Problem set 09 — Molecular evolution
Covers Ch 33–35.
Attempt before revealing. Every problem here is a null model plus a departure from it, and the arithmetic is the argument — reading it does not transfer.
Problems are roughly in order of difficulty. ★ marks the two worth returning to. Parts flagged (trap) carry a specific conceptual error that the solution names explicitly.
1. The cancellation at the heart of neutral theory
A evolves neutrally. Use the pinned rate μ = 1.2 × 10⁻⁸ per bp per generation (verified facts). Take two populations, N = 10⁴ and N = 10⁶.
(a) How many new neutral enter each population per site per generation? (b) What is the probability of any one of them, and why? (c) Multiply. State the substitution rate k for each population. (d) A coding gene in the same genome has only f₀ = 0.15 of its new mutations neutral. Two lineages separate for T = 5 × 10⁵ generations. Compute divergence for pseudogene and gene. (e) (trap) A colleague redoes (c) using "the replication error rate, 10⁻¹⁰". By what factor does the answer change, and what did they get wrong?
Solution
(a) A diploid population of N individuals carries 2N copies of each site, each mutating at μ:
N = 10⁴: 2 × 10⁴ × 1.2 × 10⁻⁸ = 2.4 × 10⁻⁴
N = 10⁶: 2 × 10⁶ × 1.2 × 10⁻⁸ = 2.4 × 10⁻²
The large population generates 100× more raw material.
(b) Under alone is a martingale, and the process absorbs at 0 or 1, so E[p∞] = p₀ = 1·P(fix) + 0·P(loss). Hence P(fix) = 1/(2N) — a new mutant is one copy out of 2N. That gives 5 × 10⁻⁵ and 5 × 10⁻⁷: exactly 100× smaller.
(c) k = 2Nμ × 1/(2N):
N = 10⁴: 2.4 × 10⁻⁴ × 5 × 10⁻⁵ = 1.2 × 10⁻⁸
N = 10⁶: 2.4 × 10⁻² × 5 × 10⁻⁷ = 1.2 × 10⁻⁸
Both equal μ. The 2N cancels algebraically: k = μ regardless of N, of bottlenecks, or of whether N is even constant. A beneficial mutation fixes with probability ≈ 2s instead, giving kb = 4Nμbs — nothing cancels, so adaptive substitution scales with N.
(d) k = f₀μ and both lineages accumulate substitutions, so d = 2kT.
pseudogene: d = 2 × 1.2 × 10⁻⁸ × 5 × 10⁵ = 0.012
gene: k = 0.15 × 1.2 × 10⁻⁸ = 1.8 × 10⁻⁹
d = 2 × 1.8 × 10⁻⁹ × 5 × 10⁵ = 0.0018
The ratio 0.0018/0.012 = 0.15 = f₀. A ratio of divergences recovers the neutral fraction — what dN/dS does in problem 3, with synonymous sites as the pseudogene.
(e) The trap, and the most expensive unit error in the subject. The answer falls 120-fold, and every divergence time built on it inflates the same way.
- 10⁻¹⁰ is per base per replication — one round of copying, after base selection, and .
- 1.2 × 10⁻⁸ is per base per generation — zygote to gamete, summing a few hundred germline divisions and including unrepaired chemical damage that was never a polymerase error.
Check: 10⁻¹⁰ × ~300 divisions ≈ 3 × 10⁻⁸ against a measured 1.2 × 10⁻⁸ — it overshoots, because 10⁻¹⁰ is an order-of-magnitude bound. And k = μ is a per-generation statement, because 1/(2N) is a per-generation fixation probability.
2. Jukes–Cantor and the nonlinearity of the correction
Two aligned sequences of L = 1,000 sites.
(a) Compute d = −¾ ln(1 − 4p/3) for p = 0.05 and p = 0.45, expressing each correction as a percentage of p. (b) Repeat for p = 0.72. Then attempt p = 0.78 and explain what happens. (c) (trap) Using Var(d) = p(1−p)/[(1 − 4p/3)²L], compute the standard error at each of p = 0.05 and 0.45. Which estimate is more precise?
Solution
(a)
p = 0.05: 4p/3 = 0.0666667 ; 1 − 4p/3 = 0.9333333 ; ln = −0.0689929
d = −0.75 × (−0.0689929) = 0.0517446 correction = +3.5%
p = 0.45: 4p/3 = 0.6000000 ; 1 − 4p/3 = 0.4000000 ; ln = −0.9162907
d = −0.75 × (−0.9162907) = 0.6872181 correction = +52.7%
A ninefold rise in observed difference produces a 13.3-fold rise in inferred substitutions. Observed difference is a compressed function of elapsed change; undoing it is super-linear.
(b) p = 0.72: ln(0.04) = −3.2188758, d = 2.4142 — a 235% correction; every site has been hit twice on average.
p = 0.78: 4p/3 = 1.04, so 1 − 4p/3 = −0.04 and the logarithm is undefined. Not a bug: under JC69 two infinitely diverged sequences differ at 3/4 of sites, so 78% is worse than random — evidence the model is wrong, not evidence of enormous distance.
(c)
p = 0.05: Var = 0.0475/(0.8711111 × 1000) = 5.4528 × 10⁻⁵ ; sd = 0.007384
p = 0.45: Var = 0.2475/(0.16 × 1000) = 1.546875 × 10⁻³ ; sd = 0.039330
The trap. "p = 0.05, its error is five times smaller" compares absolute errors on quantities differing 13-fold. Relative: 14.3% (0.007384/0.051745) against 5.7% (0.039330/0.687218) — the distant pair is more precise, and countably so: 50 sites differ at p = 0.05, giving noise ~1/√50 = 14%; 450 differ at p = 0.45, giving ~1/√450 = 4.7%, inflated to 5.7% by the correction. Precision improves with divergence until the (1 − 4p/3)² denominator explodes near saturation.
3. dN/dS: normalise or don't bother
An aligned 900 bp coding region (300 ) between two species, with 225 synonymous sites and 675 nonsynonymous sites. Observed: 45 synonymous and 27 nonsynonymous differences.
(a) (trap) A first pass reports "27 nonsynonymous versus 45 synonymous, ratio 0.60." What is wrong with it? (b) Compute pN, pS, then dN, dS and ω. Interpret. (c) Compare ω with the uncorrected pN/pS. Which class needed the bigger correction, and which way does skipping it bias ω? (d) Suppose 5 codons are genuinely under positive selection at ω = 8 and 295 sit at ω = 0.1. Compute gene-wide ω, then find how strong those 5 would have to be for gene-wide ω to exceed 1.
Solution
(a) The trap is that raw counts are not rates. The code does not offer the two classes equal opportunity: roughly one site in four is synonymous and three in four nonsynonymous, because third positions are degenerate and second positions almost never are. A gene under no selection at all accumulates about three times as many nonsynonymous changes, so the neutral expectation for the raw ratio is ~3, not 1 — meaning a raw 0.60 is already five-fold depleted. Normalising converts counts into rates per opportunity: an exposure offset in a Poisson regression, and just as non-optional.
(b)
pS = 45/225 = 0.200000
pN = 27/675 = 0.040000
dS = −0.75 × ln(1 − (4/3)(0.20)) = −0.75 × ln(0.7333333) = −0.75 × (−0.3101549) = 0.2326162
dN = −0.75 × ln(1 − (4/3)(0.04)) = −0.75 × ln(0.9466667) = −0.75 × (−0.0548082) = 0.0411062
ω = 0.0411062 / 0.2326162 = 0.1767
ω ≈ 0.18: strong — roughly 82% of amino-acid-changing removed before they could fix.
(c) Uncorrected pN/pS = 0.04/0.20 = 0.200, overstating ω by 13.2%. Synonymous needed 0.2326/0.200 = +16.3%, nonsynonymous only 0.0411/0.040 = +2.8%. The fast class saturates first, so it is the denominator that inflates, and skipping the correction biases ω upward — toward spuriously resembling relaxed .
(d) With equal synonymous opportunity per codon:
gene-wide ω = (295 × 0.1 + 5 × 8)/300 = (29.5 + 40)/300 = 69.5/300 = 0.2317
0.23 — utterly ordinary. For gene-wide ω > 1 you need 29.5 + 5x > 300, i.e. ω > 54.1 at every one of the five sites, which essentially never happens over a whole branch. This is why ω < 1 is not evidence against positive selection; the fix is site models (M1a vs M2a, M7 vs M8), which fit a mixture across codons and test whether a class with ω > 1 is needed.
4. ★ Tajima's D, and what it cannot tell you
A sample of n = 12 sequences. S = 24 segregating sites. Summed pairwise differences over all C(12,2) = 66 pairs = 297, so π = 297/66 = 4.5.
(a) What does θW measure, what does π measure, and why do they differ in sensitivity to rare variants? Compute both. (b) Compute D, using a₁ = Σ¹¹ 1/i = 3.0198773 and a₂ = Σ¹¹ 1/i² = 1.5580322. (c) Two hypothetical datasets, both S = 24: in A all sites are singletons; in B all sit at derived frequency 6/12. Compute π and D for each. (d) (trap) You write "D = −1.93: evidence of a recent ." What is wrong?
Solution
(a) Both estimate θ = 4Neμ under the null, by different weightings of the same site frequency spectrum. θW = S/a₁ counts segregating sites — every site contributes 1, whether its derived sits in 1 copy or 6. π is the mean difference between two random sequences, so a site at derived frequency i/n contributes 2i(n−i)/[n(n−1)]: zero at the edges, maximal at i = n/2. A singleton therefore adds a full unit to θW and almost nothing to π. Under E[ξi] = θ/i both integrate to θ, so the difference has expectation zero and measures the spectrum's shape, not its size.
θW = 24/3.0198773 = 7.9473 ; π = 4.5 ; π − θW = −3.4473
(b) With n = 12:
b1 = (n+1)/(3(n−1)) = 13/33 = 0.3939394
b2 = 2(n²+n+3)/(9n(n−1)) = 318/1188 = 0.2676768
c1 = b1 − 1/a1 = 0.3939394 − 0.3311393 = 0.0628001
c2 = b2 − (n+2)/(a1·n) + a2/a1²
= 0.2676768 − 14/36.2385 + 1.5580322/9.1196590
= 0.2676768 − 0.3863290 + 0.1708430 = 0.0521908
e1 = c1/a1 = 0.0628001/3.0198773 = 0.0207956
e2 = c2/(a1²+a2) = 0.0521908/10.6776912 = 0.0048878
Var = e1·S + e2·S(S−1) = 0.0207956(24) + 0.0048878(552)
= 0.4990944 + 2.6980877 = 3.1971821 sd = 1.788067
D = −3.4473/1.788067 = −1.928
D ≈ −1.93: an excess of rare variants.
(c) Variance depends only on n and S, so sd = 1.788067 in both.
A (i = 1): weight = 2(1)(11)/(12×11) = 22/132 = 0.1666667
πA = 24 × 0.1666667 = 4.0
DA = (4.0 − 7.9473)/1.788067 = −2.208
B (i = 6): weight = 2(6)(6)/132 = 72/132 = 0.5454545
πB = 24 × 0.5454545 = 13.0909
DB = (13.0909 − 7.9473)/1.788067 = +2.877
Identical θW = 7.95; π differs 3.3-fold; D swings from −2.21 to +2.88. The statistic is blind to how much variation there is and sensitive only to how it is distributed.
(d) The trap: a negative D has at least four causes and cannot separate them — a recent selective sweep, whose wiped-out region re-accumulates new and therefore rare ; population expansion, whose genealogies have long terminal branches and so many singletons; , chronically removing carrying deleterious mutations; and ordinary at the locus.
Sweep and expansion are genuinely confounded — both push the spectrum toward rare variants, and no precision on D at one locus separates them. Nor is D standard normal despite the , so "−1.93, therefore p < 0.05" is wrong too; p-values come from simulation.
The escape route is that demography acts genome-wide while selection acts locally: compare against the empirical genome-wide distribution of D; compute Fay and Wu's H, since hitchhiking drags neutral variants to high derived frequency and expansion does not; check length at frequency (iHS, XP-EHH). Background selection survives all three, which is why modern scans fit a background-selection baseline first.
5. Dating a split, and why the arithmetic overstates its own precision
1 Mb of putatively neutral autosomal sequence aligned between human and chimpanzee shows 12,000 differing sites. Use μ = 1.2 × 10⁻⁸ per bp per generation.
(a) Compute p and the JC-corrected d. (b) Compute the divergence time at a generation interval of 25 years, then at 29 years. (c) The fossil-calibrated estimate is 6–7 Mya. (trap) Give three distinct reasons your estimate carries far more uncertainty than the arithmetic suggests, with a number for each.
Solution
(a) p = 12,000/1,000,000 = 0.012
4p/3 = 0.016 ; 1 − 0.016 = 0.984 ; ln(0.984) = −0.0161294
d = −0.75 × (−0.0161294) = 0.0120970
The correction is only 0.81% — at 1.2% divergence almost no site has been hit twice.
(b) Both lineages accumulate substitutions, so d = 2μT with T per lineage:
T = 0.0120970/(2 × 1.2 × 10⁻⁸) = 504,043 generations
× 25 y = 12.6 million years × 29 y = 14.6 million years
(c) The trap is treating a chain of point estimates as a measurement.
1 — The is the whole answer, and it is contested. T scales as 1/μ. sequencing gives ~1.2 × 10⁻⁸; older phylogenetic calibrations gave ~2.5 × 10⁻⁸:
T = 0.0120970/(5 × 10⁻⁸) = 241,941 generations × 25 y = 6.0 million years
Exactly the fossil answer. The twofold disagreement in the literature is the twofold disagreement in μ, and no amount of sequence adjudicates it.
2 — Sequence divergence is older than the species split. d/(2μ) dates the coalescence of the two sequences, which happened in the ancestral population before it split; the expected extra depth is 2Nanc generations. With Nanc ≈ 65,000:
504,043 − 130,000 = 374,043 generations × 25 y = 9.4 million years
A 26% reduction from a parameter you did not measure — it shrinks the gap without closing it.
3 — Generation time is neither constant nor shared. It has varied over the interval, differs between the sexes, and human is ~80% paternal with ~1.3–1.5 extra per year of paternal age — so a change in mating system moves μ per year without moving μ per generation, and it cannot be revised independently of reason 1.
Also: the clock is overdispersed (Var/Mean ~1–35 across proteins), and calibration uncertainty does not shrink with data, since a fossil gives only a minimum age and only rate × time is identified. Report "order 6–13 My", not "12.6 My".
6. McDonald–Kreitman
| Fixed | Polymorphic | |
|---|---|---|
| Nonsynonymous | Dn = 78 | Pn = 24 |
| Synonymous | Ds = 50 | Ps = 60 |
(a) Compute the neutrality index and α. Interpret. (b) Compute the number of adaptive directly and check it against α. (c) Test significance with a χ² on the 2×2. (d) Dropping polymorphisms below 10% frequency leaves Pn = 14, Ps = 54. Recompute α. (trap) Why did it move in that direction? (e) A second gene gives Dn = 20, Ds = 30, Pn = 90, Ps = 60. Compute α and interpret.
Solution
(a)
Pn/Ps = 24/60 = 0.400000 Dn/Ds = 78/50 = 1.560000
NI = 0.400000/1.560000 = 0.256410
α = 1 − (Ds·Pn)/(Dn·Ps) = 1 − (50 × 24)/(78 × 60) = 1 − 1200/4680 = 0.7436
α ≈ 0.74: roughly 74% of nonsynonymous substitutions were fixed by positive selection. NI < 1 is the signature — amino-acid changes enriched among fixed differences relative to their share of polymorphism.
(b)
expected Dn = Ds × (Pn/Ps) = 50 × 0.400000 = 20.0
observed Dn = 78 → excess = 58 adaptive fixations
58/78 = 0.7436 ✓ matches α, which is that excess as a proportion
(c) Row totals 102 and 110, column totals 128 and 84, n = 212.
E(Dn) = 102 × 128/212 = 61.5849 E(Pn) = 102 × 84/212 = 40.4151
E(Ds) = 110 × 128/212 = 66.4151 E(Ps) = 110 × 84/212 = 43.5849
every cell deviates by the same 78 − 61.5849 = 16.4151
χ² = 16.4151² × (1/61.5849 + 1/40.4151 + 1/66.4151 + 1/43.5849)
= 269.4555 × (0.0162378 + 0.0247432 + 0.0150569 + 0.0229437)
= 269.4555 × 0.0789816 = 21.28 df = 1, p ≈ 4 × 10⁻⁶
(d) α = 1 − (50 × 14)/(78 × 54) = 1 − 700/4212 = 0.8338, up from 0.744.
The trap is reading MK as assumption-free. It is robust to demography: synonymous and nonsynonymous sites are interdigitated in the same with the same genealogy, so a bottleneck or expansion distorts both spectra identically and the distortion cancels in the ratio of ratios — exactly the confound that ruins Tajima's D. It is not robust to slightly deleterious nonsynonymous , which segregate at low frequency and almost never fix, inflating Pn without touching Dn and pushing α down. A frequency cutoff removes them preferentially, hence the increase — and since that bias scales with Ne, uncorrected α is not comparable across species.
(e)
NI = (90/60)/(20/30) = 1.500000/0.666667 = 2.25
α = 1 − 2.25 = −1.25
Negative α is not "negative adaptation". Nonsynonymous variants are over-represented among polymorphisms relative to fixed differences — a gene loaded with slightly deleterious variation that will never fix. Report "no evidence of adaptive fixation; strong evidence of deleterious polymorphism", and do not average it with positive α from other genes.
7. Discordance: duplication or incomplete lineage sorting?
The species tree is ((A, B), C).
(a) Across 5,000 loci: 3,000 give ((A,B),C), 1,000 give ((A,C),B), 1,000 give ((B,C),A). Which process, and what is the internal in units? Use P(discordant) = (2/3)e−τ. (b) A single gene places A with C at a node dated ~450 Mya, while the A–C speciation is dated ~320 Mya. Generation time ~2 years, vertebrate Ne ~ 10⁴. (trap) Could ILS produce this? (c) If each lineage independently ends up single-copy with probability q = 0.5, how often does confidently pair two ? (d) What functional- error follows from getting (b) wrong, and where does it end up?
Solution
(a) Incomplete lineage sorting, diagnosed by symmetry: 1,000 versus 1,000. ILS arises when A and B fail to coalesce in the ancestral AB population; three lineages then enter the deeper population, all three topologies become equally likely, and the two discordant ones must therefore occur equally often in expectation.
P(discordant) = 2,000/5,000 = 0.40 = (2/3)e^(−τ)
e^(−τ) = 0.60 ⟹ τ = −ln(0.60) = 0.5108 coalescent units
The internal branch is 0.51 × 2Ne generations — short relative to ancestral population size, exactly the regime where a large minority of loci disagree with the species tree, and that is the correct answer, not an error. An asymmetric split (1,500 versus 500) would instead mean gene flow.
(b) No — ruled out by three orders of magnitude. ILS can push a coalescence deeper than the species split, but the excess depth is bounded by the ancestral population's coalescent timescale, ~2Ne generations:
excess depth = 450 − 320 = 130 My = 130 × 10⁶/2 = 6.5 × 10⁷ generations
2Ne = 6.5 × 10⁷ ⟹ Ne ≈ 3.25 × 10⁷
An ancestral vertebrate Ne of 32 million against a realistic ~10⁴ — 3,000× too large. Independently: grant even a generous 50-My internal branch between the A–B and A–C splits — 5 × 10⁷ years = 2.5 × 10⁷ generations — and the stated Ne = 10⁴ gives τ = 2.5 × 10⁷/(2 × 10⁴) = 1,250, so P(discordant) = (2/3)e−1250 ≈ 9 × 10⁻⁵⁴⁴.
The trap. ILS is right at shallow, rapidly radiating splits and arithmetically impossible at deep ones. A discordant node older than the speciation it contradicts is the signature of gene duplication followed by differential loss — the gene tree is correct as a gene tree, and its duplication node is being read as a speciation node. Confirm with the second copy retained in an outgroup, and with synteny.
(c) If the ancestor carried paralogs X1 and X2 and each lineage independently ends single-copy with probability q, retaining either copy with equal chance, both are single-copy with probability q², and half of those retained different paralogs:
P(pseudoorthology) = q²/2 = 0.25/2 = 0.125 = 12.5%
One family in eight. RBH pairs A-X1 with C-X2, scores it highly, and cannot detect the error — each really is the other's best surviving hit.
(d) Function is transferred along the orthology assignment, so the error is an annotation copied from the wrong paralog — precisely the copy whose function was licensed to diverge, since duplication is where functional change happens and speciation is not. High identity does not rescue you: expression domain can diverge completely while coding sequence stays >90% identical. And the annotation does not stay put — it is deposited in a database, inherited by every downstream pipeline, and silently weights every enrichment analysis run against that genome afterwards.
8. ★ Why duplicate genes survive at all
A newly duplicated pair. Null arise per copy at rate un; mutations conferring a beneficial new function at ub, with coefficient s. Write φ = ub/un.
(a) Derive the rate at which a null mutation fixes in one copy. Note what is absent. (b) Derive P(neo) and evaluate for a vertebrate (Ne = 10⁴) and an insect (Ne = 10⁷), with s = 0.01 and φ = 10⁻⁴. (c) Derive P(sub) under duplication–degeneration–complementation with n = 2 regulatory elements; evaluate at r = ur/uc = 1 and r = 2. (d) (trap) Vertebrate genomes retain roughly 20% of pairs long-term. Show alone cannot account for that, and give the Ne it would require. (e) Combine the mechanisms into a total predicted retention. Why is a naive sum approximately right here, and what is still missing?
Solution
(a) Immediately after duplication both copies are functional and either is dispensable, so a null mutation in either is selectively neutral. By problem 1, nulls arise at 2N·un and each fixes with probability 1/(2N), giving λnull = un. What is absent is N — the duplicate decays on a clock population size cannot slow, which is why nonfunctionalisation is the default and preservation needs an active explanation.
(b) Beneficial mutations arise at 2N·ub and fix with probability ≈ 2s, so λneo = 4N ubs. For competing Poisson processes the probability the beneficial one arrives first is the ratio of rates:
P(neo) = 4Nu_b s/(4Nu_b s + u_n) = 4Nsφ/(4Nsφ + 1)
Vertebrate: 4Nsφ = 4 × 10⁴ × 0.01 × 10⁻⁴ = 0.04 → P = 0.04/1.04 = 0.0385 = 3.85%
Insect: 4Nsφ = 4 × 10⁷ × 0.01 × 10⁻⁴ = 40 → P = 40/41 = 0.9756 = 97.6%
Ohno's mechanism is efficient in large populations and nearly useless in small ones, because 2s does not shrink with N while 1/(2N) does.
(c) Give the gene 2 independently mutable regulatory elements plus a coding region whose loss kills the copy; ur is the null rate per element, uc the coding null rate.
Step 1. Neutral targets: 2 copies × (1 coding + 2 elements), total rate 2(uc + 2ur). A coding hit ends in nonfunctionalisation, a regulatory hit does not, so P(regulatory first) = 4ur/(2uc + 4ur).
Step 2. Say element 1 of copy A is gone. Copy B's element 1 is now the only source of subfunction 1, and its coding region indispensable, so mutations in either are deleterious. That leaves three neutral targets — copy A's coding region (uc), copy A's element 2 (ur), copy B's element 2 (ur) — and only the last completes the pair.
P(sub) = 2u_r/(u_c + 2u_r) × u_r/(u_c + 2u_r) = 2r²/(1 + 2r)²
r = 1: 2(1)/(3)² = 2/9 = 0.2222 = 22.2%
r = 2: 2(4)/(5)² = 8/25 = 0.3200 = 32.0%
Look at what is missing: N. preserves duplicates by degrading them, through individually neutral mutations, so it works as well at 10⁴ as at 10⁷.
(d) The trap is assuming a preserved duplicate must have acquired a new function. Neofunctionalisation predicts 3.85% against an observed 20%, and the gap cannot be tuned away, because closing it needs 4Nesφ = 0.25:
N = 0.25/(4 × 0.01 × 10⁻⁴) = 0.25/(4 × 10⁻⁶) = 62,500
An Ne of 6.25 × 10⁴, over six times the vertebrate estimate. Vertebrates are duplicate-rich and small-Ne — the combination Ohno's mechanism cannot deliver.
(e) Let the beneficial route take its share first and route the remainder into the DDC branching:
P(neo) = 0.0385
remainder = 1/1.04 = 0.9615
P(sub) joint = 0.9615 × 0.2222 = 0.2137
total retention ≈ 0.0385 + 0.2137 = 0.2521 → 25.2%
The naive sum, 0.0385 + 0.2222 = 26.1%, is within 1 percentage point, because φ = 10⁻⁴ makes the beneficial route only 4% as fast as the null route, so it barely depletes the DDC path. The same approximation would fail for the insect, where that route carries 97.6% of the flow.
25.2% predicted against ~20% observed — the right order, from a model whose term is not adaptive at all. Still missing is dosage balance: proteins in stoichiometric complexes cannot lose a copy without misassembly, so loss is deleterious rather than neutral. It dominates after whole-genome duplication, which is why WGD pairs show a ~23 My half-life against ~8 My for small-scale duplicates in the same genome.
Where you went wrong
| Error pattern | What to re-read |
|---|---|
| Used a per-replication rate where a per-generation rate belongs | Problem 1(e); verified facts |
| Read k = μ as "big populations evolve more slowly" | Problem 1(c) — the two factors of N cancel exactly |
| Treated the correction as a small constant tweak | Problem 2(a) — 3.5% at p = 0.05, 53% at p = 0.45 |
| Compared raw counts of synonymous and nonsynonymous changes | Problem 3(a) — the code offers ~3× more nonsynonymous sites |
| Read ω < 1 as evidence against positive selection | Problem 3(d) — five selected in 300 give ω = 0.23 |
| Confused what π and θW weight, or called a negative D a sweep | Problem 4(a),(c) — same S, D from −2.21 to +2.88; 4(d) — sweep and expansion are confounded |
| Quoted a divergence date to three significant figures, or forgot divergence predates the split | Problem 5(c) — it moves twofold with μ alone; 5(c2) — subtract 2Nanc generations |
| Read a negative α as evidence against adaptation | Problem 6(e) — deleterious polymorphism inflating Pn |
| Reached for ILS to explain a deep discordant node | Problem 7(b) — ILS is bounded by ~2Ne generations |
| Trusted to return | Problem 7(c) — q²/2 ≈ 12.5% pseudoorthology |
| Assumed a retained acquired a new function | Problem 8(d) — is neutral and N-independent |