# Molecular Evolution

> Neutral changes are fixed, on average, at the rate they arise, so the count of differences between two species grows steadily with time and can be dated from one calibrated split; a departure from that count is a clue to selection, not proof. This lesson shows why the clock ticks (K = mu), corrects counts for hidden changes, dates a split with a fossil-calibrated clock, sets out where the steady tick fails (varying rates, mutation rates and Ohta's nearly neutral mutations), and reads dN/dS and the McDonald-Kreitman test. Afterwards you can date a divergence from a calibration, state what the calibration assumes, and say what a test for positive selection can and cannot show.

Course: Ecology and Evolution (Genomes, development and diversification). Level: core university (third year). About 17 minutes.
Lesson page (link this one when citing): https://stemhop.com/lessons/molecular-evolution/

Builds on: Reading Phylogenetic Trees · Drift, Gene Flow and Competing Explanations · How to Test an Evolutionary Explanation

## What you'll learn

- Correct observed sequence differences for repeated substitutions and date a split with a molecular clock calibrated by a fossil or by dated samples, stating that a fossil gives a minimum age.
- State the neutral theory's claim and its expectations (K = mu for neutral mutations, fixation probability 1/2N for one new neutral copy in an ideal population), and contrast Ohta's nearly neutral theory, in which effective population size matters.
- Interpret dN/dS (omega) below, near and above 1, and explain why a McDonald-Kreitman comparison of polymorphism and divergence is a second, partly independent signal.
- Identify the checks (alignment, orthology, mutation model, synonymous-site neutrality, functional tests) needed before accepting a statistical signal of positive selection.

**Genes can roughly date a split that left no fossil, and show whether selection or chance shaped what changed.**

> **The one idea.** On average, neutral changes (no effect on survival or reproduction) become substitutions at the rate they arise, whatever the population size. So neutral differences between two species pile up steadily, and one dated split turns their count into years. A departure from that neutral count is a clue to selection, not proof.
>
> *Scope:* neutral sites in the same gene in both species, corrected for hidden changes; the dated split and the target must share a steady rate. Slightly harmful mutations break it: their rate depends on effective population size.

## Sequences as a record

The first film counted **substitutions**, [mutations](https://stemhop.com/lessons/mutation/) that have reached **fixation** (every member of the lineage carries them), assuming you compare the same gene and see every change.

The same gene means **orthologous**: both copies descend from one ancestral gene, split by speciation rather than by duplication (the gene-family copies in [How Signaling Evolved](https://stemhop.com/lessons/how-signaling-evolved/) are mostly duplicates). [Sequencing and genomes](https://stemhop.com/lessons/sequencing-and-genomes/) covers obtaining them. Call p the proportion of aligned sites that differ.

**Predict.** Two sequences differ at 30% of aligned sites (p = 0.30). Did fewer than 30 substitutions per 100 sites happen since they split, exactly 30, or more?

> **Check your prediction.** More. A site can change twice, or change back (A to G, then G to A), hiding the first change. So p underestimates the true number of substitutions, more so as sequences diverge.

A **corrected distance**, D, estimates the real number of substitutions per site. The simplest correction (Jukes and Cantor) assumes all four bases replace one another at one rate; take its form as given (ln is the natural logarithm):

> **D = -(3/4) x ln(1 - (4/3) x p)**

For p = 0.30, 1 - 0.40 = 0.60, ln 0.60 = -0.511, and D = 0.383: over a quarter more change than you see. For small p the correction is tiny: p = 0.012 gives D = 0.0121. The tree relating the sequences is itself an inference with a stated support, as in [reading phylogenetic trees](https://stemhop.com/lessons/reading-phylogenetic-trees/).

## Why the clock ticks

Emile Zuckerkandl and Linus Pauling saw in the early 1960s that hemoglobin differences grew roughly in proportion to fossil-dated time since a split. The first film showed why; here is the rate.

> **Equation card: The rate at which neutral mutations become substitutions in one lineage.**
>
>     K = (2N × mu) × (1/2N) = mu
>
> In words: New neutral mutations arise at two N mu per generation; each fixes with probability one over two N; so the substitution rate K equals the neutral mutation rate mu.
>
> - N: diploid individuals, so 2N copies of each site
> - mu: neutral mutations per site, per copy, per generation
> - K: neutral substitutions per site, per generation, in one lineage
>
> Sanity check: N = 100, mu = 10^−8: 2 × 10^−6 new mutations per generation × 1/200 = 10^−8 = mu.

The fixation chance of one new neutral copy is 1/2N only in an ideal population (constant size, random mating); [population genetics](https://stemhop.com/lessons/population-genetics/) derives it. K = mu is an average; real substitutions arrive irregularly. Two steps link K to measured distance:

- **Per year**, the clock rate r is K divided by generation time, so at the same per-generation mutation rate, long generations mean fewer substitutions per year.
- **Between two tips**, D = 2rt after a time t, if both branches share the rate r.

Motoo Kimura reached the **neutral theory** in 1968 from the other direction, arguing that protein comparisons implied too many substitutions for selection to pay for each in selective deaths. He proposed that most substitutions are neutral or nearly neutral mutations fixed by [random drift](https://stemhop.com/lessons/drift-gene-flow-explanations/), while most strongly harmful mutations are removed before they spread. The theory does not say all mutations are neutral, or that selection is unimportant for traits.

## Reading time: calibration

The **[molecular clock](https://stemhop.com/methods/molecular-clock/)** says the expected number of changes on a branch is its rate times its duration. The first film’s outside date is a **calibration**; besides a fossil, sequences sampled at sufficiently different known dates, such as virus samples or [ancient DNA](https://stemhop.com/methods/next-generation-sequencing/), can calibrate a tree.

**Worked problem (illustrative data).** Two species that fossils show split 3.0 million years ago differ by a corrected 0.012 substitutions per site in a gene. A second pair differs by 0.024 in the same gene. When did the second pair split? Assume one constant rate on all four lineages, an exact calibration date, and negligible variation in the ancestral population (see Going deeper).

1. *Set up the model.* D = rt + rt = 2rt, with r the rate per lineage and t the time since the split.
2. *Calibrate.* r = 0.012 / (2 x 3,000,000 years) = 2.0 x 10−9 substitutions per site per year.
3. *Date the target.* t = 0.024 / (2 x 2.0 x 10−9) = 0.024 / (4.0 x 10−9) = 6.0 million years. Check: 2 x 2.0 x 10−9 x 6,000,000 = 0.024.

Forgetting the factor 2 in step 3 gives 0.024 / 2.0 x 10−9 = 12 million years. Since a real fossil calibration is only a minimum, both splits may be older. Try it on the line below.

*Interactive on the lesson page:* Calibrating a molecular clock: D = 2rt

## When the tick rate varies

**Rates differ between branches**: in Drummond and colleagues’ 2006 marsupial analysis, the fastest branch of the most probable tree was estimated to evolve 2.7 times faster than the slowest. **Relaxed clocks** let each branch have its own rate. Bayesian dating commonly treats a fossil calibration as a range of possible ages, not a point; even with abundant data, the dates stay sensitive to the rate model and calibration ranges put in.

**The mutation rate itself varies.** Bergeron and colleagues in 2023 sequenced parents and offspring from 68 vertebrate species and found per-generation germline mutation rates differing 40-fold, around 10−8 per site per generation in mammals. So mu is no constant, and turning it into a per-year rate takes assumptions about neutrality and generation time.

**Some mutations are not quite neutral.** For these, the **effective population size**, Ne, matters: the size of an ideal population that drifts as strongly as the real one (the smaller Ne, the stronger drift).

**Predict.** A new mutation lowers fitness by 1 part in 10,000 for each copy carried. Is it more likely to become a substitution when Ne is 1,000 or 100,000?

> **Check your prediction.** When Ne is 1,000: chance swings dwarf so small a cost, and it fixes at about 80% of the neutral rate (simple diploid model, constant Ne; other dominance or demographic assumptions shift it). At 100,000, selection wins and it almost never fixes. This is Tomoko Ohta’s 1973 **nearly neutral theory**: many mutations are slightly deleterious, with costs of the order of 1/Ne, and their substitution rate depends on Ne.

*Figure (AI-generated illustration).* For slightly harmful mutations, drift can fix them in a small population while selection removes them in a large one, so their substitution rate depends on the effective population size. For strictly neutral mutations it does not: more mutations in a large population are offset by a smaller chance that each fixes. Image: two population fields of plain dots, a small one with six dots and a large one with twenty-four. A slightly harmful mutation, marked in gold, spreads to every dot in the small population but is removed in the large one. A bracket beneath notes that a large population makes more neutral mutations, each with a smaller chance of fixing.

## Reading selection in coding sequences

Where selection acts, substitutions depart from the neutral count: purifying selection can make amino-acid changes lag, and positive selection can make them run ahead. The second film shows two ways to catch that in a real gene.

**Film transcript: Reading Selection in Genes**

### Film transcript: Reading Selection in Genes

Where selection acts, changes depart from the neutral count. Purifying selection can make them lag; positive selection can make them run ahead. So how could you catch that in a real gene? Here is the trick. Find two kinds of change in the same gene that neutrality says should behave alike. If they part ways by more than chance allows, selection may be at work. Take a coding gene, read three letters at a time. The genetic code has several codons for most amino acids. So some changes keep the amino acid. These are synonymous. Other changes alter the amino acid. These are nonsynonymous, or replacement, changes. If synonymous changes are neutral, they are a built-in control for the replacements.

Count each kind along a branch, per site where it could happen. Call them dN and dS. Their ratio is called omega. Below one, replacements were fixed more slowly than silent changes. That suggests purifying selection removed some.

Near one is consistent with neutrality. Above one suggests positive selection. A second test uses variation inside a species. In 1991, John McDonald and Martin Kreitman applied it to the Adh gene of fruit flies. The gene makes alcohol dehydrogenase. Sort every change into a table. The columns are replacement and synonymous. One row holds variants within a species. The other holds differences fixed between species. Under neutrality, a variant within a species is a passing stage, on its way to fixation or loss. So both rows should hold the same ratio of replacement to synonymous changes. The variants row fills first, and it is mostly synonymous.

If replacements were neutral, what should the fixed row look like? Neutrality predicts the same ratio. Instead, the fixed row held more replacements than the variants predicted. McDonald and Kreitman suggested this reflected adaptive fixation. Here each kind of change is the control for the other. So this test needs no clock and no dated split. Both tests are model inferences, and each model can fail. A signal like this still needs further checks before it counts as adaptation.

Both tests ask the same question. Did one kind of change keep pace with the neutral count? A departure is a clue to selection, not proof.

From the film

- A synonymous change keeps the amino acid; a nonsynonymous, or replacement, change alters it.
- If synonymous changes are neutral, omega = dN/dS reads selection on the protein: below 1 suggests purifying selection, near 1 is consistent with neutrality, above 1 suggests positive selection.
- The McDonald-Kreitman test compares the ratio of replacement to synonymous changes among variants within a species with the same ratio among differences fixed between species.
- At Adh in fruit flies, fixed differences held more replacements than the variants predicted, which McDonald and Kreitman suggested reflected adaptive fixation.
- Each test is a model inference: a departure from the neutral count is a clue to selection, not proof, and needs further checks before it counts as adaptation.

Why count per site? Given the [genetic code](https://stemhop.com/lessons/genetic-code-and-translation/), most single-base changes in a codon alter the amino acid, so a gene offers roughly three nonsynonymous sites for every synonymous one, and raw counts would overstate amino-acid change about threefold.

*Figure (AI-generated illustration).* In a coding gene, a nonsynonymous change alters the amino acid and a synonymous change does not, because the genetic code has several codons for most amino acids. That is why dS can serve as a built-in control for dN. Image: a row of six codon blocks, each with three small squares, above a string of six coloured beads for amino acids. Two insets compare a synonymous change, where one square turns gold but the bead keeps its colour, with a nonsynonymous change, where one square turns gold and the bead changes colour.

**Worked problem, step 4.** For the second pair, a different gene has dN = 0.004 and dS = 0.040 substitutions per site. Omega = 0.004 / 0.040 = 0.10: the signature of **purifying selection** removing most amino-acid changes, though the gene still evolves; omega alone cannot rule out a few adaptive changes hidden in the average.

**Predict.** One gene has 3 codons under strong positive selection and 297 under strong constraint. Is a gene-wide omega likely to reveal the 3?

> **Check your prediction.** Probably not. A gene-wide omega averages over all sites, so 297 constrained codons usually hold it below 1 unless the three change extremely fast. Codon models can instead estimate omega for a branch or a class of sites.

A **polymorphism** is a variant that still varies within a species. The second film’s Adh counts, from three *Drosophila* species (McDonald and Kreitman, 1991):

| Adh changes | Replacement | Synonymous | Replacement : synonymous |
| --- | --- | --- | --- |
| Fixed between species | 7 | 17 | 0.41 |
| Variable within species | 2 | 42 | 0.05 |

Is the gap chance? Replacements are 7 of 24 fixed differences (29%) but 2 of 44 polymorphisms (5%); their G-test put the chance of so large a gap, if the rows matched, at P = 0.006. The table is partly independent evidence beside omega.

## When the signal misleads

Besides the power problem of gene-wide omega, each test’s model can fail:

- **[McDonald-Kreitman](https://stemhop.com/methods/population-genomic-inference/).** It needs no clock but treats synonymous changes as neutral. Slightly harmful replacement variants linger as polymorphisms but rarely fix, inflating the variable row and hiding adaptation; a change in population size can mimic the signal.
- **Omega within one population.** The differences sampled are mostly recent polymorphisms not yet sorted by selection, so omega near 1 can arise under selection and does not establish neutrality.
- **The synonymous control.** In Enterobacterales, the bacterial group that includes *E. coli*, Rahman and colleagues (2021) found non-neutral synonymous substitutions; treating all as neutral inflated omega, and corrected values averaged 80% of standard estimates.
- **Branch-site tests**, which ask whether some codons changed faster on one chosen branch, assume one base changes at a time, so one mutation hitting adjacent bases looks like rapid amino-acid change. Venkat, Hahn and Thornton (2018) simulated mammal- and fly-based alignments without positive selection; such multinucleotide mutations can make the test report it.

### Design problem

A branch-site test flags positive selection in an immune gene on one primate lineage. Before you believe it, what would you check?

> **One way to do it.** 1. *The data.* Check the alignment around the flagged codons; confirm every sequence is a true ortholog, not a **paralog** (a duplicated copy) or misassembled.
> 2. *The model.* Simulate alignments on the same tree with no positive selection but realistic multinucleotide mutations and base composition; if the test often fires, the result is weak. Check codon-usage bias, since selection on synonymous sites changes dS.
> 3. *A second line of evidence.* Build a McDonald-Kreitman table from polymorphism within the species and fixed differences from a close relative, with an **outgroup** (a more distant species) to place the fixed changes on this lineage. An excess of fixed replacements adds partly independent support, under the same synonymous-neutrality assumption.
> 4. *Function.* Reconstruct and express the ancestral and derived proteins, and measure the property that matters, such as binding to a pathogen protein. Even then, report a model inference supported by evidence, not an observation of adaptation.

## Going deeper

### The mutation-rate numbers

Bergeron and colleagues’ survey used 151 trios. The 36 mammals sampled averaged 7.97 x 10−9 mutations per site per generation, and the four classes sampled did not differ significantly overall.

### Variation already in the ancestor

In a simple model (neutral autosomal genes, a randomly mating diploid ancestral population of constant size, no gene flow after the split), two species’ gene copies trace back to an ancestral copy on average about 2Ne generations older than the species split: with Ne = 10,000 and 20-year generations, about 0.4 million years, 13% of the 3-million-year calibration. So a raw gene divergence runs old. A calibrated date can err the other way: the calibrating gene split is older too, so the rate comes out too high and the second date too young (6.0 million years, where with the same offset the species split was about 6.4).

---

From STEMhop (https://stemhop.com), free lessons by Christopher Deppmann. Lesson text © the authors. Practice questions: https://stemhop.com/lessons/molecular-evolution/
