Find open-source science resources

BEND 2.0 benchmarks the capabilities genomic models need to guide DNA design. Eight tasks cover sequence priors, mutation-effect prediction, biological context, genetic interactions, and sequence selection for desired responses. The Python package prepares experimental data, fits linear readouts to frozen model embeddings, evaluates predictions on biological holdouts, and produces scores, prediction tables, and local HTML reports. A separate interface evaluates native nucleotide probabilities. Model adapters support evaluating new models and submitting results to the public leaderboard. Requires Python 3.11 or later; model frameworks and weights are installed separately. The software is MIT-licensed. Benchmark datasets retain separate upstream terms, including CC BY 4.0 and CC BY-NC 4.0; see the data documentation.