Principal Research Scientist Vertex Pharmaceuticals, United States
Disclosure(s):
Guanyu Wang, PhD: No financial relationships to disclose
Objectives: Understanding uncertainty in human dose prediction is vital to making correct decisions in several areas of drug development (e.g., lead optimization, first-in-human dose selection, drug supply planning, etc). Point estimate of dose using mean predicted PK parameters does not account for uncertainty in predictions and may not help manage risks with clinical study planning, while simple practice of Monte Carlo simulations may overestimate uncertainty and make the predictions too wide to be useful. We developed an algorithm as a balance between the two approaches to estimate the uncertainty in dose by incorporating the uncertainty in various methods used for prediction of pharmacokinetic parameters.
Methods: The anticipated human dose depends on a number of PK parameters, and each parameter is predicted using a selection of methods. We defined uncertainty in dose by simultaneously incorporating the uncertainties propagated from individual prediction methods used for each PK parameter as follows: 1) derived uncertainty of individual PK prediction methods using a dataset of 774 literature compounds collected from 10 publications; and 2) developed a “batch log-averaging” approach to combine the uncertainties from individual prediction methods of each PK parameter to form an empirical distribution for the parameter. Dose is calculated for each parameter combination drawn from the “batch log” distribution and those doses, after a sufficient number of iterations, form a probability distribution of the anticipated human dose. 3) When communicating uncertainty, we chose the Words of Estimative Probability1 (known as WEP) as the numerical definition of confidence in dose.
Results: Uncertainties (quantified by a coefficient of variation (CV%)) of three volume of distribution prediction methods and three clearance prediction methods were derived from the dataset we built from literature. We evaluated the “batch log-averaging” algorithm along with the derived CV% on a set of example compounds. By varying certain parameters and their uncertainties, we evaluated how uncertainty in dose could be influenced by compound properties, pharmacokinetic driver of efficacy, efficacy target uncertainty, and administration schedule. The upper bound of dose was communicated as the dose with 93% cumulative probability, corresponding to the certitude level of “almost certain” in WEP.
Conclusions: The “batch log-averaging” algorithm is able to preserve the uncertainty from various measurements across multiple parameters while placing a tighter bound for the probabilities projected from all anticipated sources of uncertainty. Using case studies, we demonstrated that the approach could quantify the level of confidence in dose prediction when communicating with multidisciplinary stakeholders to support rank ordering of lead compounds, set dose and regimen in early clinical trials, and help inform material requirements for scale-up compound syntheses.
Citations: [1] Words of Estimative Probability, Stud. Intel. V8: 4, 1993
