Leveraging pleiotropy in human genetic architecture by building a map of pleiotropy using machine learning – PleioMap

Although pleiotropy, which occurs when a genetic element has a causal effect on at least two traits, is thought to play a central role in the genetic architecture of complex traits and diseases, it is a poorly understood mechanism. Here, we reexamined known concepts of human genetics through the prism of pleiotropy and we formulated the 5-pleiotropy hypothesis. Indeed, observed pleiotropy which occurs when one genetic variant affects more than one trait can stem from 5 biological mechanisms: 1) LD pleiotropy (linkage disequilibrium) 2) vertical pleiotropy (causality) 3) network pleiotropy (genetic correlation) 4) serendipitous pleiotropy (polygenicity) 5) horizontal pleiotropy (independent effects).

Our global objective is to study pleiotropy and to model the 5 types of pleiotropy using publicly available summary statistics data stemming from genome-wide association studies (GWASs). More specifically, in Workpackage 1, we will build a comprehensive framework to disentangle between the 5 states of pleiotropy by modeling pleiotropy due to relationships between traits while quantifying pleiotropy at the level of genetic variants, providing in fine a genome-wide map of pleiotropy. Strategies to achieve this first goal include i) improving on proof-of-concept method; ii) rerouting existing methods used to model relationships between complex traits and diseases; iii) build a novel statistical framework based on machine learning algorithms, notably semi-supervised learning by using additional eQTL (expression Quantitative Trait Loci) data and a colocalization method to label genetic variants for pleiotropy.
In Workpackage 2.1, we propose to study relationships between complex traits and diseases stemming from pleiotropy. Special attention will be paid to disentangling between vertical & network pleiotropies, which stem from causal relationships between traits, and the other forms of pleiotropy. In addition, after applying PleioMap to many traits, we intend to develop visualization tools to build networks of relationships between complex traits and diseases.
In Workpackage 2.2, we will study the pleiotropic effects of the genetic variants themselves and make an inventory of 5 types of pleiotropy validating or invalidating our 5-pleiotropy model. By identifying pleiotropic effects of genetic variants which we think can be much weaker than the effects identified by traditional GWASs, we do hope to be able to provide updated heritability estimates for complex traits and diseases and contribute to improving fine-mapping.
The full code to produce the PleioMap, the PleioMap itself as well as the network of relationships between traits and diseases will be made publicly available as a resource to the scientific community.

The PleioMap project is ambitious and challenging but we strongly believe that this field of research is of interest for the human genetics community and will be thriving in the coming years. We do expect PleioMap to open new avenues for additional applications such as prediction of drug side effects or off-target effects in genome editing and will provide insights into new biological mechanisms behind the shared etiology of traits and diseases. Therefore, we do hope that PleioMap and the study of pleiotropy will trickle down to allow the development of novel preventive and therapeutic strategies and towards personalized medicine applications.