Deep dive asks whether AlphaGenome Atlas's precomputed 9bn-variant lookup is a real unlock
Babayaga1664 · reddit · 2026-09-10
A non-geneticist's close read of DeepMind's AlphaGenome Atlas preprint questions its significance: the team precomputed 9bn single-base changes plus 100m indels, turning per-variant inference into lookup, plus an AVI score and motif map. Concerns: predictions aren't measurements; the headline clinical result (29.5% top-50 recall on solved GREGoR cases vs 12.5% for CADD) is retrospective; only 4 of 25 rare-variant findings nominally replicated in All of Us, none surviving Bonferroni. The DNM1 deep-intronic case stands out, but is it representative? Would clinical geneticists actually use it, and is precomputation the real unlock or framing on an incremental gain?
More from Research
- FLASH: GPU-native deformable sim learns cloth folding in minutes on one RTX 5090 — ChongZzZhang · 2026-10-26
- LEMON: zero-shot LLM misbehavior monitor flags hallucinations from attention maps — yoavartzi · 2026-10-26
- Behind the EMNLP keynote: LMs as simulated learners to generate child language acquisition hypotheses — najoungkim · 2026-10-26
- The paper behind the EMNLP 2026 keynote on bigram 'simulated learners' — najoungkim · 2026-10-26
- Laying out DeepLIFT as a standalone computation graph: from sequence to attributions directly — jmschreiber91 · 2026-10-26
- Speculative Diffusion Paper Maps the Speed-Quality Pareto Frontier via Two Tradeoffs — UCLA · 2026-10-26