A Comparison of Genomic Forecasts Based on Genotypes versus Allele Frequencies

AbstractAccelerating land use and climate change threaten to disrupt relationships between adaptive variation and environmental optima of many species. Consequently, management must increasingly identify nonlocal genetic sources for restoration programs. Genomic offset methods, such as gradientForests, have shown promise in identifying these sources using genomic data, potentially bypassing the need for traditional, time-consuming transplant experiments. However, previous studies primarily used population-level allele frequencies (AFs) for training and population mean fitness for evaluation, ignoring individual variation within populations. Here, we used simulation data to compare the accuracy of genotype- and AF-based models, factorially evaluated using both individual and population mean fitness. With more than 810,000 evaluations of such models, we found that the number of loci had little impact on model performance. As expected, population-level evaluation provided an optimistic view of predictive performance for both genomic inputs. While genotype- and AF-based models showed similar qualitative and quantitative aspects, genotype-based models improved predictions in landscapes that differed from strict environmental clines by incorporating additional loci beyond those used by AF-based models. This suggests that genotype-based models may enhance offset predictions in environments that are discontinuous and have multiple populations in geographically distant yet similar environments. We close with recommendations for future use and evaluation of these tools.