AlphaGenome 2 adds cell-type and cell-state information to DNA-effect prediction, a specialized advance that could help researchers choose experiments. A newly reviewed visual-work benchmark finds three leading agents statistically indistinguishable on its harder criteria, with substantial requirements still unmet. The November 10–11 sources distinguish promising scientific prediction from experimental validation, and benchmark scores from complete professional work.
Sources checked and limits of this review
Review since the November 11 briefing, including relevant November 10 material newly examined. Checked primary watchlists at METR, ARC Prize, Epoch AI, Google DeepMind, OpenAI and Anthropic, plus independent-research searches. Nature’s November 11 reporting supplied the AlphaGenome 2 lead, checked against DeepMind’s original announcement and official API repository. The linked AlphaGenome 2 paper exceeded the retrieval tool’s size limit, so its full methods were not reviewed and detailed numerical performance claims are not repeated here. Read the original November 10 MKA benchmark paper, including grading methods, model-comparison uncertainty and limitations. It is a lab-authored evaluation involving DeepMind’s own models, not a wholly independent replication. Its original November 10 date is preserved. No additional item was selected from the other primary watchlists or independent-paper leads checked; this is not a claim that no other research appeared. Previously covered Signal proofs, robotics, Anthropic releases and OpenAI stop controls are not redated as new findings. Coverage is bounded by accessible pages and indexing. Reported results and editorial SI interpretations are labeled separately; no human editorial approval, overall wisdom verdict or pivotal designation is claimed.
Evidence note 1 · Lab-reported scientific model and evaluation; SI relevance is editorial interpretation
AlphaGenome 2 predicts DNA effects in specific cell types and states
DeepMind’s AlphaGenome 2 takes a million-letter DNA sequence together with a target biological condition, rather than predicting only across fixed broad-tissue outputs. Trained on over 100 million cell profiles, it can use new single-cell measurements to condition predictions without retraining. DeepMind reports improved variant identification in immune-cell tests.
Our interpretation. Editorial interpretation: identifying where a DNA change acts could help scientists prioritize experiments and disease mechanisms. This is evidence about a specialized predictive tool, not a general scientist or an observed increase in discovery rates.
Main limitation. Lab-reported results, not independent replication. Predictions still need experimental validation; the tool has not been validated or approved for clinical diagnosis. Reproducing a known disease mechanism computationally is not a new treatment discovery. The full paper exceeded retrieval limits, so detailed benchmark methods and numerical comparisons were not checked here.
Next evidence. Independent tests on unseen cell states, prospective perturbation experiments, uncertainty calibration, and measured effects on experimental success and researcher time.
Evidence note 2 · Lab-authored cross-model benchmark paper, newly reviewed; SI relevance is editorial interpretation
Newly reviewed: visual-work benchmark finds no clear winner among three leading agents
DeepMind’s MKA benchmark tests 200 visually demanding professional workflows across ten sectors. On harder rubric criteria, Fable 5.2, Gemini 4 Carbon and Opus 5.5 were statistically indistinguishable. The highest score was 56.6% of weighted criteria, not 56.6% of complete jobs successfully automated.
Our interpretation. Editorial interpretation: reading diagrams and checking produced files can expose weaknesses hidden by text-heavy tests. Measuring these abilities helps assess whether agents can deliver usable work, rather than merely plausible answers.
Main limitation. Lab-authored benchmark, not independent reproduction. Each model ran once per task, so reported uncertainty omits run-to-run variation. Automated grading agreed closely with experts in a 50-task study focused on selected high-weight criteria; that does not validate every judgment. The harder subset depends on a particular earlier-model panel.
Next evidence. Independent tasks and repeat runs, broader human grading, robustness to task and judge changes, and comparisons that include total cost and fully accepted deliverables.