Machine-learning-assisted screening of key flavor compounds in pumpkins
Jiangfei Chai1,2, Xiangdong Zhang3, Jincha Cao3
1College of Food and Health, Zhejiang Agriculture and Forestry University, Hangzhou, China.
Background:
Pumpkin is an important food source. Its flavor significantly impacts consumer acceptance and market competitiveness. Machine learning (ML) can capture non-linear patterns and multivariate relationships in volatile organic compound (VOC) datasets, linking VOC profiles to sensory attributes. In this study, sensory evaluation, gas chromatography-time-of-flight mass spectrometry (GC-ToF/MS) and ML were integrated to identify the flavor profiles of raw and steamed pumpkins and the VOC markers associated with their typical sensory attributes.
Results:
Raw pumpkin's sensory profile was dominated by green and cucumber aromas, whereas that of steamed pumpkin was characterized by chestnut, creamy, sweetness, and potato aromas. Based on five classification ML algorithms, using a consensus voting strategy, 25 important VOCs were identified across raw and steamed pumpkins. Alcohols were the primary VOCs in raw pumpkin, whereas 2,3-butanedione and methional were the main ML-highlighted VOCs associated with steamed pumpkin. ExtraTrees achieved the highest performance in linking VOCs to sensory attributes. Shapley Additive Explanations (SHAP) summary plots showed that alcohol compounds contributed substantially to the cucumber and green aromas. 2,3-Butanedione and methional emerged as major drivers of creamy, sweet, and potato aromas.
Conclusion:
The study contributes to the understanding of the sensory profile and VOC composition of pumpkins. The identified key VOCs provided reference compounds for sensomics research, establishing a basis for aroma breeding and product processing. © 2026 Society of Chemical Industry.
