Comprehensive-GWAS: a pipeline for genome-wide association studies utilizing cross-validation to assess the predictivity of genetic variations

Full text

I
nt. J. Data Mining and Bioinformatics, Vol. 25, Nos. 1
/
2, 2021 17
Copyright © 2021 Inderscience Enterprises Ltd.
Leveraging machine learning to advance
genome-wide association studies
Gabrielle Dagasso
Department of Mathematics and Statistics,
Thompson Rivers University,
Kamloops, British Columbia, Canada
Email: gdagasso@gmail.com
Yan Yan
Department of Computing Science,
Thompson Rivers University,
Kamloops, British Columbia, Canada
Email: yyan@tru.ca
Lipu Wang
Department of Plant Sciences,
University of Saskatchewan,
Saskatoon, Saskatchewan, Canada
Email: lipu.wang@usask.ca
Longhai Li
Department of Mathematics and Statistics,
University of Saskatchewan,
Saskatoon, Saskatchewan, Canada
Email: longhai.li@usask.ca
Randy Kutcher
Department of Plant Sciences,
University of Saskatchewan,
Saskatoon, Saskatchewan, Canada
Email: randy.kutcher@usask.ca
Wentao Zhang
National Research Council of Canada,
Saskatoon, Saskatchewan, Canada
Email: Wentao.Zhang@nrc-cnrc.gc.ca

18 G. Dagasso et al.
Lingling Jin*
Department of Computer Science,
University of Saskatchewan,
Saskatoon, Saskatchewan, Canada
Email: lingling.jin@cs.usask.ca
*Corresponding author
Abstract: Genome-Wide Association Studies (GWAS) has demonstrated its
power in discovering genetic variations to particular traits related to
agronomically important features in crops. The typical output of a GWAS
program includes a series of Single Nucleotide Polymorphisms (SNPs) and
their significance. Currently, there is no standard way to compare results across
different programs or to select the most ‘significant’ results uniformly and
consistently. To obtain a comprehensive and accurate set of SNPs associated
with a trait of interest, we present a novel automated pipeline that leverages
machine learning for GWAS discoveries. The pipeline first performs
population structure analysis, then executes multiple GWAS software and
combines their results into a single SNP set. After that, it selects SNPs from the
set with high individual and/or joint effects with the Least Absolute Shrinkage
and Selection Operator analysis. Finally, the predictivity of the model is
assessed using cross-validation.
Keywords: genome-wide association studies; machine learning; population
structure analysis; cross-validation; LASSO; fusarium head blight.
Reference to this paper should be made as follows: Dagasso, G., Yan, Y.,
Wang, L., Li, L., Kutcher, R., Zhang, W. and Jin, L. (2021) ‘Leveraging
machine learning to advance genome-wide association studies’, Int. J. Data
Mining and Bioinformatics, Vol. 25, Nos. 1/2, pp.17–36.
Biographical notes: Gabrielle Dagasso is an undergraduate student at
Thompson Rivers University. She is finishing her Bachelor of Science with a
major in Mathematics and a minor in Computing Science in 2021. She has
been working as an Undergraduate Research Assistant with Lingling Jin since
the 3rd year of her program. She is the recipient of several research
scholarships, such as NSERC USRA and TRU Undergraduate Research
Experience Award Program.
Yan Yan is an Assistant Professor in the Department of Computing Science of
Thompson Rivers University (TRU). She received her PhD degree in
Biomedical Engineering from the University of Saskatchewan (UofS). Before
joining TRU, she worked as a Postdoctoral Fellow at the Department of
Computer Science in the University of Western Ontario and the UofS. Her
primary research interests include proteomics with a focus on peptide
sequencing, population genomics on large scale genome-wide data analysis and
machine learning in computational biology.
Lipu Wang received his PhD degree in the Department of Biology from the
University of Saskatchewan. Now, he is a Research Officer at the Crop
Development Centre, Department of Plant Sciences, University of
Saskatchewan. Previously he was a Research Associate in National Research
Council of Canada. He has extensive experience and expertise in Plant
Molecular Biology and Plant Pathology. Currently, he focuses on
characterising the molecular mechanism of Fusarium head blight (FHB)
resistance in wheat by using molecular and genomic tools. He developed an

Levera
g
ing machine learning to advance genome-wide association studies 19
LC-MS/MS-based mycotoxin test methods for FHB breeding programs and
investigates FHB resistance components through genome-wide association
study (GWAS).
Longhai Li received his PhD degree in Statistics from the University of
Toronto in 2007 and BSc degree in Statistics from the University of Science
and Technology of China in 2002. Currently, he is a Full Professor at the
University of Saskatchewan. His research activities focus on developing and
applying statistical machine learning methods for high-throughput data and
complex-structured data.
Randy Kutcher is a Professor of Plant Pathology at the Crop Development
Centre, Department of Plant Sciences, University of Saskatchewan. Previously,
he was a Research Scientist with Agriculture and Agri-Food Canada in
Saskatchewan focusing on Canola Pathology and Integrated Pest Management.
He received his BSc and MSc degrees from the University of Manitoba and
PhD degree from the University of Saskatchewan. He leads the Cereal and Flax
Pathology program; his wheat research is focused on applied strategies to
mitigate stripe rust, fusarium head blight and leaf spot diseases. He teaches an
undergraduate course and conducts extension activities with growers and
industry on field crop disease management.
Wentao Zhang is a Quantitative Geneticist from National Research Council of
Canada (NRC), is curious about key questions: how does genetic variation give
rise to phenotypic variation and how can better predict the performance of
these complex traits. His research focuses on understanding the genetic
architecture underlying complex traits include disease resistance, high yield,
better quality and good agronomics package in crops like wheat, Canola and
recently pulses by applying: (1) quantitative genetics, genomics, statistical
modelling and computational approaches for complex traits (2) Genomics
selection, machine learning and deep learning to predict complex traits.
Lingling Jin received her PhD degree in Computer Science specialised in
Bioinformatics from the University of Saskatchewan in 2017. Currently, she is
an Assistant Professor of Computer Science at the University of Saskatchewan
and an Adjunct Professor at Thompson Rivers University. Previously, she was
a Postdoctoral Research Scientist with Agriculture and Agri-Food Canada. Her
primary research interest is in computational modelling of polyploid genome
evolution and various aspects of plant bioinformatics.
This paper is a revised and expanded version of a paper entitled
‘Comprehensive GWAS: a pipeline for genome-wide association studies
utilising cross-validation to assess the predictivity of genetic variations’
presented at the ‘IEEE International Conference on Bioinformatics and
Biomedicine 2020 (BIBM 2020)’, 16–19 December 2020.
1 Introduction
With the development of next generation sequencing technology and drastically
decreasing of sequencing cost, vast amount of whole-genome data becomes available
nowadays. Genome-Wide Association Studies (GWAS) becomes powerful tools for
identifying associations between genetic variants within a species and phenotype
differences among individual samples. Typically, the genetic variants are Single

20 G. Dagasso et al.
Nucleotide Polymorphisms, also known as SNPs, which are changes of single DNA
base-pairs. The genetic data is referred to as genotypes and the phenotypic data is called
phenotypes. In most cases, the number of SNPs is much more than the number of
samples as the SNPs are across the whole genome. This makes the genotype “small-n-
large-p” data.
GWAS do not rely on previous knowledge to locate potential causal genes regions;
instead, they scan the whole genome to identify causal regions. Therefore, they have the
ability to investigate low-frequency and rare variants across the whole genome and
identify genes that could be missed by other methods. Since last decade, GWAS have
served as primary methods to reveal the relationships between genotypes and phenotypes
and made significant advancement in health, agriculture and many other fields. In
agriculture, they are widely used to dissect driving genes and biological architecture of
complex traits.
GWAS usually rely on statistical models to evaluate associations and find significant
SNPs that are related to the phenotype. Traditional GWAS perform a series of single-
SNP statistical hypothesis tests. These tests are independent for each SNP and the null
hypothesis is that there is no association between the SNP and the phenotype of interests.
If the null hypothesis is rejected, the tested SNP is reported to be associated with the
phenotype and a p-value is usually output along with the SNP to indicate the
significance. Out of all SNPs, the ones correlated with phenotypes are just a small subset,
which means that the genotype data is sparse. This property could bring complex
challenges to current GWAS methods as many of reported SNPs may not be truly related
to the phenotype. Phenotype can be binary (e.g., 1/0 representing having a particular
disease or not) or quantitative (e.g., height of individuals). When dealing with
quantitative phenotypes, linear regression is usually applied. The most used models
include Generalised Linear Models (GLMs) and Mixed Linear Models (MLMs). MLMs
are also known as Linear Mixed Models (LMMs). In this study, we will refer to them
as LMMs.
GWAS packages commonly used include PLINK (Purcell et al., 2010), BOLT-LMM
(Loh et al., 2015), FaST-LMM (Lippert et al., 2011), GCTA (Yang et al., 2011),
TASSEL (Bradbury et al., 2007), GAPIT (Lipka et al., 2012; Tang et al., 2016) and
others. All of them are based on traditional statistical models and many of them have the
options to choose between LMM, GLM and other statistical models. PLINK is a free,
open-source toolset for GWAS and population-based linkage analyses. It provides a
range of basic yet rapid computation for large biological data sets. It was first developed
for human genomes and is one of the earliest GWAS packages. Nowadays, PLINK is
considered as a standard GWAS method in many fields of study. BOLT-LMM utilises an
efficient Bayesian mixed-model and it is reported to increase the statistical power and
computational speed for expanded data sets. FaST-LMM applies a factorised log-
likelihood function in the LMM and claims to scale linearly with data size in both run
time and memory. It can be ideal for large data to achieve satisfying results. GCTA is a
tool for genome-wide complex trait analysis. By utilising LMM, it fits the contribution of
all SNPs as random effects and addresses the “missing heritability” problem of human
genomes.
TASSEL and GAPIT were developed specifically for GWAS of agricultural plants.
TASSEL was first developed and tested for maize with the purpose to handle a wide
range of insertions and deletions. Many other existing software did not consider these
types of polymorphisms before. GAPIT was developed after TASSEL and used statistical

Levera
g
ing machine learning to advance genome-wide association studies 21
methods that were similar as TASSEL. It is capable to perform both GWAS and
Genomic Selection (GS) (Goddard and Hayes, 2007). Genomic selection is a marker-
assisted selection in which genetic markers across the whole genome (typically SNPs)
are used and breeding values are predicted. GAPIT has been updated frequently to
incorporate the state-of-the-art methods. It is therefore reported to achieve the most
accurate and computationally efficient results.
2 Motivations
GWAS typically output a series of SNPs along with p-values, which are used to evaluate
the significance of SNPs associated with phenotypes of interest. A pre-defined threshold
is set to select significant SNPs for further analysis, e.g., 10–7 is a commonly used cut-off
in the literature. However, output SNPs from these programs could vary largely and their
p-values may not be comparable to each other. For example, one study investigated
different GWAS packages (Yan et al., 2018) reported that for the same input plant data,
p-values of the same SNPs output by PLINK were dramatically smaller than the ones
output by TASSEL and GAPIT. Their differences were by orders of magnitude.
Moreover, none of these programs utilises cross-validation which ensures that a model is
generalised and measures the accuracy of the models, such as GLM and LMM. Cross
validation is an important step in statistical analysis to ensure the results found are
validate and not false positives due to over estimations of the model’s accuracy (Oetting
et al., 2017). A false positive occurs when something assumed to be true is false, likewise
a false negative occurs when something assumed to be false is true. For example, when
someone tests positive for a disease but do not have it, this would be a false positive. A
common fault of some programs when dealing with false positives is to utilise either the
Bonferroni or Benjamini-Hochberg methods which may lead to false negatives and loss
of true significant SNPs (Waldmann et al., 2013).
In addition, traditional GWAS methods are criticised as causing the “missing
heritability”. It means that a large number of SNPs identified by GWAS may only
account for a proportion of the heritability of complex traits. This is because p-values
from GWAS only indicate whether SNPs are related to a phenotype or not; they cannot
tell how strongly they are related. Therefore, some individual causal SNPs fail to pass the
stringent significance thresholds because of their small effects to the phenotype (Manolio
et al., 2009; Tam et al., 2019). Thus, we cannot completely rely on GWAS results as
predictors for casual SNPs. Moreover, GWAS can be sensitive to the degree of match
between the assumed statistical model (such as GLM or LMM) and the real data
distribution. Last but not least, the single SNP analysis performed in these methods
ignores the potential correlations and joint effects among SNPs.
To overcome these limitations, machine learning can be integrated in the statistical
modelling of GWAS. Machine learning algorithms can identify relevant features from a
complex data set or make accurate predictions from such data. They have been widely
applied to whole-genome data analysis of various bioinformatics problems. When
incorporating with GWAS, it could identify high orders of interactions between SNPs
based on biological pathways and gene modules (epistasis), identify and prioritise SNPs
with small effects by evaluating simultaneously pan-genomic SNPs (Sun et al., 2020),
and serve as alternative to find casual SNPs and even predict phenotypes. Instead of

22 G. Dagasso et al.
using machine learning or traditional statistical modelling alone, combining the two
methods into a two-step model will integrate the strength of the separate methods.
The Least Absolute Shrinkage and Selection Operator (LASSO) is one of such
machine learning methods that is suitable for “big-n-small-p” data just as the genotype.
LASSO and its variations have been used to whole-genome data for feature selection and
produce a subset of SNPs with significantly large effects to the phenotype (Waldmann
et al., 2019). LASSO selects a subset of SNPs with the best joint effect explaining the
phenotype of interests. These selected SNPs can also be used in the phenotype prediction
and viewed as an alternative for genomic selection. Since LASSO can produce a reduced
set of SNPs, it could be extremely useful as a “pre-processing” step for conducting
GWAS on large genomes or polyploid crops, or as a “post-processing” step to select
relevant SNPs with high individual effects and/or high joint effects from the “significant”
SNPs reported by statistical methods like GWAS. GWAS methods often have limitations
on the scale of SNPs and could fail to find the correlated SNPs on large complex data.
The predictive model in LASSO can be validated using an independent test data set to
measure its effectiveness. In order to utilise a machine learning method as LASSO, the
whole data set is often arbitrarily split into two parts, a training data set and a test data
set. When applying cross-validation, it splits the whole data set into training and testing
repeatedly to increase the testing sample size. Cross-validation is thus usually applied to
data sets with small sample size just like the genotype data.
In this paper, we present a two-step wrapper model that integrates LASSO into
GWAS workflow and develop an automatic pipeline. It first uses traditional GWAS with
statistical models to select SNPs with their significant values, and then uses LASSO to
further select SNPs with high individual effects and/or high joint effects to enhance the
model’s results for relating SNPs to the phenotypes. In addition, the pipeline
automatically integrates population structure calculation into the model to improve the
prediction accuracy. Users do not need to generate the population structure matrix
separately and manually feed it into any GWAS program in the pipeline. Since different
GWAS programs often produce dissimilar association results, the proposed pipeline
includes multiple GWAS packages to mitigate these effects. The pipeline is publicly
available online at https://github.com/notTrivial/Comprehensive-GWAS.
3 Methods
In this section, the methods and software used in the proposed two-step wrapper model
are introduced.
The typical statistical models used in GWAS include the GLM, LMM and the
Multiple Locus Mixed Effects Model (MLMM). In this study, we select GLM and LMM.
Typically, GLM models assumes the data is independent and LMM can deal with
non-independence by adding random effects. The statistical model of a GLM can be
described as
=
y
Xe


and a LMM model can be described as
=
y
XZue




Levera
g
ing machine learning to advance genome-wide association studies 23
where y is the vector of observations,

is an unknown vector containing fixed effects
including genetic marker and population structure


Q, u is an unknown vector of
random additive effects,
X
and
Z
are the known design matrices and e is the residuals
(Bradbury et al., 2007). When the covariance between individuals is considered in the
LMM model, it is viewed as a random additive genetic effect.
Our proposed model combines several open-source software to achieve a clear and
simple GWAS workflow for end users. It conducts GWAS analysis utilising multiple
packages. The current version includes TASSEL and GAPIT as the GWAS programs.
TASSEL utilises GLM and LMM in determining associations; GAPIT uses similar
statistical models as TASSEL and claims to be 7-fold faster than it. They are particularly
popular in plant genome analysis. Since our study focuses on the plant genomes and
agriculture, the two programs are included in the pipeline by default. However, it can be
easily extended to include additional GWAS programs such as PLINK.
One important feature of the proposed model is that it integrates the automatic
calculation of population structure of the input data. Including population structure in
GWAS helps reduce false positive rate of output associations that are not related to the
phenotype (Sul et al., 2018). To determine the most likely population structure, a widely
used program, Structure, is integrated to calculate population structure matrix Q. This
enables an accurate determination of the number of sub-populations (denoted as k) in a
data set (Pritchard et al., 2000). To further verify these results, the Evanno method is
included to perform an assessment and visualise the likelihood across all tested sub-
population numbers in the data set (Evanno et al., 2005). It is implemented using the
method from the pophelper v2.3.0 package (Francis, 2019) in R. The optimal population
structure is then integrated in the model as a co-variance. Both TASSEL and GAPIT
could take into consideration of this and use the population matrix as a co-variance.
After the GWAS analysis, SNPs along with p-values reported from each program go
through a False Discovery Rate (FDR) adjustment, unless already done. The adjusted
p-values are also known as q-values. Only those SNPs with q-values less than the cut-off
are viewed as significant SNPs. In this study, the q-value=0.05 is set as the default
cut-off.
The combined significant SNPs and their q-values are then verified by LASSO with
the use of glmnet package in R (Friedman et al., 2010). This package is responsible for
fitting a GLM via penalised maximum likelihood. It is computed for both the elastic-net
and LASSO penalty paths (Friedman et al., 2010). LASSO penalises non-zero
coefficients by taking the sum of their absolute values, referred to as 1L and 2L
penalties from Ridge Regression, where it penalises the model with the sum of squared
individuals. Elastic-net combines both 1L and 2L penalties. Ridge regression is to
shrink regression coefficients for variables with minor contribution to the model
(Friedman et al., 2010). When combined with GWAS, the application of LASSO further
narrows down the SNPs found by these programs and builds a predictive model for the
phenotype. The most significant SNPs passed by LASSO are the final output from the
model.

24 G. Dagasso et al.
4 Pipeline overview
In this section, the details of the pipeline are introduced with clear explanations about
input and output data of each step.
The pipeline takes genotype and phenotype data as input and produces association
results along with visualisations and verification of association with the integration of
LASSO. The entire workflow of the pipeline is shown as a flowchart in Figure 1.
Figure 1 The flowchart depicts the overall workflow of the pipeline
The pipeline is written in R v3.6.1. To execute the pipeline, TASSEL and Structure need
to be pre-installed in the appropriate locations as described on https://github.com/
notTrivial/Comprehensive-GWAS. The rest will be installed by the pipeline
automatically at run time. It is designed to be as seamless as possible to end users to
utilise different GWAS software and adjust output from each software to be the input for
the next one. The steps taken in the pipeline are further described as follows.

Levera
g
ing machine learning to advance genome-wide association studies 25
Step 0: data pre-processing
The input data required for this pipeline is numeric genotype, non-numeric genotype,
phenotype data and a kinship matrix. The numeric genotype is used for determining the
population structure by Structure and validation by LASSO. To be specific, numeric
genotype is a matrix with rows as individuals and columns as the SNP loci. When used in
Structure, marker names and headings are removed as per its requirement. When used in
LASSO, marker names and headings are present in the file. The non-numeric genotype is
used for GWAS along with phenotypes and the kinship matrix. The phenotype data and
kinship are in TASSEL required format and both need to have two files, one with the
header and one without.
Step 1: determine optimal population structure
The Structure program (Pritchard et al., 2000) is a Bayesian clustering program that
characterises population clusters to discriminate populations, to determine population
structure and to reveal the genetic composition of individuals using molecular markers. It
uses multilocus allele frequencies to assign individuals to a pre-defined number of
populations (k). The assignment is usually run for a range of k such as 2 to 10. Multiple
repeats are carried out for each k. Each output file for each repeat of k showing the
assignment probabilities of all individuals is referred to as the Q-matrix.
The Evanno method (Evanno et al., 2005) is used to estimate the optimal number of
populations, k, based on the results generated from the Structure program (Pritchard
et al., 2000). The pophelper package (Francis, 2019) in R is the analysis tool of the
Evanno method to analyse and visualise population structure. First, Structure examines
possible numbers of sub-populations from 2 to k, where k is specified by the user, i.e.
=10k is specified in the results section. This process is repeated E times for each
possible number of sub-populations to ensure that the Evanno method is able to run. In
this study, =10E is used as default. Results are then presented to the users and the
optimal number of populations (k) is chosen. The resultant Q-matrix is the output of this
step and then will be automatically fed into the follow-up GWAS analysis.
Step 2: GWAS with traditional statistical methods
The non-numeric genotype data, phenotype data, kinship matrix K, and population matrix
Q are used as input to run GAPIT and TASSEL. The pipeline has the flexibility to choose
between GLM and LMM models as both programs support them (Bradbury et al., 2007;
Tang et al., 2016). Output of this step includes Manhattan Plots, QQ-Plots and statistical
association results from TASSL and GAPIT. They are then output to their respective
directories as shown in Figure 2.
Step 3: merge data set from statistical methods
Statistical association results from GAPIT and TASSEL are read back into the pipeline as
the input of this step. Reported p-values of each program are adjusted using FDR. Since
GAPIT can report both adjusted and un-adjusted p-values, in the current version, only
TASSEL needs this adjustment. They are then merged together and the appropriate
Manhattan plots of each statistical model (GLM or LMM) per phenotype are generated.
Only the SNPs that meet the cut-off and are found by both programs will be kept as the
output of this step and used in further verification steps.

26 G. Dagasso et al.
Figure 2 The file structure chart depicts the overall data flow of the pipeline
Step 4: LASSO and cross-validation on merged data set
The significant SNPs output from above Step are verified using the LASSO method from
the glmnet package in R to identify joint correlations. Cross-validation with 10 folds are
applied to the significant SNPs of each statistical method per phenotype. The phenotype
data is converted to discrete variables from numerical, anything over 50% is converted
to 1 and less than 50% is converted to 0. The binomial method is applied in the
validation. The coefficient plot and cross-validation plot are the output of this step and
stored under the LASSO directory.
Step 5: generate final results
The final significant SNPs passed from above step of each statistical method per
phenotype are output to respective directories. Combined with other output from
previous steps, all the results are stored under the directories as described in Figure 2.
5 Experiments and results
In this section, we describe the experiments and report the results of using the proposed
pipeline on a plant genomeic and phenotypic data set.

Levera
g
ing machine learning to advance genome-wide association studies 27
5.1 Experimental data
Fusarium Head Blight (FHB) is a fungal disease that affects both wheat and barley
worldwide. It is a serious disease that causes reduced crop yields and produces
mycotoxins which pose serious health threats to both humans and livestock (Walter et al.,
2010; Hilton et al., 1999). Therefore, detecting FHB resistance has significant meaning in
plant breeding. GWAS has been widely used to provide important insights into
the linkages of complex traits to causal SNPs and its related genes in plant breeding
(Yan et al., 2019).
In this study, 199 bread wheat varieties were selected with the lowest FHB severity
rate among a poll of 4000 varieties. 2235 SNPs were detected using wheat 90 k SNP
array (Wang et al., 2014). Three phenotypes including flowering date, plant height and
disease severity have been screened in a FHB disease greenhouse in Saskatchewan,
Canada in 2019.
5.2 Population structure determination
The pipeline’s built-in function identifies the most likely population structure of the
experimental data. Figures 3 and 4 show the number of population verified by Evanno
method. As noted in both figures, the graph plateau’s at k = 2, which supports that k = 2
is the optimal number. The Q-matrix produced by Structure with k = 2 is used in the
subsequent GWAS runs (Evanno et al., 2005). Figure 5 is a stacked barplot that shows
the population group assignment of each individual in 100 repetitions for k ranging from
2 to 7 groups of populations.
Figure 3 The likelihood distribution of k results using the pophelper package in R (Step 1 in
Figure 1). The graph shows the mean of ()( )
L
kSD

over 10 runs for each k value.
It’s plateau at k = 2, which supports the usage of its corresponding Q-matrix

28 G. Dagasso et al.
Figure 4 k plot results using the Evanno method in the pophelper package in R. k is
calculated from the absolute value of the second order rate of change of the likelihood
distribution divided by the standard deviation of the likelihood distribution (Step 1 in
Figure 1). The modal value of this distribution is the true k or the uppermost level of
structure. Here = 2k clusters because the graph plateau’s at = 2k. This supports the
usage of the Q-matrix of = 2k
5.3 GWAS results
From all three tested phenotypes, two of them, plant height and disease severity, did not
get significant results by TASSEL and GAPIT. The resultant QQ plots from them
showed that the model did not fit the data well. However, this is also likely due to the
fact that the 90 k SNP array only detects SNPs in “preselected” genomic regions that
have a higher likelihood to be associated with traits of interest without including any
irrelevant positions (Tang et al., 2016; Bradbury et al., 2007). In the following, results on
flowering date are shown and explained in details.
Figures 6 and 7 show the Manhattan plots of TASSEL and GAPIT’s GWAS results.
The SNP’s p-values are with a scale of log 10 ( )pvalue

. They are both produced by
the GLM of the two programs as The LMM found no significant results for either of
them. Both programs identified the most significant SNPs located on chromosomes 2B
and 6B.

Levera
g
ing machine learning to advance genome-wide association studies 29
Figure 5 Multiline barplots from Q-matrices from = 2k to = 7k for step 1 in Figure 1. Each
group is individually labelled

30 G. Dagasso et al.
Figure 6 Manhattan plot for flowering date by GLM of GAPIT. The most significant SNPs are
above the green line and located on chromosomes 2B and 6B (Step 2 in Figure 1)
Figure 7 Manhattan plot for flowering date by GLM of TASSEL. The most significant SNPs are
located on chromosomes 2B and 6B (Step 2 in Figure 1)
There were a total of 65 significant (q-values

0.05) SNPs predicted by the pipeline
using GLM of TASSEL and GAPIT. The q-value is adjusted by the FDR (Benjamini and
Hochberg, 1995) of the p-value. They are the merged results from the two programs.
Among the whole genome, chromosome 2B and chromosome 6B have the largest
number of SNPs which the two programs agree on. This indicates that the two regions
may have significant influence to the flowering time phenotype. The pipeline output is
summarised in the merged Manhattan plot in Figure 8. It shows the merged significant
SNPs Manhattan Plot with a scale of log 10 ( )q value

. The coloured columns help
distinguish different chromosomes. Each dot on this Manhattan plot corresponds to one
significant SNP (q-values  0.05), though some overlap to appear as a larger circle.

Levera
g
ing machine learning to advance genome-wide association studies 31
Figure 8 Manhattan plot for flowering date using the merged SNPs from both TASSEL and
GAPIT. All 65 significant SNPs (whose q-values

0.05) are illustrated in this graph.
The colours indicate different chromosomes from its neighbouring chromosomes
(Step 3 in Figure 1)
The Manhattan plot also highlights the top 3 significant merged SNPs with their names
displayed. These top 3 significant SNPs are also listed in Table 1 with their chromosome
and location information. The results show that our proposed pipeline captured
significant SNPs from both GWAS software and they could have joint effects to the
phenotype.
Table 1 Top 3 Significant SNPs after merging (Step 5 in Figure 1)
Name of SNP Chromosome Position
wsnp_RFL_Contig4668_5551137 6B 71
wsnp_Ex_rep_c71064_69904031 2B 100
wsnp_Ra_c3955_7262354 2B 116
5.4 LASSO and cross-validation
Top significant SNPs found by the glmnet of R package are shown in Table 2. They
played an important role in fitting the model and therefore were concluded to be the top
significant SNPs from the glmnet analysis. SNPs used in glmnet are only the 65
significant SNPs from the merged results of GAPIT and TASSEL. We notice that these
top SNPs are not the same as the SNPs with the lowest q-values. But they could pinpoint
new regions that researchers could focus on to find potential casual relationships to the
phenotype. As we stated in Section 2, GWAS results only reflect associations but not
how strong they are. SNPs reported by LASSO may influence the phenotype but failed to
pass the q-value threshold.

32 G. Dagasso et al.
Table 2 Top 5 significant SNPs found by glmnet in step 5 of Figure 1. These SNPs
contributed more to the development of the model fitting than the other 65 significant
SNPs from the merged results
Name of SNP Chromosome Position
Excalibur_rep_c114833_645 3D 41
D_contig25474_188 7D 161
RAC875_c5243_479 1B 31
BS00065005_51 1A 55
wsnp_Ex_c56629_58677561 5B 12
The coefficients plot from the model fitting of SNPs is shown in Figure 9, which includes
all 65 significant SNPs. Each curved line in Figure 9 corresponds to one SNP and it
shows the path of each SNPs’ coefficient against the 1L-norm. The number of non-zero
coefficients is indicated as the numbers in the x-axis on the top.
Figure 10 shows the corresponding cross-validation result. It shows the validation of
the model, the choice of two
s

, and the LASSO regularisation parameter, as indicated
by the vertical dotted lines. The x-axis on the top is the number of non-zero coefficients
just the same as the ones in Figure 9. The red dotted line is the cross-validation curve and
the error bars indicate the upper and lower standard deviation of the

. One

gives the
minimum mean cross-validated error and the other gives the most regularised model such
that the error is within one standard error of the minimum (Friedman et al., 2010).
Figure 10 also shows that the misclassification error rate is about 0.20 , which indicates
the predictivity of the selected SNPs. This highlights the significance of the SNPs from
the merged results.
Figure 9 Coefficients from LASSO for the 65 merged significant SNPs (Step 4 in Figure 1)

Levera
g
ing machine learning to advance genome-wide association studies 33
Figure 10 Cross-validation from LASSO for the 65 merged significant SNPs (Step 4 in Figure 1)
6 Conclusions and discussions
In this study, we proposed a computational method for seamless GWAS analyses with
statistically modelling and machine learning. It uses an ensemble approach that allows
easy analysis of associations between genotype and phenotype data to narrow down areas
and SNPs of “true” significance while being efficient and easy to handle. Combining
results from two well-known GWAS programs yields more promising results, which
allows overlapping analysis of multiple programs across multiple statistical models.
Automatically including the population structure analysis in these GWAS programs helps
to reduce a high rate of false positive associations for SNPs that are not linked with
phenotypes. The final step of the pipeline further verifies the significant SNPs by using
LASSO to provide better accuracy in detecting “true” significant SNPs to help narrow
down the number of resultant SNPs (Friedman et al., 2010).
The pipeline is created for GWAS analysis in one place while utilising various
software. It also has the capacity to extend and include more GWAS programs and
statistical models to be utilised in other analyses. We demonstrated the utility of this

34 G. Dagasso et al.
pipeline in associating FHB resistance related phenotypes with genomic regions in bread
wheat varieties. Significant SNPs were found in the flowering date phenotype by GLM,
whereas the other studied phenotypes had no significant SNPs found with the q-value
threshold of 0.05 using TASSEL and GAPIT. The significant SNPs in the flowering date
phenotype was verified by FDR adjusted p-values and were further verified by cross-
validation using LASSO (Friedman et al., 2010). Having this result verified by both
GAPIT and TASSEL lends credibility to the finding of these significant SNPs.
There are a few directions where the pipeline will be extended to obtain better and
more comprehensive results. For example, instead of detecting associated SNPs for each
phenotype independently, the pipeline can be extended to combine results from multiple
related phenotypes. It is known that flowering date, plant height and disease severity
rates are phenotypes related to FHB disease; therefore, genetic variants that are common
across multiple related phenotypes or unique to any specific phenotype can be aggregated
and associated with the FHB disease. Further, given that results from different GWAS
programs could vary largely in terms of output SNPs as well as their p-values, in addition
to the current q-value threshold, the pipeline can include other selection criterion such as
choosing the top significant SNPs without meeting the q-value threshold. Finally, more
GWAS packages such as PLINK can be included in the pipeline to complement with
current two software, GAPIT and TASSEL. In this case, additional file format
conversion could also be added into the pipeline, for example, to convert PLINK ped/bed
files to the VCF files. This will enable the pipeline to be further extended to synthesise
and report significant SNP sets in a customised but uniform way.
References
Benjamini, Y. and Hochberg, Y. (1995) ‘Controlling the false discovery rate: a practical and
powerful approach to multiple testing’, Journal of the Royal Statistical Society: Series B
(Methodological), Vol. 57, No. 1, pp.289–300.
Bradbury, P.J. and Zhang, Z., Kroon, D.E., Casstevens, T.M. (2007) ‘TASSEL: software for
association mapping of complex traits in diverse samples’, Bioinformatics, Vol. 23, No. 19,
pp.2633–2635.
Evanno, G., Regnaut, S. and Goudet, J. (2005) ‘Detecting the number of clusters of individuals
using the software STRUCTURE: a simulation study’, Mol. Ecol., Vol. 14, No. 8,
pp.2611–2620.
Francis, R.M. (2019) Tabulate, Analyse and Visualise Admixture Proportions from STRUCTURE,
TESS, BAPS, ADMIXTURE and Tab-Delimited q-matrices Files, R package version 2.3.0.
Friedman, J., Hastie, T. and Tibshirani, R. (2010) ‘Regularization paths for generalized linear
models via coordinate descent’, Journal of statistical software, Vol. 33, No. 1, pp.1–22.
Goddard, M.E. and Hayes, B.J. (2007) ‘Genomic selection’, Journal of Animal breeding and
Genetics, Vol. 124, No. 6, pp.323–330.
Hilton, A.J., Jenkinson, P., Hollins, T.W. and Parry, D.W. (1999) ‘Relationship between cultivar
height and severity of fusarium ear blight in wheat’, Plant Pathology, Vol. 48, No. 2,
pp.202–208.

Levera
g
ing machine learning to advance genome-wide association studies 35
Lipka, A.E., Tian, F., Wang, Q., Peiffer, J., Li, M., Bradbury, P.J., Gore, M.A., Buckler, E.S. and
Zhang, Z. (2012) ‘GAPIT: genome association and prediction integrated tool’, Bioinformatics,
Vol. 28, No. 18, pp.2397–2399.
Lippert, C., Listgarten, J., Liu, Y., Kadie, C.M., Davidson, R.I. and Heckerman, D. (2011) ‘Fast
linear mixed models for genome-wide association studies’, Nature Methods, Vol. 8, No. 10,
pp.833–835.
Loh, P-R., Tucker, G., Bulik-Sullivan, B-K., Vilhjalmsson, B.J., Finucane, H.K., Salem, R.M.,
Chasman, D.I., Ridker, P.M., Neale, B.M. and Berger, B. et al. (2015) ‘Efficient Bayesian
mixed-model analysis increases association power in large cohorts’, Nature Genetics, Vol. 47,
No. 3, pp.284–290.
Manolio, T.A., Collins, F.S., Cox, N.J., Goldstein, D.B., Hindorff, L.A., Hunter, D.J., McCarthy,
M.I., Ramos, E.M., Cardon, L.R. and Chakravarti, A. et al. (2009) ‘Finding the missing
heritability of complex diseases’, Nature, pp.747–753.
Oetting, W.S., Jacobson, P.A. and Israni, A.K. (2017) ‘Validation is critical for genome-wide
association study-based associations’, American Journal of Transplantation, Vol. 17, No. 2,
pp.318–319.
Pritchard, J.K., Stephens, M. and Donnelly, P. (2000) ‘Inference of population structure using
multilocus genotype data’, Genetics, Vol. 155, No. 2, pp.945–959.
Purcell, S., Neale, B., Todd-Brown, K., Thomas, L., Ferreira, M.A.R., Bender, D., Maller, J., Sklar,
P., de Bakker, P.I.W. and Daly, M.J. et al. (2007) ‘PLINK: a tool set for whole-genome
association and population-based linkage analyses’, The American Journal of Human
Genetics, Vol. 81, pp.559–575.
Sul, J.H., Martin, L.S. and Eskin, E. (2018) ‘Population structure in genetic studies:
confounding factors and mixed models’, PLoS Genetics, Vol. 14, No. 12. Doi:
10.1371/journal.pgen.1007309.
Sun, S., Dong, B. and Zou, Q. (2020) (2020) ‘Revisiting genome-wide association studies from
statistical modelling to machine learning’, Briefings in Bioinformatics. Doi:
10.1093/bib/bbaa263.
Tam, V., Patel, N., Turcotte, M., Bossé, Y., Paré, G. and Meyre, D. (2019) ‘Benefits and
limitations of genome-wide association studies’, Nature Reviews Genetics, Vol. 20, No. 8,
pp.467–484.
Tang, Y., Liu, X., Wang, J., Li, M., Wang, Q., Tian, F., Su, Z., Pan, Y., Liu, D. and Lipka,
A.E. et al. (2016) ‘GAPIT version 2: an enhanced integrated tool for genomic association and
prediction’, The plant genome, Vol. 9, No. 2, pp.1–9.
Waldmann, P., Ferenčaković, M., Mészáros, G., Khayatzadeh, N., Curik, I. and Sölkner, J. (2019)
‘Autalasso: an automatic adaptive lasso for genome-wide prediction’, BMC Bioinformatics,
Vol. 20, No. 1, pp.1–10.
Waldmann, P., Mészáros, G., Gredler, B., Fuerst, C. and Sölkner, J. (2013) ‘Evaluation of the lasso
and the elastic net in genome-wide association studies’, Frontiers in genetics, Vol. 4. Doi:
10.3389/fgene.2013.00270.
Walter, S., Nicholson, P. and Doohan, F.M. (2010) ‘Action and reaction of host and pathogen
during Fusarium head blight disease’, New Phytologist, Vol. 185, No. 1, pp.54–66.
Wang, S., Wong, D., Forrest, K., Allen, A., Chao, S., Huang, B.E., Maccaferri, M., Salvi, S.,
Milner, S.G. and Cattivelli, L. et al. (2014) ‘Characterization of polyploid wheat genomic
diversity using a high-density 90 000 single nucleotide polymorphism array’, Plant
Biotechnology Journal, Vol. 12, No. 6, pp.787–796.

36 G. Dagasso et al.
Yan, Y., Burbridge, C., Shi, J., Liu, J. and Kusalik, A. (2018) ‘Comparing four genome-wide
association study (GWAS) programs with varied input data quantity’, Proceedings of the
IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pp.1802–1809.
Yan, Y., Burbridge, C., Shi, J., Liu, J. and Kusalik, A. (2019) ‘Effects of input data quantity on
genome-wide association studies (GWAS)’, International Journal of Data Mining and
Bioinformatics, Vol. 22, No. 1, pp.19–43.
Yang, J., Lee, S.H., Goddard, M.E. and Visscher, P.M. (2011) ‘GCTA: a tool for genome-wide
complex trait analysis’, The American Journal of Human Genetics, Vol. 88, No. 1, pp.76–82.