See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/51959358

Flavor network and the principles of food pairing

Article  in  Scientific Reports · November 2011

DOI: 10.1038/srep00196 · Source: arXiv

CITATIONS READS
274 10,385

4 authors:

Yong-Yeol Ahn Sebastian E Ahnert

Indiana University Bloomington University of Cambridge
49 PUBLICATIONS   5,628 CITATIONS    101 PUBLICATIONS   2,264 CITATIONS   

SEE PROFILE SEE PROFILE

James P. Bagrow Albert-Laszlo Barabasi

University of Vermont Northeastern University
81 PUBLICATIONS   3,065 CITATIONS    524 PUBLICATIONS   160,318 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Network resilience View project

Foodome Project View project

All content following this page was uploaded by Sebastian E Ahnert on 28 May 2014.

The user has requested enhancement of the downloaded file.

 Flavor network and the principles of food pairing
arXiv:1111.6074v1 [physics.soc-ph] 25 Nov 2011

Yong-Yeol Ahn,1,2,3† Sebastian E. Ahnert,1,4†∗ James P. Bagrow,1,2

Albert-László Barabási1,2∗

1
Center for Complex Network Research, Department of Physics
Northeastern University, Boston, MA 02115
2
Center for Cancer Systems Biology
Dana-Farber Cancer Institute, Harvard University, Boston, MA 02115
3
School of Informatics and Computing
Indiana University, Bloomington, IN 47408
4
Theory of Condensed Matter, Cavendish Laboratory
University of Cambridge, Cambridge CB3 0HE, UK

†
These authors contributed equally to this work.
∗
To whom correspondence should be addressed.
E-mail: sea31@cam.ac.uk (S.E.A.); alb@neu.edu (A.L.B.)

Abstract
The cultural diversity of culinary practice, as illustrated by the variety of regional
cuisines, raises the question of whether there are any general patterns that determine the
ingredient combinations used in food today or principles that transcend individual tastes
and recipes. We introduce a flavor network that captures the flavor compounds shared by
culinary ingredients. Western cuisines show a tendency to use ingredient pairs that share
many flavor compounds, supporting the so-called food pairing hypothesis. By contrast, East
Asian cuisines tend to avoid compound sharing ingredients. Given the increasing availabil-
ity of information on food preparation, our data-driven investigation opens new avenues
towards a systematic understanding of culinary practice.

As omnivores, humans have historically faced the difficult task of identifying and gather-
ing food that satisfies nutritional needs while avoiding foodborne illnesses [1]. This process
has contributed to the current diet of humans, which is influenced by factors ranging from an

1
evolved preference for sugar and fat to palatability, nutritional value, culture, ease of produc-
tion, and climate [1, 2, 3, 4, 5, 6, 7, 8, 9]. The relatively small number of recipes in use (∼ 106 ,

e.g. http://cookpad.com) compared to the enormous number of potential recipes (> 1015 , see
Supplementary Information Sec S1.2), together with the frequent recurrence of particular com-
binations in various regional cuisines, indicates that we are exploiting but a tiny fraction of the

potential combinations. Although this pattern itself can be explained by a simple evolutionary
model [10] or data-driven approaches [11], a fundamental question still remains: are there any
quantifiable and reproducible principles behind our choice of certain ingredient combinations

and avoidance of others?

Although many factors such as colors, texture, temperature, and sound play an important
role in food sensation [12, 13, 14, 15], palatability is largely determined by flavor, representing

a group of sensations including odors (due to molecules that can bind olfactory receptors), tastes
(due to molecules that stimulate taste buds), and freshness or pungency (trigeminal senses) [16].
Therefore, the flavor compound (chemical) profile of the culinary ingredients is a natural start-

ing point for a systematic search for principles that might underlie our choice of acceptable
ingredient combinations.
A hypothesis, which over the past decade has received attention among some chefs and food

scientists, states that ingredients sharing flavor compounds are more likely to taste well together
than ingredients that do not [17]. This food pairing hypothesis has been used to search for novel
ingredient combinations and has prompted, for example, some contemporary restaurants to

combine white chocolate and caviar, as they share trimethylamine and other flavor compounds,
or chocolate and blue cheese that share at least 73 flavor compounds. As we search for evidence
supporting (or refuting) any ‘rules’ that may underlie our recipes, we must bear in mind that the

scientific analysis of any art, including the art of cooking, is unlikely to be capable of explaining
every aspect of the artistic creativity involved. Furthermore, there are many ingredients whose

2
main role in a recipe may not be only flavoring but something else as well (e.g. eggs’ role to
ensure mechanical stability or paprika’s role to add vivid colors). Finally, the flavor of a dish

owes as much to the mode of preparation as to the choice of particular ingredients [12, 18,
19]. However, our hypothesis is that given the large number of recipes we use in our analysis
(56,498), such confounding factors can be systematically filtered out, allowing for the discovery

of patterns that may transcend specific dishes or ingredients.

Here we introduce a network-based approach to explore the impact of flavor compounds on
ingredient combinations. Efforts by food chemists to identify the flavor compounds contained

in most culinary ingredients allows us to link each ingredient to 51 flavor compounds on av-
erage [20]1 . We build a bipartite network [21, 22, 23, 24, 25, 26] consisting of two different
types of nodes: (i) 381 ingredients used in recipes throughout the world, and (ii) 1,021 flavor

compounds that are known to contribute to the flavor of each of these ingredients (Fig. 1A).
A projection of this bipartite network is the flavor network in which two nodes (ingredients)
are connected if they share at least one flavor compound (Fig. 1B). The weight of each link

represents the number of shared flavor compounds, turning the flavor network into a weighted
network [27, 22, 23]. While the compound concentration in each ingredient and the detection
threshold of each compound should ideally be taken into account, the lack of systematic data

prevents us from exploring their impact (see Sec S1.1.2 on data limitations).
Since several flavor compounds are shared by a large number of ingredients, the resulting
flavor network is too dense for direct visualization (average degree hki ' 214). We therefore

use a backbone extraction method [28, 29] to identify the statistically significant links for each
ingredient given the sum of weights characterizing the particular node (Fig. 2), see SI for de-
tails). Not surprisingly, each module in the network corresponds to a distinct food class such as

meats (red) or fruits (yellow). The links between modules inform us of the flavor compounds
1
While finalizing this manuscript, an updated edition (6th Ed.) of Fenaroli’s handbook of flavor ingredients has
been released.

3
that hold different classes of foods together. For instance, fruits and dairy products are close to
alcoholic drinks, and mushrooms appear isolated, as they share a statistically significant number

of flavor compounds only with other mushrooms.

The flavor network allows us to reformulate the food pairing hypothesis as a topological
property: do we more frequently use ingredient pairs that are strongly linked in the flavor net-

work or do we avoid them? To test this hypothesis we need data on ingredient combinations
preferred by humans, information readily available in the current body of recipes. For gen-
erality, we used 56,498 recipes provided by two American repositories (epicurious.com and

allrecipes.com) and to avoid a distinctly Western interpretation of the world’s cuisine, we also
used a Korean repository (menupan.com) (Fig. 1). The recipes are grouped into geographically
distinct cuisines (North American, Western European, Southern European, Latin American, and

East Asian; see Table S2). The average number of ingredients used in a recipe is around eight,
and the overall distribution is bounded (Fig. 1C), indicating that recipes with a very large or
very small number of ingredients are rare. By contrast, the popularity of specific ingredients

varies over four orders of magnitude, documenting huge differences in how frequently various
ingredients are used in recipes (Fig. 1D), as observed in [10]. For example, jasmine tea, Ja-
maican rum, and 14 other ingredients are each found in only a single recipe (see SI S1.2), but

egg appears in as many as 20,951, more than one third of all recipes.

Results

Figure 3D indicates that North American and Western European cuisines exhibit a statistically
significant tendency towards recipes whose ingredients share flavor compounds. By contrast,

East Asian and Southern European cuisines avoid recipes whose ingredients share flavor com-
pounds (see Fig. 3D for the Z-score, capturing the statistical significance of ∆Ns ). The system-
atic difference between the East Asian and the North American recipes is particularly clear if we

4
inspect the P (Nsrand ) distribution of the randomized recipe dataset, compared to the observed
number of shared compounds characterizing the two cuisines, Ns . This distribution reveals that

North American dishes use far more compound-sharing pairs than expected by chance (Fig. 3E),
and the East Asian dishes far fewer (Fig. 3F). Finally, we generalize the food pairing hypothesis
by exploring if ingredient pairs sharing more compounds are more likely to be used in specific

cuisines. The results largely correlate with our earlier observations: in North American recipes,
the more compounds are shared by two ingredients, the more likely they appear in recipes. By
contrast, in East Asian cuisine the more flavor compounds two ingredients share, the less likely

they are used together (Fig. 3G and 3H; see SI for details and results on other cuisines).
What is the mechanism responsible for these differences? That is, does Fig. 3C through
H imply that all recipes aim to pair ingredients together that share (North America) or do not

share (East Asia) flavor compounds, or could we identify some compounds responsible for the
bulk of the observed effect? We therefore measured the contribution χi of each ingredient to
the shared compound effect in a given cuisine c, quantifying to what degree its presence affects

the magnitude of ∆Ns .

In Fig. 3I,J we show as a scatter plot χi (horizontal axis) and the frequency fi for each
ingredient in North American and East Asian cuisines. The vast majority of the ingredients lie

on the χi = 0 axis, indicating that their contribution to ∆Ns is negligible. Yet, we observe
a few frequently used outliers, which tend to be in the positive χi region for North American
cuisine, and lie predominantly in the negative region for East Asian cuisine. This suggests that

the food pairing effect is due to a few outliers that are frequently used in a particular cuisine,
e.g. milk, butter, cocoa, vanilla, cream, and egg in the North America, and beef, ginger, pork,
cayenne, chicken, and onion in East Asia. Support for the definitive role of these ingredients is

provided in Fig. 3K,L where we removed the ingredients in order of their positive (or negative)
contributions to ∆Ns in the North American (or East Asian) cuisine, finding that the z-score,

5
which measures the significance of the shared compound hypothesis, drops below two after
the removal of only 13 (5) ingredients from North American (or East Asian) cuisine (see SI

S2.2.2). Note, however, that these ingredients play a disproportionate role in the cuisine under
consideration—for example, the 13 key ingredients contributing to the shared compound effect
in North American cuisine appear in 74.4% of all recipes.

According to an empirical view known as “the flavor principle” [30], the differences be-
tween regional cuisines can be reduced to a few key ingredients with specific flavors: adding
soy sauce to a dish almost automatically gives it an oriental taste because Asians use soy sauce

widely in their food and other ethnic groups do not; by contrast paprika, onion, and lard is a
signature of Hungarian cuisine. Can we systematically identify the ingredient combinations
responsible for the taste palette of a regional cuisine? To answer this question, we measure

the authenticity of each ingredient (pci ), ingredient pair (pcij ), and ingredient triplet (pcijk ) (see
Materials and Methods). In Fig. 4 we organize the six most authentic single ingredients, ingre-
dient pairs and triplets for North American and East Asian cuisines in a flavor pyramid. The

rather different ingredient classes (as reflected by their color) in the two pyramids capture the
differences between the two cuisines: North American food heavily relies on dairy products,
eggs and wheat; by contrast, East Asian cuisine is dominated by plant derivatives like soy sauce,

sesame oil, and rice and ginger. Finally, the two pyramids also illustrate the different affinities
of the two regional cuisines towards food pairs with shared compounds. The most authentic
ingredient pairs and triplets in the North American cuisine share multiple flavor compounds,

indicated by black links, but such compound-sharing links are rare among the most authentic
combinations in East Asian cuisine.
The reliance of regional cuisines on a few authentic ingredient combinations allows us to

explore the ingredient-based relationship (similarity or dissimilarity) between various regional

cuisines. For this we selected the six most authentic ingredients and ingredient pairs in each

6
regional cuisine (i.e. those shown in Fig. 4A,B), generating a diagram that illustrates the ingre-
dients shared by various cuisines, as well as singling out those that are unique to a particular

region (Fig. 4C). We once again find a close relationship between North American and West-
ern European cuisines and observe that when it comes to its signature ingredient combinations
Southern European cuisine is much closer to Latin American than Western European cuisine

(Fig. 4C).

Discussion

Our work highlights the limitations of the recipe data sets currently available, and more gener-
ally of the systematic analysis of food preparation data. By comparing two editions of the same
dataset with significantly different coverage, we can show that our results are robust against data

incompleteness (see SI S1.1.2). Yet, better compound databases, mitigating the incompleteness
and the potential biases of the current data, could significantly improve our understanding of
food. There is inherent ambiguity in the definition of a particular regional or ethnic cuisine.

However, as discussed in SI S1.2, the correlation between different datasets, representing two
distinct perspectives on food (American and Korean), indicates that humans with different eth-
nic background have a rather consistent view on the composition of various regional cuisines.

Recent work by Kinouchi et al. [10] observed that the frequency-rank plots of ingredients
are invariant across four different cuisines, exhibiting a shape that can be well described by a

Zipf-Mandelbrot curve. Based on this observation, they model the evolution of recipes by as-
suming a copy-mutate process, leading to a very similar frequency-rank curve. The copy-mutate
model provides an explanation for how an ingredient becomes a staple ingredient of a cuisine:

namely, having a high fitness value or being a founder. The model assigns each ingredient a
random fitness value, which represents the ingredient’s nutritional value, availability, flavor,
etc. For example, it has been suggested that each culture eagerly adopt spices that have high

7
anti-bacterial activity (e.g. garlic) [6, 7], spices considered to have high fitness. The mutation
phase of the model replaces less fit ingredients with more fit ones. Meanwhile, the copy mecha-

nism keeps copying the founder ingredients—ingredients in the early recipes—and makes them
abundant in the recipes regardless of their fitness value.
It is worthwhile to discuss the similarity and difference between the quantities we measured

and the concepts of fitness and founders. First of all, prevalence (Pic ) and authenticity (pci ) are
empirically measured values while fitness is an intrinsic hidden variable. Among the list of
highly prevalent ingredients we indeed find old ingredients—founders—that have been used in

the same geographic region for thousands of years. At the same time, there are relatively new
ingredients such as tomatoes, potatoes, and peppers that were introduced to Europe and Asia
just a few hundred years ago. These new, but prevalent ingredients can be considered to have

high fitness values. If an ingredient has a high level of authenticity, then it is prevalent in a
cuisine while not so prevalent in all other cuisines.
Indeed, each culture has developed their own authentic ingredients. It may indicate that

fitness can vary greatly across cuisines or that the stochasticity of recipe evolution diverge the
recipes in different regions into completely different sets. More historical investigation will
help us to estimate the fitness of ingredients and assess why we use the particular ingredients

we currently do. The higher order fitness value suggested in [10] is very close to our concept of
food pairing affinity.
Another difference in our results is the number of ingredients in recipes. Kinouchi et al.

reported that the average number of ingredients per recipe varies across different cookbooks.
While we also observed variation in the number of ingredients per recipe, the patterns we found
were not consistent with those found by Kinouchi et al. For instance, the French cookbook

has more ingredients per recipe than a Brazillian one, but in our dataset we find the opposite
result. We believe that a cookbook cannot represent a whole cuisine, and that cookbooks with

8
more sophisticated recipes will tend to have more ingredients per recipe than cookbooks with
everyday recipes. As more complete datasets become available, sharper conclusions can be

drawn regarding the size variation between cuisines.

Our contribution in this context is a study of the role that flavour compounds play in de-
termining these fitness values. One possible interpretation of our results is that shared flavor

compounds represent one of several contributions to fitness value, and that, while shared com-
pounds clearly play a significant role in some cuisines, other contributions may play a more
dominant role in other cuisines. The fact that recipes rely on ingredients not only for flavor but

also to provide the final textures and overall structure of a given dish provides support for the
idea that fitness values depend on a multitude of ingredient characteristics besides their flavor
profile.

In summary, our network-based investigation identifies a series of statistically significant

patterns that characterize the way humans choose the ingredients they combine in their food.
These patterns manifest themselves to varying degree in different geographic regions: while

North American and Western European dishes tend to combine ingredients that share flavor
compounds, East Asian cuisine avoids them. More generally this work provides an example
of how the data-driven network analysis methods that have transformed biology and the social

sciences in recent years can yield new insights in other areas, such as food science.

Methods
Shared compounds

To test the hypothesis that the choice of ingredients is driven by an appreciation for ingredient

pairs that share flavor compounds (i.e. those linked in Fig. 2), we measured the mean number
of shared compounds in each recipe, Ns , comparing it with Nsrand obtained for a randomly con-
structed reference recipe dataset. For a recipe R that contains nR different ingredients, where

9
each ingredient i has a set of flavor compounds Ci , the mean number of shared compounds
2 X
Ns (R) = |Ci ∩ Cj | (1)
nR (nR − 1) i,j∈R,i6=j

is zero if none of the ingredient pairs (i, j) in the recipe share any flavor compounds. For ex-
ample, the ‘mustard cream pan sauce’ recipe contains chicken broth, mustard, and cream, none

of which share any flavor compounds (Ns (R) = 0) in our dataset. Yet, Ns (R) can reach as high
as 60 for ‘sweet and simple pork chops’, a recipe containing apple, pork, and cheddar cheese
(See Fig. 3A). To check whether recipes with high Ns (R) are statistically preferred (implying

the validity of the shared compound hypothesis) in a cuisine c with Nc recipes, we calculate
∆Ns = Nsreal − Nsrand , where ‘real’ and ‘rand’ indicates real recipes and randomly constructed
P
recipes respectively and Ns = R Ns (R)/Nc (see SI for details of the randomization pro-

cess). This random reference (null model) controls for the frequency of a particular ingredient
in a given regional cuisine, hence our results are not affected by historical, geographical, and
climate factors that determine ingredient availability (see SI S1.1.2).

Contribution

The contribution χi of each ingredient to the shared compound effect in a given cuisine c,
quantifying to what degree its presence affects the magnitude of ∆Ns , is defined by
  P !
1 X 2 X 2f i j∈c fj |Ci ∩ Cj |
χi =  |Ci ∩ Cj | − P , (2)
Nc R3i nR (nR − 1) Nc hnR i j∈c fj
j6=i(j,i∈R)

where fi represents the ingredient i’s number of occurrence. An ingredient’s contribution is

positive (negative) if it increases (decreases) ∆Ns .

Authenticity

we define the prevalence Pic of each ingredient i in a cuisine c as Pic = nci /Nc , where nci is
the number of recipes that contain the particular ingredient i in the cuisine and Nc is the total

10
 0
number of recipes in the cuisine. The relative prevalence pci = Pic − hPic ic0 6=c measures the
authenticity—the difference between the prevalence of i in cuisine c and the average prevalence

of i in all other cuisines. We can also identify ingredient pairs or triplets that are overrepresented
in a particular cuisine relative to other cuisines by defining the relative pair prevalences pcij =
0 0
Pijc − hPijc ic0 6=c and triplet prevalences pcijk = Pijk
c c
− hPijk ic0 6=c , with Pijc = ncij /Nc and Pijk
c
=

ncijk /Nc .

References

[1] Rozin, P. The selection of foods by rats, humans, and other animals. Advances in the
Study of Behavior 7, 21–76 (1976).

[2] Pfaffman, C. Olfaction and taste V, chap. Phylogenetic origins of sweet sensitivity (New
York: Academic Press, 1975).

[3] Garcia, J. & Hankins, W. G. Olfaction and taste V, chap. The evolution of bitter and the

acquisition of toxiphobia (New York: Academic Press, 1975).

[4] Drewnowski, A. & Greenwood, M. R. C. Cream and sugar: human preferences for high-
fat foods. Physiology & Behavior 30, 629–633 (1983).

[5] Diamond, J. M. Guns, germs, and steel: The fates of human societies (W.W. Norton, New
York, 1997).

[6] Billing, J. & Sherman, P. W. Antimicrobial functions of spices: why some like it hot. The

Quarterly Review of Biology 73, 3–49 (1998).

[7] Sherman, P. W. & Hash, G. A. Why vegetable recipes are not very spicy. Evolution and
Human Behavior 22, 147–163 (2001).

11
 [8] Harris, M. Good to eat: riddles of food and culture (Waveland Press, 1998).

[9] Counihan, C. & van Esterik, P. (eds.) Food and culture (Routledge, 2007).

[10] Kinouchi, O., Diez-Garcia, R. W., Holanda, A. J., Zambianchi, P. & Roque, A. C. The
non-equilibrium nature of culinary evolution. New Journal of Physics 10, 073020 (2008).

[11] Teng, C.-Y., Lin, Y.-R. & Adamic, L. A. Recipe recommendation using ingredient net-

works (2011). ArXiv:1111.3919 [cs.SI].

[12] This, H. Molecular gastronomy: exploring the science of flavor (Columbia University
Press, 2005).

[13] Johnson, J. & Clydesdale, F. Perceived sweetness and redness in colored sucrose solutions.
Journal of Food Science 47, 747–752 (1982).

[14] Shankaer, M. U. & Levitan, C. A. Grape expectations: the role of cognitive influences in

color-flavor interactions. Conscious Cogn. 19, 380–390 (2010).

[15] Zampini, M. & Spence, C. The role of auditory cues in modulating the perceived crispness
and staleness of potato chips. Journal of Sensory Studies 19, 347–363 (2005).

[16] Breslin, P. A. S. & Beauchamp, G. K. Suppression of bitterness by sodium: variation

among bitter taste stimuli. Chem. Senses 20, 609–623 (1995).

[17] Blumenthal, H. The big fat duck cookbook (Bloomsbury, London, 2008).

[18] This, H. Molecular gastronomy, a scientific look at cooking. Accounts of Chemical Re-
search 42, 575–583 (2009).

[19] McGee, H. On food and cooking: the science and lore of the kitchen (Scribner, 2004).

12
[20] Burdock, G. A. Fenaroli’s handbook of flavor ingredients (CRC Press, 2004), 5th edn.

[21] Newman, M. E. J., Barabási, A.-L. & Watts, D. J. The structure and dynamics of networks:
(Princeton University Press, 2006).

[22] Caldarelli, G. Scale-free networks: complex webs in nature and technology (Oxford Uni-

versity Press, USA, 2007).

[23] Dorogovtsev, S. N. & Mendes, J. F. F. Evolution of networks: from biological nets to the
internet and WWW (Oxford University Press, USA, 2003).

[24] Albert, R. & Barabási, A.-L. Statistical mechanics of complex networks. Rev. Mod. Phys.
74, 47 (2002).

[25] Newman, M. E. J. The structure and function of complex networks. SIAM Review 45,

167–256 (2003).

[26] Dorogovtsev, S. N., Goltsev, A. V. & Mendes, J. F. F. Critical phenomena in complex

networks. Reviews of Modern Physics 80, 1275–61 (2008).

[27] Barrat, A., Barthélemy, M., Pastor-Satorras, R. & Vespignani, A. The architecture of
complex weighted networks. Proc. Nat. Acad. Sci. 111, 3747 (2004).

[28] Serrano, M. A., Boguñá, M. & Vespignani, A. Extracting the multiscale backbone of
complex weighted networks. Proceedings of the National Academy of Sciences 106, 6483–

6488 (2009).

[29] Lee, S. H., Kim, P.-J., Ahn, Y.-Y. & Jeong, H. Googling social interactions: web search
engine based social network construction. PLoS One 5, e11233 (2010).

[30] Rozin, E. The flavor-principle cookbook (Hawthorn Books; Book Club Ed. edition, 1973).

13
Acknowledgements

We thank M. I. Meirelles, S. Lehmann, D. Kudayarova, T. W. Lim, J. Baranyi, H. This for

discussions. This work was supported by the James S. McDonnell Foundation 21st Century

Initiative in Studying Complex Systems.

Author contributions

YYA, SEA, and ALB designed research and YYA, SEA, and JPB performed research. All
authors wrote and reviewed the manuscript.

14
A Ingredients Flavor compounds B Flavor network
shrimp 1-penten-3-ol
white 2-hexenal
Shrimp scampi and tomato broil

wine 2-isobutyl thiazole

mozzarella 2,3-diethylpyrazine Prevalence
2,4-nonadienal
parmesan 3-hexen-1-ol
4-hydroxy-5-methyl...
olive 4-methylpentanoic acid
oil acetylpyrazine
allyl 2-furoate
parsley alpha-terpineol
beta-cyclodextrin
cis-3-hexenal
dihydroxyacetone
tomato
dimethyl succinate

garlic
ethyl propionate
hexyl alcohol
isoamyl alcohol
garlic isobutyl acetate
isobutyl alcohol
lauric acid
limonene (d-,l-, and dl-)
Seasoned mussels

scallion
l-malic acid
sesame oil methyl butyrate tomato Shared compounds
methyl hexanoate
starch methyl propyl trisulfide
nonanoic acid
olive
soy
sauce phenethyl alcohol
propenyl propyl disulfide
oil
nut
propionaldehyde
black
black propyl disulfide pepper
pepper p-mentha-1,3-diene
p-menth-1-ene-9-al
terpinyl acetate
sake
tetrahydrofurfuryl alcohol
mussel trans, trans-2,4-hexadienal
...

100
C 0.15
North American
D
0.12 Western European 10-1
Frequency, f(r)

Southern European
0.09 Latin American
10-2 North American
P(s)

East Asian -3 Western European

0.06 10
Southern European
0.03 10-4 Latin American
East Asian
0 10-5
0 10 20 30 1 10 100 1000
Number of ingredients per recipe (s) Rank, r

Figure 1: Flavor network. (A) The ingredients contained in two recipes (left column), to-
gether with the flavor compounds that are known to be present in the ingredients (right column).
Each flavor compound is linked to the ingredients that contain it, forming a bipartite network.
Some compounds (shown in boldface) are shared by multiple ingredients. (B) If we project the
ingredient-compound bipartite network into the ingredient space, we obtain the flavor network,
whose nodes are ingredients, linked if they share at least one flavor compound. The thickness
of links represents the number of flavor compounds two ingredients share and the size of each
circle corresponds to the prevalence of the ingredients in recipes. (C) The distribution of recipe
size, capturing the number of ingredients per recipe, across the five cuisines explored in our
study. (D) The frequency-rank plot of ingredients across the five cuisines show an approxi-
mately invariant distribution across cuisines.

15
16
Figure 2: The backbone of the flavor network. Each node denotes an ingredient, the node color indicates food category,
and node size reflects the ingredient prevalence in recipes. Two ingredients are connected if they share a significant number
of flavor compounds, link thickness representing the number of shared compounds between the two ingredients. Adjacent
links are bundled to reduce the clutter. Note that the map shows only the statistically significant links, as identified by the
algorithm of Refs.[28, 29] for p-value 0.04. A drawing of the full network is too dense to be informative. We use, however,
the full network in our subsequent measurements.
 2 10
A Many shared compounds B Few shared compounds C D 8 45.5

Shrimp 6
Coffee Beef Lemon 1
4

Z
2

6Ns
0
132 102 97 63 9 69 0
-2
-1 -4
North Western Latin Southern East North Western Latin Southern East
Ns = 102 Ns = 9 American European American European Asian American European American European Asian

0.6 0.6
E F I North American J East Asian
0.5 0.5
0.2 0.2
egg butter
0.4 0.4 cayenne
North American, East Asian,
p << 10-3 p 5 1.6 = 10-2 ginger rice
0.3 0.3
milk

P(Ns)
P(Ns)
0.1 0.1 onion
0.2 vanilla 0.2
beef
tomato cream

Frequency of use
0.1 0.1
pork
cocoa
cream cheese chicken
0 0 0 0
9.5 10 10.5 11 11.5 12 6 6.2 6.4 6.6 6.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 -0.4 -0.2 0 0.2 0.4
Ns Ns Compound contribution, ri Compound contribution, ri
1 1 2
G North American
H East Asian
K 50 North American L
all
linear regression linear regression 1
0.8 0.8 r = <0.31 40 milk
p 5 0.0027 chicken
0
30 cayenne
0.6 0.6
cream pork
butter -1

PR
PR
20 cocoa egg ginger
peanut butter beef
0.4 0.4
vanilla strawberry -2
10
cream cheese

17
0.2 0.2

z-score (shared compounds)

r = 0.59 0 -3
all
p 5 9.0 = 10-11 East Asian
0 0 -10 -4
0 30 60 90 120 0 30 60 90 120 0 5 10 15 20 25 30 0 5 10 15 20 25 30
Ns Ns Number of removed ingredients Number of removed ingredients

Figure 3: Testing the food pairing hypothesis. Schematic illustration of two ingredient pairs, the first sharing many more (A)
and the second much fewer (B) compounds than expected if the flavor compounds were distributed randomly. (C,D) To test
the validity of the food pairing hypothesis, we construct 10,000 random recipes and calculate ∆Ns . We find that ingredient
pairs in North American cuisines tend to share more compounds while East Asian cuisines tend to share fewer compounds
than expected in a random recipe dataset. (E,F) The distributions P (Ns ) for 10,000 randomized recipe datasets compared
with the real values for East Asian and North American cuisine. Both cuisines exhibit significant p-values, as estimated using
a z-test. (G,H) We enumerate every possible ingredient pair in each cuisine and show the fraction of pairs in recipes as a
function of the number of shared compounds. To reduce noise, we only used data points calculated from more than 5 pairs.
The p-values are calculated using a t-test. North American cuisine is biased towards pairs with more shared compounds while
East Asian shows the opposite trend (see SI for details and results for other cuisines). Note that we used the full network,
not the backbone shown in Fig. 2 to obtain these results. (I,J) The contribution and frequency of use for each ingredient in
North American and East Asian cuisine. The size of the circles represents the relative prevalence pci . North American and
East Asian cuisine shows the opposite trends. (K,L) If we remove the highly contributing ingredients sequentially (from
the largest contribution in North American cuisine and from the smallest contribution in East Asian cuisine), the shared
compounds effect quickly vanishes when we removed five (East Asian) to fifteen (North American) ingredients.
A North American Number of
shared compounds

B East Asian

C Co-occurrence in recipes
Latin American
Southern European corn

onion East Asian

basil

tomato cayenne
parmesan scallion
cheese
olive oil
garlic Soybean
sesame oil
macaroni
North American
soy sauce Rice
milk
milk
ginger
wheat
cane
molasses
vanilla butter
cream
Western European
thyme
egg

Figure 4: Flavor principles. (A,B) Flavor pyramids for North American and East Asian
cuisines. Each flavor pyramid shows the six most authentic ingredients (i.e. those with the
largest pci ), ingredient pairs (largest pcij ), and ingredient triplets (largest pcijk ). The size of the
nodes reflects the abundance Pic of the ingredient in the recipes of the particular cuisine. Each
color represents the category of the ingredient (see Fig. 2 for the color) and link thickness indi-
cates the number of shared compounds. (C) The six most authentic ingredients and ingredient
pairs used in specific regional cuisine. Node color represents cuisine and the link weight reflects
the relative prevalence pci of the ingredient pair.

18
Supporting Online Material
Flavor network and the principles of food pairing
by Yong-Yeol Ahn, Sebastian E. Ahnert, James P. Bagrow, Albert-László Barabási

Table of Contents
S1 Materials and methods 2
S1.1 Flavor network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
S1.1.1 Ingredient-compounds bipartite network . . . . . . . . . . . . . . . . . 2
S1.1.2 Incompleteness of data and the third edition . . . . . . . . . . . . . . . 4
S1.1.3 Extracting the backbone . . . . . . . . . . . . . . . . . . . . . . . . . 6
S1.1.4 Sociological bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
S1.2 Recipes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
S1.2.1 Size of recipes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
S1.2.2 Frequency of recipes . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
S1.3 Number of shared compounds . . . . . . . . . . . . . . . . . . . . . . . . . . 14
S1.4 Shared compounds hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . 14
S1.4.1 Null models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
S1.4.2 Ingredient contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 18

List of Figures
1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
S1 Full ingredient network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
S2 Degree distribution of flavor network . . . . . . . . . . . . . . . . . . . . . . . 3
S3 Comparing the third and fifth edition of Fenaroli’s . . . . . . . . . . . . . . . . 5
S4 Backbone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
S5 Potential biases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
S6 Coherency of datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
S7 Number of ingredients per recipe . . . . . . . . . . . . . . . . . . . . . . . . . 12
S8 The distribution of duplicated recipes . . . . . . . . . . . . . . . . . . . . . . . 13
S9 Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
S10 Null models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
S11 Shared compounds and usage . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

List of Tables
S1 Statistics of 3rd and 5th editions . . . . . . . . . . . . . . . . . . . . . . . . . 4
S2 Recipe dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

1
 S3 Coherence of cuisines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
S4 Each cuisine’s average number of ingredients per recipe . . . . . . . . . . . . . 11
S5 Top contributors in North American and East Asian cuisines . . . . . . . . . . 19

S1 Materials and methods

S1.1 Flavor network

S1.1.1 Ingredient-compounds bipartite network

The starting point of our research is Fenaroli’s handbook of flavor ingredients (fifth edition [1]),
which offers a systematic list of flavor compounds and their natural occurrences (food ingre-

dients). Two post-processing steps were necessary to make the dataset appropriate for our
research: (A) In many cases, the book lists the essential oil or extract instead of the ingredient
itself. Since these are physically extracted from the original ingredient, we associated the flavor

compounds in the oils and extracts with the original ingredient. (B) Another post-processing
step is including the flavor compounds of a more general ingredient into a more specific ingre-
dient. For instance, the flavor compounds in ‘meat’ can be safely assumed to also be in ‘beef’

or ‘pork’. ‘Roasted beef’ contains all flavor compounds of ‘beef’ and ‘meat’.

The ingredient-compound association extracted from [1] forms a bipartite network. As the

name suggests, a bipartite network consists of two types of nodes, with connections only be-
tween nodes of different types. Well known examples of bipartite networks include collabora-

tion networks of scientists [2] (with scientists and publications as nodes) and actors [3] (with
actors and films as nodes), or the human disease network [4] which connects health disorders
and disease genes. In the particular bipartite network we study here, the two types of nodes are

food ingredients and flavor compounds, and a connection signifies that an ingredient contains a

2
compound.

The full network contains 1,107 chemical compounds and 1,531 ingredients, but only 381

Figure S1: The full flavor network. The

size of a node indicates average preva-
lence, and the thickness of a link repre-
sents the number of shared compounds.
All edges are drawn. It is impossible to ob-
serve individual connections or any modu-
lar structure.

103 103 102

102 102
N(kc)
N(ki)

N(k)

101

101 101

100 100 100

100 101 102 103 100 101 102 103 100 101 102 103
Ingredient degree, ki Compound degree, kc Degree in ingredient network, k

104 104 104

N(kc) (cumulative)

103 103 103

N(ki) (cumulative)

N(k) (cumulative)

-0.5

102 102 102

101 101 101

100 100 100

100 101 102 103 100 101 102 103 100 101 102 103
Ingredient degree, ki Compound degree, kc Degree in ingredient network, k

Figure S2: Degree distributions of the flavor network. Degree distribution of ingredients in
the ingredient-compound network, degree distribution of flavor compounds in the ingredient-
compound network, and degree distribution of the (projected) ingredient network, from left to
right. Top: degree distribution. Bottom: complementary cumulative distribution. The line and
the exponent value in the leftmost figure at the bottom is purely for visual guide.

3
 3rd eds. 5th eds.
# of ingredients 916 1507
# of compounds 629 1107
# of edges in I-C network 6672 36781
Table S1: The basic statistics on two different datasets. The 5th Edition of Fenaroli’s handbook
contains much more information than the third edition.

ingredients appear in recipes, together containing 1,021 compounds (see Fig. S1). We project

this network into a weighted network between ingredients only [5, 6, 7, 27]. The weight of
each edge wij is the number of compounds shared between the two nodes (ingredients) i and
j, so that the relationship between the M × M weighted adjacency matrix wij and the N × M

bipartite adjacency matrix aik (for ingredient i and compound k) is given by:
N
X
wij = aik ajk (S3)
k=1

The degree distributions of ingredients and compounds are shown in Fig. S2.

S1.1.2 Incompleteness of data and the third edition

The situation encountered here is similar to the one encountered in systems biology: we do not
have a complete database of all protein, regulatory and metabolic interactions that are present

in the cell. In fact, the existing protein interaction data covers less than 10% of all protein
interactions estimated to be present in the human cell [9].

To test the robustness of our results against the incompleteness of data, we have performed
the same calculations for the 3rd edition of Fenaroli’s handbook as well. The 5th edition con-
tains approximately six times more information on the chemical contents of ingredients (Ta-

ble S1). Yet, our main result is robust (Fig. S3), further supporting that data incompleteness is
not the main factor behind our findings.

4
 0.3
14 55.7
0.25 12

0.2 10
8
0.15
∆Ns

Z
0.1 4
2
0.05
0
0
-2
-0.05 -4
North Western Latin Southern East North Western Latin Southern East
American European American European Asian American European American European Asian

1.8
14 45.5
1.6
1.4 12
1.2 10
1 8
0.8
∆Ns

6
Z

0.6
4
0.4
2
0.2
0 0

-0.2 -2
-0.4 -4
North Western Latin Southern East North Western Latin Southern East
American European American European Asian American European American European Asian

Figure S3: Comparing the third and fifth edition of Fenaroli’s to see if incomplete data impacts
our conclusions. The much sparser data of the 3rd edition (Top) shows a very similar trend
to that of the 5th edition (Bottom, repeated from main text Fig. 3). Given the huge difference
between the two editions (Table S1), this further supports that the observed patterns are robust.

5
 Plant
deriv
ative
s

er
les

ble oil

green bell pepp

b
eta

carrot pper
ses
Pla

olivee oiloil
g

vinegar
Ve

ustard
co lery oil rup
nts

uce

mber
vegeta

bell pe
s

i
parsnip
ola

sy

tomato juice

cucuhin
sesam

m
soy sa

turnip
rutabaga
can spirit h

lettuce
an
leek

zucc i

as mpki an
cin
dis

sc shroo
oil
le

black tea
soyb ine teant oil

ake
cassava

s
oa

beaabe

m paragun
coconut
em

n
tea

pu k be
cocffee

m
map

por
tea
oil
jasm ermi
woorsera

green

n
l

ean

onallion
onio
lemeast e pee

shiit
pepp

p ion
lim

blac
ce

ted
Nu

out
oil
el

sm ao d

roas

ls spr
u

gaam ato
nut
ho
oke

pea
pe

oil

sse
ts

c m ic
er

cac

to
y ang

dive
bru
flow

y ot
sun
on

o l
d
an

c auli a
ee

en

ble
v abbaflower
r
rap ech
es

rh uerkabi ge ge
or
be
um

rel
el
d

pe

lau

ub raut ta
nu
t
nu

rin
an
ta on pea ed

paanda
se

pis octed nuhat zelnut d me se

rleoaf lm
od um
wcaopsic t

t
o
m

y
s e arb
ko e
e

croasasted asted sesa

p q ler
h
sa hlr
ds

ch u t

ch oy a a s
roroaste
io

s e u
c

kp an
ea
ut n t

ic be
ro
to
an pecanu
e t ory
ta
po
d
peastedal mond n
be hic ve
ke
ba
w ecaw al
d
ro
c i ch
ip
n
ea
te

ckhaleney cboli
to
as
ro ta
p ashet po

ss
re
c
nuntil ed ut eed kid roc ish
le e b d rc ke
s eln e s
z d ra ate ho bean
haesamon w rticg bean ey
s amun kidn
alm red
l r
ea via

s
fishlmon h
tm ca

food
oa real
e eat
Cere

ce g
rrin

sa tty kfis
ric he

wh e bread d
brea
eat it
fa oc
hadd allo
p
sc

Sea
wh
whckwheaurt d on roe
al

flo le colm
bu
sa sh
catfiussel fish
ryeg nood
eg lt
m oked
oatrn grit mp
ma
sm wn

shri
pra
d
brea n
co
crab uob
ushi

ster salmon
kats lfish
bran popcor shel
rice rn oy
cobread tuna
oked
smweed
sea r
lobusste
rye ley
bar rice at flour
brown grain whe octop id
squ eon caviar
whole
ni sturg
caro
maflake clam h

Flowers
crayfis
corn

lavender
turkey tulip
orange flower
meat sausage flower
rose
smoked bergamot
geranium
liver gardenia
chamomile
pork violet
chicken liver
mutton
pork liver
beef pear brandy
gin
grape brandy
chicken rum
Meats

beefroasted sake
beef broth cabernet sauvignon
port wine wine
lamb whiskey
frankfurte cider
cognac
r
chicken bro
th veal wine
ouzo
brandy

ic
boiled
pork
fried chicke
beef liver n
che
apprry
ham

Alcohol
bra
le bra ndy

drinks
grilled
arm ndy
black agnac
porkbeef
e
cured sausag
pork
baco sherry
berry
brand
y

n be
er
cham
wine
wh pagne wi
re ite w ne
an
peise sherd win ine
vanpper ry e
illa m
h
la oney
c ace
n orian
cinutmegder egrd
ta nam
ca ma
ve
clo bo
neg oil

raw rin on ga blu

b
c ay d lan
ga
ec

l
tu ay
pa u t t heese

ima

ts
c rm en
ta ard e ne
rm e r

duc
so
c ba amric yo
bu ur
tte mi

g u esa

An
rfa lk
g umsco om ca
ch mem
t
Sp

s ed
rt
sta ing in pep
crwiss ch nc

pro
da be
r ch rt
eechee

c r ma ea eese he
pe e an r
se se
is lic
ro
ice

r
bslaffroen saenedise r ice
or

es
ac mroquefoertamno cheem e
s

k ilk chee chse

pe gelylicbasilory

fe
an
ho
pp a

se e

ta
bu ilk eseese
pe
sa

es
erm

ro intil
v se

ch rm t
m hege ch
tte f a e
mhervage gree

pp e
in

ee il
ccotta
t

c v u
lo n y

se k
m

fe a

e
b nnel
ar

r
bmlu
sp

an ric
fe ill

iry
aplive
ea
y

da
d lm
rm

erinb t
okiwi ngot
ba
int

paasil tro

m urr

Da
er
c ate r oran
b ilan m
k

d e

a an

o
bitterry t pear

ry
c arjora e

b
rsle

bartleberry a
litc
huckl na it
thyge n

bpaassion fru
hi
sa a g o

pblaeckacurrantge
tarrongrass

et
lem

ora
fr
blackcord grary
chayot
y

coniper bery
gua

jun strawberr
cher e

wild

uit
perus
graprine

ge
cit

r
nectae

prick
va

lim ju

pinter
wa
fig fruit

muscat
grape

raspbepe

n
okraerine
apple

tangjuice
e

grape
blackberry

papaya
lingonberry

raspberry
quince

ac
ly pear
avocado
lime

kumquat
lemon juice

strawberry
cranberry

plum
melon
lemon
orange juice

eapplon
ry

grape

rry
me

He
le
ice

rbs

Fruits

Figure S4: The backbone of the ingredient network extracted according to [10] with a signifi-
cance threshold p = 0.04. Color indicates food category, font size reflects ingredient prevalence
in the dataset, and link thickness represents the number of shared compounds between two in-
gredients.

S1.1.3 Extracting the backbone

The network’s average degree is about 214 (while the number of nodes is 381). It is very dense

and thus hard to visualize (see Fig. S1). To circumvent this high density, we use a method that
extracts the backbone of a weighted network [10], along with the method suggested in [11].
For each node, we keep those edges whose weight is statistically significant given the strength

(sum of weight) of the node. If there is none, we keep the edge with the largest weight. A
different visualization of this backbone is presented in Fig. S4. Ingredients are grouped into

6
categories and the size of the name indicates the prevalence. This representation clearly shows
the categories that are closely connected.

S1.1.4 Sociological bias

Western scientists have been leading food chemistry, which may imply that western ingredients

are more studied. To check if such a bias is present in our dataset, we first made two lists
of ingredients: one is the list of ingredients appearing in North American cuisine, sorted by
the relative prevalence pci (i.e. the ingredients more specific to North American cuisine comes

first). The other is a similar list for East Asian cuisine. Then we measured the number of flavor
compounds for ingredients in each list. The result in Fig. S5A shows that any potential bias, if
present, is not significant.

There is another possibility, however, if there is bias such that the dataset tends to list more
familiar (Western) ingredients for more common flavor compounds, then we should observe a

correlation between the familiarity (frequently used in Western cuisine) and the degree of com-
pound (number of ingredients it appears in) in the ingredient. Figure S5B shows no observable
correlation, however.

S1.2 Recipes

The number of potential ingredient combinations is enormous. For instance, one could generate

∼ 1015 distinct ingredient combinations by choosing eight ingredients (the current average per
recipe) from approximately 300 ingredients in our dataset. If we use the numbers reported in
Kinouchi et al. [12] (1000 ingredients and 10 ingredients per recipe), one can generate ∼ 1023

ingredient combinations. This number greatly increases if we consider the various cooking
methods. Regardless, the fact that this number exceeds by many orders of magnitude the ∼ 106

7
A North American Western European Southern European Latin American East Asian
300 300 300 300 300
Number of compounds

Number of compounds

Number of compounds

Number of compounds

Number of compounds
p-value = 0.01 (*) p-value = 0.07 p-value = 0.23 p-value = 0.04 (*) p-value = 0.12
250 250 250 250 250
200 200 200 200 200
150 150 150 150 150
100 100 100 100 100
50 50 50 50 50
0 0 0 0 0
0 100 200 300 0 100 200 300 0 100 200 300 0 100 200 300 0 100 200 300
B
400 400 400 400 400
Avg. compound degree

Avg. compound degree

Avg. compound degree

Avg. compound degree

Avg. compound degree

350 p-value = 0.82 350 p-value = 0.99 350 p-value = 0.53 350 p-value = 0.50 350 p-value = 0.72
300 300 300 300 300
250 250 250 250 250
200 200 200 200 200
150 150 150 150 150
100 100 100 100 100
50 50 50 50 50
0 0 0 0 0
0 100 200 300 0 100 200 300 0 100 200 300 0 100 200 300 0 100 200 300
Ingredient rank Ingredient rank Ingredient rank Ingredient rank Ingredient rank

Figure S5: Are popular, much-used ingredients more studied than less frequent foods, leading
to potential systematic bias? (A) We plot the number of flavor compounds for each ingredient as
a function of the (ranked) popularity of the ingredient. The correlation is very small compared
to the large fluctuations present. There is a weak tendency that the ingredients mainly used in
North American or Latin American cuisine tend to have more odorants, but the correlations are
weak (with coefficients of -0.13 and -0.10 respectively). A linear regression line is shown only
if the corresponding p-value is smaller than 0.05. (B) If there is bias such that the book tends
to list more familiar ingredients for more common flavor compounds, then we can observe the
correlation between the familiarity (how frequently it is used in the cuisine) and the degree of
the compound in the ingredient-compound network. The plots show no observable correlations
for any cuisine.

8
recipes listed in the largest recipe repositories (e.g. http://cookpad.com) indicates that
humans are exploiting a tiny fraction of the culinary space.

We downloaded all available recipes from three websites: allrecipes.com, epicurious.com,

and menupan.com. Recipes tagged as belonging to an ethnic cuisine are extracted and then

grouped into 11 larger regional groups. We used only 5 groups that each contain more than
1,000 recipes (See Table S2). In the curation process, we made a replacement dictionary for
frequently used phrases that should be discarded, synonyms for ingredients, complex ingredi-

ents that are broken into ingredients, and so forth. We used this dictionary to automatically
extract the list of ingredients for each recipe. As shown in Fig. 1D, the usage of ingredients
is highly heterogenous. Egg, wheat, butter, onion, garlic, milk, vegetable oil, and cream ap-

pear more than 10,000 recipes while geranium, roasted hazelnut, durian, muscat grape, roasted
pecan, roasted nut, mate, jasmine tea, jamaican rum, angelica, sturgeon caviar, beech, lilac
flower, strawberry jam, and emmental cheese appear in only one recipe. Table S3 shows the

correlation between ingredient usage frequency in each cuisine and in each dataset. Figure. S6
shows that the three datasets qualitatively agree with each other, offering a base to combine
these datasets.

S1.2.1 Size of recipes

We reports the size of the recipes for each cuisine in Table S4. Overall, the mean number of

ingredients per recipe is smaller than that reported in Kinouchi et al. [12]. We believe that it
is mainly due to the different types of data sources. There are various types of recipes: from
quick meals to ones used in sophisticated dishes of expensive restaurants; likewise, there are

also various cookbooks. The number of ingredients may vary a lot between recipe datasets. If a
book focuses on sophisticated, high-level dishes then it will contain richer set of ingredients per

9
Table S2: Number of recipes and the detailed cuisines in each regional cuisine in the recipe
dataset. Five groups have reasonably large size. We use all cuisine data when calculating the
relative prevalence and flavor principles.
Cuisine set Number of recipes Cuisines included
North American 41525 American, Canada, Cajun, Creole, Southern
soul food, Southwestern U.S.
Southern European 4180 Greek, Italian, Mediterranean, Spanish, Por-
tuguese
Latin American 2917 Caribbean, Central American, South American,
Mexican
Western European 2659 French, Austrian, Belgian, English, Scottish,
Dutch, Swiss, German, Irish
East Asian 2512 Korean, Chinese, Japanese
Middle Eastern 645 Iranian, Jewish, Lebanese, Turkish
South Asian 621 Bangladeshian, Indian, Pakistani
Southeast Asian 457 Indonesian, Malaysian, Filipino, Thai, Viet-
namese
Eastern European 381 Eastern European, Russian
African 352 Moroccan, East African, North African, South
African, West African
Northern European 250 Scandinavian

Epicurious vs. Allrecipes Epicurious vs. Menupan Allrecipes vs. Menupan

North American 0.93 N/A N/A
East Asian 0.94 0.79 0.82
Western European 0.92 0.88 0.89
Southern European 0.93 0.83 0.83
Latin American 0.94 0.69 0.74
African 0.89 N/A N/A
Eastern European 0.93 N/A N/A
Middle Eastern 0.87 N/A N/A
Northern European 0.77 N/A N/A
South Asian 0.97 N/A N/A
Southeast Asian 0.92 N/A N/A
Table S3: The correlation of ingredient usage between different datasets. We see that the differ-
ent datasets broadly agree on what constitutes a cuisine, at least at a gross level.

10
 2
Epicurious
Allrecipes
Menupan (Korean)
rand

1
- Ns
real
Ns

-1
East Southern Latin Western North
Asian European American European American

Figure S6: Comparison between different datasets. The results on different datasets qualita-
tively agree with each other (except Latin American cuisine). Note that menupan.com is a
Korean website.

North American 7.96

Western European 8.03
Southern European 8.86
Latin American 9.38
East Asian 8.96
Northern European 6.82
Middle Eastern 8.39
Eastern European 8.39
South Asian 10.29
African 10.45
Southeast Asian 11.32
Table S4: Average number of ingredients per recipe for each cuisine.

11
recipe; if a book focuses on simple home cooking recipes, then the book will contain fewer in-
gredients per recipe. We believe that the online databases are close to the latter; simpler recipes

are likely to dominate the database because anyone can upload their own recipes. By contrast,
we expect that the cookbooks, especially the canonical ones, contain more sophisticated and
polished recipes, which thus are more likely to contain more ingredients.

Also, the pattern reported in Kinouchi et al. [12] is reversed in our dataset: Western Euro-
pean cuisine has 8.03 ingredients per recipe while Latin American cuisine has 9.38 ingredients

per recipe. Therefore, we believe that there is no clear tendency of the number of ingredients
per recipe between Western European and Latin American cuisine.

Yet, there seems to be an interesting trend in our dataset that hotter countries use more
ingredients per recipe, probably due to the use of more herbs and spices [13, 14] or due to
more diverse ecosystems. (6.82 in Northern European vs. 11.31 in Southeast Asian). Figure S7

shows the distribution of recipe size in all cuisines.

A 0.15 B 0.18
North American African
Western European 0.15 South Asian
0.12 Southern European Southeast Asian
Latin American Middle Eastern
East Asian 0.12 Eastern European
0.09
Northern European
P(s)

P(s)

0.09
0.06
0.06

0.03
0.03

0 0
0 10 20 30 0 10 20 30 40
Number of ingredients per recipe (s) Number of ingredients per recipe (s)

Figure S7: Number of ingredients per recipe. North American and Western European cuisine
shows similar distribution while other cuisines have slightly more ingredients per recipe.

12
 100000

Number of recipes with D duplicates

10000

1000

100

10

1
1 10 100
Number of duplicates, D

Figure S8: If a recipe is very popular, the recipe databases will have a tendency to list more
variations of the recipe. This plot shows that there are many duplicated recipes that share the
same set of ingredients. The number of duplicates exhibits a heavy-tailed distribution.

S1.2.2 Frequency of recipes

In contrast to previous work [12] that used published cookbooks, we use online databases. Al-

though recipes online are probably less canonical than established cookbooks, online databases
allow us to study much larger dataset more easily. Another important benefit of using online
databses is that there is no real-estate issue in contrast to physical cookbooks that should care-

fully choose what to include. Adding a slight variation of a recipe costs virtually nothing to the
websites and even enhances the quality of the database. Therefore, one can expect that online
databases capture the frequency of recipes more accurately than cookbooks.

Certain recipes (e.g. signature recipes of a cuisine) are much more important than others;
They are cooked much more frequently than others. Figure S8 shows that there are many du-

plicated recipes (possessing identical sets of ingredients), indicating that popularity is naturally
encoded in these datasets.

13
S1.3 Number of shared compounds

Figure S9 explains how to measure the number of shared compounds in a hypothetical recipe

with three ingredients.

S1.4 Shared compounds hypothesis

S1.4.1 Null models

In order to test the robustness of our findings, we constructed several random recipe datasets

using a series of appropriate null models and compare the mean number of shared compounds
Ns between the real and the randomized recipe sets. The results of these null models are sum-
marized in Fig. S10, each confirming the trends discussed in the paper. The null models we

used are:

(A, B) Frequency-conserving. Cuisine c uses a set of nc ingredients, each with frequency fi .

For a given recipe with Ni ingredients in this cuisine, we pick Ni ingredients randomly

from the set of all nc ingredients, according to fi . That is the more frequently an ingredi-
ent is used, the more likely the ingredient is to be picked. It preserves the prevalence of

Shared Compounds

a 2 a
c c
b f
d g

0
1
d 2+1+0
=1
3

Binary
Figure S9: For a recipe with three ingredients, we count the number of shared compounds in
a 1 a
every possible pair of ingredients, and
c divide it by the number
c of possible pair of ingredients.
b f
d g
1
14
0
d 1+1+0
=0.67
3
 A 2
B 10

8
Frequency Frequency
conserving
rand

conserving 6
1
- Ns

Z
2
real

0
Ns

0
-2

-4
North Western Latin Southern East North Western Latin Southern East
American European American European Asian American European American European Asian
C 1
D 10
Frequency and ingredient 9 Frequency and ingredient
category preserving 8 category preserving
7
rand

6
5
- Ns

Z 4
3
real

2
1
Ns

0 0
-1
-2
-3
North Western Latin East Southern North Western Latin East Southern
American European American Asian European American European American Asian European
E 5
F 50
4 Uniform random 40 Uniform random
3
rand

30
2
- Ns

20
1
Z

10
real

0
0
Ns

-1
-2 -10

-3 -20
North Southern Western Latin East North Southern Western Latin East
American European European American Asian American European European American Asian
G 2
H 50
category preserving 40 category preserving
and uniform random and uniform random
1
rand

30
- Ns

20
0
Z

10
real

0
Ns

-1
-10

-2 -20
North Southern Latin Western East North Southern Latin Western East
American European American European Asian American European American European Asian

Figure S10: Four different null models. Although the size of the discrepancy between cuisines
varies greatly, the overall trend is stable.

15
 each ingredient. This is the null model presented in the main text.

(C, D) Frequency and ingredient category preserving. With this null model, we conserve
the category (meats, fruits, etc) of each ingredient in the recipe, and when sample ran-

dom ingredients proportional to the prevalence. For instance, a random realization of a

recipe with beef and onion will contain a meat and a vegetable. The probability to pick
an ingredient is proportional to the prevalence of the ingredient in the cuisine.

(E, F) Uniform random. We build a random recipe by randomly choosing an ingredient that

is used at least once in the particular cuisine. Even very rare ingredients will frequently
appear in random recipes.

(G, H) Uniform random, ingredient category preserving. For each recipe, we preserve the
category of each ingredient, but not considering frequency of ingredients.

Although these null models greatly change the frequency and type of ingredients in the

random recipes, North American and East Asian recipes show a robust pattern: North American
recipes always share more flavor compounds than expected and East Asian recipes always share
less flavor compounds than expected. This, together with the existence of both positive and

negative Nsreal − Nsrand in every null model, indicates that the patterns we find are not due to a
poorly selected null models.

Finally, Fig. S11 shows the probability that a given pair with certain number of shared
compounds will appear in the recipes, representing the raw data behind the generalized food-

pairing hypothesis discussed in the text. To reduce noise, we only consider Ns where there are
more than five ingredient pairs.

16
 0.9 North
0.7 American
P(n)

0.5
0.3
0.1
0.9 Western Southern
0.7 European European
P(n)

0.5
0.3
0.1
0.9 Latin East
0.7 American Asian
P(n)

0.5
0.3
0.1
0 50 100 150 0 50 100 150
Number of shared compounds Number of shared compounds

Figure S11: The probability that ingredient pairs that share a certain number of compounds also
appear in the recipes. We enumerate every possible ingredient pair in each cuisine and show the
fraction of pairs in recipes as a function of the number of shared compounds. To reduce noise,
we only used data points calculated from more than 5 pairs.

17
S1.4.2 Ingredient contributions

To further investigate the contrasting results on the shared compound hypothesis for different

cuisines, we calculate the contribution of each ingredient and ingredient pair to ∆Ns . Since
Ns (R) for a recipe R is defined as

2 X
Ns (R) = |Ci ∩ Cj | (S4)
nR (nR − 1) i,j∈R,i6=j

(where nR is the number of ingredients in the recipe R), the contribution from an ingredient

pair (i, j) can be calculated as following:

!  
1 X 2 fi fj 2
χcij = |Ci ∩ Cj | − |Ci ∩ Cj | , (S5)
Nc R3i,j nR (nR − 1) Nc2 hnR i (hnR i − 1)

where fi indicates the ingredient i’s number of occurrences. Similarly, the individual contribu-
tion can be calculated:
  P !
1 X 2 2fi j∈c fj |Ci ∩ Cj |
X
χci =  |Ci ∩ Cj | − P . (S6)
Nc R3i nR (nR − 1) Nc hnR i j∈c fj
j6=i(j,i∈R)

We list in Table. S5 the top contributors in North American and East Asian cuisines.

18
 North American East Asian
Ingredient i χi Ingredient i χi
milk 0.529 rice 0.294
butter 0.511 red bean 0.152
cocoa 0.377 milk 0.055
vanilla 0.239 green tea 0.041
cream 0.154 butter 0.041
cream cheese 0.154 peanut 0.038
egg 0.151 mung bean 0.036
Positive peanut butter 0.136 egg 0.033
strawberry 0.106 brown rice 0.031
cheddar cheese 0.098 nut 0.024
orange 0.095 mushroom 0.022
lemon 0.095 orange 0.016
coffee 0.085 soybean 0.015
cranberry 0.070 cinnamon 0.014
lime 0.065 enokidake 0.013
tomato -0.168 beef -0.2498
white wine -0.0556 ginger -0.1032
beef -0.0544 pork -0.0987
onion -0.0524 cayenne -0.0686
chicken -0.0498 chicken -0.0662
tamarind -0.0427 onion -0.0541
vinegar -0.0396 fish -0.0458
Negative pepper -0.0356 bell pepper -0.0414
pork -0.0332 roasted sesame seed -0.0410
celery -0.0329 black pepper -0.0409
bell pepper -0.0306 shrimp -0.0408
red wine -0.0271 shiitake -0.0329
black pepper -0.0248 garlic -0.0302
parsley -0.0217 carrot -0.0261
parmesan cheese -0.0197 tomato -0.0246
Table S5: Top 15 (both positive and negative) contributing ingredients to each cuisine.

19
References

[1] Burdock, G. A. Fenaroli’s handbook of flavor ingredients (CRC Press, 2004), 5th edn.

[2] Newman, M. E. J. The structure of scientific collaboration networks. Proc. Natl. Acad.

Sci. USA 98, 404 (2001).

[3] Watts, D. J. & Strogatz, S. H. Collective dynamics of ‘small-world’ networks. Natue 393,

440 (1998).

[4] Goh, K.-I. et al. The human disease network. Proc. Natl. Acad. Sci. USA 104, 8685
(2007).

[5] Newman, M. E. J., Barabási, A.-L. & Watts, D. J. The structure and dynamics of networks:
(Princeton University Press, 2006).

[6] Caldarelli, G. Scale-free networks: complex webs in nature and technology (Oxford Uni-
versity Press, USA, 2007).

[7] Dorogovtsev, S. N. & Mendes, J. F. F. Evolution of networks: from biological nets to the

internet and WWW (Oxford University Press, USA, 2003).

[8] Barrat, A., Barthélemy, M., Pastor-Satorras, R. & Vespignani, A. The architecture of
complex weighted networks. Proc. Nat. Acad. Sci. 111, 3747 (2004).

[9] Venkatesan, K. et al. An empirical framework for binary interactome mapping. Nature
Methods 6, 83–90 (2009).

[10] Serrano, M. A., Boguñá, M. & Vespignani, A. Extracting the multiscale backbone of
complex weighted networks. Proceedings of the National Academy of Sciences 106, 6483–
6488 (2009).

20
 [11] Lee, S. H., Kim, P.-J., Ahn, Y.-Y. & Jeong, H. Googling social interactions: web search
engine based social network construction. PLoS One 5, e11233 (2010).

[12] Kinouchi, O., Diez-Garcia, R. W., Holanda, A. J., Zambianchi, P. & Roque, A. C. The

non-equilibrium nature of culinary evolution. New Journal of Physics 10, 073020 (2008).

[13] Billing, J. & Sherman, P. W. Antimicrobial functions of spices: why some like it hot. The
Quarterly Review of Biology 73, 3–49 (1998).

[14] Sherman, P. W. & Hash, G. A. Why vegetable recipes are not very spicy. Evolution and
Human Behavior 22, 147–163 (2001).

21

View publication stats