TREE 2453 No.

of Pages 14

Review

Evolution in the light of fitness landscape

theory
Inês Fragata,1 Alexandre Blanckaert,1 Marco António Dias Louro,1
David A. Liberles,2 and Claudia Bank1,*

By formalizing the relationship between genotype or phenotype and fitness,

fitness landscapes harbor information on molecular and evolutionary con-
straints. The shape of the fitness landscape determines the potential for adap-
tation and speciation, as well as our ability to predict evolution. Consequently,
fitness landscape theory has been invoked across the natural sciences and
across multiple levels of biological organization. We review here the existing
literature on fitness landscape theory by describing the main types of fitness
landscape models, and highlight how these are increasingly integrated into an
applicable statistical framework for the study of evolution. Specifically, we
demonstrate how the interpretation of experimental studies with respect to
fitness landscape models enables a direct link between evolution, molecular
biology, and systems biology.

The fitness landscape – a useful concept for the study of evolution

A longstanding goal of evolutionary biology is to understand the relationship between genotype
(see Glossary), phenotype, and fitness, and its consequences for adaptation and speciation.
The theory of fitness landscapes encompasses a multitude of models that map genotype or
phenotype to fitness [1–5]. Fitness landscape models greatly simplify the study of adaptation by
reducing individuals to either genotypes or phenotypes, whose reproductive success is
determined by a single trait, namely fitness. This simplification makes evolution tractable
and potentially predictable [1,4–6]. However, reduction of the interaction of evolutionary forces
and the complexity of nature to differences in fitness in a static environment can also lead to an
incorrect assessment of possible evolutionary paths and constraints [4,7] (Box 1). Neverthe-
less, owing to its appealing simplicity, which lends itself to mathematical analysis, the concept
of fitness landscapes has been widely used. Today we can draw from a large body of fitness
landscape theory that was developed to investigate questions in speciation, adaptation, and
molecular and systems biology (Fig. 1) (e.g., [2,3,5,8–10]).

For almost a century fitness landscapes have mostly been studied theoretically. Recent
advances in next-generation sequencing now allow assessment of increasingly large experi-
mental fitness landscapes with unprecedented accuracy (e.g., [11–15]; reviewed in [5]). This
has spurred the development of methods to bridge the gap between theoretical and empirical 1
Instituto Gulbenkian de Ciência,
fitness landscapes, and created the opportunity to evaluate the classical theory (e.g., [16–19]). 2780-156 Oeiras, Portugal
2
Department of Biology and Center for
Moreover, the flexibility of the fitness landscape concept allows the relationship between Computational Genetics and
genotypes or phenotypes and fitness to be studied across biological levels (Fig. 2) [3,4,20]. Genomics, Temple University,
This flexibility is reflected in the variety of questions to which fitness landscape theory has been Philadelphia, PA 19122-1801, USA

applied, ranging from speciation to protein and gene network evolution (Fig. 1). Although the
same overarching framework underlies this work, and similar models are often derived and *Corresponding author.
analyzed, there is little exchange between these fields. Complementing other reviews, which evoldynamics@gmail.com (C. Bank).

Trends in Ecology & Evolution, Month Year, Vol. xx, No. yy https://doi.org/10.1016/j.tree.2018.10.009 1
© 2018 Elsevier Ltd. All rights reserved.
 TREE 2453 No. of Pages 14

have provided important in-depth overviews of the recent advances in either of these fields Glossary
[5,20–22], we here highlight the connections and overlaps between the fields. Specifically, we Biological network: The set of
emphasize several recent theoretical and empirical examples of interdisciplinary fitness land- elements (e.g., genes) necessary for
scape studies. These examples illustrate how fitness landscape models can be used to the realization of a biological or
ecological function, and the various
estimate the prevalence of epistasis, study the importance of neutral evolution, and address relationships (e.g., activation,
how integration of the genotype–phenotype map into the theory of fitness landscapes might inhibition) that exist between the
allow for the prediction of evolution. elements of the set.
Epistasis: Genetic background-
specific (fitness) effect of a mutation.
Epistasis in the light of fitness landscapes Effective population size: The
The term epistasis, first introduced by Bateson [23,24], describes the fitness effects of mutations hypothetical population size of a
Wright–Fisher population (panmictic
that are specific to the genetic background the mutation occurs on. From a fitness landscape point
and of constant size) that best
of view, epistasis represents a nonlinearity in the function that describes the fitness landscape. In reproduces the observed population
the following we discuss how, through the study of epistasis, fitness landscape theory has been genetics statistics.
invoked in the study of speciation, adaptation, and biological networks. Evolvability: The ability of a
biological system (a population,
individual, network, or molecule) to
Manifested as postzygotic genetic incompatibilities between species, epistasis maintains species have or produce variants that can be
barriers via selection against unfit hybrid offspring [25]. Indeed, one of the earliest descriptions of a acted upon by selection.
fitness landscape model can be found in the work of Bateson [23]. It would later become known as Fitness: A measure of the
reproductive or replicative success of
the Bateson–Dobzhansky–Muller (BDM) model [26,27], arguably the most widely accepted model
a biological entity (from molecules to
for the evolution of postzygotic reproductive isolation [28–30]. This model describes how genetic individuals). Usually, fitness-related
changes that accumulate in geographically isolated populations can be incompatible, in other phenotypes (e.g., growth rate) are
words leading to unfit hybrid offspring when brought together in the same genome. The BDM used as proxies for fitness.
Genetic drift: Random change in
model explains the buildup of reproductive isolation via genetic incompatibilities without the need allele frequencies over time in a
to cross a fitness valley, in other words without the need to invoke the fixation or bypassing of population of finite size.
deleterious intermediate steps during divergence [28]. Essentially, the BDM model specifies a Genotype: The genetic constitution
fitness landscape in which there is a ridge of fit genotypes along which the populations evolve, but of an organism.
Genotype–fitness landscape: A
where the accumulated mutations from diverged populations interact negatively epistatically. map from (usually) discrete
Various BDM incompatibilities have been identified empirically (Fig. 1) [31–33]. Theoretical studies genotypes to fitness. Genotypes tend
to represent nucleotide, amino acid,
or gene differences.
Locus: A position in the genome of
Box 1. The problem of dimensionality and the illusion of crossing valleys. an individual. Depending on the
Fitness landscapes are multidimensional. However, they are often visualized in 3D, two dimensions corresponding to focus of the study, a locus can
genotypic or phenotypic axes, and a third to fitness (but see [6]). Such a representation leads to a wrong intuition of how correspond to a single nucleotide or
populations evolve, and this has spurred criticism of the importance of the concept [7]. Specifically, the representation amino acid position, several base
of the landscape in 3D leads to a picture with local peaks separated by valleys of low fitness, implying that switching pairs of DNA sequence, or a gene.
fitness peaks requires crossing a fitness valley. This neglects the potential existence of high-fitness ridges that connect Mutation: A heritable change in the
fitness peaks along other dimensions of the landscape. To an extreme, this is illustrated in the ‘holey landscape model’ genetic sequence of an individual.
[20] (Figure I). Considering only neutral and lethal genotypes, this model captures how populations can move along Neutral: A property of a mutation
fitness ridges and end up at opposite ends of a fitness hole, leading to reproductive isolation without ever ‘crossing’ a which harbors no fitness effect for an
valley. individual, such that its frequency in
the population depends only on
The dimensionality of the fitness landscape is also important when considering experimental fitness landscapes. extraneous factors, such as genetic
Indeed, it is impossible to capture the full fitness landscape of an organism or even of a single molecule. Therefore, we drift. A mutation is conditionally
must focus on a subset of the full fitness landscape. It has been shown [18] that extrapolating any properties from the neutral if its neutrality is genetic-
local fitness landscapes requires a rigorous statistical framework because many different global landscapes can give background or environment-
rise to the same local fitness landscape. To further illustrate this point we show that apparently unconnected networks dependent.
of genes can be connected when considering the full fitness landscape (Table 1). NK model: Mathematical model that
describes a genotype–fitness
landscape with different degrees of
The number of dimensions of a ‘holey fitness landscape’ or a neutral network determines the percolation threshold Pt of
epistasis. This model is defined by
said network [2,20], pt  1/L, a measure of the connectivity of the landscape. A fitness landscape is subcritical
the parameters N, the number of loci
(i.e., composed of many small networks of fit genotypes, isolated from each other), if Pv < Pt, where Pv is the probability
in the landscape, and K, the degree
that a given genotype is viable. In this case, each genotype is trapped within its own vicinity. If Pv < Pt, a single massive
of epistasis between loci.
network exists in which all fit genotypes are connected through single mutation steps.

2 Trends in Ecology & Evolution, Month Year, Vol. xx, No. yy

TREE 2453 No. of Pages 14

Eight loci Eight loci Phenotype: A set or subset of

observable traits of an individual that
stems from the interactions between
genotype and environment.
Phenotype–fitness landscape: A
map from (usually) continuous
phenotypes to fitness. In a
multidimensional phenotype space,
each dimension is composed by a
different one-dimensional trait.
Pleiotropy: The property of a gene
to affect more than one independent
phenotypic trait.
Rough Mount Fuji model (RMF): A
model describing a genotype–fitness
landscape that is composed of an
additive component and an epistasic
component. Ruggedness is then
tuned by changing the relative
Four loci proportions of the two components.

Fitness
0

Genotypic Genotypic
dimension 1 dimension 2

Unfit genotypes Isolated fit genotypes Fit genotypes belonging

to the largest network

Fit genotypes belonging Fit genotypes belonging Fit genotypes belonging

to the second largest network to the third largest network to the fourth largest network

Figure I. Representation of a ‘holey landscape’ as a network (top) and 3D fitness landscape (bottom). We compare two
cases in which the probability, Pv = 0.2, for a genotype to be fit, is below (left panels; L = 4 and therefore Pt  0.25) and
above (right panels: L = 8 and therefore Pt  0.125) the percolation threshold. For the same proportion of fit versus unfit
genotypes, larger landscapes are easier to explore: for L = 4, 66% of the fit genotypes are in the largest network, and for
L = 8, 82% belong to the largest network. Colors indicate the different connected networks, with red indicating isolated
genotypes. In the network view (upper panels), genotypes are connected by single step mutations. The figure illustrates
how the 3D projection creates a wrong impression of fitness valleys by masking the connectivity that exists between fit
genotypes in higher dimensions.

Trends in Ecology & Evolution, Month Year, Vol. xx, No. yy 3

 TREE 2453 No. of Pages 14

Table 1. Properties of subsampled fitness landscapes.a

Pv = 0.5 Pv = 0.2 Pv = 0.1 Pv = 0.05 Pv = 0.01

Probability of having at least two 0.999 0.874 0.494 0.197 0.008

fit genotypes

Average number of networks 1.12 0.778 0.462 0.406 0.125

Probability of having at least two 0.386 0.773 0.763 0.706 0.75

isolated networks

Probability that two isolated 0.925 0.212 0.336 0.588 >5/6

networks in the subsampled
fitness landscape are connected
in the full landscape

a
We simulated a ‘holey fitness landscape’ with eight diallelic loci and subsampled fitness landscape with four loci. Pv
corresponds to the probability that a given genotype is fit, with the fitness of genotypes assigned through permutation.
Here Pt  0.125, therefore the first two columns correspond to supercritical networks, while the last three are
subcritical networks. In the subsampled fitness landscapes, we estimated the number of isolated networks of
genotypes (i.e., groups of fit genotypes that are not connected by fit genotypes). In addition, we quantified the
probability that two isolated networks were connected in the full fitness landscape. Each measure was obtained based
on 1000 simulations.

of the BDM model and its extensions (e.g., [34,35]) traditionally focus on small fitness landscapes.
This is because one needs to define 2L 1 parameters (one fitness for each genotype minus one
reference) to fully specify the fitness landscape for L diallelic loci. A minimal genome-wide model of
epistasis, the ‘holey landscape model’, was proposed by Gavrilets [20]: genotypes are either fit or
lethal, with this fitness being assigned randomly, as illustrated in Box 1. The resulting fitness
landscape is highly epistatic because the complete genotype information is necessary to know the
fitness of a given genotype, and has been used to study speciation through genetic drift. In
addition, this model highlights the importance of the dimensionality of fitness landscapes: higher-
dimensional fitness landscapes tend to be more connected (i.e., lower probability of having truly
isolated fitness peaks or ridges) than their lower-dimension counterparts (Box 1). In addition to its
application to questions in speciation, the holey landscape model was also successfully applied to
the study of RNA folding and structure [36], which illustrates the versatility of the model (also
discussed in the next section).

During adaptation to a new environment, epistasis dictates which sequences of consecutive

mutations are feasible [11]. For example, a recent study of the influenza nucleoprotein [37]
confirmed the theory that adaptive functional changes to a protein frequently destabilize the
protein structure, such that compensatory changes are necessary to restabilize the structure
around the adaptive functional change [9,38]. Gong et al. [37] reconstructed an adaptive walk,
and found that each activity-enhancing (i.e., functionally adaptive) mutation was followed by a
mutation that increased the structural stability of the protein. When the authors introduced the
activity-enhancing mutations without their stability-enhancing partners, growth of the virus was
strongly impaired. In such cases the amount of epistasis in the fitness landscape limits the
possible routes to high fitness (Box 2). In a growing body of theoretical work, several statistics
have been developed to quantify the amount and type of epistasis in a given fitness landscape.
How these relate to each other was analyzed using tunable theoretical fitness landscapes, most
notably the NK model and the rough Mount Fuji model (RMF) [16,19].

Epistasis is also prevalent in phenotype–fitness landscape models, such as Fisher's geometric

model (FGM) [8,39]. In FGM, populations evolve through mutations that are represented by
jumps in an n-dimensional phenotype space. Mutations in FGM are usually assumed to be
additive at the phenotype level. Nevertheless, mutational effects become epistatic at the fitness

4 Trends in Ecology & Evolution, Month Year, Vol. xx, No. yy

 TREE 2453 No. of Pages 14

Theore cal Experimental

examples examples
Can the Mul linear
model [84, 85, S1, S2] NA
Does it popula on Yes
consider Yes be poly- [8, 10, 18, 39,
No [18, 40, 83, S9-
morphic? FGM 41, 42, 79-81,
the phe- S11]
S3-S8]
notype?
No Holey [2, 36, 56, S12- [2, 36, S12,
Yes landscape S15] S13]

Biological [48, 56, 74,

76-78, 111, NA
No networks
Single S16,S17]
Yes molecule? Yes Biophysical [57, 62, S18-
Does it [47, 70]
model S21]
consider Does it
Genotype the whole [3, 17, 53, 54,
include the NK model [14,16]
genome? 88, S22-S24]
phenotype
? Are No Block
[S25-S28] [S29]
muta onal Presence Yes model
Yes effects of non-
No Yes
addi ve? interac ve No Rough
[54, 88, 95,
Does it No loci? mount Fuji [14,16]
Yes S24, S30]
focus on model
specific
molecules? Is epistasis House of
[95, S23, S24,
tunable? No cards [14,16]
S31-S33]
model

BDM [23, 26-28, 34, [31-33, 49,

No model S34-S43] S44-S55]

Fig. 1. Selected examples of theoretical and experimental fitness landscapes. The decision graph illustrates the differences between fitness landscape models and
provides examples of how they are used in the literature. Supplemental references (marked with S) are available in the supplemental information online. Abbreviations:
BDM model, Bateson–Dobzhansky–Muller model; FGM, Fisher's geometric model; NA, not applicable; NK model, a model defined by the number of loci in the
landscape, N, and the degree of epistasis between loci, K.

level because the relationship between phenotype and fitness is non-linear; the underlying
fitness landscape is usually assumed to be a single-peaked Gaussian distribution. Martin et al.
[40] showed that, although the epistasis between two random mutations is on average 0,
epistasis between two beneficial mutations in FGM is on average negative. Gros et al. [41]
extended this result by deriving an analytical expression of the mean epistasis between two
random mutations, which can be positive, negative, or zero depending on the sharpness of the
fitness peak in the extended FGM. These theoretical distributions of epistasis have been
successfully matched with empirical data, both for pairs of random mutations and pairs of
beneficial mutations [40].

In reality, mutations occur in genotype space rather than phenotype space. Theoretical work has
recently shown that, although the FGM (despite the above-described pattern of negative epistasis)
is relatively smooth in its phenotype–fitness map, the underlying genotype–fitness map can be
highly epistatic [10,42]. Blanquart et al. [42] derived, under the assumption of FGM, that the amount
of epistasis in an extracted genotype–fitness landscape depends on the choice of mutations. As
predicted in this work, fitness landscapes tend to be smoother when built from mutations chosen
from a realized adaptive walk (e.g., [12,43]) than those built from independent beneficial mutations
that arose in the same reference background (e.g., [44]). A multitude of other experimental studies
have reported ubiquitous epistasis in experimental fitness landscapes (e.g., [12,14,45–47]).

At the molecular level, epistasis occurs through interactions between proteins within a pathway
(e.g., [45]) or between amino acids within a folded protein structure (e.g., [46]). In such cases

Trends in Ecology & Evolution, Month Year, Vol. xx, No. yy 5

 TREE 2453 No. of Pages 14

(A) Genotype Phenotype Fitness

Gene Molecule

ACTGACCGTA Structural stability

TGACTGGCAT catalyc acvity

Biomolecular network
Between-level complexity

Chromosome Expression paern

flux control

Cell

Growth rate
metabolic homeostasis
Genome

Organism

Compeve fitness
viability
ferlity

Within-level complexity

(B) Genotype Phenotype Fitness Reference

ACTGACCGTA Expression paern Fridelander et al. [77]

TGACTGGCAT

Neyfack et al. [111]

Metabolic homeostasis
Cuypers et al. [78]

ACTGACCGTA
Compeve fitness Harmand et al. [83]
TGACTGGCAT

Fig. 2. Components and examples of genotype–phenotype–fitness maps. (A) Components of genotype-to-phenotype-

to-fitness maps: such maps are normally degenerate because multiple genomic architectures give rise to fewer phenotypic
(Figure legend continued on the bottom of the next page.)

6 Trends in Ecology & Evolution, Month Year, Vol. xx, No. yy

 TREE 2453 No. of Pages 14

epistasis constrains how different subunits of a protein, a protein complex, or a pathway can
coevolve, or how different rates of enzymatic activities need to match within a pathway [48]. The
prevalence of epistasis within a biological network and its potential role in speciation was
recently illustrated by Ono et al. [49]. From yeast populations adapted to a fungicide [50], the
authors selected six first-step beneficial mutations within the ergosterol biosynthesis pathway,
each mutation in a different gene. Consistent with the interaction of the genes in the pathway,
they observed negative epistasis between the chosen mutations. This provides a biological
explanation for the accumulation of BDM incompatibilities in isolated populations subject to
identical environments. Mutations can speed up or slow down the production of ergosterol;
hybrids that carry a mix of mutations from different populations can then randomly deviate from
the optimum production speed and thus suffer from low fitness (also [51]). Surprisingly, the sign
of the observed epistasis changed with an increasing concentration of the drug. In addition to
linking biological networks and speciation, this study is noteworthy because it demonstrates
the environment-dependence of epistasis and because the authors showed that considering
the order of the genes in the pathway was not sufficient to explain the fitness of the double
mutants. A mechanistic model based on fitness landscape theory could potentially provide
additional power to explain the experimental results in the future.

Neutral evolution on fitness landscapes

Genetic drift of neutral mutations was proposed by Kimura as the main driver of evolution at the
molecular level, whereas natural selection would mostly govern phenotypic evolution [52]. At
first sight, this proposal is in stark contrast to models of genotype–fitness landscapes (e.g., NK,
RMF [53,54]) in which no neutral mutations exist, such that evolution is governed by selection,
often under the assumption that the effective population size is infinite ([3,55], also Box 3). We
provide here various counter-examples in which fitness landscape models have been useful to
understand how neutral evolution can aid both speciation and adaptation.

Specifically, the aforementioned concept of holey landscapes reconciles the neutral theory with
the concept of fitness landscapes. Recall that holey landscapes are multidimensional
genotype–fitness landscapes in which genotypes are either fit or lethal (Box 1). Schiffman
and Ralph [56] showed recently that a holey-landscape-like architecture can emerge from
complex molecular interactions that make up the cellular environment of an organism. The
authors investigated theoretically how neutral changes could give rise to phenotypic divergence

outputs, which coalesce into some (single) measure of fitness (reviewed in [110]). Genes and their nucleotide sequences
are taken as the fundamental unit of the genotype. These can be organized into groups of genes or gene clusters, and
further into chromosomes and whole genomes. Likewise, gene products are considered to be the basic building block of
the phenotype. The interaction of multiple gene products constitutes a biological network, and these can be used to map
the physiology of a cell [78,111]. Fitness proxies are highly dependent on the biological level being probed. For example,
structural stability (reviewed in [9,110]) or catalytic activity [47] are often considered as measures of fitness at the level of a
single molecule, whereas more complex parameters, such as growth rate [47] are a frequently used measure of fitness in
whole organisms. (B) Examples of genotype–phenotype–fitness maps. (Top) In this study [77] the consensus sequence of
two copies of a transcription factor-coding gene, as well as their genomic binding sites, are represented by a sequence of
nucleotides. The effect of the transcription factors on their downstream genes depends on the number of mismatches
between the consensus sequence and the corresponding binding site. Fitness is measured relative to the expression level
of the downstream genes. (Center) In these studies [78,111] a circular genome encodes the complete molecular
machinery of a virtual cell. The phenotypic component is determined by a network of transcription factors, metabolic
enzymes, and protein pumps that govern the uptake and processing of an external resource. The reproductive success of
these virtual organisms depends on maintaining the internal concentration of the metabolized nutrient at a given value.
(Bottom) In this study [83] a strain of E. coli was grown in different concentrations of nalidixic acid and the profile of antibiotic
resistance was assessed relative to the presence or absence of mutations in the gyrA gene. The fitness of the mutants was
measured in competition assays with a resistant strain and this was found to be consistent with an extended version of
Fisher's geometric model (FGM).

Trends in Ecology & Evolution, Month Year, Vol. xx, No. yy 7

 TREE 2453 No. of Pages 14

and possibly hybrid incompatibility. In the model, the gene regulatory network of individual
organisms is represented as a system of linear differential equations, whose output determines
an oscillatory phenotype that in turn determines the fitness of an individual. The authors showed
that speciation can occur without any direct selection because different and mutually exclusive
regulatory solutions are reached through drift. Thus, their mechanistic model recapitulates
previous results obtained for holey landscapes. Second, the accumulation of neutral standing
genetic variation in a population as a result of plateaus in the fitness landscape can pave the
way for innovation after a change in the environment. For example, Draghi et al. [57] simulated
the evolution of self-replicating RNA molecules before and after an environmental shift. They
found that standing genetic variation that was neutral in the initial background, but showed
positive epistasis with subsequent mutations, facilitated the appearance of subsequent bene-
ficial mutations by spreading the population on a fitness plateau. Such positively epistatic
effects of neutral mutations were predicted for biological networks [58], and have been
identified experimentally in viral proteins [59]. Thus, depending on the shape of the fitness
landscape and how it changes across environments, neutral mutations can provide oppor-
tunities for adaptation.

Finally, rapid fluctuations in the environment might render evolution on rugged fitness land-
scapes effectively neutral if mutational effects change across environments. Such behavior
can be captured by fitness seascapes, a concept developed to quantify adaptation in
fluctuating environments [60]. Although fitness seascape models ignore neutral mutations,
they might nevertheless converge to an effectively flat (static) fitness landscape when the
environmental fluctuations are faster than the fixation time of mutations [61]. This points to an
important open question. On the one hand, it is unknown how much fitness landscapes
change across environments; on the other hand, we do not know how much natural
environments fluctuate and, thus, how much commonly used fitness measures and the
resulting experimental fitness landscapes indeed reflect reproductive success in natural
populations.

Fitness landscapes as a framework to investigate the genotype–phenotype–

fitness map
Genotype–phenotype–fitness models determine the effect of mutations at the phenotypic level
and provide a more complete approach to studying fitness landscapes [62,63]. Although the

Box 2. Predicting outcomes of evolution using fitness landscape theory.

Fitness landscape models might allow us to predict evolution [5] through the analysis of adaptive walks and the estimation of the probability of the same outcome from
different starting points on the landscape. An adaptive walk is the sequential fixation of beneficial mutations that take the population from its initial (low) fitness value to
a fitness optimum [88–90]. Most models consider adaptive walks in asexual populations under a strong-selection weak-mutation regime, assuming that beneficial
mutations are rare and, when they appear, become fixed [3,55,89]. Three types of adaptive walks have been defined (Figure II): (i) random adaptive walks [91], (ii) true
adaptive walks [55,92], and (iii) greedy adaptive walks [88]. The type of walk considered determines the predictability of evolution on a given fitness landscape.
Whereas from a given starting point populations will always reach the same adaptive peak in greedy walks, in random walks the probability of reaching the same peak
in a rugged fitness landscape is low. For true adaptive walks, in which the next mutational step depends on the relative fitness advantage of all neighboring single-step
mutants, the probability of reaching the same endpoint depends strongly on the structure of the landscape, in particular the need to transverse fitness plateaus will
decrease said probability [3,88,89].

Unsurprisingly, the shape of the landscape also affects the predictability of evolution. The higher the number of local optima in the fitness landscape the smaller the
probability that populations with different starting genotypes will reach the same optimum. The number of peaks in a rugged landscape increases with the number of
loci and the prevalence of epistasis [93,94]. An increase in the number of peaks decreases the probability of reaching the global fitness optimum, but increases the
accessibility of the landscape (i.e., that at least one monotonic path exists between a low fitness genotype and a fitness peak [95,96]). The accessibility of the fitness
landscape also increases with the number of alleles per locus [97], for example when the amino acid level is considered instead of the nucleotide level [98]. So far it has
been argued that phenotypic evolution is more easily predictable than genetic evolution [5,99]. By integrating expectations of epistasis across biological levels,
genotype–fitness landscapes could develop from an abstract construct into an effective tool for the prediction of evolutionary paths in genotype space.

8 Trends in Ecology & Evolution, Month Year, Vol. xx, No. yy

 TREE 2453 No. of Pages 14

(A) (B)
1.10 1.10

1.08 1.08

1.06 1.06

1.04 1.04

1.02 1.02

1.00 1.00
0 1 2 3 4 5 6 0 1 2 3 4 5 6
Hamming distance from parental type Hamming distance from parental type

(C)
1.10

1.08

1.06

1.04

1.02

1.00
0 1 2 3 4 5 6
Hamming distance from parental type

Figure II. Predictability of the different types of adaptive walks on a fitness landscape, demonstrated using data from [14]. In all panels, the y axis corresponds to
fitness relative to the parental type and the x axis to the Hamming distance (the number of single mutational steps) from the parental type. In a greedy walk (A), the
fittest single-step mutant neighbor is chosen deterministically. Thus different realizations starting from the ancestral type end up in the same fitness peak [88]. In a
random walk (B), a new genotype is selected among the set of all fitter genotypes that are accessible by a single mutation [3,8]. This reduces the predictability of the
adaptive walk. In a true adaptive walk (C), the new genotype is chosen with a probability proportional to the relative fitness advantage of all single-step mutant
neighbors [3,55,88]. In this walk, large fitness differences between single-step neighbors lead to a higher predictability of the walk. Adaptive walks are represented by
blue lines, and fitness peaks are marked as red circles. For each panel 500 simulations were performed starting from the ancestral type. Darker shades of blue and
red indicate higher numbers of simulations using the path and ending in the fitness peak, respectively.

relationship between genotype and phenotype is complex [64], theoretical and empirical
studies have pointed out several general properties that could be incorporated into generalized
genotype–phenotype–fitness models. First, the dimensionality of the genotype space is larger
than its phenotypic counterpart (genotype redundancy). This results in phenotypic robustness
(many genotypes code for the same phenotype), which indicates that neutral evolution at the
genotype level can play an important role in the genotype–phenotype–fitness map [6,62,63,65].
Second, experimental and theoretical studies found that a so-called genotype bias exists at the

Trends in Ecology & Evolution, Month Year, Vol. xx, No. yy 9

 TREE 2453 No. of Pages 14

Box 3. Wright's shifting balance versus the drift barrier – how population size tunes the shape of the
fitness landscape.
The effective population size Ne determines the effectiveness of natural selection; mutations with a selection coefficient s
in the interval 1/2Ne < s < 1/2Ne behave indistinguishably from a neutral mutation [100]. Thus, differences in the
population size affect evolution by altering the perceived local shape of the fitness landscape that a population is
experiencing. On the one hand, a (temporal) reduction in the population size can make a landscape more connected, as
proposed in Wright's shifting balance theory [1,101]. On the other hand, the population size limits the ability of the
population to climb a fitness peak to the top, as illustrated in the drift–barrier hypothesis [102]. With smaller population
size, larger plateaus of effectively neutral genotypes are created. In the limit of small population size, this results in a
‘holey landscape’ with only neutral and lethal genotypes.

The population size also directly affects two parameters that relate to adaptation dynamics on the fitness landscape: the
rate of introduction of new mutations in a population increases with population size, whereas the probability of fixation of
those mutations once introduced decreases [103]. In large populations, mutations that fix will segregate for a longer time
before reaching fixation and, depending upon the mutation rate, can potentially encounter other segregating mutations.
In asexual populations, this leads to clonal interference, which makes the fixation of mutations dependent on both their
time of appearance and their selection coefficient, and which violates the strong-selection weak-mutation assumption
that is the basis of many theoretical fitness landscape studies (Box 2). The complex aspects of the population size-
dependence of adaptation dynamics on fitness landscapes have been studied both theoretically [48,104–106] and
experimentally [107–109].

RNA and protein level [65–69], which means that the distribution of the number of genotypes
per phenotype tends to be non-uniform, with the most frequent phenotypes being encoded by
a large proportion of the genotypes. Although the field is far from developing a general
approach that includes the described properties and constraints, we highlight here two types
of promising studies that have combined theoretical and experimental work to give a better
understanding of the relationship between genotype, phenotype, and fitness.

In molecular and systems biology, knowledge of biophysics coupled with high-throughput

experimental methods has laid the foundation for linking theoretical to experimental fitness
landscapes at the molecular and network level. For example, various genotype–phenotype–
fitness maps have been created based on information from RNA or protein folding or from
biological networks [62,63,70–72]. With respect to RNA or protein folding, the stability or
binding specificity of a molecule can be predicted from the biophysical properties of
nucleotide or amino acid sequences. These phenotypes can be interpreted as fitness
proxies of the molecule, allowing the impact of different evolutionary forces (mainly selec-
tion and drift) on molecular evolution to be tested [9,70,73]. With respect to biological
networks, the genotype–phenotype–fitness map arises from the interactions between
network components, for example metabolic and gene-regulatory networks. These are
usually defined by a set of nodes (e.g., genes or proteins, representing the genotype) and
connections (e.g., activation or repression). The output (e.g., expression level) represents
the phenotype, which can be mapped to fitness. Several studies have focused on investi-
gating how the general properties of these networks evolve, such as the number of network
components (complexity) [74–77] or the ability to generate novel phenotypes upon mutation
(evolvability) [48,75,78]. For example, Friedlander et al. [77] modeled a system in which a
transcription factor (TF) activates two downstream genes in response to two signals, and
followed its evolution after a duplication event to investigate the conditions that favor
specialization. In this model, the affinity of the TF to its DNA target is modeled according
to biophysical principles and acts as a proxy for gene expression and fitness. The authors
observed that, for one of the TFs, specificity to environmental cues evolves first while
maintaining promiscuity in its binding-site affinity, whereas the second TF regulates only
one gene, but is promiscuous in relation to the input signal. Thus, specialization proceeded

10 Trends in Ecology & Evolution, Month Year, Vol. xx, No. yy

 TREE 2453 No. of Pages 14

through small intermediate steps in which input signals are transmitted in a non-specialized
manner to both TFs (crosstalk), with the interactions between TF-binding site specificity
evolving in a neutral-like landscape for long periods of time, before specialization
takes place. This work represents an accurate, but highly specific and parameter-heavy
(bottom-up) genotype–phenotype–fitness map [75,79].

A complementary approach was taken by integrating the genotype–phenotype map into

phenotype–fitness models such as FGM (introduced above in the section ‘Epistasis in the
Light of Fitness Landscapes’). FGM naturally incorporates several important features of the
genotype–phenotype map, such as pleiotropy, epistasis, and the non-linear mapping between
genotype, phenotype, and fitness. Owing to this flexibility, FGM has been used to study various
evolutionary questions [3,8,10,18,40,79–81]. For example, FGM has been used to reproduce
many patterns characteristic of speciation [81] and hybridization [82]. Interestingly, and in
contrast to speciation models in which epistasis must be specifically defined (e.g., BDM),
genetic incompatibilities and epistasis are emergent properties of FGM (see section ‘Epistasis
in the Light of Fitness Landscapes’). The blindness of the FGM to the underlying molecular
features makes it readily applicable to experimental studies. Recently, Harmand et al. [83]
combined experimental and theoretical approaches and explicitly included the genotype–
phenotype map in FGM. The authors showed that, to describe the change of the fitness
landscape across a gradient of antibiotic concentrations, it is necessary to take into account
that mutations can impact differently on phenotypes across environments, and that changes in
the environment might affect the height of the fitness peak and thus the rate of fitness increase.
By introducing these theoretical considerations, they were able to draw a genotype–
phenotype–fitness map that captured the specifics of antibiotic resistance evolution in E. coli.
Thus, although the original FGM does not explicitly consider the genotype level, it provides a
conceptually simple framework that naturally incorporates the genotype–phenotype map
[8,10,18,42,79].

Another, less commonly used phenotype–fitness model is the multilinear approach that is
derived from quantitative genetics. In this model, the genotype–phenotype map is directly
modeled into the fitness landscape. The distribution of phenotypes in the population is
described by the vector of mean phenotypes, and the G matrix, which is the variance–
covariance matrix of the phenotypes. The mutational architecture of the phenotype distribution
is described by the M matrix, which can take into account not only additivity but also epistatic
interactions between loci [84]. This allows the evolutionary forces to act at different levels by
shaping both the G and M matrices [84–86]. The multilinear approach is similar to FGM, but
accommodates standing genetic variation that might exist in the population, and both recom-
bination and mutation can contribute to changes in mean trait values [84,86]. Although more
realistic, the multilinear approach comes with the caveat of being computationally heavy
because it requires tracking of the G and M matrices.

In summary, much progress has been made theoretically and experimentally regarding the
mapping of genotype to phenotype to fitness. Both bottom-up [62,63,70–72] and top-
down [83] approaches have generated genotype–phenotype–fitness maps, and these have
shown that integrating the genotype–phenotype relationship increases our ability to under-
stand and predict evolution [72,87]. A major limitation of the described models is that they
are either highly specific and heavily parameterized to describe a particular molecule or
network, or that they do not consider mechanistic and molecular constraints. It will be a
challenge to bridge this gap in future work, for example by incorporating biophysical
constraints into FGM.

Trends in Ecology & Evolution, Month Year, Vol. xx, No. yy 11

 TREE 2453 No. of Pages 14

Concluding remarks Outstanding questions

In the past 20 years the concept of fitness landscapes has developed from a purely theoretical Properties of local fitness landscapes
and philosophical construct into an applied framework to analyze and interpret large datasets. are often not informative of the shape
of the global fitness landscape. Which
Fitness landscape theory can be used to integrate the increasing amount of information on properties of the global fitness land-
observed genome sequences, the biophysical properties of specific molecules, and the fitness scape can be extrapolated from one
of multiple organisms in various sets of conditions that we can now gather across the subfields or several local fitness landscapes?
How is a local fitness landscape infor-
of biology. Feeding this information into a single theoretical framework bridges the gap between
mative of its vicinity in genotype
evolutionary, molecular, and systems biology, and thus creates opportunities for interdisciplin- space?
ary collaborative approaches to the study of evolution (see Outstanding questions).
How much does a fitness landscape
In this review we show that, although at first sight fitness landscapes have been used to change across an environmental gra-
dient? Can this change (both in single-
address apparently unrelated questions across fields, there are many commonalities between
mutation effects and epistasis) be pre-
the different models and the emerging results. For example, scenarios of neutral evolution in dicted using knowledge from molecu-
biological networks often relate to the dynamics of evolution on a ‘holey landscape’ which was lar and systems biology?
originally developed to describe speciation. Similarly, explicit genotype-fitness models such as
the house-of-cards model converge to a ‘holey landscape’ in the limit of small population size. How much do the various fitness prox-
ies used in experimental studies matter
Explicit efforts to connect these different levels, for example by incorporating the genotype–
for evolution in natural populations?
phenotype map into FGM, have shown promising steps towards predicting patterns of
evolution. Further combination of the different models in a unified framework will result in more How can the accumulating mechanis-
realistic fitness landscape models that can be compared to experimental data. Integrated into a tic information on complex biological
statistical framework, this approach could improve our understanding of the relative contri- systems be integrated into a general-
izable framework for the study of adap-
butions of (intrinsic) molecular and (extrinsic) ecological constraints to adaptation and specia- tation and speciation? Is there an
tion, and, ultimately, improve our ability to predict evolution. optimal level of model complexity for
predicting evolution?
Acknowledgments
We thank the reviewers and the members of the Evolutionary Dynamics lab for their careful reading and insightful
comments on the manuscript. This work was supported by Fundação Calouste Gulbenkian. I.F. was supported by a
postdoctoral fellowship from the FCT (Fundação para a Ciência e a Tecnologia) within the project JPIAMR/0001/2016. M.
A.D.L. was supported by a PhD fellowship from the FCT (PD/BD/139217/2018).

Appendix A Supplementary data

Supplementary data associated with this article can be found, in the online version, at https://doi.org/10.1016/j.tree.2018.
10.009.

References
1. Wright, S. (1932) The roles of mutation, inbreeding, crossbreed- 10. Hwang, S. et al. (2017) Genotypic complexity of Fisher's geo-
ing and selection in evolution. Proceedings of the Sixth Interna- metric model. Genetics 206, 1049–1079
tional Congress of Genetics 356–366 11. Weinreich, D.M. et al. (2006) Darwinian evolution can follow only
2. Gavrilets, S. (2009) High-dimensional fitness landscapes and very few mutational paths to fitter proteins. Science 312, 111–
speciation. In Evolution: The Extended Synthesis (Pigliucci, M. 114
and Müller, G.B., eds), pp. 45–80, MIT Press 12. Khan, A.I. et al. (2011) Negative epistasis between beneficial
3. Orr, H. (2005) The genetic theory of adaptation: a brief history. mutations in an evolving bacterial population. Science 332,
Nat. Rev. Genet. 6, 119–127 1193–1196
4. Svensson, E. and Calsbeek, R. (2012) The Adaptive Landscape 13. Chou, H.H. et al. (2014) Mapping the fitness landscape of gene
in Evolutionary Biology, Oxford University Press expression uncovers the cause of antagonism and sign epistasis
5. de Visser, J.A. and Krug, J. (2014) Empirical fitness between adaptive mutations. PLoS Genet. 10, e1004149
landscapes and the predictability of evolution. Nat. Rev. Genet. 14. Bank, C. et al. (2016) On the (un)predictability of a large intra-
15, 480–490 genic fitness landscape. Proc. Natl. Acad. Sci. 113, 14085–
6. McCandlish, D.M. (2011) Visualizing fitness landscapes. Evolu- 14090
tion 65, 1544–1558 15. Wu, N.C. et al. (2016) Adaptation in protein fitness landscapes is
7. Kaplan, J. (2008) The end of the adaptive landscape metaphor? facilitated by indirect paths. eLife 5 (July), 1–21
Biol. Philos. 23, 625–638 16. Szendro, I.G. et al. (2013) Quantitative analyses of empirical
8. Tenaillon, O. (2014) The utility of Fisher's geometric model in fitness landscapes. J. Stat. Mech.: Theory Exp. 2013, P01005
evolutionary genetics. Annu. Rev. Ecol. Evol. Syst. 45, 179–201 17. Otwinowski, J. and Plotkin, J.B. (2014) Inferring fitness land-
9. Chi, P.B. and Liberles, D.A. (2016) Selection on protein struc- scapes by regression produces biased estimates of epistasis.
ture, interaction, and sequence. Protein Sci. 25, 1168–1178 Proc. Natl. Acad. Sci. 111, E2301–E2309

12 Trends in Ecology & Evolution, Month Year, Vol. xx, No. yy

 TREE 2453 No. of Pages 14

18. Blanquart, F. and Bataillon, T. (2016) Epistasis and the structure 45. Costanzo, M. et al. (2010) The genetic landscape of a cell.
of fitness landscapes: are experimental fitness landscapes com- Science 327, 425–431
patible with Fisher's geometric model? Genetics 203, 847–862 46. Bank, C. et al. (2015) A systematic survey of an intragenic
19. Ferretti, L. et al. (2016) Measuring epistasis in fitness land- epistatic landscape. Mol. Biol. Evol. 32, 229–238
scapes: the correlation of fitness effects of mutations. J. Theor. 47. Bendixsen, D.P. et al. (2017) Negative epistasis in experimental
Biol. 396, 132–143 RNA fitness landscapes. J. Mol. Evol. 85, 159–168
20. Gavrilets, S. (2004) Fitness Landscapes and the Origin of Spe- 48. Orlenko, A. et al. (2017) Characterizing the roles of changing
cies, Princeton University Press population size and selection on the evolution of flux control in
21. Kondrashov, D.A. and Kondrashov, F.A. (2015) Topological metabolic pathways. BMC Evol. Biol. 17, 117
features of rugged fitness landscapes in sequence space. 49. Ono, J. et al. (2017) Widespread genetic incompatibilities
Trends Genet. 31, 24–33 between first-step mutations during parallel adaptation of Sac-
22. Echave, J. and Wilke, C.O. (2017) Biophysical models of protein charomyces cerevisiae to a common environment. PLoS Biol.
evolution: understanding the patterns of evolutionary sequence 15, e1002591
divergence. Annu. Rev. Biophys. 46, 85–103 50. Gerstein, A.C. et al. (2012) Parallel genetic changes and non-
23. Bateson, W. (1909) Heredity and variation in modern lights. parallel gene–environment interactions characterize the evolu-
Darwin Mod. Sci. 85–101 tion of drug resistance in yeast. Genetics 192, 241–252
24. Phillips, P.C. (2008) Epistasis – the essential role of gene inter- 51. Kvitek, D.J. and Sherlock, G. (2011) Reciprocal sign epistasis
actions in the structure and evolution of genetic systems. Nat. between frequently experimentally evolved adaptive mutations
Rev. Genet. 9, 855 causes a rugged fitness landscape. PLoS Genet. 7, e1002056
25. Coyne, J. and Orr, H. (2004) Speciation, Sinauer 52. Kimura, M. (1983) The Neutral Theory of Molecular Evolution,
26. Dobzhansky, T. (1936) Studies on hybrid sterility. II. Localization Cambridge University Press
of sterility factors in Drosophila pseudoobscura hybrids. Genet- 53. Kauffman, S.A. and Weinberger, E.D. (1989) The NK model of
ics 21, 113 rugged fitness landscapes and its application to maturation of
27. Muller, H. (1942) Isolating mechanisms, evolution and tempera- the immune response. J. Theor. Biol. 141, 211–245
ture. Biol. Symp. 6, 71–125 54. Neidhart, J. et al. (2014) Adaptation in tunably rugged fitness
28. Orr, H. (1995) The population genetics of speciation: the evo- landscapes: the rough Mount Fuji model. Genetics 198, 699–721
lution of hybrid incompatibilities. Genetics 139, 1805–1813 55. Gillespie, J. (1984) Molecular evolution over the mutational land-
29. Butlin, R. et al. (2012) What do we need to know about specia- scape. Evolution 38, 1116–1129
tion? Trends Ecol. Evol. 27, 27–39 56. Schiffman, J.S. and Ralph, P.L. (2018) System drift and specia-
30. Seehausen, O. et al. (2014) Genomics and the origin of species. tion. bioRxiv http://dx.doi.org/10.1101/231209 [Published
Nat. Rev. Genet. 15, 176–192 online January 26, 2018]

31. Presgraves, D. (2003) A fine-scale genetic analysis of hybrid 57. Draghi, J.A. et al. (2011) Epistasis increases the rate of condi-
incompatibilities in Drosophila. Genetics 163, 955–972 tionally neutral substitution in an adapting population. Genetics
187, 1139–1152
32. Kao, K.C. et al. (2010) A genome-wide analysis reveals no
nuclear Dobzhansky–Muller pairs of determinants of speciation 58. Draghi, J.A. et al. (2010) Mutational robustness can facilitate
between S. cerevisiae and S. paradoxus, but suggests more adaptation. Nature 463, 353
complex incompatibilities. PLoS Genet. 6, e1001038 59. Kryazhimskiy, S. et al. (2011) Prevalence of epistasis in the
33. Corbett-Detig, R. et al. (2013) Genetic incompatibilities are evolution of influenza A surface proteins. PLoS Genet. 7, 1–11
widespread within species. Nature 504, 135–137 60. Merrell, H.J. (1994) The Adaptive Seascape: The Mechanism of
34. Bank, C. et al. (2012) The limits to parapatric speciation: Evolution, University of Minnesota Press
Dobzhansky–Muller incompatibilities in a continent–island 61. Mustonen, V. and Lässig, M. (2009) From fitness landscapes to
model. Genetics 191, 845–863 seascapes: non-equilibrium dynamics of selection and adapta-
35. Blanckaert, A. and Bank, C. (2018) In search of the Goldilocks tion. Trends Genet. 25, 111–119
zone for hybrid speciation. PLoS Genet. 14, 1–23 62. Schaper, S. and Louis, A.A. (2014) The arrival of the frequent:
36. Wilke, C.O. et al. (2003) Compensatory mutations cause excess how bias in genotype–phenotype maps can steer populations to
of antagonistic epistasis in RNA secondary structure folding. local optima. PLoS ONE 9, 1–9
BMC Evol. Biol. 3, 3 63. Chen, H. et al. (2018) The genotype–phenotype relationships in
37. Gong, L.I. et al. (2013) Stability-mediated epistasis constrains the light of natural selection. Mol. Biol. Evol. 35, 525–542
the evolution of an influenza protein. eLife 2, e00631 64. Pigliucci, M. (2010) Genotype–phenotype mapping and the end
38. Pollock, D.D. et al. (2012) Amino acid coevolution induces an of the ‘genes as blueprint’ metaphor. Philos. Trans. R. Soc. B:
evolutionary Stokes shift. Proc. Natl. Acad. Sci. 109, E1352– Biol. Sci. 365, 557–566
E1359 65. Ahnert, S.E. (2017) Structural properties of genotype–pheno-
39. Fisher, R.A. (1930) The Genetical Theory of Natural Selection, type maps. J. R. Soc. Interface 14, 20170275
Oxford University Press 66. Ferrada, E. and Wagner, A. (2012) A comparison of genotype–
40. Martin, G. et al. (2007) Distributions of epistasis in microbes fit phenotype maps for RNA and proteins. Biophys. J. 102, 1916–
predictions from a fitness landscape model.Nat. Genet.39, 555–560 1925

41. Gros, P. et al. (2009) The evolution of epistasis and its links with 67. Greenbury, S. et al. (2014) A tractable genotype–phenotype
genetic robustness, complexity and drift in a phenotypic model map modelling the self-assembly of protein quaternary struc-
of adaptation. Genetics 182, 277–293 ture. J. R. Soc. Interface 11, 20140249

42. Blanquart, F. et al. (2014) Properties of selected mutations and 68. Dingle, K. et al. (2015) The structure of the genotype–phenotype
genotypic landscapes under Fisher's geometric model. Evolu- map strongly constrains the evolution of non-coding RNA. Inter-
tion 68, 3537–3554 face focus 5, 20150053
43. Chou, H.-H. et al. (2011) Diminishing returns epistasis among 69. Manrubia, S. and Cuesta, J.A. (2017) Distribution of genotype
beneficial mutations decelerates adaptation. Science 332, network sizes in sequence-to-structure genotype–phenotype
1190–1192 maps. J. R. Soc. Interface 14, 20160976
44. Rokyta, D.R. et al. (2011) Epistasis between beneficial mutations 70. Jimenez, J.I. et al. (2013) Comprehensive experimental fitness
and the phenotype-to-fitness map for a ssDNA virus. PLoS landscape and evolutionary network for small RNA. Proc. Natl.
Genet. 7, e1002075 Acad. Sci. 110, 14984–14989

Trends in Ecology & Evolution, Month Year, Vol. xx, No. yy 13

 TREE 2453 No. of Pages 14

71. Du Plessis, L. et al. (2016) How good are statistical models at 92. Orr, H.A. (2002) The population genetics of adaptation: the
approximating complex fitness landscapes? Mol. Biol. Evol. 33, adaptation of DNA sequences. Evolution 56, 1317–1330
2454–2468 93. Poelwijk, F.J. et al. (2011) Reciprocal sign epistasis is a neces-
72. Bershtein, S. et al. (2017) Bridging the physical scales in evolu- sary condition for multi-peaked fitness landscapes. J. Theor.
tionary biology: from protein sequence space to fitness of organ- Biol. 272, 141–144
isms and populations. Curr. Opin. Struct. Biol. 42, 31–40 94. Crona, K. et al. (2013) The peaks and geometry of fitness
73. Stich, M. et al. (2010) Phenotypic effect of mutations in evolving landscapes. J. Theor. Biol. 317, 1–10
populations of RNA molecules. BMC Evol. Biol. 10, 46 95. Franke, J. et al. (2011) Evolutionary accessibility of mutational
74. Soyer, O.S. and Bonhoeffer, S. (2006) Evolution of complexity in pathways. PLoS Comput. Biol. 7, e1002134
signaling pathways. Proc. Natl. Acad. Sci. 103, 16337–16342 96. Hegarty, P. and Martinsson, A. (2014) On the existence of
75. Fraser, J.S. et al. (2013) From systems to structure: bridging accessible paths in various models of fitness landscapes.
networks and mechanism. Mol. Cell 49, 222–231 Ann. Appl. Probab. 24, 1375–1395
76. Yubero, P. et al. (2017) The space of genotypes is a network of 97. Zagorski, M. et al. (2016) Beyond the hypercube: evolutionary
networks: implications for evolutionary and extinction dynamics. accessibility of fitness landscapes with realistic mutational net-
Sci. Rep. 7, 13813 works. PLoS Comput. Biol. 12, 1–18
77. Friedlander, T. et al. (2017) Evolution of new regulatory functions on 98. Fragata, I. et al. (2018) The fitness landscape of the codon space
biophysically realistic fitness landscapes. Nat. Commun. 8, 216 across environments. Heredity 121, 422–437
78. Cuypers, T.D. et al. (2017) Evolution of evolvability and pheno- 99. Dettman, J.R. et al. (2012) Evolutionary insight from whole-
typic plasticity in virtual cells. BMC Evol. Biol. 17, 60 genome sequencing of experimentally evolved microbes. Mol.
79. Martin, G. (2014) Fisher's geometrical model emerges as a Ecol. 21, 2058–2077
property of complex integrated phenotypic networks. Genetics 100. Ohta, T. (1973) Slightly deleterious mutant substitutions in evo-
197, 237–255 lution. Nature 246, 96–98
80. Barton, N. (2001) The role of hybridization in evolution. Mol. Ecol. 101. Coyne, J.A. et al. (2000) Is Wright's shifting balance theory
10, 551–568 important in evolution? Evolution 54, 306–317
81. Fraïsse, C. et al. (2016) The genetics of speciation: insights from 102. Lynch, M. and Hagner, K. (2015) Evolutionary meandering of
Fisher's geometric model. Evolution 70, 1450–1464 intermolecular interactions along the drift barrier. Proc. Natl.
82. Simon, A. et al. (2018) Coadapted genomes and selection on Acad. Sci. 112, E30–E38
hybrids: Fisher's geometric model explains a variety of empirical 103. Kimura, M. (1957) Some problems of stochastic processes in
patterns. Evol. Lett. 2, 472–498 genetics. Ann. Math. Stat. 28, 882–901
83. Harmand, N. et al. (2017) Fisher's geometrical model and the 104. Jain, K. et al. (2011) Evolutionary advantage of small populations
mutational patterns of antibiotic resistance across dose gra- on complex fitness landscapes. Evolution 65, 1945–1955
dients. Evolution 71, 23–37 105. Handel, A. and Rozen, D.E. (2009) The impact of population size
84. Jones, A.G. et al. (2014) Epistasis and natural selection shape the on the evolution of asexual microbes on smooth versus rugged
mutational architecture of complex traits. Nat. Commun. 5, 3709 fitness landscapes. BMC Evol. Biol. 9, 1–10
85. Hansen, T.F. and Wagner, G.P. (2001) Modeling genetic archi- 106. Goldstein, R.A. (2013) Population size dependence of fitness
tecture: a multilinear theory of gene interaction. Theor. Popul. effect distribution and substitution rate probed by biophysical
Biol. 59, 61–86 model of protein thermostability. Genome Biol. Evol. 5, 1584–
86. Arnold, S.J. et al. (2008) Understanding the evolution and sta- 1593
bility of the G-matrix. Evolution 62, 2451–2461 107. Wahl, L.M. et al. (2002) Evaluating the impact of population
87. Lässig, M. et al. (2017) Predicting evolution. Nat. Ecol. Evol. 1, bottlenecks in experimental evolution. Genetics 162, 961–971
0077 108. Lachapelle, J. et al. (2015) Repeatability of adaptation in experi-
88. Orr, H.A. (2003) A minimum on the mean number of steps taken mental populations of different sizes. Proc. R. Soc. B: Biol. Sci.
in adaptive walks. J. Theor. Biol. 220, 241–247 282, 20143033

89. Seetharaman, S. and Jain, K. (2014) Length of adaptive walk on 109. Frickel, J. et al. (2018) Population size changes and selection
uncorrelated and correlated fitness landscapes. Phys. Rev. E: drive patterns of parallel evolution in a host–virus system. Nat.
Stat. Nonlinear Soft Matter Phys. 90, 1–12 Commun. 9, 1706

90. Heredia, J.P. et al. (2017) Selection limits to adaptive walks on 110. Walter, F. (2002) Modelling ‘evodevo’ with RNA. BioEssays 24,
correlated landscapes. Genetics 205, 803–825 1164–1177

91. Macken, C.A. and Perelson, A.S. (1989) Protein evolution on 111. Neyfakh, A.A. et al. (2006) A system for studying evolution of life-
rugged landscapes. Proc. Natl. Acad. Sci. 86, 6191–6195 like virtual organisms. Biol. Direct 1, 23

14 Trends in Ecology & Evolution, Month Year, Vol. xx, No. yy