"PRINCIPLES OF PHYLOGENETICS: ECOLOGY AND EVOLUTION" Integrative Biology 200B Spring 2011 University of California, Berkeley B.D. Mishler Feb. 3, 2011. Quantitative character evolution within a cladogram (intro; ancestral trait reconstruction; phylogenetic conservatism) I. Continuous characters – ancestral states Many traits of interest are measured on continuous or metric scales – size and shape, physiological rates, etc. Continuous traits are often useful for species identification and taxonomic descriptions; historically, they were also used in phylogenetic analysis through the use of clustering algorithms that can group taxa based on multivariate phenetic similarity. With the advent of cladistics, phylogenetic reconstruction per se has shifted entirely to discrete trait analysis, but continuous traits remain important in post-tree analysis. In principal, ordinal traits take on discrete states while continuous traits are real numbers. In reality, there is a continuum from ordered discrete traits with only a few states (1,2,3,4,5 petals) to those with enough states that we may treat them as continuous (dozens to hundreds). It is also possible to 'discretize' traits into states if breaks are observed in the distributions for different species. Distributions of continuous traits may take on any arbitrary shape, but certain distributions occur repeatedly, possibly reflecting common underlying 'natural' processes. Normal distribution: sum of many small additive effects Exponential distribution: product of many small multiplicative effects Poisson distribution: frequencies of rare events in discrete intervals etc. This becomes important when we consider whether ancestral reconstruction of continuous traits should reflect an underlying evolutionary model of the process that describes or dictates trait evolution. Traits may be transformed to better meet an appropriate distribution (e.g. log- transform). I.B. Parsimony methods for ancestral states 1. Linear or Wagner Parsimony: minimize the sum of absolute or linear changes along each branch (analogous to normal parsimony for discrete traits). The ancestral value at each node will be the median of the three values around it (two child nodes, one parent node). The root is a special case, where it will be the median of the two child nodes, as there is no parent node. 2. Squared-Change Parsimony: minimize the sum of squared changes along each branch. The ancestral value at each node will be the mean of the three values around it (two child nodes, one parent node). Weighted SCP can also be calculated, where the change along each branch is divided by its branch length before summing – a given change on a long branch is penalized less.I.C. Brownian motion and maximum likelihood ancestral states Brownian motion (BM) is the term for a random walk in a continuous valued variable. If a trait was determined by multiple, independent additive factors of small effect, and if each factor was mutating or changing at random (e.g., by drift), then the character change would constitute BM. Brownian motion is the starting point for discussions of continuous character evolution, for its simplicity and its close ties to parametric statistics based on normal distributions. In Brownian motion the size of each step is drawn from a normal distribution with mean = 0 (no trend) and variance s2 (= standard deviation s), where each step is a unit of time. When we consider Brownian motion as a process, this variance is viewed as a rate parameter, β. One of the fundamental principles of probability theory is that the variance of the sum of two random processes is the sum of their variances. In other words, if the variance of a brownian motion process is β after one time step, it will be β + β = 2β after two time steps. So the variance increases linearly with time. If you apply BM to a large number of independent random walks, with time = t along each walk, then you can probably see that the variance of the resulting values at the tips of the walks will be tβ. What is less intuitive for most of us (if you are not used to statistical thinking) is that a single value resulting from a random walk also has a variance that refers to the underlying (and unobserved) distribution from which that value has been drawn. Whether you know it or not, we all solve a maximum likelihood (ML) problem on a daily basis when we calculate the mean for a set of numbers. The mean of X (a set of numbers) is the sum of X divided by N, the number of values in X, right? Yes. Alternatively, the mean of X is the ML solution for the starting point of N random walks that end with values X. This can be solved from the following steps: 1) From the central limit theorem, we know that random walks generate values drawn from a normal distribution, with mean u and variance s2. 2) The probability of each value of X, under a normal distribution is: P(x)= exp[−(x−u)2/2s2] / s √2π 3) The ML solution for u and s2 are the values that maximize their cumulative probability over all values of x, and the cumulative probability is the product of the individual probabilities. A product of a series of values for P(x) looks pretty nasty, so instead let's take the sum of the log of P(x) (because the log of a product is the sum of the logs): log(P(x)) = [−(x−u)2/2s2] − log ( s √2π) 4) To maximize this, we can ignore the denominator (2s2) and the second term, since they will be constants. And if we are maximizing the sum of negative terms, we can instead minimize the sum of the positive terms. So the mean of X is that value which minimizes: ∑(x−u)2 Look familiar!? It's the sum of squares of X. And now there's some magic, and the sum of squares is also how we calculate s2, but we won't try to derive that as a ML problem here._________________________________ R script to solve for the mean of a set of numbers by finding the minimum of the sum of squares: 5) Try this R script to solve for the mean of a set of numbers by finding the minimum of the sum of squares: ## enter a set of numbers in xx xx = c(1,2,4,5) ## create a sequence of candidate values for the mean of xx xu = seq(1,5,by=0.1) ## create a variable to hold the sum of squares lxu = rep(NA,length(xu)) ## loop through xu and calculate the likelihood score for each candidate value as sum of
View Full Document
Unlocking...