Scholarlink

<p>A new integrated extraction strategy was established by combining natural solvent systems (NADES), microwave-assisted extraction, and machine learning techniques for optimizing the extraction of bioactive compounds from fermented watermelon rind. The results showed that solid-state fermentation significantly improved extraction rates, compared to non-fermented controls. The fermentation pretreatment comparison was conducted under identical extraction conditions, with non-fermented watermelon rind serving as the internal control. Machine learning models optimized using Bayesian optimization were developed to describe and predict three important response variables: TPC, TFC, and antioxidant activity, as functions of four extraction variables (microwave power, temperature, time, and solid-liquid ratio). The ensemble models showed outstanding predictive capabilities with test R² values of 0.9147, 0.9088, and 0.9252 for TFC, TPC, and DPPH activity, respectively, with very low overfitting (ΔR² &lt; 0.06). Importance analysis of the features showed temperature as the most important parameter in the extraction of bioactive compounds (importance: 0.842–0.885), followed by solid-liquid ratio. Simultaneous optimization using the validated models showed the optimal extraction conditions to be 62.5&#xa0;°C, 27.7&#xa0;min, 300&#xa0;W, and 30&#xa0;mg/mL, predicting values of 1.656&#xa0;mg CE/g (TFC), 19.80&#xa0;mg GAE/g (TPC), and 71.27% (DPPH activity). Solid-state fermentation enhanced extraction yields, increasing total phenolics (16.9 to 19.1&#xa0;mg GAE/g), flavonoids (0.61 to 0.74&#xa0;mg CE/g), and DPPH activity (66% to 75%). The developed ensemble models showed high accuracy (R² = 0.91–0.93; RMSE = 0.0812&#xa0;mg/g, 0.51&#xa0;mg/g, and 0.74%). Experimental validation results have confirmed the high accuracy of the model with relative errors of less than 3% for all responses. The combination of fermentation pretreatment, green extraction, and machine learning indicates promise for sustainable valorization of agricultural waste resources as sources of bioactive compounds for nutraceutical applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Ensemble machine learning with Bayesian optimization predicts bioactive extraction from fermented watermelon rind using natural deep eutectic solvents

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月25日

查看全文
  • Mostafa Khajeh,
  • Mansour Ghaffari-Moghaddam,
  • Afsaneh Barkhordar,
  • Mousa Bohlooli,
  • Didem Saloglu

摘要

A new integrated extraction strategy was established by combining natural solvent systems (NADES), microwave-assisted extraction, and machine learning techniques for optimizing the extraction of bioactive compounds from fermented watermelon rind. The results showed that solid-state fermentation significantly improved extraction rates, compared to non-fermented controls. The fermentation pretreatment comparison was conducted under identical extraction conditions, with non-fermented watermelon rind serving as the internal control. Machine learning models optimized using Bayesian optimization were developed to describe and predict three important response variables: TPC, TFC, and antioxidant activity, as functions of four extraction variables (microwave power, temperature, time, and solid-liquid ratio). The ensemble models showed outstanding predictive capabilities with test R² values of 0.9147, 0.9088, and 0.9252 for TFC, TPC, and DPPH activity, respectively, with very low overfitting (ΔR² < 0.06). Importance analysis of the features showed temperature as the most important parameter in the extraction of bioactive compounds (importance: 0.842–0.885), followed by solid-liquid ratio. Simultaneous optimization using the validated models showed the optimal extraction conditions to be 62.5 °C, 27.7 min, 300 W, and 30 mg/mL, predicting values of 1.656 mg CE/g (TFC), 19.80 mg GAE/g (TPC), and 71.27% (DPPH activity). Solid-state fermentation enhanced extraction yields, increasing total phenolics (16.9 to 19.1 mg GAE/g), flavonoids (0.61 to 0.74 mg CE/g), and DPPH activity (66% to 75%). The developed ensemble models showed high accuracy (R² = 0.91–0.93; RMSE = 0.0812 mg/g, 0.51 mg/g, and 0.74%). Experimental validation results have confirmed the high accuracy of the model with relative errors of less than 3% for all responses. The combination of fermentation pretreatment, green extraction, and machine learning indicates promise for sustainable valorization of agricultural waste resources as sources of bioactive compounds for nutraceutical applications.