--- 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-Sugon/alphagenome_dataset ---

AlphaGenome

# Model Introduction AlphaGenome is a DNA sequence model developed by Google DeepMind. It accepts DNA intervals of up to 1 Mbp as input and predicts a range of functional genomic signals for genomic 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 in JAX / Flax and supports genomic interval inference, variant effect scoring, track evaluation, and example fine-tuning workflows. This model package is accompanied by the Hugging Face dataset `OneScience-Sugon/alphagenome_dataset`, which enables rapid local validation. # Use Cases | Scenario | Description | | :---: | :--- | | Genomic interval prediction | Takes a reference-genome FASTA file, a chromosome, and genomic interval coordinates as input and outputs predicted tracks such as ATAC, DNase, CAGE, RNA-seq, and ChIP | | Variant effect scoring | Takes either a VCF file or built-in example variants as input, compares predictions for the reference and variant sequences, and produces a variant scoring table | | Track prediction evaluation | Uses validation data from the AlphaGenome dataset to compute regression evaluation metrics across different assay bundles | | Fine-tuning experiments | Validates a fine-tuning workflow using custom reference genomes, interval CSVs, and BigWig signal files | | Hugging Face / OneCode execution | After downloading the model project and accompanying dataset, quickly verifies that the scripts run correctly in a life-sciences runtime environment | # Usage Guide ## 1. OneCode Usage Try one-click AI4S development in the OneCode online environment: [Try one-click AI4S development](https://web-2069360198568017922-iaaj.ksai.scnet.cn:58043/home) ## 2. Manual Installation and Usage **Hardware Requirements** - GPU or DCU is recommended. - A CPU can be used for import checks and lightweight configuration tests; full training and inference will be slow. - DCU users must install DTK in advance. DTK 25.04.2 or later is recommended, or a OneScience-recommended version matching the current cluster. **Environment Check** - NVIDIA GPU: ```bash nvidia-smi ``` - Hygon DCU: ```bash hy-smi ``` ### Download the Model Package ```bash hf download --model OneScience-Sugon/alphagenome --local-dir ./alphagenome cd alphagenome ``` ### Install the Runtime Environment **DCU Environment** ```bash # Activate DTK and CONDA first conda create -n onescience311 python=3.11 -y conda activate onescience311 # uv installation supported 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 ``` ### Data for Inference, Evaluation, and Fine-Tuning The OneScience community has uploaded the data required for AlphaGenome inference, evaluation, and fine-tuning to Hugging Face: [OneScience-Sugon/alphagenome_dataset](https://huggingface.co/datasets/OneScience-Sugon/alphagenome_dataset). After downloading, place the data in the `data/` directory under the model package. ```bash hf download --dataset OneScience-Sugon/alphagenome_dataset --local-dir ./data ``` ### Model Weights The repository already includes `weight/alphagenome-all-folds`, and all scripts allow the model weights directory to be specified via `--model_dir`. ### Preparing Weights If using local weights, place the AlphaGenome Orbax checkpoint in the following directory: ```text weight/ alphagenome-all-folds/ _CHECKPOINT_METADATA _METADATA ... ``` When running in a shared environment, you can also reuse centrally managed model and dataset directories via environment variables: ```bash export ONESCIENCE_MODELS_DIR=/path/to/onescience/models export ONESCIENCE_DATASETS_DIR=/path/to/onescience/datasets ``` Scripts look in the following locations first: - `${ONESCIENCE_MODELS_DIR}/AlphaGenome/alphagenome-all-folds` - `${ONESCIENCE_DATASETS_DIR}/AlphaGenome` If the above environment variables are not set, the scripts fall back to the following paths under the current model package: - `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 are saved to `outputs/`. ### Variant Effect Scoring ```bash bash scripts/run_variant.sh ``` To specify a VCF input: ```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 are saved as a CSV file. ### Track Prediction Evaluation ```bash bash scripts/run_track.sh ``` You can also explicitly specify 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 It is recommended to download the Hugging Face dataset `OneScience-Sugon/alphagenome_dataset` to `data/` under the model package. The default directory structure is as follows: ```text data/ reference/ HOMO_SAPIENS/ GRCh38.p13.genome.fa GRCh38.p13.genome.fa.fai v1/ train/ ... ``` In this structure: - `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 an interval CSV file with column names `chromosome,start,end`, as well as one or more BigWig signal files. # OneScience Official 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 | # Citation & License - This repository is adapted from the open-source AlphaGenome model to support DCUs. The source code is licensed under Apache License 2.0. - For scientific use, please cite the original paper: [Advancing regulatory variant effect prediction with AlphaGenome](https://www.nature.com/articles/s41586-025-10014-0).