BEND 2.0: Benchmarking the Capabilities Needed to Design DNA
Accurate predictions of DNA changes can still miss context-specific effects and interactions between mutations.
BEND 2.0 tests DNA models across eight tasks built on five experimental datasets, including whether predictions can guide sequence design. Linear predictors trained on frozen model representations can match or beat models’ native outputs, while functional post-training substantially improves performance. In design tasks, predictions reduce the number of candidates needed to find a desired response. Substantial gaps remain, particularly when designing for specific contexts or objectives across multiple cell types.
HF Daily Papers' note
BEND 2.0 tests DNA models across eight tasks built on five experimental datasets, including whether predictions can guide sequence design. Linear predictors trained on frozen model representations can match or beat models’ native outputs, while functional post-training substantially improves performance. In design tasks, predictions reduce the number of candidates needed to find a desired response. Substantial gaps remain, particularly when designing for specific contexts or objectives across multiple cell types.
HF Daily Papers' note
score 5