1. Background and research history
For many readers, DeepMind’s “Alpha” story began with a Go board. In March 2016, AlphaGo defeated champion Lee Sedol 4-1. AlphaZero followed in 2017, learning chess, shogi and Go through self-play, starting from the rules of each game.
The work soon expanded into biology. AlphaFold 2 made a major advance in predicting proteins’ three-dimensional structures in 2020. AlphaFold 3, developed with Isomorphic Labs and introduced in 2024, extended that work to complexes involving proteins, DNA, RNA and other molecules. The scope has since expanded even further, including AlphaMissense (launched in 2023), predicting whether a change to a protein’s amino-acid sequence is likely to be harmful, and AlphaEarth (launched in 2025), combining Earth-observation data to support mapping and environmental monitoring.
AlphaGenome tackles another puzzle: we can read DNA, but what do changes in it actually do? A change outside a protein-coding region can still alter gene activity or disrupt how RNA is assembled. Introduced in June 2025, AlphaGenome predicts these molecular effects, helping scientists decide which clues to investigate next.
Its immediate promise is better-informed experiments; translating predictions into diagnoses or drugs still requires validation. This report traces the research and earlier models behind AlphaGenome, explains what makes it different, and examines what it could mean for biotech startups, pharmaceutical companies and genomics suppliers.
DNA illustration. Image Source: Google DeepMind
1.1 The biology behind the question
A genome is an organism’s complete set of DNA. In humans, it is organized into 23 pairs of chromosomes - 22 pairs of autosomes and one pair of sex chromosomes (usually XX or XY). These chromosomes carry thousands of genes, stretches of DNA that contain instructions for making proteins or functional RNA molecules.
Different genes contribute to different traits and biological functions. For example, OCA2 and HERC2 on chromosome 15 are associated with eye color and pigmentation; CFTR on chromosome 7 is linked to cystic fibrosis; and ADH1A on chromosome 4 is part of a gene family involved in alcohol metabolism.
Different Eye Colors decided by OCA2 and HERC2 genes. Image Source: Wikipedia
But knowing where genes are located is only part of the picture. Cells copy DNA into RNA, which can then be used to produce proteins or perform other functions. Much of the genome does not directly encode proteins, yet some of these noncoding regions help control when, where and how strongly genes are expressed.
Functional genomics studies how these sequences work and interact. Its central challenge is understanding not just what DNA contains, but how changes in its sequence affect biological activity, the problem AlphaGenome aims to address.
1.2 How researchers learned to read and interpret DNA
Before researchers could understand what DNA does, they first had to read it. In 1977, Frederick Sanger and colleagues, and independently Allan Maxam and Walter Gilbert, developed practical methods for determining DNA sequences. Sanger's method eventually became the standard, laying the foundation for sequencing entire genomes.
A major milestone came with the Human Genome Project (1990-2003), an international effort to map the roughly three billion DNA bases in the human genome. It produced the first human reference genome, giving scientists a common map for locating genes and comparing genetic differences. But the reference still contained gaps, and knowing the sequence did not explain what most of it actually did.
Robert Waterston presenting a detailed map of human chromosome 7 at a 2001 press conference announcing the first analyses of the draft human genome sequence. Image Source: National Human Genome Research Institute
That question helped drive ENCODE (the Encyclopedia of DNA Elements), launched in 2003. Building on the Human Genome Project, ENCODE aimed to identify functional elements across the genome, particularly regions involved in controlling gene activity. By measuring where proteins bind DNA, which regions are accessible, and where RNA is produced, it created large public datasets that helped researchers understand how the genome operates.
Meanwhile, researchers were connecting genetic variation to disease. Genome-wide association studies (GWAS), which expanded in the mid-2000s, compared DNA variants across large populations to identify associations with traits and diseases. But these associations often could not identify the responsible variant or explain the underlying biological mechanism. The NIH's Genotype-Tissue Expression (GTEx) project, launched in 2010, added another layer by mapping how genetic differences influence gene activity across human tissues, from blood to brain.
Together with advances in next-generation sequencing, these efforts transformed genomics from a reference sequence into vast datasets linking DNA to biological activity. They also created a new opportunity: train AI models on DNA sequences paired with experimental measurements, then use them to predict how genetic changes might affect gene regulation and other molecular processes. The challenge was learning enough biological context to make those predictions accurate and useful, especially for changes in noncoding DNA whose effects are often difficult to interpret.
Chapter 2. Earlier models and their contributions
As genomic datasets grew, researchers began using deep learning to predict how DNA sequences influence biological activity. Different models tackled different aspects of this challenge, from identifying regulatory patterns and predicting RNA splicing to understanding how distant DNA regions influence genes. Together, they laid the foundation for AlphaGenome, while leaving room for a model that could combine broader biological predictions with longer sequence context and finer detail.
Table 1. Key AI models preceding AlphaGenome. © 2026 Jarsy Research
Beyond their technical advances, these models changed how scientists investigate genetic variants. DeepSEA helped researchers prioritize potentially important noncoding variants identified through GWAS, moving beyond statistical associations toward possible biological explanations. SpliceAI revealed how previously overlooked mutations could create abnormal RNA splice sites, potentially contributing to rare genetic disorders. Enformer introduced transformer-based attention to identify relationships between distant regulatory regions and genes, while Borzoi showed how RNA activity and processing could vary across tissues, capturing information that a single gene-expression measurement might miss. AlphaMissense took a different approach, building on AlphaFold and evolutionary information to evaluate roughly 71 million possible protein-altering variants, providing a large-scale resource for studying genetic disease.
Borzoi: a neural network for predicting RNA-seq coverage from sequence. Image Source: Nature.com
Chapter 3. AlphaGenome's innovations and applications
Introduced in June 2025, AlphaGenome, developed by Google DeepMind, builds on earlier sequence-to-function models to predict a broader range of biological activities directly from DNA sequences.
The model can analyze roughly one million DNA letters at once and predict thousands of experimental measurements across human and mouse tissues and cell types. These include gene expression, DNA accessibility, protein binding, RNA splicing and interactions between distant genomic regions.
3.1 What makes AlphaGenome different?
AlphaGenome's main innovation is combining longer DNA context, finer prediction resolution and broader biological coverage within a single model.
A bin groups neighboring DNA positions into one prediction. Longer input provides more context but does not guarantee better accuracy, and AlphaGenome's output resolution varies by task.
RNA splicing illustrates the advance. SpliceAI already predicted splice sites, while Borzoi captured splicing-related signals through RNA coverage. AlphaGenome additionally predicts which splice sites connect and how strongly, alongside other regulatory measurements. This helps researchers develop more specific explanations for how DNA changes might disrupt RNA processing.
Table 2: Comparison of Enformer, Borzoi, and AlphaGenome. Data Source: Nature.com. © 2026 Jarsy Research
3.2 How much better is it?
In the final Nature study, AlphaGenome outperformed the strongest external comparison models on 22 of 24 genomic-track evaluations, and matched or exceeded them on 25 of 26 variant-effect evaluations.These tests covered tasks including gene expression, DNA accessibility and RNA splicing.
The results demonstrate its strength as a broad research model, but do not imply clinical diagnostic accuracy or superiority across every task. Specialized models remain competitive in some areas, and experimental validation is still necessary.
AlphaGenome model architecture, training regimes and comprehensive evaluation performance. Image Source: Nature.com
3.3 AlphaGenome Atlas: From prediction to practical research
In September 2026, Google DeepMind introduced AlphaGenome Atlas, a catalogue of precomputed predictions for roughly nine billion possible single-letter DNA changes. It also introduced the AlphaGenome Variant Impact (AVI) score, combining predictions from AlphaGenome and AlphaMissense to help researchers prioritize potentially important variants without running each calculation themselves.
One early application involved DNM1, a gene associated with severe neurological disorders. Researchers at the Broad Institute, working with the GREGoR consortium, used Atlas to identify an overlooked variant that AlphaGenome predicted would create an abnormal RNA splice site and extend the resulting protein. Follow-up experiments supported the predicted mechanism, demonstrating how AI can help turn uncertain genetic findings into testable biological explanations.
3.4 What remains difficult?
Despite its advances, AlphaGenome still faces important scientific limitations. Long-range gene regulation remains difficult to predict, and the model can miss differences between tissues and cell types. More importantly, predicting a molecular effect does not establish that a variant causes disease. Experimental validation remains essential, and AlphaGenome is not intended for direct clinical decision-making.
Commercialization also presents challenges. While research code and model weights are available, the pretrained weights carry non-commercial restrictions, with commercial access offered separately through Google Cloud. Companies building products around AlphaGenome must therefore consider licensing, computing costs and integration requirements. AlphaGenome's near-term value lies in helping researchers identify promising genetic variants, understand possible mechanisms and prioritize experiments, not replacing biological or clinical validation.
4. Impact and outlook
AlphaGenome could influence several parts of the genomics industry, particularly drug discovery, RNA therapeutics and genetic interpretation. Its main opportunity is to help researchers understand the effects of DNA changes more efficiently, potentially improving target selection and reducing unnecessary experiments. However, the commercial impact will depend on whether these predictions translate into better research decisions, validated discoveries and ultimately successful products.
4.1 Which industries could benefit most?
Table 3: Potential industry applications of AlphaGenome. Data Source: Nature.com © 2026 Jarsy Research
The strongest near-term opportunities are in research workflows where AlphaGenome can help narrow thousands of candidate variants or targets into a smaller set for experimental testing. Benefits in diagnostics, drug development and synthetic biology will require further validation.
4.2 Which companies are most exposed?
Private companies: Direct exposure to genomic research and drug discovery
Genomics-focused startups such as Nucleome Therapeutics, Deep Genomics and Variant Bio are among the most directly exposed, although in different ways. Nucleome studies noncoding DNA using proprietary 3D genomics data, which its CTO has suggested could complement AlphaGenome's predictions. Deep Genomics develops RNA therapeutics using its proprietary BigRNA model, making AlphaGenome both a potential research tool and a competitive benchmark. Variant Bio combines human genetic diversity and biological measurements to discover drug targets, where AlphaGenome could help explain the mechanisms behind genetic associations.
For these companies, the opportunity lies in combining AI predictions with proprietary data and experiments, while the risk is that increasingly accessible models make some computational capabilities less distinctive. Variant Bio's partnerships with Novo Nordisk and Boehringer Ingelheim, potentially worth hundreds of millions of dollars, demonstrate pharmaceutical demand for genetics-led discovery, although none of these cases establishes commercial adoption of AlphaGenome.
Public companies: Different opportunities across the value chain
For larger public companies, AlphaGenome's potential effects are more varied. Regeneron and Novo Nordisk could benefit from stronger genetic evidence in early drug discovery, although the financial impact would likely take years to emerge. Illumina may benefit indirectly if better interpretation increases the value of sequencing, while Tempus AI could explore applications in genomic research and variant analysis, subject to clinical-use restrictions.
Alphabet has a more direct role as AlphaGenome's developer and commercial provider. Google Cloud offers paid enterprise access, creating a potential revenue opportunity from model usage and computing infrastructure. However, AlphaGenome-specific revenue has not been disclosed, and its financial contribution to Alphabet is likely to remain small relative to the group's overall business in the near term.
Table 4. Key Companies: AlphaGenome Exposure, Competitive Positioning and Financial Overview. Data Sources: PitchBook.com, Tracxn.com, company websites. © 2026 Jarsy Research
4.3 Where could economic value emerge?
AlphaGenome's commercial impact remains unproven, but the potential is meaningful. A 2024 Nature study found that drug mechanisms supported by human genetic evidence were 2.6 times more likely to succeed clinically, highlighting the value of better genetic insights in drug discovery. However, this does not measure AlphaGenome's own impact.
Near-term benefits could include faster target selection, fewer unnecessary experiments and improved genetic interpretation. Longer-term value will depend on turning predictions into validated discoveries and successful medicines. As genomic AI becomes more accessible, companies with proprietary datasets, strong experimental capabilities and drug-development expertise may be best positioned to capture lasting value.
4.4 What's next for AlphaGenome?
Following the September 2026 release of AlphaGenome Atlas, Google DeepMind is moving toward more comprehensive genomic interpretation. Key challenges include improving long-distance gene regulation predictions, tissue-specific accuracy and understanding how regulatory elements interact. Advances in these areas could help researchers better explain how genetic changes affect biological processes.
The longer-term ambition is to move beyond predicting which DNA changes matter toward understanding why they matter and how their effects might be corrected. If successful, future models could help accelerate drug discovery, rare-disease research and even the design of DNA with desired regulatory functions. The next major breakthrough would be turning genomic predictions into reliable, experimentally validated discoveries at scale.
Further reading:
