title: Coding the BYM2 model for disconnected graphs in Stan or how I stopped worrying and learned to love PC priors
In A note on intrinsic Conditional Autoregressive models for disconnected graphs, Freni-Sterrantino et.al. show how to implement the BYM2 model for use with areal data where the graph structure of the map is not fully connected. In this talk, we present an implementation of this proposal in Stan which illustrates use of the Penalized Complexity framework. Finally, we discuss the challenges of extending this to more complicated spatio-temporal data.
time zone: -5 GMT. (currently US Eastern Daylight Time)
Hi @maxbiostat. Thanks for the offer, but I’m actually now working outside of academia and health research, and so wouldn’t have the time to focus on / contribute to this. Best of luck with pulling it together though!
The power prior (Chen and Ibrahim, 2000) enables practitioners to utilize historical data to obtain a flexible informative prior. However, the prior may be sensitive to hyperparameter specification. Typically, such sensitivity is eliminated by treating the hyperparameter as random, that is, utilizing a hierarchical model. Unfortunately, for the vast majority of models, the normalizing constant of the power prior is unavailable.
In this talk, we present computational techniques to estimate normalized power prior (NPP) models using Stan. Specifically, we utilize a two-stage process that first estimates the normalizing constant for a wide range of values, and then utilizes a second round of sampling to obtain samples from the posterior density using the normalized power prior. The methodology and software will be invaluable to practitioners and regulators who wish to incorporate prior knowledge on the basis of historical data, but are skeptical of tuning parameters.
I’m bumping this because we have not received many abstracts so far. Bear in mind this does not need to be conference-grade work: submissions of partial work or promising ideas is more than welcome too.
Bayesian modelling of organoid growth: In cystic fibrosis research (and presumably elsewhere) a common way to asses organoid growth is to compute area under the growth curve and compre this value between groups. It however turns out that using a simple hierarchical linear model on the logarithm of the size of the organoids adjusting for baseline size fits the data much better and provides more interpretable results.
Title: Summarising enzyme information from online databases using Stan and Arviz
Authors: Teddy Groves and Areti Tsigkinopoulou
Time Zone: Copenhagen (currently GMT+2)
Many systems biology applications require information about the kinetic behaviour of metabolic enzymes, which can be found in databases like BRENDA, which records measurements of kinetic parameters for a wide range of organisms, enzymes and experimental conditions. However, it is not straightforward to use this information in kinetic models as the raw measurements are often scarce, inconsistent and difficult to aggregate. For example, if a certain enzyme/substrate combination’s Michaelis constant has only been measured once, but additional measurements are available for the same substrate and a similar enzyme (say with the same first three digits of the EC number), it is not straightforward to say what is the best guess as to the true value of the
original measured quantity.
Our analysis addresses the problem of summarising the information in BRENDA about kinetic parameters using a nested multilevel statistical model implemented in Stan. We demonstrate how representing structural information like the enzyme classification number and substrate improves the model’s predictive accuracy. We present our model’s results in a web application which we hope will allow systems biologists to more easily use the BRENDA data.
Title: Automated kinetic modelling in Stan and its application to the methionine cycle
Authors: Nicholas Cowie and Teddy Groves
Time Zone: Copenhagen (GMT +2/1 depending on daylight savings)
The methionine cycle is a fundamental pathway in human metabolism. Its intermediates participate in a variety of mechanisms competing for the same resources. These functions all occur simultaneously resulting in a highly regulated pathway with approximately 6 allosteric effectors. In order to understand the operation of the methionine cycle we constructed a kinetic model as a system of ODEs, where the state variables are the metabolite concentrations and the fluxes are constrained by mechanistic rate laws. However, these systems are described as ‘sloppy parameter’ systems, where the marginal parameter distributions can be large, yet still display tight posterior predictions. Furthermore, measuring these parameters independently is obscured by measuring the system in in vitro conditions, as opposed to those experienced in vivo. This prevents the standard practice of placing tight priors on the parameter values.
Our software automates network construction, and conducts parameter sampling in Stan conditioned on in vivo measurements of metabolite concentrations, fluxes, and protein concentrations. The resultant posterior draws take advantage of the HMC algorithm to efficiently sample the highly correlated parameter space and generate sets that describe the kinetic model equally well, despite their broad marginal distributions. This approach is applied to the methionine cycle as an ideal case of a highly regulated and non-linear pathway.
Just want to make totally sure, since we have such a crazy time system in the US:
10am EST is 11am EDT, so in strict terms there was never any conflict in the above.
@maxbiostat, I assume that in fact the start is at 10am EDT (GMT -4). If, on the other hand, you really did mean 10am EST above (GMT -5), then that’s 11am Eastern Time (which until 7 November means EDT = GMT - 4).