Explore the results
How many edits to reach a target response?
We start with an edit whose effect on RNA activity has been measured.
A benchmark for
genomic design
Why?
There are far more ways to change a DNA sequence than we can test in the lab. Models could help us choose which changes to try, but we need to know how much we can rely on their suggestions.
BEND 2.0 was built around that question. We compare models using experimental measurements, including cases where the same edit behaves differently depending on the cell type or surrounding DNA. We want to understand when their predictions are useful for deciding what to test next.
Data sources
Acknowledgements
Supported by the Novo Nordisk Foundation through the Center for Basic Machine Learning Research in Life Science (NNF20OC0062606), CAZAI (NNF22OC0077058), Quantum for Life (NNF20OC0059939), and the NovoSTAR programme, and by the Pioneer Centre for AI (DNRF grant P1). Access to the Gefion supercomputer for our broader research is supported by DCAI and a Novo Nordisk Foundation grant (NNF25OC0105165).
Code & data
Evaluate a new modelGitHub ↗ Benchmark dataHugging Face ↗ Original BENDGitHub ↗Cite BEND 2.0