Target Audience
- Readers interested in the evolution of machine learning in genomics and the latest developments in genomic AI.
- Researchers and technology enthusiasts exploring biological simulation and the prediction of molecular functions directly from DNA sequences.
- Professionals interested in applying genomic AI to personalized medicine, rare disease diagnosis, and drug discovery.
Executive Summary
- AlphaGenome Atlas represents a major step in transforming machine learning-based genomic function prediction into a genome-wide scientific resource. It precomputes the predicted molecular effects of approximately 9 billion possible single-nucleotide substitutions across the human genome.
- The evolution from DeepSEA to Enformer, Borzoi, and AlphaGenome has dramatically expanded the range of biological functions, genomic distances, and molecular processes that can be predicted directly from DNA sequences.
- AlphaGenome Atlas connects variant prioritization with mechanistic interpretation, potentially accelerating rare disease research, personalized medicine, and the investigation of disease-associated genetic variants.
- However, alternative approaches such as Evo 2 and GPN-Star offer complementary strengths. AlphaGenome Atlas is not necessarily the best-performing approach for every variant prediction task.
- The next generation of genomic AI will likely integrate molecular function prediction, evolutionary constraints, and cellular context, moving toward increasingly comprehensive biological simulation.
1. Where Does AlphaGenome Atlas Fit in the History of Genomic Machine Learning?
Published on September 8, 2026, AlphaGenome Atlas: in silico mutagenesis of the entire human genome improves prioritization and interpretation of non-coding variants represents an important milestone in the evolution of machine learning-based genomics.[1]
To understand its significance, it is essential to distinguish AlphaGenome from AlphaGenome Atlas.
AlphaGenome is a predictive model, whereas AlphaGenome Atlas applies that model to single-nucleotide substitutions across the entire human genome, transforming variant effect prediction into a searchable map.
AlphaGenome Atlas precomputes predictions for approximately 9 billion possible single-nucleotide substitutions in the human genome and makes them available through a database containing approximately 1 PB (petabyte) of data.
It also introduces the AlphaGenome Variant Impact (AVI) score, which integrates predictions of non-coding regulatory effects with protein-level impact predictions from AlphaMissense. This enables both variant prioritization and interpretation of underlying molecular mechanisms.[1]
To understand where this research fits into the broader field, it is useful to review the history of genomic machine learning.
At a high level, the progression can be summarized as follows:
Annotating variants with known genomic features ↓ Predicting genomic annotations directly from DNA sequences ↓ Predicting long-range gene regulation ↓ Predicting multiple molecular phenotypes with a unified model ↓ Precomputing predictions for every possible variant across the genome
AlphaGenome Atlas represents the final stage of this progression.
First Generation: Handcrafted Features and Classical Machine Learning
Early approaches to variant prioritization relied heavily on existing genomic annotations.
Researchers collected information such as evolutionary conservation scores (PhyloP and PhastCons), transcription factor binding sites, histone modifications, DNase hypersensitivity, and expression quantitative trait loci (eQTLs).
These features were then used in statistical or classical machine learning methods, including support vector machines (SVMs) and random forests, to predict whether a variant might be pathogenic.
Representative methods included CADD, GWAVA, and Eigen.
The basic workflow was:
Variant → Existing annotations → Pathogenicity prediction
In other words, these methods predicted the consequences of variants using previously observed and curated genomic information.
A major limitation was that prediction performance could be weaker in genomic regions or cell types where relevant annotations were unavailable.
Second Generation: DeepSEA — Reading Regulatory Function Directly from DNA
A major turning point came with DeepSEA in 2015.[2]
DeepSEA used convolutional neural networks (CNNs) to predict hundreds of chromatin-related features directly from DNA sequences, including transcription factor binding, DNase sensitivity, and histone modifications.
It also estimated variant effects by comparing predictions for the reference allele and the alternative allele.
This established a fundamental workflow:
DNA sequence ↓ Deep learning ↓ Chromatin state ↓ Variant effect
This approach helped establish the sequence-to-function paradigm that remains central to modern genomic AI.
If we were to identify one direct technological ancestor of AlphaGenome Atlas, DeepSEA would arguably be more relevant than AlphaFold.
Third Generation: Basenji and Enformer — Modeling Distant Enhancers
The next major challenge was the length of genomic sequence context that models could process.
Gene expression is not determined solely by a few hundred base pairs surrounding a gene.
Enhancers located tens or hundreds of kilobases away can strongly influence gene expression.
The Basenji family of models helped address this problem, with Enformer, published in 2021, representing a particularly important advance.[3]
Enformer processes approximately 196 kb of DNA sequence and uses the attention mechanism of the Transformer architecture to model long-range interactions.
For the human genome, it predicts more than 5,000 functional genomic tracks.
This development changed the questions researchers could ask.
Previously, the question was often:
Does this variant disrupt a transcription factor binding motif?
With long-range sequence-to-function models, the question evolved into:
How does the effect of this variant propagate to the expression of a gene located far away?
This represented a transition from predicting local DNA functions to modeling broader gene regulatory systems.
Fourth Generation: Sei and Borzoi — Integrating Functions and Predicting RNA Directly
In 2022, Sei introduced a model capable of predicting 21,907 chromatin profiles and organizing them into approximately 40 sequence classes.[4]
These classes provided a more interpretable representation of regulatory functions, including enhancers, promoters, CTCF-associated regions, and cell-type-specific regulatory activity.
Meanwhile, Borzoi approached the problem from a different direction.[5]
Its defining feature was the direct prediction of RNA-seq coverage from DNA sequences.
DNA → RNA-seq coverage
Borzoi processes DNA sequences of approximately 524 kb and predicts not only transcriptional activity but also variant effects related to splicing and polyadenylation.
By this stage, two critical directions leading toward AlphaGenome had emerged:
Modeling long genomic contexts and integrating multiple regulatory modalities.
2. What Did AlphaGenome Change?
Introduced in 2025 and published in Nature in 2026, AlphaGenome substantially unified the capabilities of sequence-to-function models.[6]
AlphaGenome accepts DNA sequences of up to 1 Mb and simultaneously predicts 5,930 tracks across 11 modalities for the human genome.
Its predicted outputs include:
- RNA-seq
- CAGE
- PRO-cap
- Splicing
- DNase / ATAC
- Histone modifications
- Transcription factor binding
- Chromatin contacts
Its major achievement is the combination of:
Long context × Base-level resolution × Multimodal prediction
Previous models had different areas of specialization:
- Enformer was particularly effective at modeling long-range regulatory effects.
- Borzoi was designed for predicting RNA processing-related outcomes.
- ChromBPNet focused on chromatin accessibility.
- Specialized splice models focused on splicing predictions.
AlphaGenome moves toward integrating these different prediction targets into a single sequence-to-function model.
In the original AlphaGenome paper, the model matched or exceeded the performance of the best existing methods in 25 out of 26 variant effect prediction evaluations.[6]
This advancement is also significant from the perspective of biological simulation.
A single DNA sequence can now serve as the common input for predicting multiple molecular processes, including transcription, chromatin states, and splicing.
However, this does not mean that AlphaGenome can simulate an entire living organism.
Its primary focus remains the relationship between DNA sequences and molecular functions.
3. The Real Innovation of AlphaGenome Atlas
The central innovation of AlphaGenome Atlas is not necessarily a new neural network architecture.
Instead, it is the transformation of a variant-effect prediction model into a genome-wide scientific resource.[1]
How It Differs from Conventional Variant Effect Prediction
Traditionally, a typical workflow looked like this:
Identify candidate variants through GWAS / WGS ↓ Input variants into Enformer / Borzoi / AlphaGenome ↓ Run inference for each candidate variant
Researchers first identified variants of interest and then used a predictive model to estimate their functional consequences.
AlphaGenome Atlas changes this workflow by precomputing predictions for approximately 9 billion single-nucleotide substitutions.
Precompute predictions for approximately 9 billion SNVs ↓ Identify variants through GWAS / WGS ↓ Retrieve existing predictions through lookup
This changes how researchers can use variant effect predictions.
Instead of running a computationally intensive model for each new analysis, researchers can retrieve previously calculated results from a database.
The Connection to AlphaFold Database
This relationship closely resembles that between AlphaFold and AlphaFold Database.
AlphaGenome : AlphaGenome Atlas ≈ AlphaFold : AlphaFold Database
Just as AlphaFold helped transform protein structure prediction into a widely accessible scientific resource, AlphaGenome Atlas makes large-scale genomic variant effect predictions broadly available.
In other words, it represents a shift:
From an AI model to scientific infrastructure.
This is arguably the most important way to understand the historical significance of AlphaGenome Atlas.
4. Why Is In Silico Mutagenesis Important?
The term in silico mutagenesis (ISM) appears in the paper's title and is central to understanding the approach.
However, ISM itself is not a new methodology.
It is a computational technique in which individual nucleotides in a DNA sequence are systematically substituted, and the resulting changes in model predictions are measured.
For example, if a particular position contains an adenine (A), the following substitutions can be evaluated:
- A → C
- A → G
- A → T
The model then calculates the difference between the predicted outputs for the reference and alternative sequences.
[ \Delta f = f(\text{ALT sequence}) - f(\text{REF sequence}) ]
Here, (f) represents a model that predicts molecular functions from DNA sequences.
The difference in predictions provides an estimate of how a variant may affect transcription factor binding, gene expression, or other molecular processes.
Variants have been evaluated using this general principle since the DeepSEA era.[2]
AlphaGenome also uses ISM to visualize effects on features such as splicing motifs.[6]
Precomputing Approximately 9 Billion Variants
What makes AlphaGenome Atlas remarkable is that this technique was applied across the entire human genome.
The human genome contains approximately 3 billion base pairs.
For each position, there are three possible substitutions to a different nucleotide.
Therefore:
Approximately 3 billion bases × 3 alternative alleles ≈ 9 billion variants
This corresponds to systematically evaluating nearly every possible single-nucleotide substitution across the human genome.
The breakthrough is therefore less about inventing an entirely new computational method and more about scaling an established method to the entire genome.
It is a revolution in scale.
5. Why the AlphaGenome Variant Impact (AVI) Score Matters
Another important contribution of AlphaGenome Atlas is the AlphaGenome Variant Impact (AVI) score.[1]
AlphaGenome produces a large number of molecular function predictions.
For a single variant, the predicted effects might include:
- Decreased chromatin accessibility
- Reduced transcription factor binding
- Reduced H3K27ac
- Reduced transcription
- Altered splicing
- Changes in chromatin contact maps
These predictions are valuable for understanding molecular mechanisms.
However, large-scale genome analysis presents a practical challenge.
Whole-genome sequencing can identify enormous numbers of variants, and researchers must determine which ones deserve further investigation.
The critical question becomes:
Which variants should we investigate first?
AVI is designed to integrate multiple prediction outputs into a unified framework for prioritizing variants.
It also incorporates AlphaMissense predictions for coding variants, with the goal of enabling variant ranking across both coding and non-coding regions.
In addition, feature attribution makes it possible to interpret why a variant receives a high impact score by examining contributions from splicing, gene expression, chromatin features, and other molecular functions.
This connects two previously distinct approaches.
The first is the CADD-like approach, in which each variant is assigned a single impact or pathogenicity-related score.
The second is the AlphaGenome approach, which predicts the molecular mechanisms that a variant may affect.
AlphaGenome Atlas attempts to combine these capabilities.
It connects variant prioritization with mechanistic interpretation.
6. How AlphaGenome Atlas Connects to Personalized Medicine and Current Genomic Analysis
Modern WGS and GWAS analysis often follows a workflow similar to the following:
WGS / GWAS / Fine mapping ↓ Candidate variant identification ↓ AI-based variant prioritization ↓ Prediction of cell- or tissue-specific molecular effects ↓ Gene / pathway / mechanism inference ↓ Experimental validation using CRISPR / MPRA / patient RNA-seq
AlphaGenome Atlas can shorten the process from variant prioritization to the generation of mechanistic hypotheses.
Identifying Causal Variants in Rare Diseases
One particularly important application is rare disease research.
In an example involving the Broad Institute highlighted by Google DeepMind, variant prioritization for an unresolved disease identified a candidate variant near DNM1.[1]
AlphaGenome predicted that this variant would create a new splice site.
The prediction was subsequently confirmed experimentally.
This illustrates a workflow such as:
High AVI score ↓ Prediction of a potential splicing abnormality ↓ Validation using RNA assays
The significance extends beyond identifying variants that may be pathogenic.
Researchers can also generate specific hypotheses about why a variant may cause a molecular dysfunction.
This is particularly important for personalized medicine.
Among the many variants identified in an individual patient's genome, researchers may be able to prioritize variants associated with the disease and design experiments to investigate their molecular consequences.
Implications for Drug Discovery
The ability to predict the functional consequences of genetic variants is also relevant to drug discovery.
GWAS and other genetic studies can identify genomic regions associated with diseases.
However, identifying an associated region does not necessarily reveal which variant is responsible or which gene or regulatory mechanism connects the variant to the disease.
Technologies such as AlphaGenome Atlas may help generate hypotheses about these mechanisms by predicting effects on transcription, splicing, chromatin states, and other molecular functions.
These predictions could support research into potential drug targets and help prioritize experiments for validating disease mechanisms.
However, predicting a molecular effect is not equivalent to predicting drug efficacy or clinical outcomes.
This distinction remains one of the major limitations of current genomic AI.
7. Is AlphaGenome a Genomic Version of an LLM?
Modern genomic AI can be broadly divided into two major families.
One consists of sequence-to-function models such as AlphaGenome.
The other consists of DNA language models and genomic foundation models.
The two families differ in their training objectives and the types of information they learn.
Sequence-to-Function Models
Representative models follow this general progression:
DeepSEA → Basenji → Enformer → Borzoi → AlphaGenome
These models primarily use experimental functional genomics data as training signals, including:
- RNA-seq
- ATAC-seq
- ChIP-seq
- CAGE
- Hi-C
Their main objective is to predict biological functions from DNA sequences.
In simplified terms:
DNA → Biological function
AlphaGenome Atlas belongs to this family.
DNA Language Models and Genomic Foundation Models
The second family includes models such as:
- DNABERT
- Nucleotide Transformer
- HyenaDNA
- Caduceus
- Evo / Evo 2
- GPN-family models
These approaches use large-scale DNA sequence datasets and self-supervised learning to capture statistical and biological patterns in genomic sequences.
Their broader objective is to learn representations of what makes DNA sequences biologically meaningful or plausible.
For example, Nucleotide Transformer was trained using 3,202 human genomes and genomes from 850 species to learn general-purpose DNA representations.[7]
Evo 2, published in 2026, takes this approach to a much larger scale.[8]
It was trained on 9 trillion base pairs across all domains of life and supports a context length of 1 million tokens.
It can also predict certain variant effects without task-specific fine-tuning.
The distinction can be summarized as follows:
AlphaGenome: A supervised, biological-simulator-oriented model for predicting molecular functions.
Evo 2: A general-purpose genomic language/foundation model trained on massive DNA sequence datasets.
AlphaGenome focuses more directly on predicting molecular processes, while Evo 2 emphasizes learning broadly applicable patterns from DNA sequences.
8. AlphaGenome Is Not Necessarily the Best Model for Every Task
This is an especially important consideration when evaluating the significance of AlphaGenome Atlas.
A 2025 study comparing DNA foundation models found that sequence-to-function models such as AlphaGenome, Enformer, and Sei performed strongly on cell- and tissue-specific regulatory variant prediction tasks, including QTL-related evaluations.[9]
For example, the following performance values were reported for AlphaGenome:
- sQTL: AUC ≈ 0.80
- ipaQTL: AUC ≈ 0.86
However, on a different task involving the classification of pathogenic variants versus common variants, DNA foundation models such as Nucleotide Transformer performed better in some cases.
This suggests that the two model families learn different types of biological information.
Sequence-to-function models are particularly suited to questions such as:
Does this variant alter gene expression in a specific tissue?
Language and evolutionary models may be particularly useful for questions such as:
How unusual or biologically constrained is this sequence change from an evolutionary perspective?
The two approaches should therefore not be viewed simply as competitors.
They offer complementary capabilities.
9. GPN-Star: An Important Comparison from 2026
Among recent related studies, GPN-Star is an especially important comparison.
GPN-Star was published in Nature in September 2026, around the same time that AlphaGenome Atlas was released.[10]
Unlike AlphaGenome, which directly predicts functional genomics measurements, GPN-Star models functional constraints using evolutionary information and genomic sequence patterns.
In benchmarks evaluating pathogenic non-coding variants, GPN-Star reportedly outperformed AlphaGenome, Borzoi, and Enformer in evaluations involving OMIM and HGMD variants.
This result is important when interpreting the role of AlphaGenome Atlas.
The ability to predict molecular mechanisms does not necessarily imply the highest accuracy in predicting pathogenicity.
For example, a variant may be predicted to affect transcription factor binding or gene expression.
However, whether that molecular effect actually causes disease is a separate question.
Therefore, even in 2026, AlphaGenome Atlas does not solve every problem related to variant pathogenicity.
Models such as GPN-Star, which incorporate evolutionary constraints, remain important alternatives and complementary technologies.
10. Comparison of Major Genomic AI Models
The major developments can be summarized as follows.
| Model family | Representative model | Primary learning target | Main strength | Relationship to AlphaGenome Atlas |
|---|---|---|---|---|
| Regulatory CNN | DeepSEA | Chromatin assays | Predicting regulatory effects of variants | Direct technological ancestor |
| Regulatory deep learning | Sei | Approximately 21,000 chromatin tracks | Interpretable regulatory sequence classes | Earlier genome-wide functional mapping approach |
| Long-range sequence-to-function | Enformer | Epigenomic features and gene expression | Long-range enhancer–gene interactions | Major predecessor of AlphaGenome |
| RNA sequence-to-function | Borzoi | RNA-seq and related measurements | Expression, splicing, and APA | Direct comparison model |
| Unified sequence-to-function | AlphaGenome | Multiple functional genomics assays | 1 Mb context, multimodal prediction, base-level resolution | Foundation model underlying Atlas |
| Genome-wide resource | AlphaGenome Atlas | AlphaGenome predictions | Rapid lookup of approximately 9 billion SNVs | Genome-wide variant-effect atlas |
| DNA foundation model | Nucleotide Transformer | DNA sequences | General-purpose sequence representations | Complementary technology |
| Generative genomic foundation model | Evo 2 | 9 trillion base pairs | Sequence prediction and generation | More general-purpose AI direction |
| Evolutionary variant effect prediction | GPN-Star | Evolutionary and genomic constraints | Pathogenicity prediction | Strong complementary and competing approach |
| Cell-state sequence-to-function | scooby | Single-cell RNA / ATAC | Cell-specific variant effects | Focuses on cellular states |
| Cell-context model | GET | Regulation across multiple cell types | Generalization to unseen cell types | Addresses a limitation of AlphaGenome |
This comparison illustrates that genomic AI is not progressing along a single path.
Researchers are advancing several complementary directions, including molecular function prediction, evolutionary constraint modeling, cellular context awareness, and general-purpose DNA representation learning.
11. scooby and GET Point Toward a Different Next Step
One limitation of AlphaGenome is its dependence on the cellular and tissue contexts represented in its training data.
In living organisms, the same DNA sequence can exhibit different transcriptional and chromatin states depending on the cell type and its physiological condition.
Models such as scooby and GET are addressing this challenge.
scooby: Single-Cell Variant Effect Prediction
Published in 2025, scooby builds on Borzoi while incorporating single-cell RNA-seq and ATAC-seq data.[11]
This enables the prediction of cell-state-specific variant effects.
Earlier models often focused on molecular functions averaged across a tissue or cell population.
scooby moves toward representing differences between finer-grained cellular states.
GET: Generalization to Unseen Cell Types
GET (General Expression Transformer) uses information from 213 human cell types and aims to generalize predictions across different cellular contexts.[12]
This represents an important direction: predicting regulatory activity in cell types or conditions that were not directly represented during training.
Related work on context-aware sequence-to-function modeling published in 2026 also addresses the difficulty of extrapolating predictions to unseen cell types and experimental conditions.[13]
These developments suggest that future models will increasingly need to consider not only DNA sequences but also the cellular environment in which those sequences operate.
A possible framework is:
DNA sequence + Cell state + TF environment + Epigenetic context
This represents a transition from predicting molecular function based solely on DNA sequence toward simulations that more closely reflect the context-dependent behavior of living systems.
12. AlphaGenome Atlas as the "AlphaFold Database Moment" for Sequence-to-Function Genomics
The historical significance of AlphaGenome Atlas can be summarized as:
The AlphaFold Database moment for sequence-to-function genomics.
DeepSEA demonstrated that:
Variant effects can be predicted directly from DNA sequences.
Enformer extended this capability:
Long-range regulatory interactions can also be modeled.
Borzoi and AlphaGenome advanced the field further:
Multiple molecular functions, including expression, splicing, and chromatin regulation, can be predicted in an integrated framework.
AlphaGenome Atlas takes the next step:
These predictions can be precomputed for every possible single-nucleotide substitution across the human genome and made broadly accessible.
This represents an important transition in genomic machine learning.
Previously, researchers often needed to execute AI models whenever they wanted to evaluate particular variants.
With AlphaGenome Atlas, large-scale predictions are computed in advance and made available as a reusable scientific resource.
In this sense, AlphaGenome Atlas is not merely a prediction model.
It is an example of transforming AI-based genomic function prediction into scientific infrastructure.
13. AlphaGenome Atlas Is Not the Final Destination for Biological Simulation
The AlphaGenome Atlas research itself presents the resource as a starting point for future research rather than an endpoint.[1]
At least four major challenges remain.
Challenge 1: Predicting Gene Expression Differences Between Individuals
The first challenge concerns whether models can quantitatively explain gene expression differences between individuals using their personal genomes.
Existing sequence-to-expression models may perform well on the reference genome while showing substantially weaker performance when predicting genetic effects on gene expression across individuals.[14]
This is a critical challenge for personalized medicine.
Even if a model can predict that a variant affects a molecular function, accurately estimating the magnitude of that effect in an individual patient is a much harder problem.
Challenge 2: Cellular States and Environmental Conditions
The second challenge is accounting for cellular states and environmental conditions.
The same variant may have different consequences depending on whether a cell is under immune stimulation, in a resting state, at a particular developmental stage, or exposed to a drug.
Therefore, models that use DNA sequences alone cannot explain every biological process occurring in an organism.
Research such as scooby and GET is moving toward incorporating cellular context into prediction.[11][12]
Challenge 3: Variants Beyond Single-Nucleotide Substitutions
The third challenge involves predicting the effects of genomic variants beyond SNVs.
AlphaGenome Atlas primarily provides a comprehensive map of single-nucleotide substitutions.
However, real genomes also contain:
- Insertions and deletions (indels)
- Structural variants
- Repeat expansions
- Haplotypes involving combinations of multiple variants
The effects of these genomic changes cannot be fully addressed through a lookup table of approximately 9 billion SNVs alone.
Challenge 4: Bridging Molecular Functions and Disease
The fourth challenge is the considerable distance between predicting molecular phenotypes and predicting clinical outcomes.
A simplified representation of biological information flow is:
DNA ↓ Chromatin ↓ RNA ↓ Protein ↓ Cell ↓ Organ ↓ Disease
AlphaGenome has become particularly powerful at modeling processes in the earlier stages of this chain.
The ability to predict chromatin states, transcription, and splicing from DNA sequences has advanced considerably.
However, there is still a substantial gap between predicting:
This variant changes the ATAC signal by 0.3.
and predicting:
This variant increases this patient's probability of developing a particular disease by a specific percentage.
This remains an important limitation for both personalized medicine and drug development.
Improved molecular-level predictions do not automatically translate into reliable predictions of disease risk, treatment response, or drug efficacy.
14. The Future of Genomic AI: Convergence of Three Major Approaches
Current genomic AI research can be broadly understood through three major directions.
Direction 1: Functional Genomics AI
Representative models include AlphaGenome, Borzoi, and scooby.
These models predict molecular functions such as transcription, splicing, and chromatin states from DNA sequences.
They are particularly useful for understanding which molecular mechanisms may be affected by genetic variants.
Direction 2: Genomic Language and Foundation Models
Representative models include Evo 2 and Nucleotide Transformer.
These approaches learn general-purpose representations from large quantities of DNA sequence data.
They aim to capture broad genomic patterns rather than focusing exclusively on a particular molecular prediction task.
Direction 3: Evolutionary Constraint Models
Representative models include GPN-Star and related approaches.
These models use evolutionary constraints and genomic sequence patterns to estimate the functional importance and potential pathogenicity of variants.
These three directions should not be viewed as mutually exclusive.
Each captures different types of information and offers strengths for different prediction tasks.
A review published in August 2026 identified generalization across genetic variation and cellular contexts, along with interpretability, as major unresolved challenges for sequence-to-function models.[15]
It is therefore reasonable to expect future systems to integrate these approaches and answer three complementary questions.
- Is this variant evolutionarily unusual? How important or constrained is the sequence change from an evolutionary and biological perspective?
- Which molecular mechanisms does it disrupt? Does it affect transcription, splicing, chromatin, or another molecular process?
- In which cellular contexts does the effect occur? Which cell types, cellular states, or environmental conditions make the effect relevant?
Combining these capabilities may be a major step toward more comprehensive genomic AI.
15. Conclusion: What AlphaGenome Atlas Means for Biological Simulation and Personalized Medicine
The historical significance of AlphaGenome Atlas lies in transforming sequence-to-function genomics into a genome-wide scientific resource.
DeepSEA established the prediction of regulatory functions directly from DNA sequences.
Enformer extended modeling to long-range gene regulation.
Borzoi and AlphaGenome integrated the prediction of multiple molecular processes.
AlphaGenome Atlas builds on these advances by precomputing the predicted molecular consequences of approximately 9 billion possible single-nucleotide substitutions.
This allows researchers to more efficiently prioritize variants identified through GWAS or WGS and investigate their potential molecular mechanisms.
For rare disease research and personalized medicine, the significance extends beyond identifying variants that are likely to be pathogenic.
The technology can also help researchers formulate hypotheses about why particular variants may lead to molecular dysfunction.
In drug discovery, it may support efforts to connect disease-associated variants with molecular mechanisms, generate hypotheses about potential drug targets, and prioritize experimental validation.
However, AlphaGenome Atlas is not necessarily the strongest approach for every variant prediction task.
Evolutionary constraint models such as GPN-Star and genomic foundation models such as Nucleotide Transformer and Evo 2 offer different strengths.
Furthermore, real biological systems are highly dependent on cellular states and environmental conditions.
DNA sequence alone is not sufficient to fully predict disease development or treatment response in individual patients.
AlphaGenome Atlas should therefore not be viewed as a complete biological simulator.
Instead, it is best understood as an important foundational technology connecting DNA sequences to molecular functions.
Looking ahead, the integration of functional genomics, genomic foundation models, evolutionary constraints, and cell-state-specific prediction may enable increasingly comprehensive biological simulations spanning DNA, cells, organs, and disease.
One important qualification is that AlphaGenome Atlas was released on September 8, 2026, as a technical report/preprint and is separate from the original AlphaGenome paper published in Nature.[1][6]
Long-term independent validation of AlphaGenome Atlas remains an important next step.
Its contributions should therefore be evaluated alongside future experimental evidence and independent benchmarking results.
References
[1] Google DeepMind, “AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome,” Google DeepMind Blog, 2026.
https://deepmind.google/blog/alphagenome-atlas-a-predictive-map-of-every-possible-dna-letter-change-in-the-human-genome/
[2] Zhou, J. and Troyanskaya, O. G., “Predicting effects of noncoding variants with deep learning-based sequence model,” Nature Methods, 12, 931–934, 2015.
https://www.nature.com/articles/nmeth.3547
[3] Avsec, Ž. et al., “Effective gene expression prediction from sequence by integrating long-range interactions,” Nature Methods, 18, 1196–1203, 2021.
https://www.nature.com/articles/s41592-021-01252-x
[4] Chen, K. M. et al., “A sequence-based global map of regulatory activity for deciphering human genetics,” Nature Genetics, 54, 940–949, 2022.
https://www.nature.com/articles/s41588-022-01102-2
[5] Linder, J. et al., “Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation,” Nature Genetics, 2025.
https://www.nature.com/articles/s41588-024-02053-6
[6] Google DeepMind et al., “AlphaGenome: a unified deep learning model for sequence-to-function prediction,” Nature, 2026.
https://doi.org/10.1038/s41586-025-10014-0
[7] Dalla-Torre, H. et al., “Nucleotide Transformer: building and evaluating robust foundation models for human genomics,” Nature Methods, 2025.
https://www.nature.com/articles/s41592-024-02523-z
[8] Evo 2 Research Team, “Sequence modeling and design from molecular to genome scale with Evo 2,” Nature, 2026.
https://doi.org/10.1038/s41586-026-10176-5
[9] “Benchmarking DNA foundation models for genomic variant prediction,” Nature Communications, 2025.
https://doi.org/10.1038/s41467-025-65823-8
[10] GPN-Star Research Team, “GPN-Star: evolutionary sequence modeling for non-coding variant prediction,” Nature, 2026.
https://www.nature.com/articles/s41586-026-11005-5
[11] scooby Research Team, “Single-cell sequence-to-function modeling with scooby,” Nature Methods, 2025.
https://www.nature.com/articles/s41592-025-02854-5
[12] GET Research Team, “Generalizable prediction of cellular gene expression from regulatory sequence,” Nature, 2025.
https://www.nature.com/articles/s41586-024-08391-z
[13] “Context-aware sequence-to-function modeling,” Nature Communications, 2026.
https://www.nature.com/articles/s41467-026-75527-2
[14] “Evaluation of sequence-to-expression models for predicting genetic effects on gene expression,” Nature Genetics, 2023.
https://www.nature.com/articles/s41588-023-01574-w
[15] “Sequence-to-function models in regulatory genomics: progress and challenges,” Nature Genetics, 2026.
https://www.nature.com/articles/s41588-026-02670-3