Contents
08 — Proteins, and what genes actually do Part 1 — Molecular foundations

08 — Proteins, and what genes actually do

Before this: Ch 01 · Ch 06 · Ch 07 · Time: ~40 min

Chapter 07 left a chain of amino acids coming off a . That chain is not yet a protein in any useful sense — it is a string that has to become a machine. This chapter covers how it does that, what the machine then does, how the cell modifies and destroys it, and then dismantles the word "gene", which by that point will have stopped meaning anything precise.

What you'll be able to do

The core idea

A protein is a linear string that folds into a surface. Everything a protein does — catalyse, bind, transport, signal, hold things up — it does by presenting a specific shape with specific chemistry at a specific place. Function is a property of the surface, and the surface is determined by the sequence.

The consequence: the genome specifies shapes indirectly, by specifying strings. Enzyme activity, binding affinity, regulatory logic are all emergent from a one-dimensional encoding. That is why one base change can abolish a function or do nothing, depending entirely on whether the residue it changes is holding up a surface.

The exception, equally important: the string is not the whole specification. The same string is chemically modified after synthesis into a set of functionally distinct molecules, is degraded on a schedule of its own, and in a substantial minority of cases does not fold at all. The map from gene to function is many-to-many in both directions — which is where the word "gene" begins to come apart.


1. Twenty side chains

Every amino acid has the same backbone. What differs is one side chain. The backbone provides the chain; the side chains provide all the chemistry.

ClassResiduesSide-chain chemistryWhat it buys you
Nonpolar aliphaticAla (A), Val (V), Leu (L), Ile (I), Met (M)saturated hydrocarbonBuried in the core — the 's raw material
AromaticPhe (F), Trp (W), Tyr (Y)flat ring systemsCore packing and stacking; Tyr's –OH makes it polar and phosphorylatable
Polar unchargedSer (S), Thr (T), Asn (N), Gln (Q)hydroxyl or amideSurface, hydrogen bonding; Ser and Thr are the main phosphorylation targets
AcidicAsp (D), Glu (E)carboxylate, pKa ≈ 3.9 / 4.3Negative at pH 7.4: salt bridges, metal coordination, catalysis
BasicLys (K), Arg (R), His (H)amine / guanidinium / imidazole, pKa ≈ 10.5 / 12.5 / 6.0Positive at pH 7.4 — except His; DNA binding; Lys is the main modification target
SpecialGly (G), Pro (P), Cys (C)see belowBreak the rules

Histidine's pKa sits near 6.0, within reach of physiological pH, making it the only residue that routinely acts as both acid and base by gaining or losing a proton. That is why it is the workhorse of catalytic sites — the Ser–His–Asp triad of the serine proteases, for one — and why substituting it so often abolishes activity.

The three special residues are special for three unrelated reasons. Glycine has no side chain, just a hydrogen, so it occupies almost no space and adopts backbone geometries nothing else can; a conserved glycine usually marks a tight turn or a packing that admits nothing larger. Proline's side chain loops back onto its own backbone nitrogen, so it cannot donate a backbone hydrogen bond and its rotation is locked — it breaks helices and terminates sheets. It is punctuation. Cysteine's thiol forms the disulfide bond, the only covalent cross-link in ordinary protein structure; because the cytoplasm is chemically reducing, disulfides are effectively confined to secreted and extracellular proteins, which is a useful tell when reading a sequence.

There is a twenty-first: selenocysteine, inserted at certain UGA given a specific downstream RNA structure (Ch 07). The alphabet is not quite closed.

2. Four levels

LevelWhat it isDetermined by
Primarythe amino acid sequencethe gene
Secondarylocal repeating backbone geometry: α-helix, β-sheet, turnslocal sequence propensity
Tertiarythe full 3D fold of one chainthe whole sequence, mostly via the hydrophobic core
Quaternaryassembly of several chains into a complexinterface surfaces

The peptide bond is planar, leaving each residue exactly two rotatable backbone angles. Secondary structure is what you get when a run of residues adopts the same pair of angles. The α-helix: 3.6 residues per turn, 1.5 Å rise per residue, 5.4 Å pitch, right-handed, with a hydrogen bond from the carbonyl of residue i to the amide of i+4 — all backbone hydrogen bonds satisfied internally, which is what makes it stable. The β-sheet: strands nearly fully extended at ~3.3–3.5 Å per residue, hydrogen-bonded sideways to neighbouring strands, with adjacent side chains pointing to opposite faces.

Those geometries hand you a signal-processing result. Map each residue to a hydrophobicity value and look at the periodicity of the vector:

amphipathic α-helix, period 3.6:      β-strand, period 2:

  L K E L E K K L K E L E K K L       V S V T V K V N V
  ^     ^       ^     ^       ^       ^   ^   ^   ^   ^
  gaps of 3 and 4 — every marked      every other residue —
  leucine on the same face            alternating faces

Period ≈3.6 means one face of a helix is greasy — and because 3.6 is not an integer, the greasy positions cannot be evenly spaced: they fall as alternating gaps of 3 and 4. Such a helix lies against a membrane, packs against another helix, or forms a coiled coil (period tightening to 3.5, the classic heptad repeat, hydrophobics at the a and d positions). Period 2 means an amphipathic strand. This is a Fourier problem on a numerical vector, and it was one of the first structural inferences ever made from sequence alone.

Tertiary structure is dominated by burying hydrophobic side chains away from water and leaving polar ones exposed (Ch 01). A protein is a greasy core in a polar shell — which yields the most useful heuristic in variant interpretation: a substitution that puts a charge in the core, or changes the volume of a buried residue, is far more likely to be damaging than the same substitution on the surface.

Quaternary structure matters here for one reason: it is the structural precondition for the mechanism in §10. If a protein works alone, a bad copy is a wasted copy. If it works in a complex, a bad copy can wreck the good ones.

3. Anfinsen, Levinthal, chaperones

Christian Anfinsen denatured ribonuclease A — 124 residues, four disulfide bonds — completely, and let it refold in a test tube with nothing else present. It recovered full catalytic activity. The conclusion, worth a Nobel Prize in 1972: all the information specifying the three-dimensional structure is in the amino acid sequence. The native state is the thermodynamic minimum and folding is spontaneous.

Cyrus Levinthal showed why that cannot be taken literally. Give a 100-residue protein just three accessible backbone conformations per residue: 3¹⁰⁰ ≈ 5 × 10⁴⁷ states. Sample each in 10⁻¹³ s, about the fastest physically plausible rate, and exhaustive search takes

5 × 10⁴⁷ states × 10⁻¹³ s  =  5 × 10³⁴ s  ≈  1.6 × 10²⁷ years

against a universe roughly 1.4 × 10¹⁰ years old. Real proteins fold in microseconds to seconds. Folding is therefore not a search over a flat space: the energy landscape is funnelled, local structure forms fast, near-native contacts are stabilising, and the chain is biased downhill from the start. Anfinsen says the destination exists; the funnel says why it is reachable.

Two things break the test-tube picture in a cell. Folding is co-translational — the chain emerges N-terminus first at ~5–20 residues per second and begins folding before the C-terminus exists, so the pathway depends partly on speed. That is one real mechanism by which a synonymous change, altering elongation rate but not sequence, is not silent. And the cytoplasm is crowded: at 20–30% macromolecule by volume, a partly folded chain with an exposed hydrophobic patch is more likely to meet another partly folded chain than to be left alone. Aggregation competes with folding.

Chaperones address that second problem — Hsp70 binding and releasing exposed hydrophobic stretches in ATP-driven cycles, chaperonins such as GroEL/GroES caging a single chain so it folds in isolation. Note what this is not: chaperones add no structural information. They do not know the target fold and do not instruct. They suppress unproductive interactions and give the chain repeated attempts at the funnel. is untouched; it is the kinetics, not the thermodynamics, that needs help.

4. Domains, families, motifs

Proteins are not monolithic. They are built from domains: independently folding units, typically 100–250 residues, that recur across unrelated proteins in different combinations.

Because a domain folds by itself it can be excised and expressed alone — and, evolutionarily, shuffled. Recombination within can move a whole exon-encoded domain into a new context without disturbing either partner's reading frame — provided the module is flanked by introns of the same (Ch 06, Ch 35). Only phase-symmetric modules (0–0, 1–1, 2–2) can be inserted, duplicated or deleted frame-neutrally; an asymmetric one frameshifts everything downstream of it. That leaves a fossil record: symmetric exons are markedly over-represented among shuffled domains. Tissue plasminogen activator is the standard exhibit: a fibronectin domain, an EGF-like domain, two kringle domains and a serine protease domain, every one of which appears in other, unrelated proteins. A mosaic assembled from parts, not a design.

Biological termRough computational equivalentHow it is detected
Domaina linked library — reusable, self-contained, composableprofile HMM (Pfam), structural comparison
Familyall descendants of one ancestral moduleHMM search, phylogenetics (Ch 34)
Motifa short string constant with meaning — a tag or addressregular expression (PROSITE-style patterns)

The last distinction is worth holding. A motif is short — often 3 to 10 residues — has no independent structure, and works as a label: a nuclear localisation signal, an ER retention signal, an integrin-binding RGD, a phosphodegron marking a protein for destruction. Motifs are cheap to gain and lose by point , which makes them the fast-evolving layer of the interaction network. A domain is expensive and essentially never invented twice.

5. What the surface does

Active sites. Catalytic machinery occupies a small pocket, usually a handful of residues distant in sequence and adjacent in space — which is why active-site residues are the most conserved positions in a family, and why painting onto a structure lights up the functional site. Catalysis works by stabilising the transition state rather than the substrate: the enzyme binds that strained configuration at the top of the barrier — a transition state, not an intermediate, which would sit in an energy well and have a finite lifetime — more tightly than it binds either substrate or product, and that differential binding energy is the rate enhancement.

Binding surfaces. Protein–protein interfaces typically bury 1,200–2,000 Ų, but binding energy is not spread evenly over them. A few hotspot residues contribute most of the affinity — which is why a handful of interface abolish binding and most do nothing.

Allostery. A molecule binds at one site and changes activity at a different site, by shifting the equilibrium between conformational states. Haemoglobin: oxygen binding at one subunit raises the affinity of the others, giving a sigmoid rather than hyperbolic binding curve (Hill coefficient ≈ 2.8), so the molecule loads in the lungs and unloads in tissue over a narrow pressure range.

Allostery is what makes biological regulation possible at all. Without a way to change a protein's behaviour from a site other than its active site there is no signalling, no feedback inhibition, and no way for one molecule to control another. It is the second input port.

6. Post-translational modification: the mutable layer

A given protein molecule's sequence is fixed once it leaves the . Its chemistry is not. Covalent modification of side chains produces functionally distinct molecules from one gene product, reversibly, on timescales of seconds.

ModificationTarget residuesCharge changeTypical meaning
PhosphorylationSer, Thr, Tyr (≈86 : 12 : 2 by abundance)+2 negativeThe universal switch; reversed by
AcetylationLysremoves +1Neutralises a positive charge — loosens histone–DNA binding (Ch 23)
MethylationLys, ArgnoneChanges shape and binding surface, not charge. Lys takes 0–3 methyls: a four-state variable
UbiquitinationLysattaches a 76-residue proteinA routing tag — see §7, where the carries the meaning
GlycosylationAsn (N-linked, sequon Asn-X-Ser/Thr, X ≠ Pro); Ser/Thr (O-linked)—Secreted and surface proteins: stability, recognition, identity

Phosphorylation earns its prominence: a phosphate places two negative charges at a defined position, making or breaking a salt bridge, triggering a conformational change, or creating a binding site a partner domain recognises. Humans encode 518 protein kinases and roughly 200 phosphatases; pairing writers with erasers is what makes the modification a switch rather than a one-way mark.

Two conceptual points matter more than the catalogue. Glycosylation is not template-encoded — the glycan is built by a series of competing enzymes, so its structure depends on enzyme abundance rather than any DNA sequence. There is no gene "for" a glycan, only genes for the enzymes that build it; the ABO blood groups are exactly this (Ch 11). And some modifications are structural, not regulatory: collagen needs the prolines sitting in the Y position of its Gly-X-Y repeat hydroxylated — roughly half of all its prolines — by a vitamin-C-dependent enzyme, or its triple helix is unstable at body temperature. Scurvy is a failure.

7. Degradation, and the two-sided ledger

graph LR P["target protein"] -->|"E1 → E2 → E3<br/>(>600 human E3 ligases)"| UB["poly-ubiquitin chain"] UB -->|"K48-linked"| PR["26S proteasome<br/>unfold, thread, cleave"] UB -->|"K63-linked"| SIG["signalling, DNA repair,<br/>trafficking — <i>not</i> degradation"] PR --> AA["amino acids recycled"]

Ubiquitin is a 76-residue protein appended to lysines by a three-enzyme cascade whose last member, the E3 ligase, supplies substrate specificity — humans encode more than 600. The critical detail: chains are built by linking through one of ubiquitin's own lysines, and the determines the meaning. K48 chains route to the proteasome; K63 chains do not, acting instead in signalling and . Ubiquitin is a typed tag, not a boolean . The 26S proteasome is a barrel with its catalytic sites inside, so a substrate must be unfolded and threaded through a narrow pore — destruction is sequestered, and nothing is degraded by accident.

Now the part worth deriving. Let protein P be made at rate s and degraded first-order at rate k:

dP/dt = s − kP

Steady-state level. Setting dP/dt = 0 gives P* = s / k. Abundance is a ratio, not a synthesis rate. Doubling the degradation rate halves the protein level with no change whatever in — so a mechanism acting only on stability suffices to control expression.

Response time. P approaches a new steady state with time constant 1/k, a half-time of ln2/k — the protein's own half-life, independent of s. A stable protein cannot be regulated quickly however sharply you change its transcription; you have to wait for the existing molecules to disappear. Hence proteins that must respond fast — cyclins, , signalling intermediates — have half-lives of minutes, and buy a short time constant with continuous resynthesis.

Measured half-lives span minutes (ornithine decarboxylase, many cyclins) to years (eye-lens crystallins, collagen), with a median in cultured mammalian cells around 10–20 hours depending on cell type and method. Degrons — the N-terminal residue, PEST sequences, phosphodegrons — are the sequence features that decide where in that range a protein sits.

8. When folding fails

Misfolding causing . CFTR p.Phe508del, the commonest cystic fibrosis , makes a channel that could conduct chloride if it reached the surface, but that folds too slowly; endoplasmic-reticulum quality control degrades it before it gets there. The defect is trafficking — but the protein that is rescued to the membrane also sits there less stably and gates poorly, which is why the correctors that stabilise the fold (tezacaftor, elexacaftor) have to be paired with a potentiator that holds the channel open (ivacaftor) before function is restored.

Misfolding causing toxic . The SERPINA1 Z allele (p.Glu342Lys in classical mature-protein numbering, p.Glu366Lys in HGVS) makes α1-antitrypsin polymerise, producing both loss of function in the lung, where the missing protease inhibitor permits emphysema, and gain of function in the liver, where polymers accumulate and cause cirrhosis. One variant, two mechanisms, two organs.

Self-propagating misfolding. Amyloid — a cross-β fibril many unrelated sequences can adopt — underlies Alzheimer's (Aβ, tau), Parkinson's (α-synuclein) and Huntington's (polyglutamine-expanded huntingtin). Prions are the limiting case: misfolded PrP templates conversion of normal PrP, so the shape is infectious. Information transmitted by conformation rather than sequence is the sharpest counterexample there is to a purely sequence-based view of biology.

9. So what, exactly, is a gene?

Beadle and Tatum's 1941 Neurospora experiments produced one gene — one enzyme: mutate a gene, lose exactly one metabolic step. A superb simplification, worth a Nobel Prize, and now clearly false. It was patched to "one gene — one polypeptide" when haemoglobin turned out to have two different chains. That patch has failed too.

PhenomenonWhat breaksExample
one locus → many productsGENCODE 50 annotates 78,733 genes and 644,292 transcripts — roughly eight per gene (verified facts). Dscam1 in Drosophila has 95 alternative exons in four clusters: 12 × 48 × 33 × 2 = 38,016 possible isoforms from one locus
Non-coding RNA genesthe product is never proteintRNA, rRNA, microRNA, . GENCODE 50 annotates 58,195 non-coding genes (35,885 lncRNA + 7,608 small ncRNA + 14,702 ) against 19,442 protein-coding — non-coding outnumbers coding roughly 3 : 1. Do not reach that count by subtracting from the 78,733 total: a further 1,096 entries — 412 IG/TR segments, 665 readthrough genes and 19 artifacts — sit outside both tallies, and 1,077 of them are protein-coding (verified facts)
Nested genesone gene inside anotherOMG, EVI2A and EVI2B lie entirely within an of NF1, transcribed from the opposite strand
Overlapping genesone stretch of sequence, two genesRoutine in viruses and bacteria; present in eukaryotes on opposite strands and in shifted frames
one transcript, several proteinsThe lac operon: three proteins, one , one mRNA (Ch 21)
Distal regulationthe control element is not in the geneThe switch controlling LCT sits ~14 kb away, inside a neighbouring gene (Ch 00)
RNA editingtranscript ≠ templateADAR enzymes convert adenosine to inosine, read as G. The protein sequence is not the DNA sequence
Alternative promoters and poly(A) sitesthe boundaries are not fixedThe same locus starts and ends in different places in different tissues

Each of these was discovered after the definition was written, and each was absorbed by weakening it. What survives is not a physical object.

A gene is an , not a thing. It is a claim — made by a curation project, revised at each release — that a particular set of overlapping transcripts constitutes one coherent unit and deserves one name. Ask "where does this gene start?" and the honest answer is "in which annotation, of which build, for which transcript?" A gene is an entry in an index, not an object in memory. Ch 44 shows the index being built.

This is not nihilism, and the word stays indispensable. In transmission genetics a gene is a locus that segregates, which is perfectly well defined; in the clinic it is the unit that carries a disease association. The point is that these are different definitions which usually, but not always, coincide — and the cases where they diverge are where the interesting biology is.

10. Loss, gain, and poison

Three mechanisms by which a variant changes what a gene product does. These underpin dominance (Ch 10, Ch 11) and variant interpretation (Ch 55).

(LoF). Less or no working product — nonsense, frameshift, essential splice-site disruption, whole-gene deletion, or a that destroys the fold or active site. Complete loss is an amorph, partial loss a hypomorph (Muller's 1932 terms, still in use).

(GoF). The product does something it should not: constitutive activity (FGFR3 p.Gly380Arg locks a receptor on, causing achondroplasia), a novel activity (neomorph), altered specificity, expression in the wrong place, or a toxic aggregate.

negative (antimorph). The mutant product actively interferes with the normal product from the other . This requires the protein to act as a multimer or to compete for a site or substrate — a defective monomer working alone is merely an absent monomer.

Why loss of function is usually . Not because "one good copy compensates" — the intact allele is generally not upregulated. Because 50% of normal product is usually enough, and Kacser and Burns made that quantitative: control of flux through a multi-step pathway is distributed across all its enzymes, so each has a small flux control coefficient and the flux–activity curve saturates strongly. Halve one enzyme and flux barely moves. Recessivity is a by-product of pathway kinetics, not an adaptation.

is what happens when that argument fails: 50% is not enough, so LoF is dominant. It is characteristic of products that are rate-limiting, needed stoichiometrically in a complex, or dosage-sensitive — PAX6 (aniridia), NSD1 (Sotos syndrome), ELN (supravalvular aortic stenosis). Roughly 3,000 human genes score as strongly loss-of-function-intolerant in metrics: the population-scale way of asking the same question empirically.

graph TD Q0["Missense variant is <b>more severe</b><br/>than a whole-gene deletion"] Q0 --> A["Mechanism is <b>not</b> simple loss of function"] A --> B{"Does the protein act in<br/>a complex or multimer?"} B -->|yes| DN["<b>Dominant negative</b><br/>mutant subunit poisons<br/>wild-type subunits"] B -->|no| GOF["<b>Gain of function</b><br/>constitutive activity,<br/>new activity, or toxicity"] Q0 --> C["Corollary: a truncating variant<br/>may be <b>benign</b> in this gene"]

That diagram is the practical payoff. A predicted-null variant is strong evidence for pathogenicity only if loss of function is an established mechanism for that gene. Where disease arises by or gain-of-function mechanism, a truncating allele can be harmless — and a truncation in the final exon, which escapes , can be worse than an earlier one, because it yields a shortened protein rather than none at all.

11. Structure prediction, and what changed

Structure determination used to be the bottleneck: one protein, one crystal, months to years. CASP — a blind biennial assessment running since 1994, in which predictors submit models for structures not yet released — measured slow progress for two decades. At CASP14 in 2020, AlphaFold2 returned a median GDT_TS around 92 on a 0–100 scale, where ~90 approximates the agreement between two independent experimental determinations of the same protein. The AlphaFold database now covers over 200 million sequences, effectively all of UniProt, with recent releases adding millions of predicted dimers; AlphaFold 3 and its open competitors extend the approach to complexes with nucleic acids and small molecules.

Structure went from a scarce experimental result to an abundant computed . Be precise about what that licenses.

It gives youIt does not give you
A high-quality model of the folded stateThe conformational ensemble — allosteric proteins have several functional states; you get one
Domain boundaries and fold assignment, often better than sequence searchFunction. Structure constrains function; it does not determine it
Per-residue confidence (pLDDT) and inter-domain confidence (PAE)A physical B-factor. Low pLDDT usually intrinsic disorder — information, not failure
A strong prior on where a residue sits: core, surface, interface, active siteA reliable prediction of what a point does

That last row is the one that gets misused. These models predict the structure a folding sequence adopts; they were not trained on stability changes and characteristically return the wild-type fold, confidently, for destabilising substitutions. Variant-effect prediction is a separate problem with separate tools, and those tools are predictors, not measurements.

Two further caveats. Accuracy leans on the depth of the multiple sequence alignment, so performance degrades for sequences with few homologues — precisely the orphan-gene case where you most wanted help. And roughly a third of human proteins contain long segments with no fixed structure; for those, "predict the structure" is the wrong question, because the functional state is the disordered one.


Common misconceptions

What people thinkWhat's actually true
One gene, one proteinGENCODE 50: 78,733 genes, 644,292 transcripts. Add and the count of functionally distinct molecules is larger again
Chaperones tell a protein how to foldThey add no structural information. They suppress aggregation and give the chain repeated attempts. is untouched
Every protein has a fixed 3D structureRoughly a third of human proteins contain long disordered segments, and for many of those disorder is the functional state
Post-translational modification is fine-tuningPhosphorylation can change activity by orders of magnitude and is the switch most signalling runs on. Modified and unmodified forms are functionally different molecules
Protein level is set by Level is synthesis ÷ degradation. Halving stability halves abundance with transcription unchanged — and stability alone sets how fast the protein can respond
A truncating variant is always worse than a Only when the mechanism is . Under a or gain-of-function mechanism a truncation can be milder, or benign
means the mutant protein is strongerDominance describes the heterozygote. It arises from , dominant-negative poisoning, or — three mechanisms with different predictions
Loss of function is because the good copy works harderThe intact is generally not upregulated. Half the product is usually enough because pathway flux is buffered against changes in any single enzyme
AlphaFold solved protein foldingIt predicts structures, extremely well. It does not model the folding process, give the conformational ensemble, or reliably predict what a does
Synonymous changes cannot affect the proteinThey can alter elongation rate and therefore co-translational folding, and can disrupt . Rarely — but the mechanism is real

Worked example: why a missense can be worse than a deletion

Osteogenesis imperfecta is caused by variants in the type I collagen genes. Two classes of variant in the same gene, COL1A1, give at opposite ends of the severity range. Work out why.

The structure. Type I collagen is a triple helix of three chains: two α1(I) chains from COL1A1 and one α2(I) chain from COL1A2. It requires glycine at every third position, because only glycine's hydrogen side chain fits at the crowded helix axis:

   Gly-X-Y-Gly-X-Y-Gly-X-Y-Gly-X-Y ...
    ^       ^       ^       ^
    faces the helix axis — only -H fits

Variant A: a frameshift in COL1A1 triggering . No α1(I) from that , so the cell makes half the normal amount. Every helix it does assemble is structurally perfect; there are simply fewer. Phenotype: OI type I — mild, near-normal stature, elevated fracture rate.

Variant B: a replacing one of those glycines — a Gly→Ser substitution somewhere in the triple-helical domain. The allele is transcribed and translated normally, so half of all α1(I) chains carry the substitution. A helix draws two α1(I) chains independently from that pool:

P(both α1 chains normal)              = ½ × ½ = ¼
P(helix contains ≥1 defective chain)  = 1 − ¼ = ¾

Three quarters of all collagen molecules are defective, against 50% fewer but perfect molecules for the null allele — and defective helices are not merely inert, since they are retained, degraded, and disrupt fibril assembly, pushing effective function below the 25% the arithmetic suggests. Phenotype: severe OI — progressively deforming type III, or perinatally lethal type II for substitutions falling in the lethal regions of the helix. Severity is strongly position-dependent and depends on which residue replaces the glycine, which is why this one class of variant spans most of the OI severity range.

The general result. For a complex containing n copies of the affected subunit with mutant fraction f, the fraction of fully wild-type complexes is (1 − f)ⁿ. For a heterozygote, f = ½:

Copies of the subunit, nFully normal complexesEffectively poisoned
1 (monomer)50%50% — indistinguishable from a null allele
2 (collagen's α1)25%75%
3 (homotrimer)12.5%87.5%
4 (homotetramer)6.25%93.75%

Three conclusions to carry forward. A allele can drive effective activity far below the 50% floor a null allele imposes — "worse than nothing" is literal. Severity scales with subunit count, which is a testable structural prediction rather than a rule of thumb. And therefore "missense more severe than deletion" is diagnostic of mechanism: it is the single most informative observation you can make about a disease gene before running any experiment, and it decides whether a predicted-null variant found in a patient is strong evidence of pathogenicity or nearly meaningless.

Connections