--- title: alphagenome-fold-3 canonical_url: "https://www.modelscope.cn/models/google/alphagenome-fold-3" md_url: "https://www.modelscope.cn/models/google/alphagenome-fold-3.md" repository: google/alphagenome-fold-3 last_updated: 2026-01-29 license: other library_name: - pytorch frameworks: - pytorch language: - en downloads: 10 stars: 0 tags: - biology --- # alphagenome-fold-3 > alphagenome-fold-3 - google 在 ModelScope 开源的模型。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… google/alphagenome-fold-3 是 ModelScope 魔搭社区上的机器学习模型，采用 other 许可。 - **Repository**: google/alphagenome-fold-3 - **License**: other - **Tags**: biology - **Downloads**: 10 - **Stars**: 0 - **Last updated**: 2026-01-29 Source: https://www.modelscope.cn/models/google/alphagenome-fold-3 --- ## Description 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. The core architecture features a U-Net-style design that combines an encoder for downsampling, transformers with inter-device communication to capture long-range interactions, and a decoder for upsampling. These components feed into task-specific output heads that generate predictions at their respective assay-specific resolutions. By achieving state-of-the-art performance across diverse genomic benchmarks, AlphaGenome provides a robust framework for understanding the molecular function of DNA sequences and interpreting non-coding variation. ## Inputs and outputs * **Input:** Up to 1 Mb (2\*\*20) one-hot encoded DNA sequence, and an organism type index (representing human or mouse). * **Output:** 11 diverse modalities, including: * RNA expression (RNA-Seq, CAGE-seq and PRO-cap). * Chromatin accessibility (DNase-seq and ATAC-seq). * Histone modifications. * Transcription factor binding. * Chromatin contact maps. * Splice sites and their usage. * Splice junction coordinates and strength. ## Citation ``` @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} } ``` ## Installation To install the accompanying code necessary to run the model, please run the following: ```shell $ pip install git+https://github.com/google-deepmind/alphagenome_research.git ``` ## Usage In addition to the model, we provide a DNA model class that wraps the core model and provides a more intuitive set of functions for creating predictions, scoring variants, performing in silico mutagenesis (ISM) and more. Here's an example of making a variant prediction: ```python 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_huggingface('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() ``` ## Model Data AlphaGenome was trained to predict read coverage for a wide range of different functional genomics assays – including RNA-seq, DNase-seq, CAGE, ChIP-seq – directly from human and mouse reference genome sequences. The training data was sourced from large public consortia including [ENCODE](http://encodeproject.org/), [GTEx](https://www.gtexportal.org/), [4D Nucleome](https://4dnucleome.org/) and [FANTOM5](https://fantom.gsc.riken.jp/5/), encompassing experimental measurements of key regulatory modalities across hundreds of cell types and tissues. For more information, please refer to the [AlphaGenome manuscript](https://www.biorxiv.org/content/10.1101/2025.06.25.661532v2). Along with the release of the weights and model inference code, we are releasing the complete training, validation and test datasets used for AlphaGenome. The data is stored in compressed TFRecord format and can be loaded as follows: ```python from alphagenome.data import fold_intervals from alphagenome_research.io import dataset from alphagenome_research.model import dna_model ds_iter = dataset.create_dataset( organism=dna_model.Organism.HOMO_SAPIENS, fold_split=dna_model.ModelVersion.ALL_FOLDS, subset=fold_intervals.Subset.TRAIN, ).as_numpy_iterator() element = next(ds_iter) ``` Each element in the dataset is a tuple, where each item corresponds to a specific data bundle. The example below illustrates this structure, showing the ATAC and DNase bundles alongside the input DNA sequence, masks, and interval metadata: ```python ({"atac": "bfloat16[1052672,256]", "atac_mask": "bool[1,256]", "dna_sequence": "float32[1052672,4]", "interval/chromosome": b"chr7", "interval/end": "int64[]", "interval/start": "int64[]"}, {"dna_sequence": "float32[1052672,4]", "dnase": "bfloat16[1052672,384]", "dnase_mask": "bool[1,384]", "interval/chromosome": b"chr7", "interval/end": "int64[]", "interval/start": "int64[]"}, ... ) ``` **Notes:** * **Interval extension**: To support data augmentation via sequence shifting during training, the 1 Mb intervals were extended by 4,096 base pairs (2,048 bp on each side). * **GTEx data exclusion**: The released dataset does not contain GTEx tissue data due to licensing restrictions, though the corresponding column headers remain in the bundle’s metadata. However, users can easily extend the dataset with GTEx or other external sources using the provided interval metadata (chromosome, start, end). ## Implementation Information ### Hardware AlphaGenome training involved a two-stage process: pre-training and distillation. Pre-training was carried out on 256 [Tensor Processing Units (TPUv3)](https://docs.cloud.google.com/tpu/docs/v3) using sequence parallelism across groups of 4 interconnected chips. We leverage TPU pods (large clusters of TPUs) to provide a scalable solution for training, to enable large batch sizes which can lead to better model quality. For distillation, AlphaGenome was trained on 64 NVIDIA H100 GPUs without sequence parallelism. The evaluation of all models was carried out on NVIDIA H100 GPUs without sequence parallelism. ### Software Training was done using [JAX](https://github.com/google/jax) and [JAXline](https://github.com/google-deepmind/jaxline). JAX allows researchers to take advantage of the latest generation of hardware, including TPUs, for faster and more efficient training of large models. JAXline is a distributed JAX training and evaluation framework. It is designed to be forked, covering only the most general aspects of experiment boilerplate. ## Evaluation AlphaGenome demonstrates state-of-the-art performance across a diverse set of genomic prediction tasks. The model was evaluated using two primary approaches: 1. **Genome Track Prediction**: Assessing the ability to predict functional genomic signals (read coverage) on previously unseen DNA sequences (held-out test intervals). 2. **Variant Effect Prediction (VEP):** Assessing the ability to predict the molecular consequences of genetic variants (e.g., single nucleotide variants) by comparing predictions for reference and alternative alleles against ground-truth datasets (e.g., experimental QTL effect sizes, readouts from reporter assays). Key Highlights: * **Broad SOTA Performance:** AlphaGenome matched or outperformed the best available external models on **22 out of 24** genome track prediction evaluations. * **Variant Interpretation:** For variant effect prediction, AlphaGenome matched or exceeded top-performing external models on **25 out of 26** evaluations. * **Multimodal Capability:** Unlike specialized models, AlphaGenome jointly predicts all assessed modalities – including splicing, expression, accessibility, and 3D contact maps – within a single framework. * **Single-Pass Efficiency:** The distilled student model achieves this performance and broad coverage with a single inference pass, eliminating the need for complex model ensembling. The tables below detail the performance metrics for specific modalities and tasks. ## Benchmark Results The following table focuses on the accuracy of the pre-trained, non-distilled model in predicting genomic tracks on unseen sequences: Modality | Evaluation | Metric | Resolution | Value | Baseline Model | Relative Improvement (%) :-------------------- | :------------------------------- | :---------- | :--------- | :---- | :------------- | :----------------------- **Splicing** | Splice site classification | auPRC | 1 bp | 0.79 | DeltaSplice | 1.0   | Splice site usage | Pearson r | 1 bp | 0.86 | DeltaSplice | 6.7 **RNA expression** | RNA-seq coverage | Pearson r | 1 bp | 0.59 | Borzoi | 28.2   | RNA-seq coverage | Pearson r | 32 bp | 0.78 | Borzoi | 4.6   | RNA-seq gene expr. LFC | Pearson r | Gene | 0.57 | Borzoi | 14.7   | CAGE coverage | Pearson r | 32 bp | 0.74 | Borzoi | 4.4   | CAGE coverage | Pearson r | 128 bp | 0.71 | Enformer | \-0.3   | Alternative PA | Spearman r | Gene | 0.87 | Borzoi | 13.1 **DNA accessibility** | DNase-seq coverage | Profile JSD | 1 bp | 0.51 | ChromBPNet | 6.4   | DNase-seq coverage | Pearson r | 32 bp | 0.86 | Borzoi | 4.7   | DNase-seq coverage | Pearson r | 128 bp | 0.87 | Enformer | 2.4   | ATAC-seq coverage | Profile JSD | 1 bp | 0.46 | ChromBPNet | 1.6   | ATAC-seq coverage | Pearson r | 32 bp | 0.57 | Borzoi | 3.4   | ATAC-seq coverage | Pearson r | 128 bp | 0.72 | Enformer | 2.7 **Histone mods** | Histone ChIP-seq coverage | Pearson r | 32 bp | 0.69 | Borzoi | 3.1   | Histone ChIP-seq coverage | Pearson r | 128 bp | 0.71 | Enformer | 2.4 **TF binding** | TF ChIP-seq coverage | Pearson r | 32 bp | 0.55 | Borzoi | 5.0   | TF ChIP-seq coverage | Pearson r | 128 bp | 0.58 | Enformer | 1.4 **DNA contact maps** | Orca Contact maps | Pearson r | 4000 bp | 0.79 | Orca | 6.3   | Orca Contact maps cell type diff | Pearson r | 4000 bp | 0.42 | Orca | 42.3 And the next table details the performance of the distilled model in predicting the functional effects of genetic variants: Modality | Evaluation | Type | Metric | Value | Baseline Model | Relative Improvement (%) :-------------------- | :---------------------------- | :-------- | :--------- | :---- | :------------- | :----------------------- **Splicing** | ClinVar splice site region | Causality | auPRC | 0.57 | Pangolin | 3.7   | ClinVar noncoding | \- | auPRC | 0.66 | Pangolin | 2.9   | ClinVar missense | \- | auPRC | 0.18 | DeltaSplice | 13.7   | Splicing outlier (zero-shot) | \- | auPRC | 0.22 | Pangolin | 59.1   | Splicing outlier (supervised) | \- | auPRC | 0.28 | AbSplice | 13.0   | sQTL | Causality | auPRC | 0.76 | Pangolin | 13.9   | MFASS | \- | auPRC | 0.51 | Pangolin | \-5.7 **RNA expression** | eQTL | Direction | Spearman r | 0.49 | Borzoi | 25.5   | eQTL (zero-shot) | Causality | auROC | 0.71 | Borzoi | 5.4   | eQTL (supervised) | Causality | auROC | 0.80 | Borzoi | 15.6   | ENCODE E2G (zero-shot) | \- | auPRC | 0.75 | Borzoi | 13.0   | paQTL | \- | auPRC | 0.63 | Borzoi | 7.3 **DNA accessibility** | CAGI5 MPRA | Causality | Pearson r | 0.65 | Borzoi | 6.3   | ds/caQTL | Direction | Pearson r | 0.70 | ChromBPNet | 7.7   | ds/caQTL | Causality | auPRC | 0.52 | Borzoi | 18.0 **TF binding** | bQTL | Direction | Pearson r | 0.55 | Borzoi | 2.8   | bQTL | Causality | auPRC | 0.50 | Borzoi | 6.0 ## Usage and Limitations ### License Unless required by applicable law or agreed to in writing, all software and materials distributed here (under the [Apache 2.0 License](https://github.com/google-deepmind/alphagenome_research/blob/main/README.md) with respect to model code and [non-commercial terms](https://deepmind.google.com/science/alphagenome/model-terms) for the model parameters) 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. ### Intended usage * Non-Commercial use only: The model parameters are restricted to non-commercial use by non-commercial organizations (e.g., universities, non-profits, research institutes, and journalism). It must not be used for any commercial activities or on behalf of commercial entities. * Model derivatives: If you fine-tune or modify this model, the resulting model is classified as a "Derivative". All derivatives are subject to these exact same terms, meaning strictly no commercial use is permitted for fine-tuned versions. * Distillation: Training a new model using the outputs or predictions of AlphaGenome is also restricted; resulting models must be governed by the AlphaGenome Model Parameters Terms of Use ### Limitations * Like other sequence-based models, accurately capturing the influence of very distant regulatory elements, like those over 100,000 DNA letters away, is still an ongoing challenge. Another priority for future work is further increasing the model’s ability to capture cell- and tissue-specific patterns. We haven't designed or validated AlphaGenome for personal genome prediction, a known challenge for AI models. Instead, we focused more on characterising the performance on individual genetic variants. And while AlphaGenome can predict molecular outcomes, it doesn't give the full picture of how genetic variations lead to complex traits or diseases. These often involve broader biological processes, like developmental and environmental factors, that are beyond the direct scope of our model.