Simulation Package for Efficient Experimental Design and Inference in Microbiology

There is growing recognition within biological sciences that mathematical modelling provides powerful methods to improve our quantitative understanding of the dynamics of biological systems. Harnessing the full potential of these methods requires a complete integration of experimental data and dynamic models within the proper statistical framework. However, progress in this area has been patchy: while some fields of biology (systems biology) lead the way, others are lagging behind. Our proposal aims to develop and deliver a free computational package that will facilitate the complete integration of dynamic models and laboratory experiments, with an initial focus on research into host-pathogen interactions.
This open-source software will have two related and essential functions:
- statistical inference (SI): given a mechanistic model combining current knowledge and hypotheses about a biological system, how much information can be extracted from new experimental data about mechanisms that cannot be directly observed?
- Optimal experimental design (OED): given a mechanistic model and preliminary data, what is the best way to design an experiment within set (budgetary or technical) constraints in order to maximise the expected gain of information?
Recent progress in scientific computing has allowed the rapid development of algorithms for SI and OED, but they have been applied independently to other areas of research. Our project will deliver the first "one-stop shop" for inter-disciplinary research projects in microbiology. We will use state-of-the-art methods from applied statistics and tailor them to the specific needs of experimental biologists. An important novelty will be our focus on stochastic simulations, which allow random variations in the dynamics of a system: as in experiments with living organisms, repeats of the same procedure never yield exactly the same results. Because they capture this essential feature of real systems, stochastic models allow more reliable and accurate inference, albeit at the cost of greater computational complexity. Our many years of expertise at the interface of statistical modelling and experimental biology put us in a very strong position to tackle these challenges.
This 18-month project will enable us to develop and test the functionality of the package with two experimental systems using existing and new data, before releasing it for free and public use in inter-disciplinary biological research. The software will be delivered as a package for use within the R software, which is a free statistical platform.

We will develop Bayesian statistical tools for statistical inference (SI) and optimal experimental design (OED) for stochastic mechanistic models. Various techniques have been proposed in recent years and applied to physical sciences and systems biology. We will use our expertise in mathematical modelling and Bayesian statistical inference in microbiology to bring together a set of tools that can work together and provide useful information for experimental microbiologists.
Most existing methods for Bayesian SI with stochastic models are based on so-called Approximate Bayesian Computation (ABC), which substitute likelihood with summary statistics in order to identify simulations that are most similar to the data. Although computationally efficient, these methods can provide incomplete or even unreliable predictions if the summary statistics chosen are not sufficient. Therefore, we will also propose an alternative approach which combines stochastic simulations and observational noise to approximate the likelihood through importance sampling.
We will exploit recent progress in Bayesian OED to build an efficient tool that can be combined with our stochastic inference methods. Although the use of stochastic mechanistic models within OED has been considered theoretically (e.g. in systems biology), to our knowledge it has not been demonstrated in practice. This will be our greatest challenge, but one we are ideally placed to address. Together with our collaborators, our joint expertise in statistical modelling and experimental biology will enable us to concentrate on very specific specifications, that can then be broadened for applications to other fields.
Testing of our software will be performed with two experimental systems as part of ongoing research projects led by our close collaborators: a murine model of typhoid infection, and an invertebrate model of enteric bacterial symbiosis.

There are growing ethical concerns surrounding the use of animals in scientific research. Caught between the imperatives of animal welfare and translational research, biologists need tools and practical guidance to pursue scientific research that contributes to our health, knowledge and well-being at a cost that society can accept to bear. The 3Rs (replace, reduce, refine) provide a framework for animal research without imposing targets, encouraging the development of innovative approaches. In areas where replacement of animals is not yet possible or needs validation, scientists are required to combine state-of-the-art experimental techniques with mathematical and statistical models in order to reduce and refine the use of animals. Traditional statistical tools for experimental design, such as power calculation, have been designed to predict simple linear relationships between experimental factors and observed variables. In living systems however, non-linear responses mediated by complex processes are the norm. Mechanistic models aim to reproduce parts of these complex processes in order to generate accurate predictions of the system's response to a given experimental treatment. Developing, refining and validating these mechanistic models are essential steps towards replacement of animals in experimental research. Our package will facilitate this process by providing scientists with the tools they need to perform two crucial statistical operations: inference and experimental design. Although the theory underlying these operations in the context of mechanistic models is well established, their practical implementation for real-life problems was until recently hindered by computational limitations. These challenges have started to be overcome in the last ten years, and we are now in an ideal situation to translate the latest progress in applied statistical research into practical solutions for life scientists.
By making our package available for free in the R environment, and enabling anyone to reuse, adapt and modify it under an Open Source General Public License (www.opensource.org), we hope to be able to reach a broad scientific community. Thus our package will contribute to equipping experimental scientists with the means to carry out high-impact research within the 3Rs framework.