Nature reports: Google's new model can understand DNA variations in one second. It unifies all genomic tasks for the first time, and its performance outshines existing models.
Google DeepMind's Alpha family has once again been featured in Nature. This time, it's targeting DNA mutations.
Now, it only takes one second to precisely locate genomic sequence variations.
According to the latest report in Nature, the Google DeepMind team has currently launched a breakthrough biological model, AlphaGenome.
It can simultaneously predict thousands of functional genomic features from a DNA sequence as long as 1 megabase and evaluate the mutation effects at single-base resolution.
It comprehensively outperforms existing models in various tasks such as gene expression, splicing, and chromatin accessibility, providing a powerful tool for deciphering the genomic regulatory code.
The authors described it as a milestone in the entire biological field:
For the first time, we have a single model that unifies long-range context, base precision, and state-of-the-art performance across the entire range of genomic tasks.
In the future, AlphaGenome will also help us better understand diseases, and the "mysterious book" of cancer may finally be deciphered.
This tool will provide a key piece of the puzzle, enabling us to establish better connections to understand diseases such as cancer.
The first single model to unify genomic tasks
Interpreting the impact of genomic sequence variations has always been a core challenge in the field of biology.
In the past decade, scientists have developed dozens of AI models separately to uncover the secrets of the genome. So, is it possible to create an "all-in-one" interpretation tool?
AlphaGenome is such a model that can unify multimodal prediction, long-sequence context, and base-pair resolution in a single framework.
Inspired by U-Net, the model architecture can process a 1-megabase DNA input sequence into two types of sequence representations during the downsampling stage: one is a one-dimensional embedding corresponding to the linear genome (at 1bp and 128bp resolution), and the other is a two-dimensional embedding corresponding to the spatial interactions of genomic fragments (at 2048bp resolution).
Inside the architecture, convolutional layers model local sequence patterns, while Transformer blocks, combined with Rotary position encoding, model coarser-grained but longer-range dependencies.
Through 8 interconnected tensor processing units, training at full base-pair resolution is achieved. Then, using encoder skip connections, the 1bp resolution of the sequence is restored during the upsampling stage.
Finally, it outputs 11 modalities, including gene expression, detailed splicing patterns, chromatin states, and chromatin contact maps, covering 5930 human or 1128 mouse genomic tracks.
The model is trained through two stages: pretraining and distillation:
- Pretraining: First, generate a fold-specific model and a full-fold model using observational data. The former is trained using a 4-fold cross-validation method to prevent overfitting, and the latter is trained on the entire available genomic interval and is regarded as the teacher model for subsequent distillation.
- Distillation: Train a single student model using randomly augmented input sequences to learn the output predictions of the full-fold teacher model, achieving stronger robustness and accuracy in predicting mutation effects.
Finally, on an NVIDIA H100 GPU, the inference time of the student model can be less than one second, showing extremely high efficiency.
Comprehensively outperforms existing technologies
To evaluate the generalization ability of AlphaGenome, the research team conducted 24 genomic track evaluations, comparing the AlphaGenome model with the strongest existing models for each task.
AlphaGenome leads in 22 of them. Compared with another multimodal sequence model, Borzoi3, it shows a +17.4% relative improvement in cell-type-specific LFC prediction.
In terms of predicting mutation effects, the experimental group assembled 26 mutation effect prediction benchmarks, including gene expression, splicing, polyadenylation, enhancer-gene connections, DNA accessibility, and transcription factor binding.
Compared with the strongest existing models, it reaches or exceeds in 24 benchmarks. For example, in the directional prediction of expression QTL, it shows a 25.5% improvement compared with Borzoi3, and in accessibility QTL, it shows an 8% improvement compared with ChromBPNet10.
The results show that AlphaGenome has advantages in both multimodal and specialized unimodal tasks and can accurately simulate genomic tracks and mutation effects.
In addition, AlphaGenome has also reached the state-of-the-art level in cross-modal genomic track prediction.
The pretrained fold-specific model shows that on unseen genomic intervals, the predicted read coverage is highly consistent with the observed read coverage.
Quantitatively, there is a strong Pearson correlation coefficient (r) between the predicted and observed signals of functional genomic tracks in the human and mouse genomes, indicating good overall expression level prediction.
In terms of the splicing modality, AlphaGenome achieves, for the first time, comprehensive prediction of splicing sites, splicing efficiency, and splicing connections, demonstrating a powerful ability to predict tissue-specific alternative splicing.
Based on the multi-dimensional splicing prediction of AlphaGenome, the experimental team designed a customized mutation scoring strategy for each prediction modality and summed up the individual scores to comprehensively consider the predicted mutation effects.
In the benchmark test of splicing-related mutation effect prediction (VEP), AlphaGenome performs best in the fine-mapping of splicing QTL (sQTL) classification and achieves the highest performance in both supervised and unsupervised scenarios.
In an experiment by MFASS to evaluate whether rare mutations disrupt splicing ability, AlphaGenome's auPRC reaches 0.54, slightly lower than Pangolin's 0.51 but higher than SpliceAI and DeltaSplice (both 0.49).
In general, AlphaGenome achieves state-of-the-art splicing mutation effect prediction in 6 out of 7 benchmark tests, providing a more comprehensive view of splicing event changes and transcript structures.
Clinically, AlphaGenome can help researchers more accurately understand the underlying causes of diseases and even discover new therapeutic targets.
For example, in a study on T-cell acute lymphoblastic leukemia (T-ALL), AlphaGenome successfully deciphered the oncogenic mutations near the TAL1 gene by introducing the MYB DNA binding motif.
In addition, it can help predict the design of synthetic DNA and assist in basic DNA research. In the future, by expanding the data, AlphaGenome will produce more accurate prediction accuracy and cover a wider range of species. Scientists will only need to make fine-tuning to generate and test hypotheses more quickly.
Currently, AlphaGenome has released a preview version and plans to be officially launched. Welcome to experience it first.
Link:
https://macro.com/app/pdf/56a50ffc-120d-4a9a-87e5-a18753430f22
Code link:
https://github.com/google-deepmind/alphagenome
Reference links:
[1]https://x.com/GoogleAI/status/1937895472305152387
[2]https://deepmind.google/discover/blog/alphagenome-ai-for-better-understanding-the-genome/
[3]https://www.nature.com/articles/d41586-025-01998-w
[4]https://www.science.org/doi/10.1126/science.1259037
This article is from the WeChat official account “Quantum Bit”. Author: Lu Yu. Republished by 36Kr with permission.