What question did this study set out to answer?

The aim is to identify aroma markers specific to different bacterial strains in fermented pumpkin juice using machine learning techniques.

April 1, 2026Open Access

Machine learning-based identification of strain-specific aroma markers in fermented pumpkin juice: An integrated 1D-CNN and random forest approach

Key Points

The aim is to identify aroma markers specific to different bacterial strains in fermented pumpkin juice using machine learning techniques.
Utilized an integrated Random Forest and 1D-Convolutional Neural Network framework.
Fermented pumpkin juice at 30 °C for 12 hours using different bacterial strains.
Evaluated flavor profiles through sensory analysis with trained panelists.
Analyzed metabolic differences and volatile compounds using SHAP for interpretation.
Leu. citreum PX3-5 produced a creamy flavor, significantly enhancing phenolic and flavonoid content.
L. plantarum HY1-3 was noted for its floral notes, while Lb. brevis PL6-1 contributed earthy flavors.
Acetoin and 2-heptanone identified as key aroma markers differentiating flavor profiles.
The model demonstrated high accuracy in predicting strain-specific growth characteristics.

Abstract

This study investigated the impact of Levilactobacillus brevis , Lactiplantibacillus plantarum , and Leuconostoc citreum on the flavor and functional profiles of pumpkin juice fermented at 30 °C for 12 h. An integrated Random Forest (RF) and 1D-Convolutional Neural Network (1D-CNN) framework, interpreted via SHAP, identified strain-specific metabolic differences and volatile patterns. Under these laboratory-scale fermentation conditions, Leu. citreum PX3-5 produced a superior profile, significantly enhancing both total phenolic (54.61 mg GAE/kg) and flavonoid (79.64 mg RE/kg) contents, while enhancing creamy/buttery notes (acetoin OAV = 3.99, 2,3-butanedione OAV = 3.08). Conversely, L. plantarum HY1-3 was characterized by pronounced floral notes (β-ionone OAV = 31.07), while Lb. brevis PL6-1 imparted distinct mushroom/earthy notes (1-octen-3-ol OAV = 4.69). The integrated RF-CNN model robustly identified acetoin and 2-heptanone as differential markers. Notably, the CNN highlighted acetoin within a local associative pattern, suggesting its potential link to citrate metabolism. These chemical attributes aligned with the sensory profile of group PX3-5 as evaluated by a trained panel (n = 13), demonstrating consistency between the model-identified volatile markers and the observed sensory preference. This work provides a data-driven basis for developing value-added pumpkin products and proposes an analytical strategy for deciphering complex flavor systems in fermented plant-based foods. • Logistic models accurately described strain-specific LAB growth (R 2 =0.984-0.998). • RF-CNN identifies non-linear associative aroma patterns missed by PCA and PLS-DA. • Strain-specific metabolism dictates divergent flavor profiles (creamy vs. earthy). • Leu. citreum PX3-5 yields creamy flavor and enhances antioxidant capacity. • Acetoin and 2-heptanone differentiate fermentation profiles.

Machine learning-based identification of strain-specific aroma markers in fermented pumpkin juice: An integrated 1D-CNN and random forest approach

Key Points

Abstract

Cite This Study