From Arabidopsis to Crop Prediction: Enrichment and Limitations of Tree‐Based SHAP for Marker Prioritization in Plant Breeding

ABSTRACT Genomic prediction is widely used in plant breeding, but feature‐attribution scores from predictive models are sometimes interpreted as evidence for trait‐associated or causal loci. We evaluated whether global SHapley Additive exPlanations (SHAP) from XGBoost and LightGBM prioritize markers or genomic windows enriched for signals from mixed‐model genome‐wide association studies (GWAS). Three public datasets were analysed: flowering time at 16°C in 970 Arabidopsis thaliana accessions, standardized grain yield in 599 wheat lines and flowering time at Arkansas in 374 rice accessions. Within each dataset, prediction, out‐of‐fold SHAP and GWAS used harmonized marker panels. SHAP values were calculated on the held‐out samples of the same five cross‐validation models used to estimate predictive performance. Arabidopsis and rice were compared using physical windows, whereas wheat was analysed at marker level because reliable genomic coordinates were unavailable for its dominant DArT markers. Fivefold cross‐validation produced Pearson correlations of 0.759–0.799 in Arabidopsis , 0.373–0.565 in wheat and 0.632–0.685 in rice. The top 50 SHAP‐ranked features were enriched for the top 5% of GWAS‐ranked signals after controlling for genomic structure and marker density or frequency. In Arabidopsis , XGBoost and LightGBM recovered 7 and 8 GWAS‐ranked windows, compared with 3.31 and 3.64 conditionally expected ( and 0.0031). In rice, the corresponding overlaps were 11 and 12, compared with 4.03 and 4.59 expected ( for both). In wheat, 18 and 23 of the top 50 SHAP‐ranked markers belonged to the top 5% of mixed‐model association results; minor‐frequency‐stratified permutation tests gave for both models. Global XGBoost–LightGBM rank correlations were 0.904, 0.806 and 0.836 in Arabidopsis , rice and wheat, respectively, although overlap among the highest‐ranked subsets remained incomplete. Tree‐based SHAP therefore captures reproducible, association‐enriched predictive structure, but it is not equivalent to GWAS and should not be used as a standalone proxy for causal loci. Its most defensible role is complementary model diagnosis and candidate prioritization combined with association evidence, biological annotation and independent validation.