Target Audience#

• Researchers and professionals interested in the history of genomic AI, from DeepSEA and Enformer to Borzoi, AlphaGenome, and AlphaGenome Atlas.
• Readers who want to understand the principles, underlying assumptions, validation status, and limitations of predicting molecular function from DNA sequences.
• Those exploring the potential of genomic AI in rare disease diagnosis, precision medicine, drug discovery, and biological simulation.

Key Takeaways#

• The historical significance of AlphaGenome Atlas lies in transforming sequence-to-function genomics into a genome-wide research infrastructure. It provides precomputed predictions of the molecular effects of approximately 9 billion possible single-nucleotide substitutions across the human genome.
• The AlphaGenome Variant Impact (AVI) score is not derived from AlphaGenome predictions alone. It integrates evolutionary conservation, AlphaMissense, and information from population genetics, bringing together two research traditions: molecular function prediction and variant deleteriousness assessment.
• A fundamental assumption underlying AlphaGenome Atlas is that regulatory rules learned from DNA sequences generalize to unseen variants, and that differences in model predictions reflect actual molecular effects. However, the extent to which predicted variant effects correspond to experimentally measured causal effects remains an important question.
• AlphaGenome Atlas has demonstrated promising performance in variant prioritization and selected molecular function prediction tasks. Nevertheless, predicting inter-individual expression differences, generalizing to unseen cellular states, modeling interactions among multiple variants, and connecting molecular effects to disease causality remain incompletely validated.
• AlphaGenome Atlas is not a complete biological simulator. Rather, it is an important scientific infrastructure for predicting and experimentally investigating the relationships between genomic variation and molecular function. Future progress may come from integrating complementary approaches represented by Evo 2, GPN-Star, scooby, and GET.

1. Where Does AlphaGenome Atlas Fit into the History of Genomic AI?#

The study "AlphaGenome Atlas: in silico mutagenesis of the entire human genome improves prioritization and interpretation of non-coding variants," released in September 2026, represents an important milestone in the development of machine learning for genomics. [1]

AlphaGenome Atlas applies AlphaGenome to predict the molecular consequences of approximately 9 billion possible single-nucleotide variants (SNVs) across the human genome, organizing the results into a research resource of approximately 1 petabyte (PB).

It also introduces the AlphaGenome Variant Impact (AVI) score, which connects variant prioritization with mechanistic interpretation—the effort to understand which molecular processes a variant may affect. [1][2]

An important distinction is that AlphaGenome and AlphaGenome Atlas serve different purposes.

AlphaGenome is a machine learning model that predicts molecular properties such as gene expression, chromatin state, and splicing from DNA sequences.

AlphaGenome Atlas, by contrast, applies this model at a genome-wide scale and combines multiple sources of evidence to support variant interpretation.

However, understanding AlphaGenome Atlas solely in terms of the approximately 9 billion variants it covers risks overlooking its deeper significance.

The research sits at the intersection of at least two long-standing scientific traditions.

The first is sequence-to-function modeling, which attempts to predict molecular function directly from DNA sequence.

The second is variant prioritization, which uses evolutionary constraint and existing genomic annotations to assess the potential deleteriousness of genetic variants.

Understanding AlphaGenome Atlas requires tracing both traditions.

From Variant Annotation to Predicting Molecular Function#

The development of machine learning in genomics can be summarized as follows:

1. Annotating variants using existing genomic information.
2. Predicting genomic annotations directly from DNA sequences.
3. Predicting long-range gene regulation.
4. Predicting multiple molecular phenotypes using a unified model.
5. Precomputing variant effects across the entire genome.
6. Integrating molecular function predictions with evolutionary constraints and other evidence to prioritize and interpret variants.

AlphaGenome Atlas represents a large-scale implementation of the final two stages.

This progression does not mean that genomic AI has evolved along a single, linear trajectory.

Approaches based on evolutionary constraint, experimentally supervised molecular function prediction, and large-scale DNA pretraining continue to develop in parallel.

2. The First Research Tradition: Assessing Variant Deleteriousness Using Existing Genomic Information#

Evolutionary Conservation and Traditional Variant Prioritization#

Early approaches to variant prioritization primarily relied on previously observed or curated genomic information to estimate the pathogenicity or functional significance of genetic variants.

Representative sources of information included:

• Evolutionary conservation scores such as PhyloP and PhastCons
• Transcription factor binding sites
• Histone modifications
• DNase hypersensitivity
• Expression quantitative trait loci (eQTLs)
• Other functional genomic annotations

Statistical and machine learning methods, including support vector machines (SVMs) and random forests, were used to integrate these features and predict the impact of variants.

Representative approaches include CADD, GWAVA, and Eigen.

The underlying framework can be summarized as:

Variant → Existing annotations → Predicted functional importance or pathogenicity

One of the central concepts in this tradition is evolutionary conservation.

If a DNA sequence has remained conserved throughout evolutionary history, changes to that sequence may have been removed by natural selection because they negatively affected survival or reproduction.

Conservation therefore provides valuable evidence about the potential deleteriousness of a variant.

However, high evolutionary conservation does not directly establish pathogenicity.

Conversely, variants in poorly conserved regions may still have important functions in specific cellular contexts or diseases.

The underlying assumption of this research tradition is that patterns revealed by evolutionary history and known genomic annotations can be used to assess previously unseen variants.

Although this is a reasonable assumption, it does not necessarily explain molecular effects that vary across cell types and experimental conditions.

Another limitation is that predictions may be less informative in genomic regions or cellular contexts where functional annotations are sparse.

Genome-Wide Precomputed Scores Existed Before AlphaGenome Atlas#

Another historical point is important when positioning AlphaGenome Atlas.

The idea of precomputing variant impact scores across the genome and making them accessible as a research resource did not originate with AlphaGenome Atlas.

Earlier resources, including CADD and dbNSFP, had already provided large collections of precomputed variant scores and functional annotations.

Sei, discussed later, also introduced a genome-wide map of regulatory activity.

Therefore, the novelty of AlphaGenome Atlas does not lie simply in making precomputed scores available through a database.

Its more substantial contribution is providing genome-wide, allele-specific predictions across numerous molecular phenotypes and connecting those predictions to variant prioritization and mechanistic interpretation.

In this sense, AlphaGenome Atlas extends the traditional variant prioritization framework by considerably expanding the molecular information available for interpretation.

3. The Second Research Tradition: Predicting Molecular Function Directly from DNA Sequence#

DeepSEA: The Foundation of Sequence-to-Function Modeling#

A major turning point came with DeepSEA, published in 2015. [3]

DeepSEA used convolutional neural networks (CNNs) to predict multiple chromatin-related features directly from DNA sequence, including transcription factor binding, DNase sensitivity, and histone modifications.

It also estimated variant effects by comparing predictions for the reference allele and alternative allele.

Whereas earlier methods relied on existing genomic annotations, DeepSEA attempted to predict those annotations directly from DNA sequence.

DNA sequence → Deep learning → Chromatin state → Variant effect

This idea became a foundation of the modern sequence-to-function paradigm.

Consequently, if we were to identify a direct technological ancestor of AlphaGenome Atlas, DeepSEA would be a more appropriate choice than AlphaFold.

Basenji and Enformer: Modeling Long-Range Gene Regulation#

The next major challenge concerned the length of DNA sequence that a model could analyze, commonly called its sequence context.

Gene expression is not determined solely by the few hundred nucleotides surrounding a promoter or transcription start site.

Enhancers located tens or hundreds of kilobases away may substantially affect transcription.

The Basenji family of models addressed this challenge, with Enformer, published in 2021, representing an especially important development. [4]

Enformer takes approximately 196 kilobases (kb) of DNA sequence as input and uses the attention mechanism of the Transformer architecture to model long-range interactions.

In humans, it predicts more than 5,000 functional genomic tracks.

This development changed the scientific questions that could be asked.

Earlier models focused on questions such as whether a variant disrupts a transcription factor binding motif.

Enformer made it possible to investigate a broader question:

How might changes in distant regulatory elements influence gene expression?

This represented a transition from predicting local DNA functions to modeling larger regulatory systems.

However, the ability to process longer DNA sequences does not guarantee accurate reconstruction of every long-range regulatory interaction within those sequences.

This limitation remains relevant to AlphaGenome.

Sei: Organizing Regulatory Functions into Interpretable Features#

Sei, published in 2022, predicts 21,907 chromatin profiles and organizes them into approximately 40 sequence classes. [5]

This framework helps describe regulatory features such as enhancers, promoters, CTCF-associated elements, and cell-type-specific regulatory activity in a more interpretable way.

Sei is important not only because it predicts many molecular features, but also because it organizes those predictions into biologically meaningful categories.

It also represents an important predecessor in the development of genome-wide functional maps.

Borzoi: Directly Predicting RNA-seq Coverage#

Borzoi approaches the problem from a different direction by predicting RNA-seq coverage directly from DNA sequence. [6]

DNA → RNA-seq coverage

Using a sequence context of approximately 524 kb, Borzoi predicts aspects of transcription, splicing, and polyadenylation.

Its importance lies in demonstrating how multiple RNA-related molecular processes can be modeled from a common DNA sequence input.

By this stage, two important capabilities underlying AlphaGenome had emerged:

• The ability to process long genomic contexts
• The ability to integrate multiple regulatory modalities

4. What Did AlphaGenome Change?#

AlphaGenome, introduced in 2025 and published in Nature in 2026, substantially expanded the integration of sequence-to-function prediction tasks. [7]

AlphaGenome takes approximately 1 megabase (Mb) of DNA sequence as input and predicts 5,930 tracks across 11 modalities in humans.

Its prediction targets include:

• RNA-seq
• CAGE
• PRO-cap
• Splicing
• DNase / ATAC
• Histone modifications
• Transcription factor binding
• Chromatin contacts

Its key achievement is the simultaneous combination of long context, base-level resolution, and multimodal prediction.

Earlier models had different strengths:

• Enformer specialized in long-range regulatory prediction.
• Borzoi offered strong capabilities for RNA processing.
• ChromBPNet focused on chromatin accessibility.
• SpliceAI and related models specialized in splicing prediction.

AlphaGenome moves toward integrating these different prediction tasks into a unified sequence-to-function model.

Its Nature paper reports performance comparable to or better than the best existing models on 25 of 26 variant effect prediction evaluations. [7]

This represents meaningful progress toward predicting multiple molecular phenomena—including transcription, chromatin state, and splicing—from a shared DNA sequence input.

It is also an important development from the perspective of biological simulation.

However, AlphaGenome does not simulate an entire biological organism.

Its primary scope is the relationship between DNA sequence and molecular function.

It does not directly reconstruct a unified dynamic system covering changes in cellular state over time, intercellular interactions, environmental responses, organ physiology, or disease progression.

The AlphaGenome paper also identifies limitations concerning long-range regulation, unseen cellular contexts, and generalization to personal genomes. [7]

Therefore, improved molecular function prediction should be distinguished from the ability to faithfully reconstruct biological causality.

5. The Real Novelty of AlphaGenome Atlas#

The innovation of AlphaGenome Atlas lies less in introducing a new neural network architecture than in turning variant effect predictions into a genome-wide scientific resource. [1][2]

Precomputing Approximately 9 Billion Single-Nucleotide Substitutions#

The core computational technique underlying AlphaGenome Atlas is in silico mutagenesis (ISM).

ISM systematically changes individual nucleotides in a DNA sequence and evaluates how the model's predictions change.

For example, if the reference nucleotide at a particular position is A, three alternative substitutions can be considered:

• A → C
• A → G
• A → T

The model is evaluated on both the reference sequence and the alternative sequence.

The predicted difference is:

[ \Delta \hat{f}

\hat{f}(\text{ALT sequence})#

\hat{f}(\text{REF sequence}) ]

Here, (\hat{f}) denotes a trained model that predicts molecular function from DNA sequence.

The resulting difference can be used to estimate effects on transcription factor binding, gene expression, chromatin state, splicing, and other molecular properties.

ISM itself is not a new technique.

Variant effect prediction has relied on this basic principle since the DeepSEA era. [3]

AlphaGenome also uses ISM to investigate features such as splicing motifs. [7]

The achievement of AlphaGenome Atlas is applying this approach across the entire human genome.

The human genome contains approximately 3 billion nucleotides.

Each position has three possible alternative nucleotides.

Therefore:

Approximately 3 billion positions × 3 alternative nucleotides ≈ 9 billion SNVs

AlphaGenome Atlas precomputes the predicted molecular effects of these substitutions.

In this respect, its achievement can be described as a revolution in scale.

The Atlas study also analyzes numerous insertions and deletions (indels) observed in population data. [1]

Therefore, it would be inaccurate to interpret the research as entirely restricted to SNVs.

Nevertheless, the resource does not comprehensively cover every structural variant, repeat expansion, or complex haplotype.

From On-Demand Prediction to Precomputed Scientific Infrastructure#

Traditionally, variant effect prediction involved identifying candidate variants through GWAS or whole-genome sequencing (WGS), then running models such as Enformer, Borzoi, or AlphaGenome on those candidates.

The traditional workflow was:

1. Identify candidate variants through GWAS or WGS.
2. Submit candidate variants to a prediction model.
3. Calculate molecular effects for each variant.

AlphaGenome Atlas changes the workflow:

1. Precompute molecular effects for possible SNVs throughout the genome.
2. Identify candidate variants through GWAS or WGS.
3. Retrieve the corresponding predictions from the database.

This changes how researchers can use genomic AI.

It reduces the need to repeatedly run computationally expensive models for individual variants and makes molecular predictions more accessible.

However, the scope of the data analyzed in the paper does not necessarily correspond exactly to the publicly downloadable data or its usage conditions.

The Analogy with AlphaFold Database#

The relationship between AlphaGenome and AlphaGenome Atlas resembles that between AlphaFold and AlphaFold Database.

AlphaGenome : AlphaGenome Atlas ≒ AlphaFold : AlphaFold Database

Just as AlphaFold Database transformed protein structure prediction into a widely accessible research resource, AlphaGenome Atlas makes large-scale variant effect predictions available as scientific infrastructure.

This can be understood as a transition from an AI model to scientific infrastructure.

However, the biological prediction problems are different.

AlphaFold primarily predicts three-dimensional protein structures.

AlphaGenome Atlas predicts molecular effects—including gene expression, chromatin state, and splicing—that often depend on cell type and biological context.

Therefore, the success of AlphaFold Database should not itself be treated as evidence that AlphaGenome Atlas predictions are biologically valid under arbitrary conditions.

Making predictions available at scale and demonstrating their accuracy across biological contexts are different achievements.

6. AVI Integrates Molecular Function Predictions with Evolutionary Constraints#

Another major contribution of AlphaGenome Atlas is the AlphaGenome Variant Impact (AVI) score. [1]

AlphaGenome produces a large amount of molecular prediction data.

For example, a particular variant might be predicted to cause:

• Reduced chromatin accessibility
• Reduced transcription factor binding
• Reduced H3K27ac
• Reduced transcription
• Changes in splicing
• Changes in chromatin contact maps

These predictions can help researchers formulate hypotheses about molecular mechanisms.

However, WGS studies often require the evaluation of millions of variants.

Examining every molecular prediction individually is impractical.

Researchers therefore need to answer a more operational question:

Which variants should be investigated first?

AVI addresses this need by integrating multiple features into a unified variant prioritization score.

AVI Is Not Based on AlphaGenome Alone#

This is one of the most important points for understanding AlphaGenome Atlas.

AVI does not use only molecular function predictions from AlphaGenome.

According to the Atlas study, it integrates 18 features across several categories. [1]

Feature category	Number of features	Main purpose
AlphaGenome-derived molecular predictions	10	Effects on transcription, splicing, chromatin, and related functions
AlphaMissense and protein loss-of-function features	4	Potential effects on proteins
Evolutionary conservation	2	Evolutionary constraint
Indel-related features	2	Characteristics of insertions and deletions
Total	18	Integrated assessment of variant impact

AVI also uses population allele frequencies as proxy training labels.

Specifically, it uses data from gnomAD and distinguishes variants with frequencies above or below 0.1%, treating them as proxies for relatively neutral variants and variants more likely to have functional consequences, respectively. [1]

This strategy is based on the idea that strongly deleterious variants tend to be removed from populations by natural selection.

However, rare variants are not necessarily pathogenic.

Variant frequency is influenced by demographic history, genetic drift, mutation age, and ancestry.

Likewise, a variant causing a large molecular change does not necessarily cause disease.

Therefore, AVI does not directly observe pathogenicity.

Instead, it estimates functional importance by learning relationships among molecular effects, evolutionary constraint, and population frequency.

Its score should not be interpreted as the probability that an individual patient will develop a disease.

Compressing High-Dimensional Molecular Information into a Single Score#

The Atlas study reports an average of approximately 27,000 experiment- and condition-specific scalar predictions per variant. [1]

AVI reduces this high-dimensional information into modality-level features, including large predicted effects across cell types and tissues.

This design has considerable practical value.

It makes large numbers of variants comparable and supports efficient prioritization.

However, aggregation can also obscure important differences between tissues and cellular contexts.

For example, a variant that strongly affects liver cells and another that strongly affects neurons may receive similar aggregate impact scores.

Yet their relevance to a neurological disease could be very different.

For disease interpretation, it is therefore important to investigate not only the AVI score but also:

• Which cell types show predicted effects
• Which molecular phenotypes are affected
• Which genes may be involved
• Whether the predicted effects are positive or negative

SHAP Interpretation and Its Limits#

AVI also provides feature attribution using SHAP (SHapley Additive exPlanations). [1]

SHAP can indicate how features related to splicing, expression, chromatin, or protein impact contribute to a particular AVI score.

This connects two approaches:

• The CADD tradition of summarizing variant impact in a single score
• The AlphaGenome tradition of predicting specific molecular effects

However, SHAP explains contributions to a model's prediction.

A large SHAP contribution from splicing features does not prove that a variant actually causes a splicing defect.

Likewise, a high AVI score driven by evolutionary conservation does not identify a specific causal molecular mechanism.

AVI is therefore valuable for prioritization and hypothesis generation, but its proposed mechanisms require independent experimental validation.

7. Five Hidden Assumptions Underlying AlphaGenome Atlas#

To evaluate AlphaGenome Atlas scientifically, it is necessary to examine not only what the technology can do, but also what must be true for its predictions to be reliable.

Based on the research history from DeepSEA onward and the design of AlphaGenome Atlas, five major assumptions can be identified.

These assumptions are an analytical interpretation of the methodology, rather than a list explicitly stated by the authors.

Assumption 1: DNA Contains Regulatory Rules That Generalize to Unseen Variants#

Sequence-to-function models learn relationships between DNA sequences and experimentally measured molecular phenotypes.

Their predictive ability depends on the existence of regulatory patterns that generalize across genomic regions and previously unseen variants.

For example, if a model learns that a particular transcription factor binding motif influences a molecular phenotype, it may predict the effect of disrupting that motif.

However, actual transcription factor binding does not depend on DNA sequence alone.

It also depends on factors such as:

• Whether the transcription factor is present in the cell
• Its concentration
• Local chromatin accessibility
• Interactions with other regulatory proteins
• Cellular state and environmental conditions

DNA sequence is therefore an essential source of information, but it does not uniquely determine every molecular function in a cell.

AlphaGenome learns part of these complex relationships from many experimental datasets.

Nevertheless, it cannot fully reconstruct arbitrary cellular or environmental states.

The Atlas study explicitly identifies the absence of direct modeling of certain trans-acting mechanisms, including changes in transcription factor expression, as a limitation. [1]

Assumption 2: Differences in Model Predictions Reflect Actual Variant Effects#

The central calculation in AlphaGenome Atlas is the difference between model predictions for reference and alternative sequences:

[ \Delta \hat{Y}

\hat{f}(S_{\mathrm{ALT}})#

\hat{f}(S_{\mathrm{REF}}) ]

Here:

• (\hat{f}) is the trained prediction model.
• (S) represents a DNA sequence.
• (\Delta \hat{Y}) is the predicted molecular effect.

However, the biologically relevant quantity is the actual change caused by modifying a nucleotide under otherwise comparable genetic and experimental conditions.

Conceptually:

[ \Delta Y_{\mathrm{true}}

Y(\operatorname{do}(\mathrm{ALT}))#

Y(\operatorname{do}(\mathrm{REF})) ]

These quantities are not necessarily identical.

A model may accurately reproduce patterns of RNA-seq or ATAC-seq signals across a reference genome without accurately predicting the causal effect of introducing a variant.

This is a fundamental issue in sequence-to-function modeling.

A 2023 Nature Genetics study by Huang and colleagues evaluated four models—Enformer, Basenji2, ExPecto, and Xpresso—using personal genome and transcriptome data. [16]

The researchers found that although the models could explain some differences in expression levels across genes, they performed poorly at predicting expression differences between individuals for the same gene.

They also sometimes predicted the wrong direction of variant effects.

A related study by Sasse and colleagues highlighted additional shortcomings in personal gene expression prediction. [17]

These studies did not directly evaluate AlphaGenome Atlas.

However, they establish an important distinction:

Accurately predicting molecular phenotypes from reference sequences is not the same as accurately predicting causal effects of unseen genetic variants.

Assumption 3: The Effect of a Single Variant Can Be Evaluated Independently of Genetic Background#

The approximately 9 billion SNVs in AlphaGenome Atlas are primarily evaluated by substituting one nucleotide in a reference genome background.

Thus, each prediction largely addresses the question:

"What happens when this nucleotide is changed in the reference sequence?"

Real individuals, however, carry many variants simultaneously.

Suppose two variants, A and B, occur in the same regulatory region.

Their combined effect may not equal the sum of their individual effects:

[ \Delta Y(A+B) \neq \Delta Y(A)+\Delta Y(B) ]

Such non-additive interactions are associated with epistasis and haplotype dependence.

In principle, AlphaGenome itself can process sequences containing multiple variants.

However, simply adding single-SNV scores from Atlas does not necessarily provide an accurate estimate of the molecular consequences of an individual's complete genome.

Generalizing from isolated variants in a reference genome to combinations of variants in personal genomes requires separate validation.

Assumption 4: Evolutionary Constraint and Population Frequency Are Informative Proxies for Variant Importance#

AVI uses evolutionary conservation and population allele frequencies in addition to molecular function predictions.

Its design reflects the assumption that variants affecting function are more likely to be constrained by natural selection.

However:

[ \text{Rare variant} \not\Rightarrow \text{Pathogenic variant} ]

Many rare variants do not cause disease.

Furthermore, variants associated with disease are not necessarily subject to strong evolutionary constraint.

This relationship may be particularly complex for late-onset diseases, environmentally dependent conditions, and functions that emerged relatively recently in evolution.

AVI uses evolutionary and population-genetic information as proxies for functional importance.

Although this is a reasonable approach, the proxies should not be assumed to correspond perfectly to clinical pathogenicity.

Assumption 5: Integrating Molecular Effects Can Efficiently Identify Disease-Relevant Variants#

AlphaGenome Atlas aggregates numerous molecular predictions to rank variants.

Doing so requires reducing cell-type- and tissue-dependent information into comparable features.

Yet biological importance is not determined solely by the magnitude of a predicted effect.

A molecular change in a liver cell may have a different disease implication from an equally large change in a neuron.

Similarly, a large molecular change may have little consequence for cellular function, while a relatively small change during a critical developmental stage may be important.

Therefore, a high integrated impact score and causal relevance to a particular disease are not equivalent.

This distinction is critical when interpreting AVI scores in precision medicine and drug discovery.

8. What Has AlphaGenome Atlas Actually Validated?#

The existence of unresolved assumptions does not mean that AlphaGenome Atlas has not demonstrated practical value.

The Atlas study reports multiple evaluations using known pathogenic variants, functional assays, rare disease cases, and population-genetic analyses. [1]

The important question is what each evaluation actually demonstrates.

Evaluation Against Known Pathogenic Variants#

AVI shows promising performance in several evaluations using pathogenic and benign variants from ClinVar.

Representative results include: [1]

Evaluation category	AVI	Comparator	Comparator performance
ClinVar intronic SNVs	AUPRC 0.76	GPN-Star-V	0.44
ClinVar synonymous SNVs	AUPRC 0.57	CADD v1.7	0.35
ClinVar 3' UTR SNVs	AUPRC 0.50	GPN-Star-M	0.18
ClinVar protein-altering SNVs	AUPRC 0.90	GPN-Star-V	0.86
ClinVar 5' UTR SNVs	AUPRC 0.26	GPN-Star-V	0.27
TraitGym Mendelian noncoding variants	AUPRC 0.76	GPN-Star-M	0.77

AUPRC stands for area under the precision-recall curve.

Higher values indicate better discrimination of positive variants within the evaluated dataset.

However, AUPRC depends on the proportion of positive examples, so values from different evaluation datasets should not be compared directly.

These results indicate that AVI provides strong variant prioritization performance in several categories.

They also show that AVI is not the best-performing method in every category.

Furthermore, agreement with ClinVar classifications does not directly establish diagnostic accuracy in previously unseen patients.

Comparison with Experimental Variant Effects#

The Atlas study also compares AVI with functional measurements from saturation genome editing and related experiments.

Among ten saturation genome-editing experiments evaluated, AVI achieved the highest Spearman correlation in eight. [1]

The paper also presents examples consistent with known molecular mechanisms, including disruption of an E2F motif in the BRCA1 promoter.

These comparisons provide a different kind of evidence from ClinVar classification.

They assess whether predicted molecular effects correspond to experimentally measured consequences.

However, strong correlations in selected experiments do not establish quantitative accuracy across all genomic regions, cell types, and environmental conditions.

Moreover, three distinct questions must be evaluated separately:

• Does the model rank variants correctly?
• Does it predict the correct direction of the effect?
• Does it predict the magnitude of the effect accurately?

Success on one does not guarantee success on the others.

The DNM1 Rare Disease Case#

A study involving researchers from the Broad Institute and others identified an important candidate variant near DNM1 in an unresolved rare disease case. [1][2]

AlphaGenome predicted that the variant could create a new splice site.

The predicted splicing effect was subsequently experimentally confirmed.

The Atlas study also reports a minigene reporter assay involving a 265-nucleotide region upstream of DNM1 exon 10a, evaluated across five cell lines.

In this experiment, the integrated AlphaGenome splicing prediction and AVI splicing attribution both achieved an AUPRC of 0.943.

The comparator models SpliceAI and Pangolin also achieved strong performance, with AUPRC values of 0.940 and 0.939, respectively. [1]

This example illustrates the practical significance of AlphaGenome Atlas.

It can support not only the identification of potentially pathogenic variants, but also the development of specific hypotheses explaining how those variants may disrupt molecular function.

Nevertheless, success in one disease-related region cannot be generalized to diagnostic accuracy across all rare diseases.

Applications to Population-Genetic Association Analysis#

The Atlas study also evaluated rare noncoding variant associations using data from 54,189 UK Biobank participants and 2,028 blood proteins. [1]

Filtering variant sets using Atlas-derived functional information increased the number of conditionally independent rare noncoding association signals by 22%.

This suggests that functional predictions can improve statistical power by selecting more informative variant sets.

However, discovering more association signals does not establish the causal molecular mechanisms of every implicated variant.

The study also examined traits including height, BMI, and HbA1c, but replication in external cohorts was limited.

Among associations that could be evaluated in the All of Us cohort, some replicated at nominal significance thresholds, but none reached significance after multiple-testing correction. [1]

These findings indicate both the promise of Atlas-assisted variant selection and the need for further validation across independent populations.

What Is Validated—and What Remains Uncertain?#

Evaluation	What has been demonstrated	What remains unresolved
ClinVar comparisons	Discrimination of known pathogenic and benign variants	Diagnostic accuracy in unseen patients
Genome editing	Agreement with measured functional effects for selected variants	Quantitative accuracy across the entire genome
DNM1 case	Utility for interpreting a rare disease variant	Improvement in diagnostic yield across rare diseases
UK Biobank analysis	Improved association discovery through functional filtering	Correct causal mechanisms for individual variants
SHAP interpretation	Identification of features contributing to AVI scores	Proof of biological causal pathways
Evolutionary integration	Improved variant prioritization	Accuracy of AlphaGenome as a standalone molecular simulator
Personal genomes	Potential for future applications	Comprehensive prediction of inter-individual expression differences

An important distinction is that AVI performance should not be equated with AlphaGenome's standalone performance.

AVI incorporates information beyond AlphaGenome.

Therefore, improvements in variant classification cannot automatically be attributed entirely to the accuracy of AlphaGenome's molecular predictions.

Another consideration concerns potential overlap in the types of information used for training and evaluation.

AVI uses population frequency as a proxy training label, while ClinVar classifications may also draw on population frequency and functional evidence.

Although the Atlas study takes measures to avoid direct overlap between training and evaluation variants, related types of evidence may still influence both processes. [1]

Independent functional experiments therefore remain essential.

9. The Most Fundamental Open Question: Can We Predict Gene Expression Differences Between Individuals?#

One of the most important limitations of sequence-to-function genomics concerns the prediction of inter-individual gene expression differences.

In a 2023 study, Huang and colleagues evaluated existing sequence-to-expression models using personal genomes and RNA-seq data from 421 individuals. [16]

The study distinguished two different prediction tasks.

The first was predicting expression differences between genes.

For example, under the same cellular condition, is gene X expressed more highly than gene Y?

The second was predicting expression differences between individuals for the same gene.

For example, how much does the expression of gene X differ between individual A and individual B?

Existing models demonstrated some predictive ability for the first task.

However, performance deteriorated substantially for the second.

For the four evaluated models, including Enformer, the distribution of correlations between predicted and observed inter-individual differences was centered approximately around zero.

The study also reported cases in which the models predicted the wrong direction of genetic effects. [16]

This distinction is crucial for precision medicine.

What ultimately matters in a medical context is not simply the molecular function of a generic reference genome.

The relevant question is how a particular patient's genome affects that patient's molecular phenotype.

Of course, the 2023 study did not directly evaluate AlphaGenome Atlas.

AlphaGenome has demonstrated improved performance over earlier models on numerous tasks. [7]

Therefore, it would be inappropriate to assume that the limitations of older models apply to AlphaGenome to exactly the same extent.

However, the AlphaGenome Nature paper acknowledges that comprehensive benchmarking of personal genome predictions remains incomplete. [7]

The key question is therefore:

Can AlphaGenome accurately predict the effects of genetic variation in previously unseen individuals, rather than only predicting molecular phenotypes associated with reference sequences?

Until this question is adequately addressed, AlphaGenome Atlas should not be regarded as a validated simulator of individual-specific molecular function.

10. Additional Challenges: Cellular Context, Long-Range Regulation, and Genetic Interactions#

Beyond predicting inter-individual gene expression differences, several other important challenges remain.

Generalization to Unseen Cell Types and Environmental Conditions#

In living organisms, the same DNA sequence can have different molecular consequences depending on cell type and cellular state.

For example, a variant may have little effect in an unstimulated immune cell but produce a strong effect following inflammatory stimulation.

Developmental stage, drug exposure, stress, and hormonal environment can also influence molecular function.

AlphaGenome is trained on data from many cell and tissue types, but it does not cover every biological condition.

Furthermore, the available cell types differ across modalities such as RNA-seq, ATAC-seq, and ChIP-seq.

The Atlas study identifies these limitations in training-data coverage and generalization to unseen cellular contexts. [1]

This challenge may not be solved simply by increasing model size.

It also raises architectural questions about how cellular state and environmental information should be represented as model inputs.

Long-Range Gene Regulation#

AlphaGenome processes approximately 1 Mb of DNA sequence.

However, processing a 1 Mb sequence does not guarantee accurate prediction of every regulatory interaction within that region.

The AlphaGenome paper identifies effects of distant regulatory elements, particularly those more than 100 kb away, as an area where improvement is needed. [7]

Real gene regulation also depends on three-dimensional chromatin structure, interchromosomal interactions, and intracellular transcription factor concentrations.

Therefore, a predicted effect on a distal enhancer does not itself establish a causal enhancer–gene relationship.

Molecular Functions Missing from Training Data#

Prediction models fundamentally depend on the information available in their training data.

As a result, they inherit limitations of experimental measurement technologies.

For example, conventional poly(A)-selected RNA-seq does not adequately capture non-polyadenylated RNA species.

The Atlas study identifies RNU4-2 as an example related to this limitation. [1]

Even a highly capable model may therefore struggle to predict phenomena that are not sufficiently represented in its training data.

Improving genomic AI will require not only better model architectures, but also broader and higher-quality experimental datasets.

Variants Beyond SNVs and Haplotype Effects#

Human genomes contain many forms of variation beyond SNVs, including:

• Insertions and deletions
• Structural variants
• Copy number variations
• Repeat expansions
• Haplotypes containing multiple variants

The Atlas study includes many observed indels, but does not comprehensively cover all structural variants or complex haplotypes. [1]

The effect of one variant may also depend on other variants in the same genomic region.

Ignoring such genetic-background dependence can introduce errors when extending single-variant predictions to complete personal genomes.

Generalization Across Ancestry Groups#

AVI uses information related to allele frequencies in human populations.

However, population frequencies vary across ancestry groups and demographic histories.

A variant that is rare in one population may have a different frequency in another.

At the same time, differences in frequency do not necessarily imply differences in biological function.

The important question is whether a model trained using allele-frequency proxies can evaluate variants accurately and consistently across diverse ancestry groups.

The Atlas study also identifies the need for validation across diverse genetic ancestries. [1]

AVI Thresholds and Clinical Interpretation#

AVI provides a score for comparing variants, but a single threshold may not be optimal across all genomic regions.

The Atlas study shows that coding and splice-related variants often receive high AVI scores, while functionally relevant cis-regulatory variants may occupy lower score ranges. [1]

The authors therefore recommend considering genomic context and intended use when choosing thresholds, and prioritizing ranking over fixed cutoffs when appropriate.

AVI scores should not be directly interpreted as probabilities of pathogenicity or used as standalone clinical diagnostic criteria.

11. Is AlphaGenome a "Genome-Scale LLM"?#

Contemporary genomic AI includes two broad research traditions.

The first consists of sequence-to-function models such as AlphaGenome.

The second consists of DNA language models and genomic foundation models.

Their main differences lie in their learning objectives and training data.

Sequence-to-Function Models#

Representative models include:

DeepSEA → Basenji → Enformer → Borzoi → AlphaGenome

These models are primarily trained using experimentally measured functional genomic data, such as:

• RNA-seq
• ATAC-seq
• ChIP-seq
• CAGE
• Hi-C

Their objective is to predict biological molecular function from DNA sequence.

DNA → Biological function

AlphaGenome and AlphaGenome Atlas belong to this research tradition.

DNA Language Models and Genomic Foundation Models#

Another research tradition includes:

• DNABERT
• Nucleotide Transformer
• HyenaDNA
• Caduceus
• Evo / Evo 2
• GPN-family models

These models learn statistical and biological patterns in DNA sequences, often through self-supervised learning.

Their objective is to learn general representations and regularities in genomic sequences.

For example, Nucleotide Transformer uses 3,202 human genomes and genomes from 850 species to learn general-purpose DNA representations. [8]

Evo 2, published in Nature in 2026, represents a much larger genomic foundation model. [9]

It was trained on approximately 9 trillion nucleotides spanning all domains of life and supports context lengths of approximately 1 million tokens.

It can also predict certain variant effects without task-specific fine-tuning.

The distinction can be summarized as follows:

AlphaGenome primarily uses experimentally supervised learning to predict molecular function.

Evo 2 primarily learns generalizable patterns from large-scale genomic sequence data, supporting prediction and sequence generation.

However, these approaches need not remain separate.

Representations learned by genomic foundation models may be useful for molecular function prediction.

Likewise, models based on evolutionary information may provide complementary evidence about the significance of variants.

Rather than viewing these approaches as mutually exclusive competitors, it may be more productive to consider how they can complement one another.

12. AlphaGenome Is Not the Best Model for Every Prediction Task#

The relative strengths of genomic AI models depend on the prediction task.

A 2025 benchmark of DNA foundation models reported that sequence-to-function models such as AlphaGenome, Enformer, and Sei tended to perform strongly on cell- or tissue-specific regulatory variant prediction tasks, including QTL-related evaluations. [10]

For example, the benchmark reported the following approximate results for AlphaGenome:

• sQTL: AUC ≈ 0.80
• ipaQTL: AUC ≈ 0.86

By contrast, DNA foundation models such as Nucleotide Transformer performed better in some tasks distinguishing pathogenic variants from common variants.

These findings suggest that different model families capture different types of information.

Sequence-to-function models are particularly relevant to questions such as:

"Will this variant alter gene expression or splicing in a particular tissue?"

Language models and evolutionary models may be more useful for questions such as:

"How unusual or evolutionarily constrained is this DNA sequence change?"

However, relative performance depends on the benchmark dataset, training data, and variant category.

It would therefore be inappropriate to declare one model family universally superior.

13. GPN-Star Reveals Another Important Direction#

GPN-Star, published in Nature in September 2026, is an important comparison for AlphaGenome Atlas. [11]

Unlike AlphaGenome, which predicts experimentally measured functional genomic signals, GPN-Star learns functional constraints from evolutionary information and patterns in genomic sequences.

Its main strength lies in estimating how evolutionarily or functionally constrained a genomic position is, rather than directly predicting every downstream molecular phenotype.

The GPN-Star paper reports performance exceeding AlphaGenome, Borzoi, and Enformer in some evaluations of noncoding pathogenic variants associated with OMIM and HGMD. [11]

This is an important result because:

Predicting molecular function accurately is not the same as predicting pathogenicity most accurately.

For example, a variant may be predicted to affect transcription factor binding or gene expression, but that molecular change does not necessarily cause disease.

However, an important distinction must be made when interpreting the comparison.

The AlphaGenome-derived predictions evaluated in the GPN-Star study are not identical to the AVI model introduced by AlphaGenome Atlas.

AVI integrates AlphaGenome with evolutionary conservation, AlphaMissense, and other features.

Indeed, the Atlas study reports that AVI outperforms GPN-Star in several ClinVar evaluations.

Conversely, GPN-Star slightly outperforms AVI in some evaluations, including 5' UTR variants and TraitGym Mendelian variants. [1]

Therefore, both statements can be true:

• GPN-Star outperforms standalone AlphaGenome predictions on some benchmarks.
• AVI outperforms GPN-Star on other benchmarks.

Ignoring these differences would lead to an oversimplified conclusion.

A more informative interpretation is that molecular function predictions and evolutionary constraints both provide valuable—and potentially complementary—information for variant assessment.

AVI represents an effort to combine these sources of evidence.

14. Comparing Major Genomic AI Models#

The major research directions can be summarized as follows.

Research tradition	Representative models	Main information learned or used	Main strengths	Relationship to AlphaGenome Atlas
Traditional variant prioritization	CADD, GWAVA, Eigen	Existing genomic annotations and evolutionary conservation	Variant deleteriousness assessment	Predecessors in variant scoring
Regulatory CNN	DeepSEA	Chromatin assays	Regulatory effects predicted from sequence	Direct technological ancestor
Regulatory deep learning	Sei	Approximately 21,000 chromatin profiles	Interpretable regulatory classes	Earlier genome-wide functional map
Long-range sequence-to-function	Enformer	Epigenomic and expression data	Long-range enhancer–gene effects	Major predecessor
RNA sequence-to-function	Borzoi	RNA-seq and related data	Expression, splicing, APA	Direct comparison model
Unified sequence-to-function	AlphaGenome	Multiple functional genomic assays	1 Mb context, multimodal prediction, base-level resolution	Foundation model underlying Atlas
Genome-wide resource	AlphaGenome Atlas	AlphaGenome predictions and variant annotations	Approximately 9 billion SNV predictions, AVI, mechanistic interpretation	Genome-wide research infrastructure
DNA foundation model	Nucleotide Transformer	DNA sequences	General-purpose sequence representations	Complementary approach
Generative genomic foundation model	Evo 2	Approximately 9 trillion DNA nucleotides	Prediction and generation	General-purpose genomic AI
Evolutionary variant prediction	GPN-Star	Evolutionary history and genomic sequences	Functional constraint and pathogenicity prediction	Complementary and competing approach
Cell-state sequence-to-function	scooby	Single-cell RNA and ATAC data	Cell-specific variant effects	Focus on cellular context
Cell-context model	GET	Regulatory data across many cell types	Generalization to unseen cell types	Addresses cellular-context limitations

This comparison illustrates that genomic AI is not converging toward a single technological approach.

Research continues in parallel on molecular function prediction, evolutionary constraints, cellular context, and general-purpose genomic representation learning.

The table describes the conceptual positions of these models rather than providing an overall performance ranking.

15. scooby and GET Point Toward a Different Next Step#

One important 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 produce different transcriptional and chromatin states depending on cell identity and state.

Models such as scooby and GET address aspects of this challenge.

scooby: Single-Cell Variant Effect Prediction#

Published in 2025, scooby builds on Borzoi while modeling single-cell RNA-seq and ATAC-seq data. [12]

This enables a move toward cell-state-specific variant effect prediction.

Traditional sequence-to-function models have often focused on predicting molecular signals averaged across tissues or cell populations.

scooby attempts to capture finer differences between cellular states.

This represents progress toward models that consider not only DNA sequence, but also cellular context.

GET: Generalization to Unseen Cell Types#

GET (General Expression Transformer) uses data from 213 human cell types and aims to generalize predictions across cellular contexts. [13]

Its significance lies in attempting to predict gene regulation in cell types and conditions not represented directly in training data.

Related context-aware sequence-to-function research published in 2026 also addresses the difficulty of generalizing traditional models to unseen cell types and conditions. [14]

These studies point toward incorporating cellular and environmental information alongside DNA sequence.

DNA sequence + Cell state + TF environment + Epigenetic context

However, providing a model with information about cellular state does not automatically solve every biological prediction problem.

Cellular states themselves change over time and are influenced by surrounding cells and environmental conditions.

Predicting molecular function conditional on a cellular state must therefore be distinguished from predicting how that state evolves dynamically.

16. How Can AlphaGenome Atlas Contribute to Precision Medicine?#

Genomic analysis based on WGS and GWAS typically involves several stages between identifying candidate variants and understanding disease mechanisms.

A simplified workflow is:

1. Identify candidate variants through WGS, GWAS, or fine mapping.
2. Prioritize candidate variants using AI-based methods.
3. Predict cell- and tissue-specific molecular effects.
4. Develop hypotheses about genes, pathways, and disease mechanisms.
5. Validate those hypotheses using CRISPR, MPRA, patient RNA-seq, or other experiments.

AlphaGenome Atlas primarily supports the stages from variant prioritization to the generation of mechanistic hypotheses.

Identifying Causal Variants in Rare Diseases#

In rare disease studies, WGS may identify millions of variants in a patient.

Determining which of those variants might contribute to disease can be extremely challenging.

Noncoding variants are particularly difficult to interpret because they do not necessarily change the amino acid sequence of a protein.

AlphaGenome Atlas provides predictions about which molecular processes may be affected by candidate variants.

In the DNM1 case, for example, the analysis supported variant prioritization and generated a specific hypothesis involving the creation of a new splice site. [1]

A potential workflow is:

1. Identify high-priority candidate variants using AVI.
2. Investigate predicted molecular effects, such as abnormal splicing.
3. Test predictions through RNA assays or other functional experiments.
4. Integrate experimental results with clinical evidence.

However, AVI and AlphaGenome predictions alone cannot establish clinical pathogenicity.

The Atlas study explicitly positions the resource as a research tool whose predictions should form part of a broader evidence chain rather than serve as standalone diagnostic conclusions. [1]

Implications for Drug Discovery#

Understanding the functional effects of genetic variation is also important for drug discovery.

GWAS may identify genomic regions associated with disease without clearly identifying which variants, genes, or regulatory mechanisms are causally involved.

AlphaGenome Atlas can support hypothesis generation by predicting whether candidate variants alter transcription, splicing, or chromatin-related processes.

This may help prioritize experiments and investigate potential therapeutic targets.

However, an important distinction remains:

Predicting a disease-associated molecular change is not equivalent to predicting the therapeutic consequences of pharmacologically modifying that molecular process.

Drug response depends on additional factors, including intercellular interactions, pharmacokinetics, toxicity, compensatory mechanisms, and disease progression.

Therefore, AlphaGenome Atlas may improve research efficiency in target discovery, but it should not be interpreted as a tool that directly predicts drug efficacy or clinical treatment outcomes.

17. How Far Are We from Comprehensive Biological Simulation?#

AlphaGenome Atlas is also relevant to the broader goal of biological simulation.

However, the term "simulation" requires careful interpretation.

A simplified view of biological information flow is:

DNA → Chromatin → RNA → Protein → Cell → Organ → Disease

In reality, these processes form a complex network involving feedback and interactions rather than a simple one-way chain.

AlphaGenome has advanced prediction capabilities primarily in the earlier stages of this network.

It has improved the ability to predict chromatin state, transcription, splicing, and related molecular phenotypes from DNA sequence.

However, there remains a substantial gap between predicting molecular phenotypes and predicting disease.

For example, a predicted change of 0.3 in an ATAC-seq signal does not directly tell us how much a particular patient's risk of developing a disease will increase.

To understand this gap, biological prediction can be divided into five stages.

Stage	Prediction target	Current status
1	Predict molecular function from DNA sequence	Major progress demonstrated by AlphaGenome
2	Predict molecular changes caused by nucleotide substitutions	Central focus of Atlas, with partial experimental validation
3	Predict individual-specific molecular states from personal genomes and cellular contexts	Major unresolved challenge
4	Simulate cellular and tissue dynamics, including drug responses	Outside the direct scope of Atlas
5	Predict disease onset, progression, and treatment response for individual patients	Requires multilevel causal models and clinical validation

From this perspective, AlphaGenome Atlas should not be described as a complete biological simulator.

A more precise description is:

A research infrastructure for predicting multimodal molecular responses to DNA sequence variation.

AlphaGenome Atlas primarily provides predictions for static changes to DNA sequences.

It does not directly simulate temporal changes in cellular states, feedback among molecular systems, or organ-level dynamics.

Nevertheless, increasingly accurate molecular predictions could become important components of more comprehensive biological simulation systems.

18. What Needs to Be Validated Next?#

The next stage of AlphaGenome Atlas research should not focus solely on increasing the number of variants with available scores.

It should also determine how accurately these predictions correspond to actual biological phenomena.

Seven validation questions are particularly important.

1. Are the Directions and Magnitudes of Predicted Molecular Effects Correct?#

Previously unseen variants should be introduced experimentally, using CRISPR or related methods, and evaluated through RNA-seq, ATAC-seq, splicing assays, or other measurements.

Validation should assess not only variant ranking but also the direction and quantitative magnitude of molecular changes.

2. Can Personal Genomes Predict Inter-Individual Expression Differences?#

Independent cohorts with matched WGS and RNA-seq data should be used to evaluate how accurately models predict expression differences between individuals for the same gene.

This is especially important for precision medicine.

3. Can Predictions Generalize to Unseen Cell Types and Conditions?#

Predictions should be evaluated in cell types, inflammatory conditions, developmental stages, and drug exposure settings not represented in training data.

Comparisons with models that explicitly incorporate cellular state would also be valuable.

4. Can Models Predict Interactions Between Multiple Variants and Haplotypes?#

Experiments should compare the effects of variant A, variant B, and their combination.

This would help determine whether predicted effects are additive or whether nonlinear interactions must be modeled explicitly.

5. Which Information Sources Explain AVI's Performance Gains?#

Models using only AlphaGenome features, only evolutionary conservation, only AlphaMissense, and all available features should be compared.

Such analyses would clarify the independent contributions of molecular function prediction and evolutionary constraint.

The Atlas study already includes ablation analyses, but broader evaluations using independent datasets and application-specific benchmarks would provide additional evidence. [1]

6. How Accurately Can AVI Prioritize Previously Unclassified Variants?#

Prospective validation should focus on variants of uncertain significance (VUS), rather than relying exclusively on previously classified pathogenic and benign variants.

Predictions made before new evidence becomes available should be compared against subsequent functional and clinical findings.

Performance should also be assessed across disease categories, genomic regions, and ancestry groups.

7. Are Predicted Molecular Mechanisms Causally Correct?#

When AVI or AlphaGenome suggests that a transcription factor, enhancer, or splicing process mediates a variant's effect, the proposed mechanism should be experimentally tested.

For example, researchers could perturb the implicated transcription factor and determine whether the predicted variant effect changes or disappears.

This would help distinguish statistical feature attribution from biological causality.

Prioritizing the Most Fundamental Questions#

Among these questions, the first two are particularly fundamental.

The first asks whether predicted variant effects correspond to real molecular changes.

The second asks whether those predictions generalize to actual personal genomes.

If the first remains unresolved, AlphaGenome Atlas may still be an excellent variant ranking system, but its reliability as a quantitative molecular simulator will remain limited.

If the second remains unresolved, a substantial gap will persist between genome-wide predictions and individualized molecular prediction for precision medicine.

19. The Future of Genomic AI May Integrate Three Research Directions#

The history reviewed here suggests that genomic AI is advancing along three major directions.

Direction 1: Functional Genomics AI#

Representative models include AlphaGenome, Borzoi, and scooby.

These models predict molecular phenotypes such as transcription, splicing, and chromatin state from DNA sequence.

They are particularly valuable for investigating the molecular mechanisms affected by genetic variation.

Direction 2: Genomic Language and Foundation Models#

Representative models include Evo 2 and Nucleotide Transformer.

These approaches learn general representations from large-scale DNA sequence data.

Their potential applications extend beyond molecular function prediction to understanding genomic patterns and generating new sequences.

Direction 3: Evolutionary Constraint AI#

GPN-Star is a representative example.

This approach uses evolutionary information and genomic sequence patterns to estimate functional constraints and variant deleteriousness.

These directions are not mutually exclusive.

They learn from different sources of evidence and are often suited to different prediction tasks.

Some models already combine elements from more than one tradition.

A 2026 review of sequence-to-function research identifies generalization to genetic variation and cellular context, as well as interpretability, as major unresolved challenges. [15]

It is therefore plausible that future genomic AI systems will increasingly integrate information capable of answering three questions:

• Is the variant evolutionarily unusual? How important or constrained is the altered sequence?
• Which molecular mechanisms are affected? Does the variant influence transcription, splicing, chromatin, or another function?
• In which cellular contexts does the effect occur? Which cell types, states, or environmental conditions reveal the effect?

Beyond these questions, it will also be necessary to determine whether the predictions generalize to actual individuals and explain clinically meaningful disease mechanisms.

If these capabilities can be integrated and validated, they could contribute to more comprehensive genomic AI and biological simulation systems.

However, this remains a research direction rather than an already demonstrated achievement.

20. Conclusion: The Historical Significance and Current Position of AlphaGenome Atlas#

The historical significance of AlphaGenome Atlas lies in transforming sequence-to-function genomics into a genome-wide scientific research infrastructure.

DeepSEA demonstrated the potential to predict regulatory function and variant effects directly from DNA sequence.

Enformer extended this approach toward long-range gene regulation.

Sei organized many regulatory predictions into interpretable functional classes.

Borzoi and AlphaGenome advanced the integration of molecular phenomena such as gene expression, splicing, and chromatin state.

AlphaGenome Atlas then applied these capabilities to approximately 9 billion possible single-nucleotide substitutions, making large-scale molecular effect predictions available as a research resource.

In this sense, describing AlphaGenome Atlas as the "AlphaFold Database moment" for sequence-to-function genomics is a useful analogy.

However, the broader idea of precomputed genome-wide variant scores predates Atlas, with important predecessors such as CADD and dbNSFP.

The more fundamental novelty of AlphaGenome Atlas is that it combines genome-wide molecular predictions with evolutionary constraints and protein-impact information, connecting variant prioritization with mechanistic interpretation.

At the same time, its scientific interpretation requires care.

Strong AVI performance in distinguishing known pathogenic variants does not establish that AlphaGenome accurately simulates every molecular consequence of genetic variation.

Likewise, strong performance on reference-genome molecular prediction tasks does not automatically demonstrate accurate prediction of genetic effects in individual patients.

At least four distinct research goals should be separated:

1. Ranking variants by their potential functional importance.
2. Accurately predicting the actual molecular effects of variants.
3. Explaining causal pathways from molecular changes to disease.
4. Predicting disease progression and treatment response in individual patients.

These goals are related, but they are not equivalent.

AlphaGenome Atlas is particularly important as a technology connecting the first two.

In rare disease research and precision medicine, it may help researchers prioritize candidate variants and generate specific hypotheses about molecular mechanisms.

In drug discovery, it may contribute to the interpretation of disease-associated variants and the prioritization of targets and experiments.

However, models such as GPN-Star, Nucleotide Transformer, Evo 2, scooby, and GET offer complementary capabilities based on evolutionary constraints, foundation modeling, and cellular context.

Furthermore, biological function in living organisms depends on genetic background, cellular state, environmental conditions, temporal dynamics, and complex molecular interactions.

AlphaGenome Atlas is therefore not a complete biological simulator.

It is better understood as an important scientific infrastructure for predicting—and systematically validating—the relationship between DNA sequence variation and molecular function.

The most valuable next step may not be simply assigning scores to even more variants.

Rather, it may be determining how accurately the predicted molecular effects can be reproduced across previously unseen variants, personal genomes, cellular states, and environmental conditions.

AlphaGenome Atlas provides an important foundation for pursuing this objective.

It is also worth noting that the AlphaGenome Atlas study was released as a medRxiv preprint in September 2026, separately from the AlphaGenome Nature paper published in 2026. [1][7]

Its current evidence base relies substantially on evaluations reported by the authors, and further independent validation—particularly regarding generalization to personal genomes and unseen cellular conditions—remains important.

Perhaps the greatest significance of AlphaGenome Atlas is not merely that it provides predictions for approximately 9 billion variants, but that it establishes a foundation for systematically investigating how much of real biology those predictions can explain.

References#

[1] Cheng, J. et al., “AlphaGenome Atlas: in silico mutagenesis of the entire human genome improves prioritization and interpretation of non-coding variants,” medRxiv, preprint, 2026.
https://doi.org/10.64898/2026.09.16.26363192
https://pmc.ncbi.nlm.nih.gov/articles/PMC13596691/

[2] 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/

[3] 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

[4] 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

[5] 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

[6] 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

[7] Avsec, Ž. et al., “Advancing regulatory variant effect prediction with AlphaGenome,” Nature, 649, 1206–1218, 2026.
https://doi.org/10.1038/s41586-025-10014-0

[8] 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

[9] “Genome modelling and design across all domains of life with Evo 2,” Nature, 2026.
https://doi.org/10.1038/s41586-026-10176-5

[10] “Benchmarking DNA foundation models for genomic variant prediction,” Nature Communications, 2025.
https://doi.org/10.1038/s41467-025-65823-8

[11] Ye, C. et al., “Predicting genome-wide functional constraints with GPN-Star,” Nature, 2026.
https://doi.org/10.1038/s41586-026-11005-5
https://www.nature.com/articles/s41586-026-11005-5

[12] “Single-cell sequence-to-function modeling with scooby,” Nature Methods, 2025.
https://www.nature.com/articles/s41592-025-02854-5

[13] “Generalizable prediction of cellular gene expression from regulatory sequence,” Nature, 2025.
https://www.nature.com/articles/s41586-024-08391-z

[14] “Context-aware sequence-to-function modeling,” Nature Communications, 2026.
https://www.nature.com/articles/s41467-026-75527-2

[15] “Sequence-to-function models in regulatory genomics: progress and challenges,” Nature Genetics, 2026.
https://www.nature.com/articles/s41588-026-02670-3

[16] Huang, C. et al., “Personal transcriptome variation is poorly explained by current genomic deep learning models,” Nature Genetics, 55, 2056–2059, 2023.
https://www.nature.com/articles/s41588-023-01574-w

[17] Sasse, A. et al., “Benchmarking of deep neural networks for predicting personal gene expression from DNA sequence highlights shortcomings,” Nature Genetics, 2023.
https://www.nature.com/articles/s41588-023-01524-6