When a letter is added to a growing strand of DNA, the reaction releases a pyrophosphate. A short enzymatic chain turns that into a pulse of light. Catch the pulse, and you know a letter went on. That was 454, and it read longer stretches than its contemporaries, which mattered a great deal for assembling genomes from scratch. Roche announced it was shutting the platform down in October 2013 and serviced it until mid-2016.
SOLiD, from Applied Biosystems, used ligation rather than synthesis. It never measured a single letter at all. Each step reported a colour standing for a pair of letters, and consecutive colours overlapped by one position, so every letter ended up pinned by two overlapping pair measurements instead of one reading of its own.
Ion Torrent, from 2010, threw out the optics altogether. A letter goes on, a hydrogen ion is released, the acidity in the well shifts very slightly, and a transistor underneath reads the shift. No lasers, no cameras, no dyes, and natural nucleotides throughout. It sequenced DNA on something close to an ordinary semiconductor chip, which is why it was cheap.
Two more broke a habit all of them shared, which was to copy the DNA first and read the copies. Pacific Biosciences shines light into a hole about a hundred nanometres across, small enough to watch one enzyme add letters to one strand as it happens. Oxford Nanopore does something stranger. It threads a single strand through a protein pore and measures the change in electrical current as the strand passes.
Notice what none of them do.
Not one of them reads a letter. Each one measures something a letter causes, and works backwards. A photon. A shift in acidity. A brightness. A current.
You can see the seam if you look for it. Because 454 and Ion Torrent both encode a run of identical letters in the size of a single signal rather than in a count of separate events, eight identical letters in a row arrive as one larger event. The signal does scale with the length of the run, but only roughly, and the approximation gets worse the longer the run. Past six or eight, turning that analogue value back into a whole number is close to a coin flip. Two different physics, the same artefact, for the same reason.
The nanopore is the furthest thing from reading of any of them, and it is worth being precise about why, because the vendor language is not.
The current at any instant is not set by one letter squeezing past. It is set by several letters sitting inside the pore together, about five in the pores of a few years ago and around nine in current chemistry, with the ones nearest the middle counting for most. What comes out of the machine is a trace. The sequence is inferred from that trace by a neural network that was trained rather than derived.
Which has a consequence worth sitting with. Put the same raw signal through a newer basecaller and it can spell something slightly different. Nanopore accuracy has improved across a decade of software releases without anybody touching the chemistry, because what improved was the model doing the guessing.
Even the platform that comes closest to watching one letter at a time does not get there on physics alone. Pacific Biosciences reads the same molecule round and round and averages the passes, and that averaging is where the accuracy comes from.
The instrument is not spelling. It is measuring, and then arithmetic finishes the job.
The sequence printed on a report was never observed by anything. It was worked out, and on the newest machines it can be worked out again.
Platform mechanisms and the homopolymer limit are from Goodwin, McPherson and McCombie, "Coming of age: ten years of next-generation sequencing technologies", Nature Reviews Genetics 17, 333 to 351, 2016, with the primary descriptions in Margulies et al., Nature 437, 2005 for 454; Rothberg et al., Nature 475, 2011 for Ion Torrent; and Eid et al., Science 323, 2009 for single-molecule real-time sequencing. The homopolymer statistics are from Bragg et al., PLoS Computational Biology 9:4, 2013. The multi-base dependence of the nanopore signal and the basecaller history are from Rang, Kloosterman and de Ridder, Genome Biology 19:90, 2018, and Wang et al., Nature Biotechnology 39, 2021.