Aims:

1. Utilise the Sparse Instrumental Variable approach (SPIV) to identify combinations of genetic varinats for use as instrumental variables in Mendelian Randomization studies.

2. Construct sparse networks of associations between outcomes using the sparse random field models.

3. Use Mendelian Randomisation approaches to investigate the causal direction of the network edges.

4. Investigate robustness of the resulting structures for multiple variable selection criteria and data subsamples.

Agakov and colleagues have been developing machine learning based methods for selecting meaningful associations both in phenotypic data and, using modifications of Mendelian Randomization approaches, in data combining genotypes and phenotypes. These variable selection methods will be applied to a network of phenotypes and genome wide-data predicting such phenotypes. The methods are based on probabilistic sparse latent variable models applied either for feature selection and structure learning in layered directed networks [1, 2] or more complex networks of associations [3]. The variables in this exploration of these methods will be: BMI, blood pressure, IL6, CRP, LDL cholesterol, HDL cholesterol, triglycerides, leptin, WISC score and adiponectin.

The analysis will either be conducted by Louise Millard (who will visit Edinburgh for training) or by a member of the Edinburgh group visiting CAiTE.

[1] F. V. Agakov, P. McKeigue, J. Krohn, A. Storkey. Sparse Instrumental Variables (SPIV) for Genome-Wide Studies. In Advances in Neural Information Processing Systems 23, 2010.

[2] F. V. Agakov, P. McKeigue, J. Krohn, J. Flint. Inference of Causal Relationships between Biomarkers and Outcomes in High Dimensions, Journal of Systemics, Cybernetics and Informatics 9(6), 2011.

[3] F. V. Agakov, P. Orchard, A. Storkey. Discriminative Mixtures of Sparse Latent Fields for Risk Management, Journal of Machine Learning Research W&CP 22, 2012.