Woolf demonstrates through his own weekend project that the entire Human Genome Project workflow — which took 13 years and $2.7 billion — now fits on a desk for roughly $1,000. He walks through the stack (Zymo collection, Qiagen extraction, MinION flow cell, dorado basecalling, minimap2 alignment, clair3 variant calling) to show that every step is either a consumer kit or open-source software runnable on a single GPU.
By submitting the piece to HN where it drew 114 points and 34 comments, bilsbie surfaced the argument that this is a genuine inflection point worth developer attention — the sequencing pipeline is now something a curious engineer can run at home, not a service bureau capability.
Woolf frames the shift as a strategic story: Illumina defended short-read sequencing as a service-bureau business for two decades, while Oxford Nanopore attacked from the opposite direction with small USB-powered devices, open-source basecallers, and a bet that software improvements would close the accuracy gap. The MinION being stapler-sized and plugging into a laptop is presented as evidence that the architectural bet has paid off.
Woolf is candid that his 5-15x coverage falls well short of the 30x needed for clinical-grade variant calling, and that ClinVar returns 'variants of uncertain significance' for many findings. The experiment validates that hobbyist sequencing works for pharmacogenomic markers and ancestry cross-checks against a 23andMe SNP array, but he stops short of claiming diagnostic value.
Bradley Woolf sequenced his own DNA at home. Not spit-in-a-tube, mail-it-to-23andMe DNA — actual reads off actual nanopores, on his own hardware, with his own pipeline. The writeup walks through the wet lab (a Zymo blood collection kit, a Qiagen extraction, library prep with Oxford Nanopore's ligation kit), the sequencing (a MinION flow cell running off a laptop over USB-C), and the compute (basecalling on a desktop GPU, alignment against GRCh38, variant calling with standard open-source tools).
The pitch is unusually blunt: everything a research lab did for the Human Genome Project across 13 years and $2.7 billion now fits on a desk, costs roughly $1,000 end-to-end, and produces a usable draft genome in a weekend. The MinION itself is about the size of a stapler and plugs into a USB port. Flow cells are consumable — you get one or two runs per cell, then throw it out. The rest of the stack is software: Oxford Nanopore's `dorado` basecaller (open source, PyTorch under the hood), `minimap2` for alignment, `clair3` or `deepvariant` for variant calling, all runnable on a single consumer GPU.
Woolf's coverage was modest — somewhere in the 5-15x range depending on region, versus the 30x you'd want for clinical-grade calls — but enough to identify most common variants, run through ClinVar, and cross-check against a prior 23andMe SNP array. He found the expected pharmacogenomic markers, the expected ancestry signal, and a handful of variants of uncertain significance that ClinVar shrugs at. In other words: the experiment worked.
The interesting part isn't the biology. It's that the sequencing stack has quietly become a software problem, and the software is all open source. Illumina spent two decades defending short-read sequencing as a capital-intensive, service-bureau business. Oxford Nanopore attacked from the opposite direction — long reads, small devices, terrible raw accuracy in the early days, and a bet that basecalling accuracy would improve faster on the model side than Illumina's chemistry could shrink on the machine side. That bet is now paying off. Recent `dorado` models are hitting Q20+ (99%+) single-read accuracy on R10.4 flow cells, which is the range where consensus calls start being clinically meaningful.
What this exposes is a familiar pattern to anyone who watched GPUs eat rendering, or transformers eat NLP: when the hard problem migrates from physics to inference, the incumbent's moat evaporates on a Moore's-law-adjacent schedule. The MinION isn't cheap because it's simple — it's cheap because the accuracy budget got moved to a neural net that runs on hardware you already own. The flow cell is a $500 disposable. The intelligence lives in a checkpoint file.
The practitioner-adjacent story here is the pipeline itself. A modern nanopore workflow looks like this: raw signal (POD5 files, a columnar format that replaced the old FAST5 HDF5 mess) → basecaller (a transformer trained on labeled signal data) → aligner (minimap2, still the reference implementation for long reads) → variant caller (clair3, another neural net) → annotation (VEP, SnpEff). Every step is a CLI tool with a Python or Rust core. There's no proprietary format you can't parse, no license server, no vendor sales rep. If you can operate a Kubernetes cluster, you can operate a genome pipeline.
The HN thread on Woolf's post lit up with two flavors of reaction. One camp: this is table stakes for personal medicine and everyone should be doing it. The other: coverage matters, false positives are dangerous, and reading ClinVar entries without a genetic counselor is a great way to spiral about a variant that turns out to mean nothing. Both are right. The MinION gives you the reads; it doesn't give you the epistemics.
If you're a developer who's ever thought "bioinformatics is somebody else's problem," that framing is now out of date. The pipelines are Python and Rust. The models are transformers. The data formats are columnar. The bottlenecks are I/O, GPU memory, and the same batch-vs-stream tradeoffs you already argue about at work. Nextflow and Snakemake are the workflow orchestrators; if you've used Airflow or Dagster, you'll recognize the shape immediately. Nf-core maintains a catalog of production-grade pipelines you can `git clone` and run against your own reads.
There's a concrete opportunity in the tooling gap. The wet-lab side is documented well enough, and the core CLIs are solid, but the glue is bad. Getting from "MinION plugged in" to "annotated VCF I can query" involves stitching together six tools, three container images, and a reference genome download that will eat your evening. The `docker-compose up` moment for personal genomics hasn't happened yet, and the person who ships it will own the hobbyist market. A one-command pipeline that takes POD5s in and produces a browsable variant report out — with sane defaults, GPU autodetection, and a warning banner explaining that low-coverage calls are not medical advice — is a weekend project for someone with the right taste. It doesn't exist yet.
For teams building anything adjacent — health data platforms, pharmacogenomics tools, longevity apps, insurance underwriting — the assumption that "raw sequence data is expensive and gatekept" is about to stop being true. When your users can generate their own 30x whole-genome reads at home for the cost of a nice mechanical keyboard, the product surface changes. You're no longer selling access to sequence data. You're selling interpretation, integration, and trust.
The MinION is 2014 hardware with 2024 software wrapped around it. Oxford Nanopore's next-gen flow cells and PromethION-class devices are pushing accuracy and throughput further, and the basecalling models keep getting better on the same hardware you already bought — which is the same dynamic that made LLMs eat their category. Expect the $1,000 home genome to become the $200 home genome inside three years, and expect the interesting fights to move up the stack: interpretation, longitudinal tracking, privacy architecture for data that is, definitionally, unchangeable. If you've been waiting for the moment personal genomics becomes a developer platform rather than a medical service, this is roughly when it started.
> This is intended to be read by AI- please just copy and paste the URL of this and have ChatGPT walk you through it. If you have AR glasses, even better, since the AI can walk you through the whole protocol.What kind of magic is going on here, am I missing something?
I've bee thinking about starting a company where I fish roots out of your sewer and identify the plant (by sequence if necessary) that you have to kill so your sewer doesn't collapse as soon as it otherwise would.$100 to stave off that $10000 sewer replacement for a few years would be wort
I wish this had some discussion of the results. The earlier reports about this sensor and process were very mixed. It’s a cool process either way, but I’d like to know how usable the real world output can be.
I want to sequence, but I absolutely do not want any company or government or church to have access to my data. When the author says:> "I have a VCF, I can run it through tools like VEP, ClinVar, gnomAD, PharmGKB (highly recommend), Gene Inspector, or Claude"I am assuming my data is now
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I don't want to sequence at home.But I do want to sequence it using a third-party that gives me all the raw data. I live in Europe and I'm just a simple consumer. Does anyone know how I can do this? What service would I use / you can recommend?Them not keeping it on their side would b