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