Biplot Analysis of

Multi-Environment Trial Data

Weikai Yan
May 2006
Contact: [email protected]
Multi-Environment Trials (MET)
• MET are essential
• MET are expensive
• MET data are valuable
• MET data are not fully used

Weikai Yan2006
 Why biplot analysis?
• Biplot analysis can help understand MET
data
– Graphically,
– Effectively,
– Conveniently

Weikai Yan2006
 Outline
• Multi-environment trial (MET) data
• Basics of biplot analysis
• Biplot analysis of G-by-E data
• Biplot analysis of G-by-T data
• Better understanding of MET data
• Conclusions

Weikai Yan2006
 Multi-environment
trial data

Contact: [email protected]
 MET data is
a genotype-environment-trait
(G-E-T) 3-way table
• Multiple Genotypes
• Multiple Environments
• Multiple Traits

Weikai Yan2006
 A G-E-T 3-way table contains
many 2-way tables
• G by E: for each trait
• G by T (trait): in each environment;
across environments
• E by T: for each genotype; across
genotypes

G-E-T data >> G-E data

Weikai Yan2006
 A G-E-T 3-way table is
an extended 2-way table
• G by V:
– each E-T combination as a variable (V)
• P by T:
– each G-E combination as a phenotype
(P)

Weikai Yan2006
 A G-E-T 3-way table implies
informative 2-way tables
• Association by environment 2-way
tables
– Associations:
• among traits
• between traits and genetic markers

Weikai Yan2006
 Goals of MET data analysis

• Short-term goals:
– Variety evaluation
• Response to the environment (G x E)
• Trait profiles (G x T)
• Long-term goals:
– To understand
• the target environment (G x E)
• the test environments (G x E)
• the crop (G x T)
• the genotype x environment interaction (A x T)

Weikai Yan2006
 Basics of biplot
analysis
Most two-way tables can be
visually studied using biplots

Contact: [email protected]
 Origin of biplot
 Gabriel (1971)
 One of the most
important advances in
data analysis in recent
decades
 Currently…
 > 50,000 web pages
 Numerous academic
publications
 Included in most
statistical analysis
packages
 Still a very new
technique to most
scientists
Prof. Ruben Gabriel, “The founder of biplot”
Courtesy of Prof. Purificación Galindo
University of Salamanca, Spain

Weikai Yan2006
 What is a biplot?
• “Biplot” = “bi” + “plot”
– “plot”
• scatter plot of two rows OR of two columns, or
• scatter plot summarizing the rows OR the columns
– “bi”
• BOTH rows AND columns
• 1 biplot >> 2 plots

Weikai Yan2006
 Mathematical definition of a Biplot
Graphical display of matrix multiplication

Matrix multiplication 5

A(4, 2) B(2, 3) P(4, 3) 4 B1

 x y  b1 b2 b3 
 a1  b1 b2 b3   20  9 6 
  A2 A1
4 3 a1
   3

a 2 
3 3  x 2 3 3    a2 6 12  15
4.472 cos =

    0.8944
  2

a3 1  3  y 4 1  2  a3  10  6 9  5.0

a 4 4 0   a4 8  12 12  1 B2
P11 = 5*4.472*0.8944 = 20

Y
0
O A4
-1

-2 B3

“Inner product property” -3 A3

– Pij =OAi*OBj*cosij -4
-4 -3 -2 -1 0 1 2 3 4 5
X
– Implies the product matrix
Weikai Yan2006
 Practical definition of a biplot
“Any two-way table can be analyzed using a 2D-biplot as soon as it can be
sufficiently approximated by a rank-2 matrix.” (Gabriel, 1971)
(Now 3D-biplots are also possible…)

Matrix decomposition 5

4 E1
P(4, 3) G(3, 2) E(2, 3)
 e1 e2 e3   x y 3 G2 G1
 g1 20  9 6   g1 4 3 
 
e1 e2 e3  2

  
 g2 6  g 2  3 3    x 2  3 3 
12  15  1 E2
     

Y
 g3  10  6 9   g 3 1  3  y 4 1  2
0
O G4
 g 4 8  12 12   g 4 4 0  -1

-2
E3
G-by-E table -3
G3
-4
-4 -3 -2 -1 0 1 2 3 4 5
X

Weikai Yan2006
 Singular Value Decomposition (SVD) &
Singular Value Partitioning (SVP)
The ‘rank’ of Y, i.e.,
the minimum number Matrix Matrix
characterising “Singular values” characterising
of PC required to
fully represent Y the rows the columns

r
SVD:   aik k bkj
Yij  SVD

k 1
r

SVP:   (aik kf )( 1k f bkj )

SVP
 (0 ≤ f ≤ 1)
k 1

Rows scores Column scores

SVD = PCA? Plot Biplot Plot

Weikai Yan2006
 Biplot interpretations
r
Yij   ( aik kf )( 1k f bkj )
k 1

 Inner-product property
 Interpretations based on biplots with f = 1
 approximates YYT, the distance matrix
 Similarity/dissimilarity among row (genotype) factors
 Interpretations based on biplots with f = 0
 approximates YTY, the variance matrix
 Similarity/dissimilarity among column (environment)
factors
 Combined use of f = 0 and f = 1

(Gabriel, 2002 Biometrika; Yan, 2002, Agron J; Built in the GGEbiplot software)
Weikai Yan2006
 Biplot analysis is…
to use biplots to display
– a two-way data per se (Y),
– its distance matrix (YYT), and
– its variance matrix (YTY)
so that
– relationships among rows,
– relationships among columns, and
– interactions between rows and columns
can be graphically visualized.

Weikai Yan2006
Data centering prior to biplot analysis
• The general linear model for a G-by-E
data set (P)
– P = M + G + E + GE
• Possible two-way “tables” (Y):
• Y = P = M + G + E + GE —original data: QQE biplot
• Y = P – M = G + E + GE —global-centered (PCA)
• Y = P – M – E = G + GE —column-centered: GGE biplot
• Y = P – M – G = E + GE —row-centered
• Y = P – M – G – E = GE —double-centered: GE biplot
All models are useful, depending on the research objectives (built in GGEbiplot)
Weikai Yan2006
Data scaling prior to biplot analysis
• Different GGE biplots
• Yij = (i + ij)/sj
• Sj = 1 no scaling

• Sj = (s.d.)j all environments are equally important

• Sj = (s.e.)j heterogeneity among environments is removed

(built in GGEbiplot)
Weikai Yan2006
 Four questions must be asked
before trying to interpret a biplot

1. What is the model?

How the data were centered and scaled?
What are we looking at?
2. What is the goodness of fit?
How confident are we about what we see?
What if the data is fitted poorly?
3. How singular values are partitioned?
What questions can be asked?
4. Are the axes drawn to scale?
Are the patterns artifacts?

(All are addressed explicitly in GGEbiplot)

Weikai Yan2006
 Biplot Analysis of
G-by-E data

TEST
GENOTYPE ENVIRONMENT
EVALUATION EVALUATION

MEGA-
ENVIRONMENT
ANALYSIS

Contact: [email protected]
 Sample G-by-E data
(Yield data of 18 genotypes in 9 environments, 1993, Ontario, Canada)

Weikai Yan2006
 Before trying to interpret a biplot…
1. Model selection?
Centering = 2 (“G+GE”)
Scaling =0

2. Goodness of fit?
78%.

3. Singular value
partitioning?
SVP = 2 (environment-
metric)

4. Draw to scale?
Yes.

Weikai Yan2006
 G By E data analysis

TEST
GENOTYPE ENVIRONMENT
EVALUATION EVALUATION

MEGA-
ENVIRONMENT
ANALYSIS

• Mega-environment is a group of geographical locations that share the same (set of)
best genotypes consistently across years.
Weikai Yan2006
 Relationships among environments
The “Environment-vector” view

• Angle vs.
correlation
• The angles
among test
environments
• Environment
grouping

Weikai Yan2006
 “Which-won-where”

G7
G18
G12

G13 G8

(Crossover GE is GE that caused genotype rank changes and different “winners” in

different test environments)
Weikai Yan2006
Are there meaningful crossover GE?
The “which-won-where” view

(Crossover GE is GE that caused genotype rank changes and different “winners” in

different test environments)
Weikai Yan2006
 Are the crossover patterns*
repeatable?
• If YES…
– The target environment can be divided into multiple
mega-environments
– GE can be exploited by selecting for each mega-
environment
– GE G
• If NO…
– The target environment CANNOT be divided into
multiple mega-environments
– GE CANNOT be exploited
– GE must be avoided by testing across locations and
years
• *Not the environment-grouping patterns
• Mega-environment is a group of geographical locations that share the same (set of) best
genotypes consistently across years.
• Multi-year data are needed
Weikai Yan2006
 Classify your target environment into
one of three categories

With Crossover GE No Crossover

GE
Repeatable (2) Multiple MEs (1) Single
Select for specifically adapted
genotypes for each ME
simple ME
A single test location,
single year suffices to
Not repeatable (3) Single select a single best
variety
complex ME
Select for generally adapted
genotypes across the whole
regions across multiple years

ME: mega-environment

Weikai Yan2006
G By E data analysis

TEST
GENOTYPE ENVIRONMENT
EVALUATION EVALUATION

MEGA-
ENVIRONMENT
ANALYSIS

Weikai Yan2006
Discriminating ability and representativeness
Vector length: discriminating ability
Angle to the AE: representativeness

Average-environment axis

Average environment

Weikai Yan2006
 Ideal test environments:
discriminating and representative

Ideal test
environment

Weikai Yan2006
 Classify each test environment into
one of three categories
Discriminative Not
discriminative

Representative (2) Good for (1) Useless

selecting (more
important)

Not (3) Useful for

representative culling (less important)

• For each “good” or “useful” test environment: is it essential?

Weikai Yan2006
Vector length = discrimination
= GE = GE1 + GE2

Contribution to
Proportionate
GE

Contribution to
Non-
proportionate
GE

Weikai Yan2006
G By E data analysis

TEST
GENOTYPE ENVIRONMENT
EVALUATION EVALUATION

MEGA-
ENVIRONMENT
ANALYSIS

Weikai Yan2006
 Vector length = GGE = G + GE

Contribution To GE
(instability)

Contribution To G
(mean performance)

Weikai Yan2006
Mean vs. Stability

Weikai Yan2006
Genotype ranking on both MEAN and STABILITY

“The ideal
genotype”

Weikai Yan2006
 Genotype classification
Mean High mean Low mean
Stability performance performance

High stability Generally adapted Bad everywhere

(VERY GOOD) (VERY BAD)

Low stability Specifically Adapted Bad somewhere

(GOOD) (BAD)

Are there stability genes?!

Weikai Yan2006
 G x E data analysis summary
• 1) Mega-environment analysis
• 2) Test environment evaluation
• 3) Genotype evaluation

Important comments:
– (2) and (3) are meaningful only for a single mega-environment
– Any stability analysis is meaningful only for a single mega-
environment
– Any stability index can be used only as a modifier to the ranking
based on mean performance

Weikai Yan2006
 Other ways to view
a GGE biplot

Contact: [email protected]
Inner-product property

Weikai Yan2006
Ranking on a single environment

Weikai Yan2006
Ranking on two environments

Weikai Yan2006
Relative adaptation of a genotype

Weikai Yan2006
Compare any two genotypes

Weikai Yan2006
 Biplot analysis of
Genotype by trait data

Contact: [email protected]
 Objectives of G By T data analysis

• Genotype evaluation based on trait

profiles
• Relationship among breeding objectives

Weikai Yan2006
 Data of 4 traits for 19 covered oat
varieties (Ontario 2004)

(Background info: High yield, high groat, high protein, and low oil are desirable for milling oats)
Weikai Yan2006
Relationships among traits

Weikai Yan2006
Trait profile of each genotype

Weikai Yan2006
Trait profile of a genotype

Weikai Yan2006
Trait profile comparison between two
genotypes

Weikai Yan2006
Genotype ranking based on a trait

Weikai Yan2006
Parent selection based on trait profiles

Weikai Yan2006
Independent culling

Weikai Yan2006
 Fuller understanding
of MET data
 MET data are more informative
than you thought

Contact: [email protected]
 A G-E-T 3-way dataset contains
various 2-way tables
• G by E data
• G by T data
• E by T data:
– for each genotype; all genotypes
• G by V data:
– each E-T as a variable (V)
• P by T data:
– each G-E as a phenotype (P)
• Genetic association by environment data
• Trait association by environment data

Weikai Yan2006
 Genetic-covariate by environment biplot
(QTL by environment biplot)

Barley
Genomics
Data

Weikai Yan2006
 Trait-association by environment biplot

Oat
MET
Data

Weikai Yan2006
 Four-way data analysis
• Year…

Weikai Yan2006
 Conclusions

Contact: [email protected]
 Conclusion (1)
• “GGE biplot analysis” is an effective tool
for G by E data analysis to achieve
understandings about….
1. the target environment,
2. the test environments, and
3. the genotypes
4. stability analysis is useful only to a single
mega-environment

Weikai Yan2006
 Conclusion (2)
• “GGE biplot analysis” is an effective tool
for G by T data analysis to achieve
understandings about….
1. the interconnected plant system,
2. positively correlated traits
3. negatively correlated traits
4. the strength and weakness of the
genotypes

Weikai Yan2006
 Conclusion (3)
• “Biplot analysis” is an effective tool for
other two-way table analysis
–Marker by environment
–QTL by environment
–Gene by treatment
–Diallel cross
–…

Weikai Yan2006
 Conclusion (4)
• Biplot analysis can be VERY EASY…
– From reading data to displaying the biplot: 2 seconds
– Displaying any of the perspectives of a biplot and
changing from one to another: 1 second
– Displaying the biplot for any subset: 1 second
– Learning how to use the software and interpret
biplots: 30 minutes
– Everything can be just one mouse-click away

Weikai Yan2006
 Thank you
Contact: Weikai Yan: [email protected]
web: www.ggebiplot.com

Contact: [email protected]