MACSYS at QUT is pleased to welcome Johann Fredrik Jadebeck.  Johann is a researcher at Forschungszentrum Jülich (IBG-1) working at the intersection of systems biology, Bayesian inference, and scientific computing. His work also emphasises reproducible, high-performance computing (HPC) pipelines that bring modern probabilistic inference to mechanistic biological models.

Title: Jump, Diffuse, Average: Bayesian Computation for Metabolic Flux Inference under Model Uncertainty

Abstract:

Metabolic network models are essential for understanding the inner workings of microorganisms and have a wide range of applications in systems biology and metabolic engineering. These models represent biochemical knowledge, such as reaction stoichiometry. Mathematical models derived from this information, along with experimental data, enable the quantification of in vivo reaction rates, or fluxes [1]. Knowledge of these fluxes generates insights that facilitate everything from the rational design of efficient bioprocesses to the precise characterization of drug effects on human metabolism.

Yet metabolic flux inference is challenging: models have different scopes and variants, datasets are scarce and often only weakly informative about the “appropriate” model structure, while the model-associated flux parameter spaces are high-dimensional with a complex shape. In addition, the likelihood function may have multiple distinct modes. In this situation, a Bayesian approach is particularly attractive because it enables the propagation of uncertainty from the data via the models to the inferences. A method that integrates all plausible model structures is Bayesian model averaging (BMA) [2].

We present the computational and methodological tools necessary for metabolic flux inference with BMA. This includes all steps from model preprocessing to MCMC sampling within high-dimensional, linearly constrained flux spaces. It also includes reversible-jump MCMC [3] for sampling model variants and trans-dimensional diffusive nested sampling [4] for efficient sampling and evidence estimation, which is needed for BMA. Specifically, we examine how sensitive the inferences are to prior assumptions, motivating the development of physiological priors for metabolic network models. All concepts and algorithms are demonstrated with real-world inference tasks.

References

• [1] Niedenführ, S., Wiechert, W., & Nöh, K. (2015). Current Opinion in Biotechnology, 34, 82–90. https://doi.org/10.1016/j.copbio.2014.12.003
• [2] Hoeting, J. A., Madigan, D., Raftery, A. E., & Volinsky, C. T. (1999). Statistical Science, 14(4), 382–417. https://doi.org/10.1214/ss/1009212519
• [3] Green, P. J. (1995). Biometrika, 82(4), 711–732. https://doi.org/10.1093/biomet/82.4.711
• [4] Brewer, B. J., Pártay, L. B., & Csányi, G. (2011). Statistics and Computing, 21, 649–656. https://doi.org/10.1007/s11222-010-9198-8

Online details:

Zoom: https://qut.zoom.us/j/89951328887?pwd=4ayunK21H8OkVRsbVAmq7FLWZ15Hbb.1

Meeting ID: 899 5132 8887

Passcode: 002005