Combining different sources of information to optimise genomic prediction of complex traits

Athina Spiliopoulou, Reka Nagy, Jennifer E Huffman, Mairead Bermingham, Caroline Hayward, Igor Rudan, Harry Campbell, Alan Wright, Jim Wilson, Ricardo Pong-Wong, Chris Haley, Felix Agakov, Pau Navarro

Research output: Contribution to conferencePosterpeer-review

Abstract / Description of output

In the last decade there has been substantial progress in identifying genetic loci associated with complex phenotypes but limited progress in using genomic information to predict phenotypic performance. In this study, we assess the advantage of using previously published results to inform genomic predictors of complex traits. Our goal is to compare model performance with respect to trait architecture and latent population structure. We consider linear, additive models with different sparsity levels, either learned from the data using lasso or elastic nets, or determined a priori by considering markers and their corresponding effects from large meta-analysis association studies. We characterise the predictive signal captured by each model and explore whether prediction accuracy can be increased by combining these simpler predictors into a meta-model.

We evaluate predictive performance using height, body mass index and high density lipoproteins in two population cohorts, originating in Croatia and Scotland. We examine how to maximise prediction accuracy when the target individuals come from the same or a different population to the training samples. Our results demonstrate that between population prediction is possible using samples from the target population to perform model selection, subject to sample size and trait architecture. Furthermore, we show that a model combining the predictions from penalised regression with meta-analysis-based polygenic scores performs better than either model on its own. Our findings suggest that incorporating previous results into statistical models as well as exploiting the predictive signal from latent data structure can lead to improved predictions of complex traits.
Original languageEnglish
Publication statusPublished - 1 Jun 2014
EventThe European Human Genetics Conference, ESHG 2014 - Milano Congressi, Viale Eginardo (Gate 2), Milan, Italy
Duration: 31 May 20143 Jun 2014


ConferenceThe European Human Genetics Conference, ESHG 2014


Dive into the research topics of 'Combining different sources of information to optimise genomic prediction of complex traits'. Together they form a unique fingerprint.

Cite this