Main content area

Prediction of sensory quality in raw carrots (Daucus carota L.) using multi-block LS-ParPLS

Kreutzmann, Stine, Svensson, Vibeke T., Thybo, Anette K., Bro, Ramus, Petersen, Mikael A.
Food quality and preference 2008 v.19 no.7 pp. 609-617
Daucus carota, carrots, genotype, raw vegetables, sensory properties, prediction, dry matter content, bitterness, sweetness, terpenoids, flavor, correlation, chlorogenic acid, volatile compounds, flavor compounds, odor compounds
The relations between the sensory quality of 24 different carrot genotypes and content of dry matter, non-volatile and volatile compounds were studied using a multi-block approach called LS-parPLS. The prediction of five sensory attributes; bitterness, sweetness, terpene flavour, green flavour and carrot flavour, gave prediction errors (RMSECV) between 0.98 and 1.36 and correlation coefficients (r) between 0.60 and 0.81. The explained Y-variances were between 15.1% and 66.0%. The highest prediction error was observed for the attribute carrot flavour whereas green flavour gave the best prediction. The attributes green flavour, bitterness and terpene flavour showed fairly good predictions (r/RMSECV/% exp-Y = 0.81/0.98/66.0, 0.79/1.23/62.3 and 0.71/1.04/50.2) whereas sweetness gave an unexpected poor prediction (r/RMSECV/% exp-Y = 0.67/1.36/44.6). Non-volatile compounds found to be important predictors were chlorogenic acid (5-CQA), sucrose, 6-methoxymellein (6-MM), falcarindiol (FaDOH), and falcarinol (FaOH). The volatile compounds found to be important predictors are considered as key flavour compounds of raw carrots: terpinolene, β-pinene, sabinene, γ-terpinene, α-pinene, β-bisabolene, caryophyllene and cuparene. In general, the overall results show that the sensory quality variation in the material regarding bitterness, green flavour and terpene flavour are explained by relatively few parameters. Despite that the results revealed some reliable relationships between the sensory attributes, aroma and chemical analysis, a large variance (about 40%) in the sensory block of variables remained unexplained and still needs further investigation for an in-depth understanding of sensory quality. LS-ParPLSc is shown to be feasible for handling several types of data blocks in one regression model.