Decoding your genome with EVEE

A reading of one human genome with a genomic foundation model: what its DNA carries, and why carrying a disease variant is almost never the same as having the disease.

Topics	Life sciences · Linear probes
Models	Evo2-7B (EVEE pathogenicity probe) · gnomAD frequency gate · ClinVar / OMIM review
References	EVEE (Evo Variant Effect Explorer) · Evo2 genomic foundation model · 1000 Genomes Project

A genome is a three-billion-letter text, inherited in two copies, one from each parent. Compared against the reference sequence scientists use as a yardstick, any one person carries millions of single-letter differences.

To read those variants, this recipe uses EVEE, the Evo Variant Effect Explorer. EVEE scores how disruptive a DNA change looks by reading it through Evo2, a genomic foundation model trained on enormous amounts of genetic sequence, so it has learned what normal DNA looks like and can flag a change that breaks the pattern. EVEE is the model-based step inside a variant diagnosis pipeline: a funnel that starts from every coding and splice variant in a genome and filters down to the handful worth a human's attention. This recipe runs that pipeline on one real genome, HG00096, an anonymous volunteer from the 1000 Genomes Project, and reads back what it carries.

Key takeaways

• Silico can put life-sciences foundation models to work. It picks up specialized genomic models like Evo2 and the EVEE pathogenicity probe and runs them on a real genome, so the model machinery is available to apply rather than something the reader has to build.
• Intelligent filtering turns raw data into a reviewable shortlist. Silico chains the model's scores with ancestry-aware frequency gating and literature review, narrowing tens of thousands of variants down to the handful worth a human's attention.

How the funnel works

The funnel starts with every variant in the information-dense part of the genome: the 37,460 single-letter differences that fall in protein-coding and splice regions, the stretches that actually spell out proteins. From there it applies one filter at a time, and the pile shrinks at each step.

• Is the change damaging? The first filter keeps only variants predicted to alter or break the protein a gene builds, and sets the silent ones aside.
• Is it rare? The next filter is the decisive one. It looks up each variant's real population frequency in gnomAD, matched to the person's ancestry, and discards anything common. The logic is simple: if a variant is common, plenty of healthy people carry it, so it cannot be the cause of a rare, serious disease. This single step clears 81% of the damaging-looking variants.
• How disruptive does the model think it is? EVEE reads each survivor through Evo2-7B, a foundation model trained on enough genetic sequence to know what normal DNA looks like, and scores how badly the change breaks the pattern (probe_v3, on a 0-to-1 scale). The score measures molecular disruption and nothing else, not whether the person is ill.
• What does the literature say? The handful of rare, damaging, high-scoring variants that remain are checked against ClinVar and OMIM for known gene-disease links and inheritance, and each is labelled: carrier, drug-response, risk, uncertain, or benign.

What comes out the bottom is five findings worth talking about.

What the funnel found

Five findings survived. Three are carrier findings: a single damaged copy of a gene that would only cause illness if both copies were broken. One affects how the body handles a particular drug, and matters only if that drug is ever prescribed. The last nudges the odds of a condition up a little.

These carrier states are exactly what EVEE is built to surface. Most genes come in two copies, and for many of them one working copy is enough, so a single damaged copy causes no symptoms and would never bring someone to a clinic. EVEE catches it anyway: reading each variant straight from the DNA, it scores the damage in a silent copy that no symptom would ever reveal, turning an otherwise invisible carrier state into a finding a human can review.

Two honest limits are worth stating. This is a reading of one genome with one tool, and it looks only at the parts of the DNA that spell out proteins, so it is not a whole-genome screen. And the scores are predictions, not certainties: a carrier or risk finding describes a possibility for the future or for a child, never the person's health today.

Related recipes

Probe the copy-number axis in Evo 2 7BA linear probe and steering vector for tandem-repeat copy number in Evo 2. Finding the Tree of Life in Evo 2A faithful replication that reads bacterial phylogeny out of a genomic model's embeddings. Driving novelty in protein generation modelsA generation-time nudge that recovers novelty from a copy-prone protein design model.