NucleoBench

Environment	Unit Tests	Integration Tests
Micromamba
Docker Hub
GHCR

A large-scale benchmark for modern nucleic acid sequence design algorithms (NucleoBench), and a new design algorithm that outperforms existing designers (AdaBeam). Link to ICML GenBio 2025 workshop paper here.

This repo is intended to be used in a few ways:

1. Design a DNA sequence with selective expression in a cell-type (or any other target property in the benchmark, see list here), using the AdaBeam algorithm (or any of the ones listed here)
2. Design a DNA sequence with high binding affinity for a specific transcription factor (such as the ones listed here), using the AdaBeam algorithm (or any of the ones listed here)
3. Design a DNA or RNA sequence for a new task, using any designer (see tutorial here)
4. Run a new design algorithm on NucleoBench tasks.
5. Reproduce the NucleoBench results, using the standard tasks / designers or using your custom ones, on the Cloud using Google Batch or AWS (instructions here)

Citation

Please cite the following publication when referencing NucleoBench or AdaBeam:

@article{shor2025nucleobench,
  author    = {Shor, Joel and Strand, Erik and McLean, Cory Y.},
  title     = {{NucleoBench: A Large-Scale Benchmark of Neural Nucleic Acid Design Algorithms}},
  journal   = {bioRxiv},
  year      = {2025},
  doi       = {10.1101/2025.06.20.660785},
  url       = {https://www.biorxiv.org/content/10.1101/2025.06.20.660785}
}

Contents

• Quick Start
  • 1 minute install w/ pip
  • 3 minute install w/ docker
  • 5 minute install w/ source
  • 8 minute install w/ Google Batch inference
• Details
• FAQ

Quick Start

NucleoBench is provided via PyPi, Docker, or source.

Get started in 1 minute (pip install)

Install nucleobench on your terminal:

# Choose one.
pip install nucleobench  # optimizers and tasks
pip install nucleopt  # smaller, faster install for just optimizers

Then run in Python:

# 1. Choose a model (task).
from nucleobench import models
model = models.get_model('substring_count')
model_init_args = model.debug_init_args()
model_init_args['substring'] = 'ATGTC'
model_fn = model_obj(**model_init_args)

# 2. Choose an optimizer.
from nucleobench import optimizations
opt_obj = optimizations.get_optimization('adabeam')
opt_init_args = opt_obj.debug_init_args()
opt_init_args['model_fn'] = model_fn
opt_init_args['start_sequence'] = 'A' * 100
designer = opt_obj(**opt_init_args)

# 3. Run the designer and show the results.
designer.run(n_steps=100)
ret = designer.get_samples(1)
ret_score = model_fn(ret)
print(f'Final score: {ret_score[0]}')
print(f'Final sequence: {ret[0]}')

Output:

Step 99 current scores: [np.float64(508.0), np.float64(507.0), np.float64(506.0), np.float64(505.0), np.float64(504.0), np.float64(503.0), np.float64(503.0), np.float64(502.0), np.float64(502.0), np.float64(502.0)]
Final score: -508.0
Final sequence: AGATGTCATATATGATGTCATGTCATGTCGTCATGTCTGTCTCTCATGTATGTCATGTCTATGTCTGTCTATGTCTATGTCTATGTCATGTCTATGTCTC

This "recipe" can be found under recipes/python/adabeam_substring.py.

Get started in 3 minutes (docker image pull)

Get the image:

docker image pull joelshor/nucleobench:latest

Output:

latest: Pulling from joelshor/nucleobench
Digest: sha256:602230b568c0f15acfa7a0b6723ffb16fab6f32c37ae5b88c71763fb722ab5c3
Status: Image is up to date for joelshor/nucleobench:latest
docker.io/joelshor/nucleobench:latest

Make a directory for output:

readonly output="./output/docker_recipe/adabeam_atac"
mkdir -p "${output}"
readonly fullpath="$(realpath $output)"

Then run it:

docker run \
    -v "${fullpath}":"${fullpath}" \
    joelshor/nucleobench:latest \
    --model substring_count \
        --substring 'ATGTC' \
    --optimization adabeam \
        --beam_size 2 \
        --n_rollouts_per_root 4 \
        --mutations_per_sequence 2 \
        --rng_seed 0 \
    --max_seconds 15 \
    --optimization_steps_per_output 5 \
    --proposals_per_round 2 \
    --output_path ${fullpath} \
    --start_sequence AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

Output:

Completed round 3441 (5 steps) took 0.00s. Avg 0.00s per step.
Proposals deposited at:
	/Users/joelshor/Desktop/docker_test/output/docker_recipe/adabeam_atac/adabeam_substring_count/20250731_194857/20250731_194912.pkl

This "recipe" can be found under recipes/docker/adabeam_atac.sh.

Get started in 5 minutes (git clone)

git clone https://github.com/move37-labs/nucleobench.git
cd nucleobench
conda env create -f environment.yml
conda activate nucleobench

Now run the main entrypoint:

python -m docker_entrypoint \
    --model substring_count \
        --substring 'ATGTC' \
    --optimization adabeam \
        --beam_size 2 \
        --n_rollouts_per_root 4 \
        --mutations_per_sequence 2 \
        --rng_seed 0 \
    --max_seconds 15 \
    --optimization_steps_per_output 5 \
    --proposals_per_round 2 \
    --output_path ./output/python_recipe/adabeam_atac \
    --start_sequence AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

Output:

...
Completed round 3820 (5 steps) took 0.00s. Avg 0.00s per step.
  0%|                              | 3821/99999999 [00:14<109:01:33, 254.77it/s]
Proposals deposited at:
	./output/python_recipe/adabeam_atac/adabeam_substring_count/20250731_162119/20250731_162134.pkl

This "recipe" can be found under recipes/python/adabeam_atac.py.

Get started in 8 minutes (Google Batch inference)

Google Batch is Google's cheapest batch compute offering. It enables relatively cheap parallel compute on the cloud.

First, setup a Google Cloud project by following instructions here. You will need to activate the Google Batch API. If you want the job to write to a private bucket, you will also need to setup what's called a "Service Account" with the proper permissions. Once you have done this, you will need to collect the following information from your new project, and fill this information in runners/gcp/config.py:

1. PROJECT_ID (the project ID of the project you created above)
2. REGION (the region of the project)
3. BUCKET_NAME (the bucket of the output)
4. SERVICE_ACCOUNT_EMAIL (option: a service account for security)

Next, create the conda environment to launch jobs to the Cloud runner:

conda env create -f runners/environment.yml

Output:

Transaction finished

To activate this environment, use:

    conda activate runners

Or to execute a single command in this environment, use:

    conda run -n runners mycommand

Now run the dry run Google Batch job with the test script:

conda activate runners
python -m runners.gcp.job_launcher \
    --dry-run \
    --verbose \
    runners/testdata/adabeam_test.tsv

Output:

2025-08-28 18:00:21,651 - INFO - Row 1: Generated job name: bpnet-rad21-adabeam-001
2025-08-28 18:00:21,651 - INFO - Loaded 1 jobs from runners/testdata/adabeam_test.tsv
2025-08-28 18:00:21,651 - INFO - DRY RUN MODE - No jobs will be launched
2025-08-28 18:00:21,651 - INFO - Would launch job: bpnet-rad21-adabeam-00120250828-18-00-21

Now run the real job:

python -m runners.gcp.job_launcher \
    --verbose \
    runners/testdata/adabeam_test.tsv

Output:

2025-08-28 18:01:32,909 - INFO - Row 1: Generated job name: bpnet-rad21-adabeam-001
2025-08-28 18:01:32,909 - INFO - Loaded 1 jobs from runners/testdata/adabeam_test.tsv
2025-08-28 18:01:32,910 - DEBUG - Checking None for explicit credentials as part of auth process...
2025-08-28 18:01:32,910 - DEBUG - Checking Cloud SDK credentials as part of auth process...
2025-08-28 18:01:33,528 - DEBUG - Starting new HTTPS connection (1): oauth2.googleapis.com:443
2025-08-28 18:01:33,708 - DEBUG - https://oauth2.googleapis.com:443 "POST /token HTTP/1.1" 200 None
2025-08-28 18:01:34,087 - INFO - Successfully launched job: projects/nucleorave/locations/us-central1/jobs/bpnet-rad21-adabeam-00120250828-18-01-32
2025-08-28 18:01:34,087 - INFO - 
Job Launch Summary:
2025-08-28 18:01:34,087 - INFO - Successful: 1
2025-08-28 18:01:34,087 - INFO - Failed: 0
2025-08-28 18:01:34,087 - INFO - Successful jobs: bpnet-rad21-adabeam-00120250828-18-01-32

Finally, make sure that the job completes fully, and the output is in the right bucket.

Details

NucleoBench is a large-scale comparison of modern sequence design algorithms across 16 biological tasks (such as transcription factor binding and gene expression) and 9 design algorithms. NucleoBench, compares design algorithms on the same tasks and start sequences across more than 400K experiments, allowing us to derive unique modeling insights on the importance of using gradient information, the role of randomness, scaling properties, and reasonable starting hyperparameters on new problems. We use these insights to present a novel hybrid design algorithm, AdaBeam, that outperforms existing algorithms on 11 of 16 tasks and demonstrates superior scaling properties on long sequences and large predictors. Our benchmark and algorithms are freely available online.

Comparison of nucleic acid design benchmarks

NAME	YEAR	ALGOS	TASKS	SEQ. LENGTH (BP)	DESIGN BENCHMARK	LONG SEQS	LARGE MODELS	PAIRED START SEQS.
Fitness Landscape Exploration Sandbox	2020	4-6	9	Most <100	✅	❌	❌	✅
Computational Optimization of DNA Activity	2024	3	3	200	✅	❌	❌	✅
gRelu	2024	2	5	500K (20 edit)	❌	❌	✅	❌
Linder et al repos	2021	2	20	<600	✅	❌	❌	❌
NucleoBench (ours)	2025	9	16	256-3K	✅	✅	✅	✅

Table: Nucleic acid design from sequence benchmarks. All benchmarks prior to NucleoBench are limited either in the range of tasks they measure against, the range of optimizations they compare, or the complexity of the task.

Summary of tasks in NucleoBench

TASK CATEGORY	MODEL	DESCRIPTION	NUM TASKS	SEQ LEN (BP)	SPEED (MS / EXAMPLE)
Cell-type specific cis-regulatory activity	Malinois	How DNA sequences control gene expression from the same DNA molecule. Cell types are: precursor blood cells, liver cells, neuronal cells.	3	200	2
Transcription factor binding	BPNet-lite	How likely a specific transcription factor (TF) will bind to a particular stretch of DNA. Specific TFs: CTCF, E2F3, ELF4, GATA2, JUNB, MAX, MECOM, MYC, OTX1, RAD21, SOX6	11	3000	55
Chromatin accessibility	BPNet-lite	How physically accessible DNA is for interactions with other molecules.	1	3000	260
Selective gene expression	Enformer	Prediction of gene expression. We optimize for maximal expression in muscle cells, minimal expression in liver cells.	1	196,608 / 256 *	15,000

*Input length is 200K, but only 256 bp are edited.

Summary of designers in NucleoBench

Algo	Description	Gradient-based
Directed Evolution	Random mutations, track the best.	❌
Simulated Annealing	Greedy optimization with random jumps.	❌
AdaLead	Iterative combining and mutating of a population of sequences.	❌
FastSeqProp	Sampling and the straight-through estimator for maximal input.	✅
Ledidi	Sampling and the gumbel softmax estimator for maximal input.	✅
---
Ordered Beam	Greedy search, in fixed sequence order, with cache.	❌
Unordered Beam	Greedy search with cache.	❌
Gradient Evo	Directed Evolution, guided by model gradients.	✅
AdaBeam (ours)	Hybrid of Unordered Beam and improved AdaLead.	❌

Table: Summary of designers in NucleoBench. Above the solid line are designers already found in the nucleic acid design literature. Below the line are designers from the search literature not previously used to benchmark nucleic acid sequence design and hybrid algorithms devised in this work.

FAQ

1. How can I add a new task to NucleoBench? A: Follow this colab notebook.