Go to content oder Go to navigation


Franziska Thoma
MSc Bioinformatics / University of Potsdam

PharMetrX Research+ Program
PhD student year: 2021

University of PhD: University of Potsdam
Supervisor: Prof. Wilhelm Huisinga
Co-Supervisor: Prof. Charlotte Kloft
Mentoring I-Partner: AbbVie

PhD Project

Continuous learning by integrating reinforcement learning and data assimilation to individualize drug treatments with a special focus on model bias

Model-informed precision dosing (MIPD) is a framework for individualisation of drug dosing based on pharmacokinetic/pharmacodynamic models and patient specific data. It is particularly indicated in settings with high inter-individual variability and a narrow therapeutic window. The underlying mathematical models are typically derived using nonlinear mixed effects modeling. They are based on data from prior clinical studies. Therefore, both the model structure and estimated model parameters depend on the underlying study population.
Using this prior information in combination with individual patient data (e.g. biomarker levels, weight, age), dosing decisions can be made before the start of therapy (a priori prediction). Based on therapeutic monitoring over the course of treatment, dosing can be further individualised (a posteriori predictions).

In using models developed on study populations, one implicit assumption is that both the structural model and the parameter estimates represent the target population sufficiently well and variability in the outcome is described appropriately. We do, however, know that some model bias may be expected, as clinical studies are typically restrictive in inclusion criteria and performed in a well-controlled environment, both of which cannot be said for the real-world setting.

In my PhD, I aim to address different sources of model bias and explore several ways to adapt and individualise treatment:

  1. To compensate for parameter bias on the population level, I work on approaches for sequential updating of prior information in clinical practice using hierarchical Bayesian methods.
  2. I intend to address the appropriateness of models by identifying structural model bias and developing methods to cope with it, using stochastic differential equations and machine learning methods.
  3. To offer a more holistic view on treatments, I plan to work on integrating side-effects, co-medications and long-term outcome into the modeling framework. Additionally, my PhD focuses on the general application of data assimilation and reinforcement learning methods in MIPD, interpretability of results and adaptive adjustment of treatment strategies.


Please see the list of all publications and PhD theses.


  • 08/2021: Entering PharMetrX
  • 10/2018-08/2021: Master of Science Bioinformatics, University of Potsdam
  • 10/2013-09/2018: Bachelor of Science Biology, Humboldt Universität zu Berlin