This must be the place: deep learning local adaptation
- 1University of Oregon
Abstract
Climate change is increasingly disrupting the relationship between locally adapted populations and the environments in which they evolved, creating an urgent need for tools that connect genomic variation to climate. Common-garden and provenance trials remain the gold standard for characterizing local adaptation, but their time and resource requirements limit how broadly they can be applied. Genomic approaches provide a complementary path. Genotype--environment association (GEA) methods identify environmentally associated loci. Machine-learning models have also shown that geographic origin can be predicted directly from genotypes. Here we introduce EcoLocator, a supervised deep neural network that jointly predicts geographic location and climate of origin from genotypes. Through extensive simulations we demonstrate that EcoLocator accurately recovers geographic location and environment of origin from genotype data, and, with SHAP-based feature attribution, identifies adaptive loci more reliably than benchmark GEA methods. We apply our method to coastal Douglas-fir (Pseudotsuga menziesii var. menziesii), where EcoLocator predicts geographic origin (R2=0.75--0.83) and climate of origin (R2=0.52--0.72) under leave-one-out cross-validation. Notably, we find that climate predicted directly from genotypes outperforms climate inferred by first predicting geographic origin, showing that EcoLocator captures genotype--climate signal that cannot be recovered from geography alone. Our prediction errors fall within the tolerances used in existing seed-transfer guidelines, demonstrating that EcoLocator's predictions are ready for practical application, and our approach is readily extendable to other species and conservation contexts.
Related
- Place-based climate adaptation: How Atlantic Canadian communities translate local knowledge into action
- An extensible deep learning framework for multi-type chromatin interaction prediction through hierarchical adaptation strategies
- Neural correlates of place learning in the cane toad, Rhinella marina
- Insight Into the Genetic Architecture of Local (Mal)Adaptation in Wild and Feral Brassica rapa Populations
- Local translation couples synaptic activity to mitochondrial adaptation in dendrites
BibTeX
@misc{rodriguez2026this,
title = {{This must be the place: deep learning local adaptation}},
author = {J. Rodriguez and R. C. Cronn and S. Tittes and A. D. Kern},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.16.752190},
url = {https://www.biorxiv.org/content/10.64898/2026.09.16.752190v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - This must be the place: deep learning local adaptation
AU - J. Rodriguez
AU - R. C. Cronn
AU - S. Tittes
AU - A. D. Kern
PY - 2026
DA - 2026/09/16
PB - bioRxiv
DO - 10.64898/2026.09.16.752190
UR - https://www.biorxiv.org/content/10.64898/2026.09.16.752190v1
AB - Climate change is increasingly disrupting the relationship between locally adapted populations and the environments in which they evolved, creating an urgent need for tools that connect genomic variation to climate. Common-garden and provenance trials remain the gold standard for characterizing local adaptation, but their time and resource requirements limit how broadly they can be applied. Genomic approaches provide a complementary path. Genotype--environment association (GEA) methods identify environmentally associated loci. Machine-learning models have also shown that geographic origin can be predicted directly from genotypes. Here we introduce EcoLocator, a supervised deep neural network that jointly predicts geographic location and climate of origin from genotypes. Through extensive simulations we demonstrate that EcoLocator accurately recovers geographic location and environment of origin from genotype data, and, with SHAP-based feature attribution, identifies adaptive loci more reliably than benchmark GEA methods. We apply our method to coastal Douglas-fir (Pseudotsuga menziesii var. menziesii), where EcoLocator predicts geographic origin (R2=0.75--0.83) and climate of origin (R2=0.52--0.72) under leave-one-out cross-validation. Notably, we find that climate predicted directly from genotypes outperforms climate inferred by first predicting geographic origin, showing that EcoLocator captures genotype--climate signal that cannot be recovered from geography alone. Our prediction errors fall within the tolerances used in existing seed-transfer guidelines, demonstrating that EcoLocator's predictions are ready for practical application, and our approach is readily extendable to other species and conservation contexts.
ER -