AlphaGenome Atlas

The 30 comments (all HackerNews, many high-upvote) show a scientifically literate but skeptical audience. Multiple commenters link a real mutagenesis study (science.org 'Mutate 'em all' post plus biorxiv 2026.07.25.740675) that fuzzed a simple virus one mutation at a time: dedicated AI models made poor predictions of most mutation outcomes, half the mutations were harmful, and for half of those the mechanism is unknown. This is used to 'cast doubts about the value of the AlphaGenome predictive map,' since a human genome is far more complex than a virus. A widely echoed technical critique: 'the benchmark labels came from conservation, so you can't tell prediction from re-reading the prior' (circularity risk). One commenter cites Katie Pollard's ISMB talk concluding existing human variation is insufficient context to infer variant impact. A wet-lab researcher is blunt: 'in practice not even alpha fold can tell me if my gfp fusion will work... biology is not there yet.' Others ask whether promoter sequences and transcription-rate queries will be possible; the dataset reportedly captures non-coding DNA. A Borzoi comparison (Nature) clarifies the Atlas is a database, not a model. Consumer value is dismissed: 'Probably not any 23andMe haven't already told you about,' since consumer panels test a limited SNP set. Some credit DeepMind's track record (AlphaFold, earthquake alerts) while noting other DeepMind bio models 'have performed more poorly than other available models.' Caveat: a large share of the thread drifts into off-topic US politics (Trump vs Google, democracy debates) and bot-accusation meta, diluting product signal. Net: genuine excitement about a 1-petabyte, ~9-billion-variant precomputed resource, dominated by expert doubt about whether the predictions are actually predictive.

No aggregate ratings were found for this product during the last harvest.