The Plant Journal (2019) 97, 164–181 doi: 10.1111/tpj.

14170

SI GENOME TO PHENOME

Genome-wide association studies on the phyllosphere

microbiome: Embracing complexity in host–microbe
interactions
Kathleen Beilsmith1,†, Manus P.M. Thoen1,†, Benjamin Brachi2, Andrew D. Gloss1, Mohammad H. Khan1 and
Joy Bergelson1,*
1
Department of Ecology and Evolution, University of Chicago, 1101 E 57th St, Chicago, IL 60637, USA, and
2
BIOGECO, INRA, University of Bordeaux, 33610 Cestas, France

Received 25 July 2018; revised 8 November 2018; accepted 16 November 2018; published online 22 November 2018.
*For correspondence (e-mail jbergels@uchicago.edu).
†
These authors contributed equally to this study.

SUMMARY
Environmental sequencing shows that plants harbor complex communities of microbes that vary across
environments. However, many approaches for mapping plant genetic variation to microbe-related traits
were developed in the relatively simple context of binary host–microbe interactions under controlled condi-
tions. Recent advances in sequencing and statistics make genome-wide association studies (GWAS) an
increasingly promising approach for identifying the plant genetic variation associated with microbes in a
community context. This review discusses early efforts on GWAS of the plant phyllosphere microbiome and
the outlook for future studies based on human microbiome GWAS. A workflow for GWAS of the phyllo-
sphere microbiome is then presented, with particular attention to how perspectives on the mechanisms,
evolution and environmental dependence of plant–microbe interactions will influence the choice of traits to
be mapped.

Keywords: genome-wide association studies, microbiome, phyllosphere, host–microbe interactions,

community, environment, sequencing, phenotype, genotype, mapping.

INTRODUCTION have been tremendously fruitful, the advent of high-

Plants and microbes together make up the vast majority of
terrestrial biomass on Earth (Bar-On et al., 2018). It is hardly
surprising, then, that plant phenotypes are inextricably
shaped by their interactions with microbes. Some microbes
cause disease in plants (Mansfield et al., 2012), while others
offer protection against these microbial pathogens and
non-microbial enemies (Innerebner et al., 2011); still others
help plants utilize their abiotic environment through pro-
cesses such as nutrient acquisition (Haney et al., 2015).
Efforts to characterize the molecular biology, ecology and
evolution of these interactions have revealed that they vary
among plant species and genotypes, and thus have a
genetic basis (Smith and Goodman, 1999; Corwin and
Kliebenstein, 2017). This has largely been shown through
studies on interactions of plants with one or a few microbial
strains at a time, often in controlled and simplified environ-
ments in the laboratory or greenhouse. While these efforts
throughput environmental sequencing has led to the dis-
covery that leaves harbor a complex microbial ecosystem.
From mountain shrubs (Ruiz-Pe
rez et al., 2016) to seagrass
(Fahimipour et al., 2017), and from subarctic grass (Uroz
et al., 2016) to equatorial forest canopies (Lambais et al.,
2006), the photosynthesizing tissues of plants are colonized
by a massive and diverse set of microbial colonists, includ-
ing bacteria, fungi and phages (Lindow and Brandl, 2003;
Rastogi et al., 2013; Morella et al., 2018; Sapp et al., 2018).
Plant biologists and breeders are now confronted with the
challenge of translating the wealth of knowledge gained
from the study of individual plant–microbe interactions to
an understanding of factors shaping the microbiome. This
will require not only disentangling host–microbe from
microbe–microbe interactions, but also integrating the role
of environmental variation.
Genome-wide association studies (GWAS) are a power-
ful and flexible tool for tackling this challenge. GWAS refer

164 © 2018 The Authors

The Plant Journal © 2018 John Wiley & Sons Ltd

to the application of association mapping, a technique that
tests for statistical associations between genetic variation
at a locus and organismal phenotypes, to thousands or
millions of loci throughout the genome (Brachi et al.,
2011). Because GWAS only requires genotype and pheno-
type data across individuals, it can in theory be employed
for any property of the phyllosphere that can be quantified
and is influenced by plant genotype. Promisingly,
quantitative indices of community composition and diver-
sity vary among plant genotypes grown in common envi-
ronments (i.e. they are heritable; Peiffer et al., 2013; Horton
et al., 2014), making them suitable phenotypes to interro-
gate through GWAS. A key strength of GWAS in this con-
text is the flexibility of its underlying statistical models,
which can compare and quantify how the effects of geno-
type on phenotype vary across environmental conditions,
while also controlling for environmental noise that con-
founds associations in the field (Korte et al., 2012; Sasaki
et al., 2015). Recent computational and statistical advances
allow GWAS to gain power through modeling the relation-
ships between traits (Hackinger and Zeggini, 2017), offer-
ing an exciting avenue for leveraging the wealth of
different microbial lineage-level and community-level phe-
notypes that can be extracted from phyllosphere DNA
sequencing.
Here, we begin by reviewing how GWAS that leverage
phenotypes derived from microbial DNA sequencing data
has revealed new insight into the influence of genetic
variation among plants on the composition of the phyllo-
sphere. GWAS has also been used to characterize the
effects of host genetics on the associated microbiomes in
organisms other than plants, and we draw on a leading
system for these studies, the human microbiome, to illus-
trate the promise and challenges awaiting studies of the
plant phyllosphere. Next, along each stage of the work-
flow for GWAS of the phyllosphere, we synthesize emerg-
ing bioinformatic, statistical and conceptual advances that
are poised to enable new breakthroughs, and we empha-
size key considerations for researchers in light of these
advances. This outline of the workflow is intended to help
readers evaluate the feasibility and promise of conducting
microbiome-related GWAS studies in their own plant sys-
tems. We argue that perspectives on how plant–microbe
interactions evolve define the choice of phenotypes to be
measured, and thus may be the most important determi-
nant of the single nucleotide polymorphism (SNP) vari-
ants that GWAS identifies as mediators of host genotypic
effects on the microbiome. In particular, the number and
identity of host loci associated with the microbiome will
depend upon the choice to consider pairwise or commu-
nity-level interactions, on whether the interacting
microbes are characterized by taxonomic grouping or by
traits, and on the extent to which the environmental plas-
ticity of interactions is exposed in the sampling scheme.
© 2018 The Authors
The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
GWAS of the phyllosphere microbiome 165
PREVIOUS WORK
Most GWAS of plant–microbial interactions have focused
on how plant genotype shapes pairwise interactions with a
single microbial taxon. Representative examples can be
found in the 35 GWAS of direct pathogen challenge col-
lated by Bartoli and Roux (2017), or the 13 GWAS of dis-
ease resistance in maize collated by Xiao et al. (2017).
GWAS have also been used to probe the genetic basis of
plant mutualisms with microbes, such as the relationships
between legumes and their nodule-forming bacteria (Stan-
ton-Geddes et al., 2013; Curtin et al., 2017).
In contrast, only a single GWAS of the plant phyllo-
sphere community has been published. Horton et al.
(2014) grew a panel of 196 Arabidopsis thaliana genotypes
in the field, and characterized the leaf microbial commu-
nity by sequencing taxonomic marker genes for bacteria
and fungi. They found that additive genetic variation in the
host influenced these leaf microbes at the community
level. For the top eigenvectors produced by ordination
analysis (principal component analysis; PCA), 9% and 11%
of the variance in bacterial and fungal communities,
respectively, were explained by plant SNPs (narrow-sense
heritability; Box 1). However, this result was only obtained
after rare species represented by a small fraction of the
total sequencing reads in each community were excluded.
Without restricting the microbiome profile to the most-
sequenced taxa, heritability was no longer observed for
ordination outputs, and mapping SNPs to community-level
traits proved impossible.
Using these restricted community datasets, Horton et al.
(2014) successfully mapped both community-level traits
and the presence/absence of the most abundant individual
taxa in the leaf microbiome to host SNPs. Eigenvectors
produced by PCA of the fungal community were associ-
ated with SNPs located in candidate genes for cell wall
integrity. Species richness of the bacterial community was
associated with SNPs in a set of genes enriched for tri-
chome-related functions. But the strongest associations in
the entire study were found with the presence/absence of
one or a few taxa. The SNPs in these associations were
located in genes involved in defense, signal transduction,
kinase activity, cell walls and cell membranes. The strength
of these associations, coupled with the necessity of remov-
ing rarely sequenced taxa to observe any heritability, sug-
gests that plant influence may be best observed in the
most frequent phyllosphere colonists rather than compre-
hensive community phenotypes.
Because a larger number of GWAS have been performed
on the human and mouse microbiome, these models offer
a glimpse of the advantages and challenges that await
plant biologists as the library of microbiome-associated
variants grows. GWAS has successfully identified over 200
SNPs associated with variation in human microbiome
166 Kathleen Beilsmith et al.

Box 1 Heritability 6¼ inheritance

Inheritance and heritability are two terms that are often confused, and the potential for vertical transmission of the microbiome makes
it important to clearly distinguish the two concepts (box figure). In quantitative genetics, phenotypic variance is decomposed following
equations:

Vp ¼ Vg þ Ve þ Vge where Vg ¼ Va þ Vd þ Vi

where Vp is the total phenotypic variance (here, for a microbiome-related trait), Vg is the variance among genotypes explained by
genetics, Ve is the variance due to environmental variation, and Vg*e is the variance due to genotype by environment interactions. Vg
can be further decomposed into variance due to additive (Va), dominance (Vd) and epistatic effects (Vi). Broad sense heritability is for-
mally defined as the proportion of phenotypic variance explained by the effects of genes (host genes, in the case of the microbiota)
(Lynch and Walsh, 1998) (H2 = Vg/Vp). Narrow sense heritability, on the other hand, is defined as the proportion of the total phenotypic
variance due to additive effects (h2 = Va/Vp), and is related to the phenotypic response to selection (Falconer, 1981). Following these
definitions, a trait can be genetically determined, and therefore inherited, while not displaying heritability for three main reasons: (i)
there is no variation at the loci underlying the trait in the collection of individuals considered; (ii) the variance due to the environment
is much larger than the variance due to genetics, or (iii) strong gene 9 environment interactions obscure the signal.

In practice, the method to estimate broad sense heritability will depend on the model system. The simplest case concerns clonal or
selfing species, such as A. thaliana, where multiple replicates of the exact same genotype can be included in the same fully random-
ized experimental design. In this situation, broad sense heritability is estimated using mixed models as the ratio of phenotypic vari-
ance among genotypes over the total phenotypic variance and can be termed clonal repeatability. While this situation is powerful for
estimating broad sense heritability, great care should be taken to prevent confounding by non-genetic maternal effects. This is particu-
larly important when the trait of interest is the microbiome, due for example to potential vertical transmission through seeds. Control
of maternal effect is usually achieved by producing the seeds in homogeneous conditions, preferably growing the plants in a random-
ized design. However, if vertical transmission of the microbiome through seeds is important for the future microbiome, parental plants
used to generate experimental material should be grown in the environmental conditions in which the experiment will be performed,
maybe over several generations.
To compute heritability for non-selfing/clonal species, where replication of the exact same genotypes is not possible, one must either
compute trait correlations between parents and offspring, assume genetic similarity based on pedigrees, or compare zygotic and non-
zygotic twin pairs assuming that non-zygotic twins have a genetic similarity of 0.5 and zygotic twins have identical genetics. All these
cases suffer from the challenge that it is often impossible to control for environmental variance in fully randomized experimental
designs and/or to control for maternal effects, making heritability estimates less precise and increasing the need for very large sample
sizes. A recent solution is the computation of SNP-based or pseudo-heritability, which became possible following the rise of high-
throughput genotyping and sequencing. The idea is to estimate the proportion of phenotypic variance explained by the genome-wide
genetic similarity of individuals using mixed models (Yang et al., 2011). While the relationship of SNP-based heritability with broad

© 2018 The Authors

The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
 GWAS of the phyllosphere microbiome 167

sense and narrow sense heritability is unclear and depends on the experimental design, its estimates are directly relevant to GWAS
because they capture the amount of among-genotype variance explained by true genetic variation in the sample considered (and not
assumed from pedigrees or family structure; Lee and Chow, 2014).
Heritability estimates represent a target for the amount of variance explained by underlying loci and potentially detectable in GWAS.
However, high heritability values do not guarantee the successful identification of significant trait/marker association, as this will
depend on the genetic architecture of the trait (see ‘Association Mapping’).

In human microbiome studies to date, host genetics account for only a small portion of inter-individual variation in human micro-
biome composition. Microbiomes are only slightly more similar between identical and non-identical twins (Goodrich et al., 2014), and
a recent re-analysis of these studies estimated SNP-based heritability of the microbiome to be between 2 and 8% (Rothschild et al.,
2018). Environmental variation – including diet, lifestyle and antibiotic use – collectively has a larger effect than host genetics (David
et al., 2014; Rothschild et al., 2018). The possibility of replicated experiments profiling plants raised in one or more shared natural
environment, which can increase the heritability of the microbiome by minimizing environmental variation among individuals, is there-
fore a major advantage to GWAS of the plant phyllosphere. It is also important to note that heritability of the human microbiome is
dependent on the methods used to characterize the microbiome, so heritability might be underestimated by the taxonomic metrics
used to date if host genotype shapes microbiome function more strongly than it shapes taxonomic composition. The heritable compo-
nents of the microbiome may nonetheless have outsized effects on human phenotypes. For example, the most heritable microbial
family in one study (Christensenellaceae) co-varied in abundance with many other microbes; these same microbes altered microbiome
composition and reduced weight gain when transplanted into mice (Goodrich et al., 2014). When many traits are available – as is the
case with high-dimensional phyllosphere datasets – those that are heritable can be prioritized for further analysis. This is sensible
because GWAS is unlikely to be informative for traits with little to no heritability in a given study, and retaining such traits for GWAS
increases the experiment-wide multiple testing burden.

composition (Knights et al., 2014; Blekhman et al., 2015;
Davenport et al., 2015; Bonder et al., 2016; Goodrich et al.,
2016; Turpin et al., 2016; Wang et al., 2016; Demmitt et al.,
2017; Igartua et al., 2017; Kolde et al., 2018; Rothschild
et al., 2018), and a few genomic regions associated with
microbiome composition in mice (Org et al., 2015). The
majority of these associations are with the relative abun-
dances of individual bacterial families or genera, although
some studies report associations with overall community
composition or species richness (Wang et al., 2016). Genes
involved in metabolism and immunity are heavily repre-
sented in regions harboring significant associations.
The majority of associations found in human and mice
GWAS to date are unique to single studies (Wang et al.,
2016). The failure to consistently detect the same associa-
tions across studies may reflect differences in the genetic
ancestry of the mapping populations used in each study,
the environments inhabited by those populations (which
have an outsized effect on microbiome composition;
Box 1), or methods used to sequence and quantify micro-
bial communities. Alternatively, the paucity of shared asso-
ciations may arise because their effects on microbiome
composition are too small to permit consistent detection
with the sample sizes employed, which ranged roughly
from 100 to 1000 individuals. Some of these associations
may also fail to replicate because they are false positives.
Thus, even in humans, our understanding of the relation-
ship between the host genome and the microbiome
remains limited.
Microbiome GWAS results will not always paint a similar
picture of how host genomes influence microbial colonists,
as demonstrated by the challenges of synthesizing results
across mammalian studies of this kind. In order to
understand which parts of a pipeline must be standardized
to allow consistent results to be revealed, and in order to
confidently identify promising candidates for hypothesis
testing and application, plant biologists need to understand
how each step of a microbiome GWAS workflow can be
tuned to sample adequate variation, generate meaningful
phenotypes, and generate sufficient power to detect associ-
ations. Here we provide a workflow for investigating the
genetic bases of leaf microbial community variation among
plant genotypes or cultivars (Figure 1). We discuss the
choices facing investigators at each step, as well as their
tradeoffs.
WORKFLOW
Defining the question
Genome-wide association studies of the leaf microbiome
are exploratory experiments, and results can be difficult to
interpret. We propose that researchers interested in explor-
ing the relationships between host genetic variation and
the leaf microbial community should consider three main
points. First, GWAS provide an avenue to interrogate how
a large catalog of naturally occurring alleles across the
host genome impacts the leaf microbiota. However, this
relationship is complex due to both the genetics of the
host (i.e. millions of markers, many of them with rare alle-
les) and the diversity of the microbial communities (hun-
dreds to thousands of members, many if not all with zero
inflated distributions). Focus on a more geographically
restricted or less genetically diverse collection of plant
genotypes could lighten the statistical burden and sources
of confounding due to population stratification. Second,
concerning the microbial community itself, one might be

© 2018 The Authors

The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
168 Kathleen Beilsmith et al.

© 2018 The Authors

The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
 GWAS of the phyllosphere microbiome 169

Figure 1. How to use genome-wide association studies (GWAS) with your plant leaf microbiome.
1. THE QUESTION: Understanding the leaf microbiome from an evolutionary, mechanistic plant physiological or agricultural perspective requires different
downstream considerations. 2. PLANT GENOTYPING: Selecting the plant panel is an essential step in non-model systems. A plant panel with the appropriate
diversity and population structure must be selected. 3. DESIGN: One can sample microbiomes from plant leaves in their natural habitat, in a common garden
experiment, or in sterile environments like well plates. 4. SAMPLING: Differences in sampling methods often concern surface cleaning treatments and the time
during which leaf samples are stored before DNA is isolated. 5. PHENOTYPING: Microbiota can be quantified with amplicon or metagenomic sequencing, and
converted into relevant phenotypic traits for GWAS. 6. GWAS: Several statistical models exist to perform GWAS on the leaf microbiome of plants and estimate
the significance of associations with microbiome traits at loci across the genome, as depicted in Manhattan plots. 7. VALIDATION: Genes underlying candidate
loci can be functionally validated through post-GWAS statistical approaches, such as bioinformatic methods that use transcriptome datasets and reverse genetic
gene editing methods.

interested in finding host genes that influence microbial
community composition, or genes that enable hosts to fos-
ter specific microbial functions. This will have profound
consequences on the collection, sequencing and analysis
of leaf samples. Finally, if the ultimate goal is to gain a
mechanistic understanding of how the plant influences the
microbiome, it will be necessary to understand causality
between host factors and microbial community members/
functions. Furthermore, determining how functional valida-
tion will be achieved requires planning. For example, func-
tional validation of allelic effects estimated in field
experiments may be difficult to validate under laboratory
conditions or in replicated field experiments due to strong
environmental effects. We develop these points below.
Plant genotyping
Plant mapping panels benefit from having ample natural
variation, although the advantages of including more
diversity are limited by the density of SNP markers in the
genome. GWAS draw their ability to identify associated
loci from historical recombination events (Mitchell-Olds
and Schmitt, 2006; Bergelson and Roux, 2010), which cre-
ate new combinations of alleles at a frequency that
decreases with proximity, or linkage, on the chromosome.
Because nucleotides in close proximity are less likely to be
separated by a recombination event, causal variants can be
identified by the association of a nearby SNP marker with
a phenotype. The number of SNP markers needed to find
associations thus depends on the distance over which link-
age disequilibrium (LD) extends in that portion of the gen-
ome. Patterns of LD differ significantly between plant
species and even within single genomes as a function of
chromosomal location, insertions, repetitive regions and
methylation. In the model plant A. thaliana, LD typically
decays over a distance of 5–10 kb. In an outcrossing spe-
cies like maize, intragenic linkage in a diverse germplasm
decreases to < 0.25 only over 100–200 bp (Tenaillon et al.,
2001). Thus, some plants will require denser SNP data than
others to effectively map microbiome associations. The
amount of LD will also depend on how recently the mem-
bers of a mapping panel shared a common ancestor: gen-
omes in a local population will have experienced fewer
historical recombination events than those in a worldwide
panel. As more diverse genomes are included in a panel
and LD decreases, a denser set of SNPs will be necessary
to ensure that all the relevant variants in the genome fall
within the LD block of a marker. High-density SNP panels
already exist for a variety of plant species (Bayer et al.,
2017; Shirasawa et al., 2017; Eltaher et al., 2018).
Another challenge is the ability to find associations
between the microbiome and structural variation in plant
genomes. Even in A. thaliana, with its notably small gen-
ome, genome size varies by 10% over a fraction of the dis-
tribution range of the species (Long et al., 2013). When
SNPs and small structural variation are called by alignment
to a reference, much of this larger structural variation in
the genome will be lost. De novo alignments of genomes
in plant panels can help to recapture this variation. For
species with repetitive sequences and large rearrange-
ments, increasing sequencing coverage and longer
sequencing reads should help to resolve the issues with
generating such alignments, and allow mapping micro-
biome traits to both SNP and structural variants.
Selection of a GWAS panel also involves a tradeoff
between capturing phenotypic variation of interest and
confounding the analysis with population structure and
allelic heterogeneity (Pritchard et al., 2000; Zhao et al.,
2007). A good illustration is the case of FRIGIDA, a major
flowering time gene in A. thaliana. This gene displays
extensive allelic heterogeneity, even when only consider-
ing French A. thaliana populations (Le Corre et al., 2002).
Associations at FRIGIDA proved difficult to map using a
worldwide panel of accessions (Atwell et al., 2010); how-
ever, it displayed significant association in a modest panel
of 60 genotypes, all from one local population presenting
flowering time variation (Brachi et al., 2013). By using a
local population, the authors sampled individuals with a
more recent common ancestor, allowing less time for mul-
tiple causal alleles/genes to arise in the population. As a
result, fewer alleles were present at the causal loci, permit-
ting associations to be more readily revealed (see Wright
et al., 1999 for a similar reflection on finding the genetics
underlying rare diseases in humans). The cost, depending
on the species, is that sampling a restricted population
usually causes LD to decay over longer distances, reducing
the resolution of association analysis. This can limit the
ability of GWAS to pinpoint specific genes in species with
relatively long LD decay distances. Furthermore,

© 2018 The Authors

The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
170 Kathleen Beilsmith et al.
phenotypic variation may be limited for some traits of
interest at the local scale. Because both the extent to which
the leaf microbiome is heritable and the spatial scale at
which those heritable components vary among popula-
tions are still open questions, the appropriate scale for
mapping microbiome traits in particular is unclear.
When mapping populations collected from across a
large geographical range are favored, it is essential to cor-
rect for population structure in order to control the occur-
rence of false positives. A common way to perform this
correction, which has been successful in A. thaliana, is to
include a matrix of genotype similarity in a mixed linear
model for association testing (Kang et al., 2010). A draw-
back to such corrections is that they can erase associations
with SNPs that are strongly correlated with population
structure. The Quantitative Trait Cluster Association Test
(QTCAT) also has the potential to circumvent false posi-
tives due to demographic history. The QTCAT enables
simultaneous multi-marker associations while considering
correlations between markers, thus avoiding the need to
correct for population structure (Klasen et al., 2016). This
approach could prove useful given how little we know
about the extent to which the microbiome is correlated
with population structure.
One concerning limitation of GWAS, especially for the
microbiome, is a lack of power to detect small effects
(Nordborg and Weigel, 2008). It is easiest to detect associa-
tions genome-wide for traits coded by a small number of
loci with large effects. It is therefore not surprising that
GWAS of plant traits have been especially successful in the
identification of resistance genes, which are highly herita-
ble (genetically determined), of strong phenotypic effect,
and which act relatively autonomously (Bergelson and
Roux, 2010). In addition, evolutionary mechanisms tend to
maintain variation at resistance genes, sometime across
the host range (Stahl et al., 1999; Karasov et al., 2014),
reducing the confounding of associations by population
structure. Variation in microbial community composition is
unlikely to be influenced by host traits with such a simple
genetic basis because so many aspects of plant and cell
biology – including the cell wall, cell membrane, signal
transduction, meiosis, metabolic pathways and trichomes
(Horton et al., 2014) – appear to be involved. Given these
limitations in power, it is particularly important to select
the correct scale of phenotypic variation, genotype suffi-
cient SNPs, and find appropriate corrections for population
structure depending on the specific microbiome traits of
interest.
Design
The extent to which features of the phyllosphere micro-
biome are shaped by host genotype or the environment
is still an open question. Depending on whether they are
interested in the effects of plant genotype on microbiome
The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
composition in a single environment, average effects that
are consistent across environments, or changes in effects
across environments (i.e. genotype-by-environment, or
G 9 E, interactions), investigators can design studies to
better control environmental effects in the lab or better
quantify them in the greenhouse or field.
Environmental variation can mask observable relation-
ships between host genes and features of the phyllo-
sphere microbiome by creating differences in host gene
expression and microbial growth across replicates of the
same host genotype. Associations that might be
quenched by such variation could be recovered in studies
that use synthetic plant microbial communities to gather
phenotypes on a diverse panel of host plants in tightly
controlled environmental conditions (Bodenhausen et al.,
2014; Bai et al., 2015). A study in the model plant A. thali-
ana demonstrated that host genetic effects on the leaf
microbiome can be measured in synthetic communities.
Differences in community composition were observed
among four A. thaliana accessions 2 weeks after gnotobi-
otic leaves were sprayed with a synthetic community
composed of seven microbes from the most abundant
phyla sampled in natural leaf communities (Bodenhausen
et al., 2014).
When it occurs across sites in the field, environmental
variation can cause leaf microbial community phenotypes
to map differently to plant genotypes. This principle was
demonstrated in a recent field study of a perennial mus-
tard (Boechera stricta) grown in a variety of environments
differing in elevation, moisture, temperature, soil nutrients
and plant diversity. In B. stricta leaves of 48 lineages, both
the overall community diversity and the relative abun-
dance of specific taxonomic groups were better predicted
by site-specific rather than site-averaged host genetic
effects (Wagner et al., 2016). Thus, environmental variation
is a key variable that must be considered in order to fully
understand the plasticity of the map between host plant
genotype and leaf microbiome phenotype. Environmental
variables have been manipulated in GWAS of other envi-
ronmentally sensitive plant traits (Li et al., 2010; Wu et al.,
2018). For example, one GWAS of flowering time in
A. thaliana was replicated in growth chambers simulating
four distinct natural climates. Some polymorphisms asso-
ciated with flowering time overlapped between climates
while others were climate-specific (Li et al., 2010). Studies
like this one may be similarly useful in establishing which
features of the environment most influence the host vari-
ants affecting leaf communities. Studies using environ-
ment as a variable can be used to find polymorphisms that
have robust effects on the microbiome across environ-
ments. Further, GWEIS (gene-by-environment genome-
wide interaction study) following on such a design can be
used to find the polymorphisms that cause plasticity
(Hamza et al., 2011).
© 2018 The Authors

Unlike flowering time, plant microbiomes vary not only
with geography but also with abiotic conditions within
sampling sites. Abiotic conditions should be measured
across samples in case they affect the chosen phenotype
for microbiome GWAS. For example, a study of soybean
and wheat root microbiomes across many different field
sites found that pH and nitrate content correlated with
community diversity (Rascovan et al., 2016).
Sampling
The microbes of the phyllosphere must be sampled, identi-
fied and enumerated in order to generate phenotypes for
use in GWAS. While imaging of hybridized fluorescent
probes or colony isolation and identification with Sanger
sequencing can produce these phenotypes for bacteria,
next-generation sequencing approaches are the most high-
throughput and comprehensive options for characterizing
leaf microbiomes. The most common form of microbiome
data is sequences for variable regions of a conserved mar-
ker gene like bacterial 16S rRNA or fungal ITS. After DNA
is extracted from lysed leaf tissue, a portion of the marker
gene is amplified from the sample of environmental DNA
by polymerase chain reaction (PCR) and sequenced on a
next-generation platform such as the Illumina MiSeq or
HiSeq. An alternative to marker gene sequencing is the col-
lection of metagenomic data. In this approach, a library of
environmental DNA is prepared from leaf lysate and
sequenced without targeting specific types of microbes or
particular genes using shotgun sequencing. Many method-
ological considerations for these approaches are not speci-
fic to phyllosphere sampling of microbial communities and
have been reviewed elsewhere (Knight et al., 2018). Below
we elaborate on phyllosphere-specific sampling concerns.
A leaf is a heterogeneous environment composed of
microhabitats defined by abiotic conditions and proximity
to physical features like stomata. Fluorescence imaging
indicates that populations of some phyllosphere bacterial
species grow to different densities in these microhabitats
(Burch et al., 2013; Peredo and Simmons, 2018). These
findings support the idea that leaf microhabitats could har-
bor different microbial communities in nature (Edwards
et al., 2015). Because the leaf is heterogeneous, sampling
designs should ensure that within-leaf variation will not
confound the relationship between SNPs and microbe
communities. Microbiome samples should be collected
either from consistent positions on leaves or by homoge-
nizing the leaf tissue before extraction to ensure that any
within-leaf variation is well sampled so that it does not
confound associations between plant genotype and micro-
bial community phenotype.
Leaf tissue can be homogenized by grinding with a
mortar and pestle or by bead beating. However, first it is
useful to separate epiphytic and endophytic microbes
because transcriptome data suggest that host plants and
© 2018 The Authors
The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
GWAS of the phyllosphere microbiome 171
microbial colonists have different interactions at these
locations. For example, gene expression patterns of the
phyllosphere bacterium Pseudomonas syringae are dis-
tinct when it is grown on leaf surfaces, where genes
involved in motility are significantly more transcribed, rel-
ative to the apoplast, where genes related to iron and
nitrogen uptake are more transcribed (Yu et al., 2013). To
preserve separate inner and outer communities, vortex
washes, sonication in buffer and surface sterilization have
each been used to first collect ‘epiphytes’, although the
steps employed are highly variable between studies, and
their efficacy at removing tightly adhered surface
microbes is not well characterized.
One fundamental limitation of phyllosphere sampling is
its destructive nature. Unlike many microbiome samples
from human hosts, which can be collected repeatedly from
the same individuals via stool samples or cheek swabs, the
leaf microbiome, with its lower bacterial load, requires the
harvest of whole tissue, especially in small species or
young seedlings. Even when large plants are sampled
repeatedly, the possible effect of tissue wounds on micro-
biome assembly must also be considered. Tissue wounds
can trigger immune responses by releasing damage-asso-
ciated molecules in the plant and can also induce phyto-
hormone synthesis (Savatin et al., 2014). Arabidopsis
thaliana mutants lacking phytohormone synthesis or sig-
naling-related genes have distinct root communities in nat-
ural soils (Lebeis et al., 2015), suggesting that altered
levels of phytohormones could affect community assembly
in nature. In addition, the age and maturity of plants and
leaves at the time of sampling can influence the leaf micro-
bial community (Meaden et al., 2016; Wagner et al., 2016).
In metagenome and marker gene datasets, DNA extrac-
tion and sequencing methods can bias the measured rela-
tive abundances of microbial taxa. This bias has been
quantified by comparing relative abundances from
sequencing data to expected abundances given samples
with a defined mix of bacterial cells from diverse taxa
(Morgan et al., 2010; Fouhy et al., 2016). In marker gene
data, the bias created by the choice of target sequences for
amplification is an even more important concern. Primer
choice affects how much chloroplast DNA is amplified, the
number of reads that can be classified, and the relative
abundances of different taxa in phyllosphere sequence
samples. If amplification and sequencing methods severely
skew the abundances of taxa influenced by host genotype
in the community profile, then differences at these steps
could limit our ability to synthesize GWAS findings across
work in the model system A. thaliana and across plant sys-
tems in general.
Phenotyping
The most important aspect of any GWAS pipeline is defin-
ing which traits will be mapped to the host genome. This
172 Kathleen Beilsmith et al.
in turn determines the type of raw phenotype data to col-
lect. To help determine the type of microbiome sequencing
to perform and the way in which data will be analyzed, we
suggest focusing on three main questions.
1. Is the research question better addressed by associat-
ing host genes with microbial taxa or with the biological
functions encoded in microbial genomes?
2. If the research focuses on taxa, what mode of host–mi-
crobe interactions is more likely: diffuse interactions
between hosts and microbial communities or targeted
interactions between hosts and specific microbes in the
community?
3. How important is the environment in shaping host–mi-
crobe interactions, and do the hypotheses under consider-
ation require explicit investigation of the impact of host
genotype-by-environment interactions on microbial com-
munities?
In the following three sections, we discuss how the
answers to these three questions impact the selection of
phenotypes for microbiome GWAS, as well as how diverse
approaches can be used to build a richer picture of how
host plants shape the microbial communities of their
leaves.
Phenotyping Microbial Taxa or Microbial Functions. In-
stead of focusing on how host genotype influences a sin-
gle microbial taxon, microbiome GWAS focuses on how
host genotype affects the overall community. The most
intuitive way of characterizing microbial communities is to
identify microbial species or strains, and quantify their
abundances. In practice, the most common strategy is to
sequence PCR-amplified marker genes bearing taxonomi-
cally relevant information, and then group these sequence
reads into operational taxonomic units (OTUs) or amplicon
sequence variants (ASVs). OTUs are generated by cluster-
ing sequence reads based on similarity thresholds (West-
cott and Schloss, 2015), while ASVs are generated by
modeling and correcting sequencing errors so that even
single-nucleotide differences can be used to distinguish
biological variants (Callahan et al., 2016).
In order for the abundances of OTUs or ASVs to be
meaningful community phenotypes, two conditions must
be fulfilled. First, the genes targeted must provide suffi-
cient resolution of taxa. Although regions of 16S ribosomal
RNA (rRNA) have been used most often, they afford poor
resolution at the level of genera or finer clades. This limita-
tion can be circumvented by using faster-evolving genes to
gain deeper taxonomic resolution (Bartoli and Roux, 2017).
For example, gyraseB sequences allow pathogenic and
commensal OTUs within a microbial genus to be distin-
guished (Barret et al., 2015). The tradeoff in using less con-
served genes is that multiple or degenerate primers may
be necessary to sample all genera of interest. Second,
The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
OTU- or ASV-based taxonomic units must harbor pheno-
typic differences relevant to plant–microbe interactions. In
particular, GWAS based on these phenotypes are likely to
identify associations driven by microbial motifs conserved
within families or orders.
Obtaining sufficient taxonomic resolution with universal
primers is a key limitation of marker gene microbiome
characterization. Fortunately, untargeted whole metagen-
ome sequencing has become a viable alternative. Untar-
geted data from whole communities can be used to find
known genes or taxa in the sample or give sufficient cover-
age to reconstruct whole genomes of community mem-
bers. Untargeted data can offer finer taxonomic resolution
by covering more of each microbe’s genome. These read-
based and assembly-based approaches are described by
Knight et al. (2018), and reviewed by McIntyre et al. (2017)
and Vollmers et al. (2017), respectively. Subsequently, gen-
omes assembled de novo from metagenomic reads, or
obtained from a collection of publicly available genomes
for plant-associated microbes (Levy et al., 2018), can be
used as references to delineate and quantify the abun-
dance of up to hundreds of unique strains per bacterial
species in each phyllosphere DNA sample (Albanese and
Donati, 2017; Smillie et al., 2018). A challenge for these
techniques, however, is separation of host and microbe
DNA, as the typically larger size of host genomes limits the
coverage of small microbial genomes when sequencing
mixtures of host and microbe DNA. In addition, the com-
pleteness of reference databases limits the ability to assign
reads to genes or to assemble fragmented microbial gen-
omes by comparative methods.
Irrespective of whether amplified gene fragments or
untargeted shotgun sequencing is used to identify taxo-
nomic units, these methods enable a calculation of the rela-
tive abundances of members of the microbial communities.
However, absolute abundances provide greater information
content for understanding the interactions between hosts
and microbes (Vandeputte et al., 2017). One way to over-
come this limitation is to estimate microbial cell density in
samples using flow cytometry, for example (Props et al.,
2017). Taxa counts can then be converted to absolute abun-
dances and used to compare the population size of a partic-
ular microbe between plant genotypes.
A more inherent limitation arises from the possibility
that lateral gene transfer among microbes may obliterate
the taxonomic signal associated with ecologically impor-
tant microbial genes. Comparative studies have revealed
that bacterial genomes contain both a shared core and a
divergent accessory component. Core genes can encode
essential cell functions or clade-specific adaptive traits
(Lassalle et al., 2011). Accessory genes are acquired
through lateral transmission by conjugation, transforma-
tion and transduction (Soucy et al., 2015), and can include
genomic islands with virulence factors (Araki et al., 2006).
© 2018 The Authors

While the frequency of these transfers is difficult to quan-
tify, over evolutionary time they can contribute thousands
of genes to microbial genomes that are not conserved
among closely related taxa (Philippe and Douady, 2003).
The movement of these genes across microbial phyloge-
nies creates incongruities between the history of the trans-
ferred genes and the rest of the genome. Because laterally
transferred genes are phylogenetically incongruent, taxo-
nomic groupings based on the similarity of a conserved
marker gene will not correspond to the presence or
absence of these genes. As a result, microbiome GWAS
may be less effective than a disease case-control GWAS
for detecting resistance genes that target effector proteins,
which often show a highly variable presence/absence pat-
tern within species groups indicative of horizontal transfer.
An alternative approach to phenotyping a microbial
community is to focus on functions present in the micro-
bial community rather than taxa (Bonder et al., 2016; Levy
et al., 2018). This approach could identify associations dri-
ven by microbial traits irrespective of whether they are
found in the core or accessory genome, and regardless of
whether they are phylogenetically congruent within lin-
eages. Whole-genome sequencing of leaf microbial com-
munities will be a superior phenotyping method if
metabolic, competitive and virulence traits are more
important than phylogenetic identity for defining the effect
of a microbe within a leaf community. Furthermore, if bac-
terial genomes are well-annotated, the association of
genetic variants with specific microbial genes rather than
taxonomic assignments would give more insight into
the possible biological mechanisms underlying the
associations.
If microbial functions are to be used as a phenotype,
they can be inferred either from the microbial genomes or
from plant responses to the microbial community. In the
first approach, unassembled reads from whole-genome
sequencing are annotated (Keegan et al., 2016). The anno-
tated genes are grouped into functional systems like
biosynthetic or signaling pathways based on databases like
KEGG (Kanehisa et al., 2017). Then, the frequencies of
these systems can be compared across plant genotypes.
This approach was used to find microbial functions that
were more frequent in the rhizosphere of a fungus-resis-
tant bean cultivar than in the rhizosphere of a related, sus-
ceptible cultivar (Mendez et al., 2018). In particular,
samples from the resistant cultivar had a higher abun-
dance of microbial sequences associated with the produc-
tion of rhamnolipids and phenazines. In the second
approach, plant physiological responses that are linked to
microbial community function are measured in the pres-
ence of defined synthetic communities. An example of a
plant phenotype linked to microbial community function
appears in Paredes et al. (2018). In this study, bacterial
strains were grouped into broad functional categories
© 2018 The Authors
The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
GWAS of the phyllosphere microbiome 173
based on how they affected the response of A. thaliana to
phosphate starvation (as measured by phosphate accumu-
lation in shoots). Groups of eight or nine strains from each
functional category were then combined into synthetic
communities that were used to infect host plants; plant
phosphate accumulation was then measured. Community
membership explained over half of the variation in plant
phosphate accumulation, suggesting that this phenotype is
closely tied to microbial functions in the plant. If applied to
GWAS, plant phenotypes tied to community functions in
planta could help identify host genes responsible for struc-
turing the leaf community to favor or limit specific func-
tions instead of taxa.
Targeted Versus Diffuse Plant–Microbe Interactions. When
investigating the host genes shaping the relative abun-
dance of microbial taxonomic units (OTUs or ASVs), the
associations detected will depend upon which aspects of
the microbial community are mapped. This choice will be
driven by one’s perspective on whether plants primarily
influence their microbiomes through targeted interactions
with key taxonomic groups or diffuse interactions with
entire communities. Two emerging paradigms for plant
relationships with leaf microbial communities can be used
to illustrate the difference between these perspectives: hub
taxa versus ecosystems on a leash (Foster et al., 2017).
Hubs are identified as highly connected nodes in net-
works constructed to represent interactions between taxa
in the microbiome. Because the interactions between
microbes cannot be directly observed in nature, they are
inferred from relationships among taxa relative abun-
dances as measured by sequencing reads. The edges of
the network represent these relationships in the data, and
the nodes they connect each represent a clade of bacteria
grouped by a threshold of sequence similarity. The degree,
defined as the number of edges connecting to a node, and
measures of its centrality in a network can be used to clas-
sify a subset of microbial taxa as hubs; these hubs are
believed to be involved in some of the most important
putative microbe–microbe interactions in the plant commu-
nity. Inferring networks based on microbiome data from
amplicon sequencing is an active field of research; the
main challenge being to control for compositional effects
(i.e. biases induced by the finite number of reads per sam-
ple). Current methods rely on computing sparse correlation
or partial correlation matrices among transformed OTU rel-
ative abundances (Friedman and Alm, 2012; Kurtz et al.,
2015).
Agler et al. (2016) sequenced bacterial, fungal and
oomycete marker genes in leaves of wild A. thaliana from
five sites, as well as the leaves of three A. thaliana acces-
sions in common garden plots. They constructed a net-
work of the community members based on correlations
among relative abundances, and then used measures of
174 Kathleen Beilsmith et al.
closeness centrality, betweenness centrality, and degree to
select nodes as putative hubs. One isolate of a fungal hub
taxon, Albugo, was used to infect plants and test the
effects of hub presence on other phyllosphere biota as
measured by changes in community diversity and variabil-
ity relative to communities on control plants. Communities
containing the potential hub taxon were found to have
lower richness and to be less variable among replicates
than control communities. The authors hypothesized that
Albugo, and hub taxa in general, stabilize plant microbial
communities through strong negative interactions with
other microbes. In this framework, plant interactions tar-
geting a hub taxon can produce propagating effects that
influence the entire microbial community.
In order to find genes responsible for plant interactions
with microbes of ecological importance, investigators may
choose to apply GWAS to the relative or absolute abun-
dance of candidate hubs identified through network analy-
sis of whole-community data. If plants shape their
microbiome through pairwise interactions with hub spe-
cies that in turn shape the community through microbe–
microbe interactions, phenotypes focused on these hub
taxa may be better explained by host genotype and facili-
tate detection of the underlying genetic variants through
GWAS.
An alternative scenario is that leaf microbial communi-
ties are shaped by a large number of diffuse interactions.
Just as the cumulative action of weakly repressive miRNAs
can stabilize the transcriptome and canalize gene expres-
sion (Ma et al., 2018), small plant–microbe effects in aggre-
gate may have a significant effect on microbial load and
diversity in the phyllosphere. Plants may be selected to act
diffusely to cultivate and stabilize microbial communities if
these communities provide a benefit, either directly or indi-
rectly by preventing colonization of pathogens. This idea
has been described as the plant holding the microbial
ecosystem on a leash (Foster et al., 2017). While it is not
yet clear which measure(s) of a microbiome makes it most
resistant to invasion, or provides robust direct benefits, it
is clear that comprehensive community phenotypes, rather
than individual taxa abundances, would be the most
appropriate phenotypes for characterizing the genetic
architecture of the ‘leash’ through GWAS. Measures of
community diversity or evenness could capture some of
these effects, just as species richness captured the many
microbe–microbe interactions of the Albugo hub within the
leaf community. Temporal sampling schemes or the con-
struction of interaction networks can also allow inference
of community stability in terms of species turnover or per-
sistence.
The goal of many microbiome studies is to identify
manipulations of the microbiota that may enhance plant
traits like yield or pathogen resistance (Chaparro et al.,
2012). GWAS can be used to identify plant alleles that
The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
support the assembly of ‘good’ microbiomes that provide
these benefits to plants. Selecting appropriate phenotypes
for these GWAS will depend on whether targeted taxa or
diffusely selected community properties are thought to be
the means by which microbiomes provide benefits to
plants. For example, one way in which specific microbes
can improve pathogen resistance is by producing elicitors
of defense responses in plants. Members of the bacterial
genus Bacillus elicit plant responses that reduce the sever-
ity of subsequent pathogen infections (Kloepper et al.,
2004). There is even evidence that Bacillus subtilis can be
chemoattracted to root exudates, suggesting that the plant
recruits bacteria with protective effects (Rudrappa et al.,
2008). For cases like this, when targeted recruitment of
specific taxa for plant benefits seems likely, GWAS pheno-
types based on the abundances of just a few taxa would
best capture the plant variation relevant to the interaction.
However, there is a long record of ecological studies link-
ing increased species diversity to decreased temporal vari-
ability of biomass and species composition (McCann, 2000;
Allesina and Tang, 2012). Community evenness, or a lack
of dominance by a small number of taxa, has been linked
to lower invisibility (Hillebrand et al., 2008). For cases
where a diverse and even community is thought to prevent
invasion or limit the growth of pathogenic microbes, phe-
notypes like Faith’s phylogenetic distance or the Shannon
evenness index would be a better choice to capture rele-
vant host variation in a GWAS.
Plant–Microbe Interactions in Natural Environments. Al-
though it is well established that the environment affects
genotype–phenotype mapping in plants, it is unclear how
this effect influences the associations found with micro-
biome traits. Environmental effects may be mediated by
modifiers of the causal loci. In these cases, genome-wide
tests of interactions that establish the contributions of
both a SNP and its interaction with an environmental
variable could help to identify the modifier loci (Hamza
et al., 2011). Another possibility is that overlapping sets
of genes map to microbiome phenotypes in different
environments. While some of these genes will only show
associations in some environments, others will show
robust signal across many environments. A comparison
of significant associations identified in independent
GWAS that have been performed in multiple environ-
ments will allow the identification of environment-inde-
pendent and environment-dependent host effects on leaf
communities.
The scenario of both environmentally sensitive and envi-
ronmentally robust variation in host influence originally
appeared to apply for the trait of flowering time in A. thali-
ana (Li et al., 2010). But further study (Brachi et al., 2010)
revealed that very different environments could produce
association results with very little overlap for this trait.
© 2018 The Authors
 GWAS of the phyllosphere microbiome 175

Brachi et al. (2010) measured flowering time over 2 years Association mapping
in the field for tens of thousands of plants from natural
Linear mixed models (LMMs) are used to estimate the
lines, RILs and NILs in order to perform QTL mapping and
effects of SNPs on quantitative phenotypes, incorporating
GWAS. The phenotypes gathered from plants in the field
covariates and a matrix of relatedness to correct for popu-
mapped to different loci than those gathered from green-
lation structure (Kang et al., 2008, 2010; Widmer et al.,
house plants, with limited overlap in the genes where can-
2014). This approach has worked remarkably well when
didate SNPs fell. The GWAS on field samples detected
applied to traits with a simple genetic architecture that
many candidate genes that were not associated with flow-
were measured in a controlled greenhouse setting. For
ering time variation among accessions in the lab, including
example, GWAS of the hypersensitive response triggered
a number of circadian-clock-related genes. The unpre-
by high loads of P. syringae has pinpointed known R-genes
dictable and variable environment in the field may cause
in A. thaliana (Figure 2a; Method S1). In contrast, associa-
certain regulatory variation to be more important in com-
tion signals from LMMs applied to abundance measure-
plex traits than is revealed in the relatively stable condi-
ments of microbial taxa, which are likely to be influenced
tions of a greenhouse or incubator. If this is the case for
by many more loci, are often much weaker. For illustrative
microbiome-related host variation, then conducting pheno-
purposes, we show GWAS results for the titer of P. syrin-
typing in the field will be of particular importance to find-
gae in syringe-inoculated, greenhouse-grown plants (Fig-
ing candidate alleles with ecological relevance or utility
ure 2b), and the relative abundance of an OTU
field conditions.

(a)

–log10(P), observed
12
RPS2
10 HR (greenhouse)
–log10(P)

8
5

4 0
1 2 3 4 5 0 2 4
Chromosome –log10(P), expected

(b)
–log10(P), observed
12

10 Titer (greenhouse)
–log10(P)

5
4 0

1 2 3 4 5 0 2 4
Chromosome –log10(P), expected

(c)
12
–log10(P), observed

10 Relative abundance (field)

–log10(P)

5
4 0

1 2 3 4 5 0 2 4
Chromosome –log10(P), expected

Figure 2. Comparison of genome-wide association studies (GWAS) results for the Arabidopsis thaliana–Pseudomonas syringae interaction underscores the
challenges in GWAS of the phyllosphere.
Manhattan plots of the strength of phenotypic associations at over 200 000 single nucleotide polymorphisms (SNPs; left) and quantile-quantile plots comparing
observed P-values with those expected under a null model of no associations (right) are shown for three GWAS analyses in A. thaliana.
(a) Hypersensitive response to P. syringae DC3000::AvrRpt2 following syringe inoculation into leaves of greenhouse-grown plants (Atwell et al., 2010).
(b) Titer of P. syringae DC3000 4 days after syringe inoculation of greenhouse-grown plants (Atwell et al., 2010).
(c) Relative abundance of OTU24, which encompasses a diversity of P. syringae strains with near-identical 16S sequences, in field-grown plants (Horton et al.,
2014). To generate the figure, phenotype datasets from previously published studies were pared to 70 A. thaliana accessions shared among all studies, and
GWAS was conducted using a linear mixed model (LMM) for each trait individually (see Methods S1). The red dotted line in the Manhattan plots indicates the
genome-wide significance threshold following Bonferroni correction for the total number of SNPs (a = 0.05). A peak of significant association with hypersensitive
response falls at the gene RPS2. While titer in the greenhouse was highly heritable (pseudo-h2 = 0.88), no significant associations were recovered. The relative
abundance of OTU24 in the field was not heritable (pseudo-h2 = 0), and also lacked significant associations.

© 2018 The Authors

The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
176 Kathleen Beilsmith et al.
encompassing P. syringae in naturally colonized, field-
grown plants (Figure 2c). Unlike GWAS for the hypersensi-
tive response, these GWAS lack significant associations
after applying a stringent genome-wide correction for mul-
tiple testing (i.e. Bonferroni correction for the family-wise
error rate). A major challenge for GWAS of the microbiome
is extracting robust effects of plant genotype from associa-
tions that can be relatively weak at the level of individual
bacterial taxa.
Meta-analyses that aggregate statistical signals across
multiple analyses offer one route for strengthening signals
in microbiome GWAS. Horton et al. (2014), for example,
applied meta-analyses to GWAS on both community-level
traits and the relative abundances of individual taxa in the
community. This approach revealed that genes affecting
cell wall integrity are strongly enriched in regions harbor-
ing the top associations across multiple traits, even though
associations at these genes with the abundance of individ-
ual OTUs often failed to reach genome-wide significance in
individual GWAS. A large number of meta-analysis meth-
ods has been developed specifically for GWAS (Zhu et al.,
2018).
Multi-trait GWAS methods offer another avenue for
strengthening signals in microbiome GWAS. These meth-
ods, which jointly model a SNP’s association with many
traits rather than with each trait individually, have not yet
been applied to GWAS of the phyllosphere. The promise
of multi-trait methods for microbiome GWAS arises from a
key feature of sequence-based phyllosphere profiling: it
produces measurements of hundreds to thousands of phe-
notypes per individual (e.g. the abundance of different bac-
terial taxa, genes or predicted functions). A given SNP can
exert shared or opposing effects on different traits, or it
can affect some traits but not others. Multi-trait GWAS
methods leverage relationships among traits as additional
sources of information, and in doing so can increase map-
ping power (Korte et al., 2012; Zhou and Stephens, 2014;
Porter and O’Reilly, 2017).
Multi-trait GWAS methods typically test the hypothesis
that a given SNP has a non-zero effect size on one or more
traits against the null hypothesis that the SNP has no effect
on any trait (Zhou and Stephens, 2014). The performance
of these models depends on the genetic architecture of the
traits under investigation, and the pattern of phenotypic
and genetic correlation among them. The nuances of
model performance across a range of plausible biological
scenarios remain an active area of investigation (Porter
and O’Reilly, 2017). However, a few general patterns have
emerged. First, under some models, the largest gains in
power occur for SNPs whose effects across traits deviate
from the pattern of phenotypic correlations among those
traits. This is borne out in an empirical dataset for A.
thaliana–microbe interactions, where Wang et al. (2017)
used a bivariate GWAS model to discover a novel locus
The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
associated with the plant hypersensitive response to P. syr-
ingae that varied depending on effector genes present in
the infecting strain. Second, multi-trait GWAS offers
improved power when applied to traits that are genetically
correlated but somewhat inaccurately measured, analo-
gous to improvements in GWAS performance that result
as a consequence of phenotypes being measured across
multiple, replicate individuals of each accession (Korol
et al., 2001). This advantage may be particularly relevant to
GWAS of the phyllosphere, where heritability of the in
planta abundance of individual taxa is low, and the esti-
mated effects of each plant genotype on the abundance of
each taxon are therefore likely to be noisy when only a
handful of individuals are phenotyped per accession.
The largest barrier to implementing multi-trait GWAS
methods for high-dimensional phyllosphere phenotypes is
computational. Multi-trait GWAS analyses – particularly
those that fit a full mixed model to all traits, rather than
combining summary statistics from univariate GWAS – can
take tens to hundreds of hours to run for tens of traits.
Because runtime scales exponentially with the number of
traits analyzed, multi-trait GWAS quickly becomes intract-
able for larger numbers of traits (Korte et al., 2012). Fortu-
nately, high dimensionality is not a unique property of
microbiome phenotypic datasets. Human brain imaging,
for example, generates thousands to millions of measure-
ments per individual, which consist of both independent
and highly redundant phenotypes (Medland et al., 2014).
Motivated by these types of scenarios, approaches to syn-
thesize high-dimensional data into more meaningful phe-
notypes and to perform GWAS on extremely large
numbers of traits individually are active areas of research
(Medland et al., 2014; Ganjgahi et al., 2017) that hold pro-
mise to increase the feasibility of multi-trait GWAS of the
microbiome.
Beyond increasing mapping power, multi-trait GWAS
methods can provide a more comprehensive view of the
phenotypic consequences and evolutionary implications
of a given polymorphism. The multi-trait mixed model of
Korte et al. (2012), for example, estimates how the effect of
a given SNP is shared or differs across pairs of traits and/
or across environments. Whether alternative alleles at a
locus affect one trait or many, and whether these effects
are positively or negatively correlated among traits has
important consequences for how that locus will evolve in
natural populations that are distributed across spatially
and temporally heterogeneous environments (Frachon
et al., 2017).
After SNP effects are determined, two main fine-map-
ping methods are used to prioritize candidate causal vari-
ants from a GWAS: (i) selection based on P-values
corrected for multiple testing, and LD with the lead SNP;
and (ii) Bayesian methods that assign posterior probabili-
ties of causality to each SNP (Spain and Barrett, 2015).
© 2018 The Authors

Conditional analyses can determine if multiple weakly
linked or unlinked causal variants contribute to the associa-
tion of the same locus with a microbiome trait when com-
pared to a situation in which only one signal exists at this
locus, as has been done for GWAS on obesity in humans
(Wu et al., 2014) and nested analysis for broad-spectrum
black mold resistance in A. thaliana (Huard-Chauveau
et al., 2013). We can then derive biological meaning of can-
didate causal variants through post-GWAS bioinformatics
and empirical approaches (Gallagher and Chen-Plotkin,
2018).
Validation
Post-GWAS bioinformatics approaches for plants cur-
rently entail surveying gene ontology databases, tran-
scriptomics and epigenomic prioritization. Recent
advances in genomic annotation have allowed plant biol-
ogists to make better informed decisions on which genes
to experimentally investigate. Arabidopsis thaliana still
remains the model plant with most genomic resources
for this, including more than a thousand completely
sequenced genotypes (1001 Genomes Consortium, 2016),
transcriptomes (Klepikova et al., 2016) and a map of spe-
cies-wide methylation patterns (Schmitz et al., 2013). Sim-
ilar consortium efforts in crop species have led to
expansive genomic resources to link phenotype to geno-
type in tomato (Aflitos et al., 2014), maize (Chia et al.,
2012) and rice (Huang et al., 2012). Transcriptome analy-
ses have also been used to confirm GWAS results on har-
vest index in Brassica napus (Lu et al., 2016), resistance
to the oomycete Phytophthora infestans in potato (Muktar
et al., 2015), and a large set of biotic and abiotic stresses
in A. thaliana (Thoen et al., 2016).
After prioritizing candidate causal variants, there are still
often many potential causal variants responsible for the
association. Manipulation of host genetics is the most
robust validation of any candidate causal variant from a
GWAS. One such way in which candidate genes from an
association study on plant traits have been validated is by
testing single knockout lines, such as T-DNA insertion lines
that are readily available in A. thaliana (Thoen et al., 2016;
Vetter et al., 2016), although caution is warranted with the
use of these T-DNA insertion lines due to potential geno-
mic rearrangements misleading result interpretations (Tax
and Vernon, 2001). Phenotyping traits of interest in single
knockout lines can lead to false negatives, when allelic dif-
ferences are responsible for the association, or when the
wild-type does not have the active form of the protein.
Allele swapping can ameliorate this problem (Huard-Chau-
veau et al., 2013; Karasov et al., 2014). For gene-editing,
CRISPR-Cas9 has become widely used in transformable
plants, mainly by introducing mutations through non-
homologous end-joining of double-stranded breaks cre-
ated by CRISPR-Cas9 activity. Recently developed Cas9
© 2018 The Authors
The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
GWAS of the phyllosphere microbiome 177
variants allow for precision gene editing in plants, even in
transformation-recalcitrant species, through new delivery
methods reviewed elsewhere (Yin et al., 2017). A combina-
tion of gene-disruption platforms consisting of Tnt1 retro-
transposons, hairpin RNA-interferences constructs and
CRISPR/Cas9 nucleases, successfully validated the GWA
peak that contributed to trait variation in the symbiosis
between legumes and rhizobia in Medicago truncatula
(Curtin et al., 2017). While such techniques could also be
applied to test multiple candidate SNPs in combination,
such validation approaches cannot be applied to the most
complex of polygenic traits associated with microbiome
composition.
OUTLOOK
There are precious few studies mapping the effects of host
genetic variants on the plant microbiome. From work on
the human microbiome, however, we know that there are
few consistent association peaks across studies. At this
point, the extent to which this lack of consistency between
studies is biological versus methodological is unclear.
Inconsistencies could arise from biological differences in
the diversity of host genotypes included in the mapping
panel or the interactions of uncontrolled environmental
factors with host and microbe genotypes. They could also
arise from methodological differences in how microbiome
sequence data are reduced to phenotypes or in power due
to the sample sizes and statistical approaches used. If
broad patterns in host–microbe interactions are to be dis-
covered, it is important that care is taken in designing
future plant GWAS studies to ensure that the detected
associations can be meaningfully compared across studies.
We highlight below a few areas of basic research that we
expect will help inform successful GWAS design for leaf
microbiome traits.
Elucidating the genetic bases of leaf microbial commu-
nity variation will require that we first establish how much
of the variation in taxon abundances and microbial com-
munity properties are heritable. Without this knowledge, it
will be difficult to focus phenotyping on community traits
or members that are likely to be influenced by host genetic
variation. We must also understand the environmental
scales at which microbiomes show consistent patterns in
composition to select appropriate mapping populations.
Efforts characterizing environmental variability in the phyl-
losphere microbiome will prove invaluable in improving
the power and success of mapping efforts.
That said, the most important choice that an investigator
must make is the trait to be mapped. We still have only
limited information on which aspects of microbial commu-
nities influence their function. Are community measures
such as diversity or richness meaningful predictors of the
behavior of communities, or should one focus on particu-
lar genes, functional systems, strains or genera?
178 Kathleen Beilsmith et al.
Establishing whether the plant fitness benefits conferred
by microbes arise from community-level properties or
specific taxa will be particularly important for determining
which phenotypes can be used in GWAS to identify loci for
use in breeding or engineering for properties like yield or
pathogen resistance.
If it is found that strain-specific interactions provide most
of the selective pressure on plants and their interactions
with microbial partners, techniques that move beyond tradi-
tional GWAS to target these specific interactions will likely
prove most fruitful. For example, Wang et al. (2018) used a
two-way mixed effects model to map simultaneously both
host and pathogen genomic variation shaping their interac-
tion. They found non-overlapping regions of plant genomes
conferring resistance to Xanthomonas arboricola or Xan-
thomonas campestris. In addition, the regions conferring
resistance to the latter had different levels of specificity for
strains within the pathogen species. The same approaches
could be applied to provide an initial picture of how genes
shape the interaction between host plants and their benefi-
cial or commensal microbial inhabitants.
ACKNOWLEDGEMENTS
The authors would like to thank all members of the E&E Micro-
biome Journal club at the University of Chicago for lively discus-
sions, and Suzi Colpa for her artistic touch on the illustrations.
Funding was provided by the Dropkin Foundation, and the Univer-
sity of Chicago Biological Sciences Division.
NOTE FROM THE AUTHORS
After this review was written, a GWAS of the maize bacter-
ial phyllosphere was published. Wallace et al. (2018)
sequenced 16S transcripts from 300 maize genotypes
grown in a common field environment, and they used
these data to estimate dozens of community diversity
indices, the relative abundance of hundreds of OTUs, and
the representation of thousands of predicted metabolic
functions in community metagenomes for each leaf sam-
ple. A few percent of the metrics in each of these three
categories were heritable, many of which appeared to be
driven largely by variation in the abundance of Methylo-
bacteria. Intriguingly, functions related to the metabolism
of short-chain carbon molecules were overrepresented,
offering the hypothesis that these bacterial metabolic traits
may be an important part of the causal link between plant
genotype and phyllosphere community composition. How-
ever, GWAS identified few significant associations with
these traits, and consequently did not reveal which plant
genes or traits most strongly influence phyllosphere com-
position in maize. This study illustrates both the wealth of
insight that can be gleaned from GWAS of the phyllo-
sphere, particularly when moving beyond bacterial taxon-
omy by focusing on metagenomic features, but also the
difficulty in uncovering the links between plant genes and
The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
phyllosphere community composition in field environ-
ments. We eagerly await the future studies that extend the
approaches of Horton et al. (2014) and Wallace et al. (2018)
to larger collections of plant genotypes and additional spe-
cies, more powerful tools for quantifying the taxonomic
and functional composition of microbial communities, and
powerful multi-trait GWAS methods in hopes of overcom-
ing this challenge.
CONFLICT OF INTEREST
The authors declare no conflict of interest.
SUPPORTING INFORMATION
Additional Supporting Information may be found in the online ver-
sion of this article.
Method S1. Methods used to generate Figure 2 are reported in
Methods S1.
REFERENCES
Aflitos, S., Schijlen, E., De Jong, H. et al. (2014) Exploring genetic variation
in the tomato (Solanum section lycopersicon) clade by whole-genome
sequencing. Plant J. 80, 136–148.
Agler, M.T., Ruhe, J., Kroll, S., Morhenn, C., Kim, S.-T., Weigel, D. and
Kemen, E.M. (2016) Microbial hub taxa link host and abiotic factors to
plant microbiome variation (MK Waldor, Ed). PLoS Biol. 14, e1002352.
Albanese, D. and Donati, C. (2017) Strain profiling and epidemiology of bac-
terial species from metagenomic sequencing. Nat. Commun. 8, 2260.
Allesina, S. and Tang, S. (2012) Stability criteria for complex ecosystems.
Nature 483, 205–208.
Araki, H., Tian, D., Goss, E.M., Jakob, K., Halldorsdottir, S.S., Kreitman, M.
and Bergelson, J. (2006) Presence/absence polymorphism for alternative
pathogenicity islands in Pseudomonas viridiflava, a pathogen of Ara-
bidopsis. Proc. Natl Acad. Sci. USA 103, 5887–5892.
Atwell, S., Huang, Y.S., Vilhjalmsson, B.J. et al. (2010) Genome-wide asso-
ciation study of 107 phenotypes in Arabidopsis thaliana inbred lines.
Nature 465, 627–631.
Bai, Y., Mu€ ller, D.B., Srinivas, G. et al. (2015) Functional overlap of the Ara-
bidopsis leaf and root microbiota. Nature 528, 364.
Bar-On, Y.M., Phillips, R. and Milo, R. (2018) The biomass distribution on
earth. Proc. Natl Acad. Sci. USA 115, 6506–6511.
Barret, M., Briand, M., Bonneau, S., Pre
veaux, A., Valie re, S., Bouchez, O.,
Hunault, G., Simoneau, P. and Jacques, M.-A. (2015) Emergence shapes
the structure of the seed microbiota (HL Drake, Ed.). Appl. Environ.
Microbiol. 81, 1257–1266.
Bartoli, C. and Roux, F. (2017) Genome-wide association studies in plant
pathosystems: toward an ecological genomics approach. Front. Plant
Sci. 8, 763.
Bayer, M.M., Rapazote-Flores, P., Ganal, M. et al. (2017) Development and
evaluation of a barley 50k iSelect SNP array. Front. Plant Sci. 8, 1792.
Bergelson, J. and Roux, F. (2010) Towards identifying genes underlying eco-
logically relevant traits in Arabidopsis thaliana. Nat. Rev. Genet. 11, 867–
879.
Blekhman, R., Goodrich, J.K., Huang, K. et al. (2015) Host genetic variation
impacts microbiome composition across human body sites. Genome
Biol. 16, 191.
Bodenhausen, N., Bortfeld-Miller, M., Ackermann, M. and Vorholt, J.A.
(2014) A synthetic community approach reveals plant genotypes affect-
ing the phyllosphere microbiota. PLoS Genet. 10, e1004283.
Bonder, M.J., Kurilshikov, A., Tigchelaar, E.F. et al. (2016) The effect of host
genetics on the gut microbiome. Nat. Genet. 48, 1407–1412.
Brachi, B., Faure, N., Horton, M., Flahauw, E., Vazquez, A., Nordborg, M.,
Bergelson, J., Cuguen, J. and Roux, F. (2010) Linkage and association
mapping of Arabidopsis thaliana flowering time in nature. PLoS Genet.
6, e1000940.
© 2018 The Authors

Brachi, B., Morris, G.P. and Borevitz, J.O. (2011) Genome-wide association
studies in plants: the missing heritability is in the field. Genome Biol. 12,
232.
Brachi, B., Villoutreix, R., Faure, N., Hautekeete, N., Piquot, Y., Pauwels, M.,
Roby, D., Cuguen, J., Bergelson, J. and Roux, F. (2013) Investigation of
the geographical scale of adaptive phenological variation and its under-
lying genetics in Arabidopsis thaliana. Mol. Ecol. 22, 4222–4240.
Burch, A.Y., Finkel, O.M., Cho, J.K., Belkin, S. and Lindow, S.E. (2013)
Diverse microhabitats experienced by Halomonas variabilis on salt-
secreting leaves. Appl. Environ. Microbiol. 79, 845–852.
Callahan, B.J., McMurdie, P.J., Rosen, M.J., Han, A.W., Johnson, A.J.A. and
Holmes, S.P. (2016) DADA2: high-resolution sample inference from Illu-
mina amplicon data. Nat. Methods 13, 581–583.
Chaparro, J.M., Sheflin, A.M., Manter, D.K. and Vivanco, J.M. (2012) Manip-
ulating the soil microbiome to increase soil health and plant fertility.
Biol. Fertil. Soils 48, 489–499.
Chia, J.-M., Song, C., Bradbury, P.J. et al. (2012) Maize HapMap2 identifies
extant variation from a genome in flux. Nat. Genet. 44, 803.
Corwin, J.A. and Kliebenstein, D.J. (2017) Quantitative resistance: more
than just perception of a pathogen. Plant Cell 29, 655–665.
Curtin, S.J., Tiffin, P., Guhlin, J. et al. (2017) Validating genome-wide asso-
ciation candidates controlling quantitative variation in nodulation. Plant
Physiol. 173, 921–931.
Davenport, E.R., Cusanovich, D.A., Michelini, K., Barreiro, L.B., Ober, C. and
Gilad, Y. (2015) Genome-wide association studies of the human gut
microbiota. PLoS ONE 10, e0140301.
David, L.A., Maurice, C.F., Carmody, R.N. et al. (2014) Diet rapidly and
reproducibly alters the human gut microbiome. Nature, 505, 559–563.
Demmitt, B.A., Corley, R.P., Huibregtse, B.M., Keller, M.C., Hewitt, J.K.,
McQueen, M.B., Knight, R., McDermott, I. and Krauter, K.S. (2017)
Genetic influences on the human oral microbiome. BMC Genom. 18, 659.
Edwards, J., Johnson, C., Santos-Medell
ın, C., Lurie, E., Podishetty, N.K.,
Bhatnagar, S., Eisen, J.A. and Sundaresan, V. (2015) Structure, variation,
and assembly of the root-associated microbiomes of rice. Proc. Natl
Acad. Sci. USA 112, E911–E920.
Eltaher, S., Sallam, A., Belamkar, V., Emara, H.A., Nower, A.A., Salem,
K.F.M., Poland, J. and Baenziger, P.S. (2018) Genetic diversity and popu-
lation structure of F3:6 Nebraska winter wheat genotypes using genotyp-
ing-by-sequencing. Front. Genet. 9, 76.
Fahimipour, A.K., Kardish, M.R., Lang, J.M., Green, J.L., Eisen, J.A. and Sta-
chowicz, J.J. (2017) Global-scale structure of the Eelgrass microbiome.
Appl. Environ. Microbiol. 83, e03391–16.
Falconer, D.S. (1981) Introduction to Quantitative Genetics. New York: Long-
mans, Greens.
Foster, K.R., Schluter, J., Coyte, K.Z. and Rakoff-Nahoum, S. (2017) The evolu-
tion of the host microbiome as an ecosystem on a leash. Nature 548, 43–51.
Fouhy, F., Clooney, A.G., Stanton, C., Claesson, M.J. and Cotter, P.D. (2016)
16S rRNA gene sequencing of mock microbial populations – impact of
DNA extraction method, primer choice and sequencing platform. BMC
Microbiol. 16, 123.
Frachon, L., Libourel, C., Villoutreix, R. et al. (2017) Intermediate degrees of
synergistic pleiotropy drive adaptive evolution in ecological time. Nat.
Ecol. Evol. 1, 1551–1561.
Friedman, J. and Alm, E.J. (2012) Inferring correlation networks from geno-
mic survey data. PLoS Comput. Biol. 8, e1002687.
Gallagher, M.D. and Chen-Plotkin, A.S. (2018) The post-GWAS era: from
association to function. Am. J. Hum. Genet. 102, 717–730.
Ganjgahi, H., Winkler, A.M., Glahn, D.C., Blangero, J., Donohue, B., Kochu-
nov, P. and Nichols, T.E. (2017) Fast and powerful genome wide associa-
tion analysis of dense genetic data with high dimensional imaging
phenotypes. Nat. Commun. 9, 3254.
1001 Genomes Consortium T (2016) 1,135 Genomes reveal the global pat-
tern of polymorphism in Arabidopsis thaliana. Cell 166, 481–491.
Goodrich, J.K., Waters, J.L., Poole, A.C. et al. (2014) Human genetics shape
the gut microbiome. Cell 159, 789–799.
Goodrich, J.K., Davenport, E.R., Beaumont, M., Jackson, M.A., Knight, R.,
Ober, C., Spector, T.D., Bell, J.T., Clark, A.G. and Ley, R.E. (2016) Genetic
determinants of the gut microbiome in UK twins. Cell Host Microbe 19,
731–743.
Hackinger, S. and Zeggini, E. (2017) Statistical methods to detect pleiotropy
in human complex traits. Open Biol. 7, 170 125.
© 2018 The Authors
The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
GWAS of the phyllosphere microbiome 179
Hamza, T.H., Chen, H., Hill-Burns, E.M. et al. (2011) Genome-wide gene-
environment study identifies glutamate receptor gene GRIN2A as a
Parkinson’s disease modifier gene via interaction with coffee. PLoS
Genet. 7, e1002237.
Haney, C.H., Samuel, B.S., Bush, J. and Ausubel, F.M. (2015) Associations
with rhizosphere bacteria can confer an adaptive advantage to plants.
Nat. Plants 1, 15 051.
Hillebrand, H., Bennett, D.M. and Cadotte, M.W. (2008) Consequences of
dominance: a review of evenness effects on local and regional ecosys-
tem processes. Ecology 89, 1510–1520.
Horton, M.W., Bodenhausen, N., Beilsmith, K. et al. (2014) Genome-wide
association study of Arabidopsis thaliana leaf microbial community. Nat.
Commun. 5, 5320.
Huang, X., Kurata, N., Wei, X. et al. (2012) A map of rice genome variation
reveals the origin of cultivated rice. Nature 490, 497.
Huard-Chauveau, C., Perchepied, L., Debieu, M., Rivas, S., Kroj, T., Kars, I.,
Bergelson, J., Roux, F. and Roby, D. (2013) An atypical kinase under bal-
ancing selection confers broad-spectrum disease resistance in Arabidop-
sis. PLoS Genet. 9, e1003766.
Igartua, C., Davenport, E.R., Gilad, Y., Nicolae, D.L., Pinto, J. and Ober, C.
(2017) Host genetic variation in mucosal immunity pathways influences
the upper airway microbiome. Microbiome 5, 16.
Innerebner, G., Knief, C. and Vorholt, J.A. (2011) Protection of Arabidopsis
thaliana against leaf-pathogenic Pseudomonas syringae by Sphin-
gomonas strains in a controlled model system. Appl. Environ. Microbiol.
77, 3202–3210.
Kanehisa, M., Furumichi, M., Tanabe, M., Sato, Y. and Morishima, K. (2017)
KEGG: new perspectives on genomes, pathways, diseases and drugs.
Nucleic Acids Res. 45, D353–D361.
Kang, H.M., Zaitlen, N.A., Wade, C.M., Kirby, A., Heckerman, D., Daly, M.J.
and Eskin, E. (2008) Efficient control of population structure in model
organism association mapping. Genetics 178, 1709–1723.
Kang, H.M., Sul, J.H., Service, S.K., Zaitlen, N.A., Kong, S.-Y., Freimer, N.B.,
Sabatti, C. and Eskin, E. (2010) Variance component model to account
for sample structure in genome-wide association studies. Nat. Genet. 42,
348–354.
Karasov, T.L., Kniskern, J.M., Gao, L. et al. (2014) The long-term mainte-
nance of a resistance polymorphism through diffuse interactions. Nature
512, 436–440.
Keegan, K.P., Glass, E.M. and Meyer, F. (2016) MG-RAST, a metagenomics
service for analysis of microbial community structure and function.
Methods Mol. Biol. 1399, 207–233.
Klasen, J.R., Barbez, E., Meier, L., Meinshausen, N., Bu€ hlmann, P., Koorn-
neef, M., Busch, W. and Schneeberger, K. (2016) A multi-marker associa-
tion method for genome-wide association studies without the need for
population structure correction. Nat. Commun. 7, 13 299.
Klepikova, A.V., Kasianov, A.S., Gerasimov, E.S., Logacheva, M.D. and
Penin, A.D. (2016) A high resolution map of the Arabidopsis thaliana
developmental transcriptome based on RNA-seq profiling. Plant J. 88,
1058–1070.
Kloepper, J.W., Ryu, C.-M. and Zhang, S. (2004) Induced systemic resistance
and promotion of plant growth by Bacillus spp. Phytopathology 94,
1259–1266.
Knight, R., Vrbanac, A., Taylor, B.C. et al. (2018) Best practices for analysing
microbiomes. Nat. Rev. Microbiol. 16, 410–422.
Knights, D., Silverberg, M.S., Weersma, R.K. et al. (2014) Complex host
genetics influence the microbiome in inflammatory bowel disease. Geno-
mic Med. 6, 107.
Kolde, R., Franzosa, E.A., Rahnavard, G., Hall, A.B., Vlamakis, H., Stevens,
C., Daly, M.J., Xavier, R.J. and Huttenhower, C. (2018) Host genetic varia-
tion and its microbiome interactions within the Human Microbiome Pro-
ject. Genome Med. 10, 6.
Korol, A.B., Ronin, Y.I., Itskovich, A.M., Peng, J. and Nevo, E. (2001)
Enhanced efficiency of quantitative trait loci mapping analysis based on
multivariate complexes of quantitative traits. Genetics 157, 1789–1803.
Korte, A., Vilhja
lmsson, B.J., Segura, V., Platt, A., Long, Q. and Nordborg,
M. (2012) A mixed-model approach for genome-wide association studies
of correlated traits in structured populations. Nat. Genet. 44, 1066–1071.
Kurtz, Z.D., Mu € ller, C.L., Miraldi, E.R., Littman, D.R., Blaser, M.J. and Bon-
neau, R.A. (2015) Sparse and compositionally robust inference of micro-
bial ecological networks. PLoS Comput. Biol. 11, e1004226.
180 Kathleen Beilsmith et al.
Lambais, M.R., Crowley, D.E., Cury, J.C., B€ull, R.C. and Rodrigues, R.R.
(2006) Bacterial diversity in tree canopies of the Atlantic forest. Science
312, 1917.
Lassalle, F., Campillo, T., Vial, L. et al. (2011) Genomic species are ecologi-
cal species as revealed by comparative genomics in Agrobacterium
tumefaciens. Genome Biol. Evol. 3, 762–781.
Lebeis, S.L., Paredes, S.H., Lundberg, D.S. et al. (2015) Salicylic acid modu-
lates colonization of the root microbiome by specific bacterial taxa.
Science, 349, 860–864.
Le Corre, V., Roux, F. and Reboud, X. (2002) DNA polymorphism at the FRIGIDA
gene in Arabidopsis thaliana: extensive nonsynonymous variation is consis-
tent with local selection for flowering time. Mol. Biol. Evol. 19, 1261–1271.
Lee, J.J. and Chow, C.C. (2014) Conditions for the validity of SNP-based her-
itability estimation. Hum. Genet. 133, 1011–1022.
Levy, A., Conway, J.M., Dangl, J.L. and Woyke, T. (2018) Elucidating bacterial
gene functions in the plant microbiome. Cell Host Microbe 24, 475–485.
Li, Y., Huang, Y., Bergelson, J., Nordborg, M. and Borevitz, J.O. (2010) Asso-
ciation mapping of local climate-sensitive quantitative trait loci in Ara-
bidopsis thaliana. Proc. Natl Acad. Sci. USA 107, 21 199–21 204.
Lindow, S.E. and Brandl, M.T. (2003) Microbiology of the phyllosphere.
Appl. Environ. Microbiol. 69, 1875–1883.
Long, Q., Rabanal, F.A., Meng, D. et al. (2013) Massive genomic variation
and strong selection in Arabidopsis thaliana lines from Sweden. Nat.
Genet. 45, 884–890.
Lu, K., Xiao, Z., Jian, H. et al. (2016) A combination of genome-wide associ-
ation and transcriptome analysis reveals candidate genes controlling
harvest index-related traits in Brassica napus. Sci. Rep. 6, 36 452.
Lynch, M. and Walsh, B. (1998) Genetics and analysis of quantitative traits.
Sunderland, MA: Sinauer Associates.
Ma, F., Lin, P., Chen, Q., Lu, X., Zhang, Y.E. and Wu, C.-I. (2018) Direct mea-
surement of pervasive weak repression by microRNAs and their role at
the network level. BMC Genom. 19, 362.
Mansfield, J., Genin, S., Magori, S. et al. (2012) Top 10 plant pathogenic
bacteria in molecular plant pathology. Mol. Plant Pathol. 13, 614–629.
McCann, K.S. (2000) The diversity-stability debate. Nature 405, 228–233.
McIntyre, A.B.R., Ounit, R., Afshinnekoo, E. et al. (2017) Comprehensive
benchmarking and ensemble approaches for metagenomic classifiers.
Genome Biol. 18, 182.
Meaden, S., Metcalf, C.J.E. and Koskella, B. (2016) The effects of host age
and spatial location on bacterial community composition in the English
Oak tree (Quercus robur). Environ. Microbiol. Rep. 8, 649–658.
Medland, S.E., Jahanshad, N., Neale, B.M. and Thompson, P.M. (2014)
Whole-genome analyses of whole-brain data: working within an
expanded search space. Nat. Neurosci. 17, 791–800.
Mendez, L.W., Raaijmakers, J.M., de Hollander, M., Mendes, R. and Tsai,
S.M. (2018) Influence of resistance breeding in common bean on
rhizosphere micorbiome composition and function. ISME J. 12, 212–
224.
Mitchell-Olds, T. and Schmitt, J. (2006) Genetic mechanisms and evolution-
ary significance of natural variation in Arabidopsis. Nature 441, 947–952.
Morella, N.M., Gomez, A.L., Wang, G., Leung, M.S. and Koskella, B. (2018)
The impact of bacteriophages on phyllosphere bacterial abundance and
composition. Mol. Ecol. 27, 2025–2038.
Morgan, J.L., Darling, A.E. and Eisen, J.A. (2010) Metagenomic sequencing
of an in vitro-simulated microbial community. PLoS ONE 5, e10209.
Muktar, M., Lu€ beck, J., Strahwald, J. and Gebhardt, C. (2015) Selection and
validation of potato candidate genes for maturity corrected resistance to
Phytophthora infestans based on differential expression combined with
SNP association and linkage mapping. Front. Genet. 6, 294.
Nordborg, M. and Weigel, D. (2008) Next-generation genetics in plants. Nat-
ure 456, 720.
Org, E., Parks, B.W., Joo, J.W.J. et al. (2015) Genetic and environmental
control of host-gut microbiota interactions. Genome Res. 25, 1558–1569.
Paredes, S.H., Gao, T., Law, T.F. et al. (2018) Design of synthetic bacterial
communities for predictable plant phenotypes. PLoS Biol. 16, e2003962.
Peiffer, J.A., Spor, A., Koren, O., Jin, Z., Tringe, S.G., Dangl, J.L., Buckler,
E.S. and Ley, R.E. (2013) Diversity and heritability of the maize rhizo-
sphere microbiome under field conditions. Proc. Natl Acad. Sci. USA
110, 6548–6553.
Peredo, E.L. and Simmons, S.L. (2018) Leaf-FISH: microscale imaging of
bacterial taxa on phyllosphere. Front. Microbiol. 8, 2669.
The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
Philippe, H. and Douady, C.J. (2003) Horizontal gene transfer and phyloge-
netics. Curr. Opin. Microbiol. 6, 498–505.
Porter, H.F. and O’Reilly, P.F. (2017) Multivariate simulation framework
reveals performance of multi-trait GWAS methods. Sci. Rep. 7, 38 837.
Pritchard, J.K., Stephens, M., Rosenberg, N.A. and Donnelly, P. (2000) Asso-
ciation mapping in structured populations. Am. J. Hum. Genet. 67, 170–
181.
Props, R., Kerckhof, F.-M., Rubbens, P., De Vrieze, J., Hernandez Sanabria,
E., Waegeman, W., Monsieurs, P., Hammes, F. and Boon, N. (2017)
Absolute quantification of microbial taxon abundances. ISME J. 11,
584–587.
Rascovan, N., Carbonetto, B., Perrig, D., D
ıaz, M., Canciani, W., Abalo, M.,
Alloati, J., Gonza
lez-Anta, G. and Vazquez, M.P. (2016) Integrated analy-
sis of root microbiomes of soybean and wheat from agricultural fields.
Sci. Rep. 6, 28 084.
Rastogi, G., Coaker, G.L. and Leveau, J.H.J. (2013) New insights into the
structure and function of phyllosphere microbiota through high-through-
put molecular approaches. FEMS Microbiol. Lett. 348, 1–10.
Rothschild, D., Weissbrod, O., Barkan, E. et al. (2018) Environment domi-
nates over host genetics in shaping human gut microbiota. Nature 555,
210–215.
Rudrappa, T., Czymmek, K.J., Pare
, P.W. and Bais, H.P. (2008) Root-
secreted malic acid recruits beneficial soil bacteria. Plant Physiol. 148,
1547–1556.
Ruiz-Pe
rez, C.A., Restrepo, S. and Zambrano, M.M. (2016) Microbial and
functional diversity within the phyllosphere of Espeletia species in an
Andean high-mountain ecosystem. Appl. Environ. Microbiol. 82, 1807–
1817.
Sapp, M., Ploch, S., Fiore-Donno, A.M., Bonkowski, M. and Ros, L.E. (2018)
Protists are an integral part of the Arabidopsis thaliana microbiome. Env-
iron. Microbiol. 20, 30–43.
Sasaki, E., Zhang, P., Atwell, S., Meng, D. and Nordborg, M. (2015) “Miss-
ing” G x E variation controls flowering time in Arabidopsis thaliana.
PLoS Genet. 11, e1005597.
Savatin, D.V., Gramegna, G., Modesti, V. and Cervone, F. (2014) Wounding
in the plant tissue: the defense of a dangerous passage. Front. Plant Sci.
5, 470.
Schmitz, R.J., Schultz, M.D., Urich, M.A. et al. (2013) Patterns of population
epigenomic diversity. Nature 495, 193–198.
Shirasawa, K., Tanaka, M., Takahata, Y. et al. (2017) A high-density SNP
genetic map consisting of a complete set of homologous groups in auto-
hexaploid sweet potato (Ipomoea batatas). Sci. Rep. 7, 44 207.
Smillie, C.S., Sauk, J., Gevers, D. et al. (2018) Strain tracking reveals the
determinants of bacterial engraftment in the human gut following fecal
microbiota transplantation. Cell Host Microbe, 23, 229–240.
Smith, K.P. and Goodman, R.M. (1999) Host variation for interactions with
beneficial plant-associated microbes. Annu. Rev. Phytopathol. 37, 473–
491.
Soucy, S.M., Huang, J. and Gogarten, J.P. (2015) Horizontal gene transfer:
building the web of life. Nat. Rev. Genet. 16, 472–482.
Spain, S.L. and Barrett, J.C. (2015) Strategies for fine-mapping complex
traits. Hum. Mol. Genet. 24, R111–R119.
Stahl, E.A., Dwyer, G., Mauricio, R., Kreitman, M. and Bergelson, J. (1999)
Dynamics of disease resistance polymorphism at the Rpm1 locus of Ara-
bidopsis. Nature 400, 667–671.
Stanton-Geddes, J., Paape, T., Epstein, B. et al. (2013) Candidate genes and
genetic architecture of symbiotic and agronomic traits revealed by
whole-genome, sequence-based association genetics in Medicago trun-
catula. PLoS ONE 8, e65688.
Tax, F.E. and Vernon, D.M. (2001) T-DNA-Associated duplication/transloca-
tions in Arabidopsis. Implications for mutant analysis and functional
genomics. Plant Physiol. 126, 1527–1538.
Tenaillon, M.I., Sawkins, M.C., Long, A.D., Gaut, R.L., Doebley, J.F. and
Gaut, B.S. (2001) Patterns of DNA sequence polymorphism along chro-
mosome 1 of maize (Zea mays). Proc. Natl Acad. Sci. USA 98, 9161–9166.
Thoen, M.P.M., Davila Olivas, N.H., Kloth, K.J. et al. (2016) Genetic architec-
ture of plant stress resistance: multi-trait genome-wide association map-
ping. New Phytol. 213, 1346–1362.
Turpin, W., Espin-Garcia, O., Xu, W. et al. (2016) Association of host gen-
ome with intestinal microbial composition in a large healthy cohort. Nat.
Genet. 48, 1413–1417.
© 2018 The Authors

Uroz, S., Bue
e, M., Deveau, A., Mieszkin, S. and Martin, F. (2016) Ecology of
the forest microbiome: highlights of temperate and boreal ecosystems.
Soil Biol. Biochem. 103, 471–488.
Vandeputte, D., Kathagen, G., D’hoe, K. et al. (2017) Quantitative micro-
biome profiling links gut community variation to microbial load. Nature
551, 507–511.
Vetter, M., Karasov, T.L. and Bergelson, J. (2016) Differentiation between
MAMP triggered defenses in Arabidopsis thaliana. PLoS Genet. 12,
e1006068.
Vollmers, J., Wiegand, S. and Kaster, A.K. (2017) Comparing and evaluating
metagenome assembly tools from a microbiologist’s perspective - not
only size matters!. PLoS ONE 12, e0169662.
Wagner, M.R., Lundberg, D.S., del Rio, T.G., Tringe, S.G., Dangl, J.L. and
Mitchell-Olds, T. (2016) Host genotype and age shape the leaf and root
microbiomes of a wild perennial plant. Nat. Commun. 7, 12 151.
Wallace, J., Kremling, K.A., Kovar, L.L. and Buckler, E.S. (2018) Quantitative
genetics of the maize leaf microbiome. Phytobiomes, in press.
Wang, J., Thingholm, L.B., Skieceviciene,_ J., Rausch, P., Kummen, M., Hov,
J.R., Degenhardt, F., Heinsen, F.-A., R€ uhlemann, M.C. and Szymczak, S.
(2016) Genome-wide association analysis identifies variation in vitamin
D receptor and other host factors influencing the gut microbiota. Nat.
Genet. 48, 1396–1406.
Wang, B., Li, Z., Xu, W., Feng, X., Wan, Q., Zan, Y., Sheng, S. and Shen, X.
(2017) Bivariate genomic analysis identifies a hidden locus associated
with bacteria hypersensitive response in Arabidopsis thaliana. Sci. Rep.
7, 45 281.
Wang, M., Roux, F., Bartoli, C., Huard-Chauveau, C., Meyer, C., Lee, H.,
Roby, D., McPeek, M.S. and Bergelson, J. (2018) Two-way mixed-effects
methods for joint association analysis using both host and pathogen
genomes. Proc. Natl Acad. Sci. USA 115, E5440–E5449.
Westcott, S.L. and Schloss, P.D. (2015) De novo clustering methods outper-
form reference-based methods for assigning 16S rRNA gene sequences
to operational taxonomic units. PeerJ 3, e1487.
© 2018 The Authors
The Plant Journal © 2018 John Wiley & Sons Ltd, The Plant Journal, (2019), 97, 164–181
GWAS of the phyllosphere microbiome 181
Widmer, C., Lippert, C., Weissbrod, O., Fusi, N., Kadie, C., Davidson, R.,
Listgarten, J. and Heckerman, D. (2014) Further improvements to lin-
ear mixed models for genome-wide association studies. Sci. Rep. 4,
6874.
Wright, A.F., Carothers, A.D. and Pirastu, M. (1999) Population choice in
mapping genes for complex diseases. Nat. Genet. 23, 397–404.
Wu, Y., Gao, H., Li, H. et al. (2014) A meta-analysis of genome-wide associa-
tion studies for adiponectin levels in east asians identifies a novel locus
near WDR11-FGFR2. Hum. Mol. Genet. 23, 1108–1119.
Wu, S., Tohge, T., Cuadros-Inostroza, A.
et al. (2018) Mapping the Ara-
bidopsis metabolic landscape by untargeted metabolomics at different
environmental conditions. Mol. Plant 11, 118–134.
Xiao, Y., Liu, H., Wu, L., Warburton, M. and Yan, J. (2017) Genome-wide
association studies in maize: praise and stargaze. Mol. Plant 10, 359–
374.
Yang, J., Manolio, T.A., Pasquale, L.R. et al. (2011) Genome partitioning of
genetic variation for complex traits using common SNPs. Nat. Genet. 43,
519–525.
Yin, K., Gao, C. and Qiu, J.-L. (2017) Progress and prospects in plant gen-
ome editing. Nat. Plants 3, 17 107.
Yu, X., Lund, S.P., Scott, R.A., Greenwald, J.W., Records, A.H., Nettleton,
D., Lindow, S.E., Gross, D.C. and Beattie, G.A. (2013) Transcriptional
responses of Pseudomonas syringae to growth in epiphytic versus
apoplastic leaf sites. Proc. Natl Acad. Sci. USA 110, E425–E434.
Zhao, K., Aranzana, M.J., Kim, S. et al. (2007) An Arabidopsis example of
association mapping in structured samples. PLoS Genet. 3, e4.
Zhou, X. and Stephens, M. (2014) Efficient multivariate linear mixed model
algorithms for genome-wide association studies. Nat. Methods 11, 407–
409.
Zhu, Z., Anttila, V., Smoller, J.W. and Lee, P.H. (2018) Statistical power and
utility of meta-analysis methods for cross-phenotype genome-wide asso-
ciation studies. PLoS ONE 13, e0193256.