Population Pharmacokinetic Model

• Problem Description

  • This example is from data contained in a NONMEM tutorial. It represents a population-based pharmacokinetic (PK) model where the properties (attributes) of an individual within a population are used to model that individual’s PK response to the input of a given compound.

  • Properties used include weight, age, creatine clearance, and smoking/non-smoking status.

  • Data include PK time-course profiles for 30 subjects, all receiving the same dose of input compound (dose is not a variable in this study).

• Model Description

  • Nonlinear response function variables are optimized to find the set that yields response curves to best fit the observed data.

  • Predictions are made for two given property sets, neither of which was used to build the response function model.

  • A virtual screen is performed, whereby the population property space is searched to find the highest scoring property sets that maximize a given obective function (VSOF).
    • Objective function used in this case is the area under the curve and above a minimum therapeutic plasma concentration level (3 for this example).
    • Curves that have values above a toxicity threshold (10 for this example) are not considered.

• Results

  • Resulting curves are generated from the nonlinear response function model, for the 30 sets of population data. Model curves are labeled M1-M30. Prediction curves are generated for the 2 prediction property sets. Prediction curves are labeled P1 and P2. These resulting plots are shown here.

  • Response curves for the top 10 VSOF values are shown here.

  • Plots are generated for percent of maximum VSOF value vs property type, for the 4 different property types. These plots are shown here.

• Conclusions

  • The nonlinear response function shows an fairly accurate fit to the observed PK data, over the entire range of input property sets.

  • The response functions for the two predicted property sets appear to be in line with curves that might be expected given those curves generated from the model property sets in relative close proximity to the prediction property set.

  • The nonlinear response function model allows for virtual screening of curves that would result from property sets across the entire domain of the property values.

  • Being able to predict entire time-course curves, rather than just surrogates such as AUC and half-life, allows for more precise screening; for example, identifying those property sets that result in a maximized therapeutic area, which takes into account both intensity and duration of therapeutic concentration levels.

  • Plots of percent maximum therapeutic area vs property (individual’s attributes) allows for quick visualization of the affect that each property has on therapeutic area.