Machine learning-assisted optimization of green L-DOPA extraction from Faba bean (Vicia faba L.) leaf tissues using CatBoost and bootstrap-based uncertainty quantification.

L-3,4-dihydroxyphenylalanine (L-DOPA), a key phytochemical with therapeutic significance in Parkinson's disease management, is naturally abundant in Vicia faba L. (Faba bean) tissues. The present study developed and optimized a green extraction protocol for L-DOPA from Faba bean leaf tissues using a systematic multi-factor experimental design encompassing extraction medium (11 levels), solid-to-liquid (S/L) ratio (1:25-1:100 w/v), pre-treatment method (sonication vs. rotospin), and extraction time (10-30 min). A total of 792 observations were generated in biological triplicates. Four machine learning models were trained and benchmarked: Linear Regression, Random Forest, CatBoost Baseline, and Optuna-tuned CatBoost (40 trials, 5-fold CV, TPE sampler). Random Forest achieved the highest test R2 (0.844), while the Optuna-tuned CatBoost (R2 = 0.774) marginally exceeded the untuned baseline (R2 = 0.746). To maximise reproducibility and avoid selecting a model on the hold-out test set, the parsimonious CatBoost Baseline was prespecified as the single operational model for all downstream analyses, retaining CatBoost native categorical-feature handling and seamless SHAP integration. SHAP analysis revealed S/L ratio (mean SHAP= 3099 µg/g DW) as the dominant extraction factor, followed by extraction medium (2736 µg/g DW) and pre-treatment method (1857 µg/g DW), while extraction time showed negligible influence (623 µg/g DW; Kruskal-Wallis p = 0.31). The globally optimal condition identified by bootstrap-based uncertainty quantification (n = 200 resamples) was 10% lemon juice (v/v) at A1:100 S/L ratio with sonication for 10 min, yielding a predicted mean L-DOPA of 47,127 µg /g DW (90% CI: 38,450-53,241 µg /g DW). This study integrates ML with an all-categorical full factorial design and an uncertainty-aware bootstrap optimisation framework, providing per-condition 90% confidence intervals across the full extraction space, capabilities absent from conventional RSM approaches. All models, datasets, and an interactive Streamlit decision-support dashboard are publicly available.