AlphaGenome Research

Model Weights | Installation | Quick Start | Documentation | Community | Terms of Use

AlphaGenome is a unified DNA sequence model designed to advance regulatory variant-effect prediction and shed light on genome function. It analyzes DNA sequences of up to 1 million base pairs to deliver predictions at single base-pair resolution across diverse modalities, including gene expression, splicing patterns, chromatin features, and contact maps.

This repository provides the following research code:

• An implementation of the AlphaGenome model, written in JAX.
• An implementation of the AlphaGenome API with accompanying variant scorers.
• A dataset loader for reading AlphaGenome training data from TFRecords.
• An example colab notebook for analysing our evaluation results.

We strongly recommend using our AlphaGenome API to interact with the model without needing specialized hardware.

Installation

To install, clone a local copy of this repository and run pip install:

$ git clone https://github.com/google-deepmind/alphagenome_research.git
$ pip install -e ./alphagenome_research

This will install any required dependencies, including this repository in development mode.

Model weights

To use our pre-trained model weights, you can download them from either:

• Kaggle or
• Hugging Face

Both require accepting our non-commercial model terms. Requests are processed immediately.

Model requirements

In order to run the model, we recommend running with at least an NVIDIA H100 GPU. Please ensure CUDA, cuDNN and JAX are correctly installed; the JAX installation documentation is a useful resource in this regard.

For training, we recommend running on Tensor Processing Units (TPUs) v3 or higher.

Quick start

The easiest way to interact with the AlphaGenome model is using the provided DNA Model class. This wraps the core model and provides a more intuitive set of functions for creating predictions, scoring variants, performing in silico mutagenesis (ISM) and more.

It also provides the following factory functions to create a model instance using weights:

from alphagenome_research.model import dna_model

# To download from Kaggle:
model = dna_model.create_from_kaggle('all_folds')

# or Hugging Face:
model = dna_model.create_from_huggingface('all_folds')

Here’s an example of making a variant prediction using model weights downloaded from Kaggle:

from alphagenome.data import genome
from alphagenome.visualization import plot_components
from alphagenome_research.model import dna_model
import matplotlib.pyplot as plt

model = dna_model.create_from_kaggle('all_folds')

interval = genome.Interval(chromosome='chr22', start=35677410, end=36725986)
variant = genome.Variant(
chromosome='chr22',
position=36201698,
reference_bases='A',
alternate_bases='C',
)

outputs = model.predict_variant(
interval=interval,
variant=variant,
ontology_terms=['UBERON:0001157'],
requested_outputs=[dna_model.OutputType.RNA_SEQ],
)

plot_components.plot(
[
plot_components.OverlaidTracks(
tdata={
'REF': outputs.reference.rna_seq,
'ALT': outputs.alternate.rna_seq,
},
colors={'REF': 'dimgrey', 'ALT': 'red'},
),
],
interval=outputs.reference.rna_seq.interval.resize(2**15),
# Annotate the location of the variant as a vertical line.
annotations=[plot_components.VariantAnnotation([variant], alpha=0.8)],
)
plt.show()

Citing AlphaGenome

If you use AlphaGenome in your research, please cite using:

@article{alphagenome,
title={Advancing regulatory variant effect prediction with {AlphaGenome}},
author={Avsec, {\v Z}iga and Latysheva, Natasha and Cheng, Jun and Novati, Guido and Taylor, Kyle R. and Ward, Tom and Bycroft, Clare and Nicolaisen, Lauren and Arvaniti, Eirini and Pan, Joshua and Thomas, Raina and Dutordoir, Vincent and Perino, Matteo and De, Soham and Karollus, Alexander and Gayoso, Adam and Sargeant, Toby and Mottram, Anne and Wong, Lai Hong and Drot{\'a}r, Pavol and Kosiorek, Adam and Senior, Andrew and Tanburn, Richard and Applebaum, Taylor and Basu, Souradeep and Hassabis, Demis and Kohli, Pushmeet},
journal={Nature},
volume={649},
number={8099},
year={2026},
doi={10.1038/s41586-025-10014-0},
publisher={Nature Publishing Group UK London}
}

Acknowledgements

AlphaGenome’s model release uses the following libraries and packages:

• Abseil
• anndata
• Chex
• Einshape
• Etils
• Haiku
• huggingface_hub
• JAX
• jaxtyping
• kagglehub
• NumPy
• Orbax
• pandas
• pyarrow
• pyfaidx
• PyRanges
• TensorFlow
• typeguard

We thank all their contributors and maintainers!

License and Disclaimer

Copyright 2026 Google LLC

All software is licensed under the Apache License, Version 2.0 (Apache 2.0); you may not use this except in compliance with the Apache 2.0 license. You may obtain a copy of the Apache 2.0 license at: https://www.apache.org/licenses/LICENSE-2.0. As noted above, model weights are available via Kaggle and Hugging Face and are subject to the model terms at: https://deepmind.google.com/science/alphagenome/model-terms.

Code examples and documentation to help you use the AlphaGenome model are licensed under the Creative Commons Attribution 4.0 International License (CC-BY). You may obtain a copy of the CC-BY license at: https://creativecommons.org/licenses/by/4.0/legalcode.

Unless set out below under the Training Data, Evaluation Data or Training and Evaluation Data headings, all other materials are licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC-BY-NC). You may obtain a copy of the CC-BY-NC license at: https://creativecommons.org/licenses/by-nc/4.0/legalcode.

Unless required by applicable law or agreed to in writing, all software and materials distributed here under the Apache 2.0 or CC-BY licenses are distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the licenses for the specific language governing permissions and limitations under those licenses.

This is not an official Google product.

Training Data

FANTOM5: This data has been reprocessed. The original FANTOM5 data is made available at https://fantom.gsc.riken.jp/5/ under a CC-BY license (see link above for a copy). Citation: Lizio M, et al. Update of the FANTOM web resource: expansion to provide additional transcriptome atlases. Nucleic Acids Res. 47: D752–D758 (2019). https://doi.org/10.1093/nar/gky1099.

4D Nucleome: This data has been reprocessed using the method set out in the ‘Methods’ section of the accompanying paper. The original 4D Nucleome data is available from the 4DN Data Portal at https://data.4dnucleome.org/ and subject to the Data Use Guidelines found there. The 4DN Data Portal is part of 4DN, citation 4DN White Paper (https://www.nature.com/articles/nature23884) and 4DN Data Portal Paper (https://www.nature.com/articles/s41467-022-29697-4).

Evaluation Data

This data includes: (i) list of variants; (ii) target values; and (iii) AlphaGenome predicted score.

CAGI: CAGI data can be obtained from genomeinterpretation.org/challenges.html and is subject to the terms found here: http://www.genomeinterpretation.org/data-use-agreement.html.

GTEx v8: GTEx v8 data can be obtained from: gtexportal.org/home. The data used for the work described in this paper was obtained from: https://github.com/calico/borzoi. Please visit the GTEx Portal for the most up to date and accurate version of this data.

GTEx v8 reprocessed into EMBL-EBI eQTL catalogue: Data originally made available at the GTEx Portal (see above) with modifications made by EMBL-EBI, and provided under a CC-BY-4.0 license a copy of which can be found at https://creativecommons.org/licenses/by/4.0/legalcode. Citation: Kerimov, N., Hayhurst, J.D., Peikova, K. et al. A compendium of uniformly processed human gene expression and splicing quantitative trait loci. Nat Genet 53, 1290–1299 (2021). https://doi.org/10.1038/s41588-021-00924-w.

ChromBPNet: ChromBPNet data can be obtained at https://www.synapse.org/Synapse:syn59449898/files/. Citation: Pampari, A. et al. ChromBPNet: bias factorized, base-resolution deep learning models of chromatin accessibility reveal cis-regulatory sequence syntax, transcription factor footprints and regulatory variants. BioRxiv, 2024–12 (2025).

ClinVar: ClinVar data can be found at: https://ftp.ncbi.nlm.nih.gov/pub/clinvar/vcf_GRCh38/ subject to this Data Use Policy https://www.ncbi.nlm.nih.gov/clinvar/docs/maintenance_use/. Citation: Landrum, M. J. et al. ClinVar: improving access to variant interpretations and supporting evidence. Nucleic Acids Res. 2018 Jan 4;46(D1):D1062-D1067. doi: 10.1093/nar/gkx1153.

MFASS: MFASS data can be found at https://github.com/KosuriLab/MFASS. Citation: A Multiplexed Assay for Exon Recognition Reveals that an Unappreciated Fraction of Rare Genetic Variants Cause Large-Effect Splicing Disruptions; Chong, Rockie et al.; Molecular Cell, Volume 73, Issue 1, 183 - 194.e8.

eQTL: eQTL data is provided with a CC-BY-4.0 license a copy of which can be found here: https://creativecommons.org/licenses/by/4.0/legalcode. Citation: Kerimov, N., Hayhurst, J.D., Peikova, K. et al. A compendium of uniformly processed human gene expression and splicing quantitative trait loci. Nat Genet 53, 1290–1299 (2021). https://doi.org/10.1038/s41588-021-00924-w.

Open Targets: Open Targets data can be obtained at https://platform-docs.opentargets.org/licence and is provided with a Creative Commons 1.0 Universal license, a copy of which can be found here: https://creativecommons.org/publicdomain/zero/1.0/legalcode.

PolyA site annotations: PolyA site annotations can be obtained here: https://exon.apps.wistar.org/polya_db/v3/. The data used for this project is a reprocessed version which can be found: https://storage.googleapis.com/seqnn-share/helper/polyadb_human_v3.csv.gz. Citation: Linder, J., Srivastava, D., Yuan, H. et al. Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation. Nat Genet 57, 949–961 (2025). https://doi.org/10.1038/s41588-024-02053-6.

Training & Evaluation Data

ENCODE: This data has been reprocessed. The original ENCODE data is made available at https://www.encodeproject.org/help/getting-started/#download pursuant to the Data Use Policy at https://www.encodeproject.org/help/citing-encode/. The specific data can be found cited in the Supplementary Tables published as part of the Advancing regulatory variant effect prediction with AlphaGenome paper. The data is presented by the ENCODE Consortium, whose most recent publications are:

• ENCODE integrative analysis (PMID: 22955616; PMCID: PMC3439153)
• ENCODE portal (PMID: 41168159; PMCID: PMC12575607; PMID: 31713622; PMCID: PMC7061942)
• ENCODE uniform processing pipelines: https://www.biorxiv.org/content/10.1101/2023.04.04.535623.

GENCODE: Copyright of the released GENCODE dataset is © 2024 EMBL-EBI. A modified version of the GENCODE dataset (which can be found here: https://www.gencodegenes.org/human/releases.html), is made available with reference to the following:

• Copyright © 2024 EMBL-EBI
• The GENCODE dataset is subject to the EMBL-EBI terms of use, available at https://www.ebi.ac.uk/about/terms-of-use.
• Citation: Frankish A, et al (2018) GENCODE reference annotation for the human and mouse genome.
• Further details about GENCODE can be found at https://www.gencodegenes.org/human/releases.html, with additional citation information at https://www.gencodegenes.org/pages/publications.html and further acknowledgements can be found at https://www.gencodegenes.org/pages/gencode.html.

To prepare the dataset, the team followed the method set out here: https://github.com/google-deepmind/alphagenome/blob/main/scripts/process_gtf.py

Third-party software

Your use of any third-party software, libraries or code referenced in the materials in this repository (including the libraries listed in the Acknowledgments section) may be governed by separate terms and conditions or license provisions. Your use of the third-party software, libraries or code is subject to any such terms and you should check that you can comply with any applicable restrictions or terms and conditions before use.