bend2 1.0.1

We built BEND 2.0 to test whether genomic models can guide the design of DNA. Across eight tasks, we examine how they predict sequence function, account for cell type and genetic background, and choose sequences that produce a desired response.

BEND prepares the data, fits linear readouts to your model's frozen embeddings, and reports task scores and predictions. Python 3.11 or later is required. Your model's framework and weights are installed separately.

python -m pip install bend2

Benchmark data

The package is available now; the GitHub repository and online leaderboard are awaiting their public launch. Full documentation is included in the source distribution under Download files on this page.

Evaluate your model

With BEND installed, save an adapter as my_model.py:

import numpy as np
from bend2 import Model


def embed(sequences: list[str]) -> np.ndarray:
    return your_encoder(sequences)  # [batch, features]


model = Model("My model", embed=embed, revision="checkpoint-v1")

Replace your_encoder with your model's inference. Return one finite, fixed-width CPU NumPy vector per complete sequence, in input order, and record the actual model revision.

Start with Atlas: 7,657 short sequences, sixteen multi-cell design patterns, and data included with the package.

bend2 run my_model.py --tasks atlas_profile --output results/my-model
bend2 show results/my-model

To try the included base-composition baseline without model weights:

bend2 run composition --tasks atlas_profile --output results/composition

Tasks

Seven tasks use embeddings; prior needs nucleotide probabilities. The siraj task group runs context, background, repair_percentile and cell_selective together, sharing inference work. Inspect requirements before a larger run:

bend2 plan my_model.py --tasks siraj

Data and results

Small datasets ship with the package. Larger files download from the public BEND 2.0 data release, pinned to a fixed revision and checked against the original file sizes and checksums. Source-specific terms apply to the data; see the data guide and source terms.

Saved results can be opened without loading the model:

from bend2 import Run

run = Run("results/my-model")
scores = run.to_dataframe()
predictions = run.table("atlas_profile")

Runs stay local until you choose to share them. Use bend2 --help for commands and bend2 tasks for task requirements. The source distribution includes docs/submitting.md for preparing a future leaderboard contribution.