Résumé
The ability of species and populations to adapt to their environment faces increasing challenges as climate change accelerates. Recent methods based on genomic offset (GO) statistics aim to quantify the risk of non-adaptation of populations to future climates. While several studies have evaluated the ability of different offset statistics to predict population (mal)adaptation, the impact of the chosen climate data?which could vary in relevance and quality?remains unexplored. In this study, we tested how the choice of environmental variables impacts the predictive performance of GO statistics. Our assessment leverages a common garden experiment to evaluate how well GO values correspond to population variation in fitness. We analyzed 157 pearl millet (Pennisetum glaucum (L.) R.Br. syn Cenchrus americanus (L.) Morrone) landraces cultivated in West Africa, for which fitness proxies were measured during field trials conducted over 2?years. We calculated geometric and gradient forest GO statistics using three different historical climate datasets: bioclimatic variables from WorldClim (v.2.1), from CHELSA (v.2.1), and a dataset specifically designed for this study derived from the EWEMBI (v.1) climate database. The predictive power of these statistics varied threefold when using the mean weight of seeds as the fitness trait, from the lowest correlation with GO at R2?=?0.20, to the highest at R2?=?0.60. Climate datasets have a stronger effect on the predictive performance of the offset statistics than the chosen statistical method itself. We showed that considering climate variables that are closely linked to the crop life cycle globally improves the relationship between GO values and estimated fitness. Our results emphasize the importance of using climate metrics that are relevant to the study system, along with experimental validation to enhance the reliability of GO predictions. This will ultimately strengthen future breeding and conservation strategies aimed at mitigating the impact of climate change.