Construction of a Regression Model for a Coffee Sensory Evaluation Through the Comprehensive Analysis of Metabolites

Significance of the Topic

Comprehensive sensory evaluation of food products, such as coffee, relies on complex interactions among multiple chemical constituents rather than simple one‐to‐one relationships. Advanced metabolomic profiling combined with regression modelling presents a powerful approach to link sensory attributes with underlying metabolite patterns and accelerate quality control and product development in the food industry.

Objectives and Study Overview

This study aimed to construct a predictive regression model correlating bitterness scores of eight coffee types with their metabolite profiles. Using a widely targeted GC-MS/MS metabolomics workflow, the authors extracted and quantified hundreds of metabolites, then applied partial least squares (PLS) regression to uncover key compounds influencing perceived bitterness.

Methodology and Instrumentation

Sample Preparation and Extraction:

• Ground and roasted coffee beans (eight varieties) were weighed (~20 mg) and spiked with an internal standard (2-Isopropylmalic acid).
• Methanol/water/chloroform extraction was performed, followed by overnight vacuum drying.
• Derivatization involved methylamine in pyridine, followed by N-Methyl-N-(trimethylsilyl)trifluoroacetamide treatment.

GC-MS/MS Analysis:

• Instrument: Shimadzu Gas Chromatograph Mass Spectrometer (GC-MS/MS).
• Column: BPX-5, 30 m×0.25 mm, 0.25 µm film.
• Split injection (30:1) at 250 °C; oven ramp 60 °C to 330 °C at 15 °C/min.
• Interface and ion source temperatures: 200 °C and 280 °C, respectively.
• Data acquisition based on the Smart Metabolites Database.

Main Results and Discussion

Sensory Evaluation:

• Eight panelists rated bitterness on a five‐point scale; scores were summed per sample.
• Bitterness scores ranged from 19 to 34 across samples.

Metabolite Detection:

• 475 metabolites targeted; 192 detected across all samples after normalization to the internal standard.

PLS Regression Model:

• Simca software yielded an R2 of 0.9945 and RMSEP of 0.346, indicating high predictive accuracy.
• Positive regression coefficients (e.g., Glycine-3TMS, Arabitol-5TMS, Mannitol-6TMS) were associated with increased bitterness.
• Negative coefficients (e.g., 4-Hydroxybenzoic acid-2TMS, Glyceraldehyde-2TMS) indicated compounds that reduce perceived bitterness.

Practical Benefits and Applications

This integrated approach allows rapid identification of flavor‐active metabolites, guiding coffee roasting and blending to achieve desired sensory profiles. The model framework can be extended to other food matrices, enabling data-driven quality control, flavor optimization, and new product formulation.

Future Trends and Opportunities

• Integration of larger metabolite libraries and high‐resolution mass spectrometry for deeper coverage.
• Machine learning algorithms beyond PLS to capture nonlinear relationships.
• Real-time sensory prediction platforms linked to process control in food manufacturing.
• Cross‐validation with trained sensory panels and consumer studies to refine model robustness.

Conclusion

The study demonstrates that GC-MS/MS‐based metabolomic profiling coupled with PLS regression provides an accurate, interpretable model for predicting coffee bitterness. Key metabolites influencing taste were identified, offering actionable insights for food quality management and product innovation.

References

1. Shimadzu Corporation. Construction of a Regression Model for Coffee Sensory Evaluation through the Comprehensive Analysis of Metabolites. Application News No. M274, First Edition: Feb. 2018.