Flavor network and the principles of food pairing

Yong-Yeol Ahn,
1,2,3
Sebastian E. Ahnert,
1,4
James P. Bagrow,
1,2
Albert-L aszl o Barab asi
1,2
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 avor network that captures the avor compounds shared by
culinary ingredients. Western cuisines show a tendency to use ingredient pairs that share
many avor 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 difcult task of identifying and gather-
ing food that satises nutritional needs while avoiding foodborne illnesses [1]. This process
has contributed to the current diet of humans, which is inuenced by factors ranging from an
1
evolved preference for sugar and fat to palatability, nutritional value, culture, ease of produc-
tion, and climate [2, 3, 1, 4, 5, 6, 7, 8, 9]. The relatively small number of recipes in use ( 10
6
,
e.g. http://cookpad.com) compared to the enormous number of potential recipes (> 10
15
, 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
quantiable 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 avor, 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 avor compound (chemical) prole 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 avor 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 avor compounds,
or chocolate and blue cheese that share at least 73 avor compounds. As we search for evidence
supporting (or refuting) any rules that may underlie our recipes, we must bear in mind that the
scientic 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 avoring but something else as well (e.g. eggs role to
ensure mechanical stability or paprikas role to add vivid colors). Finally, the avor 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 ltered out, allowing for the discovery
of patterns that may transcend specic dishes or ingredients.
Here we introduce a network-based approach to explore the impact of avor compounds on
ingredient combinations. Efforts by food chemists to identify the avor compounds contained
in most culinary ingredients allows us to link each ingredient to 51 avor 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 avor
compounds that are known to contribute to the avor of each of these ingredients (Fig. 1A).
A projection of this bipartite network is the avor network in which two nodes (ingredients)
are connected if they share at least one avor compound (Fig. 1B). The weight of each link
represents the number of shared avor compounds, turning the avor 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 avor compounds are shared by a large number of ingredients, the resulting
avor network is too dense for direct visualization (average degree k 214). We therefore
use a backbone extraction method [28, 29] to identify the statistically signicant 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 avor compounds
1
While nalizing this manuscript, an updated edition (6th Ed.) of Fenarolis handbook of avor 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 signicant number
of avor compounds only with other mushrooms.
The avor 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 avor 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 worlds 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 specic 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
signicant tendency towards recipes whose ingredients share avor compounds. By contrast,
East Asian and Southern European cuisines avoid recipes whose ingredients share avor com-
pounds (see Fig. 3D for the Z-score, capturing the statistical signicance of N
s
). The system-
atic difference between the East Asian and the North American recipes is particularly clear if we
4
inspect the P(N
rand
s
) distribution of the randomized recipe dataset, compared to the observed
number of shared compounds characterizing the two cuisines, N
s
. 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 specic
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 avor 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) avor compounds, or could we identify some compounds responsible for the
bulk of the observed effect? We therefore measured the contribution
i
of each ingredient to
the shared compound effect in a given cuisine c, quantifying to what degree its presence affects
the magnitude of N
s
.
In Fig. 3I,J we show as a scatter plot
i
(horizontal axis) and the frequency f
i
for each
ingredient in North American and East Asian cuisines. The vast majority of the ingredients lie
on the
i
= 0 axis, indicating that their contribution to N
s
is negligible. Yet, we observe
a few frequently used outliers, which tend to be in the positive
i
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 denitive role of these ingredients is
provided in Fig. 3K,L where we removed the ingredients in order of their positive (or negative)
contributions to N
s
in the North American (or East Asian) cuisine, nding that the z-score,
5
which measures the signicance 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
considerationfor 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 avor principle [30], the differences be-
tween regional cuisines can be reduced to a few key ingredients with specic avors: 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 (p
c
i
), ingredient pair (p
c
ij
), and ingredient triplet (p
c
ijk
) (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 avor pyramid. The
rather different ingredient classes (as reected 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 afnities
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 avor 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 nd 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 signicantly 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 signicantly improve our understanding of
food. There is inherent ambiguity in the denition 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 tness value or being a founder. The model assigns each ingredient a
random tness value, which represents the ingredients nutritional value, availability, avor,
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 tness. The mutation
phase of the model replaces less t ingredients with more t ones. Meanwhile, the copy mecha-
nism keeps copying the founder ingredientsingredients in the early recipesand makes them
abundant in the recipes regardless of their tness value.
It is worthwhile to discuss the similarity and difference between the quantities we measured
and the concepts of tness and founders. First of all, prevalence (P
c
i
) and authenticity (p
c
i
) are
empirically measured values while tness is an intrinsic hidden variable. Among the list of
highly prevalent ingredients we indeed nd old ingredientsfoundersthat 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 tness 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
tness 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 tness of ingredients and assess why we use the particular ingredients
we currently do. The higher order tness value suggested in [10] is very close to our concept of
food pairing afnity.
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 nd 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 avour compounds play in de-
termining these tness values. One possible interpretation of our results is that shared avor
compounds represent one of several contributions to tness value, and that, while shared com-
pounds clearly play a signicant 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 avor but
also to provide the nal textures and overall structure of a given dish provides support for the
idea that tness values depend on a multitude of ingredient characteristics besides their avor
prole.
In summary, our network-based investigation identies a series of statistically signicant
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 avor
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 avor compounds (i.e. those linked in Fig. 2), we measured the mean number
of shared compounds in each recipe, N
s
, comparing it with N
rand
s
obtained for a randomly con-
structed reference recipe dataset. For a recipe R that contains n
R
different ingredients, where
9
each ingredient i has a set of avor compounds C
i
, the mean number of shared compounds
N
s
(R) =
2
n
R
(n
R
1)

i,jR,i=j
|C
i
C
j
| (1)
is zero if none of the ingredient pairs (i, j) in the recipe share any avor compounds. For ex-
ample, the mustard cream pan sauce recipe contains chicken broth, mustard, and cream, none
of which share any avor compounds (N
s
(R) = 0) in our dataset. Yet, N
s
(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 N
s
(R) are statistically preferred (implying
the validity of the shared compound hypothesis) in a cuisine c with N
c
recipes, we calculate
N
s
= N
real
s
N
rand
s
, where real and rand indicates real recipes and randomly constructed
recipes respectively and N
s
=

R
N
s
(R)/N
c
(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
i
of each ingredient to the shared compound effect in a given cuisine c,
quantifying to what degree its presence affects the magnitude of N
s
, is dened by

i
=

1
N
c

Ri
2
n
R
(n
R
1)

j=i(j,iR)
|C
i
C
j
|

2f
i
N
c
n
R

jc
f
j
|C
i
C
j
|

jc
f
j

, (2)
where f
i
represents the ingredient is number of occurrence. An ingredients contribution is
positive (negative) if it increases (decreases) N
s
.
Authenticity
we dene the prevalence P
c
i
of each ingredient i in a cuisine c as P
c
i
= n
c
i
/N
c
, where n
c
i
is
the number of recipes that contain the particular ingredient i in the cuisine and N
c
is the total
10
number of recipes in the cuisine. The relative prevalence p
c
i
= P
c
i
P
c

i

c

=c
measures the
authenticitythe 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 dening the relative pair prevalences p
c
ij
=
P
c
ij
P
c

ij

=c
and triplet prevalences p
c
ijk
= P
c
ijk
P
c

ijk

=c
, with P
c
ij
= n
c
ij
/N
c
and P
c
ijk
=
n
c
ijk
/N
c
.
References
[1] Rozin, P. The selection of foods by rats, humans, and other animals. Advances in the
Study of Behavior 7, 2176 (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, 629633 (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, 349 (1998).
[7] Sherman, P. W. & Hash, G. A. Why vegetable recipes are not very spicy. Evolution and
Human Behavior 22, 147163 (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 avor (Columbia University
Press, 2005).
[13] Johnson, J. &Clydesdale, F. Perceived sweetness and redness in colored sucrose solutions.
Journal of Food Science 47, 747752 (1982).
[14] Shankaer, M. U. & Levitan, C. A. Grape expectations: the role of cognitive inuences in
color-avor interactions. Conscious Cogn. 19, 380390 (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, 347363 (2005).
[16] Breslin, P. A. S. & Beauchamp, G. K. Suppression of bitterness by sodium: variation
among bitter taste stimuli. Chem. Senses 20, 609623 (1995).
[17] Blumenthal, H. The big fat duck cookbook (Bloomsbury, London, 2008).
[18] This, H. Molecular gastronomy, a scientic look at cooking. Accounts of Chemical Re-
search 42, 575583 (2009).
[19] McGee, H. On food and cooking: the science and lore of the kitchen (Scribner, 2004).
12
[20] Burdock, G. A. Fenarolis handbook of avor ingredients (CRC Press, 2004), 5th edn.
[21] Newman, M. E. J., Barab asi, 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. & Barab asi, 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,
167256 (2003).
[26] Dorogovtsev, S. N., Goltsev, A. V. & Mendes, J. F. F. Critical phenomena in complex
networks. Reviews of Modern Physics 80, 127561 (2008).
[27] Barrat, A., Barth elemy, M., Pastor-Satorras, R. & Vespignani, A. The architecture of
complex weighted networks. Proc. Nat. Acad. Sci. 111, 3747 (2004).
[28] Serrano, M. A., Bogu n a, 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 avor-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.
Additional information
Competing nancial interests
The authors declare no competing nancial interests.
14
Figure 1: Flavor network. (A) The ingredients contained in two recipes (left column), to-
gether with the avor compounds that are known to be present in the ingredients (right column).
Each avor 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 avor network,
whose nodes are ingredients, linked if they share at least one avor compound. The thickness
of links represents the number of avor 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 ve cuisines explored in our
study. (D) The frequency-rank plot of ingredients across the ve cuisines show an approxi-
mately invariant distribution across cuisines.
Figure 2: The backbone of the avor network. Each node denotes an ingredient, the node
color indicates food category, and node size reects the ingredient prevalence in recipes. Two
ingredients are connected if they share a signicant number of avor 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 signicant links, as
identied 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.
15
Figure 3: Testing the food pairing hypothesis. Schematic illustration of two ingredient pairs,
the rst sharing many more (A) and the second much fewer (B) compounds than expected if
the avor compounds were distributed randomly. (C,D) To test the validity of the food pairing
hypothesis, we construct 10,000 random recipes and calculate N
s
. We nd 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(N
s
) for 10,000 randomized recipe datasets compared with the real values for East Asian and
North American cuisine. Both cuisines exhibit signicant 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 p
c
i
. 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 ve (East Asian) to
fteen (North American) ingredients.
Figure 4: Flavor principles. (A,B) Flavor pyramids for North American and East Asian
cuisines. Each avor pyramid shows the six most authentic ingredients (i.e. those with the
largest p
c
i
), ingredient pairs (largest p
c
ij
), and ingredient triplets (largest p
c
ijk
). The size of the
nodes reects the abundance P
c
i
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 specic regional cuisine. Node color represents cuisine and the link weight reects
the relative prevalence p
c
i
of the ingredient pair.
16
black
pepper
garlic
tomato
olive
oil
0
0.03
0.06
0.09
0.12
0.15
0 10 20 30
P
(
s
)
Number of ingredients per recipe (s)
C D
North American
Western European
Southern European
Latin American
East Asian
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
1 10 100 1000
F
r
e
q
u
e
n
c
y
,

f
(
r
)
Rank, r
C D
North American
Western European
Southern European
Latin American
East Asian
A B Flavor network
...
propenyl propyl disulde
cis-3-hexenal
2-isobutyl thiazole
2-hexenal
trans, trans-2,4-hexadienal
1-penten-3-ol
acetylpyrazine
dihydroxyacetone
beta-cyclodextrin
isobutyl acetate
dimethyl succinate
phenethyl alcohol
limonene (d-,l-, and dl-)
terpinyl acetate
methyl hexanoate
p-mentha-1,3-diene
3-hexen-1-ol
p-menth-1-ene-9-al
alpha-terpineol
methyl propyl trisulde
propionaldehyde
ethyl propionate
nonanoic acid
4-methylpentanoic acid
tetrahydrofurfuryl alcohol
allyl 2-furoate
4-hydroxy-5-methyl...
2,3-diethylpyrazine
lauric acid
l-malic acid
isoamyl alcohol
2,4-nonadienal
methyl butyrate
isobutyl alcohol
hexyl alcohol
propyl disulde
Ingredients Flavor compounds
tomato
olive
oil
mozzarella
shrimp
parsley
parmesan
white
wine
garlic
sesame oil
starch
sake
mussel
nut
black
pepper
soy
sauce
scallion
S
h
r
i
m
p

s
c
a
m
p
i

a
n
d

t
o
m
a
t
o

b
r
o
i
l
S
e
a
s
o
n
e
d

m
u
s
s
e
l
s
Prevalence
Shared compounds
fruits
dairy
spices
alcoholic beverages
nuts and seeds
seafoods
meats
herbs
plant derivatives
vegetables
flowers
animal products
plants
cereal
Categories
Prevalence
Shared
compounds
-1
0
1
2
North
American
Western
European
Latin
American
Southern
European
East
Asian
U
N
s
A B C D
E F
G H
I J
K L
-4
-2
0
2
4
6
8
10
North
American
Western
European
Latin
American
Southern
European
East
Asian
Z
0
0.1
0.2
6 6.2 6.4 6.6 6.8
P
(
N
s
)
N
s
East Asian,
p T 1.6 10
-2
0
0.1
0.2
9.5 10 10.5 11 11.5 12
P
(
N
s
)
N
s
North American,
p << 10
-3
0
0.2
0.4
0.6
0.8
1
0 30 60 90 120
P
R
N
s
North American
linear regression
0
0.2
0.4
0.6
0.8
1
0 30 60 90 120
P
R
N
s
r = 0.59
p T 9.0 10
-11
r = [0.31
p T 0.0027
East Asian
linear regression
0
0.1
0.2
0.3
0.4
0.5
0.6
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
F
r
e
q
u
e
n
c
y

o
f

u
s
e
Compound contribution,
i
North American
milk
butter
cocoa
vanilla
cream
cream cheese
egg
tomato
0
0.1
0.2
0.3
0.4
0.5
0.6
-0.4 -0.2 0 0.2 0.4
Compound contribution,
i
East Asian
rice
beef
ginger
pork
cayenne
chicken
onion
-10
0
10
20
30
40
50
0 5 10 15 20 25 30
z
-
s
c
o
r
e

(
s
h
a
r
e
d

c
o
m
p
o
u
n
d
s
)
Number of removed ingredients
all
milk
butter
cocoa
vanilla
cream
cream cheese
egg
peanut butter
strawberry
North American
-4
-3
-2
-1
0
1
2
0 5 10 15 20 25 30
Number of removed ingredients
all
beef
ginger
pork
cayenne
chicken
East Asian
Many shared compounds
N
s
= 102
Coffee Beef
132 102 97
Few shared compounds
N
s
= 9
Shrimp
Lemon
63 9 69
45.5
North American
A
Number of
shared compounds
East Asian
B
Co-occurrence in recipes
C
corn
ginger
Southern European
Rice
East Asian
Latin American
milk
wheat
cream
vanilla
egg
butter
cane
molasses
Western European
North American
thyme
onion
tomato
garlic
olive oil
basil
parmesan
cheese
macaroni
soy sauce
Soybean
sesame oil
scallion
cayenne
tomato
garlic
cayenne
milk
wheat
butter
vanilla
egg
Supporting Online Material
Flavor network and the principles of food pairing
by Yong-Yeol Ahn, Sebastian E. Ahnert, James P. Bagrow, Albert-L aszl o Barab asi
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 . . . . . . . . . . . . . . . . . 3
S1.1.3 Extracting the backbone . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
S1.1.4 Sociological bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
S1.2 Recipes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
S1.2.1 Size of recipes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
S1.2.2 Frequency of recipes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
S1.3 Number of shared compounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
S1.4 Shared compounds hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
S1.4.1 Null models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
S1.4.2 Ingredient contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
List of Figures
S1 Full ingredient network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
S2 Degree distribution of avor network . . . . . . . . . . . . . . . . . . . . . . . . . 3
S3 Comparing the third and fth edition of Fenarolis . . . . . . . . . . . . . . . . . . 4
S4 Backbone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
S5 Potential biases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
S6 Coherency of datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
S7 Number of ingredients per recipe . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
S8 The distribution of duplicated recipes . . . . . . . . . . . . . . . . . . . . . . . . . 10
S9 Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
S10 Null models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
S11 Shared compounds and usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
List of Tables
S1 Statistics of 3rd and 5th editions . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
S2 Recipe dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
S3 Coherence of cuisines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
S4 Each cuisines average number of ingredients per recipe . . . . . . . . . . . . . . . 9
S5 Top contributors in North American and East Asian cuisines . . . . . . . . . . . . 15
1
S1 Materials and methods
S1.1 Flavor network
S1.1.1 Ingredient-compounds bipartite network
The starting point of our research is Fenarolis handbook of avor ingredients (fth edition [1]),
which offers a systematic list of avor compounds and their natural occurrences (food ingredients).
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 avor compounds in the oils
and extracts with the original ingredient. (B) Another post-processing step is including the avor
compounds of a more general ingredient into a more specic ingredient. For instance, the avor
compounds in meat can be safely assumed to also be in beef or pork. Roasted beef contains
all avor 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 between
nodes of different types. Well known examples of bipartite networks include collaboration net-
works of scientists [2] (with scientists and publications as nodes) and actors [3] (with actors and
lms 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 ingredi-
ents and avor compounds, and a connection signies that an ingredient contains a compound.
The full network contains 1,107 chemical compounds and 1,531 ingredients, but only 381
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, 8]. The weight of each edge
w
i j
is the number of compounds shared between the two nodes (ingredients) i and j, so that the
relationship between the MM weighted adjacency matrix w
i j
and the NM bipartite adjacency
Figure S1: The full avor network. The size of a
node indicates average prevalence, and the thickness
of a link represents the number of shared compounds.
All edges are drawn. It is impossible to observe indi-
vidual connections or any modular structure.
2
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 Fenarolis handbook contains much more
information than the third edition.
matrix a
ik
(for ingredient i and compound k) is given by:
w
i j
=
N

k=1
a
ik
a
jk
(S1)
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
10
0
10
1
10
2
10
3
10
0
10
1
10
2
10
3
N
(
k
i
)
Ingredient degree, k
i
10
0
10
1
10
2
10
3
10
0
10
1
10
2
10
3
N
(
k
c
)
Compound degree, k
c
10
0
10
1
10
2
10
0
10
1
10
2
10
3
N
(
k
)
Degree in ingredient network, k
10
0
10
1
10
2
10
3
10
4
10
0
10
1
10
2
10
3
N
(
k
i
)

(
c
u
m
u
l
a
t
i
v
e
)
Ingredient degree, k
i
-0.5
10
0
10
1
10
2
10
3
10
4
10
0
10
1
10
2
10
3
N
(
k
c
)

(
c
u
m
u
l
a
t
i
v
e
)
Compound degree, k
c
10
0
10
1
10
2
10
3
10
4
10
0
10
1
10
2
10
3
N
(
k
)

(
c
u
m
u
l
a
t
i
v
e
)
Degree in ingredient network, k
Figure S2: Degree distributions of the avor network. Degree distribution of ingredients in the ingredient-compound
network, degree distribution of avor 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 gure at the bottom is purely for visual guide.
3
-0.05
0
0.05
0.1
0.15
0.2
0.25
0.3
North
American
Western
European
Latin
American
Southern
European
East
Asian

N
s

-4
-2
0
2
4
6
8
10
12
14
North
American
Western
European
Latin
American
Southern
European
East
Asian
Z
55.7
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
North
American
Western
European
Latin
American
Southern
European
East
Asian

N
s

-4
-2
0
2
4
6
8
10
12
14
North
American
Western
European
Latin
American
Southern
European
East
Asian
Z
45.5
Figure S3: Comparing the third and fth edition of Fenarolis 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.
4
A
lc
o
h
o
lic
d
r
in
k
s
A
n
im
a
l
p
r
o
d
u
c
t
s
D
a
ir
y
Fruits
H
e
rb
s
S
p
ic
e
s
M
e
a
t
s
C
e
r
e
a
l
N
u
t s
a
n
d
s
e
e
d
s
P
l a
n
t s
P l a n t d e r i v a t i v e s
V
e
g
e
t a
b
l e
s
S
e
a
f
o
o
d
s
F
l o
w
e
r
s
a
l m
o
n
d
le
e
k
b
lu
e
b
e
rry
w
o o d
caviar
c o r n f l a k e
fish
mandarin
b a c o n
l e m
o n g r a s s
a
p
ric
o
t
o
liv
e
rutabaga
herring
pear brandy
p
a
rs
n
ip
b
l a
c
k
p
e
p
p
e
r
kiwi
tu
rn
ip
blue cheese
s a f f r o n
s u n f l o w e r o i l
cassava
g
re
e
n
b
e
ll p
e
p
p
e
r
m
a
n
g
o
c
a
n
e

m
o
l a
s
s
e
s
tom
ato juice
l i c o r i c e
s m
o k e
b
e
ll
p
e
p
p
e
r
gin
b
u
tte
r
p
a
rm
e
s
a
n
c
h
e
e
s
e
s e s a m
e s e e d
t a
r
r
a
g
o
n
salm
on
m a c a r o n i
s
a
g
e
c
u
rra
n
t
t
h
y
m
e
c a c a o
c
a
r
r
o
t
c u r e d p o r k
fatty fish
a n i s e s e e d
p
o
rc
in
i
p o r k s a u s a g e
h a z e l n u t
haddock
s t a r a n i s e
scallop
d
a
te
grape brandy
rum
b
itte
r o
ra
n
g
e
b
e
rry
z
u
c
c
h
in
i
c
u
c
u
m
b
e
r
w h o l e g r a i n w h e a t f l o u r
b r o w n r i c e
m
a
r j o
r a
m
le
ttu
c
e
lim
a
b
e
a
n
g r i l l e d b e e f
bartlett pear
w
o
o
d
s
p
i r i t
litchi
h
o
r s
e
r a
d
i s
h
p e a n u t o i l
g
i n
g
e
r
b
e
a
n
s
e
e
d
huckleberry
c
u
m
i n
l e n t i l
b
a
n
a
n
a
passion fruit
c
i l a
n
t r
o
b a r l e y
cod
salmon roe
c
o
c
o
a
n
u
t
sour milk
butterfat
c a p s i c u m a n n u u m
c a s h e w
p
e
a
r
h a m
catfish
yogurt
r o
o
t
b e e f l i v e r
s
h
iita
k
e
l e a f
b
a
s
i l
p
a
r
s
l e
y
t a b a s c o p e p p e r
mussel
black currant
camembert cheese
roasted onion
b
la
c
k
b
e
a
n
cheddar cheese
o
r
a
n
g
e
swiss cheese
fru
it
r y e b r e a d
black raspberry
b a l m
p
u
m
p
k
in
concord grape
c
o
f f e
e
p
e
c
a
n
juniper berry
lavender
c
r
e
a
m
f r i e d c h i c k e n
p a l m
b o i l e d p o r k
r o a s t e d a l m o n d
tulip
a
s
p
a
ra
g
u
s
v e a l
c a r d a m
o m
wild strawberry
t u r m
e r i c
m
u
s
h
r
o
o
m
c h i c k e n b r o t h
c a
y e
n
n
e
smoked fish
sake
d
i l l
prawn
rom
ano cheese
m a n d a r i n p e e l
orange flower
f r a n k f u r t e r
c o r n
brussels sprout
s
c
a
llio
n
g a l a n g a
cabernet sauvignon wine
c
re
a
m
c
h
e
e
s
e
c a r a w a y
o
n
io
n
c
e
l e
r y
o
i l
endive
r i c e b r a n
port wine
shrimp
w
a
l n
u
t
p
e
a
c
h
l a m b
b e e f b r o t h
m
a
p
l e
s
y
r
u
p
s o y b e a n o i l
crab
t a m
a r i n d
c
itru
s
prickly pear
r o a s t e d p e c a n
p
o
ta
to
p
in
e
a
p
p
le
w
a
te
rm
e
lo
n
y
a
m
muscat grape
l e
m
o
n
p
e
e
l
p o p c o r n
j a s m i n e t e a
o
kra
p e a n u t
f e
n
n
e
l
g
a
r
lic
to
m
a
to
whiskey
ta
n
g
e
rin
e
roquefort cheese
honey
grape juice
cider
b r e a d
m
ilk
cognac
y
e
a
s
t
c o r n g r i t
flower
c l o v e
wine
papaya
c i n
n
a
m
o
n
r
a
s
p
b
e
r
r
y
a
v
o
c
a
d
o
katsuobushi
ku
m
q
u
a
t
ouzo
s p e a r m i n t
p e p p e r m i n t o i l
s
tr
a
w
b
e
r
r
y
p
lu
m
fe
ta
ch
e
e
se
rose
m
e
lo
n
l a u r e l
b
a
y
ca
u
liflo
w
e
r
c
a
b
b
a
g
e
v
e
g
e
ta
b
le
o
r
a
n
g
e
ju
ic
e
p i s t a c h i o
kohlrabi
l e
m
o
n
f e
n
u
g
r e
e
k
bergamot
c
r
a
n
b
e
r
r
y
c
o
c
o
n
u
t
sauerkraut
rhubarb
b
u
tte
rm
ilk
l o v a g e
r o a s t e d b e e f
r o a s t e d n u t
brandy
c h i c k e n
cherry brandy
shellfish
c
e
le
ry
s
q
u
a
s
h
apple brandy
armagnac
c h e r v i l
blackberry brandy
n u t m
e g
p
e
a
m
i n
t
l e
m
o
n

j u
i c
e
l i m
e
o a t
v
e
g
e
t
a
b
l e

o
i l
m a l t
soybean
chickpea
o
r a
n
g
e
p
e
e
l
lard
g r e e n t e a
s
o
y
s
a
u
c
e
m
u
s
t
a
r
d
q
u
i n
c e
sherry wine
baked potato
b e e f
p o r k l i v e r
e
g
g
t e
a
e g g n o o d l e
b
e
e
t
beer
geranium
oyster
m u t t o n
r o
s
e
m
a
r y
c h i c k e n l i v e r
tuna
gardenia
r y e f l o u r
l i n g o n b e r r y
ch
ico
ry
potato chip
c
h
iv
e
b
l a
c
k
b
e
r r y
r a p e s e e d
c o r i a n d e r
m
ilk
fa
t
m a c e
a
p
p
l e
p o r k
kale
b l a c k t e a
l i v e r
smoked salmon
b e e c h
s a v o r y
g r a p e f r u i t
v a n i l l a
c o c o n u t o i l
f i g
s m o k e d s a u s a g e
champagne wine
r o a s t e d h a z e l n u t
chamomile
o
l
i
v
e

o
i
l
l i m
e
j u
i c
e
r o
a
s t e
d
p
e
a
n
u
t
p e p p e r
n
e
c
t a
r i n
e
c
h
e
e
s
e
b u c k w h e a t
h o l y b a s i l
g
r
a
p
e
m e a t
seaweed
w h i t e b r e a d
kidney bean
broccoli
bone oil
s
e
s
a
m
e
o
i l
w h e a t b r e a dw
h
e
a
t
a n g e l i c a
r i c e
p e p p e r m
i n t
lobster
cottage cheese
octopus
radish
w
atercress
squid
c e r e a l
r o a s t e d s e s a m e s e e d
a n i s e
c
h
e
r
r
y
artichoke
o a t m e a l
white wine
mung bean
g u a v a
v
i n
e
g
a
r
violet
red wine
sturgeon caviar
clam
sherry
red kidney bean
t u r k e y
crayfish
c h a y o t e
Figure S4: The backbone of the ingredient network extracted according to [10] with a signicance threshold p = 0.04.
Color indicates food category, font size reects ingredient prevalence in the dataset, and link thickness represents the
number of shared compounds between two ingredients.
the same calculations for the 3rd edition of Fenarolis handbook as well. The 5th edition contains
approximately six times more information on the chemical contents of ingredients (Table S1). Yet,
our main result is robust (Fig. S3), further supporting that data incompleteness is not the main
factor behind our ndings.
S1.1.3 Extracting the backbone
The networks 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 signicant 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 categories and the size of the
name indicates the prevalence. This representation clearly shows the categories that are closely
connected.
5
0
50
100
150
200
250
300
0 100 200 300
N
u
m
b
e
r

o
f

c
o
m
p
o
u
n
d
s
North American
A
B
p-value = 0.01 (*)
0
50
100
150
200
250
300
0 100 200 300
N
u
m
b
e
r

o
f

c
o
m
p
o
u
n
d
s
Western European
p-value = 0.07
0
50
100
150
200
250
300
0 100 200 300
N
u
m
b
e
r

o
f

c
o
m
p
o
u
n
d
s
Southern European
p-value = 0.23
0
50
100
150
200
250
300
0 100 200 300
N
u
m
b
e
r

o
f

c
o
m
p
o
u
n
d
s
Latin American
p-value = 0.04 (*)
0
50
100
150
200
250
300
0 100 200 300
N
u
m
b
e
r

o
f

c
o
m
p
o
u
n
d
s
East Asian
p-value = 0.12
0
50
100
150
200
250
300
350
400
0 100 200 300
A
v
g
.

c
o
m
p
o
u
n
d

d
e
g
r
e
e
Ingredient rank
p-value = 0.82
0
50
100
150
200
250
300
350
400
0 100 200 300
A
v
g
.

c
o
m
p
o
u
n
d

d
e
g
r
e
e
Ingredient rank
p-value = 0.99
0
50
100
150
200
250
300
350
400
0 100 200 300
A
v
g
.

c
o
m
p
o
u
n
d

d
e
g
r
e
e
Ingredient rank
p-value = 0.53
0
50
100
150
200
250
300
350
400
0 100 200 300
A
v
g
.

c
o
m
p
o
u
n
d

d
e
g
r
e
e
Ingredient rank
p-value = 0.50
0
50
100
150
200
250
300
350
400
0 100 200 300
A
v
g
.

c
o
m
p
o
u
n
d

d
e
g
r
e
e
Ingredient rank
p-value = 0.72
Figure S5: Are popular, much-used ingredients more studied than less frequent foods, leading to potential systematic
bias? (A) We plot the number of avor compounds for each ingredient as a function of the (ranked) popularity of the
ingredient. The correlation is very small compared to the large uctuations 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 coefcients 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 avor 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.
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 rst made two lists of ingredients:
one is the list of ingredients appearing in North American cuisine, sorted by the relative prevalence
p
c
i
(i.e. the ingredients more specic to North American cuisine comes rst). The other is a similar
list for East Asian cuisine. Then we measured the number of avor compounds for ingredients in
each list. The result in Fig. S5A shows that any potential bias, if present, is not signicant.
There is another possibility, however, if there is bias such that the dataset tends to list more
familiar (Western) ingredients for more common avor compounds, then we should observe a cor-
relation between the familiarity (frequently used in Western cuisine) and the degree of compound
(number of ingredients it appears in) in the ingredient. Figure S5B shows no observable correla-
tion, however.
S1.2 Recipes
The number of potential ingredient combinations is enormous. For instance, one could generate
10
15
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 10
23
in-
gredient combinations. This number greatly increases if we consider the various cooking methods.
6
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 avor 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
Regardless, the fact that this number exceeds by many orders of magnitude the 10
6
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 ingredients 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 appear 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 ower, 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
7
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 different datasets broadly
agree on what constitutes a cuisine, at least at a gross level.
-1
0
1
2
East
Asian
Southern
European
Latin
American
Western
European
North
American
N
s
r
e
a
l

-

N
s
r
a
n
d
Epicurious
Allrecipes
Menupan (Korean)
Figure S6: Comparison between different datasets. The results on different datasets qualitatively agree with each other
(except Latin American cuisine). Note that menupan.com is a Korean website.
8
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.
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 recipe; if a
book focuses on simple home cooking recipes, then the book will contain fewer ingredients 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 European
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.
0
0.03
0.06
0.09
0.12
0.15
0 10 20 30
P
(
s
)
Number of ingredients per recipe (s)
A B
North American
Western European
Southern European
Latin American
East Asian
0
0.03
0.06
0.09
0.12
0.15
0.18
0 10 20 30 40
P
(
s
)
Number of ingredients per recipe (s)
A B
African
South Asian
Southeast Asian
Middle Eastern
Eastern European
Northern European
Figure S7: Number of ingredients per recipe. North American and Western European cuisine shows similar distribu-
tion while other cuisines have slightly more ingredients per recipe.
9
1
10
100
1000
10000
100000
1 10 100
N
u
m
b
e
r

o
f

r
e
c
i
p
e
s

w
i
t
h

D

d
u
p
l
i
c
a
t
e
s
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.
Yet, there seems to be an interesting trend in our dataset that hotter countries use more ingre-
dients 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.
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 benet of using online
databses is that there is no real-estate issue in contrast to physical cookbooks that should carefully
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 duplicated
recipes (possessing identical sets of ingredients), indicating that popularity is naturally encoded in
these datasets.
S1.3 Number of shared compounds
Figure S9 explains how to measure the number of shared compounds in a hypothetical recipe with
three ingredients.
10
S1.4 Shared compounds hypothesis
S1.4.1 Null models
In order to test the robustness of our ndings, we constructed several random recipe datasets using
a series of appropriate null models and compare the mean number of shared compounds N
s
be-
tween the real and the randomized recipe sets. The results of these null models are summarized in
Fig. S10, each conrming the trends discussed in the paper. The null models we used are:
(A, B) Frequency-conserving. Cuisine c uses a set of n
c
ingredients, each with frequency f
i
. For
a given recipe with N
i
ingredients in this cuisine, we pick N
i
ingredients randomly from the
set of all n
c
ingredients, according to f
i
. That is the more frequently an ingredient is used,
the more likely the ingredient is to be picked. It preserves the prevalence of 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 random in-
gredients 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 cat-
egory 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 avor compounds than expected and East Asian recipes always share less avor
compounds than expected. This, together with the existence of both positive and negative N
real
s

Shared Compounds
a
b
a
c
d
d
c
f
g
2
1
0
2+1+0
3
=1
Binary
Figure S9: For a recipe with three ingredients, we count the number of shared compounds in every possible pair of
ingredients, and divide it by the number of possible pair of ingredients.
11
0
1
2
North
American
Western
European
Latin
American
Southern
European
East
Asian
N
s
r
e
a
l

-

N
s
r
a
n
d
A B
C D
E F
G H
Frequency
conserving
-4
-2
0
2
4
6
8
10
North
American
Western
European
Latin
American
Southern
European
East
Asian
Z
Frequency
conserving
0
1
North
American
Western
European
Latin
American
East
Asian
Southern
European
N
s
r
e
a
l

-

N
s
r
a
n
d
Frequency and ingredient
category preserving
-3
-2
-1
0
1
2
3
4
5
6
7
8
9
10
North
American
Western
European
Latin
American
East
Asian
Southern
European
Z
Frequency and ingredient
category preserving
-3
-2
-1
0
1
2
3
4
5
North
American
Southern
European
Western
European
Latin
American
East
Asian
N
s
r
e
a
l

-

N
s
r
a
n
d
Uniform random
-20
-10
0
10
20
30
40
50
North
American
Southern
European
Western
European
Latin
American
East
Asian
Z
Uniform random
-2
-1
0
1
2
North
American
Southern
European
Latin
American
Western
European
East
Asian
N
s
r
e
a
l

-

N
s
r
a
n
d
category preserving
and uniform random
-20
-10
0
10
20
30
40
50
North
American
Southern
European
Latin
American
Western
European
East
Asian
Z
category preserving
and uniform random
Figure S10: Four different null models. Although the size of the discrepancy between cuisines varies greatly, the
overall trend is stable.
12
0.1
0.3
0.5
0.7
0.9
0 50 100 150
P
(
n
)
Number of shared compounds
Latin
American
0.1
0.3
0.5
0.7
0.9
P
(
n
)
Western
European
0.1
0.3
0.5
0.7
0.9
P
(
n
)
North
American
0 50 100 150
Number of shared compounds
East
Asian
Southern
European
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.
N
rand
s
in every null model, indicates that the patterns we nd are not due to a poorly selected null
models.
Finally, Fig. S11 shows the probability that a given pair with certain number of shared com-
pounds 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 N
s
where there are more than
ve ingredient pairs.
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 N
s
. Since N
s
(R)
for a recipe R is dened as
N
s
(R) =
2
n
R
(n
R
1)

i, jR,i=j

C
i
C
j

(S2)
(where n
R
is the number of ingredients in the recipe R), the contribution from an ingredient pair
(i, j) can be calculated as following:

c
i j
=

1
N
c

Ri, j
2
n
R
(n
R
1)

C
i
C
j

f
i
f
j
N
2
c
2
n
R
(n
R
1)

C
i
C
j

, (S3)
13
where f
i
indicates the ingredient is number of occurrences. Similarly, the individual contribution
can be calculated:

c
i
=

1
N
c

Ri
2
n
R
(n
R
1)

j=i( j,iR)

C
i
C
j

2f
i
N
c
n
R

jc
f
j

C
i
C
j

jc
f
j

. (S4)
We list in Table. S5 the top contributors in North American and East Asian cuisines.
14
North American East Asian
Ingredient i
i
Ingredient i
i
Positive
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
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
Negative
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 sh -0.0458
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.
15
References
[1] Burdock, G. A. Fenarolis handbook of avor ingredients (CRC Press, 2004), 5th edn.
[2] Newman, M. E. J. The structure of scientic 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., Barab asi, 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 Univer-
sity 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., Barth elemy, 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, 8390 (2009).
[10] Serrano, M. A., Bogu n a, M. & Vespignani, A. Extracting the multiscale backbone of com-
plex weighted networks. Proceedings of the National Academy of Sciences 106, 64836488
(2009).
[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, 349 (1998).
[14] Sherman, P. W. & Hash, G. A. Why vegetable recipes are not very spicy. Evolution and
Human Behavior 22, 147163 (2001).
16