In recent times, a few colleagues at my Department and I have devoted some research effort to data management for diallel mating experiments, which we have summarised in a paper (Onofri et al., 2020) and a series of five blog posts (see here). A final topic that remains to be covered relates to the frequent possibility that these diallel experiments are repeated across years and/or locations. How should the resulting dataset be analysed?
We will start from a multi-environment full diallel experiment with 5 parental lines, in four blocks and 10 environments. The dataset is factitious and it was generated by Monte Carlo simulation, starting from the results reported in Zhang, et al. (2005; see here). The following box shows how we can load the data, after installing (if necessary) and loading the ‘lmDiallel’ package. In the same box, we use ‘dplyr’ to transform the explanatory variables into factors.
# library(devtools) # Install if necessary
# install_github("OnofriAndreaPG/lmDiallel")
library(lmDiallel)
library(dplyr)
dataset <- read.csv("https://www.casaonofri.it/_datasets/diallelMET.csv", header = T)
dataset <- dataset %>%
dplyr::mutate(across(c(Env, Block, Par1, Par2), .fns = factor))
head(dataset)
## Env Block Par1 Par2 Yield
## 1 1 1 1 1 10.78
## 2 1 1 1 2 12.78
## 3 1 1 1 3 15.23
## 4 1 1 1 4 10.66
## 5 1 1 1 5 14.42
## 6 1 1 2 1 11.84Fixed model fitting
For this full diallel experiment (withe selfs and reciprocals) we can fit the Griffing’s model 1, including the General combining Abilities (GCA), total Specific Combining Abilities (tSCA) and reciprocal effects (REC). We also include the environment effect, the block within environment effect and all interactions between genetical effects and the environment. If we regard all effects as fixed, we can code the final model either by using the lm() function, or by using the lm.diallel() wrapper in the ‘lmDiallel’ package, as shown in the box below. The two parameterisations are slightly different, although variance partitioning is totally equivalent.
dMod <- lm(Yield ~ Env/Block + GCA(Par1, Par2) + tSCA(Par1, Par2) +
REC(Par1, Par2) + GCA(Par1, Par2):Env +
tSCA(Par1, Par2):Env + REC(Par1, Par2):Env,
data = dataset)
# anova(dMod)
dMod2 <- lm.diallel(Yield ~ Par1 + Par2,
data = dataset, fct = "GRIFFING1",
Env = Env, Block = Block)
anova(dMod2)
## Analysis of Variance Table
##
## Response: Yield
## Df Sum Sq Mean Sq F value Pr(>F)
## Environment 9 10.6 1.17 0.1550 0.9978
## Env:Block 30 3195.1 106.50 14.0554 <2e-16 ***
## GCA 4 8673.7 2168.42 286.1657 <2e-16 ***
## GCA:Env 36 187.7 5.21 0.6882 0.9175
## tSCA 10 3021.3 302.13 39.8725 <2e-16 ***
## tSCA:Env 90 108.8 1.21 0.1596 1.0000
## Reciprocals 10 4379.7 437.97 57.7989 <2e-16 ***
## Reciprocals:Env 90 352.6 3.92 0.5170 0.9999
## Residuals 720 5455.8 7.58
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1Considering the ‘diallel’ object ‘dMod2, the full list of genetical parameters (in each environment) is retrieved by using the glht() function in the ’multcomp’ package. In the box below we show an excerpt of the output.
library(multcomp)
gh <- glht(linfct = diallel.eff(dMod2))
# summary(gh, test = adjusted(type = "none"))
# Estimate Std. Error t value Pr(>|t|)
# g_1:1 == 0 -3.67200 0.38929 -9.432 7.06e-10 ***
# g_2:1 == 0 -0.63850 0.38929 -1.640 0.113019
# g_3:1 == 0 1.28950 0.38929 3.312 0.002723 **
# g_4:1 == 0 0.03025 0.38929 0.078 0.938658
# g_5:1 == 0 2.99075 0.38929 7.682 3.75e-08 ***
# ts_1:1:1 == 0 1.73500 1.10109 1.576 0.127184
# ts_1:2:1 == 0 0.29150 0.80255 0.363 0.719379
# ts_1:3:1 == 0 -1.43150 0.80255 -1.784 0.086153 .
# ts_1:4:1 == 0 -2.10225 0.80255 -2.619 0.014504 *
# ts_1:5:1 == 0 1.50725 0.80255 1.878 0.071631 .
# ts_2:1:1 == 0 0.29150 0.80255 0.363 0.719379
# ts_2:2:1 == 0 1.03050 1.10109 0.936 0.357942 The ANOVA table above shows that genetical effects did not significantly interact with the environment and, therefore, we might be interested in getting estimates of average genetical effects, which can be done by using the glht() function and passing the type = "means" argument in the diallel.eff() call. An excerpt of the result is given in the box below.
gh <- glht(linfct = diallel.eff(dMod2, type = "means"))
# summary(gh, test = adjusted(type = "none"))
# Simultaneous Tests for General Linear Hypotheses
#
# Linear Hypotheses:
# Estimate Std. Error t value Pr(>|t|)
# g_1:1 == 0 -3.46259 0.12311 -28.127 < 2e-16 ***
# g_2:1 == 0 -0.55351 0.12311 -4.496 0.000127 ***
# g_3:1 == 0 0.64251 0.12311 5.219 1.89e-05 ***
# g_4:1 == 0 0.40521 0.12311 3.292 0.002868 **
# g_5:1 == 0 2.96838 0.12311 24.112 < 2e-16 ***
# ts_1:1:1 == 0 1.53104 0.34820 4.397 0.000165 ***
# ts_1:2:1 == 0 0.21247 0.25379 0.837 0.410125
# ts_1:3:1 == 0 -1.39756 0.25379 -5.507 8.87e-06 ***
# ....
# ....
# r_1:2:1 == 0 -0.20075 0.30776 -0.652 0.519943
# r_1:3:1 == 0 2.06700 0.30776 6.716 3.98e-07 ***
# r_1:4:1 == 0 -1.33550 0.30776 -4.339 0.000192 ***
# r_1:5:1 == 0 -3.97250 0.30776 -12.908 8.19e-13 ***
# ....
# ...Mixed model fitting
In most cases we might be willing to regard the environment effect as random, so that all the interactions between genetical effects and the environment are random as well. A mixed model can be fitted by using the mmer.diallel() wrapper, that is is available in the ‘lmDiallel’ package. The call is very similar to a lm.diallel() call, as shown in the box above; we can use the ‘type = “environment”’ argument to specify that we want to include random environment effects. The fit can take quite a few seconds (or minutes, depending on your device…). Fixed effects and variance components can be easily retrieved from the model object, as shown in the box below.
dMod3 <- mmer.diallel(Yield ~ Par1 + Par2,
data = dataset, fct = "GRIFFING1",
Env = Env, Block = Block, type = "environment")
dMod3$beta #fixed effects
## Trait Effect Estimate Std.Error t.value
## 1 Y (Intercept) 16.272390 0.2800709 58.1009636
## 2 Y GCA(Par1, Par2)g_1 -3.462590 0.1005471 -34.4374961
## 3 Y GCA(Par1, Par2)g_2 -0.553515 0.1005471 -5.5050326
## 4 Y GCA(Par1, Par2)g_3 0.642510 0.1005471 6.3901402
## 5 Y GCA(Par1, Par2)g_4 0.405210 0.1005471 4.0300520
## 6 Y tSCA(Par1, Par2)ts_1:1 1.531040 0.3246046 4.7166305
## 7 Y tSCA(Par1, Par2)ts_1:2 0.212465 0.2365942 0.8980143
## 8 Y tSCA(Par1, Par2)ts_1:3 -1.397560 0.2365942 -5.9069910
## 9 Y tSCA(Par1, Par2)ts_1:4 -1.866010 0.2365942 -7.8869632
## 10 Y tSCA(Par1, Par2)ts_2:2 1.946890 0.3246046 5.9977275
## 11 Y tSCA(Par1, Par2)ts_2:3 0.896115 0.2365942 3.7875606
## 12 Y tSCA(Par1, Par2)ts_2:4 -3.674085 0.2365942 -15.5290556
## 13 Y tSCA(Par1, Par2)ts_3:3 -0.187160 0.3246046 -0.5765784
## 14 Y tSCA(Par1, Par2)ts_3:4 1.134515 0.2365942 4.7951930
## 15 Y tSCA(Par1, Par2)ts_4:4 4.228940 0.3246046 13.0279727
## 16 Y REC(Par1, Par2)r_1:2 -0.200750 0.2869127 -0.6996903
## 17 Y REC(Par1, Par2)r_1:3 2.067000 0.2869127 7.2042832
## 18 Y REC(Par1, Par2)r_1:4 -1.335500 0.2869127 -4.6547268
## 19 Y REC(Par1, Par2)r_1:5 -3.972500 0.2869127 -13.8456774
## 20 Y REC(Par1, Par2)r_2:3 1.345250 0.2869127 4.6887092
## 21 Y REC(Par1, Par2)r_2:4 -1.277750 0.2869127 -4.4534460
## 22 Y REC(Par1, Par2)r_2:5 -2.285625 0.2869127 -7.9662747
## 23 Y REC(Par1, Par2)r_3:4 4.424875 0.2869127 15.4223768
## 24 Y REC(Par1, Par2)r_3:5 -0.038625 0.2869127 -0.1346229
## 25 Y REC(Par1, Par2)r_4:5 -2.149875 0.2869127 -7.4931342
dMod3$varcomp #variance components
## VarComp VarCompSE
## Env 0.00000000 0.42671998
## Env:Block 2.99662037 0.84276821
## GCA:Env -0.03826627 0.03707895
## tSCA:Env 0.00000000 0.14493517
## REC:Env 0.00000000 0.13004754
## Residuals 6.58550954 0.34464357Random model fitting
In other instances, our aim is to estimate variance components for all genetical effects and, therefore, we might like to regard all effects as random. This is easily done by changing the call to mmer.diallel() and replacing type = "environment" with type = "all".
dMod4 <- mmer.diallel(Yield ~ Par1 + Par2,
data = dataset, fct = "GRIFFING1",
Env = Env, Block = Block, type = "all")
dMod4$beta #fixed effects
## Trait Effect Estimate Std.Error t.value
## 1 Y (Intercept) 16.45538 1.920185 8.569682
dMod4$varcomp #variance components
## VarComp VarCompSE
## Env 0.00000000 0.42621427
## Env:Block 2.99608118 0.84185447
## GCA 4.11923254 3.41043097
## tSCA 4.68884763 2.14244233
## REC 5.39232126 2.44832836
## GCA:Env -0.03828058 0.03706453
## tSCA:Env 0.00000000 0.14494177
## REC:Env 0.00000000 0.13005751
## Residuals 6.58560305 0.34466898Obviously, a similar procedure can be used to fit all other diallel models to multi-environment diallel experiments.
Thanks for reading!
Andrea Onofri
Luigi Russi
Niccolò Terzaroli
Department of Agricultural, Food and Environmental Sciences
University of Perugia (Italy)
andrea.onofri@unipg.it
References
- Covarrubias-Pazaran, G., 2016. Genome-Assisted Prediction of Quantitative Traits Using the R Package sommer. PLOS ONE 11, e0156744. https://doi.org/10.1371/journal.pone.0156744
- Griffing, B., 1956. Concept of general and specific combining ability in relation to diallel crossing systems. Australian Journal of Biological Science 9, 463–493.
- Hayman, B.I., 1954. The Analysis of Variance of Diallel Tables. Biometrics 10, 235. https://doi.org/10.2307/3001877
- Zhang, Y., Kang, M.S., Lamkey, K.R., 2005. DIALLEL-SAS05: A Comprehensive Program for Griffing’s and Gardner-Eberhart Analyses. Agronomy Journal 97, 1097–1106. https://doi.org/10.2134/agronj2004.0260