Genetics: Published Articles Ahead of Print, published on January 31, 2012 as 10.1534/genetics.111.137075
1
A test for selection employing quantitative trait locus and mutation accumulation data
Daniel P. Rice*,§§ and Jeffrey P. Townsend*,§
*Department of Ecology and Evolutionary Biology, and §Program in Computational Biology and Bioinformatics, Yale University, New Haven, CT 06520
§§Current address: Department of Organismic and Evolutionary Biology, Harvard University, 26 Oxford St., Cambridge, MA 02138
Copyright 2012. 2
Running Head: QTL, mutation accumulation, and selection Keywords: Phenotype / Selection / Statistical Tests / QTL / Neutrality
Corresponding Author:
Jeffrey P. Townsend, Department of Ecology and Evolutionary Biology, Yale University,
Osborne Memorial Labs 226, 165 Prospect Street, PO Box 208106, New Haven, CT
06520
Tel: (203) 432 4646
Fax: (203) 432-5176
Email: [email protected] 3
Abstract
Evolutionary biologists attribute much of phenotypic diversity observed in nature to the action of natural selection. However, for many phenotypic traits, and especially quantitative phenotypic traits, it has been challenging to test for the historical action of selection. An important challenge for biologists studying quantitative traits, therefore, is to distinguish between traits that have evolved under the influence of strong selection and those that have evolved neutrally. Most existing tests for selection employ molecular data, but selection also leaves a mark on the genetic architecture underlying a trait. In particular, the distribution of quantitative trait locus (QTL) effect sizes and the distribution of mutational effects together provide information regarding the history of selection. Despite the increasing availability of QTL and mutation accumulation data, such data have not yet been effectively exploited for this purpose. We present a model of the evolution of QTL and employ it to formulate a test for historical selection. To provide a baseline for neutral evolution of the trait, we estimate the distribution of mutational effects from mutation accumulation experiments. We then apply a maximum likelihood-based method of inference to estimate the range of selection strengths under which such a distribution of mutations could generate the observed QTL. Our test thus represents the first integration of population genetic theory and QTL data to measure the historical influence of selection. 4
Introduction
Identifying which quantitative traits have been subject to strong selection and which have evolved under neutral or nearly neutral conditions is a challenging and important task for evolutionary biologists (Boake et al., 2002). To this end, biologists are devoting increased attention to the genetic basis and evolutionary causes of quantitative variation (Barton and de Vladar, 2009, Barton and Keightley, 2002, Lai et al., 2007). On one hand, substantial progress has been made in revealing the genetic architecture of quantitative traits (Mackay, 2001, Mackay and Lyman, 2005, Brem and Kruglyak, 2005,
Lai et al., 2007, Verhoeven et al., 2004, Zimmerman et al., 2000, Ashton et al., 2001,
Gleason et al., 2009, Gleason et al., 2002). On the other hand, it has been difficult to connect the action of microevolutionary forces detected in studies of contemporary populations to their macroevolutionary effects (Grant and Grant, 2002). Consequently, few attempts to detect natural selection currently exploit the growing body of knowledge about the sizes and directions of quantitative trait locus (QTL) effects (for counter- examples, see Rieseberg et al., 2002, Albertson et al., 2003, and Lexer et al., 2005).
Progress toward understanding the basis of quantitative genetic variation is likely to come from studying allelic variation at specific QTL (Barton and Keightley, 2002).
Identifying the genetic architecture of quantitative traits begins with mapping QTL to broad genomic regions and ends with the molecular definition of quantitative trait loci alleles. This high degree of resolution has been achieved, for instance, for some QTL in
Drosophila (Palsson et al., 2005, Mackay, 2001, Zimmerman et al., 2000, Palsson and
Gibson, 2004). QTL mapping may be coupled with further genetic dissection (e.g. fine- 5 scale mapping, disequilibrium mapping, transgenic manipulation; Mackay, 2001) to characterize the specific loci targeted by selection. Weinig et al. (2003) demonstrated that the QTL architecture behind herbivory tolerance was one of many loci of small effect, and simultaneously elucidated locus-specific evolutionary constraints, demonstrating that linking molecular genetic tools to quantitative genetic analysis and field studies in ecologically relevant settings can clarify the role of specific loci in the evolution of quantitative traits.
These QTL studies not only generate maps from genotype to phenotype, but also yield information that may be applied to the study of phenotypic evolution (Erickson et al., 2004). Behind every genetically based phenotypic difference lies an evolutionary history. Despite the evident relevance of QTL data to the study of the evolutionary process, very little theory has been developed to link data from QTL studies to population genetic parameters via the methods of molecular evolution. This lack of theoretical tools has inhibited progress in explaining the process underlying the evolution of phenotypes that are genetically dissected in empirical QTL studies.
While most tests for selection depend directly on molecular data (McDonald and
Kreitman, 1991, Depaulis et al., 2003, Mousset et al., 2004, Nurminsky et al., 2005,
Schlenke and Begun, 2003, Schlenke and Begun, 2004, Schlenke and Begun, 2005), Orr
(1998a) presented two innovative attempts to integrate QTL data and population genetic theory into tests for selection: the QTL sign test and the QTL effect size test. These tests employ a null model of the evolutionary process that produces QTL. They have been applied multiple times to empirical data (Rieseberg et al., 2002, Albertson et al., 2003,
Lexer et al., 2005, Orr, 2010). However, the QTL sign test has been deemed 6 substantially flawed by its high false-positive rate (Anderson and Slatkin, 2003) and the
QTL effect size test has been shown to provide very low power to detect selection
(Anderson and Slatkin, 2003). Moreover, the tests neither model selection explicitly nor account for the distribution of mutational effects, both of which are important to accurately model the “filtering” action of natural selection (Orr 1998b).
Accordingly, understanding the evolutionary significance of QTL observations requires evaluation of both the origin and fate of heritable variation. In particular, the evolutionary forces that result in the genetic differences we observe must be estimated
(Barton and Keightley, 2002). Here we develop a test for a history of directional selection on quantitative traits and a method for estimation of the strength of that selection. Using mutation accumulation (MA) data and the QTL effect distribution derived from a cross between two divergent populations, we infer the selective history of the trait. While existing tests (Orr, 1998) only model neutral evolution and assume selection to be at work whenever data does not fit the model’s predictions under neutrality, we model selection explicitly and consider neutrality as one potential inference on a continuum. Such a flexible model is a key first step toward unraveling the complex link between molecular evolution and quantitative trait variation.
Theory
Overview: Our framework for inference integrates information from mutation accumulation experiment (Figure 1A-B), population genetic theory on selection (Figure
1G-J), the consequent expected outcome of selection (Figure 1K-N), and the empirically obtained QTL effect size distribution (Figure 1O-R). In our schematic example, mutation 7 accumulation is depicted as resulting in increased variance of the phenotype with no directional bias (Figure 1A), or, alternatively, as resulting in an increased variance in phenotype with a downward bias (Figure 1B). From phenotypic observations obtained along the time course of a mutation accumulation experiment (Figure 1A-B), the distribution of mutational effects on phenotype can be inferred (Figure 1C-F). Data as in
Figure 1A yields an inferred distribution of mutational effects that is symmetrical around no change (Figure 1C and Figure 1D); in contrast, data as in Figure 1B yields asymmetry
(Figure 1E and Figure 1F).
We multiply the mutation effect distribution by the calculated probability of fixation as a function of the strength of selection corresponding to the phenotypic effect
(Figure 1G-J). For a neutral phenotype, fixation probabilities are equivalent across all phenotypes (Figure 1G and Figure 1I), whereas under selection, fixation probabilities depend upon the phenotype conferred (Figure 1H and Figure 1J). The product of the mutation effect distribution and the fixation probability distribution is the probability distribution of fixed phenotypic differences resulting from substitutions at quantitative trait loci (Figure 1K-N). Sampling a number of substitutions from this distribution and summing their effects (e.g. Figure 1O-R, boxes), we compare the cumulative effects with those of the observed QTL (Figure 1O-R, bars). The accord between the sampled effects and the observed effects varies with the estimated mutation effect distribution and selective regime considered: the accord will be better when the level of selection is appropriate to generate fixed effects whose cumulative effects are close to those of the observed QTL. A likelihood function (Figure 1S-T) quantifies the fit of the model and provides a test of the hypothesis of neutrality and an estimate the strength of selection. In 8 our schematic example, the MA and QTL data together could support an effectively neutral model (Figure 1A,C,D,G,H,K,L,O,P,S). Alternatively, parameterizing with the same QTL data (Figure 1O-R, bars), but with different MA data (Figure 1A, B) could yield rejection of neutrality, and thus, a conclusion that the quantitative trait has been subject to a history of selection.
To develop this inferential framework, we first present a method for the estimation of the distribution of mutational effects with data from mutation accumulation experiments, which are a well-established approach to inferring the effects of mutations
(Halligan and Keightley, 2009, Keightley, 1994, Shaw et al., 2002, GarciaDorado, 1997).
In these experiments, one or more lines of the organism under study are kept in unselective conditions, subjected to serial population size bottlenecks (Lopez and Lopez-
Fanjul, 1993). This enforced genetic drift facilitates the accumulation of mutations with minimal interference from selection. A quantitative trait may be assayed at the end of the experiment or at regular intervals to track its change over time in each population. Using the change in the trait means and the number of generations between observations, we estimate a Gaussian distribution of mutational effects (Turelli 1984, Sawyer et al. 2003,
Jones et al. 2007).
We then derive the phenotypic effect fixation function from components of classical population genetic theory (Kimura, 1962). The product of this function with the mutational effect distribution provides a distribution of QTL effect sizes. Because these components are all probabilistic, comparison of their product to estimated QTL effect sizes and the standard errors thereof constitutes a likelihood-based approach to inferring the historical regime of genetic drift or natural selection. 9
We apply this test to an example trait, sensory bristle number in the fruit fly
Drosophila melanogaster. The genetic architecture of Drosophila sensory bristle number is well-studied (Dilda and Mackay, 2002, Lai et al., 1994, Long et al., 1995, Long et al.,
1996, Macdonald et al., 2005, Mackay and Lyman, 2005, Robin et al., 2002) and the selective history on this trait has been the subject of much debate (Mackay et al., 1994,
Kearsey and Barnes, 1970, Robertson, 1955, Robertson, 1967, Mather, 1941).
Estimating the mutational effect distribution: To estimate the distribution of the phenotypic effects of mutations on the trait, we analyze the time course of a mutation accumulation experiment. Although the appropriate formal distribution of the mutational effects for any given trait is unknown (Keightley and Lynch, 2003), moments of the distribution of mutation effects on a trait can be inferred from the changes in the trait over time. For analytical tractability, we follow Turelli (1984), Sawyer et al. (2003), and Jones et al. (2007) and assume the effects of mutations on a trait follow a Gaussian distribution
(x )2 1 2 f (x;, ) e 2 , (1) 2 2
with mean and standard deviation . This notation and other notation to be used in the derivation of the model are summarized in Table 1.
The sum of k independent normally distributed random variables with the same
2 mean and variance is normally distributed with mean k and variance k . Assuming independence of mutational effects, the distribution of the total effect of k mutations is
( xk )2 1 2 f (x, , ,k) e 2k . (2) 2 k 2 10
The distribution of the total effect of the mutations that accumulate in g generations is the sum of the distributions of the effect of k mutations, as k varies from zero to infinity, weighted by the probability that k mutations occur. If mutations are independent, the number of mutations affecting the trait that occur in g generations follows a Poisson distribution:
guk Pnumber of mutations k e gu , (3) k! where u is the per generation rate of mutations affecting the trait. The sum over k of
Equation 2 weighted by Equation 3 is
k (xk )2 gu 1 2 h(x, g,u; , ) e gu e 2k . (4) k! 2 k 1 2 k
Equation 4 is an infinite sum and cannot readily be calculated exactly. In practice, however, accurate approximations for small gu may be achieved by sums of only a small number of terms normalized by the total of all terms considered. We included terms so that, for a given gu, the total Poisson weight of the terms was at least 0.999.
Inference of historical selection from QTL and MA data: Inference of historical selection from QTL and MA data requires specification of the relationship between the strength of selection and the distribution of effect sizes of mutations that fix in the population. To calculate the probability that a mutation with a particular effect would fix, we apply the solutions of selection-diffusion equations. Assuming that these equations hold, the probability of a new mutation fixing in the population is given by
Kimura (1962):
1 es Pfixation , (5) 1 e2Ns 11 where s is the selection coefficient and N is the population size.
Following Lande (1983) and Chevin and Hospital (2008), we assume that the selective value of the mutation is proportional to its quantitative effect on the trait. To avoid making assumptions about the optimal value of the trait, we assume that the fitness gradient did not change as the value of the trait changed. This assumption stands in contrast to some existing models of adaptive walks (Orr 1998, Martin and Lenormand
2006) that use a Gaussian fitness landscape. It is also a conservative assumption for our purposes: a walk toward a nearby optimum on a peaked fitness landscape will fix mutations of smaller effect than in our model, diminishing the signature of selection that our method is sensitive to. Additionally, we assume that adaptation proceeds by successive fixation of beneficial mutations (Atwood et al. 1951), as opposed to simultaneous selection at many loci. If many loci are selected simultaneously, the strength of selection on a particular locus will tend to decrease over time (Chevin and
Hospital 2008), rather than remain constant. Hence, the assumption of successive fixation is also conservative. Furthermore, we assume that the mutation has no other effects on fitness beyond its effect on the trait in question. Thus, for a mutation with phenotypic effect x,
s cx , (6) where c is a constant whose sign and magnitude determine the direction and strength of selection, respectively. A positive value of c indicates selection for an increase in the trait and a negative value of c indicates selection for a decrease in the trait. When c = 0, the trait is neutral (i.e. the selection coefficient for a mutation of any effect size is zero).
Substituting Equation 6 into Equation 5 yields 12
1 e cx , c 0 2Ncx Pfixation 1 e , (7) 1 , c 0 2N which provides the probability that a mutation of size x will fix in the population.
The probability distribution of the phenotypic effects of substitutions is determined by the distribution of mutational effects multiplied by the relative fixation probability of those effects. Accordingly, multiplying Equation 1 by Equation 7 yields the distribution of effect sizes of substitutions, given that they have fixed:
(x)2 cx 1 e 1 2 f (x) C e 2 . (8) 1 e2Ncx 2 2
To make f(x) a probability distribution, Equation 8 contains a normalization constant, C, chosen to ensure that the integral of f(x) over all x is equal to one.
Methods
Overview of the data required: To infer a mutation effects distribution, we use the time course of a mutation accumulation experiment, specifically the change in trait value and number of generations between each measurement in each line. We also use an independent estimate of the per-generation mutation rate of the trait. To calculate probabilities of fixation of mutants drawn from the effect distribution, we also use the historical effective population size of the population in question. To construct a likelihood function for the parameters underlying the observed change in the quantitative trait, we use the effect size and standard error of QTL identified by crossing individuals from two divergent populations that differ in the trait value. 13
Mutation effects distribution: Moments of the distribution of mutation effects on a trait were inferred from the changes in the trait value over the course of a mutation accumulation experiment. Given the change in the trait means, the number of generations between observations, and an independent estimate of the per generation mutation rate for the trait, a likelihood function was specified for the parameters and
of the distribution of mutation effects as
L, ln hxi , gi ,u;, , (9) i where L(…) is the log likelihood, h(…) is from Equation 4, and xi and gi are the change in the mean value of the trait and number of generations between each observation. The values of and for which this function is maximized are the maximum likelihood estimates of the parameters of the Gaussian distribution of mutational effect sizes.
Estimating selection: We also applied a maximum likelihood framework to estimate the strength of historical selection on the trait. The phenotypic effect of each
QTL is the result of one or more underlying substitutions, so we maximize the joint likelihood of c and the vector H = (H1,H2,…,Hn), where Hi is the number of substitutions at the ith locus and can be any positive integer, and n is the number of loci. There is no explicit constraint the number of mutations at each locus, nor on the directions of the underlying mutational effects, only the constraint of the specified total effect. Thus, our method can properly accommodate amalgamations of several QTL that are detected as single QTL by a QTL analysis (c.f. MacKay, 2001, MacKay and Lyman, 2005, Studer and Doebley, 2011), even if the cryptic QTL are in opposing directions, just as it accommodates multiple mutations of opposing direction. The likelihood of c and H given 14 the QTL data is the product of the joint likelihood of c and H given each locus individually:
n lc, H;q,e l c, Hi ;qi ,ei . (10) i
Here q and e are the effects and standard errors of the observed QTL. The joint likelihood of c and H given a single locus is
lc, Hi ;qi ,ei g x;qi ,ei F x;c, Hi dx , (11) where g(…) is the probability density function of a Gaussian distribution with mean qi and standard deviation ei, and F(…) is the probability density function of the sum of Hi substitutions under strength of selection c. The latter is calculated by Hi convolutions of f(x) from Equation 8 with itself.
The integral in Equation 11 has no closed form solution and is difficult to evaluate by typical numerical approaches for most values of Hi. We effectively approximated the likelihood by Monte Carlo sampling from the distribution given by F(x) and calculating average value of g(x) evaluated for each sampled effect:
x q 2 j i 2 1 2ei lc, H i ;qi ,ei gx j ;qi ,ei e . (12) j 2 2ei j
Angle brackets represent averaging over all Monte Carlo samples xj. As the number of
Monte Carlo samples increases, the right hand side of Equation 12 approaches the true likelihood.
To calculate the likelihood of a given value of c, we fixed c at a value and evaluated the right hand side of Equation 12 for a range of vectors of H. For each locus i, we started with Hi = 1, then increased Hi until we found its unimodal maximum 15 likelihood value. We took the likelihood of c to be the product over i of these likelihoods
(Equation 10). We calculated the likelihood of a dense grid of values of c in this way to obtain ĉ, the maximum likelihood value. To test the hypothesis of neutral evolution, we performed a likelihood ratio test with a null hypothesis of c = 0 and an alternate hypothesis of c = ĉ.
Sources of data: We applied our method to QTL and MA data for sensory bristle number in Drosophila melanogaster. QTL effect size data was extracted from Dilda
(2002), who had applied four different artificial selection regimes on bristle traits.
Populations HST and LST were selected for high and low numbers of sternopleural bristles respectively, while populations HAB and LAB were selected for high and low numbers of abdominal bristles. The QTL effects and standard errors from each population are presented in Figures 3A, 4A, 5A, and 6A. These QTL were assayed by crossing the artificially selected populations with a wild-type laboratory strain. Because the QTL for this trait were assayed using artificially selected strains rather than phenotypically divergent strains from wild populations, our example serves to illustrate the use of the test rather than to make an inference about the selective history of the trait in the wild.
Mutation accumulation data for sternopleural bristle number and abdominal bristle number were extracted from Paxman (1957) and Lopez and Lopez-Fanjul (1993), respectively. Paxman (1957) propagated six lines of flies to accumulate mutations for forty generations and measured the average number of sternopleural bristles in each line at five time points. The distribution of trait changes in the mean sternopleural bristle number of each line between each measurement is depicted in Figure 2A. Lopez and 16
Lopez-Fanjul (1993) maintained ninety-three MA lines for sixty-one generations, and measured the change in the number of abdominal bristles in each line at the end of the experiment. The distribution of trait changes is depicted in Figure 2B.
The per generation rate of mutation affecting sensory bristle number in
Drosophila is estimated to be 0.03 (Mackay and Lyman, 2005).
Estimating the distribution of mutational effects: To estimate the mean and standard deviation of the distribution of mutation effects μ and σ, we calculated their maximum likelihood values for each type of bristle. For each bristle trait, we evaluated
Equation 9, letting u = 0.03 (Mackay and Lyman, 2005). We let xi and gi be the trait changes and numbers of generations between measurements for all measurements in all lines. Implementing a simple hill-climbing algorithm in MATLAB, we calculated the joint maximum likelihood estimates for and (See Supplementary Materials for
MATLAB code). We estimated the standard deviations of our estimates by finding the values of and whose log likelihood was equal to the maximum log likelihood minus one (Hudson 1971).
Estimating selection: For each of the four QTL data sets, we calculated the likelihood of a range of values of c using an effective populations size N = 50 (Gurganus et al. 1999) and the maximum likelihood estimates of and for the appropriate bristle
4 trait. For each value of c and Hi, we generated 6.5x10 Monte Carlo samples from the distribution of substitution effects in MATLAB and calculated the likelihood according to Equation 12 (See supplementary materials for MATLAB code). To calculate the significance of the likelihood ratio test, we computed the test statistic 17
lc 0 D 2log . (13) lc cˆ
D is approximately chi-squared distributed with one degree of freedom (Hudson 1971).
It follows that the 95% confidence interval includes values of c with likelihoods within two log units of the maximum.
False positive rate: To estimate the false positive rate of the test, we simulated
QTL data under a neutral model of evolution. For each of the four data sets, we found the maximum likelihood value of H for c = 0. For the ith observed locus, we generated a neutral locus by summing Hi effects drawn from the distribution of mutation effects for the appropriate bristle trait. Although it entailed extensive computation, we generated
100 sets of simulated neutral QTL for each data set. We applied the likelihood ratio test for selection to these simulated data and recorded the proportion of neutral data sets for which the neutral hypothesis was rejected.
Undetected QTL: A well-documented bias in QTL studies is the failure to detect small-effect loci (Beavis 1994, Bost et al. 2001, Xu 2003). To assess whether undetected
QTL might have influenced our estimate, we applied the analytical approach of Otto and
Jones (2001) to estimate the number and average effect of undetected loci in each data set. We then reran our analysis including in each data set the estimated number of additional QTL, with their effect sizes following an exponential distribution with the estimated mean.
Sensitivity to the mutation effect distribution: To assess the sensitivity of our test to the estimates of the mutation effect distribution parameters, we perturbed μ by one standard deviation in each direction and applied the test. We then similarly perturbed σ and compared the resulting estimates and confidence intervals to our primary results. 18
Power: To estimate the power of the test with a diverse set of putatively realistic
QTL distributions, we applied the test to all subsets of the QTL from each of the four populations, using the appropriate mutation effect distribution parameters in each case.
We examined how the proportion of times the test rejected neutrality varied with the number of loci included in the subset.
Results
The distribution of mutation effects: We estimated the maximum-likelihood mean and standard deviation of the Gaussian distribution of mutational effects (Equation
9) based on the MA data of each type of sensory bristle. For sternopleural sensory bristles (Paxman 1957), the maximum likelihood estimate (MLE) of μ was -0.08±0.06 (±
1SD) and the MLE of σ was 0.23±0.05. For abdominal bristles (Lopez and Lopez-Fanjul
1993), our estimate of μ was -0.25±0.04 and our estimate of σ was 0.34±0.05. We calculated the MLE ĉ for the four QTL data sets, in each case applying the corresponding mutation effect distribution.
Selection in the HST population: Based on the nine QTL detected in the HST population (Figure 3A) and the mutation effect distribution estimated for sternopleural bristle number (Figure 3B), the likelihood ratio test (χ2 = 38.1, P < 0.001) rejects a neutral hypothesis for an increase in sternopleural bristle number. The MLE for c, ĉ, was 0.018, with a 95% confidence interval (CI) ranging from .013 to .022 (Figure 3C). None of the corresponding one hundred simulated neutral data sets based on the HST data yielded a false positive. 19
Applying Otto and Jones (2000) to assess whether undetected QTL might have influenced our HST result yielded an estimate that four loci were not detected, and that these loci had an average effect of 0.388. Including four simulated QTL following an exponential distribution with mean 0.388, the MLE ĉ remained 0.018 and the CI ranged from .013 to .022 (Figure 3D), the same result as in our main analysis.
To assess the sensitivity of our test to the estimates of the mutation effect distribution parameters, we perturbed μ by one standard deviation in each direction and applied the test. For μ = -0.144, ĉ = 0.027 with a CI ranging from 0.021 to 0.031, while for μ = -0.022, ĉ = 0.012 with a CI ranging from 0.007 to 0.016 (Figure 3E). Similarly perturbing σ, we found that for σ = 0.183, ĉ = 0.029, with a CI ranging from 0.025 to
0.034, and for σ = 0.288, ĉ = 0.005, with a CI from 0.001 to 0.01 (Figure 3F). Because confidence intervals for ĉ did not include zero, rejection of the hypothesis of neutrality underlying the evolution of the HST QTL data is robust to moderate perturbations in μ and σ.
Selection in the LST population: Based on the four QTL detected in the LST population (Figure 4A) and the mutation effect distribution estimated for sternopleural bristle number (Figure 4B), a likelihood ratio test (χ2 = 4.88, P = 0.027) rejects a neutral hypothesis for a decrease in sternopleural bristle number. In this case, the log likelihood curve does not have a maximum, but instead increases asymptotically as ĉ decreases
(Figure 4C). Conservatively using the largest sampled likelihood as the maximum likelihood and applying the likelihood ratio test, we found that ĉ < -0.005 with 95% confidence. None of the one hundred simulated neutral data sets based on the LST data yielded a false positive. 20
We estimated that four loci were not detected and that these loci had an average effect of -0.284. When we included these simulated QTL in the data, we again found ĉ < -
0.005 (Figure 4D). Perturbing μ, we found that for μ = -0.144, ĉ < 0.007, and for μ = -
0.022, ĉ < -0.017 (Figure 4E). Perturbing σ, we found that for σ = 0.183, ĉ < 0.001, and for σ = 0.288, ĉ < -0.007 (Figure 4F). Because confidence intervals for ĉ included zero, rejection of the hypothesis of neutrality underlying the evolution of the LST QTL data was not robust to moderate perturbations in μ and σ.
Selection in the HAB population: Based on the three QTL detected in the HAB population (Figure 5A) and the mutation effect distribution estimated for abdominal bristle number (Figure 5B), a likelihood ratio test (χ2 = 34.7, P < 0.001) rejects the neutral hypothesis for an increase in the number of abdominal bristles. As in the LST population, the log likelihood curve does not have a maximum, instead increasing asymptotically as ĉ increases (Figure 5C). Conservatively using the largest observed likelihood as the maximum likelihood and applying the likelihood ratio test, we found that ĉ > 0.031 with
95% confidence. Only one of the one hundred simulated neutral data sets based on the
HAB data yielded a false positive.
We estimated that five loci were not detected and that these loci had an average effect of 0.245. When we included these simulated QTL in the data, we found that ĉ >
0.037 (Figure 5D). Perturbing μ, we found that for μ = -0.294, ĉ >0.037, and for μ = -
0.215, ĉ > 0.026 (Figure 5E). Perturbing σ, we found that for σ = 0.30, ĉ > 0.034, and for
σ = 0.392, ĉ > 0.029 (Figure 5F). Because confidence intervals for ĉ did not include zero, rejection of the hypothesis of neutrality underlying the evolution of the HAB QTL data is robust to moderate perturbations of μ and σ. 21
Selection in the LAB population: Based on the three QTL detected in the LAB population (Figure 6A), and the mutation effect distribution estimated for abdominal bristle number (Figure 6B), a likelihood ratio test (χ2 = 4.00, P = 0.046) rejects a neutral hypothesis for a decrease in abdominal bristle number (Figure 6C). Conservatively using the largest observed likelihood as the maximum likelihood and applying the likelihood ratio test, we found that ĉ < 0 with 95% confidence. Only one of the one hundred simulated neutral data sets based on the LAB data yielded a false positive.
We estimated that thirteen loci were not detected and that these loci had an average effect of -0.240. When we included these simulated QTL in the data, ĉ was 0 with a 95% CI ranging from -0.032 to 0.015 (Figure 6D). Perturbing μ, we found that for
μ = -0.294, ĉ < 0.006, and for μ = -0.215, ĉ < -0.004 (Figure 6E). Perturbing σ, we found that for σ = 0.30, ĉ < 0.003, and for σ = 0.39, ĉ < -0.001 (Figure 6F). Because confidence intervals for ĉ included zero, rejection of the hypothesis of neutrality underlying the evolution of the LAB QTL data is not robust to perturbations in μ and σ.
Power: We examined the power of subsets of the observed QTL to yield a positive result. The proportion of cases where neutrality could be rejected increased monotonically with the number of QTL in the subset (Figure 7). For single QTL and the observed mutation effect distribution, the test rejected neutrality in only 43% of cases, while for data sets with six or more of these loci, the test always rejected neutrality.
Discussion
We have derived a new method for detecting the level of historical selection on quantitative traits. First, we showed how to estimate the parameters of a Gaussian 22 distribution of mutational effects from the change in the trait over the course of a mutation accumulation experiment. Then, we derived the distribution of the effect sizes of mutations that fix as a function of the regime of selection or neutrality applied to the trait. We demonstrated how to integrate these two results to yield a maximum likelihood estimate of the strength of historical selection given a set of observed QTL effect sizes and the results of a mutation accumulation experiment. Finally, we applied the test to a well-studied quantitative trait, detecting a history of artificial selection on four cases of sensory bristle numbers in D. melanogaster.
The test we have derived is the first test for selection on phenotype that incorporates information about a trait’s genetic architecture as well as the phenotypic change induced by mutation. It is also the first test to include an explicit model of how neutrality and selection shape the loci that affect a trait. Finally, our test is the first that will allow biologists to not only distinguish between neutrally evolving traits and selected traits but also to estimate the range of strengths of selection that are most likely to have produced the current phenotypic value. As such, it represents a theoretical and practical improvement on existing methods to detect selection on quantitative traits.
Application of the test to sensory bristle number in D. melanogaster demonstrated its ability to detect a history of selection. Likelihood ratio tests demonstrated sufficient power to reject neutrality in four out of four populations that had been subject to artificial selection. For the HST data, we were able to identify a MLE for a coefficient determining the strength of selection, as well as to describe its 95% CI. For the other three data sets, we were able to estimate upper or lower bounds for the strength of selection. Because the
HST data contained the largest number of QTL (9), this difference in precision is 23 expected. The HST data set was also the only one that contained some QTL with effects opposite to the actual and inferred direction of selection. Similarly, by applying the test to subsets of the data, we demonstrated that the test’s power increased with an increased number of QTL. We also assessed the false positive rate by applying the test to one hundred simulated neutral data sets based on each set of real QTL. In no case did more than one of the one hundred simulated data sets yield a false positive.
We assessed the sensitivity of our test to the estimates of the mutation effect distribution parameters by perturbing μ and σ by one standard deviation in each direction and re-applying the test. For the HST and HAB data, rejection of neutrality was robust to these perturbations. However, for the LST data and the LAB data, the result was no longer significant in the case of a more negative μ and in the case of a smaller σ. This loss of significance speaks to a perhaps unsurprisingly higher sensitivity of the test in cases where selection and mutation are acting in opposite directions (as they are in HST and
HAB), than in cases where selection and mutation are acting in the same direction. Our low false positive rate across both cases in which selection and mutation are acting in concert or in conflict indicates that this difference in sensitivity is in accord with the actual inferential difficulty and not a problem of a biased test construction.
A well-documented bias in QTL studies is the tendency to fail to detect small- effect loci (Beavis 1994, Bost et al. 2001, Xu 2003). To assess whether undetected QTL might have influenced our estimates, we applied the analytical approach of Otto and
Jones (2000) to estimate the number and effect of undetected QTL, and included simulated loci following these parameters in our analysis. For the three data sets where the test rejected neutrality, this modification did not change the significance of the result 24 and had little effect on the estimates of c. However, when we included the simulated loci in the LAB data set, where we had already bordered on insufficient power to detect the action of selection, the MLE for c was shifted to 0. For this data set in particular, the analysis indicated that there were many more undetected loci (13) than observed QTL
(5). Furthermore, the simulated loci had a mean of -0.240, almost identical to the estimate of μ for abdominal bristles, making them effectively neutral for the purposes of our test.
Therefore, application of Otto and Jones (2000) imputation can have a considerable impact when the expected number of unobserved loci is much larger than the number of detected QTL.
Wider application of this test awaits the availability of more MA and QTL data.
Unfortunately, there are few traits for which both MA and QTL analysis have been performed. Practical and ethical considerations limit the range of species for which mutation accumulation experiments may be performed. On the other hand, numerous examples of applications of Orr’s 1998 test to empirical data (Rieseberg et al., 2002,
Albertson et al., 2003, Lexer et al., 2005, Orr, 2010) demonstrate that tests for selection on phenotypic traits are in demand. Although both types of experiments can be labor- intensive and time-consuming, there are many cases where either MA or QTL data are already available, and only the corresponding partner experiment need be completed to apply this test for historical selection.
Several expansions of our model would improve its scope and precision. First, the model presented here implicitly rather than explicitly incorporates time. Because the generation and fixation of mutations take time, the amount of time a trait has had to evolve is an important factor in how many substitutions are responsible for each QTL 25 effect. Explicitly incorporating time into the model should provide more information to determine which vector H is most likely. Since H affects the likelihood of c, this additional discriminatory power would yield more precise estimates of the strength of selection. However, explicit rather than implicit incorporation of time in our framework is complex.
Our test assumes that the probability of fixation of a new mutation follows
Equation 5, which has been derived from various models of evolution (Kimura 1962,
Moran 1960, Gillespie 1974), and is a fairly general result, accommodating a variety of evolutionary scenarios for which the effective population size captures all relevant complications. Examples include cyclically varying population size (Otto and Whitlock
1997), population subdivision with symmetric migration (Pollack 1966), and deleterious mutations at completely linked loci (Charlesworth 1994). Probability of fixation results that could be used in place of Equation 5 under more diverse evolutionary scenarios are reviewed by Patwa and Wahl (2008).
Because we model the sequential mutation and fixation of alleles at different loci, our test applies to QTL identified by crossing related populations and estimates the selective regime responsible for the trait difference between the populations. The method is not applicable, as currently implemented, to QTL identified by crossing individuals from the same population. However, since selection should shape the standing variation within a population, a related test that applies to QTL variation in a single population is conceivable. Accordingly, one potential weakness of the test is that our model assumes that QTL between the divergent populations are constructed from novel mutations, without allowing some proportion of fixed differences to arise from standing variation in 26 the ancestral population. QTL arising from standing variation fix faster and are likely to be of smaller effect size than QTL arising from novel mutations (Barrett and Schluter,
2008). Consequently, our model assumption that all variation arises as new mutation is conservative for the purpose of our test: within our framework, smaller QTL effect sizes will be more consistent with neutrality.
Finally, distributions of mutational effect other than the Gaussian distribution used here should be explored. Implementing the model with other distributions would allow assessment of its sensitivity to this assumption and provide flexibility to its use as new information comes to light about the distributions of mutational effects found in nature. The central challenge in this endeavor is that other distributions lack the convenient property of the normal distribution that the sum of normally distributed random variables is also normally distributed. As currently implemented, our approach relies on this fact for tractability (Equation 2).
As a theoretical result, this test for selection on quantitative traits is important because it bridges population genetic theory and genetic architecture as revealed by QTL.
Furthermore, the test for selection outlined here provides an improvement over currently available methods for determining the regime of historical selection or neutrality that generated observed quantitative differences in phenotype. Results of the test applied to four data sets known to be the result of artificial selection for sensory bristle number in
D. melanogaster are highly encouraging, because they indicate that the test has sufficient power to be used on moderately sized QTL data sets, and also that the false positive rate is very low. Applying the test to QTL data generated by crossing two individuals representative of diverged wild populations will yield conclusions about the history of 27 selection responsible for the observed trait differences. Finally, our model provides a novel framework that holds promise for expansion in future work to incorporate such factors as time, varying probabilities of fixation, and diverse shapes of the distribution of mutational effects.
Acknowledgments
We thank Trudy Mackay and Carlos Lopez-Fanjul for graciously and helpfully digging deep in their files to provide key details of their published experiments. DPR was partially funded while performing this work by a Yale College Dean’s Research
Fellowship, a Yale-HHMI Future Scientist Summer Fellowship, and a National Science
Foundation Graduate Research Fellowship DGE-1144152. The final computations were run on the Odyssey cluster supported by the FAS Science Division Research Computing
Group at Harvard University. 28
Literature Cited
ALBERTSON, R., STREELMAN, J. & KOCHER, T. 2003. Genetic basis of adaptive shape differences in the cichlid head. Journal of Heredity, 291-301. ANDERSON, E. & SLATKIN, M. 2003. Orr's quantitative trait loci sign test under conditions of trait ascertainment. Genetics, 445-446. ASHTON, K., WAGONER, A., CARRILLO, R. & GIBSON, G. 2001. Quantitative trait loci for the monoamine-related traits heart rate and headless behavior in Drosophila melanogaster. Genetics, 283-294. ATWOOD, K. C., SCHNEIDER, L. K., & RYAN, F. J. 1951. Periodic selection in Escherichia coli. Proceedings of the National Academy of Sciences of the United States of America, 37(3), 146-55. BARRETT, R. D. H., & D. SCHLUTER. 2008. Adaptation from standing genetic variation. Trends in Ecology and Evolution 23(1), 38-44. BARTON, N. & DE VLADAR, H. 2009. Statistical Mechanics and the Evolution of Polygenic Quantitative Traits. Genetics, 997-1011. BARTON, N. & KEIGHTLEY, P. 2002. Understanding quantitative genetic variation. Nature Reviews Genetics, 11-21. BEAVIS, W. D., 1998 QTL analyses: power, precision and accuracy, pp. 145–162 in Molecular Dissection of Complex Traits, edited by A. H. Paterson. CRC Press, Boca Raton/New York BOAKE, C., ARNOLD, S., BREDEN, F., MEFFERT, L., RITCHIE, M., TAYLOR, B., WOLF, J. & MOORE, A. 2002. Genetic tools for studying adaptation and the evolution of behavior. American Naturalist, 160, S143-S159. BOST, B., VIENNE, D. de, HOSPITAL, F, MOREAU, L., & DILLMANN, C. 2001. Genetic and Nongenetic Bases for the L-Shaped Distribution of Quantitative Trait Loci Effects. Genetics, 157(4), 1773-1787. BREM, R. B. & KRUGLYAK, L. 2005. The landscape of genetic complexity across 5,700 gene expression traits in yeast. Proceedings of the National Academy of Sciences of the United States of America, 102, 1572-1577. CHARLESWORTH, B. 1994. The effect of background selection against deleterious mutations on weakly selected, linked variants. Genetics Research Cambridge 63, 213-227. CHARLESWORTH, B. 2009. Effective population size and patterns of molecular evolution and variation. Nature Reviews Genetics, 195-205. CHEVIN, L.-M., & HOSPITAL, F. 2008. Selective sweep at a quantitative trait locus in the presence of background genetic variation. Genetics, 180(3), 1645-60. DEPAULIS, F., MOUSSET, S. & VEUILLE, M. 2003. Power of neutrality tests to detect bottlenecks and hitchhiking. Journal of Molecular Evolution, S190-S200. DILDA, C. 2002. The genetic architecture of Drosophila sensory bristle number. Ph.D. dissertation, North Carolina State University. In Dissertations & Theses: Full Text [database on-line]; available from http://www.proquest.com (publication number AAT 3346505; accessed November 22, 2011). DILDA, C. & MACKAY, T. 2002. The genetic architecture of drosophila sensory bristle number. Genetics, 1655-1674. 29
ERICKSON, D., FENSTER, C., STENOIEN, H. & PRICE, D. 2004. Quantitative trait locus analyses and the study of evolutionary process. Molecular Ecology, 2505- 2522. GARCIADORADO, A. 1997. The rate and effects distribution of viable mutation in Drosophila: Minimum distance estimation. Evolution, 51, 1130-1139. GILLESPIE, J. H. 1974. Natural selection for within-generation variance in offspring number. Genetics 76, 601-606. GLEASON, J., JAMES, R., WICKER-THOMAS, C. & RITCHIE, M. 2009. Identification of quantitative trait loci function through analysis of multiple cuticular hydrocarbons differing between Drosophila simulans and Drosophila sechellia females. Heredity, 416-424. GLEASON, J., NUZHDIN, S. & RITCHIE, M. 2002. Quantitative trait loci affecting a courtship signal in Drosophila melanogaster. Heredity, 1-6. GRANT, P. & GRANT, B. 2002. Unpredictable evolution in a 30-year study of Darwin's finches. Science, 707-711. GURGANUS, M. C., NUZHDIN, S. V., LEIPS, J. W., & MACKAY, T. F. C. 1999. High-Resolution Mapping of Quantitative Trait Loci for Sternopleural Bristle Number in Drosophila melanogaster. Genetics, 152(4), 1585-1604. HALLIGAN, D. L. & KEIGHTLEY, P. D. 2009. Spontaneous Mutation Accumulation Studies in Evolutionary Genetics. Annual Review of Ecology Evolution and Systematics, 40, 151-172. JONES, A., ARNOLD, S. & BURGER, R. 2007. The mutation matrix and the evolution of evolvability. Evolution, 727-745. KEARSEY, M. J. & BARNES, B. W. 1970. Variation for metrical characters in Drosophila populations .2. Natural selection. Heredity, 25, 11-&. KEIGHTLEY, P. & LYNCH, M. 2003. Toward a realistic model of mutations affecting fitness. Evolution, 683-685. KEIGHTLEY, P. D. 1994. The distribution of mutation effects on viability in Drosophila melanogaster. Genetics, 138, 1315-1322. KIMURA, M. 1962. On probability of fixation of mutant genes in a population. GENETICS, 713-&. KIMURA, M. 1983. The neutral theory of molecular evolution, Cambridge [Cambridgeshire] ; New York, Cambridge University Press. LAI, C., LEIPS, J., ZOU, W., ROBERTS, J., WOLLENBERG, K., PARNELL, L., ZENG, Z., ORDOVAS, J. & MACKAY, T. 2007. Speed-mapping quantitative trait loci using microarrays. Nature Methods, 839-841. LAI, C., LYMAN, R., LONG, A., LANGLEY, C. & MACKAY, T. 1994. Naturally- occuring variation in bristle number and DNA polymorphisms at the scabrous locus of Drosophila melanogaster. Science, 1697-1702. LANDE, R. 1983. The response to selection on major and minor mutations affecting a metrical trait. Heredity, 50(1), 47-65. LEXER, C., ROSENTHAL, D., RAYMOND, O., DONOVAN, L. & RIESEBERG, L. 2005. Genetics of species differences in the wild annual sunflowers, Helianthus annuus and H-petiolaris. Genetics, 2225-2239. LONG, A., MULLANEY, S., MACKAY, T. & LANGLEY, C. 1996. Genetic interactions between naturally occurring alleles at quantitative trait loci and 30
mutant alleles at candidate loci affecting bristle number in Drosophila melanogaster. Genetics, 1497-1510. LONG, A., MULLANEY, S., REID, L., FRY, J., LANGLEY, C. & MACKAY, T. 1995. High-resolution mapping of genetic factors affecting abdominal bristle number in Drosophila melanogster. Genetics, 1273-1291. LOPEZ, M. & LOPEZFANJUL, C. 1993. Spontaneious mutation for a quantitative trait in Drosophila melanogaster .2. Distribution of mutant effects on the trait and fitness. Genetical Research, 117-126. MACDONALD, S., PASTINEN, T. & LONG, A. 2005. The effect of polymorphisms in the Enhancer of split gene complex on bristle number variation in a large wild- caught cohort of Drosophila melanogaster. Genetics, 1741-1756. MACKAY, T. 2001. Quantitative trait loci in Drosophila. Nature Reviews Genetics, 2, 11-20. MACKAY, T. & LYMAN, R. 2005. Drosophila bristles and the nature of quantitative genetic variation. Philosophical Transactions of the Royal Society B-Biological Sciences, 1513-1527. MACKAY, T. F. C., FRY, J. D., LYMAN, R. F. & NUZHDIN, S. V. 1994. Polygenic mutations in Drosophila melanogaster - Estimates from response to selection of inbred strains. Genetics, 136, 937-951. MARTIN, G., & LENORMAND, T. 2006. The fitness effects of mutations across environments: a survey in light of fitness landscape models. Evolution, 60(12), 2413-2427. MATHER, K. 1941. Variation and selection of polygenic characters. J. Genet. London, 41, pp. 159-193. MCDONALD, J. & KREITMAN, M. 1991. Adaptive protein evolution at the adh locus in Drosophila. Nature, 652-654. MORAN, P. A. P. 1960. The survival of a mutant gene under selection II. Journal of the Australian Mathematics Society. 1, 485-491. MOUSSET, S., DEROME, N. & VEUILLE, M. 2004. A test of neutrality and constant population size based on the mismatch distribution. Molecular Biology and Evolution, 724-731. NURMINSKY, D., DEPAULIS, F., MOUSSET, S. & VEUILLE, M. 2005. Detecting Selective Sweeps with Haplotype Tests. Selective Sweep. Springer US. ORR, H. 1998a. Testing natural selection vs. genetic drift in phenotypic evolution using quantitative trait locus data. Genetics, 149, 2099-2104. ORR, H. A. 1998b. The Population Genetics of Adaptation: The Distribution of Factors Fixed during Adaptive Evolution. Evolution, 52(4), 935 - 949. ORR, H. A. 2010. The population genetics of beneficial mutations. Philosophical Transactions of the Royal Society B-Biological Sciences, 365, 1195-1201. OTTO, S.P. & WHITLOCK, M. C. 1997. The probability of fixation in populations of changing size. Genetics 146, 723-733. OTTO, S.P. & JONES, C. D. 2000. Detecting the undetected: estimating the total number of loci underlying a quantitative trait. Genetics, 2093-2107. PALSSON, A., DODGSON, J., DWORKIN, I. & GIBSON, G. 2005. Tests for the replication of an association between Egfr and natural variation in Drosophila melanogaster wing morphology. Bmc Genetics, -. 31
PALSSON, A. & GIBSON, G. 2004. Association between nudeotide variation in Egfr and wing shape in Drosophild melanogaster. Genetics, 1187-1198. PATWA, Z. & WAHL L. M. 2008 The fixation probability of beneficial mutations. Journal of the Royal Society Interface, 5, 1279-1289. POLLAK, E. 1966. On the survivla of a gene in a subdivided population. Journal of Applied Probability, 3, 142-195. RIESEBERG, L., WIDMER, A., ARNTZ, A. & BURKE, J. 2002. Directional selection is the primary cause of phenotypic diversification. Proceedings of the National Academy of Sciences of the United States of America, 12242-12245. ROBERTSON, A. 1955. Selection in animals - Synthesis. Cold Spring Harbor Symposia on Quantitative Biology, 20, 225-229. ROBERTSON, A. 1967. The nature of quantitative genetic variation Drosophila melanogaster. Proceedings of the Mendel Centennial Symposium of the Genetics Society of heritage from Mendel, 7-11 September, 1965 Fort Collins, Colo., 265- 280. ROBIN, C., LYMAN, R., LONG, A., LANGLEY, C. & MACKAY, T. 2002. hairy: A quantitative trait locus for Drosophila sensory bristle number. Genetics, 155-164. SAWYER, S., KULATHINAL, R., BUSTAMANTE, C. & HARTL, D. 2003. Bayesian analysis suggests that most amino acid replacements in Drosophila are driven by positive selection. Journal of Molecular Evolution, S154-S164. SCHLENKE, T. & BEGUN, D. 2003. Natural selection drives drosophila immune system evolution. Genetics, 1471-1480. SCHLENKE, T. & BEGUN, D. 2004. Strong selective sweep associated with a transposon insertion in Drosophila simulans. Proceedings of the National Academy of Sciences of the United States of America, 1626-1631. SCHLENKE, T. & BEGUN, D. 2005. Linkage disequilibrium and recent selection at three immunity receptor loci in Drosophila simulans. Genetics, 2013-2022. SHAW, F. H., GEYER, C. J. & SHAW, R. G. 2002. A comprehensive model of mutations affecting fitness and inferences for Arabidopsis thaliana. Evolution, 56, 453-463. STUDER, A., & DOEBLEY, J.F. 2011. Do large effect QTL fractionate? A case study at the maize domestication QTL teosinte branched1. Genetics 188(3):673-681. VERHOEVEN, K., VANHALA, T., BIERE, A., NEVO, E. & VAN DAMME, J. 2004. The genetic basis of adaptive population differentiation: A quantitative trait locus analysis of fitness traits in two wild barley populations from contrasting habitats. Evolution, 270-283. WEINIG, C., DORN, L., KANE, N., GERMAN, Z., HAHDORSDOTTIR, S., UNGERER, M., TOYONAGA, Y., MACKAY, T., PURUGGANAN, M. & SCHMITT, J. 2003. Heterogeneous selection at specific loci in natural environments in Arabidopsis thaliana. Genetics, 321-329. XU, S. 2003. Theoretical basis of the Beavis effect. Genetics, 165(4), 2259-68. ZIMMERMAN, E., PALSSON, A. & GIBSON, G. 2000. Quantitative trait loci affecting components of wing shape in Drosophila melanogaster. Genetics, 671-683.
32
Tables
TABLE 1. NOTATION
Parameter Description x Effect size of a mutation or substitution
Mean effect size of spontaneous mutations
Standard deviation of the effect size of spontaneous mutations k Number of mutations occurring between MA measurements u Per generation mutation rate affecting the trait g Number of generations between MA measurements s Selection coefficient
N Effective population size c Strength of selection: slope of the relationship between s and x
C Normalizing constant to make Eq. 8 a probability density function n The number of loci detected
H=(H1,H2,…,Hn) Vector of the number of substitutions responsible for the QTL, where hi
is the number of substitutions responsible for QTL i q=(q1,q2,…,qn) Estimated effect sizes of the QTL
e=(e1,e2,…,en) Standard errors of the effect sizes of the QTL
33
Figure Legends
FIGURE 1. A depiction of our approach toward characterizing the neutral or selected evolution of phenotype based on mutation accumulation and quantitative trait locus data for a phenotype. A) A constructed example of phenotypic change due to unbiased mutation within seven mutation accumulation lines. B) A constructed example of phenotypic change due to downwardly-biased mutation within seven mutation accumulation lines. C and D) The distribution of mutational effects inferred from panel A.
E and F) The distribution of mutational effects inferred from panel B. G) A depiction of a probability of fixation of novel mutations that is invariant across potential phenotypic values. H) A depiction of probability of fixation of novel mutations that increases with the phenotypic value. I) Same as panel G. J) Same as penel H. K) The product of the functions depicted in panels C and G: in this case, a symmetrical distribution of fixed mutations with a zero mode. L) The product of panels D and H: in this case, an asymmetrical distribution of fixed mutations with a positive mode and positive skewness.
M) The product of the functions depicted in panels I and M: in this case, a symmetrical distribution with a negative mode. N) The product of the functions depicted in panels F and J: in this case, an asymmetrical distribution with a zero mode and positive skewness.
O) An example of a single sample of mutations (open boxes) drawn from the distribution depicted in panel K, compared to the experimentally observed positive and negative QTL
(shaded bars): in this case, the sample matched the QTL well. P) An example of a single sample of mutations (open boxes) drawn from the distribution depicted in panel L, compared to the experimentally observed positive and negative QTL (shaded bars): in this case, the sample matched the positive QTL fairly well, but matched the negative QTL poorly. Q) An example of a single sample of mutations (open boxes) drawn from the distribution in panel M, compared to the experimentally observed positive and negative
QTL (shaded bars): in this case, the sample matched the positive QTL poorly, but matched the negative QTL fairly well. R) An example of a single sample of mutations (open boxes) drawn from the distribution in panel N, compared to the experimentally observed positive 34 and negative QTL (shaded bars): in this case, the sample matched the positive and negative QTL well. The mutations drawn here “happen” to exactly match the sample drawn in panel O, to emphasize that sampling from the fixed mutation distribution is highly stochastic and requires many samples to accurately numerically integrate to yield the likelihood that the fixed mutation distribution underlies the experimentally observed QTL.
S) A plot of the likelihood across a range of values of the strength of selection: in this case, samples drawn from the distribution in panel K are identified as fitting better than samples drawn from the distribution in panel L, indicating neutral evolution of the phenotype. Parts of the plot that lie above the dashed line correspond to a 95% CI. The plot at the y-axis corresponds to neutrality, and falls within the CI, so we cannot reject a hypothesis of selective neutrality for this phenotype. T) A plot of the likelihood across a range of values of the strength of selection: in this case, samples drawn from the distribution in panel N are identified as fitting better than samples drawn from the distribution in panel M, indicating evolution of the phenotype driven by natural selection.
The plot at the y-axis corresponds to neutrality, and falls outside the CI, so we can reject a hypothesis of selective neutrality for this phenotype, and estimate that the strength of selection corresponds to the level of selection identified by the peak of the plot.
FIGURE 2. Frequency distribution of changes in bristle traits during mutation accumulation. A) Change in sternopleural bristle number between measurements in a mutation accumulation experiment by Paxman (1957). Six lines of D. melanogaster were raised in unselective conditions in which they were subjected to population bottlenecks for forty generations. The mean sternopleural bristle number was assayed five times over the course of the experiment. B) Change in abdominal bristle number in a mutation accumulation experiment by Lopez and Lopez-Fanjul (1993). Ninety-three mutation accumulation lines of D. melanogaster were maintained for sixty-one generations and then assayed for mean abdominal bristle number.
35
FIGURE 3. Selection in the HST population. A) The effects and standard errors of sternopleural bristle QTL detected in the HST population selected for high sternopleural bristle number. B) The maximum likelihood estimate of the distribution of mutational effects for sternopleural bristles. C) Log likelihood versus c, the strength of selection, given the HST QTL data and the maximum likelihood distribution of mutational effects.
The horizontal line indicates the likelihood threshold for a 95% confidence interval. D) Log likelihood versus c, given the HST QTL data plus four imputed unobserved loci. E) Log likelihood versus c, given the HST QTL data, with the estimate of μ perturbed by adding one standard deviation (crosses, dashed threshold line) and subtracting one standard deviation (circles, solid threshold line). F) Log likelihood versus c given the HST QTL data, with the estimate of σ perturbed by adding one standard deviation (crosses, dashed threshold line) and by subtracting one standard deviation (circles, solid threshold line).
FIGURE 4. Selection in the LST population. A) The effects and standard errors of sternopleural bristle QTL detected in the LST population, selected for low sternopleural bristle number. B) The maximum likelihood estimate of the distribution of mutational effects for sternopleural bristles. C) Log likelihood versus c, the strength of selection, given the LST QTL data and the maximum likelihood distribution of mutational effects.
The horizontal line indicates the likelihood threshold for a 95% confidence interval. D) Log likelihood versus c, given the LST QTL data plus four imputed unobserved loci. E) Log likelihood versus c, given the LST QTL data, with the estimate of μ perturbed by adding one standard deviation (crosses, dashed threshold line) and subtracting one standard deviation (circles, solid threshold line). F) Log likelihood versus c given the LST QTL data, with the estimate of σ perturbed by adding one standard deviation (crosses, dashed threshold line) and by subtracting one standard deviation (circles, solid threshold line).
FIGURE 5. Selection in the HAB population. A) The effects and standard errors of abdominal bristle QTL detected in the HAB population (selected for high abdominal bristle 36 number. B) The maximum likelihood estimate of distribution of mutational effects for abdominal bristles. C) Log likelihood versus c, the strength of selection, given the HAB
QTL data and the maximum likelihood distribution of mutational effects. The horizontal line indicates the likelihood threshold for a 95% confidence interval. D) Log likelihood versus c, given the HAB QTL data plus five imputed unobserved loci. E) Log likelihood versus c given the HAB QTL data, with the estimate of μ perturbed by adding one standard deviation (crosses, dashed threshold line) and by subtracting one standard deviation
(circles, solid threshold line). F) Log likelihood versus c given the HAB QTL data, with the estimate of σ perturbed by adding one standard deviation (crosses, dashed threshold line) and by subtracting one standard deviation (circles, solid threshold line).
FIGURE 6. Selection in the LAB population. A) The effects and standard errors of abdominal bristle QTL detected in the LAB population. B) The maximum likelihood estimate of distribution of mutational effects for abdominal bristles. C) Log likelihood versus c, the strength of selection, given the LAB QTL data and the maximum likelihood distribution of mutational effects. The horizontal line indicates the likelihood threshold for a 95% confidence interval. D) Log likelihood versus c, given the LAB QTL data plus thirteen imputed unobserved loci. E) Log likelihood versus c given the LAB QTL data, with the estimate of μ perturbed by adding one standard deviation (crosses, dashed threshold line) and subtracting one standard deviation (circles, solid threshold line). F) Log likelihood versus c given the LAB QTL data, with the estimate of σ perturbed by adding one standard deviation (crosses, dashed threshold line) and subtracting one standard deviation (circles, solid threshold line).
FIGURE 7. Power versus the number of loci detected. We applied the likelihood ratio test to all subsets of the four QTL data sets, with the appropriate mutation effect distribution.
The y-axis represents the proportion of subsets with a given number of loci for which the test rejected the hypothesis of neutrality. FIGURE 1 A) B) Phenotype Phenotype
Time in mutation Time in mutation accumulation accumulation
C) D) E) F) Frequency x x x x G) H) I) J) Fixation probability = = = = K) L) M) N) Frequency
O) P) Q) R) QTL
Phenotypic effect Phenotypic effect Phenotypic effect Phenotypic effect
S) T) Likelihood Likelihood
Strength of selection Strength of selection FIGURE 2
A) B) FIGURE 3
A) 3 B) 2
2
1
1 Effect
−1 Frequency
−2
−3 −1 0 1 QTL Effect C) D) −20 −20
−25 −25
−30 −30
Log likelihood −35 −35
−40 −40 0 0.01 0.02 0.03 0 0.01 0.02 0.03
E) F) −20 −20
−25 −25
−30 −30
Log likelihood −35 −35
−40 −40 0 0.01 0.02 0.03 0 0.01 0.02 0.03 Strength of selection (c) Strength of selection (c) FIGURE 4
A) 3 B) 2
2
1
1 Effect
−1 Frequency
−2
−3 −1 0 1 QTL Effect C) 0 D) 0
−2 −2 Log likelihood
−0.2 −0.1 0 −0.2 −0.1 0
E) 0 F) 0
−2 −2 Log likelihood
−0.2 −0.1 0 −0.2 −0.1 0 Strength of selection (c) Strength of selection (c) FIGURE 5
A) 3 B) 2
2
1
1 Effect
−1 Frequency
−2
−3 −1 0 1 QTL Effect C) D)
−5 −5
−15 −15 Log likelihood
0 0.05 0.1 0 0.05 0.1
E) F)
−5 −5
−15 −15 Log likelihood
0 0.05 0.1 0 0.05 0.1 Strength of selection (c) Strength of selection (c) FIGURE 6
A) 3 B) 2
2
1
1 Effect
−1 Frequency
−2
−3 −1 0 1 QTL Effect C) 0 D) 0
−2 −2 Log likelihood
−0.2 −0.1 0 −0.2 −0.1 0
E) 0 F) 0
−2 −2 Log likelihood
−0.2 −0.1 0 −0.2 −0.1 0 Strength of selection (c) Strength of selection (c) FIGURE 7