Project overview

Population pharmacokinetic (PopPK) modeling is a cornerstone of modern drug development and clinical pharmacology, but current approaches rely heavily on manual model specification, iterative optimization, and expert-driven trial-and-error. Classical estimation methods such as SAEM or FOCE-I can require substantial computational time and specialized expertise, limiting accessibility.

This doctoral project aims to continue developping AutoPopPK, an AI-assisted framework for the automated inference of population pharmacokinetic models directly from concentration–time data. The project uses modern deep learning architectures combined with mechanistic pharmacokinetic models to infer both individual-level and population-level PK parameters, including inter-individual variability, in a very fast, robust, and interpretable manner. 

Current state of the project: a neural network architecture has been developed in python (pyTorch). The model was trained to perform multiple tasks including the prediction of concentration–time profiles and the inference of population-level parameters for a simple pharmacokinetic model (one-compartment model with first-order absorption and elimination). The model demonstrates robust performance across a wide range of conditions, including varying numbers of observations per subject, varying numbers of subjects per study, and observation noise levels of up to 20%.

Immediate challenges:

• Develop a strategy which may involve adapting the neural network architecture and loss function such that the model is able to :
  • recognize various structural models (1, 2 or 3-compartment models)
  • handle multiple dose administrations per subjects
  • recognize delays in drug absorption
  • handle concentration measures that are below the level of quantification
  • be robust to outliers
• Perform model validation using different sources of real-world population pharmacokinetics data

Scientific objectives

The central objective of this PhD is to rethink population PK modeling as an inference problem solvable through learned amortized inference, while preserving the interpretability and extrapolability of classical pharmacometric models.

Specific research objectives include:

1. Design of hierarchical AI architectures for PopPK inference
   • Develop neural architectures capable of learning study-level and subject-level pharmacokinetic information (e.g., typical values, individual deviations, variability structures).
2. Integration of mechanistic PK structure into machine learning models
   • Embed classical pharmacokinetic models (e.g., compartmental ODE systems) within neural inference frameworks.
   • Ensure that inferred parameters correspond to interpretable PK quantities (kₐ, CL, Vc, Ω), enabling scientific interpretation and extrapolation.
3. Automated and scalable inference
   • Replace iterative likelihood-based estimation (e.g., SAEM, FOCE-I) with amortized inference, where parameter estimation is achieved through a single forward evaluation of a trained model.
   • Quantify gains in computational efficiency and robustness relative to classical PopPK workflows.
4. Uncertainty quantification and diagnostics
   • Develop methods to estimate parameter uncertainty, identify influential observations and potential outliers, and generate diagnostic plots analogous to those used in standard PopPK practice.
   • Compare AI-based uncertainty estimates with classical inference results.
5. Validation on simulated and real-world datasets
   • Benchmark AutoPopPK against established PopPK tools using controlled simulation studies.
   • Apply the framework to real clinical datasets to assess generalizability and practical relevance.

Expected contributions

This project is expected to contribute:

• A novel paradigm for population PK modeling, shifting from iterative optimization to fast, learned inference.
• A bridge between machine learning and pharmacometrics, combining interpretability and mechanistic structure with modern AI techniques.
• A practical tool for non-experts, with the long-term vision of an application where a standard NONMEM-style dataset serves as input and the output consists of estimated parameters, uncertainty measures, outlier detection, and diagnostic plots.
• Methodological advances relevant to broader applications in quantitative systems pharmacology and biomedical modeling.

Why this is a PhD-level project

This project goes well beyond software development. It involves:

• fundamental questions about identifiability, inference, and uncertainty in hierarchical time-series models,
• methodological innovation at the intersection of statistics, pharmacometrics, and machine learning,
• and extensive validation against established scientific standards.

The work is expected to result in multiple peer-reviewed publications in pharmacometrics, quantitative pharmacology, and applied machine learning.

Project overview

This doctoral project aims to develop a quantitative, systems-level modeling framework to investigate the role of Epstein–Barr virus (EBV) infection in the initiation and progression of multiple sclerosis (MS). Strong epidemiological evidence supports a causal link between EBV exposure and MS, yet the mechanistic pathways connecting viral persistence, immune dysregulation, and neuroinflammation remain incompletely understood. This project seeks to bridge this gap by constructing mechanistically grounded mathematical models that link biological processes across multiple scales.

Scientific objectives

The primary objective of the project is to formulate and analyze mechanistic models describing the dynamic interactions between EBV, immune cell populations (such as B cells, T cells and microglia), and inflammatory processes implicated in MS. These models will be designed to capture both within-host viral dynamics and system-level immune responses, while accounting for inter-individual variability and uncertainty.

Specific objectives include:

• Developing differential equation–based models to describe EBV latency, immune surveillance, and immune-mediated tissue damage.
• Integrating longitudinal clinical and biological data to inform and constrain model parameters.
• Investigating how variability in immune response or viral control may influence MS risk, disease progression, or relapse dynamics (analysis of trajectories).
• Using the models to explore hypothetical intervention scenarios, such as antiviral or immunomodulatory strategies, and to assess their potential impact on disease trajectories.

Methodological approach

The project will rely on mathematical modeling techniques commonly used in systems biology and pharmacology, including systems of ordinary and stochastic differential equations. Parameter estimation, uncertainty quantification, as well as dynamical and sensitivity analyses will be employed to link models to data and to identify key drivers of system behavior.

The modeling framework will emphasize biological interpretability and mechanistic consistency, ensuring that model components correspond to identifiable biological processes. Rather than focusing on prediction alone, the models will be used to generate testable hypotheses and to provide insight into the causal mechanisms underlying disease development.

Expected outcomes and impact

This project is expected to deliver:

• A mechanistic modeling framework for studying EBV–immune system interactions in the context of MS.
• Quantitative insights into the biological processes linking viral persistence and immune dysregulation to MS pathogenesis.
• A virtual experimentation platform to support hypothesis testing and exploration of therapeutic strategies.
• Generalizable modeling approaches applicable to other immune-mediated or infectious diseases.

By advancing mechanistic understanding of MS through quantitative modeling, this project will contribute to fundamental knowledge in neuroimmunology and support the rational design of future experimental and clinical studies.