Contents
35 — Genome evolution, duplication and orthology Part 7 — Molecular evolution

35 — Genome evolution, duplication and orthology

Before this: Ch 33 · Ch 34 · Ch 27 · Time: ~55 min

Chapters 33 and 34 treated genes as fixed objects whose sequences and diverge. That assumption is false over deep time. Genes are duplicated, lost, fused, moved, imported from other species, and occasionally created from nothing. This chapter is about the genome as a mutable data structure rather than a fixed-length string.

What you'll be able to do

The core idea

Point cannot invent a new gene. It can only edit an existing one, and an existing gene is usually doing something, so most edits are rejected. The way out of this trap is redundancy: make a second copy first, and one copy is then free to be edited without cost.

Almost every gene family in every genome is the visible residue of that process. But redundancy is a wasting asset. A second copy is selectively invisible precisely because it is redundant, so the mutations that destroy it are neutral, and the default fate of a is decay into a . Gene family evolution is a race between decay and the acquisition of a reason to exist.

The quantitative content of this chapter is that race, written down and solved.


1. Three ways to duplicate

MechanismHowStructural signature
Unequal / Misaligned repeats pair during ; crossover yields one chromatid with a duplication and one with a deletion (Ch 18)Tandem array, copies adjacent and same-orientation; reciprocal deletion product exists
RetrotranspositionProcessed mRNA reverse-transcribed and inserted, using machinery (Ch 19)Retrogene: intronless, poly-A tail, flanking target-site duplications, lands anywhere — usually on a different , usually without its
Whole-genome duplication (WGD)Failed meiotic reduction produces unreduced gametes; autopolyploidy or allopolyploidyEvery gene duplicated at once; doubled synteny blocks genome-wide

The three are not interchangeable. Tandem duplication copies the gene and its neighbourhood, including its regulatory elements, so the copy starts life expressed in the same place at the same time — the setup for . Retrotransposition copies only the spliced message, so the copy arrives naked, with no promoter and no ; most retrogenes are dead on arrival, and the ones that survive did so by capturing a nearby promoter, which usually gives them a different expression pattern. WGD everything simultaneously, including every member of every protein complex, which turns out to be the key to why WGD duplicates are retained differently from single-gene duplicates.

2. The fates of a duplicate pair

Ohno's 1970 framing: a is a spare part, freed from selective , on which evolution can experiment. Three outcomes.

graph TD D["duplicate pair<br/>born redundant"] D -->|"null mutation fixes<br/>in one copy"| NF["<b>nonfunctionalisation</b><br/>pseudogene — the usual outcome"] D -->|"beneficial mutation<br/>fixes in one copy"| NEO["<b>neofunctionalisation</b><br/>new function; needs 4N<sub>e</sub>sφ &gt;&gt; 1"] D -->|"complementary degenerative<br/>mutations fix in both"| SUB["<b>subfunctionalisation</b> (DDC)<br/>both copies now required<br/>— entirely neutral"] D -->|"stoichiometry constrains<br/>relative copy number"| DOS["<b>dosage balance</b><br/>loss of either copy is deleterious"]

Nonfunctionalisation, and why it wins by default

Immediately after duplication both copies are functional and either is dispensable. So a null in either copy is neutral. Under neutrality the rate equals the (Ch 33): with 2N copies of the locus segregating, null mutations arise at rate 2N·un and each fixes with probability 1/(2N), so nulls fix at rate

\lambda_{\text{null}} = u_n

independent of population size. The duplicate is on a clock it cannot see.

Neofunctionalisation, and why it needs a large population

Now let ub be the rate of mutations conferring a useful new function with s. Beneficial mutations arise at rate 2N·ub and each fixes with probability ≈ 2s (Haldane's approximation, Ch 27):

\lambda_{\text{neo}} = 2N u_b \cdot 2s = 4N u_b s

These are competing Poisson processes, so the probability that wins the race is the ratio of rates:

P(\text{neo}) = \frac{4Nu_bs}{4Nu_bs + u_n} = \frac{4Ns\varphi}{4Ns\varphi + 1}, \qquad \varphi \equiv \frac{u_b}{u_n}

Everything hangs on 4Nesφ. Set φ = 10⁻⁴ — there are roughly ten thousand ways to break a gene for every way to give it a useful new job — and s = 0.01:

LineageNe4NesφP(neo)
Vertebrate~10⁴0.043.8%
Insect / unicellular eukaryote~10⁷4098%

So Ohno's mechanism is efficient in large populations and nearly useless in small ones. But vertebrates — small Ne — retain duplicates in abundance. Something else is preserving them.

Subfunctionalisation and the DDC model

Force et al.'s duplication–degeneration–complementation model supplies it. Suppose the gene has n independently mutable regulatory subfunctions — separable driving expression in separate tissues (Ch 22) — plus a coding region whose loss kills the whole gene. Let ur be the null rate per regulatory element and uc the coding null rate. Take n = 2 and follow the process.

Step 1 — the first fixation. Neutral targets: 2 copies × (1 coding region + 2 elements). Total rate 2(uc + 2ur). A coding hit ends the process in nonfunctionalisation; a regulatory hit does not. So

P(\text{regulatory first}) = \frac{4u_r}{2u_c + 4u_r} = \frac{2u_r}{u_c + 2u_r}

Say it knocked out element 1 of copy A.

Step 2 — the constraint has changed. Copy B's element 1 is now the only source of subfunction 1, so mutations there are deleterious and no longer fix. Likewise copy B's coding region is now indispensable. The remaining neutral targets are copy A's coding region (uc), copy A's element 2 (ur), and copy B's element 2 (ur). Only the last produces complementary loss:

P(\text{sub} \mid \text{step 1}) = \frac{u_r}{u_c + 2u_r}

Result. With r ≡ ur/uc:

P(\text{sub}) = \frac{2u_r}{u_c + 2u_r} \cdot \frac{u_r}{u_c + 2u_r} = \frac{2u_r^2}{(u_c + 2u_r)^2} = \frac{2r^2}{(1+2r)^2}

Check the limits: r → 0 (no separable regulatory modules) gives 0; r → ∞ (coding region immune) gives 1/2. At r = 1, P(sub) = 2/9 ≈ 22%, and P(sub) rises with n.

Look at what is missing from that formula: N. preserves duplicate genes by degrading them, through mutations that are individually neutral. It therefore works exactly as well in a population of ten thousand as in a population of ten million — which is why it, and not Ohno's adaptive mechanism, is the plausible default explanation for the duplicate-rich genomes of small-Ne vertebrates. Preservation by decay is the counterintuitive core of this chapter.

Twenty-two percent versus four percent, for a vertebrate. And the two mechanisms are not exclusive: a subfunctionalised pair is preserved long enough for neofunctionalisation to become possible later.

Dosage balance

A fourth mechanism, and the one after WGD. Many proteins act in stoichiometric complexes. If a subunit's copy number changes alone, the complex misassembles and the excess subunit aggregates — so losing one copy of a duplicated subunit is deleterious, not neutral. The prediction is a reciprocal retention pattern, and it is observed: genes encoding complex subunits, ribosomal proteins, and signalling components are preferentially retained after WGD (where the whole complex doubled together) and preferentially lost after single-gene duplication (where it did not). Dosage sensitivity is also why is so damaging (Ch 20).

3. Dating duplications: the Ks clock

Synonymous divergence Ks between two is approximately clock-like (Ch 33). Divergence accumulates on both lineages, so for a duplication of age t with synonymous rate rs per site per year:

= 2 r_s t \quad \Longrightarrow \quad t = \frac{K_s}{2 r_s}

Now derive the expected shape of a genome-wide Ks histogram. are born at some steady rate β per gene per unit time and lost at rate λ, so the density of surviving pairs of age t is f(t) ∝ β·e−λt. Change variables with t = Ks/(2rs):

f(K_s) \propto \frac{\beta}{2r_s} \exp\!\left(-\frac{\lambda K_s}{2 r_s}\right)

Steady duplication and loss produce a monotonic exponential decay in Ks, and nothing else. Therefore any peak — an excess of pairs sharing one age — requires a burst of simultaneous duplications. That is the standard evidence for whole-genome duplication, and it is how the ancient WGDs in plant lineages were found. Two caveats: Ks saturates above roughly 1–1.5, blurring old events beyond dating, and between paralogs resets Ks toward zero, making some pairs look spuriously young.

4. Three gene families worth knowing

Globins are the textbook case because every step is legible. The α-globin cluster sits on 16, the β-globin cluster on chromosome 11 — a duplication old enough that a chromosomal rearrangement has since separated them. Within each cluster, tandem duplication produced further copies, and the genes sit in the order in which they are switched on:

chr11  5'--[HBE1]---[HBG2]-[HBG1]---[HBD]--[HBB]--3'
         embryonic    fetal            adult
chr16  5'--[HBZ]---[HBM]-[HBA2]-[HBA1]-[HBQ1]--3'
         embryonic         adult

That is and regulatory evolution in one picture: fetal haemoglobin binds oxygen more tightly than adult haemoglobin, which is how a fetus extracts oxygen across the placenta. The products diverged; so did the switches that select between them. Both clusters also contain — the third fate, preserved in place.

Olfactory receptors are the largest human gene family: roughly 800 loci, of which only about 400 retain intact reading frames. Over half the family is pseudogenised, against roughly 1,100 functional receptors in mouse. This is not degeneration in the pejorative sense; it is relaxed . Primates shifted to vision, the cost of losing any one receptor fell toward zero, and nonfunctionalisation — always the default — was no longer opposed.

HOX clusters record whole-genome duplication directly. The invertebrate chordate amphioxus has one cluster; humans have four (HOXA, HOXB, HOXC, HOXD) on four different chromosomes, in the same internal gene order. Two rounds of WGD near the base of the vertebrates — the 2R hypothesis — explain 1 → 2 → 4 exactly, and the same 1:4 pattern appears across many other vertebrate families. Teleost fish underwent a third round (3R, dated variously between about 230 and 330 million years ago) and carry up to seven clusters; salmonids added a fourth (~80 Mya) and are still rediploidising. Plants have done this repeatedly and recently enough that WGD peaks are stacked in many angiosperm Ks histograms.

5. Concerted evolution

Tandem arrays behave strangely: copies within one species are more similar to each other than to their in a sister species. That is the opposite of what independent divergence predicts, and the phylogenetic signature is unmistakable:

independent divergence            concerted evolution (observed for rDNA)
 ┌─ human copy1 ─┐                 ┌─ human copy1 ─┐
 ├─ mouse copy1 ─┘ (by position)   ├─ human copy2 ─┘ (by species)
 ├─ human copy2 ─┐                 ├─ mouse copy1 ─┐
 └─ mouse copy2 ─┘                 └─ mouse copy2 ─┘

The cause is continual homogenisation within the array by and unequal (Ch 18). If the per-copy homogenisation rate exceeds the , variation is erased faster than it accumulates and the array evolves as a unit.

The human rDNA arrays are the standard example: a ~43 kb repeat unit tandemly arrayed in the nucleolar organiser regions of the five acrocentric (13, 14, 15, 21, 22), a few hundred copies per haploid genome and highly variable between individuals. Hundreds of identical rRNA genes is exactly what a cell that must build millions of needs, and concerted evolution is what keeps them identical. These arrays were among the last regions of the human genome to be assembled (Ch 45).

Contrast birth-and-death evolution, the alternative regime: new copies arise by duplication, old ones die by pseudogenisation, and members diverge freely. Olfactory receptors, immunoglobulins and MHC genes follow this pattern. The diagnostic is the same phylogeny: birth-and-death families group by , not by species.

6. Orthology and paralogy, rigorously

Fitch's definitions are about events, not similarity, and nearly every practical error comes from forgetting that.

TermDefinitionConsequence
Homologs whose most recent common ancestor is a speciationUsually the best candidate for shared function
Homologs whose MRCA is a duplicationFunction may have diverged deliberately
In-paralogParalogs that duplicated after a given speciation eventBoth are co-orthologs of the single gene in the other species
Out-paralogParalogs that duplicated before that speciationNot co-orthologs; pairing them across species is an error
OhnologParalogs generated by whole-genome duplicationDosage-balanced, syntenically detectable
XenologHomologs whose history includes horizontal transfer contradicts species tree
graph TD A["ancestral gene"] -->|"<b>duplication</b>"| B["gene α"] A -->|"<b>duplication</b>"| C["gene β"] B -->|"speciation"| B1["human α"] B -->|"speciation"| B2["mouse α"] C -->|"speciation"| C1["human β"] C -->|"speciation"| C2["mouse β"] C1 -->|"<b>duplication</b>"| D1["human β1"] C1 -->|"<b>duplication</b>"| D2["human β2"]

In that tree: human α and mouse α are orthologs. Human α and human β are out-paralogs, and so are human α and mouse β. Human β1 and β2 are in-paralogs relative to the human–mouse split, and both are co-orthologs of mouse β — orthology is a one-to-many relation, not a bijection.

Inference

MethodHowFails when
(RBH)a ∈ A and b ∈ B are orthologs iff each is the other's top BLAST/DIAMOND hitDifferential gene loss; rate asymmetry; one-to-many after WGD (RBH returns at most one pair); partial assemblies; domain shuffling inflating scores
Synteny-basedRequire conserved gene-order context around the candidate pairRearrangement-rich lineages; very deep comparisons where synteny has eroded
Tree-basedBuild a gene tree, reconcile it against the species tree, label each internal node duplication or speciationGene-tree error propagates; needs good alignments and taxon sampling; expensive

RBH's sharpest failure is pseudoorthology from differential loss. Suppose the ancestor had paralogs X1 and X2; lineage A lost X2 and lineage B lost X1. RBH confidently pairs A-X1 with B-X2 and labels them orthologs. They are paralogs, and their divergence predates the speciation. If each lineage independently ends up single-copy with probability q — retaining either paralog with equal chance — then both lineages are single-copy with probability q², and half of those cases retain different paralogs, so the probability of reciprocal differential loss is q²/2, which for q = 0.4 is 8%. This is not an edge case.

The ortholog conjecture

The reason any of this matters practically: essentially all functional of non-model genomes is transferred along orthology assignments (Ch 44). That rests on the ortholog conjecture — that orthologs retain function better than paralogs at equal sequence divergence, because a duplication event is where functional change is licensed and a speciation event is not.

It has been contested. A 2011 analysis using Gene Ontology annotations found paralogs more functionally similar than orthologs; the rebuttal was that GO annotations are biased, since within-species paralogs are typically annotated by the same experiments in the same lab and inherit each other's terms. Later analyses using expression data, which does not have that bias, support the conjecture — but weakly. The defensible summary: orthologs are on average somewhat more functionally conserved than paralogs, the effect is smaller than the confidence with which annotation pipelines apply it, and one-to-one orthologs are far safer than members of large expanded families. Treat a transferred annotation as a prior, not a fact.

7. Synteny, rearrangement and breakpoint reuse

Synteny in modern usage means conserved gene order along a . It decays over time through , , fissions and fusions, but slowly enough that human and mouse — ~90 million years apart — still share a few hundred conserved blocks.

Nadeau and Taylor's 1984 random breakage model assumed rearrangement breakpoints fall uniformly, which predicts an exponential distribution of conserved-segment lengths — and from a handful of markers it correctly estimated ~180 human–mouse segments. Then whole-genome comparison showed the model was wrong in an interesting way: breakpoints cluster and are reused far more often than uniformity permits. Breakpoint regions are enriched for and repetitive elements, which is mechanistically sensible, because between dispersed repeats is precisely what generates rearrangements. Genome architecture is not a passive substrate; it biases its own future rearrangements. The same repeat-mediated fragility drives recurrent structural variation in human disease and in cancer (Ch 56).

8. Sideways, and from nothing

Horizontal gene transfer (HGT) moves genes between contemporaneous organisms rather than from parent to offspring. In prokaryotes it is routine, by three mechanisms: transformation (uptake of naked environmental DNA), transduction (accidental packaging by a bacteriophage), and conjugation (direct plasmid transfer through a pilus). Consequences:

In eukaryotes HGT is rare but real — well-supported cases include bdelloid rotifers, plant-to-plant transfer between parasites and hosts, and a handful of metabolic genes acquired by insects. It also has an unusually bad evidence record. The 2015 report that a tardigrade genome was ~17% foreign was, on reanalysis of an independent assembly, almost entirely bacterial contamination; a contemporaneous claim of extensive HGT into the human genome did not survive scrutiny either. The prior on a surprising eukaryotic HGT claim should be low until contamination is excluded — by an independent assembly, by and -usage evidence that the gene is integrated into a host , and by phylogenetic placement.

De novo gene birth — protein-coding genes originating from previously non-coding sequence — was long dismissed on the argument that a random ORF is useless. It happens anyway. The route is a continuum rather than a jump: pervasive low-level transcription of intergenic regions, occasional translation of the resulting transcripts, and selection acting on those few that happen to help, gradually converting a "proto-gene" into a gene. Confirmed cases now exist in Drosophila, yeast, mouse and human, typically identified by finding an intact ORF in one lineage aligned to unambiguously non-coding, non-deleted syntenic sequence in outgroups. Young de novo genes are short, weakly expressed, tissue-restricted (often testis), and mostly not essential — which is what the proto-gene model predicts. Estimated rates remain uncertain; the phenomenon is not.

9. Genome size, junk, and two kinds of pangenome

Eukaryotic genome size spans five orders of magnitude and correlates with nothing about organismal complexity — the . Some salamanders and lungfish carry genomes tens of times larger than ours; the largest measured eukaryotic genome, reported in 2024 for a fork fern, is on the order of 160 Gb. What genome size does correlate with is content (~46% of the human genome, reference/verified-facts.md), and inversely with .

Lynch's mutational-hazard hypothesis explains that second correlation without invoking any benefit to extra DNA. Excess DNA imposes a small deleterious cost: every added base is another target at which a can disrupt something, plus a replication cost. A stretch of excess DNA with k positions where mutation causes harm carries a cost of roughly

s \approx -u \cdot k \quad \Longrightarrow \quad |N_e s| \approx k \cdot N_e u

So the quantity that decides the outcome is Neu — and it has to be evaluated per lineage, because u is not a universal constant. The human rate is 1.3 × 10⁻⁸ per bp per generation, a figure that sums ~300 cell divisions plus unrepaired chemical damage. A bacterial generation is one replication, so the bacterial rate is the per-replication fidelity, ~10⁻¹⁰ per bp (verified facts) — about a hundredfold lower. Carrying the human number across to bacteria is the standard way to get this section wrong by two orders of magnitude. Selection can act only when |Nes| ≳ 1 (Ch 27):

LineageNeu per bp per generationNeuk needed for |Nes| ≳ 1
Bacteria~10⁸~10⁻¹⁰ (one generation = one replication)~10⁻²~100
Human~10⁴1.3 × 10⁻⁸ (per generation, ~300 divisions)1.3 × 10⁻⁴~7,700

A bacterium therefore purges an insertion whose mutational hazard is a hundred-odd positions; we do not notice one of several thousand. Run the human through it: k = 30 critical positions (splice donor, acceptor, branch point) gives s ≈ −3.9 × 10⁻⁷ and |Nes| ≈ 0.0039 — invisible, so it to . The same k = 30 in a bacterium gives |Nes| ≈ 0.3: also under the barrier, but short of it by a factor of three rather than a factor of 250.

Be honest about the size of the effect. Ne differs ~10,000-fold between the two lineages, but u differs ~100-fold in the opposite direction, so the net difference in what selection can resolve is ~100-fold, not 10,000-fold. That is still ample to separate a streamlined genome from a bloated one, and the sign of the difference is set by population size — but the drift barrier is the term, not the only one.

Genome bloat is therefore not an adaptation and not a failure. It is what happens when a lineage's population size falls far enough that selection can no longer see the cost of carrying junk. The C-value paradox dissolves into population genetics (Ch 39 develops the composition side).

Reductive evolution is the mirror image. Obligate endosymbionts of insects have the smallest cellular genomes known — down to roughly 110–160 kb and fewer than 200 genes, an order of magnitude below free-living bacteria. The mechanism is not streamlining by selection for efficiency: these lineages have tiny Ne (transmitted through a handful of cells per host generation, with no recombination), so drift dominates, Muller's ratchet turns, genes are among the early losses, and any gene made redundant by the host is deleted irreversibly. Small Ne expands eukaryotic genomes and shrinks endosymbiont ones because the deletion bias in bacteria runs the other way — in both cases drift, not selection, sets the outcome.

Two things called "pangenome"

Sequence many strains of one bacterial species and the gene content differs enormously. The pan-genome partitions into a core genome present in all strains and an accessory genome present in some — the latter largely acquired by HGT, and the reason two E. coli strains can share only around 60% of their genes while one causes no disease and the other kills you.

Whether the pan-genome is bounded is an empirical question with a clean answer. If the number of new genes contributed by the Nth genome sequenced decays as n(N) = kN−α, then

P(N) \approx P(1) + \int_1^N kx^{-\alpha}dx = P(1) + \frac{k\left(N^{1-\alpha}-1\right)}{1-\alpha}

which diverges as N → ∞ when α < 1 (an open pan-genome) and converges to P(1) + k/(α−1) when α > 1 (closed). Species with broad ecological range and active HGT — E. coli, Streptococcus — are open: every additional genome still contributes new genes.

This word collides. In Ch 45 the human "" is a graph-structured reference representing many so that reads carrying non-reference sequence can be aligned. The bacterial pan-genome is a statement about gene content variation between strains. One is a data structure for alignment; the other is a biological claim about a species' gene repertoire. Same word, unrelated concepts.


Common misconceptions

What people thinkWhat's actually true
A gene is a spare that evolution can freely improveIts redundancy is exactly what makes null neutral. The default outcome is decay to a , and preservation requires an active explanation
Duplicates are preserved because one acquires a new function requires 4Nesφ ≫ 1 and is inefficient in vertebrates. preserves duplicates by purely neutral degradation, and dosage balance preserves them by making loss deleterious
are just "the same gene in another species"Orthology is defined by the event at the MRCA, not by similarity or by best-hit score. It is frequently one-to-many, and the highest-scoring cross-species hit is often not the ortholog
give you orthologsRBH returns at most one pair, so it silently discards co-orthologs, and differential loss makes it confidently report as orthologs — plausibly in tens of percent of families
Large genomes belong to complex organismsGenome size spans five orders of magnitude with no complexity correlation. It tracks TE content and inverse Ne
Streamlined bacterial genomes prove selection for efficiencyIn endosymbionts with tiny Ne, reduction is driven by and deletion bias, not by selection for economy. Their genomes are degrading, not optimising
HGT into animals is common — look at the tardigradeThat result was contamination. Eukaryotic HGT is real but rare, and reported cases require an independent assembly before belief
The human and the bacterial pan-genome are the same ideaUnrelated. One is a graph reference for alignment; the other is core-vs-accessory gene content across strains

Worked example: reading a Ks histogram

A plant genome with 30,000 protein-coding genes. All pairs are collected and binned by synonymous divergence. Given: synonymous rate rs = 6.1 × 10⁻⁹ per site per year.

pairs
 900 |*
     | *
 450 |   *
     |      *                                    ┌───┐  3,300 pairs
 225 |         *                                 │   │
     |             *   *    *    *    *    *  ┌──┘   └──┐
   0 +---+---+---+---+---+---+---+---+---+---+---+---+---+--  Ks
      0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.85 0.9 1.0

1. Characterise the background. Counts halve every ΔKs = 0.1 (900 → 450 → 225), so the background is exponential — consistent with steady duplication and loss, as derived in §3. Convert that halving interval to time:

Δt = ΔKs / (2rs) = 0.1 / (2 × 6.1 × 10⁻⁹) = 0.1 / 1.22 × 10⁻⁸ = 8.2 × 10⁶ years

So small-scale in this genome have a half-life of about 8.2 My — the same order as the ~4 My reported across several eukaryotic genomes.

2. Test whether the bump at Ks = 0.85 is real. Extrapolate the background: from 0.1 to 0.85 is 0.75/0.1 = 7.5 halvings, and 27.5 = 181.0, so

expected background = 900 / 181.0 ≈ 5 pairs, observed 3,300 pairs

An excess of about 660-fold. Steady duplication cannot produce this. It is a whole-genome duplication.

3. Date it.

t = Ks / (2rs) = 0.85 / (1.22 × 10⁻⁸) = 6.97 × 10⁷ ≈ 70 million years ago

4. Retention. Mind the denominator: the 30,000 genes seen today are 3,300 surviving pairs (6,600 genes) plus 23,400 singletons, and each singleton is the survivor of a pair. So the pre-WGD ancestor had 23,400 + 3,300 = 26,700 genes, and the WGD created 26,700 pairs — not 30,000, which is the post-loss count:

retention = 3,300 / 26,700 = 12.4%

5. Apparent loss rate of the WGD duplicates. If loss were a constant-rate process, 0.124 = e−λt:

λ = −ln(0.124) / 70 My = 2.091 / 70 = 0.0299 per My → t½ = ln2/λ = 0.693 / 0.0299 = 23 My

6. Interpret the discrepancy. WGD duplicates show a 23 My half-life; small-scale duplicates in the same genome show 8.2 My. Two non-exclusive readings, and both are real effects:

The methodological lesson generalises: an exponential fit to survivors is only a half-life if the hazard is constant, and in duplicate-gene loss it demonstrably is not.

Connections