Inferring genotype-phenotype maps using attention-based models

Predicting phenotypes from genotypes remains a fundamental challenge in genetics. Traditional approaches in quantitative genetics typically address this problem using linear regression frameworks. These models generally assume that the genetic architecture of complex traits can be represented by an additive structure where loci contribute independently, sometimes extended to include pairwise epistatic interactions. However, such formulations often fail to capture higher-order epistasis and subtle gene-environment interactions. Recent advances in machine learning, particularly attention-based models, provide a powerful alternative. Originally developed for natural language processing, attention mechanisms are well suited for modeling context-dependent relationships and have achieved remarkable success in predicting protein structure and function. We have applied attention-based architectures to quantitative genetics to evaluate their ability to predict phenotypes from genotypes using both simulated datasets with increasing epistatic complexity and experimental data from a quantitative trait locus mapping study in budding yeast. We find that attention-based models achieve superior out-of-sample performance in epistatic regimes compared with standard methods. We have also extended this framework to a multi-environment setting and show that attention-based architectures enable effective transfer learning for phenotype prediction in new environments with limited training data.