Comprehensive benchmarking of Markov chain Monte Carlo methods for dynamical systems.

BACKGROUND: In quantitative biology, mathematical models are used to describe and analyze biological processes. The parameters of these models are usually unknown and need to be estimated from experimental data using statistical methods. In particular, Markov chain Monte Carlo (MCMC) methods have become increasingly popular as they allow for a rigorous analysis of parameter and prediction uncertainties without the need for assuming parameter identifiability or removing non-identifiable parameters. A broad spectrum of MCMC algorithms have been proposed, including single- and multi-chain approaches. However, selecting and tuning sampling algorithms suited for a given problem remains challenging and a comprehensive comparison of different methods is so far not available. RESULTS: We present the results of a thorough benchmarking of state-of-the-art single- and multi-chain sampling methods, including Adaptive Metropolis, Delayed Rejection Adaptive Metropolis, Metropolis adjusted Langevin algorithm, Parallel Tempering and Parallel Hierarchical Sampling. Different initialization and adaptation schemes are considered. To ensure a comprehensive and fair comparison, we consider problems with a range of features such as bifurcations, periodical orbits, multistability of steady-state solutions and chaotic regimes. These problem properties give rise to various posterior distributions including uni- and multi-modal distributions and non-normally distributed mode tails. For an objective comparison, we developed a pipeline for the semi-automatic comparison of sampling results. CONCLUSION: The comparison of MCMC algorithms, initialization and adaptation schemes revealed that overall multi-chain algorithms perform better than single-chain algorithms. In some cases this performance can be further increased by using a preceding multi-start local optimization scheme. These results can inform the selection of sampling methods and the benchmark collection can serve for the evaluation of new algorithms. Furthermore, our results confirm the need to address exploration quality of MCMC chains before applying the commonly used quality measure of effective sample size to prevent false analysis conclusions.

Using Bayesian-PBPK modeling for assessment of inter-individual variability and subgroup stratification.

PURPOSE: Inter-individual variability in clinical endpoints and occurrence of potentially severe adverse effects represent an enormous challenge in drug development at all phases of (pre-)clinical research. To ensure patient safety it is important to identify adverse events or critical subgroups within the population as early as possible. Hence, a comprehensive understanding of the processes governing pharmacokinetics and pharmacodynamics is of utmost importance. In this paper we combine Bayesian statistics with detailed mechanistic physiologically-based pharmacokinetic (PBPK) models. On the example of pravastatin we demonstrate that this combination provides a powerful tool to investigate inter-individual variability in groups of patients and to identify clinically relevant homogenous subgroups in an unsupervised approach. Since PBPK models allow the identification of physiological, drug-specific and genotype-specific knowledge separately, our approach supports knowledge-based extrapolation to other drugs or populations. METHODS: PBPK models are based on generic distribution models and extensive collections of physiological parameters and allow a mechanistic investigation of drug distribution and drug action. To systematically account for parameter variability within patient populations, a Bayesian-PBPK approach is developed rigorously quantifying the probability of a parameter given the amount of information contained in the measured data. Since these parameter distributions are high-dimensional, a Markov chain Monte Carlo algorithm is used, where the physiological and drug-specific parameters are considered in separate blocks. RESULTS: Considering pravastatin pharmacokinetics as an application example, Bayesian-PBPK is used to investigate inter-individual variability in a cohort of 10 patients. Correlation analyses infer structural information about the PBPK model. Moreover, homogeneous subpopulations are identified a posteriori by examining the parameter distributions, which can even be assigned to a polymorphism in the hepatic organ anion transporter OATP1B1. CONCLUSIONS: The presented Bayesian-PBPK approach systematically characterizes inter-individual variability within a population by updating prior knowledge about physiological parameters with new experimental data. Moreover, clinically relevant homogeneous subpopulations can be mechanistically identified. The large scale PBPK model separates physiological and drug-specific knowledge which allows, in combination with Bayesian approaches, the iterative assessment of specific populations by integrating information from several drugs.

Statistical analysis of chemical transformation kinetics using Markov-Chain Monte Carlo methods.

For the risk assessment of chemicals intentionally released into the environment, as, e.g., pesticides, it is indispensable to investigate their environmental fate. Main characteristics in this context are transformation rates and partitioning behavior. In most cases the relevant parameters are not directly measurable but are determined indirectly from experimentally determined concentrations in various environmental compartments. Usually this is done by fitting mathematical models, which are usually nonlinear, to the observed data and such deriving estimates of the parameter values. Statistical analysis is then used to judge the uncertainty of the estimates. Of particular interest in this context is the question whether degradation rates are significantly different from zero. Standard procedure is to use nonlinear least-squares methods to fit the models and to estimate the standard errors of the estimated parameters from Fisher's Information matrix and estimated level of measurement noise. This, however, frequently leads to counterintuitive results as the estimated probability distributions of the parameters based on local linearization of the optimized models are often too wide or at least differ significantly in shape from the real distribution. In this paper we identify the shortcoming of this procedure and propose a statistically valid approach based on Markov-Chain Monte Carlo sampling that is appropriate to determine the real probability distribution of model parameters. The effectiveness of this method is demonstrated on three data sets. Although it is generally applicable to different problems where model parameters are to be inferred, in the present case for simplicity we restrict the discussion to the evaluation of metabolic degradation of chemicals in soil. It is shown that the method is successfully applicable to problems of different complexity. We applied it to kinetic data from compounds with one and five metabolites. Additionally, using simulated data, it is shown that the MCMC method estimates the real probability distributions of parameters well and much better than the standard optimization approach.

An elastic network model to identify characteristic stress response genes.

Exposing eukaryotic cells to a toxic compound and subsequent gene expression profiling may allow the prediction of selected toxic effects based on changes in gene expression. This objective is complicated by the observation that compounds with different modes of toxicity cause similar changes in gene expression and that a global stress response affects many genes. We developed an elastic network model of global stress response with nodes representing genes which are connected by edges of graded coexpression. The expression of only few genes have to be known to model the global stress response of all but a few atypical responder genes. Those required genes and the atypical response genes are shown to be good biomarker for tox predictions. In total, 138 experiments and 13 different compounds were used to train models for different toxicity classes. The deduced biomarkers were shown to be biologically plausible. A neural network was trained to predict the toxic effects of compounds from profiling experiments. On a validation data set of 189 experiments with 16 different compounds the accuracy of the predictions was assessed: 14 out of 16 compounds have been classified correctly. Derivation of model based biomarkers through the elastic network approach can naturally be extended to other areas beyond toxicology since subtle signals against a broad response background are common in biological studies.