Why we built Expressive
The pitch for most consumer genetic-health products is roughly: "we'll sequence your DNA and tell you what to eat, what to take, and what to worry about."
That pitch sounds appealing right up until you ask the question every honest researcher already asks. Tell us what to eat, based on which study? Tell us what to take, anchored to which trial? Tell us what to worry about, given which p-value, and replicated by how many independent cohorts?
When you dig in, most of those answers don't exist. The recommendation came from a single underpowered study. Or a meta-analysis that conflated rare-disease findings with common variants. Or no study at all, just a vibe.
So we built something different. Expressive surfaces what current research actually says about each variant in your DNA, effect sizes, study quality, replication status, population data, and lets you decide what to do with it. Every claim links to its source. Every recommendation has to anchor on a specific PubMed ID, a PharmGKB clinical annotation, or a curated row from the NHGRI-EBI GWAS Catalog. Empty array is the correct output when the evidence doesn't support a concrete action, and most of the time, it does. This is what evidence-based genomics looks like when nobody is trying to sell you a supplement at the end of it.
The four guardrails
We came in with four constraints that everything else has to satisfy:
- Every claim is anchored. A recommendation that can't cite a specific paper, a specific PharmGKB level, or a specific GWAS catalogue entry doesn't ship. No "studies suggest" without saying which study.
- Hedge appropriately for thin evidence. A single underpowered cohort doesn't get the same confidence language as four replicated meta-analyses. When the literature is ambiguous, we say so out loud, with the same words a careful clinician would use.
- Bound Mendelian findings to Mendelian variants. The trap that ruins a lot of consumer genetics is conflating rare disease-causing variants in a gene with common GWAS hits in the same gene. They almost never share the same effect. If the literature for your variant talks about both, we say so explicitly.
- Refuse to invent. If the evidence doesn't support a concrete actionable change, we return zero recommendations and explain why. We'd rather show you "no clinical-grade actions warranted for this variant" than tell you to take magnesium because the model couldn't think of anything better. We don't prescribe, we describe.
These aren't preferences. They're load-bearing. A platform whose product is "we don't guess about your biology" cannot have a build process that tolerates guessing about its own outputs. Every shortcut you take there becomes the screenshot somebody posts to discredit the platform.
What we're shipping
Right now the public-facing corpus has plain-English research summaries for ~74,000 genetic variants and ~39,000 human genes. Each variant page anchors back to the original PubMed studies. Each gene page lays out what the literature actually says about its protein, its known disease associations, and what changes when you carry a variant in it.
Underneath the public corpus, we're building the personalized layer: upload your raw genetic file once (23andMe, AncestryDNA, MyHeritage, VCF), we map your variants to the research via dbSNP identifiers, and you get an action plan grounded in your actual genome, not a one-size-fits-all template. This is privacy-first DNA analysis: your genome stays yours, and the parsing happens on your terms.
The action plan v2 we're rolling out shortly does something most products don't: it consolidates recommendations across the variants you carry, surfaces contradictions when two variants pull in opposite directions (it happens, folate metabolism is the classic example), and lets you resolve them yourself with the underlying citations in hand.
What we're not
We're not a clinic. We don't diagnose. We don't replace your doctor. Anything that genuinely belongs in front of a physician, Mendelian disease findings, pathogenic ClinVar entries, pharmacogenomic flags for specific prescriptions, gets routed to "discuss with your doctor" with the relevant context, not turned into a lifestyle bullet.
We're also not a paywall around public data. The variant pages and gene pages are crawlable, searchable, and citable. The reason is partly philosophical (research that's anchored to PubMed should be accessible) and partly strategic (a public corpus that gets cited by other researchers, by LLMs, and by the medical community is the strongest moat we can build).
What you can do today
- Browse the public catalogue. Every variant page lists its trait associations, the studies, the genotype frequencies, and what we think the honest reading is. Start at /snp/ or /gene/.
- Sign up at the beta. You'll get the personalized layer when we open up access.
- Read along here. We'll be writing about what we ship, what we learn, and what we choose not to do.
We're early. The corpus will keep getting denser. The action plan will keep getting smarter. But the four guardrails stay the same. That's the bet.
Want updates when we ship new variant pages or a research deep-dive? Read the latest issue or get notified about early access.