The integration of Mendelian genetics into evolutionary
biology in the early twentieth century allowed a better
understanding of a broad range of biological phenom-
ena and unified several previously isolated fields. Despite
the enormous success of the modern synthesis, certain
key issues have remained unanswered. Most notably,
although evolutionary biology successfully interprets
molecular and cellular phenotypes as a result of diverse
evolutionary forces that acted in the past, it rarely builds
an explicit theoretical framework to predict potential
routes of evolution1,2. Why is this issue important? First,
it could help to establish the degree to which evolution
is repeatable. Although long-term microbial evolution-
ary experiments have provided numerous examples of
parallel phenotypic and genetic evolution3, it is unclear
how predictable evolution is at the level of genomes and
molecular networks. Second, such a framework has the
potential to permit informed decisions in medicine4,
biotechnology5 and environmental issues6. For exam-
ple, although in vitro methods have been developed to
forecast the evolution of antibiotic resistance to newly
developed drugs at the protein level7, no such general
tool exists for larger subsystems or whole organisms.
In this Review, we demonstrate that it is possible to
predict, rather than simply interpret, past evolution by
synthesizing evolutionary theory, systems biology and
molecular data. Even under constant selection, predict-
ing evolutionary change is challenging for two main
reasons. First, evolution is a complex mixture of deter-
ministic and chance events: the occurrence, order and
fixation of mutations in populations are all partially sto-
chastic. Second, predicting evolution requires a detailed
knowledge of the range of available mutations and their
fitness effects, an issue that could be best addressed by
combining organism-specific mechanistic models and
large-scale mutational analyses. There are three layers of
prediction that we consider in this Review (TABLE 1): pre-
dicting the distribution of mutational effects and epista-
sis (that is, parameters that influence many central issues
in evolutionary genomics); explaining the driving forces
of sequence and expression evolution on a genomic
scale; and understanding why particular evolutionary
trajectories are realized, whereas others are not.
The recent availability of systematic gene-deletion
studies8,9, genome-scale epistatic interaction maps10 and
detailed mutation analyses of individual proteins11 pro-
vides valuable insights into these problems. However,
most established experimental approaches are limited,
either because they focus on individual genes instead of
large gene networks, or because they study a restricted set
of environments and mutation types (TABLE 1). Systems
biology can help to resolve these issues by allowing
the analysis of large cellular subsystems and provid-
ing molecular explanations with clear links to changes
in environmental conditions. The Review focuses on
genome-scale models of microbial metabolic networks12
owing to their large-scale, predictive power coupled with
mechanistic insights and wide usage.
The Review starts with a brief summary of meta-
bolic network modelling; we emphasize the data used
for model reconstructions, the model-building steps
and the reliability and limitations of these models. We
then discuss how these models can be used to study
the three layers of prediction described above (TABLE 1).
Last, we demonstrate how computational and experi-
mental approaches can be more tightly integrated by
*Synthetic and Systems
Biology Unit, Institute of
Biochemistry, Biological
Research Center, Temesvári
krt. 62, H‑6726 Szeged,
Hungary.
‡Cambridge Systems Biology
Centre and Department
of Genetics, University of
Cambridge, Cambridge
CB2 3EH, UK.
§ Departments of
Bioinformatics (CMBI) and
Systems Biology (CSBB),
Nijmegen Centre for
Molecular Life Sciences,
Radboud University
Nijmegen Medical Centre,
P.O. BOX 9101, 6500 HB
Nijmegen, The Netherlands.
Correspondence to C.P. 
e‑mail: cpal@brc.hu
doi:10.1038/nrg3033
Published online 2 August 2011
Epistatic interaction
Non-independent effect of
mutations on a phenotype.
Epistasis is negative when a
genotype with two mutations
has a lower phenotype value or
positive when it has a higher
value than would be expected
from the product of the single
mutant values.
Systems-biology approaches for
predicting genomic evolution
Balázs Papp*‡, Richard A. Notebaart§ and Csaba Pál*
Abstract | Is evolution predictable at the molecular level? The ambitious goal to answer
this question requires an understanding of the mutational effects that govern the
complex relationship between genotype and phenotype. In practice, it involves
integrating systems-biology modelling, microbial laboratory evolution experiments and
large-scale mutational analyses — a feat that is made possible by the recent availability
of the necessary computational tools and experimental techniques. This Review
investigates recent progresses in mapping evolutionary trajectories and discusses the
degree to which these predictions are realistic.
M O D E L L I N G
REVIEWS
NATURE REVIEWS | GENETICS VOLUME 12 | SEPTEMBER 2011 | 591
© 2011 Macmillan Publishers Limited. All rights reserved

Graph-theoretical
approaches
The study of graphs. A graph
provides an abstract
representation of a biological
or physical system in which
components are represented
by nodes that are connected to
each other by edges (links).
considering new technological advances in the fields
of experimental evolution, genome engineering and
automated model reconstruction.
Genome-scale metabolic models
Properties and advantages. The three layers of predic-
tions listed above require biologically detailed compu-
tational models that estimate the impact of mutations
and environmental changes on fitness. Models that are
most suited for evolutionary studies are based on sound
biochemical principles: they should capture the func-
tional states of the cell and compute phenotypes (for
example, growth rate) that serve as fitness correlates.
These models include detailed kinetic models of specific
metabolic pathways13 or regulatory circuits (for example,
the cell cycle14), logical models of signalling networks15
and constraint-based models of genome-scale metabolic
networks16 (TABLE 2). Organism-specific kinetic models
are generally highly accurate and realistic but require
detailed experimental data, which are rarely available for
large systems. However, constraint-based models allow
integration of high-throughput post-genomic data but
generally offer no information about metabolite concen-
trations or about the temporal dynamics of the system
(but see REFS 17–19).
Genome-scale metabolic models have been useful,
as they rely on high-quality metabolic network recon-
structions12. These reconstructions are primarily based
on a sequenced genome and are generally built manually
using information from metabolic databases — such as
KEGG20 and BRENDA21 — and the primary literature.
Next, the network reconstruction is converted into a
mathematical model that can be analysed using con-
straint-based approaches (BOX 1). By mimicking nutrient
conditions used in prior experimental studies, the model
is validated against high-throughput data, and existing
discrepancies are resolved by new sets of experiments.
Genome-scale metabolic models therefore have at
least two conceptual advantages over other approaches
(TABLE  2). First, whereas most other modelling
approaches focus on small-scale biochemical systems
(that is, individual pathways), constraint-based models
aim to calculate the metabolic behaviour of moderately
large systems (that is, 600–1,300 genes). Thus, these
models allow comparisons to be made with the results
of high-throughput genomic data. Second, in contrast
to most statistical or graph-theoretical approaches22, these
models are far more detailed and realistic, as they infer
the functional states of the network as a function of
nutrient availability in the environment.
Genome-scale metabolic models have already
proven to be successful in several applications: distin-
guishing between essential and non-essential genes
across environmental conditions23; identifying epistatic
interactions24; predicting growth properties25; guiding
metabolic engineering16; and charting the functional
dependence (coupling) between genes26. However, sev-
eral important problems remain to be addressed (BOX 1),
not least because these models generally do not incor-
porate enzyme kinetic information and cannot cap-
ture the nonlinear relationship between enzyme level
and metabolic flux (but see REFS 18,19,27,28). These
are the major reasons why predicting minor mutational
effects on enzyme activity or predicting metabolite
concentrations remains challenging.
Model applications. Despite these current limitations,
evolutionary biologists have recognized the potential of
these models and have used them to reach three goals.
First, they have been used to estimate the overall patterns
of epistasis29–31 and distribution of mutational effects32,33,
and second, the models can infer interspecies differences
in metabolic gene content and hence can explain general
trends of genome evolution34,35. But perhaps the most
inspiring aspect of this framework is its capacity to make
specific and reliable predictions on the outcome of meta-
bolic evolution, both in short-term laboratory evolution
and on macroevolutionary time scales.
Distribution of mutational effects and epistasis
Recent large-scale gene deletion analyses demon-
strated that mutations with weak phenotypic effects are
Table 1 | Three major issues in evolutionary systems biology
Layer of
prediction
Importance Current state of
knowledge
Difficulty New tools and knowledge needed
for integration
Distribution
of mutational
effects and
epistatic
interactions
General architecture of
adaptation. Robustness
against mutations
Wealth of systematic
gene deletion studies and
epistasis maps
Existing fitness landscape models
are not biologically detailed.
High-throughput experiments are
restricted to a few environmental
conditions, or they only consider null
mutations
Realistic systems-biology models offer
new predictions on mutational effects
and mechanistic insights. New types
of experimental data (for example,
fitness profiling of point mutations or
gene overexpression studies)
General
patterns
of genome
evolution
Evolutionary forces
driving protein and
expression divergence,
gene loss, horizontal
gene transfer and gene
duplicability
Impact of post-genomic
features (for example, gene
expression or network
position)
No clear relationship between
fitness and the post-genomic gene
features studied
Predictions and measuring most
relevant physiological data (for
example, range of neutrality, optimal
gene activity or physiological
coupling between genes)
Specific
evolutionary
trajectories
Relative importance of
chance and necessity
in evolution. Predictive
tools for applications
Map of adaptive landscape
for single proteins. Insight
from experimental
evolutionary studies
Difficult to map adaptive landscapes
for large cellular subsystems
empirically. Interpretations
dominate over predictions
Modelling the outcome of adaptive
evolution at the molecular level. New
experimental technologies to map
adaptive landscapes
R E V I E W S
592 | SEPTEMBER 2011 | VOLUME 12 www.nature.com/reviews/genetics
© 2011 Macmillan Publishers Limited. All rights reserved