MetrumRG ACoP Workshop

This workshop provides a guided hands-on experience in performing efficient reproducible and traceable Bayesian PPK analyses. We will walk through both NONMEM-based and Stan-based workflows. Such workflows are facilitated by the use of MeRGE (Metrum Research Group Ecosystem) open-source tools that have been extended to support fully Bayesian analyses with either NONMEM or Stan. Sometimes you may want more flexibility than offered by NONMEM. We illustrate that by implementing a PPK model employing shrinkage priors for covariate effects using Stan/Torsten.

Presenters: William Gillespie, Ph.D.; Timothy Waterhouse, Ph.D.; Curtis Johnston, Pharm.D.; Seth Green, M.S.

1. Brief Review of Bayesian principles and methods: 

We will cover several key topics, including Bayes' Rule, the Bayesian modeling and inference process, computation for Bayesian modeling, and the key challenge of sampling from high-dimensional probability distributions in Bayesian modeling and inference. Additionally, the workshop will discuss the general computational approach of posterior simulation and provide a brief introduction to Markov chain Monte Carlo simulation.

2. Outline of Bayesian data analysis workflow: 

We will provide a structured overview of the steps involved in conducting Bayesian data analysis.

3. Overview of the MeRGE Tools: 

We will introduce the MeRGE tools, highlighting their functionality and how they support the integration of pharmacometric and systems pharmacology models. This overview will demonstrate how these tools facilitate efficient model development, analysis, and decision-making in complex biomedical research.

4. Data and base PPK model used for the examples: 

We will cover the specific datasets and the foundational population pharmacokinetic (PPK) model that will be used in the practical examples.

5. Bayesian workflow for PPK modeling using NONMEM:

We will provide an overview of how to apply Bayesian analysis to PPK modeling within NONMEM. This will start with an introduction to some conceptual issues such as prior selection and output interpretation.  It will then cover key steps including creating and submitting models, performing MCMC and model fitting diagnostics, and developing models iteratively, either by copying an FOCE model or starting from scratch.

6. Bayesian workflow for PPK modeling using Stan/Torsten: 

We will cover the Bayesian workflow for population pharmacokinetics (PPK) modeling using Stan and Torsten. We’ll start with creating and submitting a model, including initial model object creation and annotation. Next, we’ll summarize model outputs, perform MCMC diagnostics, and evaluate model fitting metrics. The workflow includes traceable updates, copying, and revising models. Enhancements such as adding correlated random effects, switching to non-centered parameterization, incorporating covariates, and using shrinkage priors like spike-and-slab priors will also be covered.

We will cover generated quantities models, which allow for simulations without refitting the data. Using generated quantities, we can perform leave-one-group-out cross-validation (LOGO-CV) to assess model performance. We’ll also discuss how to summarize modeling outputs using run_log and stan_summary_log to compare multiple models effectively.

7. Closing Discussion: 

We will address choosing the right tool for PPK modeling, comparing NONMEM with Stan/Torsten. We’ll provide additional resources and highlight MRG expos for further exploration. Some topics not covered include models requiring numerical solutions of ODEs and within-chain parallel computation with Stan/Torsten. Finally, we’ll preview upcoming features and developments in the field.