AlphaGenome PyTorch

A PyTorch port of AlphaGenome, the DNA sequence model from Google DeepMind that predicts hundreds of genomic tracks at single base-pair resolution from sequences up to 1M bp.

We strive to make it an accessible, readable, and hackable implementation — for integrating into existing PyTorch pipelines, fine-tuning on custom datasets, and building on top of.

Installation

Installation from PyPI:

pip install alphagenome-pytorch

Installation from repo:

pip install git+https://github.com/genomicsxai/alphagenome-pytorch

For fine-tuning (incl. BigWig data loading):

pip install alphagenome-pytorch[finetuning]  # adds pyBigWig, pyfaidx

Use with coding agents

If you use a coding agent, docs/alphagenome-usage.md is an agent-agnostic guide to running the model and pulling out predictions for a specific assay, cell type, or resolution, and docs/finetuning/ covers fine-tuning.

For Claude Code, install them as a plugin so the agent picks them up automatically in any project:

/plugin marketplace add genomicsxai/alphagenome-pytorch
/plugin install alphagenome@alphagenome

This adds two skills — one for predictions, one for fine-tuning — so you can ask things like "get DNase predictions from GM12878 at 128bp", "write a wrapper to get all the K562 predictions", or "fine-tune on my ATAC BigWigs with LoRA".

Quick Start

import torch
import numpy as np
from alphagenome_pytorch import AlphaGenome

# Load pretrained model
model = AlphaGenome.from_pretrained('alphagenome.pt', device='cuda')

# Create one-hot encoded DNA sequence in NLC format (batch=1, length=131072, channels=4)
# Channels: A=0, C=1, G=2, T=3
sequence = np.random.randint(0, 4, size=(1, 131072))
dna_onehot = torch.tensor(np.eye(4)[sequence], dtype=torch.float32).cuda()

# Inference (handles dtype casting, returns float32 outputs)
outputs = model.predict(dna_onehot, organism_index=0)  # organism: 0=human, 1=mouse

The weights for this port are available on Hugging Face.

Output structure

Each genomic-track head returns a dict mapping resolution → tensor:

outputs['atac'][1]           # (1, 131072, 256)   ATAC-seq  at 1 bp
outputs['atac'][128]         # (1, 1024,   256)   ATAC-seq  at 128 bp
outputs['dnase'][1]          # (1, 131072, 384)   DNase     at 1 bp
outputs['cage'][128]         # (1, 1024,   640)   CAGE      at 128 bp
outputs['chip_histone'][128] # (1, 1024,   1152)  ChIP-hist at 128 bp only

Contact maps are returned as a single tensor (no resolution dict):

outputs['contact_maps']      # (1, 64, 64, 28)   3D chromatin contacts

Splice heads return dicts of tensors:

outputs['splice_sites']['probs']  # (1, 131072, 5)  splice site classes

Padding

Track dimensions are padded (e.g. ATAC has 167 real human tracks but the tensor has 256 channels). Real tracks come first; the rest are zeros. Use named_outputs=True to auto-strip padding:

from alphagenome_pytorch.named_outputs import NamedOutputs, TrackMetadataCatalog
catalog = TrackMetadataCatalog.load_builtin(organism=0)
model.set_track_metadata_catalog(catalog)

named = model.predict(dna_onehot, organism_index=0, named_outputs=True)
named.atac[1].shape                  # (1, 131072, 167)  — padding removed
named.atac[1].tracks[-1].track_name  # 'UBERON:0015143 ATAC-seq'

# Filter by metadata
named.rna_seq[128].select(strand='+')
named.chip_tf[128].select(transcription_factor='CTCF')
named.atac[1].select(biosample_type='tissue', ontology_curie='UBERON:0015143')

Extracting Embeddings

Use model.encode() to get embeddings without running prediction heads — useful for building custom heads or analyzing representations:

# Get embeddings (128bp only for efficiency)
emb = model.encode(dna_onehot, organism_index=0, resolutions=(128,))
emb['embeddings_128bp']  # (B, 1024, 3072) at 128bp

Fine-tuning

Train a new head on your data with frozen trunk (linear probing) or with LoRA adapters:

from alphagenome_pytorch import (
    AlphaGenome, TransferConfig, load_trunk, prepare_for_transfer,
)

# Load trunk, freeze, add custom heads
model = AlphaGenome()
model = load_trunk(model, 'alphagenome.pt')
model = prepare_for_transfer(model, TransferConfig(
    mode='lora',
    new_heads={'atac': {'modality': 'atac', 'num_tracks': 1}},
    lora_rank=8,
))

The easiest way to start with fine-tuning is the agt finetune CLI (installed with the package):

# LoRA fine-tuning
agt finetune --mode lora --lora-rank 8 \
    --genome hg38.fa --modality atac --bigwig *.bw \
    --train-bed train.bed --val-bed val.bed \
    --pretrained-weights alphagenome.pt

# LoRA + Locon on the last 4 encoder convs before attention
agt finetune --mode lora+locon --lora-rank 8 \
    --locon-rank 4 --locon-targets down_blocks.4,down_blocks.5 \
    --genome hg38.fa --modality atac --bigwig *.bw \
    --train-bed train.bed --val-bed val.bed \
    --pretrained-weights alphagenome.pt

# Multi-GPU (torchrun needs a module target)
torchrun --nproc_per_node=4 -m alphagenome_pytorch.cli finetune --mode lora ...

agt finetune and python scripts/finetune.py are the same code path with the same flags; the script is kept as a compatibility shim (it only works from a repo clone).

Note that Locon targets are explicit by design. You can use the syntax down_blocks.3,down_blocks.4,down_blocks.5 to choose targets.

See examples/notebooks/finetune_linear_probe.ipynb for an example of linear probing on ATAC-seq data.

Numerical Parity with JAX

This port is validated against the original JAX model, including per-head and full forward pass output comparisons as well as loss values and gradients.

See a compiled ARCHITECTURE_COMPARISON.md for some technical details.

Model Outputs

Tracks column shows the number of real human tracks (without padding). Dimension is the raw output tensor size — padding fills the gap. When using named_outputs=True, padding is stripped by default. See named outputs guide for details.

See more information about model outputs in the official AlphaGenome documentation.

Example Notebooks

• Demo — Basic inference and JAX comparison
• Variant Scoring — Effect prediction
• In Silico Mutagenesis — ISM analysis
• TAL1 Mutation Example - TAL1 variant effect and ISM (Figure 6 from AlphaGenome)
• Fine-tuning — ATAC-seq linear probing
• Fine-tuning — MPRA (encoder-only)

Citation

@article{avsec2026alphagenome,
  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 others},
  journal={Nature},
  volume={649},
  number={8099},
  pages={1206--1218},
  year={2026},
  publisher={Nature Publishing Group UK London}
}

@article{avsec2025alphagenome,
    title = {AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model},
    author = {Avsec, {\v Z}iga and Latysheva, Natasha and Cheng, Jun and ...},
    year = {2025},
    journal = {bioRxiv},
    doi = {10.1101/2025.06.25.661532}
}

Acknowledgements

We acknowledge Phil Wang, Miquel Anglada-Girotto, and Xinming Tu as developers of an older AlphaGenome PyTorch port unrelated to this repo. Note that the PyPI namespace is now linked to this repo.

License

This project is a port of the google-deepmind/alphagenome_research repository licensed under the Apache License, Version 2.0:

Copyright 2026 Google LLC

The model parameters, output, and any derivatives thereof remain subject to Google DeepMind’s AlphaGenome Model Terms.

This port is licensed under the Apache License, Version 2.0 (Apache 2.0):

Copyright 2026 Danila Bredikhin, Martin Kjellberg, Christopher Zou, Alejandro Buendia, Xinming Tu, Anshul Kundaje

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this except in compliance with the License. Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.