MBI Videos

Joined-up modelling in ecology (and biosciences in general!): likelihood, Bayesian, and all that

  • video photo
    Drew Purves

    Models have long been of interest in the biological sciences, but traditionally very different kinds of models have been used for different purposes: simple models of ecological ideas to generate general insights, complex simulation models to explore whether such insights might stand up in the face of reality, and simple statistical models (often referred to as ‘methods’ or ‘routines’ – think ANOVA, PCA) were used to extract conclusions from data. And although all three kinds of models have occasionally been used to make predictions (for example, about dynamics into the future, or the response of something to an intervention), none are really suitable for this purpose because they either lack constraints from data (the first two kinds of model) or are unlikely to capture the non-linear dynamics observed in reality (the third kind). But thanks to relatively recent advances in computing power, data availability and computational statistics, it is now becoming possible to pursue a ‘joined up’ approach to modelling all across the biological sciences. This approach merges ideas-rich models with data via Likelihood / Bayesian approaches, to both increase our understanding of nature, and to begin to enable us to predict it. 


    In the first hour of this virtual seminar, Dr. Drew Purves will give an extended overview of the above, introducing the notion of Likelihood and Bayesian approaches, and try to provide an honest overview of the pros and cons of these approaches. Dr. Purves will explain ‘Metropolis Hastings MCMC sampling’, why it’s exciting (it is – really!) and introduce a particular sampler called ‘Filzbach’ that his group has developed, giving examples of the environmental science that it has enabled in his group (e.g. modelling the global carbon cycle) and in their sister group (e.g. modelling MHC Class I in the immune system, synthetic biology). Since many of you will be interested in environmental problems, Dr. Purves will also provide a tour of ‘FetchClimate’ an unnamed prototype browser-based tool for doing the whole data-model-predictions pipeline (available here if you want to try before hand: http://fetchclimate.cloudapp.net/ ). 


    In the second half, Dr. Mark Vanderwel, formerly a postdoc at MSRC in Cambridge and now at the University of Florida, will take you through every step of some example analyses using ‘FilzbachR’, which allows you to run your whole analysis from R, and even write the model itself in R. This seminar assumes some basic knowledge in R. By the end of the session you will feel equipped enough to at least try this approach on your own problems.

  • video photo
    Mark Vanderwel

    Models have long been of interest in the biological sciences, but traditionally very different kinds of models have been used for different purposes: simple models of ecological ideas to generate general insights, complex simulation models to explore whether such insights might stand up in the face of reality, and simple statistical models (often referred to as ‘methods’ or ‘routines’ – think ANOVA, PCA) were used to extract conclusions from data. And although all three kinds of models have occasionally been used to make predictions (for example, about dynamics into the future, or the response of something to an intervention), none are really suitable for this purpose because they either lack constraints from data (the first two kinds of model) or are unlikely to capture the non-linear dynamics observed in reality (the third kind). But thanks to relatively recent advances in computing power, data availability and computational statistics, it is now becoming possible to pursue a ‘joined up’ approach to modelling all across the biological sciences. This approach merges ideas-rich models with data via Likelihood / Bayesian approaches, to both increase our understanding of nature, and to begin to enable us to predict it. 


    In the first hour of this virtual seminar, Dr. Drew Purves will give an extended overview of the above, introducing the notion of Likelihood and Bayesian approaches, and try to provide an honest overview of the pros and cons of these approaches. Dr. Purves will explain ‘Metropolis Hastings MCMC sampling’, why it’s exciting (it is – really!) and introduce a particular sampler called ‘Filzbach’ that his group has developed, giving examples of the environmental science that it has enabled in his group (e.g. modelling the global carbon cycle) and in their sister group (e.g. modelling MHC Class I in the immune system, synthetic biology). Since many of you will be interested in environmental problems, Dr. Purves will also provide a tour of ‘FetchClimate’ an unnamed prototype browser-based tool for doing the whole data-model-predictions pipeline (available here if you want to try before hand: http://fetchclimate.cloudapp.net/ ). 


    In the second half, Dr. Mark Vanderwel, formerly a postdoc at MSRC in Cambridge and now at the University of Florida, will take you through every step of some example analyses using ‘FilzbachR’, which allows you to run your whole analysis from R, and even write the model itself in R. This seminar assumes some basic knowledge in R. By the end of the session you will feel equipped enough to at least try this approach on your own problems.

View Videos By