second threshold assessed was determined by Receiver Operating Characteristic (ROC) curve analysis to optimize both sensitivity and specificity of the threshold. Sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) were calculated for both thresholds.
Results: Regression analysis indicated a positive, statistically significant association between rat vs. human fe values (r2=0.34, p<0.0001) and dog vs. human fe values (r2=0.21, p=0.0007). The association between monkey and human fe data was not significant (r2=0.024, p=0.432). A two-fold error analysis was performed and 81%, 82%, and 61% fell within the two-fold error window for rat, dog, and monkey, respectively (Figure 1). When a fe cut off of ≥30% was used for all animal models, sensitivities for all animal models ranged between 76% and 87% and specificities ranged from 30% to 67%. Individual animal model sensitivity, specificity, accuracy, PPV, and NPV values are noted in Figure 2. ROC curve analysis optimized sensitivity and specificity and suggested that cutoffs of 40%, 45%, and 10% for rats, dogs, and monkeys, respectively, may better predict a ≥30% fe in humans. The cutoffs from the ROC curve analysis resulted in sensitivities ranging from 74% to 96% and specificities ranging from 63% to 83%. Individual animal model sensitivity, specificity, accuracy, PPV, and NPV values using the ROC curve analysis cutoffs of 40% for rats, 45% for dogs, and 10% for monkeys, are reported in Figure 3.
Conclusion: Animal models, especially dogs and rats, appear predictive of human fe values for poorly metabolized drugs. The predictive ability of monkeys was the worst of the three animal models investigated. Linear regression analysis poorly predicted human disposition as determined by the low r2 values in all animal models below 0.5. Two-fold error analysis indicated rats and dogs better predicted human fe within a two-fold range than monkeys. Binary classification analysis showed that when a fe cutoff of ≥30% was used for both animal models and humans, good sensitivity was achieved; however, specificity was poor. ROC curve analysis yielded new fe cutoffs for animal models, which improved specificity for all animal models, and sensitivity for monkeys. The determined fe cutoffs of 40% in rats, 45% in dogs, and 10% in monkeys were predictive of a high fe of ≥30% in humans. These cutoffs could be useful in predicting human pharmacokinetics earlier on in drug development including if the drug is likely poorly metabolized in humans and its classification as either BDDCS Class 3 or 4.