README.md

---
license: apache-2.0
tasks:
  - genomic-sequence-modeling
frameworks:
  - jax
language:
  - en
  - zh
tags:
  - OneScience
  - Life Sciences
  - Genomics
  - DNA Sequence Model
  - Variant Effect Prediction
  - AlphaGenome
datasets:
  - OneScience/alphagenome_dataset
---

<p align="center">
  <strong>
    <span style="font-size: 30px;">AlphaGenome</span>
  </strong>
</p>

# Model Introduction

AlphaGenome is a DNA sequence model proposed by Google DeepMind. It can take DNA intervals up to 1 Mbp as input and predict multiple classes of genomic functional signals for track prediction and regulatory variant effect scoring.

Paper: Advancing regulatory variant effect prediction with AlphaGenome  
https://www.nature.com/articles/s41586-025-10014-0

# Model Description

AlphaGenome is implemented based on JAX / Flax and supports genomic interval inference, variant effect scoring, track evaluation, and fine-tuning examples. This model package is accompanied by the ModelScope dataset `OneScience/alphagenome_dataset`, which can be used for quick local verification.

# Applicable Scenarios

| Scenario | Description |
| :---: | :--- |
| Genomic interval prediction | Input a reference genome FASTA, chromosome, and interval coordinates, and output predicted tracks for ATAC, DNase, CAGE, RNA-seq, ChIP, and other signals |
| Variant effect scoring | Input a VCF or built-in example variants, compare prediction differences between reference and variant sequences, and generate a variant scoring table |
| Track prediction evaluation | Use validation data from the AlphaGenome dataset to calculate regression evaluation metrics for different assay bundles |
| Fine-tuning experiments | Use a custom reference genome, interval CSV, and BigWig signal files to verify the fine-tuning workflow |
| ModelScope / OneCode runtime | After downloading the model project and accompanying dataset, quickly verify script connectivity in a biology-domain runtime environment |



# Usage Instructions

## 1. OneCode Usage

You can experience intelligent one-click AI4S programming through the OneCode online environment:

[Click to experience intelligent one-click AI4S programming](https://web-2069360198568017922-iaaj.ksai.scnet.cn:58043/home)

## 2. Manual Installation and Usage

**Hardware Requirements**

- GPU or DCU runtime is recommended.
- CPU can be used for import checks and small-configuration connectivity verification, but full training and inference are slow.
- DCU users need to install DTK in advance. DTK 25.04.2 or later is recommended, or the OneScience-recommended version matching the current cluster.





**Environment Check**

- NVIDIA GPU:

```bash
nvidia-smi
```

- Hygon DCU:

```bash
hy-smi
```

### Download the Model Package

```bash
modelscope download --model OneScience/alphagenome --local_dir ./alphagenome
cd alphagenome
```

### Install the Runtime Environment

**DCU Environment**

```bash
# First activate DTK and Conda
conda create -n onescience311 python=3.11 -y
conda activate onescience311
# Supports uv installation
pip install onescience[bio-dcu] -i http://mirrors.onescience.ai:3141/pypi/simple/  --trusted-host mirrors.onescience.ai
```

After installation, return to the model package directory:

```bash
cd ./alphagenome
```

### Training and Inference Data Introduction

The OneScience community has uploaded the data required for AlphaGenome inference, evaluation, and fine-tuning to ModelScope: [OneScience/alphagenome_dataset](https://modelscope.cn/datasets/OneScience/alphagenome_dataset). After downloading, place the data in the `data/` directory of the model package.

```bash
modelscope download --dataset OneScience/alphagenome_dataset --local_dir ./data
```

### Training Weights

The repository includes `weight/alphagenome-all-folds`, and all scripts also support specifying model weights through `--model_dir`.

### Prepare Weights

If using local weights, place the AlphaGenome Orbax checkpoint in the following directory:

```text
weight/
  alphagenome-all-folds/
    _CHECKPOINT_METADATA
    _METADATA
    ...
```

If running in a shared runtime environment, you can also reuse a unified directory through environment variables:

```bash
export ONESCIENCE_MODELS_DIR=/path/to/onescience/models
export ONESCIENCE_DATASETS_DIR=/path/to/onescience/datasets
```

The scripts will preferentially read:

- `${ONESCIENCE_MODELS_DIR}/AlphaGenome/alphagenome-all-folds`
- `${ONESCIENCE_DATASETS_DIR}/AlphaGenome`

If the above environment variables are not set, the defaults under the current model package are read:

- `weight/alphagenome-all-folds`
- `data/`

### Interval Inference

```bash
bash scripts/inference.sh
```

Equivalent Python command example:

```bash
python scripts/run_inference.py \
  --fasta_path ./data/reference/HOMO_SAPIENS/GRCh38.p13.genome.fa \
  --model_dir ./weight/alphagenome-all-folds \
  --chromosome chr19 \
  --start 10587331 \
  --end 11635907 \
  --output_dir ./outputs
```

Inference results will be saved to `outputs/`.

### Variant Effect Scoring

```bash
bash scripts/run_variant.sh
```

When specifying VCF input, you can use:

```bash
python scripts/run_variant_scoring.py \
  --vcf_path ./data/example.vcf \
  --fasta_path ./data/reference/HOMO_SAPIENS/GRCh38.p13.genome.fa \
  --model_dir ./weight/alphagenome-all-folds \
  --output_dir ./outputs_variant
```

Scoring results will be saved as CSV files.

### Track Prediction Evaluation

```bash
bash scripts/run_track.sh
```

You can also explicitly specify the data and output paths:

```bash
python scripts/run_track_prediction_eval.py \
  --model_dir ./weight/alphagenome-all-folds \
  --model_version ALL_FOLDS \
  --data_dir ./data/v1/train \
  --output_path ./outputs_track/eval_results.csv
```

### Fine-tuning Example

```bash
python scripts/run_finetuning.py \
  --fasta_path ./data/reference/HOMO_SAPIENS/GRCh38.p13.genome.fa \
  --regions_csv ./data/finetune_regions.csv \
  --bigwig_paths ./data/sample_atac.bw \
  --output_dir ./finetuned_model \
  --num_steps 1000 \
  --batch_size 2
```

# Data Format

The ModelScope dataset `OneScience/alphagenome_dataset` is recommended to be downloaded to `data/` under the model package. The default structure is as follows:

```text
data/
  reference/
    HOMO_SAPIENS/
      GRCh38.p13.genome.fa
      GRCh38.p13.genome.fa.fai
  v1/
    train/
      ...
```

Where:

- `reference/HOMO_SAPIENS/GRCh38.p13.genome.fa` is the human reference genome FASTA.
- `.fai` is the FASTA index file.
- `v1/train/` is the data directory used for track prediction evaluation.
- Custom fine-tuning also requires preparing an interval CSV file with column names `chromosome,start,end`, as well as one or more BigWig signal files.

# Official OneScience Information

| Platform | OneScience Main Repository | Skills Repository |
| --- | --- | --- |
| Gitee | https://gitee.com/onescience-ai/onescience | https://gitee.com/onescience-ai/oneskills |
| GitHub | https://github.com/onescience-ai/OneScience | https://github.com/onescience-ai/oneskills |

# Citations and License

- This repository is based on the AlphaGenome open-source model and provides DCU adaptation. The related source code uses the Apache License 2.0.
- For scientific research, cite the original paper: [Advancing regulatory variant effect prediction with AlphaGenome](https://www.nature.com/articles/s41586-025-10014-0).