Post date: Nov 15, 2013 8:54:0 PM
I re-ran the BSLMM analysis in gemma with two longer chains each with 2 million steps and a 500 thousand step burnin,
gemma -g geno_glaTrtAc.txt -p pheno_glaTraAc.txt -bslmm 1 -n 2 -o surv_glaTrtAc1 -rpace 20 -w 500000 -s 2000000
gemma -g geno_glaTrtAc.txt -p pheno_glaTraAc.txt -bslmm 1 -n 1 -o wgt_glaTrtAc1 -rpace 20 -w 500000 -s 2000000
gemma -g geno_glaTrtMs.txt -p pheno_glaTraMs.txt -bslmm 1 -n 2 -o surv_glaTrtMs1 -rpace 20 -w 500000 -s 2000000
gemma -g geno_glaTrtMs.txt -p pheno_glaTraMs.txt -bslmm 1 -n 1 -o wgt_glaTrtMs1 -rpace 20 -w 500000 -s 2000000
gemma -g geno_slaTrtAc.txt -p pheno_slaTraAc.txt -bslmm 1 -n 2 -o surv_slaTrtAc1 -rpace 20 -w 500000 -s 2000000
gemma -g geno_slaTrtAc.txt -p pheno_slaTraAc.txt -bslmm 1 -n 1 -o wgt_slaTrtAc1 -rpace 20 -w 500000 -s 2000000
gemma -g geno_slaTrtMs.txt -p pheno_slaTraMs.txt -bslmm 1 -n 2 -o surv_slaTrtMs1 -rpace 20 -w 500000 -s 2000000
gemma -g geno_slaTrtMs.txt -p pheno_slaTraMs.txt -bslmm 1 -n 1 -o wgt_slaTrtMs1 -rpace 20 -w 500000 -s 2000000
The results are in an output directory. My initial assessment is that results from the two chains are similar, but not quite as similar as I would like. Thus, I want to run longer chains and probably run three. At this point I will move to the dorc cluster for the analyse. Thus, I am moving the melGemma directory from Analyses to projects.
I wrote a script to summarize the genetic architecture results (summarizeGemma.pl). It basically combines the results from the two chains and provides posterior estimates. These preliminary results are in the file performGenArch.txt. They are also pasted below. I also have an R script to plot the results (plotArch.R).
character population plant parameter median q05 q95
surv gla Ac pve 0.1666624 0.0334592785 0.38894389
surv gla Ac pge 0.273896 0 0.900697805
surv gla Ac pi 0.00025573875 1.64558555e-05 0.0022379911
surv gla Ac n_gamma 21 0 182
surv gla Ms pve 0.3188502 0.104449215 0.572692565
surv gla Ms pge 0.2502141 0 0.884261355
surv gla Ms pi 0.0004768617 1.74836385e-05 0.00309096435
surv gla Ms n_gamma 38 0 250
surv sla Ac pve 0.8622349 0.7978201 0.9194744
surv sla Ac pge 0.9904566 0.9599446 0.999487
surv sla Ac pi 0.0002477627 0.0001789577 0.000338773
surv sla Ac n_gamma 24 18 29
surv sla Ms pve 0.10798045 0.0083019483 0.3405008
surv sla Ms pge 0.5019325 0 0.953210745
surv sla Ms pi 0.0001244681 1.5846332e-05 0.0018166082
surv sla Ms n_gamma 10 0 143
wgt gla Ac pve 0.37072775 0.086202737 0.649956325
wgt gla Ac pge 0.42366245 0.00580980245 0.94973699
wgt gla Ac pi 0.0003864875 1.9176272e-05 0.0024834994
wgt gla Ac n_gamma 30 1 194
wgt gla Ms pve 0.28257125 0.0251571375 0.720133975
wgt gla Ms pge 0.51652515 0.00627870765 0.960069295
wgt gla Ms pi 0.0002556043 1.8149086e-05 0.0023544626
wgt gla Ms n_gamma 20 1 181
wgt sla Ac pve 0.11865425 0.026708081 0.270771055
wgt sla Ac pge 0.26789645 0 0.8946898
wgt sla Ac pi 8.953675e-05 1.53336575e-05 0.00256071709999999
wgt sla Ac n_gamma 6 0 197
wgt sla Ms pve 0.44982005 0.033359743 0.9806428
wgt sla Ms pge 0.31036995 0 0.9084853
wgt sla Ms pi 0.0001858147 1.72807035e-05 0.00296390185
wgt sla Ms n_gamma 14 0 223
As with previous attempts to understand trait genetics, there is considerable uncertainty in parameters, particularly for adult weight which has a smaller sample size (dead larvae do not have adult weights). Nonetheless, there are a couple of interesting patterns. Most notably:
1. We explain a greater proportion of variance in survival in larvae reared on their natal host plant (GLA on Ms and SLA on Ac). This is particularly true for SLA on Ac (PVE > 0.8). This is counter to my original expectation that the narrow sense heritability of fitness components should be higher on a novel resource. In other words, consistent selection in an environment should remove variants that lower fitness in that environment. With that said, these individuals were reared in the lab, so perhaps this argument doesn't apply.
2. Almost all of the variation in survival for SLA on Ac that is explained by genetic variation is explained by individual SNV effects. In fact there are around 15-20 SNVs with very high posterior inclusion probabilities (> 0.5, several > 0.9). I have looked at these individual loci some. They tend to be loci with relatively low minor allele frequency such that there are only a handful of heterozygotes and most individuals are homozygous for the reference allele (the more common allele in these cases). In this population x treatment combination most individuals survived, but those that died are more likely to be heterozygous at these loci (thus rare heterozygotes are found in the few individuals that did not make it). This is interesting and actually consistent with expectations if selection acts against such variants (but not strongly enough so to get rid of them completely... perhaps the butterflies are just too drifty). I need to look at this more, but it is certainly interesting.