Grouping Plasma Pharmacokinetic Profiles Using Kohonen Self-organizing Maps

Pharmacokinetic data from large clinical trials is often highly variable, even in well-controlled trials involving homogeneous cohorts. Common sources of such variability are genetic polymorphism in the CYP enzymes, developmental and demographic differences, and dietary or lifestyle choices. Pharmacokinetic scientists are tasked with quantitating variability in these data sets and identifying mechanistic sources of variability. Rapid processing of data to identify and graphically depict trends is standard practice in pharmacokinetic modeling and when implemented appropriately can effectively guide modeling efforts. Kohonen self-organizing mapping is an unsupervised clustering tool used in many scientific fields to organize and group data. Using the software ADMET PredictorTM from Simulations Plus, Kohonen mapping was applied to highly variable plasma concentration versus time (Cp-time) data from 93 subjects selected for a particular trial using criteria intended to minimize variability. These criteria are typical for the pharmaceutical industry: narrow range of age, body weight, and body mass index. Conditions of the clinical trial were also controlled to limit variability, including, for example, restricting and controlling dietary intake, concomitant medications, and physical activity. Prior to grouping, Cp-time profiles were characterized by peak concentrations which varied by three orders of magnitude, with the time to reach peak concentration ranging from one hour to over 20 hours. Applying Kohonen mapping to the data yielded Cp-time profile groups organized by shared features such as sharp early peaks, sharp late peaks, rounded late peaks, and profiles with multiple peaks. In this example, the 93 profiles were grouped into six easily identified clusters. The smallest groupings contained two profiles and the largest grouping contained 27 profiles. This grouping positively impacted the pharmacokinetic analysis by rapidly identifying relevant subgroups among subjects. In this case, through subsequent simulation it was determined that tablet disintegration effects were responsible for pharmacokinetic profiles of subjects in some subgroups, but that intersubject differences in physiology were responsible for Cp-time profile in other subgroups. This paper demonstrates that Kohonen mapping can be a useful tool for clinical pharmacologists, pharmacokinetic scientists, and toxicologists who are faced with increasingly complicated clinical trials, highly technical formulations, and ever-shrinking deadlines. Application of this technique to pharmacokinetic data analysis can improve efficiency and productivity in the already burdened pharmaceutical industry.