BRENNEN

Industries / Health / Pharmaceuticals / Biotech

Reading and writing DNA

This work covers reading the order of the letters in DNA (sequencing), making DNA to order from chemicals or enzymes (synthesis), and cutting DNA at a chosen spot (editing).

7

Milestones achieved, from 1977 to 2022.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1977

    Reading DNA with chain terminators

    Fred Sanger's group at Cambridge used altered building blocks called dideoxy nucleotides. When the copying enzyme adds one, the growing DNA strand stops at a known letter, so the sequence can be read from the lengths of the pieces. They tested it on the DNA of a virus called phiX174 and reported that it was faster and more accurate than their earlier "plus and minus" method.

    ebi.ac.uk

    Credit

    F. Sanger, S. Nicklen and A. R. Coulson, in PNAS

  2. 1981

    Machines that write DNA

    Serge Beaucage and Marvin Caruthers at the University of Colorado described stable chemical building blocks called deoxynucleoside phosphoramidites. Each one joins to a DNA chain held on a solid support in under a minute. Beaucage later wrote that this became "the chemistry that catapulted automated DNA synthesis into many laboratories". Caruthers helped found Applied Biosystems and entered the US National Inventors Hall of Fame in 2018.

    Lessonthe earlier method, Robert Letsinger's, used reagents that broke down in moist air, and that blocked automation.

    garfield.library.upenn.eduinvent.org

    Credit

    S. L. Beaucage and M. H. Caruthers

  3. 2001

    A first draft of the human genome

    An international public team published a draft human genome sequence in Nature in February 2001 and made the data free to all. NHGRI says the project ran from October 1990 to April 2003 and puts its cost at between USD 500 million and USD 1 billion.

    Lessonthe reference called finished in 2003 still left about 8% of the genome unsequenced, mostly long repeated stretches, by the count of the team that later closed the gaps.

    ebi.ac.ukgenome.govgenome.gov

    Credit

    International Human Genome Sequencing Consortium

  4. 2008

    Short reads by the billion

    Solexa and Illumina scientists published a whole human genome read in short pieces of 35 letters. The DNA was fixed to a glass surface and read with dyed nucleotides that stop the copying for one step at a time. Each letter was read more than 30 times on average, and the team found about 4 million single-letter differences. NHGRI estimates that a draft human genome cost about USD 14 million in 2006, just over USD 4,000 in mid-2015 and under USD 1,500 by late 2015.

    ebi.ac.ukgenome.gov

    Credit

    D. R. Bentley, S. Balasubramanian and colleagues (Illumina and Solexa)

  5. 2012

    Programmable DNA cutting

    Martin Jinek, Jennifer Doudna, Emmanuelle Charpentier and colleagues showed that a bacterial protein called Cas9 cuts both strands of DNA at a site chosen by a short guide RNA, and that the guide could be made as a single molecule. This is the basis of CRISPR genome editing. Doudna and Charpentier received the 2020 Nobel Prize in Chemistry.

    ebi.ac.uksnexplores.org

    Credit

    Jinek and colleagues, in Science

  6. 2018

    Ultra-long reads from a handheld device

    A nanopore sequencer reads DNA by measuring changes in electric current as a single strand threads through a tiny hole. A UK and US team used Oxford Nanopore's handheld MinION to read 91.2 billion letters from a human cell line, with single reads up to 882,000 letters long. That let them assemble the 4-million-letter MHC immune gene region in one piece and close gaps in the reference genome.

    ebi.ac.uk

    Credit

    M. Jain, S. Koren, K. H. Miga, M. Loose and colleagues

  7. 2022

    A complete human genome

    The Telomere-to-Telomere (T2T) Consortium published T2T-CHM13, a sequence of 3.055 billion letters. It has no gaps on any chromosome except Y and adds nearly 200 million letters that the earlier reference lacked.

    Lessonlong reads were needed to finish the repeated regions that short reads could not place.

    ebi.ac.uk

    Credit

    S. Nurk, A. M. Phillippy and the T2T Consortium, in Science, April 2022

Watch for

  1. Same-day genomes, new sequencer

    Roche launched AXELIOS 1, a research-only sequencer. It copies DNA into longer stand-in molecules called Xpandomers and reads them through nanopores. Roche claims whole-genome sequencing in a day, reads of about 1,500 letters and "better cost efficiency". The release gives no prices and no outside data.

    Achieved whenindependent labs publish speed, accuracy and cost results on standard reference samples.

    roche.com

    Claimed by

    Roche, June 2026

  2. 50,000 letters made by enzymes

    Ansa Biotechnologies sells DNA up to 50,000 letters long, made by enzymes rather than harsh chemicals, and calls it "the longest sequence-perfect synthetic DNA commercially available". It quotes prices from USD 0.28 per letter pair and delivery in about 25 days, and says chemical synthesis "rarely produces sequences beyond 10 kb" (10,000 letters). Each product is checked only by Ansa's own sequencing.

    Achieved whencustomers or independent labs publish sequencing results confirming error-free 50,000-letter products at scale.

    drugdiscoverynews.com

    Claimed by

    Ansa Biotechnologies, launch release (month unconfirmed; see Checks)

  3. A USD 60 human genome

    Ultima Genomics' chief executive, Gilad Almogy, says new consumables called Solaris 2.0 Max bring the cost of a human genome read 30 times over to USD 60. The report notes that the release "does not define the costs included".

    Achieved whenan outside lab or core facility publishes a fully costed price near that figure and says what it includes.

    runtimewire.com

    Claimed by

    Ultima Genomics, October 2026

Open

  1. 01

    A rolling test for DNA models

    A public leaderboard for DNA language models, AI models trained on raw DNA letters, such as Arc Institute's Evo 2, GPN-MSA and Nucleotide Transformer. Each model is scored only on variants from ClinVar, the US public archive of DNA changes labelled harmful or harmless, whose labels first appeared after that model's training data was collected. It is rebuilt with each new ClinVar release.

    The brief
    Credit · unclaimed
  2. 02

    A dated sequencing-cost dataset

    An open spreadsheet that picks up NHGRI's cost-per-genome series where it stopped. It records dated costs per billion letters and per human genome read 30 times, for each sequencing platform, with maker list prices and the prices university core facilities publish kept apart, and a citation on every row.

    The brief
    Credit · unclaimed

Open 01 · the brief

A rolling test for DNA models

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public leaderboard for DNA language models, AI models trained on raw DNA letters, such as Arc Institute's Evo 2, GPN-MSA and Nucleotide Transformer. Each model is scored only on variants from ClinVar, the US public archive of DNA changes labelled harmful or harmless, whose labels first appeared after that model's training data was collected. It is rebuilt with each new ClinVar release.

Why it is still open

The largest recent benchmark found, the Brandes Lab's VEP-eval, scores 242,132 ClinVar variants but is fixed to the February 2026 release, with no split by date and no updates. A time-split ClinVar set does exist, built by Manuel Corpas (6,929 variants, cut off on 29 November 2025), but it tests chat models such as Claude Sonnet 4.5 and GPT-5.2, not DNA models.

github.comzenodo.org

Built with what exists

Monthly ClinVar releases from NCBI; open model weights (Evo 2 is on Hugging Face in sizes from 1 billion to 40 billion parameters, with up to 1 million letters of context); VEP-eval's scoring notebooks; one rented cloud GPU; Python. Open DNA models with million-letter context are the advance that lets a small team score many models across the genome.

cdn.jsdelivr.net

  1. First step

    Compare two ClinVar releases six months apart, list the variants with two or more review stars that are new in the later one, and score them with VEP-eval's Evo 2 notebook.

  2. Done looks like

    A public page and file listing at least 5 DNA models, each scored by AUROC (a standard accuracy score from 0.5, chance, to 1, perfect) only on variants first labelled after its stated data cutoff, with the variant list, ClinVar release and code published, and refreshed from at least two later ClinVar releases.

Open 02 · the brief

A dated sequencing-cost dataset

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open spreadsheet that picks up NHGRI's cost-per-genome series where it stopped. It records dated costs per billion letters and per human genome read 30 times, for each sequencing platform, with maker list prices and the prices university core facilities publish kept apart, and a citation on every row.

Why it is still open

NHGRI's cost page was last updated on 16 May 2023, its data table is labelled May 2022, and it does not split costs by platform. Mathurin Dorel's web tracker lists many platforms but uses list prices only, shows no update date or download, and has incomplete rows for the newest machines.

genome.govmathurind.github.io

Built with what exists

Public maker price pages and core facility price sheets; a CSV file in a public Git repository; a short Python script that recomputes cost per genome the way NHGRI did (a 3,000-million-letter genome at a stated coverage).

  1. First step

    Write down NHGRI's stated method, then collect current price sheets from five core facilities on different continents for Illumina NovaSeq X, MGI DNBSEQ-T7, Element AVITI and Ultima UG 100.

  2. Done looks like

    A CSV covering at least 15 platforms, every row with a date, a source link, whether the price is a list or core-facility price, and what it includes, with two snapshots at least three months apart.