Contents
39 — Genome landscapes and the C-value paradox Part 9 — Genomics

39 — Genome landscapes and the C-value paradox

Before this: Ch 03 · Ch 19 · Ch 35 · Time: ~40 min

Part 9 is about reading genomes. This chapter is about what you find when you have read one and zoom all the way out — the coarse statistics of the object your aligners, assemblers and variant callers will spend the next six chapters fighting with.

Two facts organise everything here. Genome size varies across life by six orders of magnitude and predicts nothing about the organism. Gene number barely varies at all. Both are more interesting than they first appear, and the second is the harder problem.

What you'll be able to do

The core idea

Nobody designed a genome size. Genome size is the equilibrium of a stochastic accretion process — insertion by transposition and duplication, removal by deletion — weakly opposed by selection. Where selection is efficient, the insertions get purged and the genome stays lean. Where selection is inefficient, they to and the genome bloats.

Selection's efficiency is set by . So genome size is, to a first approximation, a readout of demography rather than of biology. That is why it carries no information about complexity: you are looking at a fossil record of a lineage's population history, not at a parts list.

And the parts list, when you finally get it, is disappointingly similar everywhere. A nematode 1 mm long has about as many protein-coding genes as you do. Complexity is not in the number of components. It is in the wiring — which is combinatorial, and therefore cheap in sequence and expensive in explanation.


1. Six orders of magnitude, and no signal

The C-value is the DNA content of one haploid set, historically measured in picograms and now in base pairs. Human 1C ≈ 3.1 Gb.

OrganismHaploid genomeProtein-coding genes
Carsonella ruddii (insect endosymbiont)~0.16 Mb~180
Escherichia coli K-124.6 Mb~4,400
Saccharomyces cerevisiae12 Mb~6,000
Caenorhabditis elegans100 Mb~20,000
Arabidopsis thaliana135 Mb~27,000
Drosophila melanogaster180 Mb~14,000
Takifugu rubripes (pufferfish)~400 Mb~19,000
Homo sapiens3.1 Gb19,442
Allium cepa (onion)~16 Gb—
Triticum aestivum (bread wheat, hexaploid)~16 Gb~107,000 (high-confidence)
Protopterus aethiopicus (marbled lungfish)~130 Gb—
Tmesipteris oblanceolata (fork fern)160.45 Gb—

Stated as a hypothesis test, the is the emphatic rejection of

H₀: genome size increases with organismal complexity (or with gene number).

Three separate features of the data kill it, and it is worth being precise about which:

No monotone relationship across clades. A pufferfish carries roughly your gene count in an eighth of your DNA. A lungfish carries forty times your DNA. An onion carries five times.

Within-clade variance swamps between-clade variance. This is the sharper observation. Species inside a single genus differ several-fold in genome size while being nearly indistinguishable in anatomy and gene content — Allium species, Drosophila species, plethodontid salamanders. Whatever drives genome size operates on a timescale far shorter than morphological divergence, which rules out any explanation running through organismal complexity.

The largest genomes are not in the most complex organisms. They are in ferns, lilies, lungfish and salamanders.

The onion test. T. Ryan Gregory's rhetorical device, and the most efficient bad-argument filter in the field: whatever function you propose for non-coding DNA, explain why an onion needs five times more of it than a human — and why one Allium species needs several times more than its close relative. Any theory of universal function must survive that question, and almost none do. Keep it in your pocket for §5.

Gregory also argues the word "paradox" should be retired. The original paradox — how can genome size not track gene number? — was resolved once non-coding DNA was discovered. What remains is the C-value enigma: not whether the extra DNA is non-genic (it is), but why lineages differ so wildly in how much they accumulate, and whether the amount has any consequence. That question is open.

2. Resolving the paradox: it is nearly all repeat

Write genome size as a sum:

G  ≈  (N_genes × L̄_gene)  +  G_repeat  +  G_other-intergenic

Across eukaryotes, N_genes varies over roughly one order of magnitude. G varies over five. Almost the entire variance is in G_repeat — transposable-element copies and their degraded remains, plus tandem arrays — with a secondary contribution from mean gene length, because expand in large genomes too (Ch 06).

The relationship is close to mechanical: plot repeat content against genome size across species and you get a tight positive relationship, because repeats are the only component with an intrinsic replication mechanism of its own. A is a sequence whose copy number can increase without any benefit to the host (Ch 19); everything else in the genome changes copy number only through duplication events that are comparatively rare.

Whole-genome duplication supplies the second mechanism — a step change rather than accretion, routine in plants and recurrent in vertebrate ancestry (Ch 35). Bread wheat's 16 Gb is partly three genomes in one nucleus; the onion's 16 Gb is not, and is mostly repeat.

Genome size, then, is a repeat-load statistic. That is the resolution, and it immediately raises the population-genetic question of why repeat load differs so much between lineages, which is §8.

3. The deeper puzzle: the G-value paradox

Now look at the gene-count column again, restricted to animals:

C. elegans   (959 somatic cells, 1 mm)      ~20,000 protein-coding genes
D. melanogaster                             ~14,000
T. rubripes                                 ~19,000
H. sapiens   (~3 × 10¹³ cells)               19,442

This is the G-value paradox (Hahn and Wray's term), and it is a much worse problem than the . Genome size not tracking complexity is easy once you know most of the genome is inert. Gene number not tracking complexity attacks the intuition directly: these are the functional units, and a human has no more of them than a millimetre-long worm.

The 2001 draft-genome estimate of 30,000–40,000 human genes was already a downward revision from the 100,000 many expected, and it has fallen further to 19,442 in the current (verified-facts). The number kept shrinking while the organism stayed as complicated as ever.

The resolution has three components, all of them combinatorial — which is why they are invisible in a gene count.

multiplies products per gene. The large majority of human multi-exon genes are alternatively spliced. The scale shows up in the annotation: 644,292 transcripts across 78,733 annotated genes is an average of

644,292 / 78,733 ≈ 8.2 transcripts per annotated gene

(averaged over all gene classes, so protein-coding genes are above this and single-exon small RNA genes below). The record holder is not human: Dscam in Drosophila can generate 38,016 distinct isoforms from one locus by mutually exclusive exon choice. One gene, more protein variants than the fly has genes.

Regulatory elements multiply contexts per gene. ENCODE has catalogued on the order of a million candidate cis-regulatory elements in the human genome — roughly fifty per protein-coding gene (Ch 22). Each is a condition under which the gene is used. The same 19,442 genes deployed in a thousand cell types under distinct regulatory logic is a vastly larger space than 19,442.

Post-transcriptional and post-translational layers multiply again — RNA stability, localisation, microRNA control (Ch 24), and protein modification.

For programmers. Gene count is lines of code. Complexity lives in the call graph, not the line count, and the interesting difference between two systems of similar size is almost always in how the parts are composed. This is also the reason human and chimpanzee proteins are nearly identical while the organisms are not — King and Wilson made exactly this argument in 1975, before anyone could sequence a regulatory element to check it.

4. The human genome, decomposed

The pinned (GENCODE Release 50):

CategoryCount
Protein-coding genes19,442
Long non-coding RNA genes35,885
Small non-coding RNA genes7,608
14,702
IG/TR segments, readthrough genes and artifacts (tabulated separately)1,096
Total annotated genes78,733
Total transcripts644,292

The first four categories sum to 77,637, not 78,733. The remaining 1,096 are 412 immunoglobulin and T-cell-receptor segments, 665 readthrough genes and 19 artifact entries; 1,077 of them GENCODE counts as protein-coding, but it tabulates all 1,096 separately. That is why the non-coding count has to be summed from its parts rather than obtained by subtracting 19,442 from 78,733: the subtraction sweeps those 1,096 coding entities into the non-coding tally and returns 59,291 instead of the correct 58,195.

Two derivations you should be able to reproduce.

Coding sequence is a rounding error. Take a typical protein at ~400 amino acids, needing 1,200 bp:

19,442 × 1,200 bp ≈ 23.3 Mb ≈ 0.75% of 3.1 Gb

Add untranslated regions and alternative exons and the exonic fraction reaches ~1–2%. The figure is soft because "exon" is annotation-dependent, which is exactly why the derivation matters more than the number.

Non-coding genes outnumber coding genes about 3:1.

(35,885 + 7,608 + 14,702) / 19,442 = 58,195 / 19,442 ≈ 3.0

Set against that, ~46% of the genome is transposable-element derived, and more sensitive detection pushes the repeat-derived share past two-thirds.

Those two facts are the whole "" argument in miniature, and they point in opposite directions. Ohno's 1972 coinage rested on a sound inference — population genetics limits how much sequence a species can maintain against , so most of a large genome cannot be under selection. The error was lexical: "junk" collapsed not protein-coding into not doing anything, and 43,493 non-coding RNA genes — the 58,195 total minus the 14,702 pseudogenes, which are junk by anyone's definition and were Ohno's own paradigm case — plus a million candidate regulatory elements say that collapse is wrong. But the over-correction is equally wrong, and it is the one currently in fashion. Resolving which is right requires deciding what "function" means, and that turns out to be the entire argument.

5. The function debate, taken seriously

In 2012 the ENCODE consortium reported that ~80% of the human genome participates in at least one biochemical event in at least one cell type — being transcribed, bound by a protein, marked by a particular histone modification, or sitting in accessible chromatin. Much press rendered this as " is dead". The backlash was ferocious, and it was substantially correct.

The critique, argued most bluntly by Graur and colleagues and more carefully by Doolittle and by Eddy, has three parts:

Biochemical activity is expected of inert DNA. initiates spuriously; bind degenerate motifs by the millions (Ch 01 §5). A genome in an accessible nucleus will show pervasive low-level and binding whether or not any of it matters. The null expectation for the ENCODE assays applied to random sequence is not zero, and the 80% figure was reported without that null.

The threshold was permissive and the union was taken over everything. "At least one event in at least one of many cell types" is a union over a large number of assays; the fraction showing reproducible, cell-type-specific, quantitatively substantial activity is far smaller.

Two different concepts of function were conflated. This is the load-bearing objection.

Causal-role functionSelected-effect function
Question askedWhat does this sequence do?What is this sequence for?
EvidenceBiochemical activity in an assaySequence conserved by
ProvenanceCummins-style functional analysisWright-style etiological account
Human genome estimate~80% (ENCODE)~5–10% (comparative genomics)
Failure modeCounts noise as functionMisses lineage-specific and recently evolved function
Can a qualify?Yes, if transcribedNo

Both definitions are legitimate; they answer different questions. My heart pumps blood (its selected effect) and also makes a thumping noise (a causal role that is not what it is for). ENCODE measured causal roles. The claim that the result overturned junk DNA required the selected-effect sense, and the assays cannot supply it.

The comparative estimate. Align many mammalian genomes and ask which positions change more slowly than the neutral rate. That measures purifying selection directly and is the estimate with evolutionary meaning. It puts the constrained fraction of the human genome at roughly 5–10% — deeper alignments across large numbers of mammals push toward the upper end of that band, and all such estimates are lower bounds on function, because sequence whose presence matters but whose identity does not (spacers, some structural DNA) is invisible to a substitution-rate test, as is anything that evolved too recently to have accumulated a signal.

The mutational-load bound. There is a second, independent line of argument that does not require any alignment, only the and a fertility . It is worked through fully in the worked example below and lands in the same range.

The onion test. Apply it: if 80% of the human genome is functional, the corresponding fraction of the onion's 16 Gb implies an onion needs several times more functional DNA than a human. The test does not disprove anything on its own — it is a plausibility check — but any argument for near-universal function must answer it, and the ENCODE framing never did.

The honest conclusion. "What fraction of the human genome is functional?" has no answer until you fix a definition, and the two available definitions differ by an order of magnitude. Under the causal-role definition, a majority of the genome does something measurable. Under the selected-effect definition — the one that connects to evolution, to disease, and to the question of why the sequence is there — the answer is single-digit to low-double-digit percent. Both numbers are defensible; only one of them is a claim about why the genome looks the way it does. When you read a functional-genomics result, the first question to ask is which definition is doing the work (Ch 49, Ch 52).

6. Repetitive DNA: the classes that break different tools

Repeats are classified by arrangement, not origin.

Tandem repeats — copies adjacent, head to tail

ClassUnitArray sizeWhy it matters
Satellite171 bp (alpha) and others0.1–5 Mb, pericentromeric blocks, Yq. Unassemblable before long reads — most of the ~8% newly resolved by
Minisatellite (VNTR)10–100 bp0.5–30 kbCopy-number-variable; basis of 1980s DNA fingerprinting. are a special case
Microsatellite (STR)1–6 bptens of bp>10⁶ loci, ~3% of the genome. Mutate at 10⁻³–10⁻⁴ per locus per generation by polymerase slippage

STRs earn their own paragraph because their mechanism gives them two unusual uses. Polymerase slippage during replication adds or removes whole repeat units, so a locus becomes highly polymorphic in copy number within a few generations:

locus with a 4 bp unit (AGAT), two alleles in one individual

allele 1   ...CTTA [AGAT]×11 GTCC...     11 repeats
allele 2   ...CTTA [AGAT]×14 GTCC...     14 repeats
                    ^^^^^^^
           genotype recorded as 11,14 — a two-number summary of one locus

Forensics exploits exactly this. ~20 unlinked highly polymorphic STR loci, treat them as independent, and multiply the -frequency products: the random-match probability falls below 10⁻¹⁸ while the assay stays cheap and works on degraded DNA. The statistics are elementary; the power comes entirely from the having generated enough alleles per locus (Ch 57).

Repeat-expansion disease exploits nothing — it is the same mechanism running away. Above some threshold length, slippage becomes strongly biased toward expansion, and the repeat grows across generations, producing : earlier onset and greater severity down a . HTT (CAG, Huntington disease) and FMR1 (CGG, fragile X) are the canonical cases, each with a well-defined allele ladder — though the rungs are locus-specific, and conflating them is a common error. HTT runs normal / intermediate / reduced- / full-penetrance (≤26 / 27–35 / 36–39 / ≥40 CAG); FMR1 runs normal / intermediate / / full mutation. There is no HTT "premutation". Short-read sequencing systematically underestimates long expansions, because you cannot measure a repeat longer than your read (Ch 54).

Interspersed repeats — copies scattered

These are transposable-element relics: LINEs (~21%, dominated by L1 at ~17%), SINEs (~13%, dominated by Alu at ~11%), LTR/endogenous retroviral elements (~8%), and DNA transposons (~3%, none currently active in humans). Class percentages come from different and are approximate; the pinned total for TE-derived sequence is ~46%. Ch 19 covers the biology; the consequence here is that a 150 bp read from a young, low-divergence AluY copy may have many near-identical potential origins, so the aligner emits it with mapping quality 0 and every downstream caller discards it. Be precise about the scope of that claim: Alu copies have diverged 5–20% from one another over ~65 My, so most Alu-derived reads do map uniquely — measured mappability puts ~89% of annotated transposable-element sequence in the uniquely mappable class at 100 bp paired-end. The mapping problem is concentrated in the recently active subfamilies (AluYa5, AluYb8 and kin), not in Alu as a class (Ch 42).

Segmental duplications — the class that causes disease

(low-copy repeats) are blocks >1 kb at >90% sequence identity, present in two or more genomic locations. They cover roughly 5–7% of the human genome depending on assembly and identity threshold, and the complete T2T assembly revised the figure upward because duplications are precisely what short-read assemblies collapse (Ch 43).

Their importance is mechanical. Two near-identical blocks in the same region can misalign during , and recombination between the misaligned copies — non-allelic , — deletes or everything between them:

normal chromosome     ---[SD-A]-----GENES-----[SD-A']---
                            \___ misaligned pairing ___/

NAHR products         ---[SD-A]---                       deletion
                      ---[SD-A]-----GENES-----GENES-----[SD-A']---   duplication

Because the breakpoints are fixed by the positions of the duplications, the resulting rearrangements are recurrent: the same interval, the same size, in unrelated patients. That is the signature of NAHR and it distinguishes these disorders from random structural variation. 22q11.2 deletion syndrome (~3 Mb, flanked by the LCR22 blocks), the PMP22 duplication and reciprocal deletion at 17p12 causing Charcot–Marie–Tooth 1A and hereditary neuropathy with liability to pressure palsies, Williams–Beuren at 7q11.23 and Smith–Magenis at 17p11.2 are all the same mechanism at different loci (Ch 20).

7. Texture: density, isochores, CpG islands, deserts, NUMTs

Zoom to a single arm and the landscape is anything but uniform. Schematically:

position (Mb)   0        10        20        30        40        50
                |---------|---------|---------|---------|---------|
gene density    ▂▂▁▁▁▁▁▁▁▁▁▁▁▁▂▄▆█████▆▄▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▄▆██▆▄▂▁
GC content      ▃▃▃▂▂▂▂▂▂▂▂▂▂▃▄▅▆▇▇▇▇▆▅▄▃▂▂▂▂▂▂▂▂▂▂▂▂▂▂▃▄▅▆▇▇▆▅▃▂
LINE-1 density  ▆▇▇███████▇▇▆▅▄▃▂▂▂▂▃▄▅▆▇▇███████▇▇▆▅▄▃▂▂▂▃▄▅▆▇▇
Alu density     ▂▂▁▁▁▁▁▁▁▁▂▃▄▅▆▇███▇▆▅▄▃▂▁▁▁▁▁▁▁▁▁▁▁▁▂▃▄▅▆▇██▇▅▃
replication     ─── late ──────────  early  ─── late ──────  early
                |________________|            |___________|
                   GENE DESERT                 GENE DESERT
                     (~12 Mb)                    (~9 Mb)
                                (schematic, not a real locus)

Every track here is correlated with every other, and the correlations are the point.

Isochores. The genome averages ~41% GC, but that average is built from long, relatively homogeneous domains — Bernardi's isochores, L1 (<37% GC) through H3 (>53% GC) — spanning hundreds of kilobases. GC-rich isochores are gene-dense, short-introned, Alu-rich, early-replicating and open; GC-poor isochores are the reverse and are L1-rich. Chromosome 19 runs at ~26 protein-coding genes per Mb against a genome average of ~6.

The explanation is not selection for thermal stability, the original guess. It is : when recombination resolves a mismatched heteroduplex, repair is biased toward G and C, so high-recombination regions are pushed GC-ward with no consequence at all. A bias — a non-adaptive, recombination-driven distortion of which is transmitted — that produces a pattern indistinguishable from selection until you test for it properly (Ch 33).

. The dinucleotide CG is strongly depleted genome-wide — observed frequency roughly a quarter of what base composition predicts — because cytosine in a CpG is usually methylated, and 5-methylcytosine deaminates to thymine, producing a G:T mispair in which both bases are perfectly normal DNA. Repair falls to the TDG and MBD4 glycosylases of , which are far less efficient than the uracil glycosylase that clears the U:G product of deaminating an unmethylated cytosine — uracil announces itself as foreign, thymine does not (Ch 17). CpG is therefore a mutational sinkhole, and it has been draining for a hundred million years.

Against that background, regions where CpG is not depleted stand out, and the standard definition of a CpG island is a straightforward statistical filter:

length > 200 bp
GC content > 50%
observed CpG / expected CpG > 0.6      where expected = (N_C × N_G) / L

Roughly 30,000 such islands exist, and they mark the of a majority of human genes. The reason they survived is that they are usually unmethylated, so they never entered the deamination sinkhole — meaning the island is a fossil of the region's methylation history, not merely a compositional oddity. Their methylation state is a primary readout in epigenome profiling (Ch 23, Ch 49), and their GC extremity is a recurring source of dropout in sequencing (Ch 40).

. Intervals of hundreds of kilobases to megabases with no protein-coding gene account for something like a quarter of the genome. They come in two flavours, and the distinction matters: variable deserts, which are lineage-specific and largely unconstrained, and stable deserts, which are conserved across vertebrates, packed with regulatory elements, and sit beside developmental . A conserved gene desert is not empty — it is the regulatory input space of the gene next door, which is why disease variants can act on a gene a megabase away (Ch 52).

. Fragments of have inserted into the nuclear genome repeatedly over evolutionary time. Hundreds of these nuclear mitochondrial DNA segments exist in the human reference, from tens of bp to several kb, and new ones arise at an appreciable rate — on the order of one per few thousand births. They are a pure analysis hazard:

read from a NUMT (nuclear)         ...GATCACAGGTCTATCACCCTATTAACCAC...
true mtDNA sequence                ...GATCACAGGTCTATCACCCTATTAACCAC...
                                                            ^
                            NUMT carries an ancient substitution here

aligner sees two near-identical targets, picks the wrong one
  → apparent low-frequency mtDNA heteroplasmy that does not exist
  → apparent nuclear variant that is really a mitochondrial read

Any pipeline that calls mitochondrial variants — and any that reports low-frequency — must model NUMTs explicitly, usually by aligning to a combined reference and discarding reads that map ambiguously (Ch 46).

8. Why sizes differ: bacteria, streamlining and mutational hazard

Bacterial genomes look nothing like this landscape. Typical coding density is 85–90%, intergenic gaps are tens of base pairs, are essentially absent, and genes with related functions are transcribed as single polycistronic units — (Ch 21). Gene number scales almost linearly with genome size, which is precisely the relationship that fails in eukaryotes.

BacteriaEukaryotes (large-genome)
Coding density85–90%1–2% exonic
Genes per Mb~900~6 (human)
Intronsrareubiquitous, often long
unitpolycistronic operonone gene, many isoforms
Repeat contentlow, actively purgedup to and beyond 50%
Gene number vs genome size~linearuncorrelated

Two directions of small genome, often confused, need separating:

Streamlining by selection. Free-living marine bacteria with enormous census abundances — Prochlorococcus, Pelagibacter — have shed nearly all non-coding DNA. With plausibly around 10⁸–10⁹ — orders of magnitude above any multicellular eukaryote, though far below census abundance and genuinely hard to estimate — even the tiny metabolic cost of replicating an extra kilobase is visible to selection, and it is removed.

Do not read Nₑ off the census count. Marine bacteria are astronomically abundant, and that abundance has repeatedly been translated straight into an enormous Nₑ. It does not translate: and population structure sever the two, and the gap is routinely several orders of magnitude. Prochlorococcus is the cautionary case. A -accumulation experiment put its Nₑ on the order of 10⁷ — smaller than that of less abundant, large-genome relatives — with the implication that its genome reduction was driven partly by rather than purely by selection. The textbook flagship for streamlining is the one example currently under attack.

Reduction by drift. Insect endosymbionts like Carsonella ruddii have genomes an order of magnitude smaller still, and their genomes are degraded: , lost , high substitution rates, absurd base composition. They are transmitted vertically through tiny bottlenecks, so effective population size is minuscule, selection is impotent, and genes are lost by drift rather than trimmed by selection. Same outcome — a small genome — from opposite causes.

That contrast is the whole content of Lynch's mutational-hazard hypothesis, which is the serious population-genetic answer to the C-value enigma:

graph TD NE["<b>effective population size</b> N<sub>e</sub><br/>(and mutation rate)"] EFF["<b>efficacy of selection</b><br/>selection sees a variant only when |s| ≫ 1/N<sub>e</sub>"] INS["insertions: TE copies, introns,<br/>duplications — each carries a small<br/>mutational hazard, s ≈ −10⁻⁶ to −10⁻⁸"] FIX["slightly deleterious insertions<br/><b>drift to fixation</b>"] PURGE["slightly deleterious insertions<br/><b>purged</b>"] BIG["<b>large, repeat-rich genome</b><br/>vertebrates, plants, salamanders"] SMALL["<b>compact genome</b><br/>free-living bacteria, yeast, pufferfish"] NE --> EFF INS --> EFF EFF -->|"small N<sub>e</sub>"| FIX --> BIG EFF -->|"large N<sub>e</sub>"| PURGE --> SMALL

The argument is entirely quantitative and should feel familiar (Ch 27). A non-coding insertion is not neutral — it carries a small replication and mutational cost, giving a perhaps on the order of 10⁻⁶ to 10⁻⁸. Selection can only act on variants with |s| substantially greater than 1/Nₑ. Multicellular eukaryotes have Nₑ in the 10⁴–10⁵ range, so 1/Nₑ ≈ 10⁻⁴–10⁻⁵ — larger than the cost. The insertions are invisible to selection and drift to . Free-living bacteria with Nₑ ≈ 10⁸–10⁹ have 1/Nₑ at or below the cost, and the same insertions are removed.

Genome size is therefore a readout of demographic history, not of biology — the claim from the core idea, now derived. It also explains the within-genus variance that killed the complexity hypothesis in §1: Nₑ changes on a timescale of thousands of generations, far faster than morphology.

The hypothesis is contested — the selection coefficients are hard to measure, and there are competing accounts based on cell size, nutrient economics and DNA replication cost. Treat it as the best-developed quantitative framework rather than as settled.

Organellar genomes, and the paradox in miniature

Mitochondria and chloroplasts carry their own , and they replay the whole argument at small scale. Human mtDNA is 16,569 bp encoding 37 genes, with essentially no non-coding sequence beyond the ~1.1 kb control region — extreme compaction (Ch 03). Chloroplast genomes are larger and remarkably uniform across plants, around 120–160 kb.

Then plant mitochondrial genomes: 200 kb to several megabases, hundreds of times the size of the human mitochondrial genome, encoding roughly the same tiny number of proteins. The extra sequence is repeat, intron and captured DNA. The C-value enigma, complete with an onion test of its own, inside an organelle whose entire gene content you could list on an index card.


Common misconceptions

What people thinkWhat's actually true
Bigger, more complex organisms have bigger genomesThe . An onion has 5× the human genome; a pufferfish has ⅛ of it with the same gene count; a fork fern has 160 Gb. Size tracks repeat load and duplication history
The C-value paradox is unresolvedThe paradox — size not tracking gene number — was resolved by non-coding DNA. The enigma (why lineages differ so much, and whether it matters) is what remains open
Humans have far more genes than simple animalsC. elegans has roughly as many. Complexity is combinatorial: , regulation, and context, none of which appear in a gene count
ENCODE showed 80% of the genome is functionalIt showed ~80% shows some biochemical activity in some cell type — a causal-role claim with no null model. Under the selected-effect definition the estimate is ~5–10%
The 5–10% constrained estimate proves the rest is junkIt is a lower bound. Sequence whose presence matters but whose identity does not, and anything recently evolved, is invisible to a substitution-rate test
is a discredited ideaOhno's inference — that a species cannot maintain unlimited sequence against — was correct and is still correct. The word was bad; the population genetics was not
Non-coding means non-functionalNon-coding genes outnumber coding ones ~3:1, and regulatory, structural and RNA-gene sequence includes a great deal that is unambiguously functional. Do not over-correct: for the ~36,000 annotated genes specifically, function is demonstrated for a small minority and most show little sequence . "Annotated as a gene" is not evidence of function
Repeats are all the same problemEach class breaks a different tool: interspersed repeats destroy mapping uniqueness, tandem arrays defeat assembly, cause recurrent disease-associated rearrangements by
A is emptyConserved gene deserts are dense with regulatory elements and sit beside developmental genes. Emptiness is a statement about , not about function
Small genomes mean efficient organismsSmall genomes arise both from selection (streamlined marine bacteria, huge Nₑ) and from (degraded endosymbionts, tiny Nₑ). Same outcome, opposite causes
Mitochondrial reads come from mitochondriaHundreds of make nuclear reads look mitochondrial and vice versa — the standard source of phantom

Worked example: an upper bound on the functional fraction

Can 80% of the human genome be functional in the selected-effect sense? Test it with nothing but the and arithmetic. This is the Graur-style mutational-load argument, stated with its assumptions exposed.

Step 1 — per zygote. From the pinned rate of 1.1–1.3 × 10⁻⁸ per bp per generation across a 6.2 Gb genome:

6.2 × 10⁹ × 1.1 × 10⁻⁸ ≈ 68
6.2 × 10⁹ × 1.3 × 10⁻⁸ ≈ 81        → call it ~70 new SNVs per zygote

Step 2 — how many land in functional sequence and matter. Let f be the functional fraction and δ the probability that a inside functional sequence is deleterious. The deleterious mutation rate per zygote is

U = 70 × f × δ

Step 3 — turn U into a fertility requirement. Under the classical Haldane–Muller model, with deleterious mutations acting independently and multiplicatively on , mean fitness relative to a mutation-free is e^−U. For the population to hold constant size, each couple must produce, on average,

R = 2 e^U      children

Step 4 — evaluate the ENCODE claim. Set f = 0.80 and, generously to ENCODE, δ = 0.5:

U = 70 × 0.80 × 0.5 = 28
R = 2 × e²⁸ ≈ 2 × 1.4 × 10¹² ≈ 3 × 10¹² children per couple

Three trillion children per couple. The claim is not merely wrong, it is wrong by twelve orders of magnitude.

Step 5 — invert it into a bound. Instead of testing a value of f, solve for the largest one the arithmetic permits. Take a pre-modern completed family size of ~5 children, so R ≤ 5 and e^U ≤ 2.5, giving U ≤ ln 2.5 = 0.92:

f ≤ 0.92 / (70 × δ)

δ = 0.5   →   f ≤ 0.026    (2.6%)
δ = 0.2   →   f ≤ 0.065    (6.5%)
δ = 0.1   →   f ≤ 0.131    (13.1%)

Step 6 — read the result honestly. The bound lands at a few per cent to ~13% depending on δ, bracketing the comparative-genomics estimate of 5–10% from a completely independent direction — no alignments, no scores, just mutation and demography. Two methods sharing no assumptions and agreeing is worth more than either alone.

Now the caveats, because a bound is only as good as its model:

Even with every caveat applied generously, f = 0.8 is unreachable. That is the useful conclusion, and it is robust to everything the model gets wrong.

Connections