Problem set 11 — Statistical genomics
Covers Ch 51–56.
Attempt before revealing. Every problem here is ten lines of code, and the point is that the number surprises you. Reading the solution first replaces the surprise with a nod.
Roughly in order of difficulty. ★ marks the two worth the most time.
1. Where 5 × 10⁻⁸ comes from
(a) Derive it by Bonferroni from FWER 0.05 over ~1 × 10⁶ effectively independent common-variant tests, and give the two-sided |Z| cut-off. (b) Your array has 700,000 variants and you impute to 20 million. Should α be 0.05/(7 × 10⁵) or 0.05/(2 × 10⁷)? Neither — why? (c) African-ancestry are half as long, so M_eff roughly doubles. Compute the threshold, and the actual FWER you incur by using 5 × 10⁻⁸ anyway. (d) WGS pushes M_eff to ~10⁷. Compute the threshold, the FWER penalty, and the change in the Z cut-off. Comment on the last one.
Solution
(a) α = FWER / M = 0.05 / 10⁶ = 5 × 10⁻⁸
Two-sided, so Pr(Z > c) = α/2 = 2.5 × 10⁻⁸:
c = Φ⁻¹(1 − 2.5 × 10⁻⁸) = 5.4513, equivalently χ²₁ = 5.4513² = 29.717
(b) M is the number of effectively independent tests in the population's LD structure, not rows in your file.
- 0.05/(7 × 10⁵) = 7.1 × 10⁻⁸ is too lax: the array proxies for everything it tags.
- 0.05/(2 × 10⁷) = 2.5 × 10⁻⁹ is too strict: 40 markers at r² > 0.95 are one test, not 40.
Trap: you cannot buy power by testing fewer variants — the truth about a variant does not depend on what else you looked at, which is why candidate-gene studies did not replicate (Ch 51 §5).
(c) α = 0.05 / (2 × 10⁶) = 2.5 × 10⁻⁸, c = 5.5733.
Using 5 × 10⁻⁸ over 2 × 10⁶ independent tests:
FWER = 1 − (1 − 5 × 10⁻⁸)^(2×10⁶) = 1 − e^(−0.1) = 0.0952
The nominal 5% is really 9.5%. (Published proposals land nearer 1–2 × 10⁻⁸.) A fixed threshold is anti-conservative for African-ancestry samples and conservative elsewhere — the group with most to gain from discovery has its error controlled least well.
(d) α = 0.05 / 10⁷ = 5 × 10⁻⁹, c = 5.8472.
FWER at 5 × 10⁻⁸ = 1 − e^(−10⁷ × 5×10⁻⁸) = 1 − e^(−0.5) = 0.393. Rare variants sit in weak LD with everything, so each adds nearly a whole test — hence WGS proposals of 5 × 10⁻⁹ to 1 × 10⁻⁸.
The comment: a tenfold tightening of α moved the cut-off only from 5.4513 to 5.8472 — 0.396, or 7%. The tail is thin enough that order-of-magnitude arguments about M barely move the number you test against, which is how one round figure survived two decades. But the FWER moved by 2× and 8× while Z moved 7%. It is a convention, not a constant.
2. LD from a haplotype table
Two SNPs, rs1 (A/a) and rs2 (B/b). 400 phased :
rs2 = B rs2 = b total
rs1 = A 144 96 240
rs1 = a 16 144 160
total 160 240 400
(a) Haplotype and frequencies. (b) Compute D, D', r². (c) With X = 1[carries A] and Y = 1[carries B], show by direct computation that r² is the squared Pearson correlation of X and Y. (d) rs2 is causal but you only rs1. A study of 25,000 is adequately powered on rs2 directly; how many at the tag? (e) Could a denser array find a perfect proxy for rs2 at rs1's ?
Solution
(a) Divide by 400: p_AB = 0.36, p_Ab = 0.24, p_aB = 0.04, p_ab = 0.36.
p_A = 0.36 + 0.24 = 0.60, p_a = 0.40; p_B = 0.36 + 0.04 = 0.40, p_b = 0.60
(b) D = p_AB − p_A p_B = 0.36 − (0.60)(0.40) = 0.36 − 0.24 = 0.12
Cross-product check: (0.36)(0.36) − (0.24)(0.04) = 0.1296 − 0.0096 = 0.12 ✓
D > 0, so D_max = min(p_A p_b, p_a p_B) = min(0.36, 0.16) = 0.16:
D' = 0.12 / 0.16 = 0.75 r² = D² / (p_A p_a p_B p_b) = 0.0144 / (0.60 × 0.40 × 0.40 × 0.60) = 0.0144 / 0.0576 = 0.25, r = 0.50
(c) Both indicators are Bernoulli:
E[X] = 0.60, Var(X) = (0.60)(0.40) = 0.24 E[Y] = 0.40, Var(Y) = (0.40)(0.60) = 0.24 E[XY] = Pr(X=1, Y=1) = p_AB = 0.36
Cov(X, Y) = 0.36 − (0.60)(0.40) = 0.12 = D
r = 0.12 / √(0.24 × 0.24) = 0.12/0.24 = 0.50, r² = 0.25 ✓
D is not analogous to a covariance; it is the covariance of the two columns of the haplotype matrix — so r² is the fraction of variance in one column explained by the other, which is what (d) uses.
(d) R²_tag = r²·R²_causal and λ ≈ N·R² (Ch 29 §5), so equal power needs
N_tag = 25,000 / 0.25 = 100,000
Four times the cohort. The r² ≥ 0.8 array-design convention accepts 25% inflation and no more.
(e) Trap. No, and density cannot help. r² = 1 requires D² = p_A p_a p_B p_b, forcing p_A = p_B. Here they are 0.60 and 0.40, so even at maximum LD:
r²_max = D_max² / 0.0576 = 0.0256 / 0.0576 = 0.444
A common variant can never be a perfect proxy for one of different frequency, however tightly bound — which is why D' = 1 is a terrible tagging criterion.
3. Population stratification, with no causal effect anywhere ★
A 50:50 mixture of two non-interbreeding subpopulations.
| Subpop 1 | Subpop 2 | |
|---|---|---|
| Individuals | 10,000 | 10,000 |
| Risk- | 0.10 | 0.50 |
| Disease prevalence | 5% | 25% |
Inside each subpopulation the variant has exactly zero effect.
(a) Compute Cov(g, y) pooled, g the 0/1/2 dosage and y the 0/1 status; check against 2w(1−w)(p₁ − p₂)(μ₁ − μ₂). (b) Build the pooled case/control table and compute the odds ratio. (c) Compute χ² and comment on the p-value. (d) Stratify: the odds ratio within each subpopulation. (e) A collaborator proposes recruiting ten times as many people. What happens?
Solution
(a) Within a subpopulation g and y are independent, so E[gy] = E[g]E[y] there.
Subpop 1: E[g] = 2(0.10) = 0.20, E[y] = 0.05 → E[gy] = 0.010 Subpop 2: E[g] = 2(0.50) = 1.00, E[y] = 0.25 → E[gy] = 0.250
Pooled with w = 0.5: E[g] = 0.60, E[y] = 0.15, E[gy] = 0.5(0.010) + 0.5(0.250) = 0.130
Cov(g, y) = 0.130 − (0.60)(0.15) = 0.04
Formula: 2(0.5)(0.5)(0.10 − 0.50)(0.05 − 0.25) = 0.5 × (−0.40)(−0.20) = 0.04 ✓
Non-zero because both differences are. Ancestry is a common cause of and — a confounder in the strict sense.
(b) Cases: 500 and 2,500 (3,000); controls 9,500 and 7,500 (17,000). With no effect each group carries the allele at its own subpopulation's frequency.
Case risk alleles = 2(500)(0.10) + 2(2,500)(0.50) = 100 + 2,500 = 2,600 of 6,000 Control risk alleles = 2(9,500)(0.10) + 2(7,500)(0.50) = 1,900 + 7,500 = 9,400 of 34,000
risk other total
cases 2,600 3,400 6,000 freq 0.4333
controls 9,400 24,600 34,000 freq 0.2765
OR = (2,600 × 24,600)/(3,400 × 9,400) = 63,960,000 / 31,960,000 = 2.001
(c) χ² = N(ad − bc)² / [(a+b)(c+d)(a+c)(b+d)]
ad − bc = 63,960,000 − 31,960,000 = 3.2 × 10⁷, squared = 1.024 × 10¹⁵ Denominator = (6,000 × 34,000)(12,000 × 28,000) = (2.04 × 10⁸)(3.36 × 10⁸) = 6.8544 × 10¹⁶
χ² = 40,000 × 1.024 × 10¹⁵ / 6.8544 × 10¹⁶ = 597.6, Z = 24.4, p ≈ 6 × 10⁻¹³²
A hundred and twenty-four orders of magnitude past 5 × 10⁻⁸, for a variant with no effect on anything. The p-value is not lying — genotype and phenotype are associated here — it just has no opinion about why.
(d) Subpop 1: 100/1,000 vs 1,900/19,000 → OR = (100 × 17,100)/(900 × 1,900) = 1.000 Subpop 2: 2,500/5,000 vs 7,500/15,000 → OR = (2,500 × 7,500)/(2,500 × 7,500) = 1.000
χ² = 0 in each. The signal vanishes once the confounder is in the model, because the signal was the confounder.
(e) The trap. It gets worse. Stratification is bias, not variance: the estimate does not shrink with n, only the standard error does, so the non-centrality grows linearly in N.
χ² at 200,000 people = 597.6 × 10 = 5,976
Every other error in genomics improves with more data. A bigger study makes a stratification artefact more significant, not less. Campbell and colleagues (2005) reported p < 10⁻⁶ between a SNP at LCT and height, in a panel that passed the stratification tests of the day: LCT frequency varies steeply across Europe, and so does height.
4. Power, and where it goes when the allele gets rare
5,000 cases and 5,000 controls. The non-centrality of the 1-df test is
λ = N · φ(1−φ) · Var(g) · ln(OR)²
with N the total sample, φ the case fraction, Var(g) = 2p(1−p).
(a) Power to detect OR = 1.20 at MAF 0.30, α = 5 × 10⁻⁸. (b) Repeat at MAF 0.05 and 0.005, holding OR and N fixed. (c) At MAF 0.0005, how many minor- copies are in the study, and what is the smallest p-value any test could return? (d) Thirty loss-of-function variants in one gene, each at MAF 0.0005 with OR = 2.0 in the same direction. Compare single-variant power with a at α = 0.05/20,000.
Solution
(a) ln(1.20) = 0.182322, so ln(OR)² = 0.0332412. Balanced: φ(1−φ) = 0.25. Var(g) = 2(0.30)(0.70) = 0.42.
λ = 10,000 × 0.25 × 0.42 × 0.0332412 = 1,050 × 0.0332412 = 34.90, E[Z] = √34.90 = 5.908
Power = Pr(Z > 5.4513 | mean 5.908) = Pr(N(0,1) > −0.4566) = 0.676
(b) Only Var(g) changes.
MAF 0.05: Var(g) = 0.095 → λ = 2,500 × 0.095 × 0.0332412 = 7.895, E[Z] = 2.810, power = Pr(N(0,1) > 2.641) = 0.0041
MAF 0.005: Var(g) = 0.00995 → λ = 2,500 × 0.00995 × 0.0332412 = 0.827, E[Z] = 0.909, power = Pr(N(0,1) > 4.542) = 2.8 × 10⁻⁶
So 68% → 0.41% → 0.0003%. Power scales essentially as p, since 2p(1−p) → 2p — but not gently. It is a tail probability, √λ against a fixed 5.45, so a sixfold drop in MAF takes 68% to under half a percent.
(c) Minor-allele copies = 2Np = 2 × 10,000 × 0.0005 = 10. Under the null each carrier is a case with probability ½, so the most extreme outcome conceivable is all 10 being cases:
p_min = 2⁻¹⁰ = 9.8 × 10⁻⁴
Four orders short of 5 × 10⁻⁸ whatever the effect size and whatever the test — an information bound (Ch 54 §2), not a modelling weakness. Generally 2⁻ᵏ ≤ 5 × 10⁻⁸ requires k ≥ 25 carriers.
Trap: at MAF 0.005 there are 100 carriers, past that floor, and power is still 0.0003%. Clearing the floor is necessary, not sufficient.
(d) Single variant at MAF 0.0005, OR = 2.0: ln(2)² = 0.480453.
λ = 2,500 × 0.0009995 × 0.480453 = 1.201, E[Z] = 1.096, power = 6.6 × 10⁻⁶
Burden test: aggregate = 30 × 0.0005 = 0.015, so Var = 2(0.015)(0.985) = 0.02955 — 29.6× any single component. Effects are same-signed, so the numerator (Σβⱼvⱼ)² accumulates rather than cancels.
λ = 2,500 × 0.02955 × 0.480453 = 35.49, E[Z] = 5.958
The threshold also relaxes: ~20,000 genes, α = 2.5 × 10⁻⁶, two-sided c = 4.708.
Power = Pr(Z > 4.708 − 5.958) = Pr(N(0,1) > −1.250) = 0.894
From 0.0007% to 89% on the same 10,000 people: aggregation buys carriers (10 → 300 alleles), and gene-level testing buys a threshold two orders cheaper. If half the masked variants had OR < 1 the numerator would cancel to zero, which is why a bidirectional gene needs (Ch 54 §7).
5. Winner's curse and the replication you did not budget for
, SD 1. Discovery: N = 100,000, MAF 0.20, true effect β = 0.02 SD.
(a) Compute R², λ, E[Z], and power to reach 5 × 10⁻⁸. (b) Given the variant is reported, compute the expected published effect, using E[Z | Z > c] = μ + φ(c − μ)/(1 − Φ(c − μ)). (c) How large must an independent replication be for 80% power at one-sided nominal 0.05, with direction prespecified? (d) A colleague powers the replication on the published effect. What N do they pick, what power do they get, and how does the shortfall relate to (b)?
Solution
(a) Var(g) = 2(0.20)(0.80) = 0.32
R² = β²Var(g) = 0.0004 × 0.32 = 1.28 × 10⁻⁴ (0.0128% of trait variance) λ = 100,000 × 1.28 × 10⁻⁴ = 12.80, E[Z] = 3.5777
Power = Pr(Z > 5.4513 | mean 3.5777) = Pr(N(0,1) > 1.8736) = 0.0305
A typical common-variant effect, in 100,000 people, is found one time in thirty-three.
(b) That 3% is the mechanism: the estimate clears 5.4513 only when noise pushes it up, so conditioning on discovery conditions on a favourable draw. With c − μ = 5.4513 − 3.5777 = 1.8736:
φ(1.8736) = e^(−1.7552)/√(2π) = 0.17287/2.5066 = 0.068967 1 − Φ(1.8736) = 0.030493 (the power from (a) — it must be) Inverse Mills ratio = 0.068967/0.030493 = 2.2617
E[Z | detected] = 3.5777 + 2.2617 = 5.8395, inflation = 5.8395/3.5777 = 1.632
Published β̂ ≈ 0.02 × 1.632 = 0.0326; variance explained inflated by 1.632² = 2.66×, since R² goes as β². This is truncation, not p-hacking: it happens in a perfectly honest study with a correct model, and bites hardest on the most marginal findings.
(c) The curse does not travel; in a fresh sample E[Z] returns to the truth and scales as √N.
Required E[Z] = 1.6449 + 0.8416 = 2.4865 N_rep = 100,000 × (2.4865/3.5777)² = 100,000 × 0.48302 = ≈ 48,300
Half the discovery cohort — which is why replication uses a nominal threshold and prespecified direction rather than 5 × 10⁻⁸ twice, which would reject nearly every true finding.
(d) The trap. Powering on β̂ = 0.0326:
R̂² = (0.0326)² × 0.32 = 3.4008 × 10⁻⁴ N = 2.4865² / (3.4008 × 10⁻⁴) = 6.1827 / (3.4008 × 10⁻⁴) = ≈ 18,180
The actual non-centrality at that N uses the true R²:
λ = 18,180 × 1.28 × 10⁻⁴ = 2.327, E[Z] = 1.5255 Actual power = Pr(Z > 1.6449 − 1.5255) = Pr(N(0,1) > 0.1194) = 0.452
Forty-five percent, not eighty — and the study then "fails to replicate" something real. Exactly:
48,300 / 18,180 = 2.66 = (inflation)² = the inflation in variance explained
Effects inflate by 1.63, so replication sample sizes must inflate by 1.63² = 2.66, since power depends on β². Never power a replication on the discovery estimate, and shrink discovery betas before building a (Ch 53 §2).
6. From a polygenic score to an actual number ★
Disease with lifetime prevalence K = 2%. A score standardised to N(0,1) explains ρ² = 0.16 of variance in liability: affected if L = ρS + √(1−ρ²)·U exceeds T = Φ⁻¹(1−K), with U ~ N(0,1) independent of S.
(a) Derive Pr(affected | S = s) and verify it integrates back to K. (b) Compute absolute risk at the 50th, 90th and 99th percentiles. Taking the top-decile average as 7.35%, express the top decile as a relative risk and an absolute risk, and state what fraction of all cases it contains. (c) The score is applied in an ancestry absent from training. Tag–causal LD is preserved (r = 0.9) at 50% of loci, halved (r = 0.45) at 30%, destroyed (r = 0) at 20%. Compute the factor by which R² falls, then recompute the top-decile risk. (Its average at the reduced ρ² is 5.01%.) (d) Name the remaining mechanisms — three more in the tabulation, plus one easily forgotten — and quantify the -frequency one.
Solution
(a) Affected means L > T:
Pr(ρs + √(1−ρ²)U > T) = Pr(U > (T − ρs)/√(1−ρ²)) = Φ((ρs − T)/√(1−ρ²))
ρ = 0.400, √(1−ρ²) = √0.84 = 0.916515, T = Φ⁻¹(0.98) = 2.053749.
a ≡ ρ/√(1−ρ²) = 0.400/0.916515 = 0.436436 b ≡ T/√(1−ρ²) = 2.053749/0.916515 = 2.240824
Pr(affected | S = s) = Φ(0.436436 s − 2.240824)
Check: E[Φ(aS − b)] = Φ(−b/√(1+a²)) = Φ(−2.240824/1.091089) = Φ(−2.053749) = 0.0200 ✓ — the curve integrates back to the prevalence, which catches every sign error at once.
(b) s = 0, +1.2816, +2.3263 at the 50th, 90th and 99th percentiles:
50th: 0 − 2.240824 = −2.2408 → Φ = 1.25% 90th: 0.436436(1.2816) − 2.240824 = 0.5593 − 2.2408 = −1.6815 → Φ = 4.63% 99th: 0.436436(2.3263) − 2.240824 = 1.0153 − 2.2408 = −1.2255 → Φ = 11.02%
Top decile at 7.35%:
- Relative risk vs population average = 7.35/2.00 = 3.67×
- Absolute increase = 7.35% − 2.00% = 5.35 percentage points
- 92.6% of the top decile never develop the disease
- Cases falling in the top decile = 0.10 × 0.0735 / 0.02 = 36.7%
"Nearly four times the average risk" and "screening this group still misses 63% of cases" describe the same table.
Trap: do not evaluate the decile at its midpoint (95th percentile, 6.39%); Φ is convex there, so the true average of 7.35% exceeds it. The same flatness makes rarer diseases give larger ratios attached to smaller absolute risks.
(c) In standardised units the fits ŵⱼ = r_disc,j βⱼ on each tag. Applying those weights where the tag–causal correlation is r_tgt,j, with tags approximately independent:
Cov(S, y)_target = Σ r_disc,j r_tgt,j βⱼ² , Var(S)_target = Σ r_disc,j² βⱼ² R²_target / R²_discovery = [ mean(r_disc·r_tgt) / mean(r_disc²) ]²
mean(r_disc·r_tgt) = 0.9 × [0.50(0.90) + 0.30(0.45) + 0.20(0)] = 0.9 × 0.585 = 0.5265 mean(r_disc²) = 0.81, ratio = 0.5265/0.81 = 0.65, so R² falls by 0.65² = 0.4225
New ρ² = 0.16 × 0.4225 = 0.0676, ρ = 0.260, √(1−ρ²) = √0.9324 = 0.965609 a′ = 0.269260, b′ = 2.126896 90th percentile: Φ(0.269260 × 1.281552 − 2.126896) = Φ(−1.781825) = 3.74% (was 4.63%) Top decile 5.01% (was 7.35%) → RR 2.51× (was 3.67×), cases captured 25.1% (was 36.7%)
Same disease, same score, different people. And where LD flips, r enters negatively and the variant actively subtracts.
(d) Allele-frequency differences. A locus contributes wⱼ²·2pⱼ(1−pⱼ) to score variance. A tag at MAF 0.20 contributes 2(0.20)(0.80) = 0.32; the same tag at MAF 0.02 contributes 2(0.02)(0.98) = 0.0392 — a factor of 0.1225, an 88% collapse. Loci carrying 15% of discovery score variance now carry 0.15 × 0.1225 = 1.8%. The score's mean and variance shift, breaking calibration even where ranking survives; recalibration makes the number honest, not more informative.
Effect-size heterogeneity. α = a + d(q − p) is frequency-dependent by construction (Ch 30), and G×E and push the same way — the estimand itself differs, independent of measurement.
Environment and trait definition. Different exposure distributions, diagnostic thresholds, ascertainment and access to care mean the thing being predicted is not the same trait, so even perfect weights mispredict.
And a fifth, easily forgotten. Residual in the discovery GWAS puts a non-causal component into the weights (Ch 51). It is a property of that cohort's structure, does not transfer, and can transfer with the wrong sign.
The framing, routinely got wrong. None of these says biology differs between groups; each is a property of whose genomes were in the training data. Run the same pipeline on an African-ancestry cohort and the score works there and degrades in Europeans (Ch 53 §7).
7. Classifying a variant under ACMG 2015
This case is constructed. The variant and every observation are invented to exercise the framework; nothing here is a claim about a real residue.
A proband with long QT syndrome. Panel sequencing returns one candidate in KCNQ1, a gene with definitive validity for LQT1. Evidence:
- Absent from v4 (0 of ~1.6 × 10⁶ ).
- REVEL = 0.85 — the ClinGen-calibrated Moderate bin is [0.773, 0.932).
- Segregates with disease through 3 informative meioses.
- Three other in-silico predictors also call it damaging.
- A 2011 case report calls it "likely disease-causing".
Points: Supporting 1, Moderate 2, Strong 4, Very Strong 8. Posterior = π·X^(N/8)/(π·X^(N/8) + 1 − π), with π = 0.10, X = 350.
(a) Compute AF_max for LQTS via KCNQ1 with D = 1/2,000, g = 0.35, a = 0.05, ψ = 0.6. Is absence from gnomAD informative? (b) Which codes apply, at what strength? Total the points. (c) Compute the posterior and the class; confirm against the 2015 combining rules. (d) Which two listed observations contribute nothing, and why? (e) A calibrated deep mutational scan is abnormal at PS3_Moderate. Reclassify.
Solution
(a) For a condition, AF_max = D·g·a/(2ψ):
AF_max = (5 × 10⁻⁴ × 0.35 × 0.05)/(2 × 0.6) = 8.75 × 10⁻⁶ / 1.2 = 7.29 × 10⁻⁶
In 1.6 × 10⁶ alleles that is ≈ 11.7 alleles, so the database resolves the maximum credible frequency and zero observations means something: the rule-of-three bound on 0/(1.6 × 10⁶) is 3/(1.6 × 10⁶) = 1.9 × 10⁻⁶, below AF_max.
Informative but weak, so ClinGen downgrades PM2 to Supporting — almost every variant is absent from almost every database, and rarity is what benign and pathogenic rare variants share.
(b)
PM2_Supporting absent from gnomAD, below AF_max Supporting +1
PP3_Moderate REVEL 0.85, calibrated Moderate bin Moderate +2
PP1 segregation, 3 informative meioses Supporting +1
---
4 points
(c) OddsPath = 350^(4/8) = 350^0.5 = 18.708
Posterior = (0.10 × 18.708)/(0.10 × 18.708 + 0.90) = 1.8708/2.7708 = 0.675
0.10 < 0.675 < 0.90 → .
By the 2015 rules: 1 Moderate + 2 Supporting. Likely pathogenic needs ≥3 Moderate, or 2 Moderate + ≥2 Supporting, or 1 Moderate + ≥4 Supporting. None is met → VUS ✓ — the point ladder was reverse-engineered from these rules (Ch 55 §3), so this is not luck.
A VUS licenses no clinical action and no testing of relatives. "Being cautious" about a VUS is acting on a VUS.
(d) The two traps.
The three extra predictors contribute nothing. They share training features and alignments and are in several cases built from one another, so counting their agreement violates the conditional independence that makes multiplying likelihood ratios legal. Use one calibrated predictor, once, at its calibrated strength — which PP3_Moderate already is.
The 2011 case report contributes nothing. That is PP5, withdrawn by ClinGen: it launders someone else's unexamined conclusion into your evidence, and if that conclusion rested on the same prediction you already counted, you count it twice. BP6 is withdrawn symmetrically.
(e) Add PS3_Moderate, +2 → 6 points.
OddsPath = 350^(6/8) = 80.92, posterior = 8.092/8.992 = 0.900 → Likely pathogenic (the 0.90 boundary, essentially exactly). By the 2015 rules: 2 Moderate + 2 Supporting → LP ✓
One experiment, two different clinical reports, identical variant. Note the margin: drop either Supporting criterion and it is 5 points, posterior 0.812, VUS again. A call on a tier boundary should be reported as one.
Currency note. This is Richards et al. 2015 as refined by ClinGen SVI and expert panels, still the operative standard. ACMG v4 is in draft, previewed at the 2025 Clinical Genomics Meeting: it consolidates evidence codes so related observations cannot be double-counted structurally — exactly the problem in (d) — and subdivides the VUS band. The five tiers survive, but a report issued today against a draft standard is not interpretable by whoever receives it.
8. Clonal architecture from VAFs
Tumour with matched normal. Purity ρ = 0.70, all loci autosomal so CN_n = 2.
| Local tumour CN | log₂ ratio | ||
|---|---|---|---|
| TP53 p.(Arg175His) | 0.35 | 2 | 0.00 |
| KRAS p.(Gly12Asp) | 0.34 | 2 | 0.00 |
| CDKN2A p.(Trp110Ter) | 0.54 | 1 | −0.62 |
| SMAD4 p.(Arg361His) | 0.28 | 2 | 0.00 |
| ARID1A p.(Gln456fs) | 0.14 | 2 | 0.00 |
Use φ = VAF × [ρ·CN_t + (1−ρ)·CN_n] / (ρ·m), with m mutant copies per carrying cell.
(a) Compute φ for TP53, KRAS, SMAD4, ARID1A assuming m = 1. Clonal or subclonal? (b) Compute φ for CDKN2A naively (CN_t = 2), then correctly. What does the naive answer tell you, and how does the log₂ ratio decide between the two explanations? (c) Reconstruct the clonal ordering, using the pigeonhole rule. (d) The purity caller had returned ρ = 0.50. Show the data refute it.
Solution
(a) At a locus the bracket is ρ(2) + (1−ρ)(2) = 2 regardless of ρ, so with m = 1, φ = VAF × 2/0.70 = VAF × 2.857143.
- TP53: 0.35 × 2.857143 = 1.000 → clonal
- KRAS: 0.34 × 2.857143 = 0.971 → clonal (within noise of 1)
- SMAD4: 0.28 × 2.857143 = 0.800 → subclonal, 80% of tumour cells
- ARID1A: 0.14 × 2.857143 = 0.400 → subclonal, 40% of tumour cells
TP53 and SMAD4 differ by seven points of VAF, yet one is in every tumour cell and the other in four fifths. VAF is not interpretable; CCF is.
(b) Naive: φ = 0.54 × 2.857143 = 1.543. A CCF above 1 is meaningless — a mutation cannot be in more tumour cells than exist. Exactly two candidate fixes:
- m = 2 (copy-neutral LOH): φ = 0.54 × 2/(0.70 × 2) = 0.771
- CN_t = 1 (wild-type copy deleted): bracket = 0.70(1) + 0.30(2) = 1.30, so φ = 0.54 × 1.30/0.70 = 0.54 × 1.857143 = 1.003
The log₂ ratio decides. A deletion at ρ = 0.70 gives observed copy number 1.30, so log₂(1.30/2) = log₂(0.65) = −0.62 — exactly what is reported; two retained copies would give 0.00. So φ(CDKN2A) = 1.00 → clonal, second lost by deletion: Knudson's two hits (Ch 56 §5).
The lesson: a CCF above 1 is not an error to clamp. It is the estimator telling you a copy-number or multiplicity assumption is wrong, and which orthogonal measurement to read.
(c) Three clusters: {1.00, 0.97, 1.00}, {0.80}, {0.40}. The first is the founding clone. Could SMAD4 (0.80) and ARID1A (0.40) be sibling branches?
Pigeonhole rule: siblings occupy disjoint sets of cells, so their CCFs sum to at most 1.
0.80 + 0.40 = 1.20 > 1 → not siblings. One is nested inside the other, and the larger is ancestral. The frequencies alone force the topology:
normal cell
│ TP53 p.(Arg175His), KRAS p.(Gly12Asp),
│ CDKN2A nonsense + deletion of the second allele
▼
founding clone CCF = 1.00
│ SMAD4 p.(Arg361His)
▼
subclone B CCF = 0.80
│ ARID1A p.(Gln456fs)
▼
subclone C CCF = 0.40 (nested inside B)
Two caveats. This rests on the infinite-sites assumption — each position mutates at most once — which fails at CpG hotspots and wherever LOH has erased a mutated allele. And it is one biopsy: bulk sequencing cannot resolve subclones below ~5–10% CCF, precisely where a pre-existing resistant clone sits before treatment selects for it.
(d) At ρ = 0.50 the diploid m = 1 factor becomes 2/0.50 = 4.000:
φ(TP53) = 0.35 × 4.000 = 1.40 — impossible φ(SMAD4) = 0.28 × 4.000 = 1.12 — impossible
Two independent diploid loci above 1 is not a multiplicity coincidence, it is a purity error. At a diploid locus with m = 1, VAF = ρφ/2 ≤ ρ/2, so
ρ ≥ 2 × VAF_max → here 2 × 0.35 = 0.70
The highest-VAF clonal diploid mutation is a free purity calibrator.
Where you went wrong
| Error pattern | What to re-read |
|---|---|
| Divided 0.05 by the number of variants in your file | 1(b) — the threshold prices the genome, not the file |
| Used 5 × 10⁻⁸ regardless of ancestry or assay | 1(c)–(d) — FWER 9.5% and 39% |
| Read D' = 1 as "good tag", or forgot the 1/r² penalty | 2(d)–(e) — r² = 1 needs matched frequencies |
| Believed a bigger sample fixes confounding | 3(e) — stratification is bias, not variance |
| Treated a tiny p-value as evidence of causation | 3(c) — p = 6 × 10⁻¹³² with zero causal effect |
| Assumed enough carriers means enough power | 4(c) — 100 carriers, power 0.0003% |
| Powered a replication on the published effect size | 5(d) — the penalty is the square of the inflation |
| Quoted a decile relative risk without the baseline | 6(b) — 3.67× is 5.35 percentage points |
| Called failure a biological difference | 6(d) — a property of the training data |
| Counted several in-silico predictors as separate evidence | 7(d) — not conditionally independent |
| Read a as a , or clamped CCF to 1 | 8(a)–(b) — CCF > 1 tells you the copy state |