Multivariate dissection of quantitative trait gene networks
Lead Research Organisation:
University of Oxford
Department Name: Wellcome Trust Centre for Human Genetics
Abstract
Common diseases such as asthma, anxiety, depression, heart disease, susceptibility to diabetes and obesity have a complex basis. An individual‘s susceptibility to each of these diseases is the result of multiple genetic and environmental factors combining together. A key goal of medical research is to dissect the genetic, biochemical, physiological and environmental causes of these susceptibilities in order to understand how the disease has arisen, ultimately with a view to developing new treatments. The typical approach to identifying a genetic basis is to correlate single measures of disease state (e.g. level of anxiety) with the observed genetic variation among subjects. However, this approach too often fails to exploit the fact that many diseases or apparently different characteristics relating to the same disease are not independent but are themselves correlated because of a shared genetic, biochemical or physiological basis, or because they are linked in a causal chain. The full picture of the disease state, and therefore the most complete way to infer its basis, is accessible only through the analysis of all relevant measures simultaneously. I will develop statistical methods to accomplish this for a rodent model of anxiety using measurements taken from a large population of outbred mice.
Technical Summary
I propose to develop statistical methods for the multivariate analysis of disease phenotypes. In particular I will investigate the genetic architecture of anxiety and to explore how phenotypic measures of anxiety in rodents are related. My work will be based around the multivariate analysis of multiple correlated anxiety traits in a set of 2,000 heterogeneous stock (HS) mice. Specifically I aim to:
1. Develop multivariate analytical techniques to increase the power to detect and fine-map quantitative trait loci (QTLs). To achieve this I will adapt multivariate methods including Partial Least Squares Regression, commonly used in chemometrics, to the problem of mapping QTLs in genetically outbred animals. I will focus on using the existing dataset of HS, simulated data from the related Collaborative Cross (CC) design, and on real data from the CC as it becomes available. I will develop and test these methods in collaboration with Dr Gary Churchill at the Jackson Laboratory.
2. Test the causal relationship between phenotypes and their detected genetic correlates using probabilistic graphical networks based on structural equation modelling (SEM). I will train in SEM methodology at Virginia Commonwealth University under Prof Michael Neale, developer of the SEM language Mx and lead statistician on the Virginia Adult Twin Study of Psychiatric and Substance Use Disorders.
3. Refine hypotheses about candidate genes by incorporating transcript data. In collaboration with Dr Chris Holmes (Oxford), I will develop Bayesian SEMs to test the causal relationship between transcripts colocalizing with QTLs defined for multivariate phenotypes in outbred animals.
4. Use the methods above to infer plausible models for the causal network underpinning anxiety and characterize the relative uncertainty of these competing models. Many modelling strategies involve choosing between competing models that explain the observed data equally well. A more robust approach is to explicitly account for model uncertainty through model averaging techniques. In collaboration with Dr Holmes, I will build on my previous experience in this area, using both Bayesian and frequentist approaches to identify those relationships in the anxiety network that are robust to perturbations in the data or minor changes in model formulation, and are thus most likely to represent biological targets suitable for further investigation.
Methods developed during the project, although primarily applied to the HS and CC designs, will be generalizable to a variety of other mapping populations and model organisms.
1. Develop multivariate analytical techniques to increase the power to detect and fine-map quantitative trait loci (QTLs). To achieve this I will adapt multivariate methods including Partial Least Squares Regression, commonly used in chemometrics, to the problem of mapping QTLs in genetically outbred animals. I will focus on using the existing dataset of HS, simulated data from the related Collaborative Cross (CC) design, and on real data from the CC as it becomes available. I will develop and test these methods in collaboration with Dr Gary Churchill at the Jackson Laboratory.
2. Test the causal relationship between phenotypes and their detected genetic correlates using probabilistic graphical networks based on structural equation modelling (SEM). I will train in SEM methodology at Virginia Commonwealth University under Prof Michael Neale, developer of the SEM language Mx and lead statistician on the Virginia Adult Twin Study of Psychiatric and Substance Use Disorders.
3. Refine hypotheses about candidate genes by incorporating transcript data. In collaboration with Dr Chris Holmes (Oxford), I will develop Bayesian SEMs to test the causal relationship between transcripts colocalizing with QTLs defined for multivariate phenotypes in outbred animals.
4. Use the methods above to infer plausible models for the causal network underpinning anxiety and characterize the relative uncertainty of these competing models. Many modelling strategies involve choosing between competing models that explain the observed data equally well. A more robust approach is to explicitly account for model uncertainty through model averaging techniques. In collaboration with Dr Holmes, I will build on my previous experience in this area, using both Bayesian and frequentist approaches to identify those relationships in the anxiety network that are robust to perturbations in the data or minor changes in model formulation, and are thus most likely to represent biological targets suitable for further investigation.
Methods developed during the project, although primarily applied to the HS and CC designs, will be generalizable to a variety of other mapping populations and model organisms.