The community-function landscape of microbial consortia

Alvaro Sanchez1,2, Djordje Bajic1, Juan Diaz-Colunga1, Abigail Skwara1, Jean Vila1 &
Seppe Kuehn3,4

1
Department of Ecology & Evolutionary Biology & Microbial Sciences Institute, Yale
University, New Haven, Connecticut, USA. 2Department of Microbial Biotechnology,
CNB-CSIC, Campus de Cantoblanco, Madrid, Spain. 3 Center for the Physics of Evolving
Systems, The Unviersity of Chicago. Chicago, Illinois, USA. 4Department of Ecology and
Evolution, The University of Chicago, Chicago, Illinois, USA.

# Correspondence: A.S: [email protected], JDC: [email protected]

, DB: [email protected] , A.Sk: [email protected], JV: [email protected]
SK: [email protected] ,

Abstract
Quantitatively linking the composition and function of microbial communities is a major
aspiration of microbial ecology. It is also a critical step in the path towards engineering
synthetic consortia and manipulating natural microbiomes. The functions of microbial
communities are collective properties that emerge from a complex web of molecular
interactions between individual cells, which in turn lead to population-level interactions
among strains and species. Incorporating this complexity into predictive models has been
highly challenging. A similar problem of predicting phenotype from genotype has been
addressed for decades in the field of quantitative genetics, leading to advances in the
fields of protein and molecular engineering. By analogy to the genotype-phenotype
landscape, an ecological community-function (or structure-function) landscape could be
defined that maps community composition and function. In this piece, we present an
overview of our current understanding of these community landscapes, their uses,
limitations, and open questions. We argue that exploiting the parallels between both
landscapes could bring powerful predictive methodologies from evolution and genetics
into ecology, providing a boost to our ability to engineer and optimize microbial consortia.

1
Introduction.

Microorganisms have colonized every habitat on earth, forming complex and diverse
ecosystems that play critical roles throughout the biosphere. Besides their many
environmental roles, microbial communities have also been harnessed for
biotechnological applications at least since the dawn of the neolithic revolution. The
biotechnological applications of microbial consortia are growing from their traditional roles
in food and drink (Belda et al., 2017; Blasche et al., 2021; May et al., 2019; Wolfe et al.,
2014), to contemporary uses in biofuel production (Alper and Stephanopoulos, 2009;
Jiang et al., 2020; Minty et al., 2013; Senne de Oliveira Lino et al., 2021), the valorization
of discarded plant materials (Hu et al., 2017; Maleki et al., 2018; Weng et al., 2008),
bioremediation (Piccardi et al., 2019; Swenson et al., 2000a; Zanaroli et al., 2010), crop
fertilization (Baas et al., 2016, 2020), and many more (Ergal et al., 2020; Macia et al.,
2016; Roell et al., 2019; Sgobba and Wendisch, 2020).

Relative to monocultures, microbial communities offer multiple advantages in

biotechnology. Among these, they permit specialization and division of labor (Roell et al.,
2019; Thommes et al., 2019) avoiding physiological and cellular tradeoffs and other
constraints that limit the efficiency of many biochemical processes. Communities may
also contain much more genetic diversity than one would find in a single organism due to
genome size limitations. This diversity can enable communities to remain resilient to
perturbations that single strains might not survive (Erkus et al., 2013). Finally, microbial
consortia form spontaneously through evolutionary and ecological processes that are
very difficult to avoid, even when a monoculture is started from a single isogenic
population and propagated under otherwise sterile laboratory conditions (Good et al.,
2017; Kinnersley et al., 2009; Rozen and Lenski, 2000). Even in environments supplied
with a single limiting resource, diversity and coexistence always seem to find a way,
suggesting that a community is the natural endpoint of microbial systems both in natural
and synthetic conditions (Dal Bello et al., 2021; Estrela et al., 2021a, 2021b; Goldford et
al., 2018; Mancuso et al., 2021). Learning how to manipulate and engineer microbial
consortia is therefore critical to realizing the biotechnological potential of microorganisms.

2
Despite growing interest, our ability to engineer microbial consortia lags behind
bioengineering efforts in other biological systems at or below the organismal level, such
as proteins (Arnold, 2019; Lu et al., 2022) or metabolic and genetic networks (Wendisch
et al., 2006; Yang et al., 2021). One major reason is the nested hierarchical complexity
present in a consortium. Specifically, the collective properties and services provided by
microbial consortia (i.e. their "functions") emerge from the contributions of individual
community members and their interactions with one another and their environment. The
physiological traits of individual taxa dictate interactions, and these traits depend on
genomic diversity, regulatory variation, and life-history. Community functions then
emerge from the collective action of these interactions, which are often non-linear and
historically contingent. This means that parsing the mapping from structure to function
from a detailed accounting of each process in the community is an immense task even
for relatively simple consortia. Amidst this complexity, how are we to approach the
problem of community design and control?

Mapping community composition to function can draw inspiration from protein

engineering. The field of molecular engineering has very similar goals and has
encountered similar challenges. For instance, protein engineers seek to design enzymes
with desirable catalytic activities (Bornscheuer et al., 2012; Chica et al., 2005; Kuchner
and Arnold, 1997). The catalytic rate of an enzyme is encoded in its sequence of amino
acids, and it is also a collective property of the enzyme that arises from a large number
of local and long-range biophysical interactions between its amino acids. These
interactions give rise to the folded structure of the enzyme and govern its stability and
intermolecular dynamics. Engineering every possible amino acid interaction to produce a
desired enzymatic function is obviously daunting, but even the simpler task of predictively
connecting sequence with function has been a major open challenge in biophysics.
However, this has not precluded our ability to engineer and optimize enzymatic function
(Arnold, 2019; Bornscheuer et al., 2012; Chica et al., 2005). In the process of
understanding the connection between amino acid sequence and function, protein
engineering has benefited greatly from insights provided by the theory of fitness

3
landscapes (Bloom and Arnold, 2009; Romero and Arnold, 2009; Tracewell and Arnold,
2009). Perhaps the most successful example is the development of directed evolution,
which has enabled the top-down engineering of different kinds of proteins (Arnold, 2019).
Directed evolution involves an assisted exploration of the genotype-phenotype map in
search for genotypes of desired or optimized functionality (this map is often referred to as
the fitness landscape in the context of directed evolution where an objective function -i.e.
"fitness" can be externally imposed). This assisted search is implemented through a
process that mimics that of evolution by the iterative application of sequence
randomization followed by selection on expressed phenotypes (Arnold, 2019).

In addition to the algorithmic explorations of fitness landscapes, a complementary

approach has been to infer the principles of protein design by examining the statistics of
sequence variation in naturally occurring proteins - effectively learning the landscape from
extant variation. This approach has enabled the synthetic design of functional enzymes
(Russ et al., 2020), inferring folds (Morcos et al., 2011), and insights into evolvability
(Stiffler et al., 2015) and allostery (Raman et al., 2016). One key insight from this body of
work is that within the astronomically large space of possible protein sequences, natural
functional proteins inhabit a much lower-dimensional subspace (Halabi et al., 2009). This
result means that engineering proteins does not require an exhaustive search of
sequence space (an impossible task) but instead a constrained search within a low-
dimensional subspace. It is intriguing to note that this inherent low-dimensionality is also
observed in biological systems at higher scales of organization, such as behavior
(Berman et al., 2013; Jordan et al., 2013), developmental programs (Alba et al., 2021),
and even microbial communities (Raman et al., 2019).

Can we extend the theory of fitness landscapes to study and engineer microbial
community function? An important challenge is that, unlike molecular systems, microbial
communities are made up of multiple self-replicating individual genotypes, each
possessing their own fitness landscapes. It is therefore not immediately obvious how the
idea of fitness landscapes may be extended to entire communities. In particular, any
notion of fitness at the community level is not clearly defined given the independent

4
replication of genotypes rather than communities as a whole. While this is true,
community-level selection can be applied under artificial conditions, where an arbitrary
fitness function can be applied (Blouin et al., 2015; Sánchez et al., 2021; Swenson et al.,
2000b). More broadly, for a landscape to exist it is not necessary that the scalar property
that is being mapped to the composition of the community be defined in terms of fitness;
it can instead be any collective function of the community (Chang et al., 2021a; Mueller
Ulrich G. et al.; Wright et al., 2019).

In recent years, a small but growing body of work has started to extend the theory of
fitness landscapes to communities and suggested ways in which it may help us guide the
design of microbial consortia (Baranwal et al., 2022; Gould et al., 2018; Sánchez et al.,
2021; Sanchez-Gorostiaga et al., 2019; Senay et al., 2019). Examples range from fruit-
fly microbial consortia whose function is the host's lifespan and other life-history traits
(Arora et al., 2020), to sugarcane biorefinery consortia whose function is the amount of
ethanol produced during a single-batch fermentation (Senne de Oliveira Lino et al., 2021).
These and other studies (Bittleston et al., 2020; Clark et al., 2021; Eble et al., 2021;
George and Korolev, 2021; Gopalakrishnappa et al., 2022; Gould et al., 2018; Sanchez-
Gorostiaga et al., 2019; Senay et al., 2019; Xie and Shou, 2021; Xie et al., 2019) have
formally defined the structure-function (or composition-function, or community-function)
landscape as the empirical map between community composition and function in a given
habitat and set of conditions. The structure of a microbial consortium is given by the list
of all it genotypes g = {g1,g2,...,gn} and their respective abundances xg = {x1,x2,...,xn}. If a
molecular fitness landscape is a map between a genotype g (where g represents the DNA
sequence of the molecule) and a quantitative phenotype P (i.e. P(g)), a community
structure-function landscape can be conceptualized as the map between the abundance
vector xg and a collective function F of the consortium F(xg).

To make this concept useful and productive, it is critical that we identify and understand
the similarities and differences between structure-function landscapes and molecular
fitness landscapes. The goal of this paper is to synthesize our current understanding of
the community structure-function landscape, highlighting promising directions and open

5
questions. We start by drawing parallels between genetic interactions (epistasis) in simple
genetic landscapes, and their ecological analogs in simple structure-function landscapes.
We then discuss how various concepts from fitness landscape theory may be generalized
to communities. Finally, we discuss under what conditions an ecological structure-
function landscape is defined, so that a collective property of interest can be said to
depend uniquely on species composition. Our focus is eminently practical, and we focus
on those ideas and methods from fitness landscape theory that, in addition to providing
ecological insights, may help us guide our efforts to engineer and manage community
services and functions. We also highlight how “landscape thinking” (Wagner, 2019) may
provide a helpful theoretical framework to help us conceptualize the challenges
associated with engineering microbial consortia.

A simple example of landscape thinking in community function: An ecological

parallel to epistasis. To develop our intuition of how fitness landscape theory may be
extended to microbial communities, it is useful to start from the simplest scenario. The
simplest genotype-phenotype map consists of two mutations, a->A and b->B, which
define four possible genotypes: the "wild-type" (ab), the two single-mutants: (Ab and aB),
and the double mutant (AB) (Fig. 1A). One then needs a null model that describes how
both mutations combine their effects when they act independently on the phenotype.
Typically, the null model assumes that mutations act additively on the phenotype (or
multiplicatively, depending on the scale). The deviation between the phenotype of the
double mutant AB and its expected value under the null, interaction-free model, is known
as the pairwise "epistasis" between those mutations (Fig. 1A). Thus defined, epistasis
gives us a metric of interactions between mutations.

Interactions can similarly be defined in other combinatorial systems that are not genetic,
and in fact the term"epistasis" has been used to describe systems as diverse as drug
interactions (Tekin Elif et al., 2016; Wood et al., 2012) or combinations of stressors
(Beppler Casey et al., 2016), among others (Tekin et al., 2018). In recent years, we (and
others) have extended it to ecological systems as well (Eble et al., 2021; Gould et al.,
2018; Guo and Boedicker, 2016; Sanchez, 2019; Sanchez-Gorostiaga et al., 2019; Senay

6
et al., 2019; Senne de Oliveira Lino et al., 2021), and the underlying idea was already
present in earlier efforts to model the emergence of community function (Chen et al.,
2009; Eng and Borenstein, 2019).

Fig. 1. Species interactions create non additive effects on community function. (A) In population
genetics, two mutations A and B are said to interact when their phenotypic effects do not combine additively
(or multiplicative, depending on the scale). This interaction is quantified by the deviation from additivity
(referred to as the epistasis, 𝜀). (B) Empirical measurements have found that the function of pairwise
microbial co-cultures is often described by the sum of the functions in monoculture, as exemplified here by
the amylolytic activities (in hr-1) of monocultures and pairwise co-culture of B. mojavensis and B.
thuringiensis (Data from (Sanchez-Gorostiaga et al., 2019)). Other pairs, however, exhibit marked
deviations. For instance, the pair formed by B. thuringiensis and P. polymyxa (C) has an amylolytic rate
that far exceeds the expected value if both species acted independently. Three different types of interactions
may cause this deviation from the situation where species functional contributions are additive (D). For
instances, the enzymes and other molecules secreted by each species may interact with one another either
enhancing or limiting their amylolytic activity (Biochemical interactions, E). Alternatively, a species may
promote (or suppress) the growth of its partner, limiting the size of its population and thus, potentially, its
net expression of amylases (F). Finally, a population of one species may impact the per-capita expression
of amylases by another, similarly impacting the net production of this function (G).

7
In ecology, the simplest type of consortium is one containing just two different genotypes,
g1 and g2. We could inoculate identical habitats with either cells from just one of those
genotypes (g1), the other (g2), or both (g1 and g2), and measure a function of interest of
each habitat after some defined incubation time. We could then establish a null model
that would describe the function of the pairwise consortium if both species did not interact
with one another in any way (Sanchez-Gorostiaga et al., 2019). By analogy with the
epistasis concept in genetics, the deviation between the function of the pairwise
consortium and the expected value under the null model, which assumes no interactions,
is defined as the functional interaction between both genotypes, an ecological equivalent
of epistasis.

To illustrate this idea, in Fig. 1B-C, we present a recent empirical example of a simple
structure-function landscape. In this example, drawn from ref. (Sanchez-Gorostiaga et al.,
2019), the function of interest is the rate of starch degradation by extracellularly amylase
enzymes secreted by different strains of the phylum Bacillota. Biochemical modeling tells
us that these enzymes should combine additively, a point that was confirmed empirically
(Sanchez-Gorostiaga et al., 2019). Therefore, in the absence of any interactions the
amylolytic rate function of any consortium should be the sum of the functions of each
genotype in monoculture. Indeed, many genotype pairs were very well described by this
interaction-free model (e.g., as shown in Fig. 1B, the one formed by B. mojavensis and
B. thuringiensis). The (surprising) effectiveness of simple additive models has been
reported in other systems, as a recent study showed similar success with an additive
regression model for predicting fluxes of nitrate and nitrite through synthetic denitrifying
communities (Gowda et al., 2022). Interestingly though, other genotype pairs in the starch
degrading communities, deviated markedly from the additive model (Fig. 1C), indicating
the existence of strong, pairwise functional interactions between them. These interactions
indicate the presence of “epistasis”-like interactions in these simple community-function
landscapes (Sanchez-Gorostiaga et al., 2019).

8
What is the mechanistic basis of these pairwise interactions? In general, functional
interactions may arise from three different mechanisms (Fig. 1D-F) (Sanchez-Gorostiaga
et al., 2019). First, the functional contributions of each community member may interact
with each other. For instance, going back to the secreted enzyme example that is serving
as illustration, enzymes secreted by two species may act independently on the substrate,
in which case their catalytic rates will be additive. However, some enzymes act
synergistically on their substrate, as is the case of endo- and exo-cellulases: the former
create new substrates for the latter, reaching an activity together that is higher than the
sum of each of them separately (Kim et al., 2014). The enzymes secreted by each species
may also act antagonistically, for instance by aggregating (and therefore inhibiting) one
another. These deviations from additivity may be called “abiotic” interactions, as they
would occur even if no cells were present. The second type of interaction involves
changes in the amount contributed by a given genotype to the community function. For
instance, a genotype may either promote or inhibit the per-capita functional contribution
by another genotype, altering its behavior. These “behavioral” interactions may include
chemical signaling from one species that modifies the behavior of another (Mickalide and
Kuehn, 2019). Alternatively, a genotype may affect the growth (and therefore the total
number of cells in the population) of another genotype. These “population” interactions
can also alter the collective function of the ecosystem in a context-dependent manner.
The three types of interactions summarized in Fig. 1 D-F can be separated empirically
(Sanchez-Gorostiaga et al., 2019).

High-Order Functional Interactions. In communities with more than two species,

functional interactions may be more complex than pairwise (Guo and Boedicker, 2016;
Mickalide and Kuehn, 2019; Sanchez-Gorostiaga et al., 2019; Senne de Oliveira Lino et
al., 2021). Consider, for instance, the example provided in Fig. 2, where the structure-
function landscape comprising every combinatorial consortia of three amylolytic bacteria
is given (Sanchez-Gorostiaga et al., 2019). This landscape shows that co-culturing P.
polymyxa with B. mojavensis or B. subtilis increases function beyond what we might
expect from the additive model, indicating the presence of strong pairwise interactions.
Yet, the beneficial effect of adding both B. mojavensis or B. subtilis to P. polymyxa is

9
negligible, as there is no additional benefit of adding those strains. This "diminishing
returns'' effect indicates that the same genotype (e.g. B. subtilis) that is functionally
"beneficial" when added to with P. polymyxa alone is functionally neutral when added to
a consortium formed by P. polymyxa and B. mojavensis. The functional effect of adding
a species to a consortium is thus different when two species, as opposed to one, are
present. This would be the canonical definition of high-order epistasis if, instead of
species and their functional effect, we were talking about mutations and their fitness effect
(Poelwijk et al., 2019; Sanchez, 2019).

Besides the example discussed above, high-order functional interactions (HOFIs) have
been observed in the production of ethanol by sugarcane biorefinery consortia (Senne de
Oliveira Lino et al., 2021), the extension of a host lifespan by Drosophila gut microbiome
consortia (Gould et al., 2018), the metabolic activity of synthetic consortia (Guo and
Boedicker, 2016), and, more recently, gene expression in simple defined communities
(Morin et al., 2022). Just as they do in fitness landscapes, high-order functional
interactions could have profound implications for the topography of structure-function
landscapes. For instance, in sugarcane biorefinery consortia, HOFIs have been found to
tone down the predominantly negative effects of pairwise interactions between bacteria
on the net ethanol yield (Senne de Oliveira Lino et al., 2021). Based on pairwise
interactions alone, we would have expected that as bacterial biodiversity increases in our
bioreactors the ethanol yield would have collapsed. Yet, the opposite was true, and while
most pairs of bacteria had negative effects in the ethanol yield, this detrimental effect
vanished as communities increased in richness, reaching average levels that were
comparable to those of pure yeast monocultures (Senne de Oliveira Lino et al., 2021).
Despite this and other recent attempts to characterize high-order functional interactions
(Eble et al., 2021; Gould et al., 2018; Sanchez-Gorostiaga et al., 2019), our understanding
of the effect and implications of HOFIs is still very incomplete. When do they complicate
and when do they simplify the navigability of structure-function landscapes? How do they
affect the number and stability of functionally stable equilibria? These are still open
questions, representing an open frontier in functional microbial ecology.

10
Fig. 2. High-order Functional Interactions in microbial consortia. We show an example of a third-order
interaction that shapes the function of microbial consortia, in this case leading to diminishing returns.
Adding either B. subtilis (S) or B. mojavensis (M) to a monoculture of P. polymyxa (P) dramatically
enhances its function through a pairwise functional interaction. Yet, when we add either B. subtilis or B.
mojavensis to the co-culture of P. polymyxa with the other partner, their impact on function is either neutral
or negative. This shows that the functional effect of adding a species to a consortium may be different when
a second species is present, indicating the existence of a High-Order Functional Interaction (HOFI).

An ecological parallel to global epistasis and the emergence of simple Functional

Effect Equations. Building predictive models of the structure-function landscape from
the bottom-up, by combining additive, pairwise, third-order interactions, etc., is generally
challenging. There is no guarantee that the complexity of interactions ends at the second
or third-order (Sanchez-Gorostiaga et al., 2019), so the number of interactions that one
would need to measure in order to build a predictive model of the landscape can blow up.
An alternative is provided by defining global functional interactions in a way that is inspired
by recent developments in quantitative genetics. Genetic interactions can be partitioned

11
as the sum of a "global epistasis" effect, where the fitness effect of a mutation is predicted
by the fitness of the genetic background, and an "idiosyncratic epistasis" effect, which
captures the part of the fitness effect of a mutation that depends on the genetic
background while being independent of the background fitness (Reddy and Desai, 2021)
(Fig. 3A).

Can we extend this way of partitioning interactions to microbial consortia? In recent work,
we have found that, indeed, the functional effects of adding a species to a consortium
does often scale linearly with the function of the background consortium, in a way that is
very similar to what has been observed in genetic systems (Díaz-Colunga et al., 2022).
Examples include diminishing returns, as well as increasing costs, accelerated returns,
and others (Fig. 3B). The existence of these global functional interaction patterns appears
to be rather general in ecosystems as we also found them in plant and algal communities
(Díaz-Colunga et al., 2022). Importantly, different species within a consortium tend to
have different "Functional Effect Equations" describing their unique, global functional
interaction patterns (Díaz-Colunga et al., 2022). How the particular global functional
patterns exhibited by a species depend on its traits, as well as the traits of the species it
interacts with, is still not well understood. In addition, it will be interesting to understand
how this simple “global” epistasis emerges from the pairwise and potentially higher-order
interactions in the consortia, extending and complementing the work that is currently
being done to understand the origins of global epistasis in genetic fitness landscapes
(Husain and Murugan, 2020; Otwinowski et al., 2018; Reddy and Desai, 2021; Wei and
Zhang, 2019).

12
Fig. 3. An analog to global epistasis explains the functional effect of adding new species to microbial
consortia. (A) Research in quantitative genetics has shown that the fitness effect of a mutation often
depends linearly on the fitness of the genetic background where it arises. Epistasis can thus be partitioned
as the sum of a global component captured by such a linear fit (red), and an idiosyncratic component, not
predictable from the fitness of the genetic background alone, represented by the residuals of that fit (green).
Data from ref. (Khan et al., 2011). (B) An ecological parallel to global epistasis can be formulated: the
effect on ecosystem function resulting from the addition of a species to a community (an ecological
background) often scales linearly with the function of the ecological background itself. Species can have

13
less beneficial (or more deleterious) functional effects in backgrounds with higher functions (red lines), or
vice-versa (blue lines). These regressions that capture the functional effect of adding a species to a gamut
of different consortia have been termed “Functional Effect Equations” (FEEs) (Díaz-Colunga et al., 2022).
In some cases, the functional effect of a species may be dominated by an idiosyncratic component rather
than a global one (black lines). Data corresponds to butyrate production by synthetic gut microbial
communities (Baranwal et al., 2022), biomass in plankton communities (Ghedini et al., 2022), above ground
biomass in multi-species plant communities (Kuebbing et al., 2015), xylose oxidation by soil bacterial
communities (Langenheder et al., 2010) and amylase secretion in bacterial consortia (Sanchez-Gorostiaga
et al., 2019) (C) These functional trends can be exploited to predict the function of any combinatorial
assemblage of species, and thus reconstruct entire ecological structure-function landscapes from a small
subset of empirical observations. Data corresponds to the amount of pyoverdine secreted from newly
assembled microbial consortia, whose function was predicted from a simple model that concatenates the
Functional Effect Equations (FEEs) of each species as discussed in (Díaz-Colunga et al., 2022).

The usefulness of the structure-function landscape concept. An important

consequence of the existence of these predictive Functional Effect Equations is that they
make it possible to predict with reasonable accuracy how adding a given species to a
consortium will change its function. This illustrates what may be one of the most important
benefits of bringing the concept of a structure-function landscape from genetics to
ecological research: that we could apply the arsenal of analytical and statistical tools that
have been developed in genetics to infer and navigate these landscapes. For instance,
several machine learning methodologies have been developed in recent years to infer a
full genotype-phenotype landscape from a small subset of measured genotype -
phenotype relationships. These methods have found impressive success in predicting
biological function from DNA sequence under constant environmental conditions
(Romero et al., 2013; Tareen et al., 2022; Tonner et al., 2021). Adapting and applying
these methodologies to microbial consortia is an exciting prospect (Baranwal et al., 2022),
and its feasibility is encouraged by the success of simpler inference approaches. For
instance, we have recently tested the predictive power of a simple model consisting of
"stitching together" the Functional Effect Equations of all community members (Díaz-
Colunga et al., 2022). This very simple approach does an excellent job at predicting

14
various community functions for the full set of all possible consortia one may form with a
defined set of taxa. Importantly, the ability to predict the full structure-function landscape
makes it possible to identify the community compositions that will maximize and minimize
these functions, paving the way to engineering community functions from the bottom-up.
The application of machine learning and neural networks to reconstruct community-
function landscapes from a limited set of observations is still in its infancy. However,
promising results are being published (Baranwal et al., 2022), and the success of earlier
regression-based approaches to predict the landscape of small consortia (Chen et al.,
2009) is also an encouraging sign.

The landscape perspective allows one to approach the problem of community design from
a statistical point of view. We propose that, from this perspective, the complex hierarchy
of processes discussed above that influence the structure-function landscape might yield
to simple descriptions. Indeed, our recent work suggests that taking this perspective can
uncover simple rules for mapping genomes to phenotypes (Gowda et al., 2022) and
community composition to emergent function (Díaz-Colunga et al., 2022). Despite these
advances, we do not yet have a clear picture of the topography of these structure-function
landscapes and this will be important if what we wish is to optimize communities using
evolutionary engineering approaches.

The topography and navigability of an ecological structure-function landscape. The

topography of a fitness landscape gives us a measure of its navigability by either evolution
or other assisted search processes. Smooth single-peak landscapes are navigated more
easily than rugged ones since there are a larger number of adaptive paths connecting a
given genotype to the global fitness peak (Aguilar-Rodríguez et al., 2017; Nahum et al.,
2015). Smoothness is high when different mutations act independently, whereas
ruggedness increases in the presence of interactions between mutations (epistasis). In
particular, strongly positive interactions between deleterious mutations (reciprocal sign
epistasis), play a key role in determining landscape navigability as they are necessary for
the presence of multiple fitness peaks (Poelwijk et al., 2011). In multi-peaked fitness

15
landscapes evolutionary algorithms can become trapped on local optima and fail to find
the global fitness peak.

The simplest evolutionary algorithms used to navigate fitness landscapes involve an

iterative two-step process consisting of a selection of the mutants of highest fitness,
followed by sequence randomization. These belong to the class of “hill-climbing” search
algorithms, which work particularly well for smooth landscapes. Rugged fitness
landscapes with many distinct peaks, on the other hand, are more challenging to search
through a hill-climbing approach (George and Korolev, 2021; Wittmann et al., 2021),
because local information is not informative globally. By the same logic, the ruggedness
of the ecological community-function landscape will also determine its navigability using
analogous hill-climbing search algorithms, such as the directed evolution approaches
reviewed in (Sánchez et al., 2021). For example, consider one configuration of a
community that gives rise to a function which is locally a maximum, meaning that any
small change in composition reduces function. In a rugged landscape there will be many
such optima and understanding the structure (community composition) to function map
at one peak will not in general be informative as to the structure-function map at another
peak. This means that those genotypes whose changing relative abundances have the
greatest impact on function can and will be distinct from one local optimum to another. In
principle any directed evolution algorithm may thus get stuck on a sub-optimal community
and fail to find the optimal configuration of genotypes.

Learning the landscape: A complementary approach to directed evolution for exploring

the structure-function landscape is to attempt to learn the landscape via either regression
or more sophisticated machine learning methods. In this approach one collects data on a
large number of communities comprised of diverse genotypes and measures the function
of interest. Learning the landscape then amounts to performing a regression with the
following form: yi = F(xgi) where F is a proposed functional form stipulated by the
regression being used (e.g. linear model, Random Forest) and yi is the measured function
(degradation rate, pathogen inhibition etc) for community with composition xgi. Such an
approach differs from a directed evolution approach because it posits a specific functional

16
form for the structure-function landscape. This statistical approach faces the challenges
of any inference problem, including overfitting and model misspecification.

Just as with the directed evolution approach, in a situation where the landscape is
exceedingly rugged the regression approach will face challenges because the
contributions of each genotype to the function may depend strongly on the community
composition. In this scenario, any local optimum may be well approximated by a model,
but this model may dramatically fail to predict function (Otwinowski and Plotkin, 2014) in
the neighborhood of a different local optimum where the impact of adding or removing a
given genotype may be very different and where the model has not been trained. Consider
as an example- a set of species with a modest number of 50 genotypes. The full space
of all possible communities comprising these genotypes is 250 or 1015 possible
communities. If a space of this size is truly rugged and contains many local optima,
learning the structure-function map would require enumerating each optimum and the
genotypes that impact function around it, one by one. Even for 50 genotypes, this is a
daunting task which may be feasible in theory but in practice it is prohibitive, even
computationally. It is therefore crucial to ask what controls the ruggedness of these
landscapes and what is known about how rugged they might be.

The navigability of structure-function landscapes may be connected with global

functional effects. In simple models of landscapes, such as the well-known Kaufmann
NK-mode (Kauffman and Weinberger, 1989; du Plessis et al., 2016), the frequency of
random epistatic (non-additive) interactions determines the ruggedness, with increasing
epistasis driving more rugged landscapes. Critically, epistasis in the NK-model is random,
with any site in a genotype equally likely to have an epistatic interaction with any other
site. In the community structure-function context, high levels of epistasis would be
analogous to many random, strong interactions between genotypes that impact function
non-additively. Given the small handful of cases where a structure-function landscape
has been enumerated, we simply cannot say yet if this type of epistasis is prevalent in
community structure-function landscapes. This remains an important open question that
should be addressed in future work.

17
However, recent studies on landscapes in proteins have revealed that ruggedness is not
a necessary outcome of many strong epistatic interactions. Instead, some proteins have
strong epistasis and smooth landscapes. How can this be? In proteins this occurs when
a single “soft mode” dominates the physical dynamics of the system (Husain and
Murugan, 2020). To understand what this means consider the normal modes of a protein,
i.e. the coherent motions of all atoms in the protein in response to a perturbation. These
modes, or oscillations, have different stiffness which dictates how they respond when the
system is perturbed. We can think of a soft mode as a specific set of coherent motions of
all atoms in a protein that are soft – in this case any perturbation to the protein causes
the system to excite that mode.

Experimental studies of proteins with soft mechanical modes have shown that mutations
cause physical deformations along that soft mode (Husain and Murugan, 2020; Leo-
Macias et al., 2005). In essence, the protein can respond to any perturbation, be it
physical or mutational, in only one way – along the soft mode. In the limit of small
perturbations, any two perturbations simply add up to nudge the system along the soft
mode. Thus, mutations are roughly additive in their impact on the physical locations of
atoms in the protein. Epistasis is defined not in terms of the physical deformation of the
protein, but instead as the impact of pairs of mutations on a function such as the catalytic
activity or thermal stability. Both of these are complex functions of the physical locations
of all atoms, so even though the impact of each mutation on physical locations is roughly
additive, their impacts on thermal stability or catalysis are epistatic (Otwinowski et al.,
2018). However, and this is crucial, when a system possesses a soft mode, this strongly
constrains the epistatic interactions between mutations in the system because the
impacts of mutations are highly correlated (Husain and Murugan, 2020). Remarkably, the
very same logic applies to gene regulatory networks. In this case, a network with a soft
mode responds to diverse perturbations with a common change in the pattern of gene
expression. In essence, the response of the regulatory network is constrained to be low-
dimensional. Low-dimensional landscapes present in systems with soft modes are less
rugged and facilitate more rapid evolution that does not get trapped in local optima.

18
Returning to community-function landscapes in microbial communities, if the functional
interactions between genotypes are random then we expect the landscape will be hard
to navigate and directed evolution or landscape learning methods will face challenges.
However, what if the community structure-function landscape possesses a soft mode as
described above? In the community context, what would this entail? One analogy to the
protein example above could be to consider the abundances of genotypes as analogs to
physical locations of atoms in the protein. In this case, a soft mode would manifest as a
coherent variation in abundances along, for example, a single dominant principle
component. Perturbations to the community would then be constrained to drive
abundance dynamics primarily along that mode. We note that such modes of variation
have been observed in simple communities of a few species (Frentz et al., 2015; Hekstra
and Leibler, 2012), and more recently also in host associate microbiomes (Raman et al.,
2019). In analogy to protein function, community function can and often is a non-linear
function of abundances. In this case, the pattern of epistatic interactions between
genotypes will be non-random and constrained. In this situation we could expect a
structure-function landscape that is not rugged but instead smooth, potentially learnable
via regression, and navigable by directed evolution.

We stress that the above sketch of how the theory of fitness landscapes in proteins or
gene regulatory networks might map to communities is at present speculative. Our goal
here is to propose plausible scenarios for what might control the ruggedness of these
landscapes given the many insights provided by fitness landscape theory applied to
proteins, gene circuits, and other biological systems defined at lower levels of biological
organization.

Does community composition uniquely determine community function? Before we

end, we would like to address what may appear to be the proverbial elephant in the room.
While we hope we have convinced the reader that learning the map between community
composition and function may have a transformative impact in our ability to understand
and engineer microbial consortia, it may not be immediately obvious that such a map will

19
necessarily always exist. To what extent does ecological function measured at a given
time depend on the composition of a community at that same time? This question is more
nuanced than it might appear at first sight. For instance, an important function of microbial
consortia is the production of extracellular molecules, from metabolites to secreted
enzymes. The change in concentration of these secreted molecules depends on the rate
at which they are produced, which indeed depends on the abundances of different
members of the consortium as well as on their respective per-capita production rates (Fig.
1). However, the concentration of secreted molecules also depends on the rates of
molecular degradation, biochemical inactivation, diffusion out of the volume or area of
interest, and other degradative processes which eliminate the target molecule and which
do not necessarily depend on the current state of the community. This creates conditions
for which the current state of the function of a community depends not just on its current
composition, but rather on the history of assembly. This idea is perhaps best illustrated
through a simple mathematical model.

We can formally model the rate of accumulation of an extracellular molecule (say, an

enzyme E) in a volume of interest as:

dE/dt=h(z,xg)-𝜆(z,E)

Where h(.) represents the rate of enzyme secretion as a function of the collection of
environmental parameters z and the present species abundance vector xg, and 𝜆(.)
represents the net rate of enzyme loss through all possible pathways. The latter should
depend on the enzyme concentration E as well as on the environmental parameters
captured in z (which may include the concentration of proteases, enzyme inhibitors, or
other environmental parameters affecting the stability of the enzyme such as the pH). Of
course, the environment and genotypes obey their own equations, which are an extension
to higher dimensions of those introduced by Lewontin:

dxg/dt = r(z,xg)
dz/dt = k(z,xg)

20
Where r(.) and k(.) denote the dynamical equations governing the temporal evolution of
xg and z, respectively. In general, there is no reason to expect that, if we integrate those
equations, we should find that E is an explicit or even implicit mathematical function of xg
alone, or even a function of xg and z. This reflects the fact, which should be true for many
community-level traits, that the function of consortium at a given time is not, in general,
uniquely defined by its composition at that time. Rather, it should be a result of the
particular dynamical process of community assembly (i.e. the assembly history) that has
led the community to its current compositional state and, similarly, of the dynamical
history of the environmental parameters captured in z.

Fig. 4. Conditions for the existence of an ecological structure-function map. For illustrative purposes,
we use as an example a hypothetical case where we stabilize a community from a diverse initial pool of
species through periodic transfers in the laboratory, and we measure an enzymatic function akin to the
amylolytic activity discussed as an example in the main text (see also Fig. 1). The barplot shows the
composition of the community at each transfer. The structure function map will exist in three scenarios: 1)
if we map initial composition to function, assuming that the ecological dynamics are reproducible; 2) if we
map final composition to function in a dynamically stable community, e.g. transfers 6-10; and 3) if we map

21
final composition to function in an unstable community, but only if the functional dynamics (enzyme
concentration E and environmental parameters affecting its activity z, see main text) are fast compared to
population dynamics.

Does that mean that a function that uniquely maps community structure to function does,
in general, not exist? It seems to follow from the above argument that, in general, it does
not. Yet, there exist many important limits and cases of practical utility for which the
function of a community at a given point in time can indeed be uniquely defined by its
composition at that time. To illustrate these important scenarios, let us go back to the
example given above, where the function of interest is the concentration of a target
extracellular enzyme. For a structure-function landscape to be well-defined in this case,
there should exist a function E(xg) that provides a 1:1 map between the concentration of
secreted enzyme at a given time and the community composition at that time. One limit
where this function exists occurs when the dynamics of E and z are very fast compared
with the (population) dynamics of xg. In this limit, xg is approximately constant in the
timescale required for E and z to equilibrate, and therefore E (and z) will find a local
equilibrium for every value of xg before this changes significantly. Without loss of
generality, let us consider the simple case where 𝜆(E,z)=𝜆(z)E. In the separation of
timescales limit, we find that the form of the structure-function landscape is
E(xg)=h(z*,xg)/𝜆(z*), where the relevant environmental variables captured in z also
equilibrate rapidly, generally (though not necessarily) reaching a unique value (z*) for
each xg. In this case, and save for special circumstances such as when there exist
memory effects or hysteresis in the per-cell contribution to function, causing non-
linearities in k(xg,z*), every xg may be associated with a unique value of E.

Although separation of timescales is a rather stringent limit that applies only to a narrow
range of real-life scenarios, it brings up a larger point: that although the structure-function
landscape is not defined in general, it may exist when communities are in steady-state.
For many biotechnological applications, communities may be maintained in (or close to)
steady-state by either placing them in a continuous culture device or through serial
passaging (Fig. 4). In chemostats, both species composition and all environmental

22
parameters should reach steady state. Going back to the example discussed above, the
concentration of our target enzyme E should be independent of assembly history and
uniquely linked to the equilibrium concentration of xg (save for the hysteretic situation
discussed in the previous paragraph). In the case of serially passaged consortia,
empirical communities have been generally found to converge to a state of “generational
stability” (Doulcier et al., 2020; Xie et al., 2019), at least when the passaging is done
under constant conditions (Chang et al., 2021b; Estrela et al., 2021a, 2021c; Goldford et
al., 2018; Sánchez et al., 2021).

Another situation of interest in biotechnology is single-batch synthetic communities.

These can be formed by co-inoculating multiple community members at defined initial
abundances in a bioreactor. This consortium is then incubated for a given time period, at
the end of which the function of interest is measured. Here, the requirement for having a
well-defined structure-function landscape is that the population dynamics of the
consortium within the batch are highly reproducible and converge deterministically to the
same final community state at the time of harvest. In this case, the entire within-batch
dynamics including both environmental and species abundance variables are uniquely
determined by the starting abundances of the members of the consortium. Thus, each
initial community state xg will be characterized by a single value of the function (i.e. E) at
the time of harvest, which defines a 1:1 map between both (Fig. 4). Beyond any specific
assumptions regarding the model above there is also empirical evidence for this last
scenario. The primary evidence for this is the remarkable reproducibility and determinism
of community structure and dynamics during community assembly. For example, a
reproducible succession of three functional guilds reliably occurs on polysaccharide
particles in marine communities (Enke et al., 2019). This suggests that given a specific
niche to colonize, and a sufficiently diverse regional species pool, the structure of the
assembled community is reproducible. This empirical observation suggests, but does not
prove, that there are convergent ecological solutions to well-defined functional problems
– degrading polysaccharides in this case. Similar results are observed in glucose and
other small molecule enrichments (Estrela et al., 2021a; Goldford et al., 2018) and
detailed more broadly in surveys of the functional classes of bacteria in the marine

23
microbiome (Louca et al., 2016, 2018). Likewise, host associated communities also
exhibit highly conserved metagenomic structure from host to host (Louca et al., 2016,
2018), suggesting that the functional landscape is a well-defined object with structure
being tightly and reliably linked to function. What remains is to learn this mapping
quantitatively and to leverage that knowledge to design and predict community behavior.

Conclusion. It should be obvious to the reader that we are merely scratching the surface
of a very rich and we believe potentially rather fruitful line of inquiry. Parallelisms between
exploration of fitness landscapes in evolutionary engineering and the exploration of
structure-function landscapes may provide important insights to our understanding of the
mapping between community composition and function, and our ability to engineer
microbial consortia. The field of quantitative genetics has built powerful methodologies to
reconstruct and navigate genotype-phenotype maps, and it also has developed a strong
conceptual and theoretical framework to understand the origins of these genetic
landscapes. Extending these methods and ideas from quantitative genetics and computer
science into microbial ecology could radically improve our ability to understand and
engineer the function of microbial communities. We shall be most satisfied if this review
contributes to stimulating some of these efforts.

References

Aguilar-Rodríguez, J., Payne, J.L., and Wagner, A. (2017). A thousand empirical

adaptive landscapes and their navigability. Nature Ecology & Evolution 1, 0045. .

Alba, V., Carthew, J.E., Carthew, R.W., and Mani, M. (2021). Global constraints within
the developmental program of the Drosophila wing. Elife 10.
https://doi.org/10.7554/eLife.66750.

Alper, H., and Stephanopoulos, G. (2009). Engineering for biofuels: exploiting innate
microbial capacity or importing biosynthetic potential? Nat. Rev. Microbiol. 7, 715–723. .

Arnold, F.H. (2019). Innovation by Evolution: Bringing New Chemistry to Life (Nobel
Lecture). Angew. Chem. Int. Ed Engl. 58, 14420–14426. .

Arora, J., Mars Brisbin, M.A., and Mikheyev, A.S. (2020). Effects of microbial evolution
dominate those of experimental host-mediated indirect selection. PeerJ 8, e9350. .

24
Baas, P., Bell, C., Mancini, L.M., Lee, M.N., Conant, R.T., and Wallenstein, M.D. (2016).
Phosphorus mobilizing consortium Mammoth P(TM) enhances plant growth. PeerJ 4,
e2121. .

Baas, P., Bell, C., Mancini, L., Lee, M., Wallenstein, M.D., and Conant, R.T. (2020). In
vitro selection of a microbial consortium predictive of synergistic functioning along
multiple ecosystem scales.

Baranwal, M., Clark, R.L., Thompson, J., Sun, Z., Hero, A.O., and Venturelli, O.S.
(2022). Recurrent neural networks enable design of multifunctional synthetic human gut
microbiome dynamics. Elife 11. https://doi.org/10.7554/eLife.73870.

Belda, I., Zarraonaindia, I., Perisin, M., Palacios, A., and Acedo, A. (2017). From
Vineyard Soil to Wine Fermentation: Microbiome Approximations to Explain the “terroir”
Concept. Front. Microbiol. 8, 821. .

Beppler Casey, Tekin Elif, Mao Zhiyuan, White Cynthia, McDiarmid Cassandra, Vargas
Emily, Miller Jeffrey H., Savage Van M., and Yeh Pamela J. (2016). Uncovering
emergent interactions in three-way combinations of stressors. J. R. Soc. Interface 13,
20160800. .

Berman, G.J., Choi, D.M., Bialek, W., and Shaevitz, J.W. (2013). Mapping the
stereotyped behaviour of freely-moving fruit flies.

Bittleston, L.S., Gralka, M., Leventhal, G.E., Mizrahi, I., and Cordero, O.X. (2020).
Context-dependent dynamics lead to the assembly of functionally distinct microbial
communities. Nat. Commun. 11, 1440. .

Blasche, S., Kim, Y., Mars, R.A.T., Machado, D., Maansson, M., Kafkia, E., Milanese,
A., Zeller, G., Teusink, B., Nielsen, J., et al. (2021). Metabolic cooperation and
spatiotemporal niche partitioning in a kefir microbial community. Nat Microbiol 6, 196–
208. .

Bloom, J.D., and Arnold, F.H. (2009). In the light of directed evolution: pathways of
adaptive protein evolution. Proc. Natl. Acad. Sci. U. S. A. 106 Suppl 1, 9995–10000. .

Blouin, M., Karimi, B., Mathieu, J., and Lerch, T.Z. (2015). Levels and limits in artificial
selection of communities. Ecol. Lett. 18, 1040–1048. .

Bornscheuer, U.T., Huisman, G.W., Kazlauskas, R.J., Lutz, S., Moore, J.C., and
Robins, K. (2012). Engineering the third wave of biocatalysis. Nature 485, 185–194. .

Chang, C.-Y., Vila, J.C.C., Bender, M., Li, R., Mankowski, M.C., Bassette, M., Borden,
J., Golfier, S., Sanchez, P.G.L., Waymack, R., et al. (2021a). Engineering complex
communities by directed evolution. Nat Ecol Evol https://doi.org/10.1038/s41559-021-
01457-5.

Chang, C.-Y., Vila, J.C.C., Bender, M., Li, R., Mankowski, M.C., Bassette, M., Borden,

25
J., Golfier, S., Sanchez, P.G.L., Waymack, R., et al. (2021b). Engineering complex
communities by directed evolution. Nature Ecology & Evolution 1–13. .

Chen, Y., Lin, C.-J., Jones, G., Fu, S., and Zhan, H. (2009). Enhancing biodegradation
of wastewater by microbial consortia with fractional factorial design. J. Hazard. Mater.
171, 948–953. .

Chica, R.A., Doucet, N., and Pelletier, J.N. (2005). Semi-rational approaches to
engineering enzyme activity: combining the benefits of directed evolution and rational
design. Curr. Opin. Biotechnol. 16, 378–384. .

Clark, R.L., Connors, B.M., Stevenson, D.M., Hromada, S.E., Hamilton, J.J., Amador-
Noguez, D., and Venturelli, O.S. (2021). Design of synthetic human gut microbiome
assembly and butyrate production. Nat. Commun. 12, 3254. .

Dal Bello, M., Lee, H., Goyal, A., and Gore, J. (2021). Resource-diversity relationships
in bacterial communities reflect the network structure of microbial metabolism. Nat Ecol
Evol 5, 1424–1434. .

Díaz-Colunga, J., Skwara, A., Vila, J.C.C., Bajić, D., and Sánchez, Á. (2022). Emergent
ecosystem functions follow simple quantitative rules.

Doulcier, G., Lambert, A., De Monte, S., and Rainey, P.B. (2020). Eco-evolutionary
dynamics of nested Darwinian populations and the emergence of community-level
heredity. Elife 53433. .

Eble, H., Joswig, M., Lamberti, L., and Ludington, W.B. (2021). High dimensional
geometry of fitness landscapes identifies master regulators of evolution and the
microbiome.

Eng, A., and Borenstein, E. (2019). Microbial community design: methods, applications,
and opportunities. Curr. Opin. Biotechnol. 58, 117–128. .

Enke, T.N., Datta, M.S., Schwartzman, J., Cermak, N., Schmitz, D., Barrere, J.,
Pascual-García, A., and Cordero, O.X. (2019). Modular Assembly of Polysaccharide-
Degrading Marine Microbial Communities. Curr. Biol. 29, 1528–1535.e6. .

Ergal, İ., Gräf, O., Hasibar, B., Steiner, M., Vukotić, S., Bochmann, G., Fuchs, W., and
Rittmann, S.K.-M.R. (2020). Biohydrogen production beyond the Thauer limit by
precision design of artificial microbial consortia. Commun Biol 3, 443. .

Erkus, O., de Jager, V.C.L., Spus, M., van Alen-Boerrigter, I.J., van Rijswijck, I.M.H.,
Hazelwood, L., Janssen, P.W.M., van Hijum, S.A.F.T., Kleerebezem, M., and Smid, E.J.
(2013). Multifactorial diversity sustains microbial community stability. ISME J. 7, 2126–
2136. .

Estrela, S., Vila, J.C.C., Lu, N., Bajic, D., Rebolleda-Gomez, M., Chang, C.-Y., Goldford,
J.E., Sanchez-Gorostiaga, A., and Sanchez, A. (2021a). Functional attractors in

26
microbial community assembly. Cell Systems In Press. .

Estrela, S., Sanchez-Gorostiaga, A., Vila, J.C., and Sanchez, A. (2021b). Nutrient
dominance governs the assembly of microbial communities in mixed nutrient
environments. Elife 10, 2020.08.06.239897. .

Estrela, S., Sánchez, Á., and Rebolleda-Gómez, M. (2021c). Multi-Replicated

Enrichment Communities as a Model System in Microbial Ecology. Front. Microbiol. 12,
760. .

Frentz, Z., Kuehn, S., and Leibler, S. (2015). Strongly Deterministic Population
Dynamics in Closed Microbial Communities. Phys. Rev. X 5, 041014. .

George, A.B., and Korolev, K.S. (2021). Ecological landscapes guide the assembly of
optimal microbial communities.

Ghedini, G., Marshall, D.J., and Loreau, M. (2022). Phytoplankton diversity affects
biomass and energy production differently during community development. Functional
Ecology 36, 446–457. https://doi.org/10.1111/1365-2435.13955.

Goldford, J.E., Lu, N., Bajić, D., Estrela, S., Tikhonov, M., Sanchez-Gorostiaga, A.,
Segrè, D., Mehta, P., and Sanchez, A. (2018). Emergent simplicity in microbial
community assembly. Science 361, 469–474. .

Good, B.H., McDonald, M.J., Barrick, J.E., Lenski, R.E., and Desai, M.M. (2017). The
dynamics of molecular evolution over 60,000 generations. Nature
https://doi.org/10.1038/nature24287.

Gopalakrishnappa, C., Gowda, K., Prabhakara, K.H., and Kuehn, S. (2022). An

ensemble approach to the structure-function problem in microbial communities.
iScience 25, 103761. .

Gould, A.L., Zhang, V., Lamberti, L., Jones, E.W., Obadia, B., Korasidis, N.,
Gavryushkin, A., Carlson, J.M., Beerenwinkel, N., and Ludington, W.B. (2018).
Microbiome interactions shape host fitness. Proc. Natl. Acad. Sci. U. S. A. 115,
E11951–E11960. .

Gowda, K., Ping, D., Mani, M., and Kuehn, S. (2022). Genomic structure predicts
metabolite dynamics in microbial communities. Cell 0.
https://doi.org/10.1016/j.cell.2021.12.036.

Guo, X., and Boedicker, J. (2016). High-Order Interactions between Species Strongly
Influence the Activity of Microbial Communities. Biophys. J. 110, 143a. .

Halabi, N., Rivoire, O., Leibler, S., and Ranganathan, R. (2009). Protein sectors:
evolutionary units of three-dimensional structure. Cell 138, 774–786. .

Hekstra, D.R., and Leibler, S. (2012). Contingency and statistical laws in replicate

27
microbial closed ecosystems. Cell 149, 1164–1173. .

Hu, J., Xue, Y., Guo, H., Gao, M.-T., Li, J., Zhang, S., and Tsang, Y.F. (2017). Design
and composition of synthetic fungal-bacterial microbial consortia that improve
lignocellulolytic enzyme activity. Bioresour. Technol. 227, 247–255. .

Husain, K., and Murugan, A. (2020). Physical constraints on epistasis. Mol. Biol. Evol.
https://doi.org/10.1093/molbev/msaa124.

Jiang, Y., Dong, W., Xin, F., and Jiang, M. (2020). Designing Synthetic Microbial
Consortia for Biofuel Production. Trends Biotechnol. 0.
https://doi.org/10.1016/j.tibtech.2020.02.002.

Jordan, D., Kuehn, S., Katifori, E., and Leibler, S. (2013). Behavioral diversity in
microbes and low-dimensional phenotypic spaces. Proc. Natl. Acad. Sci. U. S. A. 110,
14018–14023. .

Kauffman, S.A., and Weinberger, E.D. (1989). The NK model of rugged fitness
landscapes and its application to maturation of the immune response. J. Theor. Biol.
141, 211–245. .

Khan, A.I., Dinh, D.M., Schneider, D., Lenski, R.E., and Cooper, T.F. (2011). Negative
Epistasis Between Beneficial Mutations in an Evolving Bacterial Population. Science
332, 1193–1196. https://doi.org/10.1126/science.1203801.

Kim, I.J., Lee, H.J., Choi, I.-G., and Kim, K.H. (2014). Synergistic proteins for the
enhanced enzymatic hydrolysis of cellulose by cellulase. Appl. Microbiol. Biotechnol. 98,
8469–8480. .

Kinnersley, M.A., Holben, W.E., and Rosenzweig, F. (2009). E Unibus Plurum: genomic
analysis of an experimentally evolved polymorphism in Escherichia coli. PLoS Genet. 5,
e1000713. .

Kuchner, O., and Arnold, F.H. (1997). Directed evolution of enzyme catalysts. Trends
Biotechnol. 15, 523–530. .

Kuebbing, S.E., Classen, A.T., Sanders, N.J., and Simberloff, D. (2015). Above- and
below-ground effects of plant diversity depend on species origin: an experimental test
with multiple invaders. New Phytol. 208, 727–735. .

Langenheder, S., Bulling, M.T., Solan, M., and Prosser, J.I. (2010). Bacterial
biodiversity-ecosystem functioning relations are modified by environmental complexity.
PLoS One 5, e10834. .

Leo-Macias, A., Lopez-Romero, P., Lupyan, D., Zerbino, D., and Ortiz, A.R. (2005). An
analysis of core deformations in protein superfamilies. Biophys. J. 88, 1291–1299. .

Louca, S., Jacques, S.M.S., Pires, A.P.F., Leal, J.S., Srivastava, D.S., Parfrey, L.W.,

28
Farjalla, V.F., and Doebeli, M. (2016). High taxonomic variability despite stable
functional structure across microbial communities. Nat Ecol Evol 1, 15. .

Louca, S., Polz, M.F., Mazel, F., Albright, M.B.N., Huber, J.A., O’Connor, M.I.,
Ackermann, M., Hahn, A.S., Srivastava, D.S., Crowe, S.A., et al. (2018). Function and
functional redundancy in microbial systems. Nature Ecology & Evolution
https://doi.org/10.1038/s41559-018-0519-1.

Lu, H., Diaz, D.J., Czarnecki, N.J., Zhu, C., Kim, W., Shroff, R., Acosta, D.J., Alexander,
B.R., Cole, H.O., Zhang, Y., et al. (2022). Machine learning-aided engineering of
hydrolases for PET depolymerization. Nature 604, 662–667. .

Macia, J., Manzoni, R., Conde, N., Urrios, A., de Nadal, E., Solé, R., and Posas, F.
(2016). Implementation of Complex Biological Logic Circuits Using Spatially Distributed
Multicellular Consortia. PLoS Comput. Biol. 12, e1004685. .

Maleki, N., Safari, M., and Eiteman, M.A. (2018). Conversion of glucose-xylose mixtures
to pyruvate using a consortium of metabolically engineered Escherichia coli. Eng. Life
Sci. 18, 40–47. .

Mancuso, C.P., Lee, H., Abreu, C.I., Gore, J., and Khalil, A.S. (2021). Environmental
fluctuations reshape an unexpected diversity-disturbance relationship in a microbial
community. Elife 10, 2020.07.28.225987. .

May, A., Narayanan, S., Alcock, J., Varsani, A., Maley, C., and Aktipis, A. (2019).
Kombucha: a novel model system for cooperation and conflict in a complex multi-
species microbial ecosystem. PeerJ 7, e7565. .

Mickalide, H., and Kuehn, S. (2019). Higher-Order Interaction between Species Inhibits
Bacterial Invasion of a Phototroph-Predator Microbial Community. Cell Syst 9, 521–
533.e10. .

Minty, J.J., Singer, M.E., Scholz, S.A., Bae, C.-H., Ahn, J.-H., Foster, C.E., Liao, J.C.,
and Lin, X.N. (2013). Design and characterization of synthetic fungal-bacterial consortia
for direct production of isobutanol from cellulosic biomass. Proc. Natl. Acad. Sci. U. S.
A. 110, 14592–14597. .

Morcos, F., Pagnani, A., Lunt, B., Bertolino, A., Marks, D.S., Sander, C., Zecchina, R.,
Onuchic, J.N., Hwa, T., and Weigt, M. (2011). Direct-coupling analysis of residue
coevolution captures native contacts across many protein families. Proc. Natl. Acad.
Sci. U. S. A. 108, E1293–E1301. .

Morin, M.A., Morrison, A.J., Harms, M.J., and Dutton, R.J. (2022). Higher-order
interactions shape microbial interactions as microbial community complexity increases.

Mueller Ulrich G., Juenger Thomas E., Kardish Melissa R., Carlson Alexis L., Burns
Kathleen M., Edwards Joseph A., Smith Chad C., Fang Chi-Chun, Des Marais David L.,
and Shade Ashley Artificial Selection on Microbiomes To Breed Microbiomes That

29
Confer Salt Tolerance to Plants. mSystems 0, e01125–21. .

Nahum, J.R., Godfrey-Smith, P., Harding, B.N., Marcus, J.H., Carlson-Stevermer, J.,
and Kerr, B. (2015). A tortoise-hare pattern seen in adapting structured and
unstructured populations suggests a rugged fitness landscape in bacteria. Proc. Natl.
Acad. Sci. U. S. A. 112, 7530–7535. .

Otwinowski, J., and Plotkin, J.B. (2014). Inferring fitness landscapes by regression
produces biased estimates of epistasis. Proc. Natl. Acad. Sci. U. S. A. 111, E2301–
E2309. .

Otwinowski, J., McCandlish, D.M., and Plotkin, J.B. (2018). Inferring the shape of global
epistasis. Proc. Natl. Acad. Sci. U. S. A. 115, E7550–E7558. .

Piccardi, P., Vessman, B., and Mitri, S. (2019). Toxicity drives facilitation between 4
bacterial species. Proc. Natl. Acad. Sci. U. S. A. 116, 15979–15984. .

du Plessis, L., Leventhal, G.E., and Bonhoeffer, S. (2016). How Good Are Statistical
Models at Approximating Complex Fitness Landscapes? Mol. Biol. Evol. 33, 2454–
2468. .

Poelwijk, F.J., Tănase-Nicola, S., Kiviet, D.J., and Tans, S.J. (2011). Reciprocal sign
epistasis is a necessary condition for multi-peaked fitness landscapes. J. Theor. Biol.
272, 141–144. .

Poelwijk, F.J., Socolich, M., and Ranganathan, R. (2019). Learning the pattern of
epistasis linking genotype and phenotype in a protein. Nat. Commun. 10, 4213. .

Raman, A.S., White, K.I., and Ranganathan, R. (2016). Origins of Allostery and
Evolvability in Proteins: A Case Study. Cell 166, 468–480. .

Raman, A.S., Gehrig, J.L., Venkatesh, S., Chang, H.-W., Hibberd, M.C., Subramanian,
S., Kang, G., Bessong, P.O., Lima, A.A.M., Kosek, M.N., et al. (2019). A sparse
covarying unit that describes healthy and impaired human gut microbiota development.
Science 365. https://doi.org/10.1126/science.aau4735.

Reddy, G., and Desai, M.M. (2021). Global epistasis emerges from a generic model of a
complex trait. Elife 10, 2020.06.14.150946. .

Roell, G.W., Zha, J., Carr, R.R., Koffas, M.A., Fong, S.S., and Tang, Y.J. (2019).
Engineering microbial consortia by division of labor. Microb. Cell Fact. 18, 35. .

Romero, P.A., and Arnold, F.H. (2009). Exploring protein fitness landscapes by directed
evolution. Nat. Rev. Mol. Cell Biol. 10, 866–876. .

Romero, P.A., Krause, A., and Arnold, F.H. (2013). Navigating the protein fitness
landscape with Gaussian processes. Proc. Natl. Acad. Sci. U. S. A. 110, E193–E201. .

30
Rozen, D.E., and Lenski, R.E. (2000). Long-Term Experimental Evolution in Escherichia
coli. VIII. Dynamics of a Balanced Polymorphism. Am. Nat. 155, 24–35. .

Russ, W.P., Figliuzzi, M., Stocker, C., Barrat-Charlaix, P., Socolich, M., Kast, P., Hilvert,
D., Monasson, R., Cocco, S., Weigt, M., et al. (2020). An evolution-based model for
designing chorismate mutase enzymes. Science 369, 440–445. .

Sanchez, A. (2019). Defining Higher-Order Interactions in Synthetic Ecology: Lessons

from Physics and Quantitative Genetics. Cell Syst 9, 519–520. .

Sánchez, Á., Vila, J.C.C., Chang, C.-Y., Diaz-Colunga, J., Estrela, S., and Rebolleda-
Gomez, M. (2021). Directed Evolution of Microbial Communities. Annu. Rev. Biophys.
50, 323–341. .

Sanchez-Gorostiaga, A., Bajić, D., Osborne, M.L., Poyatos, J.F., and Sanchez, A.
(2019). High-order interactions distort the functional landscape of microbial consortia.
PLoS Biol. 17, e3000550. .

Senay, Y., John, G., Knutie, S.A., and Brandon Ogbunugafor, C. (2019). Deconstructing
higher-order interactions in the microbiota: A theoretical examination.

Senne de Oliveira Lino, F., Bajic, D., Vila, J.C.C., Sánchez, A., and Sommer, M.O.A.
(2021). Complex yeast–bacteria interactions affect the yield of industrial ethanol
fermentation. Nat. Commun. 12, 1498. .

Sgobba, E., and Wendisch, V.F. (2020). Synthetic microbial consortia for small
molecule production. Curr. Opin. Biotechnol. 62, 72–79. .

Stiffler, M.A., Hekstra, D.R., and Ranganathan, R. (2015). Evolvability as a function of

purifying selection in TEM-1 β-lactamase. Cell 160, 882–892. .

Swenson, W., Arendt, J., and Wilson, D.S. (2000a). Artificial selection of microbial
ecosystems for 3-chloroaniline biodegradation. Environ. Microbiol. 2, 564–571. .

Swenson, W., Wilson, D.S., and Elias, R. (2000b). Artificial ecosystem selection. Proc.
Natl. Acad. Sci. U. S. A. 97, 9110–9114. .

Tareen, A., Kooshkbaghi, M., Posfai, A., Ireland, W.T., McCandlish, D.M., and Kinney,
J.B. (2022). MAVE-NN: learning genotype-phenotype maps from multiplex assays of
variant effect. Genome Biol. 23, 98. .

Tekin, E., Yeh, P.J., and Savage, V.M. (2018). General Form for Interaction Measures
and Framework for Deriving Higher-Order Emergent Effects. Frontiers in Ecology and
Evolution 6, 166. .

Tekin Elif, Beppler Casey, White Cynthia, Mao Zhiyuan, Savage Van M., and Yeh
Pamela J. (2016). Enhanced identification of synergistic and antagonistic emergent
interactions among three or more drugs. J. R. Soc. Interface 13, 20160332. .

31
Thommes, M., Wang, T., Zhao, Q., Paschalidis, I.C., and Segrè, D. (2019). Designing
Metabolic Division of Labor in Microbial Communities. mSystems 4.
https://doi.org/10.1128/mSystems.00263-18.

Tonner, P.D., Pressman, A., and Ross, D. (2021). Interpretable modeling of genotype-
phenotype landscapes with state-of-the-art predictive power.

Tracewell, C.A., and Arnold, F.H. (2009). Directed enzyme evolution: climbing fitness
peaks one amino acid at a time. Curr. Opin. Chem. Biol. 13, 3–9. .

Wagner, A. (2019). Life Finds a Way: What Evolution Teaches Us About Creativity
(New York: Basic Books.).

Wei, X., and Zhang, J. (2019). Patterns and Mechanisms of Diminishing Returns from
Beneficial Mutations. Mol. Biol. Evol. 36, 1008–1021. .

Wendisch, V.F., Bott, M., and Eikmanns, B.J. (2006). Metabolic engineering of
Escherichia coli and Corynebacterium glutamicum for biotechnological production of
organic acids and amino acids. Curr. Opin. Microbiol. 9, 268–274. .

Weng, J.-K., Li, X., Bonawitz, N.D., and Chapple, C. (2008). Emerging strategies of
lignin engineering and degradation for cellulosic biofuel production. Curr. Opin.
Biotechnol. 19, 166–172. .

Wittmann, B.J., Yue, Y., and Arnold, F.H. (2021). Informed training set design enables
efficient machine learning-assisted directed protein evolution. Cell Syst 12, 1026–
1045.e7. .

Wolfe, B.E., Button, J.E., Santarelli, M., and Dutton, R.J. (2014). Cheese rind
communities provide tractable systems for in situ and in vitro studies of microbial
diversity. Cell 158, 422–433. .

Wood, K., Nishida, S., Sontag, E.D., and Cluzel, P. (2012). Mechanism-independent
method for predicting response to multidrug combinations in bacteria. Proc. Natl. Acad.
Sci. U. S. A. 109, 12254–12259. .

Wright, R.J., Gibson, M.I., and Christie-Oleza, J.A. (2019). Understanding microbial
community dynamics to improve optimal microbiome selection. Microbiome 7, 85. .

Xie, L., and Shou, W. (2021). Steering ecological-evolutionary dynamics to improve

artificial selection of microbial communities. Nat. Commun. 12, 6799. .

Xie, L., Yuan, A.E., and Shou, W. (2019). Simulations reveal challenges to artificial
community selection and possible strategies for success. PLoS Biol. 17, e3000295. .

Yang, D., Park, S.Y., and Lee, S.Y. (2021). Production of Rainbow Colorants by
Metabolically Engineered Escherichia coli. Adv. Sci. 8, e2100743. .

32
Zanaroli, G., Di Toro, S., Todaro, D., Varese, G.C., Bertolotto, A., and Fava, F. (2010).
Characterization of two diesel fuel degrading microbial consortia enriched from a non
acclimated, complex source of microorganisms. Microb. Cell Fact. 9, 10. .

33