40 — Sequencing technologies
Before this: Ch 02 · Ch 04 · Ch 36 · Ch 39 · Time: ~75 min
Statistics needed: S2 Distributions · S3 Sampling and estimation
Everything in Parts 9–11 is downstream of this chapter. Alignment, assembly, and are all attempts to recover a true sequence from a noisy, biased, fragmentary sample of it. You cannot reason about those algorithms without knowing exactly what the instrument did and how it fails.
A note on the numbers. Every platform specification below is Tier C in
reference/verified-facts.md: vendor figures, best-case, and fast-moving. They are given as ranges, verified 2026-08-13. They will be wrong within a year or two. The mechanisms will not be — read for those.
What you'll be able to do
- Explain what physical quantity each sequencing platform actually measures, and derive its error profile from that physics rather than memorising it
- Distinguish redundancy across molecules from redundancy within one molecule, and say which errors each one averages away and which survive both
- Derive the Lander–Waterman model, compute the depth needed to leave a given fraction of a genome uncovered, and explain why real data violates the model
- Justify 30× as the WGS convention and calculate why somatic calling needs an order of magnitude more
- Explain as an occupancy problem, and say when a high duplicate rate means "sequence less" versus "change the prep"
- Choose between WGS, exome capture and amplicon panels — and between short and long reads — for a stated biological question, and name the failure mode each choice buys
- Say why an orthogonal confirmation must fail differently from the assay it confirms, and what independent evidence would be needed to believe a new platform's claims
The core idea
Every sequencing technology ever built solves the same problem in the same shape. DNA is a 3-billion-character string; no instrument can read it end to end. So all of them:
- Shatter the string into fragments,
- Transduce each fragment into a physical signal — photons, or current, or migration time — and decode that signal into letters,
- Reconstruct the original computationally.
Three design choices distinguish the platforms, and everything else follows from them:
| Choice | Consequence |
|---|---|
| How long a fragment can be read | What repeats you can resolve, what structural variation you can see, whether you can assemble |
| What the signal is, and how it is decoded | The error mode — substitutions vs , random vs context-dependent |
| Whether the molecule is amplified first | Whether you have bias and , and whether chemical modifications on the original survive to be measured |
And there is one unifying statistical idea that runs through the whole chapter. Accuracy can be bought in exactly three currencies:
- Chemistry — make the single measurement better.
- Redundancy across molecules — sequence deeper. Averages out sequencing error but cannot fix an error that was already in the molecule you sequenced.
- Redundancy within a molecule — read the same physical molecule several times and average. This is what makes long reads accurate, and it is a fundamentally different operation from depth.
Hold on to that third one. It is the single most important idea in modern sequencing.
1. Sanger sequencing, and why it refuses to die
The mechanism. extends a growing strand by attacking the incoming nucleotide with the 3'-hydroxyl of the last one (Ch 04). A dideoxynucleotide (ddNTP) lacks that 3'-OH. Incorporate one and the chain stops dead.
Run a normal extension reaction spiked with a small fraction of ddNTPs, each of the four carrying a different fluorescent dye. Every molecule terminates at a random position, so the product is a population of fragments of every possible length, each labelled by its last base:
template 3'- T A C G G A T C C A G T -5'
terminating base
primer -> A A (green)
primer -> A T T (red)
primer -> A T G G (black)
primer -> A T G C C (blue)
primer -> A T G C C C (blue)
...
Separate that population by length in a capillary of polymer gel, one base of resolution, and read the dye colour as each fragment passes the detector. The order of colours is the sequence. Modern instruments run ~96 capillaries in parallel; the output is a chromatogram — four intensity traces against time.
What it gives you: ~500–1,000 bp per read, very high per-base accuracy in the clean middle of the trace, and a signal a human can inspect directly.
What it costs: one reaction per read. There is no parallelism beyond the number of capillaries. The cost per base is roughly six orders of magnitude worse than a short-read instrument.
Why it is still the confirmation standard. Not nostalgia — orthogonality. When a short-read pipeline reports a variant, the plausible explanations are: it is real; the was wrong; the library prep introduced it; the base caller mis-decoded it; a sample was swapped. Sanger shares almost none of those failure modes. Different chemistry, different detection physics, different amplification, no alignment step, and a fresh aliquot of DNA. A confirmation is only worth something if it fails differently from the thing it confirms. That principle — verify with an assay whose errors are independent — is worth generalising well beyond genomics.
Its own limits are real: it reads a population average, so it detects a minor only down to roughly 15–20% of molecules (missing low-level ), and an insertion or deletion in one allele desynchronises the two overlapping traces from that point on.
The Human Genome Project was the Sanger era's monument: an international public consortium, roughly 1990–2003, a draft in 2001 and a "finished" sequence in 2003, at a total programme cost on the order of a few billion dollars, of which the sequence generation itself was some hundreds of millions. (Approximate teaching figures — the accounting depends entirely on what you charge to the project.) The genome was not actually complete: the last ~8% — , acrocentric short arms, — waited until in 2022 and long reads (Ch 45).
2. Second generation: parallelism as the whole idea
The insight that ended the Sanger era is not chemical. It is that if you immobilise the fragments on a surface so each occupies a distinct address, you can run a billion reactions simultaneously and read them all with a camera. Throughput becomes (features per unit area) × (cycles) — an optics and surface-density problem, which is exactly the kind of problem that scales.
2.1 Cluster generation
A single fluorophore is too dim to see against background. So each fragment is amplified in place into a clonal patch of ~1,000 identical copies that fluoresce in lockstep.
The flow-cell surface carries a dense lawn of two oligonucleotides, complementary to the two adapters ligated onto every library fragment. A fragment anneals by one end, is extended, denatured, then bends over and primes on a neighbouring lawn oligo — a bridge. Isothermal cycling repeats this; one strand type is then cleaved away, leaving a clonal cluster of single-stranded templates all in the same orientation.
fragment binds bridge forms extend + denature after ~30 cycles
┌────────── ┌─────┐ ┌──────┐ ┌──────┐ ▒▒▒▒▒▒▒▒▒▒▒
│ │ │ │ │ │ │ ▒ clonal ▒
──┴──┬──┬──┬── ──┴──┬──┴──┬── ──┴──┬───┴──┴──┬─── ▒ cluster ▒
lawn oligos lawn lawn ▒▒▒▒▒▒▒▒▒▒▒
Patterned flow cells replace the random lawn with an ordered array of nanowells, and use a kinetic trick (one template entering a well amplifies fast enough to exclude any second) to put exactly one cluster per well. This maximises density but introduces its own artefact: free adapters in the mix can prime the wrong template, so a read gets attributed to the wrong sample — index hopping. The fix is to index both ends with unique pairs, so a hop produces an unassigned combination rather than a silent misassignment.
2.2 Sequencing by synthesis with reversible terminators
Now read all clusters at once. Each cycle:
- Flood with all four nucleotides. Each carries a 3' blocking group — so exactly one incorporates per template per cycle — and a cleavable fluorescent dye.
- Wash away the unincorporated.
- Image the surface in the relevant channels. Every cluster reports its next base.
- Chemically cleave the dye and the 3' block. Repeat.
One base per cycle. Read length equals cycle count. Which raises the obvious question: why stop at 150?
2.3 Dephasing: why read length is capped
Because the cluster is an ensemble, and ensembles lose synchrony.
Two failures occur at low per-cycle probability. A strand fails to incorporate or fails to deblock, and falls one base behind (phasing). A strand incorporates a nucleotide whose block was defective, and runs one base ahead (prephasing). Neither strand is removed — it keeps extending, permanently out of register, contributing the wrong base's signal to every subsequent image.
If the per-cycle probability of falling out of register is p, the fraction of the cluster still in after n cycles is (1−p)ⁿ:
| p | n = 100 | n = 150 | n = 300 | n = 600 |
|---|---|---|---|---|
| 0.001 | 90% | 86% | 74% | 55% |
| 0.002 | 82% | 74% | 55% | 30% |
| 0.005 | 61% | 47% | 22% | 5% |
The signal at cycle n is a mixture: mostly the correct base, plus contamination from positions n−1 and n+1. Base callers explicitly model and deconvolve phasing, which is why 150 cycles works at all — but the correction's variance grows with n, so the quality score falls monotonically along the read. That decay is not a defect to be fixed; it is intrinsic to reading an ensemble. Cumulative photodamage and dye carryover compound it.
This is the structural reason short reads are short. Not chemistry — statistics of synchrony loss in a population.
2.4 The error profile, and why it matters downstream
| Property | Behaviour | Downstream consequence |
|---|---|---|
| Substitutions dominate | A mis-decoded colour call | Fine for SNVs at depth; the aligner's edit model can lean on this |
| are orders of magnitude rarer | The 3' block enforces one base per cycle | Indel calls are trustworthy when the read aligns |
| Quality decays along the read | Dephasing (§2.3) | Per-base quality strings are load-bearing, not decoration (Ch 41) |
| GC bias from amplification | under-represents both GC-rich and AT-rich fragments | is not uniform; CNV calling from depth needs GC correction |
| Homopolymers handled well | Terminators enforce single incorporation | Unlike historical pyrosequencing platforms, which could not count them |
2.5 Paired-end and mate-pair: buying long-range information without long reads
Paired-end. Sequence both ends of a 300–600 bp fragment. You get two 150 bp reads with a known relative orientation and an approximately known separation.
fragment |<--------------- ~500 bp insert --------------->|
================================================
read 1 ----------> <---------- read 2
(150 bp) (150 bp)
Three things this buys, all of which matter more than the extra bases:
- Anchoring. If one mate falls in a repeat and its partner does not, the pair is placed by the unique mate. This is the single largest gain — it extends usable mapping into regions a single 150-mer cannot address (Ch 42).
- A structural-variant signal. The insert-size distribution is measured empirically per library. A pair mapping 5 kb apart in a library whose mode is 500 bp is evidence for a deletion; wrong orientation is evidence for an ; a pair split across is evidence for a . calling from short reads is largely the statistics of this distribution's tails.
- Error correction by overlap, when the insert is shorter than twice the read length.
Mate-pair. Circularise a 2–20 kb fragment with a biotinylated junction adapter, shear the circle, pull down only the fragments containing the junction, and sequence outward. You get a pair whose two mates were originally kilobases apart — long-range from a short-read instrument. The cost is a fiddly prep, chimeric artefacts, and low library complexity.
Mate-pair has largely been displaced, but note what displaced it: not only long reads, but also linked reads (many short reads sharing a barcode from one long molecule) and proximity ligation (Ch 50). All three are the same idea — attach long-range information to short reads by molecular bookkeeping. The idea outlives every implementation of it.
3. Third generation: read one molecule, in real time
Two changes at once. No amplification, so no bias, no PCR , and any chemical modification on the original bases survives to be measured. And no ensemble, so no dephasing — read length is limited only by how long a molecule you can deliver intact.
The price: you are measuring a single molecule, so the raw signal is noisy.
3.1 PacBio SMRT and the circular-consensus trick
The zero-mode waveguide. To watch one polymerase you must illuminate a volume small enough that free fluorescent nucleotides in it are countable. A ZMW is an aperture tens of nanometres across in a metal film — smaller than the wavelength of the excitation light, so light cannot propagate through it. Only an evanescent field penetrates, ~30 nm deep. The observation volume is ~10⁻²¹ L, which makes single-molecule detection possible at the micromolar nucleotide concentrations the polymerase actually needs.
One polymerase sits at the bottom. Nucleotides carry a fluorophore on the phosphate that gets cleaved off during incorporation, so each incorporation is a pulse of a few milliseconds, after which the dye diffuses away. The output is a pulse train in time — colour and duration. The gaps between pulses (interpulse duration) are slowed by chemically modified template bases, which is how kinetic detection of base modifications works.
SMRTbell. Ligate hairpin adapters to both ends of a duplex fragment and it becomes a topological circle. The polymerase goes round, and round, reading both strands repeatedly.
╭───── insert (~15–25 kb) ─────╮
hairpin hairpin
╰──── complementary strand ────╯
polymerase read: pass1 → pass2 → pass3 → ... → passk
consensus: ────────── one HiFi read ──────────
Here is the statistics that makes it work. Raw per-pass error is roughly 10–15%, dominated by — and, crucially, it is close to independent between passes, because each pass is a fresh stochastic traverse of the same template. Take a naive majority vote over k passes at per-pass error ε = 0.12:
| Passes k | P(majority wrong) | ≈ Phred |
|---|---|---|
| 5 | 1.4 × 10⁻² | Q19 |
| 9 | 2.1 × 10⁻³ | Q27 |
| 15 | 1.3 × 10⁻⁴ | Q39 |
Majority vote is only a lower bound; the real consensus is a probabilistic model over the pulse train (and must first align the passes, since indel errors mean the columns are not given). But the shape is the point: error falls exponentially in passes — roughly 1.6-fold per additional pass at ε = 0.12 — and depth cannot do this. (That constant per-pass factor is visible in the table above, and it is the Chernoff exponent: P ≈ e^(−k·D(½‖ε)), with D(½‖0.12) = 0.43.) Sequencing the same locus at 15× across fifteen different molecules averages out sequencing error but faithfully reproduces any error already present in those molecules; fifteen passes over one molecule averages out the read-out noise on that specific physical object.
The trade is arithmetic: a fixed polymerase read length must be divided between insert size and pass count. Longer inserts, fewer passes, lower accuracy. HiFi is the operating point where inserts of ~15–25 kb still get enough passes for >99.9% consensus accuracy.
What does not average out: systematic errors. Contexts where the polymerase is biased the same way on every pass — certain homopolymers, some repeat structures — survive consensus. Independence is an assumption, and it is the assumption that fails first.
3.2 Nanopore: sequencing as a signal-decoding problem
A protein pore sits in an electrically resistive membrane with a voltage across it. Ions flow; you measure the current, sampled at ~4–5 kHz. A motor protein ratchets single-stranded DNA through the pore at a few hundred bases per second. Whichever bases occupy the narrow constriction partially block the ions, so the current depends on them.
current (pA)
▁▁▁▄▄▄▄▄▁▁▁▁▁▆▆▆▆▂▂▂▂▂▂▇▇▇▃▃▃▃▁▁▁▅▅▅▅▅
└── one k-mer's worth of dwell ──┘
pore constriction sees ~5–9 bases at once, not 1
Two facts define everything about this platform.
First: the observable is a function of a , not of a base. With ~0.34 nm between bases and a constriction of comparable scale, several bases contribute to the current at any instant. So you are not reading letters; you are observing a noisy function of a sliding window over the sequence, and must invert it. That is a sequence-labelling problem, and it is solved as one — historically by an HMM over k-mer states, now by neural sequence models trained on labelled current traces.
The consequence is unusual and worth stating plainly: the decoder is software, and it keeps improving. Re-basecalling archived raw signal with a newer model produces better reads from data you already own. No other platform has that property.
Second: reading the same molecule twice. In duplex mode both strands of one duplex pass through the pore consecutively; combining two near-independent observations of the same sequence is the same redundancy trick as circular consensus, with k = 2 but a much better per-observation accuracy.
As of 2026-08-13, with R10.4.1 pores and V14 chemistry: simplex about 99.75% (Q26) on the current high-accuracy basecalling models, against a vendor record of Q28 (99.8%); duplex is usually quoted about Q30 (>99.9%), though the vendor's own accuracy page no longer publishes a duplex figure, so treat that one as unpinned. Note how much of this is the decoder, not the chemistry: the pore and kit designations have not changed while the accuracy figure moved from the Q20 that gave "Q20 chemistry" its name to Q26. Accuracy on this platform has improved faster than any other in the last few years; check current figures rather than trusting these.
Three genuine capabilities that follow from the physics:
- Ultra-long reads. Tens of kb routine, >100 kb achievable. The limit is molecular biology — high-molecular-weight extraction, wide-bore pipette tips, never vortexing — not the instrument.
- Native base modifications. 5-methylcytosine, 5-hydroxymethylcytosine and N6-methyladenine perturb the current trace, so the basecaller emits methylation calls alongside base calls. One library yields sequence, methylation and long-range phasing simultaneously, with no bisulfite conversion and no separate assay (Ch 49). This is not a marginal convenience; it changes what a single experiment can answer.
- Real-time, programmable acquisition. Data streams as it is produced, so the instrument can decide mid-read whether it wants a molecule and reverse the voltage to eject it — target enrichment implemented entirely in software.
Residual errors concentrate in homopolymers and in some methylated contexts, and skew toward indels rather than substitutions — the opposite of Illumina, which is precisely why the two platforms combine well.
4. The new entrant: sequencing by expansion
Roche AXELIOS 1, launched 29 June 2026, research-use-only.
The SBX attacks is the one identified in §3.2: bases are 0.34 nm apart, sensing apertures are not, so any nanopore reads a smeared . Sequencing by expansion sidesteps this by not reading DNA at all.
A template-directed synthesis builds a surrogate polymer — an Xpandomer — in which each incorporated nucleotide contributes a bulky reporter joined by a compacted linker. Cleaving the backbone lets the linkers unfurl, expanding the molecule roughly fiftyfold in length. The reporters are now spaced far enough apart that each occupies the sensing region alone. That Xpandomer is threaded through a nanopore on a reusable CMOS sensor array.
DNA A C G T A 0.34 nm spacing — reporters overlap
│ │ │ │ │
Xpandomer A──C──G──T──A ~50× expanded — one reporter at a time
Vendor-quoted figures as of 2026-08-13: reads in a ~400–600 bp range in the short-read (simplex) mode, extending to ~1,500 bp under favourable sample and library-prep conditions, same-day whole genomes, a reusable sensor rather than a consumable flow cell. Be careful quoting a single number here: Roche runs simplex and duplex modes with different operating points, and its own published pages are not consistent with each other on read length.
Be clear about the epistemic status: independent benchmarks do not yet exist. Everything above is mechanism plus manufacturer claim. What would need to appear before you plan work around it:
- Precision and recall on community benchmark samples at matched , from labs with no commercial relationship to the vendor, including hard regions — , homopolymers, GC extremes.
- A characterised error mode, not just an aggregate quality score. Whether errors are substitution- or -like, and whether they are context-dependent, determines which downstream tools work.
- Fully loaded cost per genome including sensor lifetime, not reagent cost.
It belongs in this chapter regardless of how it fares commercially, because the idea is clean: decouple the physical scale of the sensor from the physical scale of the polymer. Every nanopore approach fights that constraint.
5. The comparison
All figures Tier C, verified 2026-08-13, ranges deliberately. Specifications move; check current ones before designing an experiment.
| Sanger | Illumina SBS | PacBio HiFi | Nanopore | Roche SBX | |
|---|---|---|---|---|---|
| Signal | Dye colour vs migration time | Fluorescence image, 1 base/cycle | Fluorescence pulse train, real time | Ionic current trace | Current trace from expanded polymer |
| Read length | 500–1,000 bp | 2×150 bp typical (2×300 available) | ~15–25 kb | 10s of kb routine; >100 kb achievable | ~400–600 bp short-read mode; up to ~1,500 bp quoted |
| Accuracy | Very high per read | ~Q30+, ~0.1% error | >99.9% consensus | Simplex ~99.75% (Q26), vendor record Q28; duplex ~Q30 (no current vendor figure) | Not independently benchmarked |
| error | Trace ambiguity, low-level blindness | Substitutions; quality decays along read | Residual systematic in homopolymers | Indels in homopolymers, some methylated contexts | Unknown |
| Throughput | One read per capillary-reaction | Very high — 25B flow cell ≈ 8 Tb | ~100–120 Gb per SMRT Cell in ~24 h at 15–20 kb inserts (~35–100 Gb for shorter) | Scales from a USB device to a rack | Same-day WGS claimed |
| Amplified? | Yes ( before) | Yes (clusters; PCR-free libraries still cluster) | No | No | Synthesis, no clonal amplification |
| Native modifications | No | No | Kinetic detection | Yes, directly from signal | Not established |
| Cost per base | Worst by orders of magnitude | Lowest | Middle | Middle; lowest capital cost | Unknown |
| Best at | Confirming one locus | Depth-hungry, cost-sensitive work: WGS at scale, , arrays' successor | Accurate assembly, SVs, phasing | Ultra-long reads, methylation, , real time | TBD |
| Worst at | Anything at scale | Repeats, SVs, phasing, modifications | Cost per base | Homopolymer-dense sequence | Unproven |
6. Coverage, depth, and the Lander–Waterman model
Two different things are called "" and conflating them causes real errors:
- Depth — how many reads overlap a given base. Written 30×.
- — what fraction of the genome has at least one read.
The model. Take N reads of length L from a genome of length G. Assume start positions are independent and uniform. A read covers base x if it starts in the L-wide window ending at x, so
The number of reads covering x is Binomial(N, L/G), and with N large and L/G tiny this is Poisson with
which is exactly the mean depth. Hence the central results:
Statistics: the binomial-to-Poisson limit this derivation turns on — and this same coverage calculation checked against a real alignment — is S2 §2.
And Lander & Waterman's contig result: with N reads and requiring an overlap fraction θ to detect a join, the expected number of islands is N e^{−C(1−θ)}; for mere coverage (θ = 0) it is N e^{−C}.
For a 3.1 Gb haploid human genome:
| Depth C | e^(−C) | Expected uncovered bases | Expected islands (150 bp reads) |
|---|---|---|---|
| 1× | 0.368 | 1.14 Gb | 7.6 million |
| 5× | 6.7 × 10⁻³ | 20.9 Mb | 697,000 |
| 10× | 4.5 × 10⁻⁵ | 141 kb | 9,400 |
| 15× | 3.1 × 10⁻⁷ | 949 bp | 96 |
| 30× | 9.4 × 10⁻¹⁴ | 0.0003 bp | ~10⁻⁴ |
Note the third column against the fourth at 10×: essentially the whole genome is touched, and yet it falls into ~9,400 disconnected islands. Covered is not assembled (Ch 43).
Why 30×, when the model says 15× is plenty
Because the model answers the wrong question and its assumptions are false. Four corrections, in increasing order of importance:
1. You must observe both , not one base. At a heterozygous site with local depth d, the probability that every read came from the same parental is 2·(½)^d:
| Local depth d | P(one haplotype only) | × ~3M het sites |
|---|---|---|
| 5 | 6.3 × 10⁻² | ~187,000 sites |
| 10 | 2.0 × 10⁻³ | ~5,900 sites |
| 20 | 1.9 × 10⁻⁶ | ~6 sites |
| 30 | 1.9 × 10⁻⁹ | ~0.006 sites |
Missing a heterozygote is a silent false negative — the worst error class in clinical genetics, because it looks like a clean result.
2. It is the left tail of the depth distribution that matters, not the mean. The table above is indexed by local depth. A genome at mean 30× still has bases at 8×.
3. Reads are not uniformly placed. GC bias from amplification, chromatin and extraction effects, and mappability all violate the uniformity assumption. Real depth distributions are overdispersed relative to Poisson — negative-binomial is the usual working model — and the low-depth tail is far fatter than e^(−C) predicts.
Statistics: overdispersion, and why a rate that varies from place to place turns Poisson into a , are covered in S2 §5.
4. Not all reads are usable. Reads in repeats get mapping quality near zero; low-quality bases are filtered. Effective depth is materially below nominal depth, which is why the callable fraction of a 30× short-read human genome is ~95%, not the 1 − 10⁻¹³ the Poisson model promises. That missing 5% is a mappability problem, not a depth problem, and buying more of the same reads does not fix it — which is the entire argument for long reads and for references (Ch 45).
So: 30× is an empirical operating point — where SNV sensitivity and precision plateau for short-read calling — not a derived one. Naming the model and then naming the reasons it under-predicts is the honest way to teach it.
Why somatic needs far more
In germline work the variant allele fraction is 0.5 or 1.0, set by biology. In cancer it is set by the sample. A heterozygous in a subclone occupying 20% of cells, in a biopsy that is 50% tumour, appears at
Variant-supporting reads at depth d are Binomial(d, f) ≈ Poisson(df). To see at least 3 with 95% probability you need λ ≈ 6.3, so d ≈ 126× — and that is only to see it, before any question of distinguishing it from error. Section 8's worked example finishes the calculation and shows why the naive threshold collapses.
Below roughly f = 10⁻³, depth stops helping altogether, because you are re-reading the same input molecules. The binding becomes the number of distinct molecules in the tube — a circulating-tumour-DNA sample may contain only a few thousand copies of the locus. The answer there is molecular consensus: tag each original molecule with a unique identifier before amplification, sequence its descendants, and collapse them to a single high-confidence observation. Redundancy within a molecule again, achieved by bookkeeping instead of physics.
7. Library preparation: bias, duplicates, and going PCR-free
A library is a collection of DNA molecules carrying the adapters a platform requires. Fragment (sonication, enzymatic digestion, or transposase-based tagmentation that fragments and adds adapters in one step), size-select, repair the ends, ligate adapters with sample indices, optionally amplify, quantify.
are two or more reads descended from the same original molecule. They are identified by identical mapping coordinates — for paired-end reads, both outer coordinates, which is far more specific than a single end — or exactly, with unique molecular identifiers.
They matter because they are not independent observations. Count them as independent and you inflate apparent depth; worse, a polymerase error in an early PCR cycle is propagated into many reads and can look like a confidently supported variant. Convention is to mark rather than delete them (Ch 46): the duplicate rate is itself a QC signal you do not want to discard.
Duplicate rate is an occupancy problem. With M distinct molecules in the library and N reads sampled, the expected number of distinct molecules observed is M(1 − e^(−N/M)), so
At N = 310 million read pairs: a library of M = 1 billion molecules gives 14% duplicates; one of M = 200 million gives 49%. A high duplicate rate is a diagnostic of low library complexity — too little input DNA — not of over-sequencing, and the fix is upstream.
Amplification bias compounds exponentially. PCR efficiency depends on GC content and secondary structure, and the difference is applied every cycle. A 5% per-cycle efficiency gap over 10 cycles is 1.05¹⁰ ≈ 1.63× difference in final representation; a 10% gap over 12 cycles is 1.1¹² ≈ 3.1×. That is entirely enough to erase a real copy-number signal or manufacture a false one.
PCR-free libraries omit amplification: ligate adapters and sequence. The cost is input DNA — hundreds of nanograms to micrograms rather than nanograms. What you buy is flat across GC content (recovering GC-rich first exons and that PCR libraries drop out), essentially no duplicates, better accuracy, and depth that means what it says for copy-number analysis. PCR-free is the standard for population-scale WGS wherever sample input allows — the 3,202-sample high-coverage 1000 Genomes release was 30× short-read data of this kind.
A separate artefact worth naming: optical duplicates, where one cluster is read as two or a well is seeded twice. Physically distinct from PCR duplicates, and distinguishable because they are adjacent on the flow cell.
8. Targeted sequencing: exomes and panels
If you only care about part of the genome, spend your reads there.
Exome capture. Synthesise biotinylated baits complementary to annotated exons, hybridise them to the library, pull down with streptavidin beads, wash, amplify, sequence. Coding sequence is ~1–2% of the genome, so a fixed read budget buys ~50–100× the depth per target.
Amplicon panels. -amplify a defined set of targets with specific primer pairs. Enormous depth on small regions, fast, cheap, tolerant of degraded input.
| WGS | Exome capture | Amplicon panel | |
|---|---|---|---|
| Territory | Everything | ~1–2% (annotated coding) | kb to a few Mb |
| Depth per unit cost | Lowest | Middle | Highest |
| uniformity | Best (PCR-free) | Poor — capture efficiency varies with GC and bait design | Poor and structured — primer-dependent |
| CNV from read depth | Yes | Weakly, needs matched-batch | No — PCR normalises depth away |
| Yes | Essentially no | No | |
| Non-coding / regulatory variants | Yes | No | No |
| , mitochondrial genome | Yes | No | Only if targeted |
| New gene added to the list | Re-analyse in silico | New capture kit | New primers, revalidate |
| Signature failure mode | Cost, data volume, interpretation burden | Systematic dropout of GC-rich first exons | under a primer-site variant |
Two points that people learn the hard way.
For exomes, the mean depth is the wrong statistic. What matters is the fraction of target bases at ≥20×, because capture efficiency is systematically, reproducibly uneven. The exons that drop out are not random — GC-rich first exons of a number of clinically important genes are chronic offenders, and they drop out the same way in every sample, so no amount of batch-level QC them.
For amplicon panels, allele dropout is the dangerous one. A variant sitting under a primer binding site can prevent that amplifying at all. The result is a confident, high-depth, homozygous-reference call at a site where the patient is actually heterozygous for a pathogenic variant. Depth provides no protection — every one of those thousands of reads came from the other allele.
Statistics: why more reads shrink sampling error but never touch a systematic bias like this is S3 §7, which uses allele dropout as its worked example.
The long-run trend favours WGS even for coding questions, because cost per answer is what matters: one assay captures SNVs, , CNVs, SVs, repeat expansions, mitochondrial variants and non-coding regulatory sequence, and a new gene list is a re-analysis rather than a new kit.
9. Cost, and the Moore's-law comparison
The comparison is made constantly and is usually made badly, so state it precisely.
Moore's law is a roughly two-year doubling of transistor count — about 10× per 6–7 years. Cost per human genome tracked something like that until about 2007, then fell by roughly three orders of magnitude in about five years as second-generation instruments displaced capillary sequencing (NHGRI's series: ≈$7.1M in late 2007 to ≈$5,900 in 2012, a factor of ~1,200), then resumed a slower decline. The full four orders from that 2007 figure — down to ~$700 — was not reached until about 2019–2020, i.e. thirteen years rather than five. As of 2026 a 30× human genome at scale is somewhere in the low hundreds to about a thousand dollars depending on platform, volume and what you count. (Approximate teaching figures, 2026-08-10 — treat as an order of magnitude, not a quote.)
Three caveats that make the headline comparison mostly rhetorical:
- "Cost per genome" conventionally means reagents and instrument time. It excludes sample collection, consent, labour, storage, compute, analysis, clinical interpretation and reporting. For clinical work those dominate, and they have not fallen four orders of magnitude.
- Cost is per base, not per answer. A cheap platform that cannot resolve the repeat you care about has infinite cost for your question.
The binding moved. Generating sequence stopped being the hard part around
- Everything from Ch 44 onward exists because interpretation
became the bottleneck — and interpretation costs have, if anything, risen.
10. Choosing a platform: by question, not by preference
Do not ask "which platform is best". Ask what the data must physically support.
| Question | The property that decides it |
|---|---|
| Population-scale variation | Cost per base; uniformity of |
| Rare-disease diagnosis, negative on exome | Territory — non-coding, SVs, |
| Assembling a genome with no reference | Read length relative to the longest repeat (Ch 43) |
| Phasing variants onto | Read length relative to the distance between the variants |
| Methylation and sequence from one sample | Whether the platform reads the original, unamplified molecule |
| Isoform-resolved transcriptome | Whether a read spans the whole transcript (Ch 47) |
| Pathogen identification in the field | Instrument size, power, and time to first result |
| Subclonal somatic variant | Distinct input molecules, and error suppression, not depth |
The honest answer is frequently two platforms: long reads to establish structure, phasing and assembly; short reads for cheap depth on top. Hybrid was mandatory before long-read consensus accuracy improved, and it remains the default for anything where both structure and depth matter.
Common misconceptions
| What people think | What's actually true |
|---|---|
| A sequencer reads bases | It measures a physical signal — photons, migration time, ionic current — and a decoder infers bases. Sequencing errors are decoding errors, which is why a better decoder improves data you collected years ago |
| 30× means every base is read 30 times | It is a mean over a non-uniform, overdispersed distribution. The left tail is what determines whether you call a heterozygote |
| Long reads are less accurate than short reads | Raw single-pass accuracy is lower. Consensus accuracy — HiFi, or duplex nanopore — is comparable or better, and read placement accuracy was always better. The 2015 comparison is not the 2026 one |
| More depth always improves a call | Only if the extra reads are independent observations. of the same input molecule add nothing, and an error present in the original molecule is reproduced by every read of it |
| Sanger is obsolete | It is the orthogonal confirmation standard because it fails differently. A confirmation sharing failure modes with the thing it confirms is worthless |
| Q30 means one error per 1,000 bases in that read | Q is an estimated per-base error probability, and its calibration is a modelling claim that can itself be wrong. Headline platform Q figures are typically a mode or median over a run, not a guarantee |
| sequences all the exons | Capture efficiency is uneven and systematically so. Some clinically important GC-rich exons drop out in every sample, in the same way |
| duplicates should be deleted to save space | They should be marked. The duplicate rate is a library-complexity measurement, and deleting the reads destroys it |
| A high duplicate rate means you sequenced too deep | It means the library had too few distinct molecules. The fix is more input DNA or fewer PCR cycles, not fewer reads |
Worked example: designing and sanity-checking a run
A tumour/normal pair. Genome 3.1 Gb haploid, paired-end 2×150.
(a) Reads for 30× on the normal. Each pair contributes 300 sequenced bases.
That is 310 million pairs, 93 Gb of sequence per sample.
(b) What the Poisson model predicts. At C = 30, P(uncovered) = e⁻³⁰ = 9.4 × 10⁻¹⁴, so 0.0003 bases of the genome are expected to be untouched. The observed callable fraction will be about 95%. The 5% gap is entirely mappability, non-uniformity and filtering — §6's corrections 3 and 4. Sequencing to 60× would barely move it; a long-read platform or a reference would.
(c) Heterozygote detection. At local depth 30, P(all reads from one ) = 2 × 2⁻³⁰ = 1.9 × 10⁻⁹. Across ~3 million heterozygous sites that is 0.006 expected failures — negligible. At local depth 10 it is 2.0 × 10⁻³, giving ~5,900 sites seen as homozygous. The mean is fine; the tail is what you are paying 30× to control.
(d) Depth for the tumour. Target a subclonal heterozygous : subclone at 20% of cells, biopsy 50% tumour, so f = 0.5 × 0.5 × 0.2 = 0.05. Alt-supporting reads ~ Poisson(df). Require P(≥3 alt reads) ≥ 0.95:
(e) Why 3 reads is not a variant call. Take a per-base error rate of 0.1%, split across three possible wrong bases: ε ≈ 3.3 × 10⁻⁴ per specific alternate . At d = 130, λ = 130 × 3.3 × 10⁻⁴ = 0.043.
There are 3.1 × 10⁹ positions × 3 alternate alleles = 9.3 × 10⁹ chances:
A typical solid tumour carries on the order of 10³–10⁴ somatic SNVs genome-wide. So a naive "≥3 supporting reads" rule yields on the order of ten to a hundred false positives for every true one — about 40× in the middle of that range, and worst for the quiet tumours you would most like to characterise.
This is the entire justification for what Ch 46 does: a matched normal to subtract variation and shared artefacts, an explicit error model rather than a threshold, a panel of normals to catch recurrent site-specific artefacts, and filters on strand bias, position within the read and mapping quality. The instrument's error rate is not a nuisance parameter — at the depths cancer genomics requires, it is the term.
Connections
- Back to: Ch 02 — the 3'-OH whose absence stops a Sanger chain, and the base-pairing that makes template-directed synthesis possible · Ch 04 — polymerase is the engine of Sanger, Illumina and PacBio alike · Ch 36 — and electrophoresis, both used here · Ch 39 — the repeat content that sets the read length you need
- Forward to: Ch 41 — , Phred scores and where these error models are recorded · Ch 42 — placing reads, and why paired ends and read length change what is placeable · Ch 43 — the islands of §6, turned into a graph problem · Ch 45 — the 5% that depth cannot reach · Ch 46 — the Bayesian machinery §8 demands · Ch 47 and Ch 48 — the same chemistry pointed at RNA · Ch 49 — native modification calling · Ch 56 — where §6's somatic depth argument is cashed out · Ch 57