Artificial selection. Long before Darwin and Wallace, farmers and breeders were using the idea of selection to cause major changes in the features of their plants. We present a new analysis of the power of artificial selection experiments to detect and localize quantitative trait loci. This analysis uses a simulation framework. A simulation tool was developed for estimating the power to detect artificial selection acting directly on single loci. The simulation tool should be.
of The Selection Power Artificial
These characteristics are favored in this environment so these bacteria can continue to thrive, and make you ill. Different environments favor different traits and so natural selection has taken place! A timeline showing a decrease in the number of bacteria over time when taking an antibiotic. What is artificial selection or selective breeding?
There are other types of selection, in addition to natural selection, that are out there in the world. Think about some decisions you make about the types of pets you want or what kind of foods you prefer to eat. An example of artificial selection - Dog breeding. Around 30, to 40, years ago, humans began domesticating wolves.
Nowadays, these domesticated animals are what we call dogs! Domestication is the act of separating a small group of organisms wolves, in this case from the main population, and select for their desired traits through breeding. Over thousands of years, the domestication of wolves resulted in the loss of some of the more aggressive traits, like the instinctual, defensive behavior in the presence of humans barking or howling, bearing their teeth, poising to attack, or running away , and the size and shape of their teeth.
Now humans select for a variety of traits in dogs based on personal preference and companionship, instead of as a way to increase human survival. A timeline showing how dogs became domesticated over a long period of time due to artificial selection. Dog breeding is a perfect example of how humans select for desirable or fashionable traits. There are three different types of breeds that exist:. Purebred is a type of dog that comes from a lineage of the same dog breed and that has never mated with another breed.
For example, a purebred german shepherd is all german shepherd and nothing else. A cross-breed dog is a dog that was the offspring of two different types of purebreds.
The resulting offspring would be a cross-breed of half german shepherd, half husky. Finally, mixed-breeds are a combination of multiple breeds, where their parents were not purebreds. There are too many possible combinations to count! In purebreds, since there is only one lineage, these mistakes are often more apparent and can make purebred dogs prone to certain diseases.
An example of artificial selection - Genetically modified organisms. Recently we have started to artificially select traits at a molecular level where we mix DNA from different plant or animal species to make genetically modified organisms GMOs. To genetically modify an organism, genetic information or, the blueprint of the organism is added or removed, or replaced by the information from another organism that has a trait we desire.
If you could identify the genetic information that coded for drought resistance from another plant, then you could insert that into the blueprints of your corn species to make it more resistant to drought!
Cartoon showing how drought-sensitive corn is bred with drought-resistant corn to produce drought-resistant offspring. GMOs are used in agriculture to help crops become more resistant to drought, cold, salinity, pests and diseases. This is advantageous for us because it allows us to feed our growing population by doing agriculture in places that are usually less than ideal or not possible. With more areas to do agriculture, we have larger agricultural production to feed ourselves.
Common misconceptions about evolution. Evolution is not the same as adaptation or natural selection. Imagine a scenario where one trait might be highly advantageous in one environment, but highly detrimental in another. A good example of this is the fur color of mice. In the forest, it will be more likely that mice take on a darker color to match the earth. Note that the true region, consisting of sites within 10 kb of either SNP, is too small to be seen at this scale; thus, the size of the orange bars represents the local false-positive rate at this threshold.
Also note that the threshold of 0. A Twenty generations of selection; B 20 generations of selection, 4 generations per selection event 80 generations total. We calculated the false-positive rate in this way with radii of 10 kb, kb, and 1 Mb.
We found that using a radius of kb or 1 Mb led to true regions that covered a substantial portion of the genome in the cases where there were a large number of QTL.
Using akb radius for the detection region gave the most interpretable results. With regards to power, the simplest method is to calculate the proportion of QTL detected. However, we feel that a more relevant measure of power is the proportion of the genetic variance in the initial population that is explained by the detected QTL. This measure appropriately gives greater weight to QTL responsible for more of the genetic variance.
Figure 3, A and B illustrates the difference between these two measures of power: Comparison of three methods for calculating and interpreting power and false-positive rate. A Power is measured by the proportion of QTL detected, and false-positive rate is measured by the proportion of neutral variant sites detected. B Power is measured by the proportion of genetic variance in the founder population explained by the detected QTL, and false-positive rate is measured as in A.
C Power is measured as in B, and false-positive rate is measured by the proportion of the neutral genome covered by the detection region. Going forward, we use the kb radius for the detection region, together with measuring power as the proportion of variance explained, to provide the most interpretable results.
The forward simulator forqs efficiently simulates the entire genome of each individual by tracking haplotype chunks. In our simulation framework, we use two forward simulations.
The first simulation creates a mixed population from the founder haplotypes, after which we generate the random trait architecture. The second simulation represents the selection experiment. Individuals in the final populations are mosaics of individuals in the mixed population, which are in turn mosaics of the founders. We implemented a custom program to handle this two-step propagation of neutral variation. Given founder sequences, the mixed population, and the final population, the program calculates the allele frequency in the final population of each variant in the genome.
After calculating allele frequencies for each population, we calculate the allele frequency difference D for each high—low population pair under consideration multiple pairs for the replication analyses. After sorting the D values, we begin with the highest D value and iteratively decrease the threshold to obtain data equivalent to a receiver operating characteristic ROC curve power and false-positive rates for that simulation run.
We note that this step depends on the method for calculating power and false-positive rate, so we performed it once for each method we described above in Calculation of power and false-positive rate.
We obtain average ROC curves by calculating the average power over replicate simulation runs at regularly spaced false-positive rates, where the power for a particular run at a given false-positive rate is obtained by linear interpolation between points on its ROC curve. To investigate the power increase due to the use of haplotype-based allele frequency estimates, we needed to simulate pooled sequence reads from a population, followed by haplotype frequency estimation in sliding windows across the genome, using the harp method and software detailed in Kessner et al.
Because this procedure is computationally expensive, and due to the large number of simulations involved in this study, it was not feasible to do this for each simulated experiment. As an alternative, we obtained empirical error distributions, which we later used to add random errors to true allele frequencies. Because errors in the haplotype frequency estimation depend on the length scale of recombination, we ran replicate neutral simulations for varying numbers of generations from 40 to Haplotypes surrounding selected QTL are expected to be longer than in neutral regions, so our empirical error distributions are conservative.
Haplotype frequency estimation was performed with harp in overlapping sliding kb windows within a single 1-Mb region. We then calculated allele frequencies at variant sites within the region, using read counts. By considering allele frequencies in bins of size 0. Similarly, we also derived allele frequency estimates from the local haplotype frequencies, from which we obtained frequency-dependent empirical error distributions for each generation count.
We performed forward simulations of populations, using the program forqs Kessner and Novembre , which models whole genomes of individuals and selection on quantitative traits. Thus, initial allele frequencies of QTL were randomly distributed according to the allele frequency distribution of DGRP variant sites. Similarly, linkage disequilibrium patterns reflect the patterns observed in the DGRP populations. To simulate various trait architectures, we considered 12 scenarios by simulating 2, 5, 10, or QTL at initial heritability levels of 0.
Following previous theoretical and empirical studies Orr ; Otto and Jones ; Mackay ; Thornton et al. These scenarios span settings that are relatively straightforward for genetic mapping such as an oligenic trait with 2 QTL and a high heritability of 0. Our simulations began with several generations of neutral mixing, emulating laboratory procedures that use inbred founder lines to create larger experimental populations with increased genetic variation and reduced linkage disequilibrium e.
Using this procedure we simulated populations of various sizes, which we used as the initial populations for the artificial selection simulations. To first illustrate the importance of modeling quantitative traits explicitly, we show examples of allele frequency trajectories of a focal QTL contributing to a quantitative trait under truncation selection in comparison to a single locus with two alleles with a constant additive selection coefficient.
For these simulations, we depart from our general procedure described above and fix the starting allele frequency of a focal QTL to 0. The remaining QTL are chosen with a random distribution on starting allele frequencies and effect sizes see Methods. Selection is assumed to be in the direction of increasing trait values. As seen in Figure 4, A and C , the allele frequency trajectories of the focal QTL under truncation selection are qualitatively different from the trajectories under a constant selection coefficient.
While a strong selection coefficient leads to nearly deterministic allele frequency trajectories that monotonically increase, trajectories of a focal QTL under strong truncation selection are dependent on the underlying trait architecture. This effect can be seen in the trajectories where the focal QTL decreases in frequency at first, due to repulsion linkage disequilibrium i. Qualitative differences between fixed selection coefficient and truncation selection on a quantitative trait.
A focal QTL exhibits fundamentally different behavior under a constant selection coefficient, compared to truncation selection on a quantitative trait. Shown are allele frequency trajectories A and realized selection coefficient distributions B of a locus under two different selection coefficients 0. C and D show the same for a focal QTL effect sizes 0. In addition, once an allele with a constant selection coefficient reaches high frequency, it only gradually increases in the final generations before finally going to fixation.
In contrast, the focal QTL under truncation selection tended to become fixed quickly after reaching high frequency in the population. This behavior is not surprising, because after a few generations of selection, the upper tail of the population trait value distribution is highly enriched for individuals carrying high-effect variants. To further illustrate these qualitative differences, we analyzed the realized selection coefficient of the trajectories, which represents the selection coefficient that would result in a given single-generation allele frequency change under a deterministic model see Methods.
Under a constant selection coefficient, the mean realized selection coefficient tracks the true selection coefficient closely during the selection phase, after which it decreases to zero during the drift phase Figure 4B. Under truncation selection, the behavior of the mean realized selection coefficient depends on the underlying genetic architecture of the trait. When the effect size is low, the realized selection coefficient increases each generation—this is because selection acts on the larger-effect QTL first and then has a greater effect on the focal QTL after the larger-effect QTL have reached fixation.
On the other hand, when the effect size is higher, the focal QTL experiences very strong selection initially, decreasing as the focal QTL rises to fixation Figure 4D. Another effect of the underlying trait architecture can be seen in the fixation times of the focal QTL Figure 5.
For a given effect size and heritability, the fixation time of the focal QTL increases with total number of QTL due to interference. While these two forms of simulation explicit QTL simulation vs. Effect of genetic architecture on fixation times. Fixation times of a focal QTL for a trait under truncation selection decrease with increasing effect size and heritability.
In all of the following analyses, we examine the power to detect QTL through allele frequency differences at variant sites. There is currently no consensus regarding the choice of test statistic for the analysis of artificial selection experiments. Both for simplicity and because of its use in practice Parts et al. We calculate D for each variant site in the genome, and we call a site detected if the D value exceeds a threshold value.
By varying the threshold, we obtain ROC curves showing the relationship between power true positive rate and the false-positive rate. Due to linkage and strong selection, detected QTL will generally have neighboring neutral variants whose allele frequency differences also exceed the detection threshold. Because of this, detection and localization of a QTL are necessarily intertwined.
In an actual experimental setting, the entire genomic region surrounding the significantly diverged loci would often be chosen for follow-up studies. We explored several methods for calculating and interpreting power and false-positive rate. We present our results using the method that we found to be most interpretable, which we summarize here see Methods for full details on the different methods. For a given D -value threshold, we determine a detection region that consists of all variants within a specified radius of any variant above the threshold 10 kb for the results presented here; see Figure 2 for an illustration.
Power is calculated as the proportion of genetic variance in the founder population explained by the QTL located within the detection region. We define the true region to consist of all variants within the specified radius of any QTL i. The false-positive rate is calculated as the proportion of the neutral genome covered by the detection region.
Thus, a false-positive rate of 0. We note also that the false-positive rate represents the combined size of regions surrounding loci above the threshold; the detection region surrounding a single locus will be one to two orders of magnitude smaller than this i. Use of this technique presumes that allele frequency differences between the high and low lines will be more pronounced at QTL contributing to the trait than, for example, differences between the high line and a control population that has been evolving neutrally.
To compare the power obtained by a divergent selection experiment to the power obtained by selection in a single direction, we simulated three populations originating from a single founder population, where one population was selected for high values, one selected for low values, and one allowed to evolve neutrally. Comparison between the high and low populations leads to a substantial increase in power over the comparison between the high and neutral populations.
Increase in power due to divergent selection. Comparison of two populations divergently selected for extreme values of a trait has greater power to detect QTL than comparison between selected and control populations. Another technique available in artificial selection experiments is the use of replicate high and low populations to increase confidence that allele frequency differences between diverged populations are due to selection rather than genetic drift.
We then calculated the average power to detect QTL, using subsets of the data representing population replicates from one to five pairs. We found that using two replicate populations substantially increases power to detect QTL Figure 7.
For example, at the low false-positive rate neutral genome proportion of 0. Adding further replicate populations continues to increase power, but with diminishing returns.
Increase in power due to replicate populations. Adding replicate pairs of divergently selected populations increases power to detect QTL. It is well known that selection on a single locus, defined by a selection coefficient s , acts more efficiently in larger populations, as can be seen in the dependence of fixation probabilities and fixation times on the population-scaled selection coefficient 4 N s Ewens In all cases we found a substantial increase in power to detect and localize QTL as we increased the population size.
Increase in power due to population size. We next investigated the effects of the length of the experiment and the strength of selection on the power to detect QTL.
Each simulation ran for 80 generations, and we examined snapshots of each population at generations 20, 40, 60, and In these simulations, the populations consisted of individuals, and the trait had 10 QTL, with a heritability of 0. However, the lower effective population size induced by strong selection results in the fixation of many neutral variants, which are then falsely detected as QTL.
Effects of length of experiment and selection strength. In addition, letting the experiment run for a greater number of generations at this lower selection pressure increases the maximum power attained. These observations suggest that recombination plays a large role in the power to detect and localize QTL: Increased recombination will also reduce interference between QTL, which should allow lower-effect QTL to be selected and detected.
This led us to investigate the effects of recombination further in our subsequent analyses. To investigate the effects of recombination on the power to detect and localize QTL, we first considered the effect of increasing the recombination rate of the simulated individuals.
We found that increasing the recombination rate does indeed increase the power to distinguish QTL from neutral variants, for all trait architectures Figure 10, A and D. While in practice it is not feasible to experimentally increase the recombination rate of individuals, the experiment can be designed to increase the opportunity for individual chromosomes to recombine and thus decrease linkage between QTL and neighboring neutral variants.
The power to detect and localize QTL depends on the extent of recombination experienced by the populations. A and D Power increases with the recombination rate of the organisms under selection. B and E Additional generations of initial mixing and C and F additional generations between rounds of selection similarly increase power.
One way to decrease linkage disequilibrium in the population is to allow more generations of initial neutral mixing in the founder population, before selection starts e. We simulated scenarios with varying numbers of generations of neutral mixing 20, 40, 60, 80 , with the other parameters as above. We found that increasing the number of generations of initial neutral mixing increases power Figure 10, B and E. An alternative way to decrease linkage disequilibrium is to intersperse extra generations of neutral mixing between rounds of selection.
To test this idea, we simulated scenarios where selection was carried out every one, two, three, or four generations, with random mating during the generations where selection did not take place other parameters as above. We further illustrate the effect of extra recombination on allele frequency differences in Figure 2.
Taken together, these results show that the ability to detect and localize QTL depends crucially on recombination, both to decrease interference between QTL and to break up linkage between QTL and nearby neutral variants. In all of our power analyses, we have assumed that we are able to calculate the allele frequencies at all variant sites perfectly. However, in actual experiments, the allele frequency at a locus will be estimated by considering the sequence read counts that cover the site.
These estimates are prone to errors from two sources: The error in measurement of allele frequencies will lead to a loss of power to detect and localize QTL.
In the case where founder sequences are known, previous studies Long et al. This work suggested that improved allele frequency estimates could be derived from local haplotype frequency estimates. We investigated the effect of allele frequency estimation error on power by first obtaining empirical error distributions for the two estimation methods see Methods.
In all cases, we found that the haplotype-based allele frequency estimates led to improved power over the read-count-based estimates.
Using founder haplotype information estimation improves power. A If founder sequence information is available, estimating local haplotype frequencies leads to better allele frequency estimates and an increase in power over that of estimates obtained from raw read counts.
B Error in local haplotype frequency estimates increases with the number of generations due to recombination; however, there is a still a net gain in power from extra generations of neutral mixing.
Our analyses from the previous section showed that increasing the amount of recombination led to an increase in power. On the other hand, increased recombination results in shorter haplotypes. This leads to greater errors in local haplotype frequency estimates and the allele frequency estimates derived from them, with a subsequent decrease in power.
To investigate the relative magnitude of these two counteracting effects, we added two scenarios: We introduced random errors as above, using empirical error distributions that take into account the increased number of generations. Thus, we do not expect haplotype-derived allele frequency estimates to lead to increased power in scenarios where the typical length of haplotype chunks is shorter than the window size used for the local haplotype frequency estimation. We have presented a new analysis of the power of artificial selection experiments to detect and localize loci contributing to a quantitative trait.
In this analysis, we explicitly model whole genomes of individuals, quantitative traits, and selection based on individual trait values, using a novel simulation framework. We showed that population genetic simulations based on loci with constant selection coefficients do not fully capture the dynamics of QTL contributing to a trait under artificial selection and that the trait architecture plays a large role in these dynamics.
In addition, explicit modeling of selection on a quantitative trait has several other advantages. For example, simulated experiments can be parameterized and results can be reported using the standard quantitative genetics concepts of effect size, genetic variance, and heritability.
Also, scenarios such as divergent selection can be simulated in a straightforward manner. Finally, the behavior of a QTL under artificial selection is dependent both on the experimental design proportion of individuals selected each generation and on the trait architecture effect size and linkage to other QTL.
While the selection coefficient conflates these parameters into a single number, our simulation framework allows these parameters to be separated and investigated independently. Our results show the important role that recombination plays in the ability to identify QTL.
Recombination not only reduces interference between QTL, but also decreases linkage disequilibrium between QTL and neighboring neutral loci. In fact, one can view the artificial selection experiment as a classification problem. The classification viewpoint is useful when thinking about how to design a selection experiment.
For example, we found that experiments allowing more opportunity for recombination, either during initial neutral mixing of the founder haplotypes or between selection events, have greater power to detect and localize QTL.
Similarly, experiments with weaker selection greater proportion selected each generation over a longer period of time will have greater power than shorter experiments with strong selection. We note that the improved mapping resolution afforded by additional recombination is analogous to the use of recombinant inbred lines see Crow for a history or advanced intercross lines Darvasi and Soller in traditional QTL mapping studies.
While recombination increases the ideal power of an artificial selection experiment, it also decreases the ability to use founder haplotype information to obtain more accurate allele frequency estimates. We showed that the haplotype-based estimates still result in a net increase in power in experimental scenarios where the scale of recombination is larger than the window used for estimating local haplotype frequencies.
This suggests that when choosing the window size for such an analysis, one should take into consideration the expected scale of recombination based on the experimental design. Similar to the findings of Kofler and Schlotterer and Baldwin-Brown et al.
Additionally, our simulation framework allowed us to quantify the increase in power due to bidirectional selection.
We also note that adding generations of initial neutral mixing in the founder population is in some ways similar to increasing the number of founder haplotypes, in that it places QTL on multiple genetic backgrounds. Our results regarding the increase in power due to additional initial mixing are thus consistent with the findings of both of these groups that increasing the number of founder haplotypes increases power to detect and localize QTL.
Also in agreement with Baldwin-Brown et al. From the classification viewpoint again, power depends on the ability of the experiment to allow QTL to differentiate in selected populations while keeping allele frequencies at neutral loci constant. Continuing the experiment beyond the point where the majority of QTL have differentiated will lead to increased fixation of linked neutral loci and hence lower power. In contrast to both Kofler and Schlotterer and Baldwin-Brown et al.
Evolution Basics: Artificial Selection and the Origins of the Domestic Dog
Slow though the process of selection may be, if feeble man can do much by his powers of artificial selection, I can see no limit to the amount of change, to the. Artificial selection, also called "selective breeding”, is where humans select for . The distribution of citations of Academic papers is in a power law distribution (a. Artificial selection is a process of genetic modification of farm animal species that . Given the intrinsic power of this method to assess genetic pleiotropy, it might.