Phylogenetics - Phylogenetic Methods and Accuracy

Methods of Phylogenetic Inference Introduction Phylogenetic inference is the process of reconstructing evolutionary history from biological data. Scientists use various methods to build evolutionary trees that best explain patterns of similarity and difference among organisms. Each method makes different assumptions about how evolution works and uses different criteria to evaluate which tree is "best." Understanding these methods is essential because the choice of method can affect the resulting tree topology and our conclusions about evolutionary relationships. The Parsimony Principle Definition: Parsimony is a fundamental principle in phylogenetics that states: the tree requiring the fewest evolutionary changes to explain the observed data is preferred. The logic behind parsimony is intuitive: if we observe characters (traits, DNA sequences, morphological features) distributed among taxa, we want to find the evolutionary tree that explains this distribution with the minimum number of change events. For example, if a specific mutation appears in only three closely related species, it's more parsimonious to assume that mutation occurred once in their common ancestor rather than three separate times independently. To apply maximum parsimony, you: Consider all possible tree topologies for your set of taxa Count the total number of character state changes required for each tree Select the tree (or trees) with the lowest total number of changes A key limitation: Parsimony assumes that evolution is relatively slow and that similar character states indicate recent common ancestry. However, under conditions of rapid evolution, parsimony can perform poorly—particularly due to a problem called long-branch attraction, which we'll discuss later. Distance-Based Methods: Neighbor Joining Distance-based methods work fundamentally differently from parsimony. Instead of counting character changes, they use overall similarity (or dissimilarity) between taxa to build trees. The concept: Distance methods calculate a pairwise distance matrix—essentially, a table showing how different each organism is from every other organism. The more different two organisms are, the larger their distance value. The tree is then constructed by clustering organisms based on these distances. Neighbor Joining, developed by Saitou and Nei (1987), is the most widely used distance method. Here's how it works: Begin with all taxa as separate lineages Identify the pair of taxa with the smallest distance (most similar to each other) Join them together with a branch point Calculate distances from this new node to all remaining taxa Repeat until all taxa are connected Neighbor joining has become popular because it doesn't require assuming that evolution proceeds at a constant rate (unlike older distance methods such as UPGMA). This makes it more applicable to real biological data where different lineages often evolve at different speeds. Maximum Likelihood Methods The fundamental principle: Maximum likelihood asks: "Which tree makes the observed data most probable?" Rather than assuming the simplest explanation, likelihood methods calculate the probability of seeing your actual data under different tree models and select the tree that makes your data most likely. To use maximum likelihood, you must: Assume a model of character change (for DNA sequences, this might specify the rates at which different types of mutations occur) For each possible tree topology, calculate the probability of observing your data given that tree and your chosen model Select the tree with the highest probability (the "maximum likelihood" tree) Why this matters: Maximum likelihood has a key advantage over parsimony—it explicitly accounts for the evolutionary process. If you know that certain mutations are more common than others (for instance, transitions occur more frequently than transversions in DNA), you can incorporate this knowledge into your analysis. This makes likelihood methods more statistically sophisticated, though also more computationally intensive. The choice of evolutionary model is critical. Different models make different assumptions about substitution rates, and using an inappropriate model can lead to incorrect tree inference. Bayesian Inference and Markov Chain Monte Carlo The Bayesian approach: Bayesian phylogenetics combines prior information about what trees are plausible with the likelihood of the observed data to calculate posterior probabilities—the probability of each tree being correct given your data and your prior assumptions. Mathematically, this follows Bayes' theorem: $$\text{Posterior Probability} = \frac{\text{Likelihood} \times \text{Prior Probability}}{\text{Probability of Data}}$$ Unlike maximum likelihood, which finds a single "best" tree, Bayesian methods generate a probability distribution across many possible trees. Trees that are more probable receive higher posterior values. The computational challenge: With even moderate numbers of taxa, the number of possible trees becomes astronomical. Bayesian phylogenetics solves this using Markov Chain Monte Carlo (MCMC) sampling—a technique that walks through tree space, proposing changes to the current tree and accepting or rejecting these changes based on their probability. Over time, MCMC explores the landscape of reasonable trees and builds up an estimate of the posterior probability distribution. A practical advantage: Bayesian analyses naturally provide measures of uncertainty. Rather than getting a single tree with some bootstrap support values (discussed below), you get a full distribution showing which clades are strongly supported and which are uncertain. Phenetic Distance Methods and Overall Similarity Phenetic approaches take a fundamentally different philosophical stance than other methods. Rather than trying to infer evolutionary history explicitly, phenetic methods assume that overall phenotypic similarity (or genetic similarity) correlates with evolutionary closeness. Phenetic methods: Calculate overall similarity/dissimilarity between organisms Assume similarity reflects evolutionary relationships Construct trees that group similar organisms together Are computationally simple and fast While neighbor joining (described above) is sometimes considered phenetic because it uses distance data, pure phenetic methods make fewer assumptions about the evolutionary process and focus simply on clustering based on similarity. These methods fell somewhat out of favor because similarity can be misleading—unrelated organisms might appear similar due to convergent evolution, while closely related organisms might appear different due to rapid change. Taxon Sampling: A Critical Practical Consideration Definition: Taxon sampling refers to which species or populations you choose to include in your phylogenetic analysis. For most large clades, you cannot realistically sequence or examine every single species. Instead, you sample representative taxa—a manageable subset intended to capture the diversity of the group. The question is: how does this sampling choice affect your results? The effect of poor sampling: Inadequate or biased taxon sampling can lead to serious errors in tree inference, most notably: Long-branch attraction: This occurs when two distantly related lineages are incorrectly grouped together because: They have accumulated many independent changes since their last common ancestor These changes make them appear more similar to each other than to their true relatives This is particularly problematic if intervening taxa (those that would clarify the true relationships) are missing from your sample For example, if you sample species A, species B, and species C, but they actually evolved as (A,(B,C)), you might mistakenly infer (C,(A,B)) because B and C have undergone rapid evolution and accumulated many independent substitutions making them appear more similar. Practical implication: Sampling more taxa, especially key transitional or intermediate forms, generally improves tree accuracy and prevents artifacts like long-branch attraction. Assessing Tree Reliability and Support Once you've inferred a tree, a critical question follows: how confident should you be in this result? Several statistical approaches address this. Bootstrap Resampling The bootstrap method, introduced by Felsenstein (1985), provides a statistical measure of how well different parts of your tree are supported by your data. Here's how it works: Take your original dataset (e.g., DNA sequence alignment) Randomly resample characters (nucleotide positions) with replacement, creating a new dataset of the same size Perform phylogenetic inference on this resampled dataset Repeat steps 2–3 many times (often 100–1000 times) For each node in your original tree, count how often that same grouping appears in the bootstrap replicates The percentage of times a clade appears is its "bootstrap support value" A bootstrap value of 95% for a particular node means that clade appeared in 95% of the resampled analyses—generally considered strong support. Values below 70% are typically considered weak support. Why this works: Bootstrap mimics the uncertainty you'd expect from having incomplete data. By resampling characters, you create slightly different datasets that might yield slightly different trees. If a clade appears consistently across these variations, it's robust; if it disappears in many resampled datasets, it's sensitive to minor data variation. Alternative Resampling Methods and Tree Comparison The jackknife is similar to bootstrap but removes characters rather than resampling with replacement. Other resampling approaches provide complementary measures of support. Consensus trees combine multiple trees into a single representation: Strict consensus: Shows only clades that appear in all input trees Majority-rule consensus: Shows clades appearing in more than 50% of trees These are useful when you have many equally parsimonious trees or when comparing results across different methods Tree distance metrics quantify how different two trees are from each other. The Robinson-Foulds metric counts the number of bipartitions that differ between two labeled trees, providing a numerical measure of topological difference. Measuring Homoplasy Homoplasy refers to similar character states that evolved independently rather than being inherited from a common ancestor. High homoplasy in your data makes phylogenetic inference difficult. The consistency index and retention index measure homoplasy levels: Consistency Index: The ratio of the minimum possible changes to the actual number of changes required. Values closer to 1.0 indicate less homoplasy Rescaled Consistency Index: A modified version that corrects for the number of taxa These indices help you evaluate whether your data strongly supports a single tree topology or whether multiple trees might be equally plausible. Model Selection for Molecular Data When analyzing DNA or protein sequences, you must select a substitution model—a mathematical description of how characters change over time. Different models make different assumptions about substitution rates and frequencies. Information-theoretic criteria guide model selection: Akaike Information Criterion (AIC): Balances fit to data against model complexity Bayesian Information Criterion (BIC): Similar to AIC but penalizes complexity more heavily These approaches help you avoid overfitting (choosing an overly complex model that fits noise rather than signal) while ensuring your model is realistic enough to capture important evolutionary processes. The choice of substitution model can affect which tree is favored, making model selection an important step in likelihood and Bayesian analyses. Comparing Methods: Practical Considerations Different phylogenetic methods can produce different trees from the same dataset. Which should you use? Parsimony is useful for morphological data and works well when homoplasy is low, but can struggle with long branches and rapid evolution Distance methods are fast and work well with large datasets, but lose information by reducing sequences to single distance values Likelihood and Bayesian methods are statistically rigorous and account for the evolutionary process explicitly, but are computationally demanding and require model selection Best practice: Compare results across methods. If different approaches agree, you have stronger confidence in the result. If they disagree, investigate why—the disagreement often reveals important features of your data or limitations of particular methods.