Research in our group is centered around two themes, with the second being a means to achieve the first.
Understanding the genetic architecture and improving prediction accuracy of complex traits.
The technological advances in terms of genotyping and sequencing that started in the early 2000s have revolutionized quantitative genetics. In particular, Genome-Wide Association Studies (GWAS) have enabled the discovery of many molecular variants associated with a plethora of complex traits in humans and model species. However, many challenges still remain. We are particularly interested in addressing the following:
(a) The associated variants only explain a small proportion of the total genetic variance for most complex traits.
(b) Little is known about context-dependent effects as the vast majority of GWAS have focused only on additive effects of genetic variants.
(c) Phenotype prediction accuracy from genotypes is low for most complex traits.
(d) A lot remains to be understood about the mechanisms by which genetic variants influence complex traits by affecting intermediate phenotypes (e.g., gene expression levels, metabolite levels).
Our research has provided new insights into all these issues. For example, we have shown that it is fundamental to account for the genetic architecture (including non-additive gene action) to achieve high phenotypic prediction accuracy from genotypes (Morgante et al., 2018). We have also shown that including multiple layers of information (i.e., transcriptomic, metabolomic, and functional annotation) into prediction models is an effective way to increase accuracy, compared with using only genotypes. This is particularly useful when large sample sizes are not available (Morgante et al., 2020; Zhou et al., 2020). We have also contributed to studies that derived networks connecting genetic variants with gene expression and metabolite variation (Everett et al., 2020; Zhou et al., 2020).
Developing statistical genetic methods.
For many applications, new statistical methods are either needed or can provide additional insights and/or better performance than existing methods. Thus, our group is also interested in developing statistical methods for genetic applications. Recently, we have developed mr.mash, a flexible and efficient method for Bayesian multivariate, multiple regression that can learn patterns of effect sharing across responses from the data. We showed that mr.mash can achieve higher accuracy than existing methods in the task of predicting multi-tissue gene expression from genotypes (Morgante et al., 2023). Although it was developed with this genetic application in mind, mr.mash is suitable for any application where multivariate multiple regression is appropriate. mr.mash is distributed as an R package available at https://github.com/stephenslab/mr.mash.alpha.
