The AlphaGenome research team. From left to right: Pushmeet Kohli, VP, Science & Strategic Initiatives, Google DeepMind, Žiga Avsec, Research Scientist Google DeepMind, and Natasha Latysheva, Research Engineer.
A new deep learning model developed by researchers at Google DeepMind represents what experts are calling a major milestone in genomic artificial intelligence. AlphaGenome, published in Nature on 28 January 2026, can predict the functional consequences of DNA sequence variations across the vast non-coding regions of the genome that have long challenged scientists.
The human genome contains approximately three billion base pairs, yet only around two percent directly encode proteins. The remaining 98 percent – the non-coding regions – play crucial roles in regulating gene activity and harbour many variants linked to disease. Understanding how changes in these sequences affect biological function has remained one of biology’s most formidable challenges.
AlphaGenome addresses a fundamental limitation that has constrained previous computational approaches. Existing models have been forced to trade off between processing long DNA sequences and achieving fine-grained predictions. Models capable of capturing long-range genomic interactions, such as enhancer-promoter connections spanning hundreds of thousands of base pairs, typically produced outputs at resolutions of 32 to 128 base pairs – too coarse to detect subtle but important regulatory features.
“Deep learning models that predict functional genomic measurements from DNA sequences are powerful tools for deciphering the genetic regulatory code,” the authors write in the paper. “Existing methods involve a trade-off between input sequence length and prediction resolution, thereby limiting their modality scope and performance.”
AlphaGenome overcomes this constraint through a novel architecture that processes one million base pairs of DNA sequence whilst generating predictions at single-nucleotide resolution. The model simultaneously predicts 5,930 human or 1,128 mouse genomic signals across 11 distinct modalities, including gene expression, RNA splicing, chromatin accessibility, histone modifications, transcription factor binding, and three-dimensional chromatin contact maps.
The research team, led by Žiga Avsec, Natasha Latysheva, and Pushmeet Kohli, conducted extensive evaluations comparing AlphaGenome against the strongest available specialised models for each task. The results demonstrate consistent improvements across virtually all assessments.
For genome track prediction tasks – where models must accurately predict experimental measurements from DNA sequences alone – AlphaGenome outperformed external models on 22 of 24 evaluations. More critically for clinical applications, the model matched or exceeded performance on 25 of 26 variant effect prediction benchmarks.
Particularly notable improvements were observed in predicting expression quantitative trait loci (eQTLs), variants associated with gene expression changes. AlphaGenome achieved a 25.5 percent relative improvement in predicting the direction of eQTL effects compared to the previous best model, Borzoi. This enhanced accuracy translates to substantial practical gains: at a threshold yielding 90 percent sign prediction accuracy, AlphaGenome recovered more than twice as many GTEx eQTLs as Borzoi.
One of AlphaGenome’s most significant innovations is its ability to explicitly model splice junctions – the points where RNA molecules are cut and rejoined during processing. Many genetic diseases, including spinal muscular atrophy and certain forms of cystic fibrosis, result from splicing errors.
“One of the main ways genetic variants cause disease is by disrupting splicing, a process that produces mature RNA sequences by excising introns and ligating exons at splice junctions,” the authors explain. Previous models could predict splice sites but not the actual junctions formed or competition between alternative splicing outcomes.
The splice junction predictions proved particularly valuable for interpreting pathogenic variants. When tested against ClinVar, a database of clinically relevant genetic variants, AlphaGenome’s composite scores outperformed existing methods across all variant categories examined.
Independent researchers have responded enthusiastically to the development. Dr Robert Goldstone, Head of Genomics at the Francis Crick Institute, comments: “DeepMind’s AlphaGenome represents a major milestone in the field of genomic AI. This level of resolution, particularly for non-coding DNA, is a breakthrough that moves the technology from theoretical interest to practical utility, allowing scientists to programmatically study and simulate the genetic roots of complex disease.”
He adds: “The model performs exceptionally well on tasks that might be expected to be governed by rigid ‘grammatical’ rules written in the DNA, such as splice site prediction. In these areas, it is poised to immediately replace older standard tools.”
Professor Ben Lehner, Head of Generative and Synthetic Genomics at the Wellcome Sanger Institute, comments: “AlphaGenome is a great example of how AI is accelerating biological discovery and the development of therapeutics. Identifying the precise differences in our genomes that make us more or less likely to develop thousands of diseases is a key step towards developing better therapeutics.”
However, Lehner also sounds a note of caution: “AlphaGenome is far from perfect and there is still a lot of work to do. AI models are only as good as the data used to train them. Most existing data in biology is not very suitable for AI – the datasets are too small and not well standardised.”
To illustrate AlphaGenome’s utility for understanding disease mechanisms, the researchers analysed mutations affecting the TAL1 oncogene in T-cell acute lymphoblastic leukaemia. The model successfully predicted that cancer-associated mutations would activate TAL1 expression by creating binding sites for the MYB transcription factor – replicating experimentally verified disease mechanisms.
“The ability of AlphaGenome to simultaneously score variant effects across all modalities accurately recapitulates the mechanisms of clinically relevant variants near the TAL1 oncogene,” the authors note.
The model was trained on publicly available functional genomics data from major international consortia including ENCODE, GTEx, 4D Nucleome, and FANTOM5. These datasets span hundreds of human and mouse cell types and tissues, covering gene expression, splicing, chromatin accessibility, histone modifications, transcription factor binding, and three-dimensional chromatin conformation.
The researchers acknowledge important constraints. Accurately capturing the influence of regulatory elements more than 100,000 base pairs distant remains challenging. The model has not been validated for personal genome prediction, and its predictions of molecular consequences do not directly translate to complex trait or disease phenotypes.
Professor Rivka Isaacson, Professor of Molecular Biophysics at King’s College London, comments: “This work is an exciting step forward in illuminating the ‘dark genome’. We still have a long way to go in understanding the lengthy sequences of our DNA that don’t directly encode the protein machinery whose constant whirring keeps us healthy.”
The researchers designed AlphaGenome’s architecture to be extendable with additional data and new assay types. Potential applications include rare disease diagnostics through improved interpretation of variants of uncertain significance, synthetic biology applications such as tissue-specific enhancer design, identification of cancer driver mutations in non-coding regions, and RNA therapeutics development.
Professor Marc Mansour of University College London comments: “AlphaGenome will be invaluable to researchers studying non-coding variants in any disease-trait association, synthetic biology, and functional genomics.”
The authors conclude: “AlphaGenome provides a powerful and unified model for analysing the regulatory genome. It advances our ability to predict molecular functions and variant effects from DNA, offering valuable tools for biological discovery and enabling applications in biotechnology.”
AlphaGenome is now available for non-commercial research use via an API, with plans for full model release to follow. < https://deepmind.google.com/science/alphagenome/ >
Avsec, Ž., Latysheva, N., Cheng, J., et al. (2026). Advancing regulatory variant effect prediction with AlphaGenome. Nature. https://doi.org/10.1038/s41586-025-10014-0
Prins Hendrikstraat 1
5611HH Eindhoven
The Netherlands
info@clinlabint.com
PanGlobal Media is not responsible for any error or omission that might occur in the electronic display of product or company data.
