We plan to develop new methods for the prediction of phenotypic traits from genetic data. Such traits can include the likelihood of specific inherited diseases as well as other highly heritable, quantitative traits like the human body size. The aim is to design and train machine learning methods that surpass the current state of the art at least for some genetic architectures.
Genome-wide association studies (GWAS) have identified large sets of associated SNPs in the last decades. For highly heritable traits accurate phenotype predictions are theoretically possible, given enough data. In practice, however, predictions are currently not as accurate as expected if the trait is complex (i.e. it depends on a large number of genetic variants). This is known as the problem of the missing heritability.
As a (partial) solution to the problem, we propose the application of custom models, that consider (higher-order) interactions between (possibly large numbers of) genetic variants. Current state of the art models usually do not do that.
If accurate and robust prediction models can be found, clinical tools that assist doctors in long term prediction of inherited diseases and enable treatments at very early stages can be developed.
The project duration is scheduled for about 3 years.
