Inferring geographic origin in genebank accessions from genome‐wide SNP variation: A barley case study

Accurate geographic metadata are essential for using genebank accessions in spatial analyses, but many historical records lack precise coordinates or contain uncertain location information. We developed SNP2GPS, a plant-genomics workflow for inferring geographic origin from genome-wide single-nucleotide polymorphism data and supporting passport metadata curation in barley (Hordeum vulgare L.) collections. SNP2GPS integrates genotype preprocessing, neural network-based coordinate prediction, repeated model training, uncertainty estimation, spatial consistency checks, visualization, and extraction of approximate coordinates from free-text locality descriptors. Models were trained on curated passport coordinates from the Leibniz Institute of Plant Genetics and Crop Plant Research barley collection and applied to accessions lacking precise geolocation data. On an independent test set of 200 withheld accessions, the base model achieved a mean prediction error of 383 km and a median error of 58 km; an augmented model using additional country-centroid labels achieved a mean error of 427 km and a median error of 56 km. Country-centroid augmentation increased geographic coverage in undersampled regions, though at the cost of reduced robustness for some samples. SNP2GPS also identified accessions with large discrepancies between recorded and predicted locations, flagging candidates for metadata review. Prediction uncertainty was lower in wild barley than in domesticated material, consistent with a stronger coupling between genetic structure and geography in wild populations. SNP2GPS provides an open, uncertainty-aware framework for imputing incomplete geolocation metadata, identifying inconsistent passport records, and delivering reusable predicted coordinates through downloadable outputs and web-based visualization.