Protein engineering is shifting from model-based variant scoring to deciding what to measure next. This exposes a limitation of fixed-dataset benchmarks, which rank predictors without showing when their advantage survives as evidence accumulates.
ORIGINAL PAPER
Decision benchmarking reveals when protein foundation models add experimental value
WHAT WE KNOW
We therefore introduce decision benchmarking across 217 ProteinGym assays, 2. 47 million variants and 95 predictors to evaluate whether models and measurements improve decisions at matched cost. Zero-shot accuracy explained less than one-fifth of the variation in discovery after calibration.
WATCH NEXT
Review the primary source, validate the main result, and establish whether any listed-company transmission is direct.
EVIDENCE
What the evidence supports so far
Research signals
Classification pending.
What remains unverified
The full methodology, effect size, and limitations still require analyst review.
Company impact remains unverified until a direct economic transmission is established.