Contents
10 — Mendel and single-gene inheritance Part 2 — Transmission genetics

10 — Mendel and single-gene inheritance

Before this: Ch 09 · Ch 08 · Time: ~45 min

Statistics needed: S1 Probability · S2 Distributions

What you'll be able to do

The core idea

Mendel reverse-engineered a data structure and a sampling process from nothing but counts.

The structure: each individual carries two copies of each heritable factor, and they do not blend or contaminate each other — a factor that vanishes in one generation reappears intact in the next. The process: each gamete receives exactly one of the two, chosen with probability ½, and fertilisation combines two independent draws.

So the at a locus is an unordered pair; transmission is a fair Bernoulli draw over that pair; and the is a many-to-one function of the genotype, which is what makes the underlying variable partly unobservable. Everything in this chapter is that model, plus its corrections.

The remarkable thing is that this was inferred in the 1860s from seed counts, four decades before anyone knew existed and nearly a century before anyone knew what a gene was made of. Mendel was not describing a mechanism. He was fitting a generative model, and the model turned out to be physically instantiated in .


1. The experiment, read as an experimental design

Mendel worked on the garden pea, Pisum sativum, across eight growing seasons from the mid-1850s, publishing in 1866. Nearly every choice he made is one you would make today.

Design choiceWhy it matteredWhat breaks without it
True-breeding linesA line that produces only round seeds, generation after generation, is homozygous. The starting are known exactlyWithout a known prior, F2 ratios are unidentifiable
Discrete, binary traitsRound or wrinkled. Yellow or green. No measurement error, no judgement call, no scaleContinuous traits hide the underlying 1:2:1 in noise — see Part 6
A plant that both selfs and outcrossesPea flowers self-pollinate inside a closed keel, so lines stay pure with no effort; but the anthers can be removed and foreign pollen applied by handYou cannot control the mating design
One factor at a time, then combinationsMonohybrid crosses first, dihybrid only afterConfounded factors, uninterpretable ratios
CountingHe tabulated thousands of individuals per crossThis is the actual innovation. Predecessors described offspring; Mendel enumerated them
Reciprocal crossesSame result whether the pollen or the ovule parent carried the traitCannot rule out maternal effects or sex- (Ch 13)

He screened 34 pea varieties, kept 22 that bred true, and worked with seven binary traits (tabulated with his counts in §3). Two of them — seed shape and cotyledon colour — are scoreable on the seed itself, which is already F2 tissue while still sitting on the F1 mother plant. That is why those two crosses have counts in the thousands and the rest have counts in the hundreds.

Notation. P is the true-breeding parental generation; F1 the first filial generation; F2 the offspring of F1 × F1 (in pea, usually an F1 self). The notation is meaningless unless P is genuinely true-breeding.

2. Vocabulary, defined once

TermDefinition
LocusA position in the genome. Plural loci. Not the same as a gene, though for this chapter treat it as one
One of the alternative sequences found at a locus
The pair of alleles an individual carries at that locus (or, loosely, the whole sequence)
The observable characteristic
HomozygousTwo identical alleles: AA or aa
HeterozygousTwo different alleles: Aa
Only one copy present, with no partner — a male at most loci (Ch 13), or either sex where one homolog carries a deletion. A hemizygous locus has nothing to be or to; whatever allele is there is expressed
DominantDescribes an allele whose phenotype appears in the heterozygote
RecessiveDescribes an allele whose phenotype appears only in the homozygote
Monohybrid / dihybrid crossA cross tracking one locus / two loci
Wild typeThe reference allele or phenotype. A convention, not a value judgement

Two conventions worth fixing now. Uppercase denotes the dominant allele, lowercase the recessive, of the same letter — A and a are alleles of one locus; A and B are different loci. And A_ means "AA or Aa" — the genotypes indistinguishable by phenotype.

Critically: dominance is a property of a pair of alleles with respect to a particular phenotype, at a particular level of description. It is not an intrinsic property of an allele. The same allele can be recessive for disease, codominant for a protein assay, and dominant for a physiological stress response (Ch 11).

3. The first law: segregation

Mendel's seven monohybrid F2 counts:

Trait formn formnRatio
Seed shaperound5,474wrinkled1,8502.96
Cotyledon colouryellow6,022green2,0013.01
Flower colourpurple705white2243.15
Ripe pod shapeinflated882constricted2992.95
Unripe pod colourgreen428yellow1522.82
Flower positionaxial651terminal2073.14
Stem lengthtall787dwarf2772.84
Total14,9495,0102.98

Every F1 resembled one parent exactly — no blending. The other form vanished, then returned in the F2 at close to one quarter.

The model that produces this is one line of probability. Let the F1 be Aa. A gamete carries A with probability ½ and a with probability ½. Two gametes are drawn independently, so the zygote distribution is the product measure:

P(AA)=\tfrac14,\quad P(Aa)=\tfrac12,\quad P(aa)=\tfrac14

That is 1:2:1 in genotype. The map collapses AA and Aa onto the same value, giving 3:1 in phenotype.

graph TD P["<b>P</b> &nbsp; AA &nbsp;×&nbsp; aa<br/><i>true-breeding</i>"] F1["<b>F1</b> &nbsp; all Aa<br/><i>all show the dominant phenotype</i>"] F2["<b>F2</b> &nbsp; 1 AA : 2 Aa : 1 aa<br/><i>3 dominant : 1 recessive</i>"] P -->|"each parent gives one allele"| F1 F1 -->|"self, or F1 × F1"| F2

The law of . The two at a locus separate during gamete formation, each gamete receiving exactly one, each with probability ½. Nothing about one allele changes the other while they coexist.

The physical event is anaphase I of , where the two are pulled to opposite poles (Ch 09). Mendel had no access to this. He inferred the sampling process from its marginal distribution.

Note what the 3:1 costs you: the phenotype is a lossy projection of the genotype. Two-thirds of dominant-phenotype F2 individuals are heterozygous, and no amount of looking at them will reveal it. Recovering the hidden state requires either a breeding experiment (§7) or sequencing.

4. Punnett squares, and why to abandon them immediately

The lays out one parent's gametes on each axis and fills the cells:

A (½)a (½)
A (½)AA ¼Aa ¼
a (½)Aa ¼aa ¼

This is an outer product of two gamete distributions, drawn out in full. That is fine for one locus and a disaster for any more:

Heterozygous loci nGamete types 2ⁿPunnett cells 4ⁿF2 classes 3ⁿF2 classes 2ⁿ
12432
241694
3864278
4162568116
5321,02424332
101,0241,048,57659,0491,024

Enumerating the full joint distribution to answer a question about one marginal is exactly the mistake you would never make in code.

Two rules do all the work instead:

From an Aa × Aa cross the per-locus numbers you need are just these:

QueryProbability
genotype AA¼
genotype Aa½
genotype aa¼
phenotype (A_)¾
phenotype (aa)¼

Everything else is products and sums of those five numbers.

5. The second law: independent assortment — and exactly where it applies

Cross a true-breeding round-yellow line to a true-breeding wrinkled-green one. The F1 are all round and yellow. Self them, and Mendel's F2 counts were:

ObservedExpected (9:3:3:1)
round, yellow315312.75
round, green108104.25
wrinkled, yellow101104.25
wrinkled, green3234.75
Total556556

You do not need a 4×4 grid to get 9:3:3:1. If the two loci are independent, the joint phenotype distribution is the product of two 3:1 marginals:

(3:1)\otimes(3:1) \;=\; 9:3:3:1

and in general the n-locus F2 phenotypic ratio is the expansion of (3+1)^n, summing to 4ⁿ.

The physical basis is which way each happens to face at metaphase I. For two loci on different pairs, the orientation of one bivalent tells you nothing about the other:

Two loci on DIFFERENT chromosome pairs — orientations are independent

   arrangement 1                  arrangement 2
   A ─────── a                    A ─────── a
   B ─────── b                    b ─────── B
        ↓                              ↓
  gametes: AB, ab                gametes: Ab, aB

  both arrangements equally likely
     →  AB : Ab : aB : ab  =  1 : 1 : 1 : 1

Scope. is not a general law of inheritance. It holds for loci on different chromosomes, and approximately for loci far apart on the same chromosome — far enough that crossovers between them are near-certain and the recombinant fraction approaches ½. For loci close together it is simply false: they are linked, and travel together. That failure is not an annoyance; it is the entire basis of genetic mapping (Ch 14) and of association studies a century later (Ch 29).

Which raises the obvious question about Mendel. Pea has seven chromosome pairs and he studied seven traits — but the genes are not one per chromosome; they fall on four of the seven. Blixt's 1975 analysis found that of the 21 possible pairwise combinations, four involve genes on the same chromosome, and three of those four are far enough apart that no would be detectable at his sample sizes. That leaves exactly one pair that should have shown linkage — and Mendel never published a cross using it. Whether that is luck or selective reporting is the same question §11 raises about his numbers.

6. Scaling up: a trihybrid without a 64-cell grid

Cross AaBbCc × AaBbCc, three independently assorting loci. The has 64 cells. Do not draw it.

Phenotypic ratio. Expand (3+1)^3 = 27 + 9\cdot3 + 3\cdot3 + 1:

A_B_C_  27      A_B_cc   9      A_bbcc  3      aabbcc  1
                A_bbC_   9      aaB_cc  3
                aaB_C_   9      aabbC_  3
                                                  total 64

Any specific query, by multiplication. No grid, no enumeration:

QuestionComputationAnswer
P( for all three)¾ · ¾ · ¾27/64 ≈ 0.42
P( for all three)¼ · ¼ · ¼1/64 ≈ 0.016
P( exactly AABbcc)¼ · ½ · ¼1/32
P(dominant for A and B, recessive for C)¾ · ¾ · ¼9/64
P(recessive at ≥1 locus)1 − (¾)³37/64 ≈ 0.58
P(homozygous at all three loci)(½)³1/8

The last one is worth pausing on: at each locus, P(AA) + P(aa) = ½, so full homozygosity across n loci is (½)ⁿ — which is why inbred lines take many generations to fix (Ch 28).

This is the whole technique. Factorise the query across loci, evaluate each factor from the five-number table in §4, multiply. It runs in O(n) where the grid runs in O(4ⁿ).

7. Test cross and backcross

The F2 problem: a round-seeded plant is RR or Rr and you cannot tell by looking. The test cross solves it — cross the unknown to a homozygous (rr).

Why that particular partner: the tester can only contribute r. Its gamete distribution is a point mass, so it adds no variance and no ambiguity, and the offspring are a direct, unbiased sample of the unknown parent's gametes.

Unknown isCrossOffspring
RRRR × rrall round
RrRr × rr½ round, ½ wrinkled

Any wrinkled offspring at all proves the parent was Rr. Absence of wrinkled offspring is evidence, not proof, and its strength is computable. Under Rr, the probability that n offspring are all round is (½)ⁿ, so the likelihood ratio favouring RR after n all-round offspring is 2ⁿ: seven offspring gives 128:1.

Statistics: why evidence is measured by a ratio of likelihoods rather than by either likelihood on its own is covered in S6 §4; turning that ratio into a probability needs a prior, S1 §5.

Compare that with the lazy alternative of selfing the unknown. Under Rr, a self gives wrinkled offspring with probability ¼, so n all-round offspring has probability (¾)ⁿ:

DesignP(no recessive offspring | heterozygous)n needed to get below 5%
Test cross (× rr)(½)ⁿ5
Self (× Rr)(¾)ⁿ11

That factor of two in required sample size, per locus, is why test crosses exist.

Dihybrid test cross. AaBb × aabb gives 1:1:1:1 if the loci assort independently. Because the tester contributes a fixed gamete, each offspring phenotype is a readout of one parental gamete — which makes this the standard design for measuring directly (Ch 14).

Backcross. Crossing an F1 to either parental line. A test cross is the special case where that parent is the recessive homozygote. Backcrossing to the parent is uninformative about the F1's but is the workhorse of strain construction: repeated backcrossing to a recipient line moves one into an otherwise uniform genetic background (Ch 37).

8. Family composition: the binomial

Offspring of the same cross are independent draws — separate meioses, separate fertilisations — so the number affected in a sibship of n is \text{Binomial}(n, p) with p fixed by the cross.

Statistics: the binomial, its two load-bearing assumptions, and the binomial coefficient are covered in S2 §1, with this exact sibship calculation at §1.1.

For Aa × Aa and a trait, p = ¼. In a family of four:

P(k=1) = \binom{4}{1}\left(\tfrac14\right)^1\left(\tfrac34\right)^3 = 4\cdot\frac{27}{256}=\frac{27}{64}\approx 0.42
P(k\ge 1) = 1-\left(\tfrac34\right)^4 = \frac{175}{256}\approx 0.68

Distinguish ordered from unordered queries, because they differ by the binomial coefficient. "The first two children affected and the last two not" is (\tfrac{1}{4})^2(\tfrac{3}{4})^2 = 9/256. "Exactly two of four affected, in any order" is \binom{4}{2} times that, = 54/256.

Two consequences that matter more than the arithmetic:

has no memory. A couple who have had three children with a recessive condition still face probability ¼ for the fourth. The gambler's fallacy is the single most common error in genetic counselling conversations, and the biology is unambiguous: each meiosis samples afresh.

Observed ratios in human families are biased, and predictably so. Families are usually ascertained because an affected child exists, which conditions on k ≥ 1 and throws away every Aa × Aa family that by chance had none. The naive estimate of p comes out well above ¼. Correcting for this is a real statistical problem (Ch 12, Ch 15) and the reason human analysis is harder than pea genetics.

9. Dominant does not mean common

This is the load-bearing correction of the chapter.

does not mean common, strong, advantageous, healthy, or more likely to be inherited. It means one thing: the heterozygote's resembles one homozygote's rather than falling between them. It is a statement about the map from to phenotype, and it carries no information whatever about , severity, or .

The counterexamples are stark:

InheritanceFrequencyEffect
HTT CAG expansion (Huntington's)DominantPrevalence of the order of 5 per 100,000 in European-ancestry populations; published estimates range roughly 2–10Fatal adult-onset neurodegeneration
FGFR3 p.(Gly380Arg) (achondroplasia)DominantOf the order of 1 in 25,000 births; ~80% arise as new Skeletal dysplasia
ABO O alleleThe most common ABO allele in nearly every human population, typically 0.6–0.7; group O is the commonest phenotype worldwide, around 45%None
CFTR F508del (cystic fibrosis)Recessive ≈ 1 in 25 in Northern European ancestrySevere multisystem disease — but only in homozygotes

Frequency is set by mutation, selection, and migration (Ch 27) — an entirely different set of processes from the ones that set dominance. The two are orthogonal.

They are not, however, uncorrelated in one specific direction, and the reason is worth deriving. A deleterious dominant allele is exposed to selection in every carrier, so it is removed roughly as fast as mutation supplies it and stays rare. A deleterious recessive is invisible to selection in heterozygotes, and at low frequency almost every copy is in a heterozygote. Under Hardy–Weinberg (Ch 26) the ratio of carriers to affected individuals is

\frac{2pq}{q^2} = \frac{2p}{q} \approx \frac{2}{q}\quad\text{for small } q

For cystic fibrosis in a Northern European population, q ≈ 0.02: about 100 carriers for every affected individual, and 98% of CFTR F508del copies sitting in people selection cannot see. That is how a severely deleterious allele reaches a frequency a lethal dominant never could.

10. What dominance is, molecularly

Why is the default for loss-of-function ? Not by convention — by enzyme kinetics.

Model the flux J through a pathway step as a saturating function of the amount of enzyme E:

J(E) = J_{\max}\,\frac{E}{E+K}

A heterozygote for a null allele makes half the enzyme, so

\frac{J(E/2)}{J(E)} = \frac{E+K}{E+2K}
Operating pointFlux retained at half dose
E = 100K (far into saturation)99.0%
E = 10K91.7%
E = K66.7%
E = 0.1K (well below saturation)52.4%

Most enzymes sit well into saturation, and in a multi-step pathway the control over flux is distributed across many enzymes, so no single one has much leverage. Halving it changes almost nothing measurable, the heterozygote looks wild type, and the null allele is therefore recessive. This is the Kacser–Burns argument (1981), and it settled a long dispute: Fisher had proposed that dominance evolved through selection on modifier genes; Wright objected that such selection is far too weak; the physiological answer is that dominance of the functional allele mostly needs no explanation at all. It falls out of saturation kinetics.

All four of Mendel's traits whose molecular basis was known before 2025 fit this exactly — every recessive is a :

TraitGene productThe recessive allele
Seed shapestarch-branching enzyme IDisrupted by a transposon-like insertion. Less amylopectin → sucrose accumulates → the seed takes up extra water and shrivels on drying (Ch 19)
Cotyledon colourSGR ("stay-green"), required to degrade chlorophyllLoss of function — chlorophyll is retained, cotyledons stay green
Flower coloura bHLH in the anthocyanin pathwayLoss of function — no pigment, white flowers
Stem lengthgibberellin 3β-hydroxylaseReduced-activity variant — less bioactive gibberellin, dwarf plants

The remaining three (pod colour, pod shape, flower position) were resolved in 2025; the yellow-pod allele turns out to be a ~100 kb deletion beside the chlorophyll synthase gene. A 160-year-old dataset finished being explained about a year ago.

The four ways to be dominant

graph LR V["one altered copy,<br/>one wild-type copy"] V --> H["<b>Haploinsufficiency</b><br/>50% product is not enough"] V --> G["<b>Gain of function</b><br/>new or constitutive activity"] V --> D["<b>Dominant negative</b><br/>mutant poisons the complex"] V --> T["<b>Toxic species</b><br/>aggregation, novel damage"]
MechanismWhy one copy sufficesExample
The gene is dosage-sensitive — it sits outside the saturating regime, or its product is needed in fixed stoichiometry. Transcription factors, ribosomal proteins and structural subunits are enriched herePAX6 (aniridia), SHOX, RUNX1. scores such genes computationally (Ch 55)
The question is not how much protein but what it does. A constitutively active receptor signals regardless of the normal copyFGFR3 p.(Gly380Arg) in achondroplasia — >95% of cases carry the identical nucleotide change
negativeThe mutant subunit is incorporated into multimers and ruins them. Type I collagen is α1(I)₂α2(I)₁, so with a heterozygous COL1A1 variant and equal expression only (½)² = ¼ of trimers have two normal α1 chains — ¾ of the collagen is defective from a 50% mutant allele doseOsteogenesis imperfecta. Note the is worse than a null allele, which merely halves normal collagen
Toxic speciesThe variant product does damage the wild type cannot undoHTT CAG expansion: ≥40 repeats is fully penetrant, 36–39 shows reduced

The row is the one to internalise: it explains the counterintuitive clinical fact that a variant can be far more severe than a complete deletion of the same gene.

11. Mendel's numbers are too good

Look again at §5. Expected 312.75, observed 315. Expected 34.75, observed 32. Compute the goodness of fit:

\chi^2 = \sum \frac{(O-E)^2}{E} = 0.0162 + 0.1349 + 0.1013 + 0.2176 = 0.470

with 3 degrees of freedom, giving p ≈ 0.93. Nothing wrong with that on its own — a large p means only that the data fit.

Statistics: the χ² distribution, where the degrees of freedom come from (including this 9:3:3:1 case), and why its expected value equals its df are covered in S2 §4; what a p-value does and does not license is S4 §3.

Fisher (1936) pooled the goodness-of-fit statistic across all of Mendel's reported experiments and obtained χ² = 41.6 on 84 degrees of freedom. The expected value of χ² is its degrees of freedom. Mendel came in at half of it.

The relevant probability is the lower tail: the chance of a fit this close or closer is about 3 × 10⁻⁵ — roughly one in 35,000. (Fisher quoted the complementary figure, P = 0.99993, from the tables of his day.)

Note the shape of the argument, because it generalises. The hypothesis under test is not "Mendel's theory is wrong". The test conditions on the theory being right and asks whether the residuals are as large as the sampling model requires — and data can fail it by being too tidy. Candidate explanations include unconscious stopping rules (counting until the ratio looked right), an assistant supplying the answer he believed was wanted, and genuine ambiguity in scoring seeds. No evidence of deliberate fabrication has ever surfaced, and the conclusions themselves have replicated for 160 years. Ch 12 treats this properly.

The transferable lesson: a p-value near 1 is as much a diagnostic as a p-value near 0. Both say the data do not look like a draw from the model you claimed.


Common misconceptions

What people thinkWhat's actually true
are more common, or strongerDominance describes only the heterozygote's . Huntington's is dominant and rare; ABO O is and the most common allele in nearly every population
A dominant allele overpowers or suppresses the recessive oneBoth alleles are transcribed and translated normally. The recessive one usually just fails to produce a working product, and half the normal amount of the working one is enough
A 3:1 ratio means 3 of every 4 offspringIt is the expected value of a multinomial. A sibship of four is very often 4:0 or 2:2. P(exactly 3:1 in a family of four) is only 27/64
is a law of inheritanceIt is a law about loci on different . Linked loci violate it flagrantly, and that violation is the basis of all genetic mapping
are how genetics problems are solvedThey are an outer product drawn by hand. They scale as 4ⁿ. Multiply per-locus probabilities instead
After three affected children, the next is saferEach is an independent draw. The probability is unchanged. This is the gambler's fallacy with a family attached
Observed human sibships confirm the ¼ ratio directlyThey do not, because families are ascertained through an affected child. The raw estimate is biased upward and must be corrected
A trait that skips a generation must be recessiveUsually, but incomplete produces the same pattern from a dominant allele (Ch 11)
A complete gene deletion is the worst possible variant variants are often more severe, because the mutant product actively poisons the wild-type product. Compare a COL1A1 null (mild) with a COL1A1 glycine substitution (severe)
Mendel's laws describe how genes workThey describe how alleles are transmitted. They say nothing about what a gene does, and Mendel's model is agnostic about mechanism — which is why it survived the discovery of DNA intact

Worked example

Three independently assorting pea loci. (Mendel's own symbols are R/r, I/i and Le/le; I use R, Y, T here for legibility.)

Setup. True-breeding RRYYTT × true-breeding rryytt.

Step 1 — the F1. Each parent contributes one per locus. Every F1 is RrYyTt: round, yellow, tall.

Step 2 — F1 gametes. Three heterozygous loci, independently assorting, so 2³ = 8 gamete types, each with probability ⅛:

RYT   RYt   RyT   Ryt   rYT   rYt   ryT   ryt

Step 3 — F2 proportions. The would be 8 × 8 = 64 cells. Factorise instead. Per locus, P() = ¾ and P() = ¼.

Phenotype classProductFraction
round, yellow, tall¾·¾·¾27/64
round, yellow, dwarf¾·¾·¼9/64
round, green, tall¾·¼·¾9/64
wrinkled, yellow, tall¼·¾·¾9/64
round, green, dwarf¾·¼·¼3/64
wrinkled, yellow, dwarf¼·¾·¼3/64
wrinkled, green, tall¼·¼·¾3/64
wrinkled, green, dwarf¼·¼·¼1/64

27:9:9:9:3:3:3:1, which sums to 64. ✓

Step 4 — expected counts. In 640 F2 plants: 270 round-yellow-tall, 90 in each of the three single-recessive classes, 30 in each double-recessive class, 10 wrinkled-green-dwarf.

Step 5 — a specific . P(RRYyTt) = P(RR) · P(Yy) · P(Tt) = ¼ · ½ · ½ = 1/32. Twenty of the 640.

Step 6 — a compound query. "Round and tall, but not homozygous at the Y locus":

P = \underbrace{\tfrac34}_{R\_}\cdot\underbrace{\tfrac12}_{Yy}\cdot\underbrace{\tfrac34}_{T\_} = \frac{9}{32}

Step 7 — the test cross. Cross an F1 to rryytt. The tester contributes only ryt, so offspring phenotypes read out F1 gametes one-for-one: eight classes in 1:1:1:1:1:1:1:1. Any departure from that — and in a real pea it would depart, because two of Mendel's loci are syntenic — is evidence of (Ch 14).

Step 8 — a family-composition question. Among 6 randomly chosen F2 seedlings, what is P(exactly 2 wrinkled)? Each seedling is an independent Bernoulli trial with p = ¼:

\binom{6}{2}\left(\tfrac14\right)^2\left(\tfrac34\right)^4 = 15\cdot\frac{1}{16}\cdot\frac{81}{256} = \frac{1215}{4096} \approx 0.297

Step 9 — test the fit. Suppose you actually observe, in 640 plants: 262, 96, 88, 94, 27, 34, 29, 10. Compute Σ(O−E)²/E against the expected counts from Step 4, with 8 − 1 = 7 degrees of freedom. A moderate χ² supports ; a large one points to linkage, viability differences, or scoring error — and a suspiciously small one points at whoever collected the data (Ch 12).

Connections

Back to:

Forward to: