Lab Newsletter — July 28, 2026: Two Ledgers

A year ago the genome-model story was all headline: Evo 2 predicting BRCA1 pathogenicity at >90% accuracy, AlphaGenome scoring a variant a second across the non-coding 98%. 2026 is the quieter, more useful chapter — the audit. The question has shifted from “what can the model do?” to “what survives when you force every claim through a held-out test set with an honest baseline?”

📒 Two ledgers, one honest baseline

A sharp 2026 analysis proposes reading this field with two separate ledgers: a capability ledger (what a model can demonstrably do at scale — what the press release reports) and a validity ledger (“what holds up when you pass each claim through an independent test set with an honest baseline”). Kept apart, the verdict is refreshingly specific: genomic foundation models are genuinely maturing for variant-effect prediction, but they fail for perturbation prediction and mechanistic interpretability — there, “five foundation models plus two other deep networks failed to beat simple linear baselines.” And the leaderboards are unstable: “the same model can be a breakthrough in one paper and an underperformer in another,” because rankings reshuffle by task. Why it matters for the lab: this is the Virtual Cell Challenge lesson at the genome layer — a held-out benchmark against a baseline you must actually beat is how you tell signal from froth. It’s the discipline our BioImage Model Zoo builds around: models the community can benchmark, not just admire.

🧬 The right model for the variant class

“One model to rule them all” is quietly dying, and that’s progress. On coding missense variants, the specialists still win — Evo 2’s 40B and 7B “ranked fourth and fifth, behind AlphaMissense, ESM-1b, and GPN-MSA.” On non-coding and splicing variants, the foundation models lead — Evo 2 sets the state of the art, AlphaGenome matched or beat external models on 24 of 26 evaluations. The honest framing: zero-shot scores are now “good enough to contribute evidence under an ACMG-style framework” but “not good enough to act alone.” A March 2026 preprint sharpens the caution, reporting systematic blind spots in Evo 2 — weaknesses on short-range signals like codon-usage bias and a counter-intuitive drop on the most severe variants — enough, its authors argue, to “challenge current claims of zero-shot pathogenicity prediction” (preprint, not yet peer-reviewed). Why it matters for the lab: and there’s a third axis beyond accuracy — access. Evo 2 is “one of the largest fully open AI models in any domain”; AlphaGenome is “API-only and non-commercial.” For a regulated lab, open weights you can freeze and audit are “a materially different proposition” — which is the whole bet behind Hypha, BioEngine and ImJoy.

🔍 Don’t just score the variant — explain it, and open it

The most encouraging response to the audit is to build for it. EVEE (Evo Variant Effect Explorer, Goodfire × Mayo Clinic, Apr 2026 preprint) trains interpretable probes on Evo 2 (7B) embeddings, reports 0.997 AUROC across variant types (generalizing zero-shot to indels), and — the part that matters — turns each prediction into a natural-language explanation of which biological signals a variant disrupts. It ships as a public web tool with precomputed calls for all 4.2 million ClinVar variants, and even as an MCP server so an LLM agent can query it directly. Its own disclaimer stays honest: “computational predictions are not diagnoses.” Why it matters for the lab: a model that reasons and explains and can be queried by an agent is precisely the shape of the tools we build — the BioImage.IO chatbot, Agent-Lens. An interpretable, open, agent-callable variant oracle is our whole thesis wearing a lab coat.

The froth is being sorted from the signal by the most old-fashioned instrument there is: a held-out test set. That’s not bad news for genome AI — it’s the field growing up. And the tools that come out ahead are the open, benchmarked, explainable ones. On that scoreboard, at least, we’ve been playing the long game all along.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.