BEND 2.0 benchmark tests if DNA foundation models can guide design
TL;DR
- BEND 2.0 spans eight tasks built on five experimental datasets to test whether genomic model predictions can guide DNA sequence design, not just score edits.
- Linear readouts fitted to frozen model representations match or exceed the models' native outputs and transfer across experimental assays and genes.
- Strong average-effect prediction can hide weaker context-specific performance, and accurate combined-mutation prediction does not imply accurate interaction prediction.
A new benchmark of genomic foundation models reports that current systems can help narrow candidate DNA sequences but "still leave substantial room for better design choices, particularly for context-specific and multi-cell objectives."
BEND 2.0, posted to arXiv on November 2, 2026 by Andreas Bjerregaard and colleagues, bundles "eight tasks built on five experimental datasets" and asks whether model predictions can actually "guide sequence design," not just score edits.
Two findings carry the paper. "Linear readouts fitted to frozen representations can match or exceed models' native outputs," the authors write, and transfer across assays and genes. But "strong prediction of average effects can conceal weaker prediction of context-specific effects, and accurate prediction of combined mutation effects does not imply accurate prediction of their interactions."
The authors frame the result in measured terms: genomic models "now display signs of general utility but have yet to saturate on performance in even simple tasks." The abstract names no individual model's score; a public leaderboard is online.
Originally reported by paper
Read the original article →Original headline: BEND 2.0 Finds Genomic Foundation Models Stall on the DNA Design Tasks That Actually Matter