Thinking about how organisations perform in an AI-enabled world

AyEye Today · 12 November 2026

The model knows which cell to ask. Biology still has to answer

A sweeping ultramarine current gathers one chartreuse and oxblood knot from dispersed dark marks across scraped warm ground.
Original artwork: The Specific Context. Chiappe × OpenAI.

Cell-state conditioning improves genetic triage, but predictions still need experimental and clinical proof.

The story, brieflyEvidence · Interpretation

The argument

AlphaGenome 2 can make a research queue more relevant by predicting where a DNA variant acts. It cannot make that queue conclusive.

What is new

DeepMind added a tissue, cell type or state as an input throughout the model, trained with profiles sampled from 107 million human cells.

Why it matters

On one held-out immune-cell benchmark, recall rose from 16% to 30% at 90% precision, helping researchers select a more specific first experiment.

LecXie’s practical interpretation

Record the model version, biological condition, confidence threshold, retained alternatives and experiment that could overturn the prediction. Let negative results update the assessment.

Evidence & limits

The benchmark and model come from DeepMind’s preprint. High precision still misses most variants; condition embeddings are observational, and the system has no clinical approval.