By Maximilian Naumow — I started with 23andMe in 2018 and would decide differently today. This is the guide I would have needed back then.
Key Findings
- SNP chips (23andMe/Ancestry) read less than 0.1% of your genome. Whole Genome Sequencing (WGS) at 30× coverage reads all 3 billion base pairs an average of 30 times. For AI-assisted analysis, WGS gives 5-10× more data points — that’s the real leverage in 2026.
- 23andMe still exists, after Chapter 11 in March 2025 acquired by TTAM Research Institute for $305M. Raw data export works normally. If you have old 23andMe data: export now, save locally.
- WGS providers 2026 (prices for 30× standard): Dante Labs (Italy) EUR 399 (sales down to EUR 169) · DNA Complete (ex-Nebula) $495 · Sequencing.com $379. Range: $169-495.
- Genome raw data does NOT belong in cloud LLMs. GDPR Article 9 data, uniquely identifiable. For analysis: local database + Claude API with strict retention settings, or local LLM.
- Statistical reality: Most SNP effects have odds ratios of 1.05-1.3. Nearly irrelevant for individual decisions. Exceptions: monogenic variants (BRCA, LDLR) and high-impact pharmacogenomics (CYP2D6).
- LLMs hallucinate excessively on genome questions — without grounding on dbSNP/ClinVar/SNPedia, answers are unusable.
Why this matters now
Two things happened together in 2025-2026:
- 23andMe filed Chapter 11 (March 2025). The company continues, but the future is nonprofit-driven. Millions of genetic raw records change hands — a good moment to think about your own data sovereignty.
- Whole Genome Sequencing became affordable. What cost $10,000 in 2018 is under $500 in 2026. Even below $200 during flash sales.
That’s the point where “knowing your own genome” is no longer a Silicon Valley luxury but a realistic baseline dataset. Combined with a local health-data hub and AI interpretation, you get something no commercial app delivers: structured genome queries against your own dataset, without cloud silos, without provider interpretation.
SNP chip vs. WGS — the difference is dramatic
If you’ve ever done 23andMe or Ancestry: you got a SNP chip. Not a genome. A sample.
| 23andMe / Ancestry (SNP chip) | Whole Genome Sequencing (30×) | |
|---|---|---|
| Positions read | ~600,000-1 million SNPs | ~3 billion base pairs |
| Genome coverage | < 0.1 % | ~99.9 % |
| Rare variants | Not captured | Fully captured |
| Coverage depth | 1× per position | 30× per position (99.98% accuracy) |
| Raw file size | ~15 MB (.txt) | ~90-120 GB (FASTQ, compressed ~30 GB) |
| AI data points | ~700k | ~3-4 million variants |
| Price (2026) | $99-199 | $169-495 |
The price delta today is so small that the choice is obvious: WGS or nothing. A SNP chip in 2026 is like a 5-megapixel camera in 2026 — technically possible, but why?
Provider landscape 2026
If you’re starting today, go WGS. Concrete options:
Dante Labs (Italy, EU-based) 30× WGS for EUR 399 standard. Flash sales 3-4× per year down to EUR 169. NovaSeq X platform (newer Illumina chemistry, higher Q30 scores than NovaSeq 6000). FASTQ, BAM, and VCF downloadable in customer portal. Pragmatic for EU users: no US import, EUR pricing, GDPR-compliant operation.
DNA Complete (formerly Nebula Genomics) $495 for 30× WGS. Plus $99 for FASTQ access (standard package only ships VCF). Parent company ProPhase Labs Chapter 11 in September 2025 — company still operates, but long-term future unclear.
Sequencing.com $379 for 30× WGS plus microarray. Additionally offers a platform with analysis apps. Interesting for “no setup, app does everything” users — but vendor lock-in.
23andMe / Ancestry (SNP chip) If you already have an account: download raw data, save locally. Do not buy new in 2026.
Selfdecode, MyHeritage, tellmeGen, etc. Third-party analysis platforms. Use your uploaded raw data. Vendor lock-in, monthly costs. Not recommended for the data-sovereignty approach.
The data-sovereignty point (that almost everyone overlooks)
Genome data is a special category:
- Uniquely identifiable: Even 100 SNPs are enough to identify you uniquely among all humans on Earth.
- GDPR Article 9: Highest protection tier for “special categories of personal data” (health + genetics).
- Not changeable: You can change a leaked password. Not a leaked genome.
- Family implications: Your raw data reveals partial genotypes of your parents, siblings, children — even if they never took a test.
Consequence for AI analysis: Genome raw data has no business in ChatGPT, Claude web UI, or any cloud LLM with “we may use your data for training” terms. Not partially. Not “just the interesting SNPs.”
The safe path:
- Raw data stays local — on your own machine, in a SQLite DB or directly as a VCF file.
- LLM interpretation via API with
zero-retentionflag. Anthropic offers this for Claude via API (Zero Data Retention for qualified users). OpenAI has similar options for Enterprise. - Ideally a local LLM (Ollama with Llama-4, LM Studio with Mistral) — then nothing leaves your hardware.
- Grounding on official databases: dbSNP, ClinVar, SNPedia with explicit version tags. Don’t let the LLM guess from memory.
The practical workflow (2026)
Concretely, how to build your genome into a sensible AI-assisted setup in 2026:
Step 1 — Get raw data Order WGS (Dante Labs or comparable). After ~4-6 weeks you receive FASTQ + VCF files for download. Save locally, encrypted (LUKS on Linux, FileVault on macOS, VeraCrypt cross-platform).
Step 2 — Local database
Parse the VCF file into a SQLite DB. Minimal structure: chromosome, position, ref_allele, alt_allele, genotype, rs_id, zygosity. A Python script with pyvcf or cyvcf2 does it in 20 lines.
Step 3 — Select interesting loci Not all 3 million variants are relevant. Every use case has core loci:
- Pharmacogenomics: CYP2D6, CYP2C19, CYP2C9, VKORC1 (warfarin), TPMT
- Strength sport: ACTN3 (rs1815739), ACE (rs4646994), MSTN
- Cognitive markers: APOE (rs429358, rs7412), COMT (rs4680), BDNF (rs6265)
- Cardio: LPA (rs10455872), PCSK9, LDLR
- Methylation: MTHFR (rs1801133, rs1801131), MTR, MTRR
Pre-built SNP lists for common categories: SNPedia, Promethease as a commercial alternative, OpenSNP for community resources.
Step 4 — Build the grounding layer Don’t let the LLM answer from memory. Build a simple query pipeline:
- User question → LLM identifies relevant SNPs
- Script queries your local SNP DB for your genotypes
- Second script call queries ClinVar/SNPedia snapshot for evidence
- LLM synthesizes an answer from the evidence, not from memory
That’s more work than “just ask ChatGPT,” but it separates “informed interpretation” from “dangerous coffee-ground reading.”
Step 5 — Integrate into OpenClaw / biohub
If you already run openclaw-biohub (my self-hosted health-data hub), you can build a genome adapter alongside: new directory under pipeline/adapters/genome/, VCF parser, snp_variants table. Then your OpenClaw agent can query your SNPs via SKILL.md and interpret them in context with your other health data.
The real payoff: connect pharmacogenomics SNPs with your supplement correlations. If your CYP2D6 genotype is “poor metabolizer” and ashwagandha shows no effect on you (see supplement correlations in the biohub setup): that’s a mechanism, not coincidence.
What you can realistically learn from SNPs
Honest expectation-setting, sorted by usefulness:
High utility (clear clinical relevance):
- Monogenic disease variants: BRCA1/2 (breast cancer), familial hypercholesterolemia (LDLR), HFE (hemochromatosis). If a pathogenic variant is present, that has clinical consequences. But: go to a specialist here, not to an LLM.
- High-impact pharmacogenomics: CYP2D6 poor metabolizers metabolize ~25% of all prescription drugs differently. Clinically relevant for codeine, tramadol, SSRIs, beta-blockers, tamoxifen.
- Nutrient absorption: LCT (lactase persistence), HLA-DQ2/DQ8 (celiac risk), COL1A1 (collagen synthesis).
Medium utility (informative, not actionable):
- Strength-sport markers: ACTN3 R/R says “likely fast-twitch dominant.” But training overrides this massively. Net effect for training planning: minimal.
- APOE: e3/e4 heterozygous has ~2× elevated Alzheimer baseline risk. Real, but lifestyle factors (sleep, exercise, cardiovascular health) dominate prognosis significantly.
- MTHFR C677T: Not the boogeyman TikTok makes it out to be. Homozygous T/T = 30% lower enzyme activity. Consequence: folate-rich nutrition or methylated folate supplementation. Nothing more.
Low utility (statistically measurable, practically irrelevant):
- Most common SNPs with “elevated risk for X” have odds ratios of 1.05-1.3. Interesting at population level, noise for you as an individual.
- Personality SNPs, “warrior vs worrier” genotypes, extraversion markers: effect sizes so small that chance is more likely than causation.
Zero or negative utility:
- “DNA-based nutrition recommendations” from commercial providers. Meta-analyses show no better outcomes than generic recommendations.
- “Genetic fitness programs”. Marketing, not science.
- “Genomic personality analysis”. Bullshit.
What LLMs get wrong on genome questions
From practical experience (not theory): the most common LLM hallucinations in genome interpretation:
- Inverted effects: LLM claims “T allele protects” when actually “T allele is the risk allele.” Common because primary literature is ambiguously written.
- Fabricated rs IDs: LLM cites “rs1234567” that doesn’t exist or has nothing to do with the claimed gene.
- Outdated assignments: SNPs occasionally get re-annotated, studies get retracted. LLMs stay stuck on old data without knowing it.
- Ignoring multi-locus effects: LLM analyzes one SNP in isolation, ignoring that 5 other SNPs in the same pathway cancel the effect.
- Transferring population data to individuals: “This variant is associated with +15% diabetes risk” doesn’t mean “you have +15% diabetes risk.”
These errors are avoidable with a grounding layer (always ClinVar/dbSNP snapshot as ground truth), but without it, LLM-based genome interpretation is dangerously unreliable.
Who this is for
Yes, if you:
- Take chronic medication or likely will in the future (pharmacogenomics utility is real)
- Have family history of monogenic diseases (BRCA, LDLR, HFE)
- Already run a QS dataset and want genome as another layer
- Have command-line comfort and value local-first
No, if you:
- Expect “what my genome says about my optimal life path” — the answer is “very little”
- Expect fitness or nutrition recommendations from SNPs
- Aren’t willing to work with odds ratios and confidence intervals
- Find the idea that a test tells you “what you are” exciting — that’s closer to esoterica than science
Verdict
Genome testing in 2026 is no longer the Silicon Valley statement kit of 2018. For $200-400 you get whole genome sequencing. For no additional cent you get LLM interpretation — if you’re willing to build the grounding layer yourself.
What doesn’t work: dump raw data into ChatGPT and ask “what does this mean?” What works: keep raw data local, query against ClinVar/SNPedia, use Claude or a local LLM as a synthesizer for the evidence.
The honest payoff is smaller than marketing promises — and larger than skeptics say. For pharmacogenomics alone, it pays off after a single avoided “wrong antidepressant reaction” episode. For everything else: don’t expect the oracle, just one more data point.
If you already run a health-data hub like openclaw-biohub: a genome adapter is the logical next step. Then SNPs aren’t an isolated trivia dataset but become part of your integrated health analysis.
Disclaimer: Not medical advice. Genome interpretation for clinically relevant questions (hereditary diseases, drug dosing, cancer risk) belongs to a human geneticist or specialist. This article describes self-experiment frameworks and AI tooling, not medical recommendations. LLM output on genome questions without primary-database grounding is unreliable — do not use as a diagnostic substitute.
Updated: July 16, 2026 (initial publication)