Content uploaded by Uri Alon
All content in this area was uploaded by Uri Alon
Content may be subject to copyright.
Biological Networks: The Tinkerer as an Engineer
This viewpoint comments on recent advances in understanding the design principles
of biological networks. It highlights the surprising discovery of “good-engineering”
principles in biochemical circuitry that evolved by random tinkering.
Franc¸ois Jacob pictured evolution as a tink-
erer, not an engineer (1). Engineers and tink-
erers arrive at their solutions by very different
routes. Rather than planning structures in ad-
vance and drawing up blueprints (as an engi-
neer would), evolution as a tinkerer works
with odds and ends, assembling interactions
until they are good enough to work. It is
therefore wondrous that the solutions found
by evolution have much in common with
good engineering design (2). This Viewpoint
comments on recent advances in understand-
ing biological networks using concepts from
Biological networks are abstract represen-
tations of biological systems, which capture
many of their essential characteristics. In the
network, molecules are represented by nodes,
and their interactions are represented by edg-
es (or arrows). The cell can be viewed as an
overlay of at least three types of networks,
which describes protein-protein, protein-
DNA, and protein-metabolite interactions. In-
herent in this description is suppression of
detail: many different mechanisms of tran-
scription regulation, for example, may be de-
scribed by a single type of arrow. Further-
more, the interactions can be of different
strengths, so there should be numbers or
weights on each arrow (3). Whenever two or
more arrows converge on a node, an input
function needs to be specified (for example,
AND or OR gates) (4, 5). At present, many of
the connections, numbers and input functions
are not known. However, something can still
be learned even from the very incomplete
networks currently available (6–8). First, the
network description allows application of
tools and concepts (9) developed in fields
such as graph theory, physics, and sociology
that have dealt with network problems before
(see D. Bray on pg. 1864 in this issue).
Second, biological systems viewed as net-
works can readily be compared with engi-
neering systems, which are traditionally de-
scribed by networks such as flow charts and
blueprints. Remarkably, when such a com-
parison is made, biological networks are seen
to share structural principles with engineered
networks. Here are three of the most impor-
tant shared principles, modularity, robustness
to component tolerances, and use of recurring
The first principle, modularity (10–12), is
an oft-mentioned property of biological net-
works. For example, proteins are known to
work in slightly overlapping, coregulated
groups such as pathways and complexes. En-
gineered systems also use modules, such as
subroutines in software (13) and replaceable
parts in machines. The following working
definition of a module is proposed based on
comparison with engineering: A module in a
network is a set of nodes that have strong
interactions and a common function. A mod-
ule has defined input nodes and output nodes
that control the interactions with the rest of
the network. A module also has internal
nodes that do not significantly interact with
nodes outside the module. Modules in engi-
neering, and presumably also in biology,
have special features that make them easily
embedded in almost any system. For exam-
ple, output nodes should have “low imped-
ance,” so that adding on additional down-
stream clients should not drain the output to
existing clients (up to some limit).
Why does modularity exist in biological
networks? It is important to realize that not
all networks that evolve by tinkering are
modular. A well-studied example is computer-
science neural networks (NNs). NNs are a set
of interconnected nodes, each of which has a
state that depends on the integrated inputs
from other nodes (14 ). As do protein signal-
ing networks, NNs function to process infor-
mation between input and output nodes (15).
In a way analogous to biological networks,
NNs are optimized by an “evolutionary” tink-
ering process of adding and removing arrows
and changing their weights until the NN per-
forms a given computational goal (gives the
“correct” output responses to input signals).
Unlike biological networks, however, NNs
are nonmodular. They typically have a highly
interconnected architecture in which each
node participates in many tasks. Viewed in
this perspective, the modularity of biological
networks is puzzling because modular struc-
tures can be argued to be less optimal than
NN-style, nonmodular structures. After all,
modules greatly limit the number of possible
connections in the network, and usually a
connection can be added that reduces modu-
larity and increases the fitness of the network.
This is the reason that NNs almost always
display a nonmodular design. A clue to the
reason that modules evolve in biology can be
found in engineering (16 ). Modules in engi-
neering convey an advantage in situations
where the design specifications change from
time to time. New devices or software can be
easily constructed from existing, well-tested
modules. A nonmodular device, in which ev-
ery component is optimally linked to every
other component, is effectively frozen and
cannot evolve to meet new optimization con-
ditions. Similarly, modular biological net-
works may have an advantage over non-
modular networks in real-life ecologies,
which change over time: Modular networks
can be readily reconfigured to adapt to new
conditions (16, 17 ).
The second common feature of engineer-
ing and biological networks is robustness to
component tolerances. In both engineering
and biology, the design must work under all
plausible insults and interferences that come
with the inherent properties of the compo-
nents and the environment. Thus, Escherichia
coli needs to be robust with respect to tem-
perature changes over a few tens of degrees,
and no circuit in the cell should depend on
having precisely 100 copies of protein X and
not 103. This point has been made decades
ago for developmental systems (17, 18) and
metabolism (2, 19, 20). The fact that a gene
circuit must be robust to such perturbations
imposes severe constraints on its design:
Only a small percentage of the possible cir-
cuits that perform a given function can per-
form it robustly. Recently, there have been
detailed experimental-theoretical studies
that demonstrate how particular gene cir-
cuits can be robust, for example, in bacte-
rial chemotaxis (21, 22) and in fruit-fly
The third feature common to engineering
and biological networks is the use of recurring
circuit elements. An electronic device, for ex-
ample, can include thousands of occurrences of
circuit elements such as operational amplifiers
and memory registers. Biology displays the
same principle, using key wiring patterns again
and again throughout a network. Metabolic net-
works use regulatory circuits such as feedback
inhibition in many different pathways (24). The
transcriptional network of E. coli has been
shown to display a small set of recurring circuit
elements termed “network motifs” (25). Each
network motif can perform a specific informa-
Department of Molecular Cell Biology and Depart-
ment of Physics of Complex Systems, Weizmann In-
stitute of Science, Rehovot, Israel 76100. E-mail:
N ETWORKS IN B IOLOGY
26 SEPTEMBER 2003 VOL 301 SCIENCE www.sciencemag.org1866
tion processing task such as filtering out spuri-
ous input fluctuation (25), generating temporal
programs of expression (3, 25) or accelerating
the throughput of the network (2, 26). Recently,
the same network motifs were also found in the
transcription network of yeast (7, 27 ). It is im-
portant to stress that the similarity in circuit
structure does not necessarily stem from circuit
duplication. Evolution, by constant tinkering,
appears to converge again and again on these
circuit patterns in different nonhomologous sys-
tems (25, 27, 28), presumably because they
carry out key functions (see Perspective (29)
STKE). Network motifs can be detected by
algorithms that compare the patterns found in
the biological network to those found in suitably
randomized networks (25, 27). This is analo-
gous to detection of sequence motifs as
recurring sequences that are very rare in
Network motifs are likely to be also
found on the level of protein signaling net-
works (30). Once a dictionary of network
motifs and their functions is established,
one could envision researchers detecting
network motifs in new networks just as
protein domains are currently detected in
the sequences of new genes. Finding a se-
quence motif (e.g., a kinase domain) in a
new protein sheds light on its biochemical
function; similarly, finding a network motif
in a new network may help explain what
systems-level function the network per-
forms, and how it performs it.
Will a complete description of the biological
networks of an entire cell ever be available?
The task of mapping an unknown network is
known as reverse-engineering (3, 31–33).
Much of engineering is actually reverse-
engineering, because prototypes often do not
work and need to be understood in order to
correct their design. The program of molecular
biology is reverse-engineering on a grand scale.
Reverse engineering a nonmodular network of
a few thousand components and their nonlinear
interactions is impossible (exponentially hard
with the number of nodes). However, the spe-
cial features of biological networks discussed
here give hope that biological networks are
structures that human beings can understand.
Modularity, for example, is at the root of the
success of gene functional assignment by ex-
pression correlations (11, 34 ). Robustness to
component tolerances limits the range of pos-
sible circuits that function on paper to only a
few designs that can work in the cell. This can
help theorists to home in on the correct design
with limited data (21–23). Network motifs de-
fine the few basic patterns that recur in a net-
work and, in principle, can provide specific
experimental guidelines to determine whether
they exist in a given system (25). These con-
cepts, together with the current technological
revolution in biology, may eventually allow
characterization and understanding of cell-wide
networks, with great benefit to medicine. The
similarity between the creations of tinkerer and
engineer also raises a fundamental scientific
challenge: understanding the laws of nature that
unite evolved and designed systems.
References and Notes
1. F. Jacob, Science 196, 1161 (1977).
2. M. Savageau, Biochemical Systems Analysis: A Study
of Function and Design in Molecular Biology (Addi-
son-Wesley, Reading, MA, 1976).
3. M. Ronen, R. Rosenberg, B. I. Shraiman, U. Alon, Proc.
Natl. Acad. Sci. U.S.A. 99, 10555 (2002).
4. C. H. Yuh, H. Bolouri, E. H. Davidson, Science 279,
5. Y. Setty, A. E. Mayo, M. G. Surette, U. Alon, Proc. Natl.
Acad. Sci. U.S.A. 100, 7702 (2003).
6. D. Thieffry, A. M. Huerta, E. Perez-Rueda, J. Collado-
Vides, Bioessays 20, 433 (1998).
7. T. I. Lee et al., Science 298, 799 (2002).
8. E. H. Davidson et al., Science 295, 1669 (2002).
9. S. H. Strogatz, Nature 410, 268 (2001).
10. L. H. Hartwell, J. J. Hopﬁeld, S. Leibler, A. W. Murray,
Nature 402, C47 (1999).
11. J. Ihmels et al., Nature Genet. 31, 370 (Aug, 2002).
12. E. Ravasz, A. L. Somera, D. A. Mongru, Z. N. Oltvai,
A.-L. Baraba´si, Science 297, 1551 (2002).
13. C. R. Myers, arXiv: cond-mat/0305575 (2003).
14. J. J. Hopﬁeld, Proc. Natl. Acad. Sci. U.S.A. 79, 2554
15. D. Bray, J. Theor. Biol. 143, 215 (1990).
16. H. Lipson, J. B. Pollack, N. P. Suh, Evolution 56, 1549
17. J. Gerhart, M. W. Kirschner, Cells, Embryos, and Evo-
lution: Toward a Cellular and Developmental Under-
standing of Phenotypic Variation and Evolutionary
Adaptability (Blackwell Science Inc, Oxford, 1997).
18. C. H. Waddington, Nature 150, 563 (1942).
19. H. Kacser, J. A. Burns, Symp. Soc. Exp. Biol. 32,65
20. M. Savageau, Nature 229, 542 (1971).
21. N. Barkai, S. Leibler, Nature 387, 913 (1997).
22. U. Alon, M. G. Surette, N. Barkai, S. Leibler, Nature
397, 168 (1999).
23. A. Eldar et al., Nature 419, 304 (2002).
24. D. Fell, Understanding the Control of Metabolism
(Portland Press, London, 1997).
25. S. S. Shen-Orr, R. Milo, S. Mangan, U. Alon, Nature
Genet. 31, 64 (2002).
26. N. Rosenfeld, M. B. Elowitz, U. Alon, J. Mol. Biol. 323,
27. R. Milo et al., Science 298, 824 (2002).
28. G. C. Conant, A. Wagner, Nature Genet. 34, 264 (2003).
29. A. Wagner, Sci. STKE 2003, pe41 (2003).
30. C. V. Rao, A. P. Arkin, Annu. Rev. Biomed. Eng. 3, 391
31. M. E. Csete, J. C. Doyle, Science 295, 1664 (2002).
32. A. Arkin, P. Shen, J. Ross, Science 277, 1275 (1997).
33. T. S. Gardner, D. di Bernardo, D. Lorenz, J. J. Collins,
Science 301, 102 (2003).
34. M. B. Eisen, P. T. Spellman, P. O. Brown, D. Botstein,
Proc. Natl. Acad. Sci. U.S.A. 95, 14863 (1998).
Social Insect Networks
Jennifer H. Fewell
Social insect colonies have many of the properties of adaptive networks. The simple
rules governing how local interactions among individuals translate into group be-
haviors are found across social groups, giving social insects the potential to have a
profound impact on our understanding of the interplay between network dynamics
and social evolution.
The formal exploration of social insect col-
onies as networks is in its infancy. Howev-
er, social insects such as wasps, ants, and
honeybees provide a powerful system for
examining how network dynamics contrib-
ute to the evolution of complex biological
systems. Social insect colonies (and social
groups generally) have key network attributes
that appear consistently in complex biological
systems, from molecules through ecosystems;
these include nonrandom systems of connectiv-
ity and the self-organization of group-level
phenotypes (1–3). Colonies exhibit multi-
ple levels of organization, yet it is still
possible to track individuals, making these
societies more accessible to experimen-
tal manipulation than many other com-
How can viewing insect societies as net-
works shape our understanding of social orga-
nization and evolution? First, they have become
one of the central model systems for exploring
self-organization: the process by which interac-
tions occurring locally between individuals
produce group-level attributes. Self-organi-
zation in a social insect colony produces
emergent properties: social phenotypes that
are greater than a simple summation of
individual worker behaviors (2). The basic
rules generating these dynamics are broad-
ly applicable across taxa whose members
show social behavior, and they produce
ubiquitous patterns of social organization,
including mass action responses, division
of labor, and social hierarchies (2, 4 ).
School of Life Sciences, Arizona State University,
Tempe, AZ 85287–1501, USA. E-mail: j.fewell@asu.
N ETWORKS IN B IOLOGY
www.sciencemag.org SCIENCE VOL 301 26 SEPTEMBER 2003 1867