Effects of using deep learning to predict the geographic origin of barley genebank accessions on genome–environment association studies

TL;DR

This paper explores using deep learning to predict the geographic origins of barley samples to improve genome-environment association studies for crop resilience.

Contribution

The novel use of neural networks to impute geographic origins and integrate them into genome-environment association studies is presented.

Findings

Neural networks accurately predicted geographic origins but sometimes produced ecologically implausible results.

Imputed data in GEA partially overlapped with regular GEA but also revealed new genomic regions near flowering time genes.

Increasing sample size via imputation did not significantly improve GEA performance as expected.

Abstract

Genome–environment association (GEA) is an approach for identifying adaptive loci by combining genetic variation with environmental parameters, offering potential for improving crop resilience. However, its application to genebank accessions is limited by missing geographic origin data. To address this limitation, we explored the use of neural networks to predict the geographic origins of barley accessions and integrate imputed environmental data into GEA. Neural networks demonstrated high accuracy in cross-validation but occasionally produced ecologically implausible predictions as models solely considered geographical proximity. For example, some predicted origins were located within non-arable regions, such as the Mediterranean Sea. Using barley flowering time genes as benchmarks, GEA integrating imputed environmental data (\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$N=11,032$$\end{document}N=11,032) displayed partially concordant yet complementary detection of genomic regions near flowering time genes compared to regular GEA (\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$N=1,626$$\end{document}N=1,626), highlighting the potential of GEA with imputed data to complement regular GEA in uncovering novel adaptive loci. Also, contrary to our initial hypothesis anticipating a significant improvement in GEA performance by increasing sample size, our simulations yield unexpected insights. Our study suggests potential limitations in the sensitivity of GEA approaches to the considerable expansion in sample size achieved through predicting missing geographical data. Overall, our study provides insights into leveraging incomplete geographical origin data by integrating deep learning with GEA. Our findings indicate the need for further development of GEA approaches to optimize the use of imputed environmental data, such as incorporating regional GEA patterns instead of solely focusing on global associations between allele frequencies and environmental gradients across large-scale landscapes.

The online version contains supplementary material available at 10.1007/s00122-025-05003-w.

Figures