Complex visualised flavour worlds - information from four domains presented simultaneously

Added on by Michael Bom Frost.

Edited for Nordic Food Lab’s blog by Michael Bom Frøst

Manuscript authors: Marta Bevilacqua, Barbara Santos Silva, Michael Bom Frøst, Benedict Reade, Kristen Rasmussen de Vasquez, Andra Tolbus, Mikael Agerlin Petersen, Rasmus Bro


Spice mixes are cornerstones of kitchens all over the world. The distinct flavours that is most often used in a cuisine are basic principles that determine that cuisine, so-called flavours principles (E Rozin & Rozin, 1981; Elizabeth Rozin, 1983). Some years ago we set out to characterise a good pool of 29 spice mixes of various origins –some classical ones from different parts of the world, some experimental ones that we ourselves at the lab had created and one that was donated to us from noma. The outcome of the sensory descriptions from 26 kind food professionals that donated their time and delicate palates to us is described previously in Calibrating Flavour part I. There are many pros and some cons of the sensory methods that we used (Frøst, Giacalone, & Rasmussen, 2015). It is very fast to duo, and can be carried out in almost any location that is relatively quiet and without too much sensory distraction such as smells and fragrances. On the other hand, the results will not be as precise as a traditional descriptive analysis carried out by a trained sensory panel. But for our purpose, to get an overview of the interrelations between a large group of spice mixes, it was precise enough. The results are extremely useful, and were created with a very small financial burden to us. 

But there was much more to this audacious adventure. At University of Copenhagen’s Department of Food Science, the spice mixes were analysed by Headspace Gas Chromatography under conditions that simulate the conditions during eating a food, where the aromatic volatiles of the food enter the nose via the back of the throat (the retronasal pathway, (Shepherd, 2006)). Further, the aroma molecules were identified by mass spectroscopy (so-called GC-MS, Hübschmann, 2015).

The purpose of this was to study and visualise information of very different origins, but all characterising the same 29 spice mixes. There are 4 types of information that we can link:

  1. The sensory map (Map) – Using projective mapping to characterize the samples. These measurements are collected in a matrix of dimension 29 rows x 52 columns (26 evaluators x 2 coordinate axes).
  2. Sensory descriptors (Words) - In addition to the positions from the projective mapping, there are also a number of words that the sensory respondents (the food professionals) have attached to the spice mixes. After some textual analysis of the original set of words (totalling 545 different words), this set of data consists of a matrix with 29 rows and 113 columns – 113 different and relevant words that escribe the sensations of the spice mixes.
  3. Aroma profiling (Odour chemicals) - Through headspace sampling, GC-MS analysis was used to obtain an aroma profile for each spice mix. The aroma profile consisted of the relative integrated peak area (relative concentration) of 122 different odorous compounds. Some of the compounds have available sensory descriptors, collected in a smaller matrix (29 x 25 named Odour descriptors).
  4. Meta-data – name (Recipe), expressed solely as the fraction of each ingredient, the cultural identity or origin (geographical place or producer) and the base (oil, aqueous, dry, fermented and dairy). A matrix of 29 x 99 variables.

The data blocks are presented in figure 1. 

Figure 1: Overview of all data blocks available

By developing new visualization approaches and using tools from data fusion, we analysed, visualised, and explored these complex data structures. We used these tools to investigate the fundamental differences and commonalities in the set of spice mixes.  It is a work of very interdisciplinary nature, requiring data analysists, flavour chemists, food professionals and sensory scientists. The process was a feast for geeks, with wild and imaginative discussions, bouncing ideas and swimming in the data. However, the harsh realities hid us hard when submitting the manuscript to scientific journals. Because of the interdisciplinary nature, it was difficult for others to appreciate all aspects of the work, and we had it turned down from four different journals. Finally we had to realise that the ideas and results we presented to the scientific community would not be published in peer-reviewed scientific journals. With this blog post we present you the idea and the concept, and the opportunity to delve deeply into our investigation by reading the full manuscript here, where all details about the chemical and data analytical procedures are available, and the results are discussed in details. In addition, we find that the data are of a unique character. Thus we give access to the data, so that others may use them for further scrutiny. The data can be found here.

Excerpts of results

The projective mapping and the aroma profiling by GC-MS provide two complementary means to quantify the differences between spice mixes, and can be used to perform a combined analysis that shows to what extent sensory results are already contained in the aroma analysis and vice versa. The full data are very complex to present visually, and a low-rank sub-space of the total variation is obtained by multivariate data analysis, Principal Component Analysis (PCA) to be precise. PCA extracts the most important variance in the data, component by component. It allows focus on the most important part of the variance, in the analysis of the patterns of samples and the variables that describe them.

In figure 2, the first two and most important components of the PCA model are shown. These components explain 50% of the variability of the data. In particular, the score plot with all of the samples colored by origin (geographical or developed by), are plotted on the top left corner of the figure (Plot A). Plot B shows the words that the evaluators have used for describing the same samples. Plot C shows a loading plot of the volatile odour compounds. Finally, the recipe plot (Plot D) shows how the mixes can be interpreted on the basis of their ingredients. There is a robust relationship across all the matrices, demonstrating that the different matrices extract similar patterns about the spice mixes.


Figure 2A: Visual representation of the most important variance in the four data blocks
Sensory map and Sample origin group designated by simlar symbols (origins: Africa;  Asia;  Britain;  Central America;  Nordic Food Lab or noma  Scandinavia;  South America;  Southern Europe).  See more in full manuscript.

Figure 2B Sensory Words. The descriptors used by the panel to characterise the sensory properties of the spice mixes.

Figure 2C: Odour chemicals and descriptors: The different odorous compounds, and in some cases the sensory descriptors that are known to be associated with the compounds.

Figure 2D: Recipes: Expressed as the fraction of each ingredient, the cultural identity or origin (geographical place or producer) and the base (oil, aqueous, dry, fermented and dairy).

The aroma analysis using gas chromatography (Figure 2 plot C), allows us to dig further into the data. Observation of plot C shows a loading plot of the volatile odour compounds. On the same plot on top of the odour chemicals, the odour descriptors are shown. The plot shows how most of the odours are gathered in the upper-right part of the plot. This can be interpreted as a general trend of increased intensity of aroma, going from the samples in the bottom-left part of the scores plot towards those in the upper-right part. The dry base consists mainly of mixes of pure spices that have been ground up, which contain a high concentration of many highly volatile compounds, e.g. terpenes. The fermented mixes will undoubtedly have a considerable content of acids and umami compounds which are not captured by the GC analysis, but will contribute to the taste of the mixes. The oil-based mixes appear to contain a medium level of aroma compounds, but this may be due to higher aroma retention in the oil under the conditions for trapping the aroma compounds.  Finally, the recipe plot (Figure 2, plot C) shows how the paste grouping can be interpreted on the basis of their ingredients. One can see how the ingredients naturally reflect the groupings such as blueberry in the left part, juniper in the upper, and chipotle in the lower part. A very detailed analysis of the patterns is provided in the full manuscript.

It is a comprehensive way to fuse sensory projective mapping data, gastronomic information (recipes), and aroma profiling (gas chromatographic data). The method also allows for inclusion of additional data if available. More efforts are needed to help minimize the barriers that result from this highly cross-disciplinary experiment. There is a need for the ability to communicate across many fields of expertise such as chemistry, flavour research, gastronomy, mathematics, and sensory science. Finding a common language that allows open and creative communication is of paramount importance for the advancement of the field.

We made a large effort to improve the visualization of the problem by enhancing the readability of the most important loadings in all the loadings plots. However, work is still required and in the future we intend to expand on this, taking advantage of modern tools that computer science offers, such as interactive visualizations that allow a deeper exploration of complex data for which static plots are sometimes inadequate.

Acknowledgements and thanks
We kindly acknowledge the 26 persons that gave their valuable time to evaluate the spice mixes. In addition, we acknowledge noma for the donation of Ants and juniper mix. We acknowledge Kiki Sontiyart, Ana Caballero and Guillemette Barthouil for their contribution to the creation of the spice mixes. And lastly we thank the anonymous reviewers that helped us improve the manuscripts along the way.


Acar, E., Rasmussen, M. A., Savorani, F., Næs, T., & Bro, R. (2013). Understanding data fusion within the framework of coupled matrix and tensor factorizations. Chemometrics and Intelligent Laboratory Systems, 129, 53–63.

Frøst, M. B., Giacalone, D., & Rasmussen, K. K. (2015). Alternative methods of sensory testing: Working with chefs, culinary professionals and brew masters. Rapid Sensory Profiling Techniques and Related Methods: Applications in New Product Development and Consumer Research.

Hübschmann, H.-J. (2015). Handbook of GC/MS: Fundamentals and Applications (3rd ed.). Wiley & Sons.

Rozin, E. (1983). Ethnic Cuisine: The Flavor-Principle Cookbook. Brattleboro: The Steven Green Press.

Rozin, E., & Rozin, P. (1981). Culinary themes and variations. Natural History, 90, 6–14.

Shepherd, G. M. (2006). Smell images and the flavour system in the human brain. Nature, 444(7117), 316–321.