Deep Learning for Analyzing Population Structure and Genotype-Phenotype Mapping

The goal of this project is to bring new power to agricultural and conservational genetics through the use of the artificial intelligence subfield called deep learning. This will be done for two main applications.

The first one is exploring and reconstructing population structure based on genetic data, including visualizations of the genetic variation.

Understanding the degree and type of variation can help in identifying subpopulations that are to be preserved in order to maximize the variation in the genetic pool, whether it is a matter of wildlife or a controlled breeding program.

The models can also be used to elucidate seemingly dissimilar populations being genetically close, or vice versa, which can help explain their agricultural production properties.The second application is models for predicting and exploring genotype-phenotype relationships.

In current breeding programs, estimates of breeding value based on pedigree-estimated or measured genetic similarity between individuals form the gold standard.

The goal is then to predict the underlying genetic phenotype contribution, rather than the actual phenotype of a single individual.

Our models will cope with complex genotype-trait-covariate interactions in a superior way, and they can also predict the expectation in phenotype results for specific matings, using simulations to identify individuals carrying complementary genetic advantages. One example is that the model should predict the heterosis advantages of F1s.