Scholarlink

<p>Essential oils (EOs) are chemically complex natural matrices whose quality and bioactivity are governed by structurally diverse secondary metabolites. Conventional techniques such as gas chromatography–mass spectrometry (GC–MS) provide detailed compositional profiling, yet they remain costly, labor-intensive, and unsuitable for real-time monitoring. Here, we present a multimodal chemometric framework that integrates UV–Vis–NIR spectroscopy (190–1100 nm) with low-cost metal oxide gas sensors to quantitatively predict EO metabolite concentrations using machine learning. Multivariate analyses employing <i>t</i>-distributed stochastic neighbor embedding (<i>t</i>-SNE), uniform manifold approximation and projection (UMAP), and correlation mapping revealed chemically coherent clustering of <i>Cistus ladanifer</i> EO samples and biochemical associations between sensor responses and metabolite families, highlighting the richness of the fused feature space. Metabolites identified by GC–MS were grouped into seven functional classes including terpenic hydrocarbons, sesquiterpenic hydrocarbons, alcohols, aldehydes, ketones, esters, and residuals. These groupings guided targeted regression modeling. Data validation was performed against GC–MS quantified metabolites. Among the tested algorithms, Ridge regression achieved the highest predictive performance (<i>R</i><sup>2</sup> = 0.999). Lasso regression followed with <i>R</i><sup>2</sup> = 0.971, favoring sparsity at the expense of completeness. Partial least squares algorithm failed to capture variance in the high-dimensional multimodal dataset. Feature attribution based on Shapley values demonstrated that accurate predictions required the joint contribution of distributed spectral bands and complementary sensor responses, underscoring the necessity of multimodal fusion for resolving chemically heterogeneous and low-abundance metabolites. This work establishes a scalable, non-destructive, and real-time strategy for EO profiling, with broad implications for traceability, sustainable cultivation, and smart agriculture, and illustrates the transformative role of machine learning in chemometric exploration of natural products.</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

Machine Learning Unlocks Chemometric Profiling of Plant-Derived Metabolites Using Spectral and Gas Sensor Fingerprints

  • 期刊论文

  • 发表日期：

    2025年11月19日

查看全文
  • Ali Ahmad,
  • Francisco Javier Diaz,
  • Sandra Sendra,
  • Jaime Lloret

摘要

Essential oils (EOs) are chemically complex natural matrices whose quality and bioactivity are governed by structurally diverse secondary metabolites. Conventional techniques such as gas chromatography–mass spectrometry (GC–MS) provide detailed compositional profiling, yet they remain costly, labor-intensive, and unsuitable for real-time monitoring. Here, we present a multimodal chemometric framework that integrates UV–Vis–NIR spectroscopy (190–1100 nm) with low-cost metal oxide gas sensors to quantitatively predict EO metabolite concentrations using machine learning. Multivariate analyses employing t-distributed stochastic neighbor embedding (t-SNE), uniform manifold approximation and projection (UMAP), and correlation mapping revealed chemically coherent clustering of Cistus ladanifer EO samples and biochemical associations between sensor responses and metabolite families, highlighting the richness of the fused feature space. Metabolites identified by GC–MS were grouped into seven functional classes including terpenic hydrocarbons, sesquiterpenic hydrocarbons, alcohols, aldehydes, ketones, esters, and residuals. These groupings guided targeted regression modeling. Data validation was performed against GC–MS quantified metabolites. Among the tested algorithms, Ridge regression achieved the highest predictive performance (R2 = 0.999). Lasso regression followed with R2 = 0.971, favoring sparsity at the expense of completeness. Partial least squares algorithm failed to capture variance in the high-dimensional multimodal dataset. Feature attribution based on Shapley values demonstrated that accurate predictions required the joint contribution of distributed spectral bands and complementary sensor responses, underscoring the necessity of multimodal fusion for resolving chemically heterogeneous and low-abundance metabolites. This work establishes a scalable, non-destructive, and real-time strategy for EO profiling, with broad implications for traceability, sustainable cultivation, and smart agriculture, and illustrates the transformative role of machine learning in chemometric exploration of natural products.