Machine learning techniques for continuous genetic assignment of geographic origin of forest trees

Origin tracking is important to ensure use of the right seed source and trade with legally harvested timber. Additionally, it can help to reconstruct human-caused historical long-distance seed transfer and to spot mislabelling in forest field trials. So far, genetic assignment approaches were mostly discrete, assigning test samples to predefined groups. The main limitation of this approach is the justification of these discrete groups when genetic variation across the landscape is actually continuous. Here, we compare the accuracy of five continuous assignment methods. Specifically, we test a nearest neighbour method (NN), direct gaussian process regression (GPR-D) using the radial basis kernel function, grid based gaussian process regression (GPR-G) applying the Matérn kernel function, genomic prediction (GP) and deep learning (DL), using two genome-wide single nucleotide polymorphism (SNP) datasets of trees from across Europe. The first dataset comprises 30,000 SNPs from 865 European beech (Fagus sylvatica) trees, the second dataset consists of 381 SNPs from 1,883 pedunculate oak (Quercus robur) trees. The accuracy, as measured by the geographic distance between true and predicted locations, was highest for the GPR-G and DL methods with the beech dataset with a median distance of only 55 km and 76 km, respectively. For the oak data GPR-G and DL also performed best with median distances of 263 km and 278 km, respectively. The relative error (distance/max distance among tree pairs) was below 8% for 90% of all samples for the best method for both datasets. We detected 35 individuals and 10 groups as outliers in the beech data and 27 individuals and 18 groups in the oak data. These outliers may be caused by mislabelling or historical human-caused long distance seed transfer. We discuss the differences in performance of the approaches and highlight future applications and potential for further improvements.