# When does accounting for gene–environment interactions improve complex trait prediction? A case study with Drosophila lifespan

## TL;DR

This study explores when considering gene-environment interactions improves predictions of complex traits like lifespan in fruit flies, finding that it helps only when the same genotypes are in both training and test data.

## Contribution

The study identifies specific prediction scenarios where gene-environment interactions improve accuracy in complex trait prediction.

## Key findings

- Gene-environment interactions explained 8% of lifespan variance in Drosophila.
- Models with G×E improved prediction accuracy only when the same genotypes were in both reference and test populations.
- Such scenarios are common in agriculture but rare in human studies.

## Abstract

Gene–environment interactions (G×E) have been shown to explain a non-negligible proportion of variance for a plethora of complex traits in different species, including livestock, plants, and humans. While several studies have shown that including G×E can improve prediction accuracy in agricultural species, no increase in accuracy has been observed in human studies. In this work, we sought to investigate the scenarios in which accounting for G×E is expected to improve prediction accuracy. Model organisms are useful for studying G×E, since environments can be defined precisely, and genotypes can be replicated across environments, which are ideal conditions to minimize confounding in G×E analyses. Thus, we used data from an experiment in Drosophila melanogaster, where researchers measured lifespan in different environments for unrelated inbred lines (i.e. genotypes). We used three different cross-validation (CV) scenarios that mimic different relationships between reference and test populations, and fitted a few statistical models with and without including G×E. The results showed that G×E explained 8% of lifespan variance. Despite that, models accounting for G×E improved prediction accuracy only in CV scenarios where the same genotypes are observed in both the reference and test populations. While these scenarios are common in agriculture, where individuals of the same family or variety appear in both populations, they are not commonly encountered in human studies, where individuals are unrelated. Thus, our work shows in which prediction scenarios we can expect improvements by accounting for G×E, and may provide a potential reason (among others) for results of human studies.