Contents
57 — Genomics in practice Part 12 — Applications and ethics

57 — Genomics in practice

Before this: Ch 46 — Variant calling · Ch 53 — Polygenic scores · Ch 55 — Clinical variant interpretation · Time: ~55 min

Statistics needed: S1 Probability · S6 Likelihood and Bayes

What you'll be able to do

The core idea

Every application in this chapter runs the same machine. A measurement — a , a read count, a set of lengths — changes a belief about a state of the world. The measurement supplies a likelihood ratio: how much more probable is this observation if the state holds than if it does not. It does not supply the belief.

A genomic test never returns a fact about the individual in front of you. It returns a likelihood ratio, and the answer you actually want is that ratio multiplied by a prior the assay cannot see.

Statistics: turning a prior and a likelihood into a posterior is S1 §5, and why the answer swings on a base rate the assay knows nothing about is S1 §8 — assumed here, not re-derived.

That sentence explains the whole success-and-failure pattern. Where the prior is high — a critically ill infant with an unexplained syndrome, an isolate from a suspected outbreak, a family with a known pathogenic variant — modest evidence is decisive and the technology looks miraculous. Where the prior is low — population screening for a rare condition, predicting one person's complex-disease risk — even excellent evidence leaves you deeply uncertain, and the technology looks like it is over-promising. It isn't lying; the arithmetic is not what people expect.

The second half of the pattern is whether the population a model was trained on matches the population it is applied to. Agriculture gets this right by construction; human get it wrong by historical accident. The contrast in §7 is the most instructive thing here, because the estimator is identical and only the design differs.


1. The rare-disease diagnostic path

Most rare diseases are genetic, and most are individually so rare that no clinician sees two cases. The historical result was the : years of specialists and single-gene tests ending in no answer. Sequencing inverted the search — instead of asking "does this patient have disease X?" one gene at a time, sequence everything and ask which of ~19,442 protein-coding genes (verified facts) carries a variant that explains this .

graph TD PH["Deep phenotyping<br/><i>HPO terms, not free text</i>"] --> TEST FH["Pedigree<br/><i>inheritance mode narrows the search</i>"] --> TEST TEST["Exome or genome sequencing<br/><i>proband, ideally as a trio</i>"] --> CALL["Variant calling (Ch 46)"] CALL --> FILT["Filter: frequency, predicted effect,<br/>segregation, gene–phenotype match"] FILT --> CLASS["ACMG/AMP classification (Ch 55)"] CLASS -->|P / LP| DX["Diagnosis: counsel, test relatives,<br/>act if actionable"] CLASS -->|VUS or nothing| REA["Reanalysis in 12–24 months<br/><i>the knowledge base moves, the data doesn't</i>"] REA --> EXTRA["RNA-seq, long reads, methylation<br/>episignature, cross-centre matchmaking"]

— proband plus both parents — is the largest single design improvement. It identifies de novo variants directly rather than inferring them, and de novo variants are the commonest cause of severe paediatric disease (Ch 54). It also collapses the candidate list: a variant carried by a healthy parent is rarely causing the child's severe early-onset condition.

Realistic yield is 30–40% for exome or genome in rare paediatric disease, with genome plausibly ahead — though read the comparison the way this chapter asks you to read everything else. The current meta-analysis (Pandey et al., Genet Med 2025) pools just three within-cohort studies for the head-to-head: 30.6% for genome against 23.2% for exome, an odds ratio of 1.7 with a 95% interval of 0.94–2.92, P = .13. The intervals overlap heavily and the difference is not statistically significant; three of the authors are Illumina employees or stockholders. An eight-point gap quoted without those three facts is exactly the kind of number this chapter exists to teach you to interrogate. Yield in any case depends far more on who was selected than on the platform: a well-phenotyped syndromic paediatric cohort exceeds 50%, an adult neuropathy cohort may not reach 15%.

So the majority of tests are negative, and a negative is not a null result. It excludes a large class of explanations and deposits a dataset that can be reanalysed — reanalysis after 12–24 months adds several points of yield with no new sample, because what changed was the literature, not the patient. Genome over exome buys , deep-intronic splice variants and ; long reads add the segmental-duplication genes short reads cannot map or (Ch 40). What a targeted repeat genotyper actually does with a "negative" genome is lab 11, run on real 1000 Genomes data.

2. When the clock is the constraint

In intensive care, a diagnosis arriving in six weeks is often useless — the decision it would have informed has been made. So the same test is re-engineered around latency: expedited consent, courier transport, a prioritised flow cell, automated -driven prioritisation, and a clinician on call to interpret. Published turnaround runs from a median around 7 days in routine rapid programmes to demonstration cases under 24 hours, with diagnostic yield in critically ill infants around 35–40% — higher than the general figure, because admission to a NICU with an unexplained presentation is itself a strong prior.

The metric that matters is not diagnosis but change in management: in published series roughly three-quarters of diagnosed infants had care altered — a drug started or stopped, surgery avoided, a metabolic diet begun, or a redirection to palliative care because a lethal diagnosis was now certain rather than suspected. That last outcome is a real benefit and is routinely miscounted as a failure.

3. Screening is not diagnosis: NIPT and the arithmetic

The mechanism

Maternal plasma contains cell-free DNA — short, -length (~150–200 bp) fragments from dying cells. In pregnancy a portion, the fetal fraction f, comes from apoptotic placental trophoblast; typically f ≈ 4–20%, median around 10% at 10–13 weeks. Note what that says: the assay measures the placenta, not the fetus.

Non-invasive prenatal testing is then a counting experiment. Sequence millions of cfDNA fragments, map them, and count how many land on 21 — about 1.5% of mappable fragments. If the fetus is trisomic, the fetal portion carries three copies instead of two, so the expected chr21 share becomes

(1-f)\cdot 1 \;+\; f\cdot \tfrac{3}{2} \;=\; 1 + \tfrac{f}{2}

times its euploid value: at f = 0.10, a 5% excess in a category holding 1.5% of reads. With N mapped fragments and chr21 share p, the count is binomial and the excess has

z \;=\; \frac{p\,f/2}{\sqrt{p(1-p)/N}} \;=\; \frac{f}{2}\sqrt{\frac{Np}{1-p}}

Statistics: the binomial count and its variance p(1−p)/N, and the normal approximation that lets you read the excess as a z, are covered in S2 §1 and §3.

Put f = 0.10, p = 0.015, demand z = 4: N ≈ 4.2 × 10⁵ mapped fragments — a rounding error of a genome, which is why is cheap. But N scales as 1/f²: at f = 0.04 you need ~2.6 × 10⁶, and by f = 0.02 counting noise is no longer the limit — GC bias and library variance are. Hence the industry floor: below ~4% fetal fraction, labs decline to call.

One consequence is easy to miss. Fetal fraction falls with maternal weight (more maternal cfDNA diluting the same signal) and is lower in 13 and 18 (smaller placentas), so "no-call" results are enriched for — a failed test is weak positive evidence, not a neutral outcome.

The arithmetic

NIPT for trisomy 21 has pooled sensitivity around 99.7% and a false-positive rate around 0.04%. Marketing calls this "over 99% accurate"; patients hear "if it says positive, it is". Take 10,000 pregnancies at a prior of 1 in 1,000:

                    trisomy 21     euploid      total
  screen positive        9.97          4.00      13.97
  screen negative        0.03      9,986.00   9,986.03
  ------------------------------------------------------
  total                 10.00      9,990.00  10,000.00

  PPV = 9.97 / 13.97 = 71.4%     NPV = 9986.00 / 9986.03 = 99.9997%

A 99.7%-sensitive, 99.96%-specific test returns a positive that is wrong 28% of the time. Nothing is broken: the false positives are drawn from a pool 999 times larger. Move the prior to 1 in 100 and the same assay gives TP 99.7, FP 3.96, PPV = 96% — identical laboratory, 25 points of difference. Quoting NIPT performance without a population is therefore meaningless, and one result letter means different things to two patients.

The negative predictive value is extraordinary in both cases. A negative NIPT is genuinely informative; a positive NIPT is a reason to do a different test. That asymmetry is the definition of a .

Where the false positives come from

Some are statistical. The interesting ones are not — the assay measured something real that was not the fetus.

SourceWhat is actually true
Confined placental The placenta genuinely carries the trisomy; the fetus does not. The measurement is correct about the tissue it sampled
Vanishing twinA demised co-twin's placenta is still shedding cfDNA
Maternal CNV or mosaicismThe duplication, deletion, or age-related X loss is the mother's — and most cfDNA is hers
Occult maternal malignancyTumour cfDNA is aneuploid; multiple discordant aneuploidies are a recognised rare cancer presentation

Confirmation is by a diagnostic procedure: chorionic villus sampling (~10–13 weeks) or amniocentesis (~15 weeks on), with , QF- or chromosomal microarray; procedure-related loss is on the order of 1 in 500 or better in contemporary series. CVS samples villi — placenta again — so confined placental mosaicism survives into it; amniocentesis samples fetal-origin amniocytes and does not. Hence the rule professional bodies restate constantly: no irreversible decision on a screening result alone.

4. Screening across the lifespan

Newborn screening is mostly not genomic: a dried blood spot, tandem mass spectrometry, 35–60 conditions chosen on Wilson–Jungner logic — serious, detectable presymptomatically, and treatable, with earlier treatment better. Replacing biochemistry with sequencing breaks that logic in a specific way. Biochemistry measures the disease process; sequencing measures a whose in an unselected population is unknown, because every published penetrance estimate came from families ascertained because someone was affected (Ch 55). a million well babies and some number are labelled with a condition they will never develop.

Current programmes make that trade-off explicit rather than pretending it away. The UK Generation Study (Genomics England with the NHS) is sequencing 100,000 newborns for more than 500 genes covering 200+ conditions (the list grew by 97 genes and 48 conditions in April 2026), restricted to conditions treatable in childhood, returning no adult-onset findings and no carrier status; recruitment began in October 2024 and runs to the end of 2026, with results returned into 2027 and about 1 in 100 babies expected to be suspected of carrying a detectable treatable condition — a suspicion, not a diagnosis, which the NHS then has to confirm. BabySeq and GUARDIAN in the US explore adjacent design points. The open questions are not technical: which conditions qualify, who consents for someone who cannot, what happens to the data for eighty years, and whether a healthy child labelled at risk has been harmed.

Carrier screening asks whether a couple could have an affected child, and it has quietly executed this curriculum's ancestry lesson. Panels were historically ancestry-targeted — specific founder variants offered to specific groups — which fails because self-reported ancestry is a poor proxy for which founder variants someone carries, and fails completely for people of admixed background. Guidance has moved to pan-ethnic panels defined by carrier-frequency thresholds. is continuous and measurable; the social categories people report are a different object, and building a clinical protocol on the second is a design error with a predictable failure mode.

Preimplantation genetic testing applies these tests to a five-cell trophectoderm biopsy after whole-genome amplification. Amplification causes , so PGT-M does not simply genotype the variant — it genotypes linked markers and calls the , robust to losing any single locus, which is why the family's must be established in advance (Ch 14). PGT-A uses low-pass sequencing for copy number and is genuinely contested: the biopsy samples trophectoderm rather than inner cell mass, embryos reported as mosaic have produced healthy children, and randomised evidence for improved live birth per cycle started is equivocal. Selecting embryos on is the live controversy — the relevant quantity is variance among siblings, which is small, so the expected gain from choosing among a handful of embryos is a fraction of a standard deviation, on an uncertain causal model, with unknown correlated effects. Professional societies say it is not ready (Ch 53, Ch 58).

5. Pharmacogenomics: the part that should have been easy

Most drugs are cleared by hepatic enzymes, dominated by the cytochrome P450 family; activity varies between people, much of that variation is genetic, and exposure is roughly inversely proportional to clearance. Three mechanisms produce three different failures: prodrug activation ( → no drug effect; → overdose), active-drug clearance (loss of function → accumulation and toxicity), and immune recognition, which is not metabolism at all — a specific HLA presents the drug or a drug-modified peptide to T cells.

GeneDrugMechanismWhat the variant does
CYP2C19clopidogrelprodrug activationPoor metabolisers make little active metabolite; antiplatelet effect fails and stent thrombosis risk rises. Alternatives (prasugrel, ticagrelor) are not CYP2C19-dependent
CYP2D6codeine, tramadolprodrug → morphinePoor metabolisers get no analgesia. Ultrarapid metabolisers (gene duplication) generate morphine fast enough to cause respiratory depression — deaths in children after tonsillectomy and in breastfed infants drove contraindications
TPMT, NUDT15azathioprine, mercaptopurineinactivation of active metabolitesLoss of function → thiopurine accumulation → life-threatening myelosuppression at standard doses. Poor metabolisers need roughly a tenth of the dose
DPYD5-fluorouracil, capecitabinecatabolism of the active drugDeficient patients suffer severe, sometimes fatal toxicity from a first standard dose. The EMA has recommended pre-treatment testing since 2020 and NHS England requires it; US uptake has been slower
HLA-Babacavirimmune (peptide–HLA–TCR)HLA-B*57:01 carriers develop hypersensitivity that can be fatal on re-challenge. Prospective randomised evidence; screening is standard of care worldwide. (HLA-B*15:02 and carbamazepine is the same mechanism)

Three structural points. Allele definitions are , not single variants — the star-allele system names combinations (*1 is reference, CYP2D6 *4 is a splice-disrupting haplotype), a diplotype maps to an activity score, and the score maps to a class. Genotyping CYP2D6 is genuinely hard: whole-gene deletions, duplications up to a dozen copies, and hybrid genes with the neighbouring CYP2D7 — a real application for copy-number-aware calling and long reads (Ch 46). And is not phenotype: a genotypic normal metaboliser on a strong inhibitor behaves as a poor metaboliser — .

Pre-emptive versus reactive. Reactive testing orders the genotype when the prescription is written and pays in days of delay. Pre-emptive testing genotypes a panel once, stores it, and fires decision support at the moment of prescribing. Pre-emptive wins on arithmetic — most people are eventually prescribed a drug with pharmacogenomic guidance, and a genotype never expires. CPIC exists because the bottleneck was never the science: its guidelines do not argue about whether an association is real, they specify what to do with a given diplotype in a form that can be encoded. That is the artefact missing from most translational genomics.

The honest account of the twenty-year lag has five parts and none of them is biology:

  1. Evidence-hierarchy mismatch. The evidence was pharmacokinetic and observational; payers wanted randomised trials with hard clinical endpoints — slow, costly, and awkward when the mechanism is not in dispute. A large European randomised implementation study eventually reported roughly a 30% reduction in clinically relevant adverse drug reactions from a 12-gene pre-emptive panel, but not until the 2020s.
  2. Workflow, not knowledge. A PDF in a chart does nothing. The result must be discrete, machine-readable, present before the prescription, and attached to an alert that fires once and correctly — competing for attention against every other interruptive alert, and losing.
  3. A result whose interpretation mutates. The genotype is fixed for life; its clinical meaning is not, and health systems are not built to re-issue old results under new rules.
  4. Reimbursement. for pre-emptive panels has been patchy and inconsistent between systems, because the cost lands years before the benefit and often in a different budget.
  5. Population representation. Star-allele panels were largely defined in European-ancestry samples. CYP2D6 and CYP2C19 allele spectra differ substantially across ancestries, and the NUDT15 variants that dominate thiopurine risk in East and South Asian populations were characterised long after TPMT. A panel genotyping only common European alleles has genuinely lower sensitivity elsewhere — a technical, fixable, long-unfixed inequity.

6. Pathogen genomics

Bacterial genomes are megabases and viral genomes kilobases, so sequencing an isolate costs almost nothing. The substrate is a phylogeny of thousands of near-identical genomes (Ch 34), and the clock runs fast enough relative to transmission that a transmission chain leaves resolvable structure.

Outbreak reconstruction. Isolates from a genuine recent chain differ by a handful of substitutions; isolates merely sharing a species differ by hundreds. The underrated use is the negative one: a suspected ward outbreak whose isolates prove phylogenetically unrelated was never an outbreak, and the response can be stood down. The positive use — near-identical isolates across wards, hospitals or months — finds transmission nobody suspected, often via an environmental reservoir. The caveat is structural: a pathogen phylogeny is not a transmission tree. Within-host diversity means a donor transmits a sample of their own population; unsampled intermediates make A→B look direct when it was A→X→B; and pathogen coalescence predates the transmission events. Genomic evidence bounds a transmission hypothesis, it does not adjudicate one.

Resistance prediction from sequence. For some organisms this is nearly a lookup — acquired resistance sits on discrete genes (mecA in MRSA, carbapenemase genes in Enterobacterales) and presence/absence is the prediction. For others it is point in the target, such as rpoB for rifampicin in Mycobacterium tuberculosis. TB is the flagship, because culture-based susceptibility testing takes weeks for an organism that doubles in a day: the WHO catalogue of resistance-associated mutations (2nd edition, 2023) was built from more than 52,000 isolates with matched sequence and across 67 countries and lists over 30,000 variants for 13 drugs. It also shows the method's boundary. The first edition contained zero variants associated with bedaquiline resistance; the second contains 86. The biology did not change — the evidence base caught up. A catalogue encodes only what has been observed, which makes predicting resistance far safer than predicting susceptibility: absence of a known resistance mutation is not evidence of absence of an unknown one.

Wastewater surveillance changes the sampling unit from patient to population. One influent sample integrates a whole sewershed, removing test-seeking bias and capturing asymptomatic infection with no clinical encounter. It yields a quantitative signal (copies per capita over time) and a compositional one: sequence the mixture, read frequencies at lineage-defining sites, and solve a constrained least-squares problem for lineage proportions. Limits are real — shedding varies enormously between people, nucleic acid decays in transit, sewersheds rarely match administrative boundaries, and there is no denominator.

SARS-CoV-2 was the largest real-time demonstration the field has had: more than fifteen million genomes deposited within a few years, feeding variant designation, vaccine strain selection and continuous tracking of lineage displacement. It also exposed the failure mode. Sequencing capacity was distributed extremely unevenly, so the "global" phylogeny was a biased sample of who could afford to sequence. Genomic surveillance measures the surveillance system as much as the pathogen.

7. Agriculture, and why genomic prediction works there

Marker-assisted selection came first: find a marker in with a large-effect and select on it. It works for oligogenic traits — a single disease-resistance locus is exactly this — and fails for yield, milk or growth, which are polygenic in the sense Part 6 describes.

Genomic selection (Meuwissen, Hayes and Goddard, 2001) drops the mapping step. a reference population with and markers, fit all markers simultaneously with shrinkage — ridge regression, equivalently a mixed model on a genomic relationship matrix, or a Bayesian variable-selection prior when large-effect loci exist — and predict a genomic estimated for candidates never phenotyped. Shrinkage when markers outnumber animals, and its mixed-model equivalent, are covered in S7 §6 and §8; this chapter assumes them. What deserves your attention is why it works so much better here than the identical machinery does in human genetics.

Genomic selection in breedingHuman
Training vs targetSame breed, often the same herd book; the reference is re-estimated annually as new progeny are phenotypedTraining overwhelmingly European-ancestry biobanks, target frequently not; accuracy decays with genetic distance from the training sample
LD structureSmall Nₑ → long, consistent LD; marker–QTL is stable across the populationLarge Nₑ → short LD blocks whose extent and phase differ between populations, so a stops tagging (Ch 29)
EnvironmentDesigned, measured, replicated — ration, photoperiod, randomised plots. G×E is estimated on purposeUncontrolled, unmeasured, entangled with ancestry, geography and socioeconomic position
PhenotypePrecise and repeated, often across many progeny of one sireSelf-report, EHR codes, or one clinic measurement
What is predictedA breeding value — additive merit transmitted to progeny, realised as an average over hundreds of offspringOne person's trait or risk, where residual variance is the entire problem
FeedbackRealised gain is measured each generation and the model refitNo comparable closed loop

The dairy result is the cleanest demonstration in applied quantitative genetics. Because a bull can be genotyped at birth rather than waiting years for daughters to be milked, the sire generation interval in US Holsteins fell from roughly seven years to under two and a half, and annual genetic gain for several traits roughly doubled after genomic selection was adopted in 2009. Response to selection is R = h²S per generation, where S is the selection differential (Ch 31); the breeder's form ΔG = i·r·σA / L is the one to reason with here, with i = S/σP the standardised intensity, r the accuracy of the estimated breeding value and L the generation interval. Genomic selection did not raise . It raised r — the accuracy with which a young, unphenotyped animal's breeding value can be estimated, which under parent average alone is poor — allowed higher intensity i among candidates now cheap to , and slashed L. All three enter the annual rate directly.

The lesson for human genetics is the opposite of the one usually drawn. Genomic prediction working spectacularly in cattle and poorly across human populations is not evidence that populations differ biologically in any relevant way. Prediction accuracy is a property of the study design — matched training and target populations, controlled environments, measured phenotypes, short generations. failure in human polygenic scores is caused by whom we sampled, and it is fixed by sampling differently.

Gene editing (Ch 38) is now entering these programmes, and the regulatory question everywhere is whether regulation attaches to the product or the process. The US Coordinated Framework splits oversight between USDA, FDA and EPA and largely exempts plant edits achievable by conventional breeding; in April 2025 the FDA approved the gene edit behind PIC's PRRS-resistant pig, among the first gene-edited livestock approvals in the US. The UK's Genetic Technology (Precision Breeding) Act 2023 became operational through secondary legislation and the first crops have come through: Rothamsted drilled a precision-bred Camelina sativa under a release notice in May 2026, and its low-asparagine, low-acrylamide wheat was granted precision bred status in July 2026. Application to animals is paused pending welfare evidence. The EU regulated such organisms under its 2001 GMO directive for two decades, a position the CJEU confirmed for mutagenesis techniques in 2018 — and then moved. On 17 June 2026 the European Parliament gave final approval to a New Genomic Techniques regulation splitting NGT plants into two categories: category 1, edits judged equivalent to what conventional breeding could have produced, which follow a verification procedure instead of the full GMO authorisation and risk-assessment regime, and category 2, which stay inside that regime. It enters into force twenty days after publication in the Official Journal and applies two years later.

So the framing to retire is "one technique, three philosophies". All three systems now key, at least in part, on the edit rather than the process — and the EU's two-tier structure closely parallels the UK's Precision Breeding Act. What still differs is institutional and temporal: three US agencies under a coordinated framework against a single statute either side of the Channel, and a UK regime already issuing release notices against an EU one that does not bite until 2028. The interesting question is no longer which philosophy each jurisdiction holds, but how fast a technology forces convergence on the product-versus-process question — and what happens to trade in the years while they are out of step.

8. Forensics: likelihood ratios and non-consenting relatives

STR profiling uses loci where a short motif is repeated a variable number of times, so differ in length and can be sized by capillary electrophoresis after multiplex . STRs beat SNPs here for three reasons that still hold: high polymorphism (many alleles per locus, so each contributes large exclusion power), short amplicons that survive degraded samples, and thirty years of legacy database. CODIS expanded from 13 to 20 core loci in 2017.

The random match probability is computed exactly as Part 5 would have you do it: per-locus frequencies under Hardy–Weinberg (Ch 26), multiplied across loci under approximate equilibrium, with a θ (FST) correction because a suspect and the true source are likelier to share a subpopulation than the naive product assumes (Ch 28). At 20 loci, RMPs of 10⁻²⁰ or smaller are routine for a complete single-source profile, reported as a likelihood ratio:

LR = \frac{P(\text{profile} \mid H_p:\ \text{the suspect is the source})}{P(\text{profile} \mid H_d:\ \text{someone else is the source})}

Statistics: reading a likelihood ratio as a unit of evidence, and the odds form that multiplies it by a prior, are covered in S6 §4.

For a clean single-source match the numerator is essentially 1 and LR = 1/RMP. And here the core idea reappears in a courtroom: the LR is not the probability the defendant is guilty. Posterior odds = prior odds × LR. Asserting that a 10⁻²⁰ match probability means a 10⁻²⁰ chance of innocence swaps the conditionals — the . Two corrections follow: relatives are not "unrelated persons" (a full sibling's match probability is many orders of magnitude higher, so Hd must name a relative when one is plausible), and a database trawl has a different prior structure than a match to a suspect identified independently.

Mixtures are the genuine difficulty. Most casework is a mixture of two or more contributors at low template with PCR artefacts:

locus D3S1358 — electropherogram peak heights (RFU)

allele:      14      15      16      17      18
           ┌───┐                   ┌───┐
   3200    │   │           ┌───┐   │   │
           │   │   ┌───┐   │   │   │   │
   1100    │   │   │   │   │   │   │   │   ┌─┐
    140    │   │   │   │   │   │   │   │   │ │  ← stutter? minor contributor?
           └───┘   └───┘   └───┘   └───┘   └─┘     drop-in?

requires estimating jointly: how many contributors, in what proportions, which peaks are true alleles versus PCR stutter, which true alleles dropped out below threshold, and which are drop-in contamination. Continuous probabilistic genotyping (STRmix, TrueAllele) models peak heights with an explicit stochastic PCR model and computes the LR by MCMC — right in principle. What remains hard: the assumed number of contributors is an input that materially changes the output; different packages on identical data have produced LRs differing by orders of magnitude; and source-code disclosure has been repeatedly litigated, so a number presented as objective may not be independently reproducible. For low-template, high-contributor mixtures, the LR is model-dependent and should be presented that way.

Familial searching deliberately looks for a near-match in an offender database, implying a first-degree relative. It works, and it has a distributional consequence: offender databases over-represent some communities, so familial searching extends that over-representation to relatives who were never arrested.

Forensic investigative genetic genealogy (FIGG) is a different technique. Generate a dense SNP profile from the crime-scene sample, upload it to a consumer genealogy database that permits law-enforcement matching, and find identity-by-descent segments shared with distant relatives — third to fifth cousins suffice. Then build trees from birth, marriage and census records and triangulate to people in the right place at the right time. The genetics narrows millions to a few dozen families; genealogy and police work do the rest. The famous case: the Golden State Killer, unidentified since the 1970s, was linked to Joseph James DeAngelo through a GEDmatch upload, arrested in April 2018, and pleaded guilty in 2020. Hundreds of cases have followed.

The privacy problem is structural and does not dissolve with better policy hygiene. Because IBD sharing extends to distant cousins, a database holding a small fraction of a population can identify most of it — published analysis found ~2% returns a third-cousin-or-closer match for the large majority of a population. You cannot opt out of a database your cousins joined; one uploader's consent exposes information about hundreds of people who never heard of the service. Governance has lagged: a US Department of Justice interim policy (approved September 2019, effective 1 November 2019) restricts federal investigative genealogy to violent crime and public-safety cases, requires other leads to be exhausted and investigators to identify themselves, and permits searching only services that notify users — but binds federal agencies and grantees, not the state and local police who are most users. GEDmatch moved to opt-in after public backlash and was later sold; a few states require judicial authorisation, most do not.

Both positions are serious. Violent crimes with no other lead have been solved and wrongly convicted people exonerated. And the method conducts, without warrant or notice, a search implicating non-consenting third parties, unevenly distributed across a population. Ch 58 takes the argument further; this chapter insists only that you see the exposure as a property of IBD, not a policy choice.

9. Ancient DNA: damage as a signature

DNA in a dead organism degrades in chemically specific ways, and that specificity is what makes the field possible. Depurination breaks the backbone, so fragments shorten with age — typically tens of base pairs. Cytosine deaminates to uracil, preferentially in single-stranded overhangs at fragment ends, and a polymerase reads uracil as thymine. In a conventional double-stranded library this gives a diagnostic asymmetry: elevated C→T at 5′ ends and its complement G→A at 3′ ends, both decaying roughly exponentially inward.

misincorporation frequency vs distance from fragment end

  0.30 ┤ ●                                        ● ← G→A at 3′ end
       │  ●                                      ●
  0.20 ┤   ●                                    ●
       │    ●●                                ●●
  0.10 ┤      ●●●                          ●●●
       │         ●●●●●●●●●●●●●●●●●●●●●●●●●●
  0.00 ┼────────────────────────────────────────────
        5′ 1  3  5  7  9 ...        ... 9  7  5  3  1 3′
        ↑ C→T at 5′ end

This is authentication. A modern contaminant — excavator, curator, technician — has long fragments and no damage, so the two read populations are separable, and in hard cases you restrict analysis to damaged reads, trading depth for certainty. Enzymatic UDG treatment excises uracils and removes both the errors and the evidence, which is why libraries are often built both ways. Everything else is physical: positive-pressure clean rooms far from any product, full-body suits, bleach and UV, and dual-indexed libraries so index hopping is detectable (Ch 40). Endogenous content is often under 1% — most reads are soil bacteria — with the petrous temporal bone and tooth cementum the exceptions.

What it established, none of which could have been reached another way:

And the humility. Preservation favours cold, dry, high-latitude sites, so the record maps preservation as much as history — tropical Africa and South and Southeast Asia are drastically under-sampled, which is a sampling artefact, not a statement about where history happened. Sequenced individuals are few and often from selective burials. Admixture-graph models are underdetermined: several topologies fit the same summary statistics, and proportions depend on the outgroups chosen. Most importantly, archaeological cultures are not populations and populations are not languages — inferring a language shift from a genetic turnover is an argument, not a measurement. These results are recruited into nationalist claims about who was "originally" somewhere with depressing regularity, and naming ancient genetic clusters after modern countries invites exactly that (Ch 58).

10. Metagenomics, eDNA, and the compositional trap

Sequence a community rather than an organism and you get metagenomics: 16S rRNA amplicon surveys (cheap, taxonomic, -biased, roughly genus-level) or shotgun sequencing (strain resolution and gene content, at the cost of depth and host-read removal).

The statistical trap is easy to state and routinely ignored. Sequencing yields relative abundances that sum to one. The data live on a simplex, not in ℝᴰ: if one taxon blooms, every other taxon's proportion falls whether or not anything happened to it, so pairwise correlations are spurious by construction and differential-abundance tests have inflated false-discovery rates. Remedies are log-ratio transforms (CLR/ILR, Aitchison geometry) or genuine absolute quantification — spike-ins, flow cytometry, total-load . Honestly: recent work finds the commonly recommended do not fully eliminate the artefact, and measuring total load is the only clean fix. Zeros compound it, because a zero count is unobserved, not absent.

The state of causal claims should be stated plainly. There are thousands of published microbiome–disease associations and very few demonstrations of causation. Cross-sectional designs are confounded by diet, medication (metformin and proton-pump inhibitors have among the largest single effects on gut composition), and stool consistency/transit time, one of the strongest covariates in the data — and reverse causation is entirely plausible, since being ill changes what you eat and what your gut does. What does support causation: gnotobiotic-animal transfer, faecal microbiota transplant for recurrent Clostridioides difficile infection (the one unambiguous clinical success, now with approved live biotherapeutics), and with host-genetic instruments — which is instrument-poor, because host explains little microbiome variance, with LCT/Bifidobacterium and FUT2 among the few robust exceptions. The gap between association and causation is a study-design gap, not a mystery.

Environmental DNA applies the same sequencing to water, soil, sediment or air, amplifying a barcode locus (COI for invertebrates, 12S rRNA for vertebrates) and matching a reference library. It detects rare and cryptic species without capturing anything and scales beyond any field team: invasive-species early warning, fish stock assessment, and — joining this section to the last — that two-million-year-old Greenland ecosystem. Its limits mirror its strengths: detection is not presence here, now (DNA is transported and persists), reference libraries are incomplete, and read counts are a poor abundance estimate because PCR amplification is not uniform across templates.

11. Direct-to-consumer testing

What is actually measured. With few exceptions a consumer test is a genotyping array, not sequencing: roughly 600,000–1,000,000 pre-selected probe positions chosen to tag common variation via LD (Ch 29). Three consequences follow directly and explain nearly every reported problem.

  1. It cannot see what is not on the chip. Most rare pathogenic variants are absent by design.
  2. Array calling depends on cluster separation, which depends on having enough carriers in the calling batch. A singleton lands in no cluster and is called badly — the mechanism behind the widely cited finding that about 40% of variants reported in consumer raw-data files, when referred for clinical confirmation, were false positives (Tandy-Connor et al., 2018), and that some variants labelled "increased risk" by third-party services were benign and common in population databases.
  3. of a gene is not coverage of a condition. A test genotyping three founder BRCA1/BRCA2 variants common in Ashkenazi Jewish populations, out of thousands of known pathogenic variants in those genes, gives a "negative" that is close to uninformative for someone with a family history. FDA authorised such reports as class II devices with special controls — including required non-diagnostic language — precisely because they are narrow.

Ancestry versus health. Ancestry inference is a supervised assignment against a reference panel: global proportions (model-based clustering or projection) or local ancestry (an HMM along each assigning segments to panels). The percentages are estimates relative to a chosen panel and an implied time depth, so they change when the company updates its panel — which users experience as "my ancestry changed" and read as unreliability. It is not unreliable; it was never measuring what they thought.

is a continuous, measurable statement about which reference populations the segments of your genome most resemble. Race is a social classification, historically contingent and different between societies. They are not the same object, they are not interchangeable, and no amount of precision in the first licenses a claim about the second.

The most technically reliable feature of these products is also the most consequential: relative matching by IBD segment sharing works very well. It is why these databases became forensic resources (§8), and why they routinely reveal misattributed parentage, undisclosed adoption and donor conception — outcomes arriving without counselling, to people who never took a test.

Third-party reinterpretation stacks two independent failure modes: the analytical one above (the call is wrong) and an interpretive one (the call is right but the classification is wrong, outdated, or from a source that never applied ACMG criteria — Ch 55). Hence the unambiguous clinical rule: any consumer result with medical implications is confirmed in an accredited laboratory before it is acted on.

Regulation and custody. The FDA regulates health-related DTC claims through the device pathway and has authorised a limited, specific set; ancestry and "wellness" claims sit largely outside it, and third-party interpretation services are effectively unregulated. These companies are not HIPAA covered entities, so US health-privacy law does not reach them, and prohibits genetic discrimination in health insurance and employment but not in life, disability or long-term-care insurance. The custodial point was then demonstrated concretely: 23andMe filed for bankruptcy in March 2025, the genomic data of millions of customers became an asset in a bankruptcy estate, and it was sold with court approval in July 2025 to TTAM Research Institute, a non-profit founded by the company's co-founder, over a competing pharmaceutical bid. Whatever one thinks of the outcome, the lesson generalises: a privacy policy is a promise made by an owner who may not be the owner tomorrow (Ch 58).


Common misconceptions

What people thinkWhat's actually true
A "99% accurate" screening test is right 99% of the timeAccuracy figures are conditional on disease status. The probability you want is conditional on the result, and depends on prevalence: a 99.7%/99.96% test has a 71% PPV at a prevalence of 1 in 1,000
tests the fetusIt counts cell-free DNA, whose non-maternal portion comes from placental trophoblast. Confined placental is a true measurement of the wrong tissue
A negative genome sequence means there is no genetic causeIt means no causal variant was identified with current knowledge, methods and . Reanalysis of the same data adds diagnoses every year
failed to translate because the science was weakThe mechanisms are among the best established in the field. It failed on workflow, reimbursement, EHR integration, evidence-hierarchy mismatch and unrepresentative panels — and still isn't , because inhibitors cause
A pathogen phylogeny shows who infected whomIt bounds transmission hypotheses. Within-host diversity, unsampled intermediates and coalescence predating transmission all break the identity
Genomic selection works in cattle and not humans because human genetics is more complexIt works because training and target populations match, environments are controlled, phenotypes are precise, and the prediction is a averaged over many progeny. Same estimator, different design
A random match probability of 10⁻²⁰ means a 10⁻²⁰ chance of innocenceThat swaps the conditionals — the . Posterior odds = prior odds × LR, and Hd must name a plausible alternative source, including relatives
DNA mixture interpretation is objectiveIt requires assuming the number of contributors and modelling stutter and dropout. Different software on identical data has produced LRs differing by orders of magnitude
You control whether your DNA is in a searchable databaseIBD sharing extends to distant cousins; ~2% database coverage returns a third-cousin-or-closer match for most of a population. Your relatives' choices set your exposure
Consumer tests sequence your genomeThey genotype ~10⁶ pre-chosen array positions. Rare variants are mostly absent and unreliably called — hence ~40% false positives among raw-data variants sent for confirmation
An ancestry percentage is a fact about youIt is an estimate relative to a chosen reference panel and time depth, and it changes when the panel changes. And is not race

Worked example: two positives, one calculation

Part A — a screen-positive result for a rare microdeletion

A 29-year-old has an expanded cell-free DNA including 22q11.2 deletion syndrome. It is positive. What is the probability the fetus is affected?

Step 1 — the three numbers.

QuantityValueSource
Prior (birth prevalence)≈ 1/4,000 = 2.5 × 10⁻⁴Population epidemiology — not the test
Sensitivity0.978Wapner et al. 2015 — SNP-based assay validation, mostly contrived plasma mixtures
False-positive rate0.0076Same source; later prospective series report different figures (Ravi et al. 2018: 0.90 and 0.0026)

Step 2 — likelihood ratio. LR_+ = 0.978 / 0.0076 = 128.7

Step 3 — prior odds to posterior odds.

\text{prior odds} = \frac{2.5\times10^{-4}}{1-2.5\times10^{-4}} = 2.5006\times10^{-4}
\text{posterior odds} = 128.7 \times 2.5006\times10^{-4} = 0.0322 \;\Rightarrow\; PPV = \frac{0.0322}{1.0322} = \mathbf{3.1\%}

Step 4 — check by counting. Per 100,000 pregnancies:

  affected        = 100,000 × 1/4000   =     25.00
  true positives  = 25.00   × 0.978    =     24.45
  unaffected      = 100,000 − 25       = 99,975.00
  false positives = 99,975  × 0.0076   =    759.81
  ----------------------------------------------------
  all positives                        =    784.26
  PPV = 24.45 / 784.26                 =      3.12%

About 1 positive in 32 is a true positive — and note this test has sensitivity nearly as high as the 21 screen in §3 (97.8% against 99.7%) and a PPV more than twenty times worse, because the prior is 4× lower and the false-positive rate 19× higher. Those two factors multiply to 76, which is the ratio of the posterior odds.

Step 5 — which number is in charge? When the prior π is much smaller than FPR/sensitivity, the denominator is dominated by false positives:

PPV \;\approx\; \frac{\pi \cdot \text{sens}}{\text{FPR}}

PPV is linear in prevalence, inversely proportional to the false-positive rate, and nearly indifferent to sensitivity. Holding prevalence fixed:

FPRFalse positives per 100,000PPV
0.76%759.83.1%
0.05%50.032.8%
0.01%10.071.0%

Raising sensitivity from 97.8% to a perfect 100% moves the first row from 3.12% to 3.19%. Cutting the false-positive rate 76-fold — a specificity of 99.24% to 99.99%, which is barely a change in specificity and an enormous one in false positives — moves it to 71%. For rare conditions, specificity is the whole game — which is where assay development concentrates, and why a laboratory's observed PPV in a clinical series is the number to ask for.

Step 6 — the decision. This is a reason to offer a diagnostic test, not a diagnosis: CVS or amniocentesis with chromosomal microarray. Parental testing matters here too — a substantial share of positive 22q11.2 cfDNA screens reflect a copy-number variant carried by the mother, whose DNA is most of the sample.

Part B — the same arithmetic in a courtroom

A degraded sample yields a partial STR profile: 5 usable loci, random match probability 10⁻⁶. It is trawled against a database of 2 × 10⁷ profiles and returns one match. Expected coincidental matches in the database alone:

2\times10^{7} \times 10^{-6} = \mathbf{20}

You expected about twenty adventitious matches and found one hit. The LR of 10⁶ is arithmetically correct and, against prior odds appropriate to "one of twenty million people, no other evidence", is nowhere near identification. Had the same 5-locus profile matched a suspect identified independently by non-genetic evidence — putting perhaps a few hundred people in the frame — prior odds of ~10⁻² would give posterior odds of ~10⁴, a very different conclusion from identical genetic data.

The genetic evidence did not change. The prior did. That is the same sentence as Step 5, and it is the whole chapter.

Connections