How Illumina Sequencing Works: From DNA Fragment to Finished Genome

Illumina's sequencing chemistry has turned up in over 369,000 published studies, and if a genome, exome, or transcriptome got sequenced anywhere in the world this week, there's a decent chance it happened on one of these machines. That's not because the underlying idea is exotic. It's built from things most life science students already know: PCR, complementary base pairing, and fluorescence. Illumina's contribution was figuring out how to run millions of copies of that idea in parallel, on a single glass slide, cycle after cycle, photographing the result each time.

That approach is called sequencing by synthesis (SBS). It comes down to four repeatable stages: turning DNA into a sequencing-ready library, copying each fragment into a dense clonal cluster, reading each cluster one base at a time using reversible-terminator chemistry, and converting the resulting images into readable sequence data.

A Short History

The core chemistry didn't originate at Illumina. It came out of Cambridge in the late 1990s, developed by Shankar Balasubramanian and David Klenerman and commercialized through a company called Solexa. Illumina acquired Solexa in 2007 and released the first Genome Analyzer not long after.

At the time, the standard method was Sanger sequencing: reading one DNA fragment at a time, producing long and highly accurate reads, but slowly and at a high cost per base. Illumina's platform traded some of that per-read length and accuracy for scale, reading hundreds of millions to billions of short fragments at once instead of one long fragment. That trade-off turned out to fit most real questions in genomics better than anyone expected, and short-read sequencing by synthesis has been the dominant next-generation sequencing (NGS) platform ever since, even though it's no longer the only serious option (more on that in the limitations section below).

Step 1: Building the Sequencing Library

Before anything gets sequenced, DNA (or RNA, reverse-transcribed into cDNA) has to become a "library": a pool of fragments, each one carrying the right adapter sequences on both ends so the flow cell can grab onto it later.

The steps, in order:

Fragmentation. Genomic DNA is sheared into pieces, usually somewhere between 200 and 800 base pairs depending on the application, either mechanically (sonication) or enzymatically. Illumina's Nextera-based kits use a faster method called tagmentation, which fragments the DNA and attaches adapters in a single enzymatic step using a transposase.
End repair and A-tailing. Fragment ends are made blunt, and a single adenine base is added to each 3' end.
Adapter ligation. Illumina adapters, carrying a matching single thymine overhang, are ligated onto both ends of every fragment. These adapters do two jobs: they let the fragment bind to the flow cell, and they carry the primer-binding sites the sequencer will use later.
Indexing. A short, sample-specific index sequence gets added during PCR. This is what makes multiplexing possible: dozens to hundreds of samples can be pooled into one run and sorted back out computationally afterward, based on which index each read carries.
Amplification and cleanup. The library is PCR-amplified (or, in PCR-free protocols, left unamplified to avoid PCR bias), then size-selected and quantified before it's loaded onto the instrument.

Step 2: Cluster Generation (Bridge Amplification)

A single DNA molecule doesn't give off enough fluorescent signal for a camera to pick up, so before any base gets read, each library fragment has to be copied into a tight cluster of identical copies, all anchored to the same spot.

This happens on the flow cell, a glass slide whose surface is coated with a dense lawn of two types of short oligonucleotides, complementary to the two adapter sequences on the library fragments. When the library flows across the surface, each fragment's adapter binds to a matching oligo.

From there, cluster generation runs through a repeating cycle:

The bound fragment is extended by a polymerase, copying it into a new strand tethered to the surface.
The original template washes away, leaving a single strand anchored at one end.
That strand bends over, and its free end anneals to a nearby, complementary surface oligo, forming a bridge.
A polymerase extends across the bridge, creating a double-stranded arch.
The bridge is denatured, leaving two single strands anchored close together.

Step 3: Reading the Bases

A sequencing primer anneals next to each cluster, at the known adapter sequence, and the flow cell floods with DNA polymerase and all four nucleotides at once, each one carrying two modifications: a fluorescent tag (a different color per base) and a reversible terminator group blocking further extension. This step is what gives sequencing by synthesis its name.

Because all four labeled bases are present together in every cycle, they compete for incorporation the way they would during natural replication, which is part of why the raw accuracy is high. There's no systematic bias toward whichever base happens to get added first.

Each cycle runs through four steps:

Incorporation. At every cluster, the polymerase adds exactly one labeled, terminated base, matching the template. The terminator stops the reaction there, so only one base gets added per round.
Wash. Unincorporated nucleotides are flushed off the flow cell.
Imaging. Lasers excite the fluorescent tags, and a camera photographs the entire flow cell, recording which color lit up at each cluster's position.
Cleavage. The fluorescent tag and the terminator group are chemically cleaved off, freeing the strand's 3' end for the next round.

That cycle repeats once per base, for as many cycles as the run calls for, commonly somewhere between 50 and 300. Every cluster on the flow cell gets read in parallel during every cycle. The sequencer isn't working through fragments one at a time; it's rephotographing millions of them simultaneously, one position further along each time.

Most runs today are paired-end: once the first read finishes, the newly built strands strip away, the original strand bridge-amplifies again in place, and a second read runs from the opposite end of the same fragment. Paired-end data makes it considerably easier to align short reads to a reference genome accurately, and to catch insertions and deletions that a single-direction read might miss.

How the Chemistry Has Evolved

The version above, four distinct fluorescent colors, one per base, all imaged separately, is the original four-channel approach, and some instruments still use it. Illumina has since introduced faster and cheaper variants:

Two-channel chemistry (MiniSeq, MiSeq, NextSeq 500/550, HiSeq X): uses two dyes instead of four. One base shows up red, one green, one in both colors, one in neither. Fewer images per cycle means faster, cheaper runs.
One-channel chemistry (iSeq 100): pushes this further, using a single dye and distinguishing bases by intensity pattern instead of color, which lets the whole optical system shrink down to a CMOS chip similar to what's in a smartphone camera.
XLEAP-SBS chemistry (NextSeq 1000/2000, NovaSeq X series): the newest generation, built around re-engineered polymerases and new dye chemistry rather than just optical shortcuts. It cuts cycle times, pushes typical base-call quality from the older Q30 standard toward Q40, and is a major reason the NovaSeq X Plus reaches the throughput it does.

Step 4: From Light to Letters

Raw output from a sequencer isn't DNA sequence. It's an enormous stack of images, and turning that into usable data takes several more steps:

Base calling. For every cluster, at every cycle, the recorded color and intensity get converted into a base call (A, T, C, or G) plus a quality score, written out per read as FASTQ files, the format nearly every downstream tool expects. Illumina reports quality on the Phred scale: a Q30 score means roughly a 1-in-1,000 chance that base is wrong (99.9% accuracy), and Q40 means roughly a 1-in-10,000 chance (99.99% accuracy).
Demultiplexing. Reads get sorted back into their original samples, based on the index sequences added during library prep.
Alignment. Each short read is mapped to its most likely position in a reference genome, using aligners like BWA or Bowtie2, or assembled from scratch by finding overlaps between reads when there's no reference to align against.
Downstream analysis. From there, the pipeline depends on the question being asked: variant calling for genetic disease, differential expression for RNA-seq, taxonomic classification for metagenomics, or STR and SNP genotyping for a forensic panel.

Illumina bundles much of this into a platform called DRAGEN, which runs on dedicated hardware or in the cloud to handle alignment, variant calling, and compression fast enough to keep up with a NovaSeq X Plus's output instead of becoming the bottleneck itself.

The Instrument Lineup

Illumina sells the same underlying SBS chemistry across a range of instruments sized for very different sample volumes. As of 2026, the current lineup looks roughly like this:

Platform	Throughput tier	Example output	Max read length	Typically used for
iSeq 100	Entry-level	1.2 Gb, up to 4M reads, ~17.5 hr run	2×150 bp	Small genomes, teaching labs, quick viral/microbial checks
MiniSeq	Entry-level	Similar tier to iSeq, slightly higher	2×150 bp	Small targeted panels, low-volume labs
MiSeq / MiSeq i100 Series	Benchtop	Up to 15 Gb, ~25M reads (classic MiSeq)	2×300 bp (Illumina's longest standard read)	16S/amplicon microbiome work, small genomes, clinical microbiology, forensic STR/SNP panels via MiSeq FGx
NextSeq 550	Benchtop	Moderate output; doubles as an array scanner	2×150 bp	Exomes, transcriptomes, methylation arrays
NextSeq 1000/2000	Benchtop, higher	~90–120 Gb for a 2×150 bp run, scaling higher on larger flow cells	Up to 2×300 bp	Exomes, transcriptomes, single-cell work, moderate WGS
NovaSeq 6000	Production-scale	~6 Tb, ~20 billion reads per run	2×250 bp	Large cohort studies, population genomics, consortium projects
NovaSeq X / X Plus	Ultra-high-throughput	8–10.5 Tb (X) or 16–21 Tb (X Plus), 26–35 billion reads per flow cell	2×300 bp	Biobank-scale sequencing, 20,000+ genomes a year

A few things the table doesn't capture: Illumina has fully retired its HiSeq line (HiSeq 1000 through HiSeq X), with the NovaSeq series covering that role now. The classic MiSeq is being phased out in favor of the newer MiSeq i100 Series, though Illumina has committed to supporting existing MiSeq instruments through the end of 2029. Several platforms also have dedicated in vitro diagnostic (Dx) versions, MiSeqDx, NextSeq 550Dx, NovaSeq 6000Dx, cleared for clinical rather than research use.

What It's Used For

A short list covers most of what actually runs on these machines:

Whole genome and exome sequencing. Population-scale reference datasets and clinical labs hunting for the genetic cause of a rare disease both run on the same core workflow.
RNA sequencing. Measuring which genes are active, how much, and which splice variants show up, from tumor biopsies to plant stress studies.
Targeted gene panels. Hereditary cancer panels, pharmacogenomic panels, and other disease-specific panels sequenced at much higher depth and lower cost than a whole genome would allow.
Single-cell sequencing. Profiling gene expression cell by cell rather than averaging across an entire tissue sample.
Metagenomics and microbiome studies. 16S rRNA amplicon sequencing, long the MiSeq's signature application, plus shotgun metagenomics for characterizing whole microbial communities.
Non-invasive prenatal testing. Screening cell-free fetal DNA in maternal blood for chromosomal abnormalities without an invasive procedure.
Oncology and liquid biopsy. Detecting circulating tumor DNA, profiling tumor mutations to guide treatment, monitoring for relapse.
Infectious disease surveillance. Whole-genome sequencing of pathogens to track outbreaks and emerging variants, the kind of work that scaled up enormously during global COVID-19 genomic surveillance.

Forensic genomics is worth walking through separately, because what MPS adds there isn't just more throughput, it's information that length-based methods physically can't see.

Standard forensic DNA profiling relies on capillary electrophoresis (CE): PCR-amplifying STR markers and separating the products by size on a genetic analyzer. CE tells you how long each STR allele is, but nothing about what its internal repeat sequence looks like, and two alleles of identical length can carry different internal sequences (an isoallele) that CE simply can't tell apart.

Illumina built a dedicated platform for closing that gap: the MiSeq FGx, launched in 2014, paired with the ForenSeq DNA Signature Prep Kit. In a single reaction, ForenSeq amplifies up to 231 forensically relevant markers, 27 autosomal STRs, 7 X-STRs, 24 Y-STRs, and close to 100 identity-informative SNPs, with optional panels for ancestry and externally visible phenotype prediction layered on top. Because MPS reads the actual sequence of each repeat instead of just its length, it resolves isoalleles that CE would score as identical, and it hands mixture-interpretation software more information to work with when separating overlapping contributor profiles than length alone ever could. Illumina spun this business out into a separate company, Verogen, in 2017; QIAGEN completed a full acquisition of Verogen in January 2023 and continues developing forensic MPS products built on Illumina's underlying SBS chemistry.

Strengths

Accuracy. Raw per-base accuracy above 99.9% (Q30) is standard, pushing toward Q40 with XLEAP-SBS chemistry, and the four-way base competition built into every SBS cycle limits the kind of systematic bias that sequential-addition methods are prone to.
Cost and throughput at scale. A NovaSeq X Plus run can sequence a human genome to research-grade coverage for around $200 in reagents, at a capacity of more than 20,000 genomes a year.
Scalability. The same core chemistry runs on instruments spanning three orders of magnitude in output, from a 1.2 Gb desktop box to a multi-terabyte production system, so a lab can match its platform to its actual sample volume instead of over- or under-buying capacity.
A mature ecosystem. Nearly two decades of use have built an enormous surrounding toolkit: aligners, variant callers, QC software, and reference databases that all assume Illumina-style short paired-end reads as a baseline input.
Multiplexing. Dual-indexing lets a single run carry samples from dozens to hundreds of different people or experiments, sorted back out computationally afterward.

Limitations

Short reads. Even the longest standard Illumina reads (2×300 bp on the MiSeq) are short next to many repetitive regions, structural variants, and full-length transcript isoforms. A 150 bp read that lands inside a long repeat often can't be placed on the genome with any confidence.
No native modification detection. Standard SBS only reports base identity, not whether a cytosine is methylated or otherwise chemically modified. Detecting methylation needs an extra step, usually bisulfite conversion, added before library prep, which takes time and introduces its own biases.
PCR-related bias. Most library prep protocols include a PCR step, which can under- or over-represent GC-rich or GC-poor regions and introduce duplicate reads that have to be flagged and removed computationally.
Turnaround time. Sample to result usually takes a day or more once library prep, cluster generation, sequencing, and analysis are all counted. That's fine for research and most clinical work, but too slow for point-of-care diagnostics or fast-moving outbreak response.
Limited structural context. Because Illumina reads short, independent fragments, phasing (working out which variants sit on the same physical chromosome copy) and large structural rearrangements are often invisible or ambiguous without additional long-range data.

These limitations are exactly what pushed the field toward long-read platforms like Oxford Nanopore and PacBio, which read tens of thousands of bases from a single native DNA molecule (in Nanopore's case, in real time as the molecule passes through a nanopore) and can call base modifications directly, no bisulfite conversion required. Long reads solve the repeat-spanning and phasing problems Illumina struggles with, historically at a higher raw per-base error rate, though that gap has been narrowing. In practice, plenty of labs run both: short-read Illumina data for cheap, accurate, high-depth variant calling, with long-read data layered on top wherever the structural questions short reads can't answer.

Where It's Headed

The clearest trend right now is cost and scale. Illumina has marketed the NovaSeq X Plus as enabling a "$200 genome," and XLEAP-SBS chemistry is the main lever behind that: faster cycles, higher quality scores, and reagents stable enough to ship at room temperature instead of needing a cold chain.

The other trend worth watching is convergence with long-read sequencing rather than competition against it. Illumina's Complete Long Reads kits, introduced alongside the NovaSeq X series, generate long-range genomic information using Illumina's own short-read chemistry rather than requiring a separate instrument, aimed at closing some of the phasing and structural-variant gap without a second platform in the lab. Expect more hybrid approaches along these lines: short-read accuracy paired with long-read context, increasingly on a single machine.

And clinically, the direction is from research tool toward routine test. Dx-cleared versions of the MiSeq, NextSeq 550, and NovaSeq 6000 already exist for diagnostic use, and as costs keep falling, sequencing keeps moving from something a research lab does occasionally to something ordered as part of standard patient care, starting in oncology and rare disease and expanding from there.

The Bottom Line

Illumina sequencing didn't win out because it was the most elegant way to read DNA. Sanger sequencing arguably reads it more directly, and long-read platforms now read longer, more natural molecules without ever breaking them apart. It won because reading a billion short fragments in parallel, cheaply enough and accurately enough, turned out to answer most of the questions biology and medicine actually ask. Whether that stays true depends on how fast the alternatives close the cost and throughput gap. For now, though, if a genome got sequenced somewhere in the world this week, there's a good chance it went through one of these machines.