Automated machine learning based prediction of monounsaturated and polyunsaturated fatty acids in Brassica napus L. using field-level agronomic data

Rapeseed ( Brassica napus L.) oil quality depends on monounsaturated (MUFA) and polyunsaturated (PUFA) fractions, but routine GC-FID profiling is slow and costly. This study develops fast, field-data–driven models to predict five individual unsaturated fatty acids (palmitoleic, oleic, eicosenoic, linoleic, α-linolenic acids) and two indices (MUFA, PUFA) and uses explainable AI to identify agronomic drivers of oil composition. We compiled 649 plot-level observations from four split-plot experiments (2016–2023, Karaj, Iran; 27 genotypes, three replicates). Eleven routine agronomic, physiological and management predictors were recorded. Seed oil was measured by low-field NMR and fatty-acid composition by GC-FID; MUFA and PUFA were calculated from FAMEs. For each trait we trained an independent TPOT (genetic-programming AutoML) regressor on an 80/20 train/test split and interpreted fitted models with SHAP (global importance and beeswarm plots). Trait-specific AutoML pipelines achieved strong test performance: PALK MAE= 0.020 (R²=0.992); OLA 0.152 (0.986); ECO 0.134 (0.971); LINO 0.457 (0.971); LINK 0.253 (0.977); MUFA 0.198 (0.860); PUFA 0.540 (0.849). SHAP consistently identified sowing date, plant water-status (leaf and crown relative water content, stomatal resistance) and reproductive traits (siliques per plant, grains per silique) as dominant, trait-specific drivers. TPOT AutoML models trained on routine field measurements accurately predicted individual fatty-acid percentages and, via SHAP explanations, consistently identified actionable agronomic levers principally sowing date, plant water status and reproductive sink traits that can inform targeted sampling, management and breeding. • TPOT AutoML predicts five fatty-acid traits and two aggregate indices from routine field data. • High accuracy: R² up to 0.99 for individual fatty acids. • SHAP pinpoints key drivers: sowing date, water status, sink strength. • Workflow enables low-cost prioritization of GC assays and breeding. • Dataset: 649 plot-level observations (2016–2023) across 27 genotypes.