A blog post titled "How I sequenced my genome at home" by Seth Showes spent most of this week on the Hacker News front page. The claim, delivered with receipts, is that he ran his own whole-genome sequencing in roughly 72 hours on a $3,200 Oxford Nanopore MinION at his kitchen table, and that Claude wrote the gene-targeting panel he used to get clinically-relevant depth on the regions that mattered to him. The part of the story worth slowing down on is the third sentence, because the AI's role is narrow, specific, and correct in a way that generalises.

How It Actually Works

Four steps stack into a home genome run, and each has a piece that has only become tractable for an amateur in the last two years.

The Sequencer

The Oxford Nanopore MinION is a pocket-sized device - about the size of a large USB stick - that reads DNA by threading single strands through protein pores embedded in a flow cell and measuring the tiny electrical disruptions as each nucleotide passes. Those disruptions are converted to A/C/G/T calls by a separate piece of software. The device retails for $3,200. A single-use flow cell is another $900 on top, the library-prep reagents (chemistry that attaches adapters to your DNA so it can thread through the pores) are $610 for six reactions, and incidentals - pipette tips, wash buffers, ethanol - add roughly $50 per run. Showes itemises $1,100 all-in per run. The first run costs $4,300 because the hardware and bulk reagents are paid upfront.

The wet-lab portion - extracting DNA, preparing a library, loading the flow cell - takes Showes around four hours. The software portion runs overnight. The single most common failure mode, he writes, is "air in the flow cell." Source: unsplash.com

Adaptive Sampling

Nanopore's distinguishing trick is that the device can decide which reads to keep and which to reject in real time. As a DNA strand starts to thread through a pore, the basecaller checks the first few hundred bases against a reference panel. If those bases map to a region you care about, sequencing continues. If they don't, the pore voltage reverses, the fragment is ejected, and the pore is freed for the next molecule. This is called adaptive sampling, and it is the feature that makes targeted home sequencing practical - instead of reading every fragment in the sample and getting shallow coverage everywhere, you concentrate the device's capacity on the genes that matter.

What adaptive sampling requires is a BED file - a plain-text list of chromosome, start-position, end-position tuples defining the regions of interest. This is where Claude comes in.

Claude Writes the Panel

Showes describes using Claude to identify which genes are relevant to a given clinical question, then to generate the BED file with GRCh38 coordinates (GRCh38 is the current human reference genome assembly) with a 100kb padding in each direction to capture regulatory regions, with overlapping ranges merged, and with a validation check to keep the total target size under 5% of the genome (the cap needed for enough coverage on a single flow cell). He calls this "one of the best uses of an LLM in a project like this," noting that the knowledge required is spread across four separate databases (UCSC Genome Browser, Ensembl, OMIM, CPIC) and is tedious rather than difficult.

"If you care about a specific set of genes - say, pharmacogenes or the genes behind a family autoimmune condition - this is the path."

The output is a file of 50 to 200 lines, each line a chromosome and two numbers. Claude doesn't do the sequencing, doesn't interpret the results, and doesn't replace a geneticist. It does the one part of the pipeline that's a pure knowledge-aggregation task, and does it in minutes rather than the hours an unfamiliar user would spend navigating four databases.

Basecalling and Alignment

The MinION produces ~49GB of raw electrical-signal data per run in a format called pod5. Oxford Nanopore's basecaller, Dorado, converts that signal into sequence reads and writes a BAM file (a compressed, indexed format for aligned reads) of around 6GB. The reads are aligned to GRCh38 with minimap2, a widely-used open-source aligner. At this point the user has a whole-genome BAM with 10x average coverage - one-third of what a clinical lab would call adequate - plus a targeted-panel BAM with 30-50x coverage on the genes of interest, which is clinical-grade for most variant-calling work.

The Concrete Example

Showes' motivation isn't abstract. His sister received a liver transplant due to complications of an autoimmune condition that runs in the family. He wanted to know which of the relevant susceptibility loci he carries. The panel he built with Claude covered the HLA region (the main immune-signaling gene cluster, densely implicated in autoimmune disease), the pharmacogenes that affect how immunosuppressants are metabolised, and the specific loci associated with his sister's diagnosis.

The Claude-generated BED file in Showes' pipeline covered the HLA region and pharmacogenes relevant to his family's autoimmune history. The photo is from the Cancer Genomics Research Laboratory, not Showes' kitchen - but the workflow is the same at both scales, which is the story. Source: unsplash.com

The output of his run is an aligned BAM file, not a diagnosis. To go from aligned reads to clinically-meaningful variants requires a separate pass through variant-calling software and - critically - interpretation by someone who can contextualise the variants against current medical literature. Showes explicitly flags this:

"Sequencing your own genome on your kitchen table does not constitute a clinical diagnostic test. You absolutely should not make health decisions based on what comes out."

Home Sequencing vs Clinical Sequencing

The two columns aren't substitutable. Clinical labs do a slower, more expensive, more-validated version of the same fundamental measurement, with accreditation chains that exist specifically because the error rate matters when the output drives medical decisions. The home version is cheaper, faster, and good enough to satisfy scientific curiosity. The distinction is not a technical one. It's a credentialing one.

Why It Matters Now

Three reasons this shows up in April 2026 and not April 2022.

The MinION Mk1D and R10.4.1 flow cells landed in 2024 and dropped the raw-accuracy error rate to under 1%. Before that point, the nanopore error rate was high enough that a home pipeline produced data too noisy to be worth interpreting. The hardware generation now in Showes' kitchen is the first one where the answer to "is this data any good" is yes for a hobbyist.

Apple's M3 Ultra Mac Studio and the rise of local-inference AI tooling brought the compute requirement down to consumer gear. The 100GB storage and roughly 128GB of RAM required to basecall a single run is expensive by laptop standards, but it's one machine in the corner of a bedroom, not a rack in a datacentre. Oxford Nanopore published a reference configuration for running the full pipeline on a single Mac in early 2026, and Showes' work is the first widely-circulated demonstration of it running.

Claude is now good enough at narrow bioinformatics tasks that the knowledge-aggregation step is no longer a bottleneck. This is the quiet part. Five years ago, "generate a BED file covering the HLA region plus known pharmacogenes with 100kb padding" required a graduate-level literature search and maybe a day of work cross-referencing databases. Today it's a single prompt to an LLM that has read those databases during training. Anthropic's Claude for Life Sciences positioning is the commercial framing of the same capability.

The combined effect is that the limiting resource for home sequencing is no longer equipment cost or expert knowledge. It is regulatory posture. A kitchen genome isn't a clinical diagnosis, and the person who confuses the two has a small but non-zero chance of making a consequential medical decision on data that hasn't been validated for that purpose. The EA Forum has been cataloguing the same question in more cautious language for most of 2025, and Showes' post is the trigger that'll force the question back into mainstream circulation.

What To Read Next

• Claude for Life Sciences: Anthropic's commercial positioning
• Oxford Nanopore's distributed sequencing thesis
• Our /ultrareview coverage for another example of an AI workflow that's narrow, correct, and priced as a premium feature

The story is about one person who knew what he was doing, used a tool appropriately, and wrote it up carefully. The worry sitting underneath is that the narrow-and-correct use case Showes shows is indistinguishable, at the blog-post level, from the broad-and-wrong use case someone will try next.

Sources:

• How I sequenced my genome at home - Seth Showes, iwantosequencemygenomeathome.com
• Hacker News discussion - front page, 21 April 2026
• Biohacker claims to have sequenced their own genome at the kitchen table - Tom's Hardware, 21 April 2026
• Oxford Nanopore MinION product page - Oxford Nanopore Technologies
• Oxford Nanopore meets Apple's M3 silicon - Oxford Nanopore blog
• Claude for Life Sciences - Anthropic
• DIY Biology: A risk in the making? - EA Forum