Molecular and morphological data in phylogenetic inference

Make corresponding observations comparable

Morphological analyses score defined structures or other traits in living and fossil samples. Molecular analyses compare corresponding sequence positions. Alignment proposes which positions correspond despite insertions, deletions and differing lengths. A column of an alignment is therefore a comparison claim, not just text placed in a neat row.

For a species-level question, investigators also examine which gene copies are being compared. Mistaking duplicated copies for the same inherited locus can mislead the analysis even if each sequence was read correctly. Contamination, missing samples and uncertain alignments likewise matter before tree calculation begins.

More data can add useful evidence, but repeating a systematic error does not remove it. Linked sites and genes may not be independent; different regions may have different histories. Combining regions is a modelling choice to assess, rather than an automatic guarantee of a better species tree.

Parsimony asks for the smallest change count

Under an equal-cost binary parsimony rule, each change between 0 and 1 counts one step. For each candidate topology, find the smallest number of steps that can explain the observed tip states. This is an inference criterion, not a rule requiring evolution to choose the simplest actual history.

Our unrooted four-tip example has A = 0, B = 0, C = 1 and D = 1. On the A/B versus C/D split, assigning 0 to the A/B internal node and 1 to the C/D node gives one change on the joining edge. On A/C versus B/D, each pair contains both states. At least two changes are needed. The third split, A/D versus B/C, also requires two.

This one character therefore favours the first split under the supplied rule. It does not determine a root or a divergence date, and it cannot exclude repeated change in the actual history. Parsimony can be misled when homoplasy and unequal evolutionary patterns do not fit its assumptions.

Likelihood asks how well a model could generate the data

A likelihood calculation asks for the probability of the observed data conditional on a proposed tree and a change model. It is not, by itself, the probability that the tree is true. Maximum-likelihood analysis compares candidate trees and ordinarily estimates model parameters and branch lengths; our small illustration keeps them fixed.

The illustration uses two character states, 0 and 1. At one chosen internal node each state has probability one half. On each of the five unrooted-tree edges, the end states match with probability 0.9 and differ with probability 0.1. These transition probabilities are assigned; 0.1 is not a measured mutation rate or a count of all substitutions along an edge.

For each possible pair of internal states, multiply the initial-state probability and all five edge transition probabilities. Then add the four terms because those internal states are unobserved alternatives. For the 0,0,1,1 tip pattern, this gives likelihood 0.0401 on the A/B versus C/D split and 0.0081 on A/C versus B/D. The first is about 4.95 times the second under exactly these assumptions.

A real sequence model generally has more states and parameters and must accommodate relevant evolutionary differences. Changing the model, samples or alignment can change the comparison. Model fit and sensitivity checks complement a high score; a single favourable calculation is not a universal validation.

Support describes an analysis and its uncertainty

A conventional character bootstrap draws alignment columns or other character units with replacement, producing a replicate dataset of the same size. The sampled taxa remain in place. Repeating the analysis on many replicates gives the fraction in which a specified branch or clade is recovered.

That frequency is tied to the dataset, resampling scheme and analysis. It is not automatically a posterior probability that the clade is true. Bayesian posterior support is a different quantity, conditional on its model, priors and analysis. Neither label removes model error, misleading sampling or gene-history conflict.

Low support can justify leaving a connection unresolved. High support can still accompany a biased result. Examine which relationship is scored, how evidence was resampled and whether reasonable alternatives change the conclusion. Presenting uncertainty helps readers use a tree responsibly.

Worked example

Check your understanding

1. Why does adding the 0,1,0,1 character create a parsimony tie between the first two splits?

   Show reasoning

   The first pattern costs 1 versus 2; the second costs 2 versus 1. Equal-weight sums are 3 versus 3. Neither wins from these two characters under the stated rule, and the tie should not be hidden by choosing a preferred drawing.

2. What does the likelihood ratio 0.0401/0.0081 establish?

   Show reasoning

   The observed single-site pattern is about 4.95 times as probable under the first fixed tree/model combination as under the second. It does not make the first tree 4.95 times as likely to be true without further assumptions and inferential treatment.

3. If a split occurs in 8 of 10 bootstrap outputs, what is scored and what uncertainty remains?

   Show reasoning

   Its recovery frequency is 80% under that resampling and analysis. The data, independence assumptions, alignment, model and sampling can still be wrong or incomplete; the fraction is not automatically a probability of biological truth.

What to carry forward

A phylogenetic result depends on comparable evidence, an explicit inference rule and a support measure whose meaning is preserved.