<p>Fig seed oil, long overlooked in traditional&#xa0;fig&#xa0;processing, has recently gained attention for its rich profile of&#xa0;lipid-bioactive&#xa0;compounds and exceptional antioxidant&#xa0;capacity. This emerging natural oil shows considerable promise in cosmetic and nutraceutical applications and may become a commercially viable alternative to prickly pear seed oil. To enable large-scale screening of this oil,&#xa0;we developed&#xa0;predictive models using non-destructive mid-FTIR spectroscopy&#xa0;combined with machine learning algorithms. A dataset of 222 fig seed oil samples was analyzed for&#xa0;total phenolic content (TPC), total flavonoid content (TFC),&#xa0;and antioxidant activity using DPPH and ABTS assays.&#xa0;We&#xa0;compared three regression&#xa0;approaches: multilayer perceptron (MLP), automated machine learning (AutoML), and genetic algorithm-enhanced partial least squares regression (GA-PLSR) in conjunction with four spectral preprocessing strategies (raw spectra, Savitzky-Golay smoothing,&#xa0;standard normal variate (SNV), and multiplicative scatter correction (MSC)) to&#xa0;optimize&#xa0;prediction performance. The MLP model combined with MSC or SNV preprocessing&#xa0;delivered&#xa0;the most reliable predictions,&#xa0;achieving&#xa0;blind test <i>R</i><sup>2</sup> values&#xa0;exceeding&#xa0;0.82 for all biochemical parameters. Antioxidant activities (DPPH and ABTS) were predicted with the highest accuracy&#xa0;(<i>R</i><sup>2</sup> &gt; 0.85), likely&#xa0;attributable&#xa0;to strong spectral signatures in the 1000–1800&#xa0;cm<sup>−1</sup> and 2800–3000&#xa0;cm<sup>−1</sup> regions. Total phenolic content was also well-predicted, with SNV preprocessing improving <i>R</i><sup>2</sup> from 0.794 to 0.823. Total flavonoid content proved more challenging&#xa0;to predict (<i>R</i><sup>2</sup> ≈ 0.44–0.74), probably due to weaker spectral signals and higher analytical variability. These results&#xa0;demonstrate&#xa0;that MSC and SNV preprocessing&#xa0;substantially&#xa0;enhance model robustness and that antioxidant-related parameters are particularly&#xa0;amenable&#xa0;to FTIR-based MLP prediction. This approach provides a rapid, cost-effective, and environmentally sustainable alternative to conventional wet chemistry methods, supporting high-throughput quality assessment and valorization of underutilized plant resources in accordance with green chemistry principles.</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

FTIR–Machine Learning Tandem for Predicting Antioxidant Bioactives in Fig Seed Oil: A Pathway to High-Throughput Screening

  • Charaf Ed-dine Kassimi,
  • Souhaila Bouchelta,
  • Souhaila Hadday,
  • Ibtissame Guirrou,
  • Ahmed Irchad,
  • Fedoua Diai,
  • Lhoussain Hajji,
  • Lahcen Hssaini

摘要

Fig seed oil, long overlooked in traditional fig processing, has recently gained attention for its rich profile of lipid-bioactive compounds and exceptional antioxidant capacity. This emerging natural oil shows considerable promise in cosmetic and nutraceutical applications and may become a commercially viable alternative to prickly pear seed oil. To enable large-scale screening of this oil, we developed predictive models using non-destructive mid-FTIR spectroscopy combined with machine learning algorithms. A dataset of 222 fig seed oil samples was analyzed for total phenolic content (TPC), total flavonoid content (TFC), and antioxidant activity using DPPH and ABTS assays. We compared three regression approaches: multilayer perceptron (MLP), automated machine learning (AutoML), and genetic algorithm-enhanced partial least squares regression (GA-PLSR) in conjunction with four spectral preprocessing strategies (raw spectra, Savitzky-Golay smoothing, standard normal variate (SNV), and multiplicative scatter correction (MSC)) to optimize prediction performance. The MLP model combined with MSC or SNV preprocessing delivered the most reliable predictions, achieving blind test R2 values exceeding 0.82 for all biochemical parameters. Antioxidant activities (DPPH and ABTS) were predicted with the highest accuracy (R2 > 0.85), likely attributable to strong spectral signatures in the 1000–1800 cm−1 and 2800–3000 cm−1 regions. Total phenolic content was also well-predicted, with SNV preprocessing improving R2 from 0.794 to 0.823. Total flavonoid content proved more challenging to predict (R2 ≈ 0.44–0.74), probably due to weaker spectral signals and higher analytical variability. These results demonstrate that MSC and SNV preprocessing substantially enhance model robustness and that antioxidant-related parameters are particularly amenable to FTIR-based MLP prediction. This approach provides a rapid, cost-effective, and environmentally sustainable alternative to conventional wet chemistry methods, supporting high-throughput quality assessment and valorization of underutilized plant resources in accordance with green chemistry principles.