Target Audience

  • Researchers and readers interested in the evolution of machine learning in genomics and the historical significance of AlphaGenome Atlas.
  • Those seeking a deeper understanding of the implicit assumptions, predictive accuracy, limitations, and unresolved questions surrounding DNA sequence-to-function models.
  • Those interested in applying genomic AI to precision medicine, rare disease diagnosis, drug discovery, and biological simulation.

Summary

  • The fundamental assumption underlying AlphaGenome Atlas is that the molecular consequences of a single-nucleotide substitution can be inferred from differences in a trained model's predictions, and that these predictions can generalize to unseen variants, cells, and individuals. This assumption encompasses several distinct hypotheses, including the predictability of molecular function from DNA sequence, the correspondence between predicted differences and actual intervention effects, and the ability to represent cellular context and genetic background.
  • AlphaGenome Atlas represents a convergence of two historical research traditions: sequence-to-function modeling, from DeepSEA through Enformer, Borzoi, and AlphaGenome, and approaches that evaluate variant importance using evolutionary conservation and genetic constraint. In particular, the AlphaGenome Variant Impact (AVI) score integrates AlphaGenome's molecular predictions with evolutionary conservation, AlphaMissense, and population genetic information.
  • Performance in variant prioritization must be distinguished from the ability to accurately predict molecular changes. The original Atlas study reports promising results using ClinVar, genome-editing experiments, rare disease cases, and UK Biobank data. However, these results do not establish that the causal molecular effects of every predicted variant have been experimentally validated.
  • The most fundamental unresolved challenge is whether sequence-based rules learned from reference genomes generalize to actual variant effects in individual genomes, cellular states, and environmental conditions. Major open questions include inter-individual gene expression differences, unseen cellular states, distal regulation, interactions among multiple variants, and causal relationships between molecular changes and disease.
  • AlphaGenome Atlas is best understood not as a completed biological simulator, but as a multimodal molecular response prediction resource for DNA sequence variants. Future progress will depend on quantitatively validating predicted molecular effects using independent perturbation experiments and individual-level genomic data.

1. What Does AlphaGenome Atlas Implicitly Assume?

AlphaGenome Atlas represents an important step toward transforming machine learning-based genome function prediction into a genome-wide research infrastructure.

Its technological lineage extends from DeepSEA through Enformer and Borzoi to AlphaGenome, within the broader sequence-to-function modeling paradigm.

The AlphaGenome Atlas study, released in September 2026, precomputed the predicted molecular effects of approximately 9 billion possible single-nucleotide substitutions across the human genome using AlphaGenome.[1]

It also introduced the AlphaGenome Variant Impact (AVI) score, which integrates molecular function predictions, evolutionary conservation, predicted protein effects, and other information to connect variant prioritization with molecular mechanism interpretation.

This represents a significant achievement in machine learning genomics.

However, examining its historical development reveals several assumptions that are not always explicitly articulated.

The most fundamental assumption is that the molecular consequences of changing a single nucleotide can be inferred from differences in the predictions of a trained model, and that these inferences can generalize to previously unseen variants, cells, and individuals.

This overarching assumption can be divided into at least five hypotheses.

Hypothesis Implicit Assumption Key Question Requiring Validation
Hypothesis 1: Molecular function can be predicted from DNA sequence Gene expression and chromatin states can be explained primarily from cis-regulatory DNA sequence information Do predictions generalize to unseen sequences and cellular conditions?
Hypothesis 2: Predicted differences reflect actual variant effects Differences between REF and ALT predictions correspond to the direction and magnitude of experimentally observed changes Do predictions agree with actual effects of genome editing?
Hypothesis 3: Complex biological contexts can be compressed Numerous molecular phenotypes, cellular states, and genetic backgrounds can be approximated by learned models and scores Is biologically important information lost during this compression?
Hypothesis 4: Molecular function and evolutionary constraint indicate variant importance Functional predictions, conservation, and population frequency can help prioritize deleterious variants Can functional importance be distinguished from pathogenicity?
Hypothesis 5: Molecular changes can inform disease research Predicted functional effects can guide the identification of candidate causal variants and experimental targets Can causal relationships between molecular changes and disease be established?

Importantly, these five hypotheses have not been validated to the same extent.

AlphaGenome Atlas has demonstrated promising results in variant prioritization and in predicting certain molecular effects.

However, generalization to individual-specific gene expression, previously unseen cellular states, and disease probabilities remains insufficiently validated.

Another important point is that the AVI score is not simply an output of AlphaGenome.

It also incorporates evolutionary conservation, AlphaMissense, and training based on population genetic information.

Therefore, high AVI performance must be evaluated separately from the question of whether AlphaGenome accurately simulates molecular function from DNA sequence. [1][2]

2. Implicit Assumptions Revealed by the History of Genomic AI

Understanding the historical significance of AlphaGenome Atlas requires considering not only improvements in predictive performance, but also the evidence on which different models base their predictions.

Two important research traditions have developed within machine learning genomics.

The first aims to predict experimentally measurable molecular functions directly from DNA sequence.

The second evaluates the functional importance or deleteriousness of genetic variants using evolutionary conservation and population genetic information.

AlphaGenome Atlas can be understood as a technology that connects these two traditions.

2.1 Reinterpreting the Research History Through the Sources of Biological Knowledge

The history of genomic AI can be organized according to the information each approach learns from or uses, together with its underlying assumptions.

Research Tradition Basic Principle Implicit Assumption
PhyloP, PhastCons Evolutionarily conserved sequences are functionally important The history of natural selection contains information about functional importance
CADD, GWAVA Integrate diverse annotations to evaluate variants Known genomic features can be used to infer the deleteriousness of unseen variants
DeepSEA, Sei Learn regulatory function directly from DNA sequence DNA contains reusable regulatory grammar
Enformer Predict gene regulation using long-range sequence context Distal regulatory effects can be inferred from sufficiently long DNA sequences
Borzoi, AlphaGenome Integrate multiple molecular phenotypes Diverse molecular processes can be predicted from shared sequence information
AlphaGenome Atlas / AVI Integrate molecular predictions with evolutionary and other constraints A combined functional impact score can improve variant prioritization
GPN-Star Learn functional constraints from phylogenetic and multispecies sequence information Evolutionary history contains information useful for predicting deleteriousness

This perspective highlights that AlphaGenome Atlas does not simply replace earlier approaches.

Instead, two major research traditions converge in the AVI score.

The functional genomics lineage

DeepSEA → Enformer → Borzoi → AlphaGenome

This tradition predicts transcription, splicing, chromatin states, and other molecular functions from DNA sequence.

The evolutionary and variant assessment lineage

PhyloP / PhastCons → Integrated scores such as CADD → Variant assessment technologies such as AlphaMissense

This tradition evaluates variant importance and deleteriousness using evolutionary constraint, functional annotations, and predicted protein effects.

AVI combines information from these traditions to support variant prioritization and molecular interpretation.[1]

An important historical qualification is necessary here.

The idea of precomputing genome-wide variant scores and making them searchable did not originate with AlphaGenome Atlas.

Earlier resources, including CADD and dbNSFP, already provided precomputed variant scores. Sei also presented a genome-wide map of regulatory activity.

The distinctive contribution of AlphaGenome Atlas is therefore not precomputation alone.

Rather, its novelty lies in providing large-scale, allele-specific predictions across diverse molecular modalities and connecting those predictions to variant prioritization and mechanistic interpretation.

2.2 Implicit Assumption 1: DNA Contains a Generalizable Regulatory Grammar

This assumption is shared by sequence-to-function models developed since DeepSEA.

DeepSEA established an important foundation for modern genomic deep learning by predicting transcription factor binding, chromatin features, and histone modifications directly from DNA sequence.[3]

For example, when a particular DNA sequence contains a transcription factor binding motif, a model learns relationships between that sequence and experimentally measured binding or expression patterns.

The implicit assumption is that sequence rules learned in one genomic region will remain applicable, at least to some extent, to other regions and previously unseen variants.

However, several biological complications arise.

  • Transcription factor binding depends not only on sequence motifs, but also on whether the relevant transcription factor is present in the cell.
  • The same motif can have different effects depending on surrounding sequences and chromatin states.
  • Gene expression involves nonlinear interactions among multiple regulatory factors.
  • The functional role of a sequence may change across developmental stages or stimulation conditions.

AlphaGenome learns some of these complex relationships.

However, specifying DNA sequence alone does not uniquely determine regulatory function under every possible cellular condition.

The AlphaGenome Atlas paper also explicitly identifies its inability to directly model trans-acting mechanisms, including changes in transcription factor expression, as a limitation.[1]

Consequently, the extent to which learned regulatory grammar generalizes across cellular states and environmental conditions remains an important validation challenge.

2.3 Implicit Assumption 2: Predicted Differences Correspond to Experimental Intervention Effects

The central computation underlying AlphaGenome Atlas is the difference between predictions for a reference allele (REF) and an alternative allele (ALT).

[ \Delta\hat{Y}=\hat f(S_{\mathrm{ALT}})-\hat f(S_{\mathrm{REF}}) ]

Here:

  • (\hat f) represents the trained prediction model.
  • (S_{\mathrm{REF}}) represents the reference sequence.
  • (S_{\mathrm{ALT}}) represents the variant sequence.
  • (\Delta\hat Y) represents the predicted molecular effect.

However, the quantity researchers ultimately want to understand is the actual molecular change caused by substituting a nucleotide under the same experimental conditions and genetic background.

[ \Delta Y_{\mathrm{true}}= Y(\operatorname{do}(\mathrm{ALT}))

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

These two quantities are not necessarily equivalent.

Even if a model accurately predicts RNA-seq or ATAC-seq patterns along the reference genome, this does not guarantee that its predicted differences correspond to the actual causal effects of genome editing.

This distinction is historically important.

A 2023 Nature Genetics study by Huang and colleagues evaluated four existing sequence-based models and found that although they achieved reasonable accuracy in predicting differences in expression levels across genes, their ability to predict inter-individual differences in expression for the same gene was weak. The models could also predict the wrong direction of variant effects.[4]

A separate study by Sasse and colleagues also identified shortcomings in predicting individual gene expression from DNA sequence.[5]

These studies did not directly evaluate AlphaGenome Atlas.

Nevertheless, they highlight a fundamental challenge inherited by the sequence-to-function modeling paradigm.

Predicting molecular function along a reference genome and predicting the causal effect of introducing a genetic variant are different problems.

2.4 Implicit Assumption 3: The Effect of a Single Variant Can Be Evaluated Independently of Other Genetic Backgrounds

The approximately 9 billion SNVs in AlphaGenome Atlas are primarily evaluated by changing one nucleotide at a time within the reference genome.

Consequently, each prediction primarily addresses the question:

"What happens if this nucleotide is changed within the reference sequence background?"

However, real individuals carry many genetic variants simultaneously.

For example, when two variants, A and B, occur within the same regulatory region, their combined effect may satisfy:

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

This is the problem of epistasis and haplotype dependence.

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

However, simply combining the scores of individual SNVs registered in the Atlas does not necessarily provide an accurate prediction of the effects of a person's complete genome.

This creates a substantial gap between reference-genome-based predictions and individual-genome modeling.

2.5 Implicit Assumption 4: Rare Variants Provide a Training Signal for Identifying Functionally Important Variants

This is an important feature of AVI.

AVI is not constructed solely from AlphaGenome's molecular function predictions.

According to the Atlas paper, it integrates the following information.[1]

  • 10 molecular function features derived from AlphaGenome
  • 4 features associated with AlphaMissense and protein loss of function
  • 2 evolutionary conservation features
  • 2 insertion/deletion-related features

AVI also uses population allele frequency information during training to distinguish variants expected to have larger functional impacts from those expected to be relatively neutral.

Specifically, the model uses gnomAD population frequencies, treating rare variants as proxy labels for functional importance.

This approach is grounded in the population genetic principle that strongly deleterious variants are more likely to be removed from populations by natural selection.

However, the following relationship does not necessarily hold:

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

A variant can be rare for reasons unrelated to pathogenicity, including its time of origin, demographic history, genetic drift, and ancestral population structure.

Likewise, disease-causing variants are not necessarily subject to strong evolutionary constraint.

AVI therefore assumes that proxies for evolutionary deleteriousness can also serve as useful proxies for clinical or molecular importance.

This is a reasonable approach, but not a universally valid assumption.

2.6 Implicit Assumption 5: High-Dimensional Molecular Information Can Be Compressed into a Single Score

The Atlas paper reports that an average of approximately 27,000 experiment- and condition-specific scalar predictions are computed for each variant.[1]

AVI aggregates this information across modalities and uses features such as the maximum predicted effect for each modality across cell types and tissues.

This is operationally convenient, but it can also result in the loss of biologically important information.

Consider two hypothetical variants.

Variant A Variant B
Hepatocytes Large decrease in expression No effect
Neurons No effect Large decrease in expression
Maximum effect magnitude Large Large

When only the maximum effect magnitude is considered, variants A and B may appear similarly important.

However, when investigating a neurological disease, their biological relevance can be very different.

Furthermore, even when SHAP is used for AVI feature attribution, SHAP explains the contributions of features to the model's score. It does not establish the underlying biological causal pathway.

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

  • The cell types in which an effect is predicted
  • The genes and molecular phenotypes predicted to be affected
  • The direction of the predicted changes
  • The magnitude of the predicted effects

3. Challenges, Limitations, and Unvalidated Aspects of AlphaGenome Atlas

When assessing AlphaGenome Atlas, it is important to distinguish among different types of unresolved issues.

One category consists of acknowledged limitations explicitly identified in the original publications.

Another consists of unvalidated questions for which current evaluations do not provide sufficient evidence.

A third category includes questions for which promising results exist, but broader applicability requires further validation.

Based on the literature available as of October 2026, the major issues can be summarized as follows.

Challenge Current Assessment Importance
Predicting inter-individual expression differences Not sufficiently benchmarked for AlphaGenome itself Very high
Predicting the direction and magnitude of actual variant effects Partially validated; broad generalization remains uncertain Very high
Generalization to unseen cell types and stimulation conditions Acknowledged limitation Very high
Trans-acting effects and distal enhancers Acknowledged limitation High
Generalization from single SNVs to multiple variants Limited validation scope High
Calibration of AVI pathogenicity and deleteriousness predictions Not established as actual clinical probabilities Very high
Multimodal integration and interpretability Partially validated; causal interpretation remains unresolved High
Missing data for noncoding RNAs and other biological processes Acknowledged limitation High
Generalization across genetic backgrounds and ancestral populations Requires sufficient independent validation High
Translation into clinical diagnosis and drug discovery outcomes Research stage Very high

The importance ratings reflect the assessment presented in this article and are not rankings assigned by the original authors.

3.1 The Most Fundamental Challenge: Predicting Reference-Genome Function Is Different from Predicting Individual Variant Effects

This is one of the most important limitations to consider.

The AlphaGenome Nature paper reports strong predictive performance across multiple molecular modalities.

However, the authors also explicitly state that prediction using individual genomes has not yet been benchmarked.[2]

In a 2023 study, Huang and colleagues evaluated models including Enformer using paired genomic and transcriptomic data from 421 individuals.[4]

They found a substantial difference between performance in predicting expression levels across genes and performance in predicting differences in expression of the same gene across individuals.

For the latter task, the distribution of correlation coefficients was centered near zero.

This distinction can be understood through two prediction tasks.

Task A: Predicting expression differences between genes

Under the same biological conditions, can a model predict whether gene X or gene Y is more highly expressed?

Existing models have demonstrated some predictive ability for this task.

Task B: Predicting expression differences between individuals

For the same gene X, can a model predict how much its expression differs between individual A and individual B?

Existing models have shown substantial limitations for this task.

Task B is particularly important for precision medicine.

The central question in precision medicine is not simply the function of an average reference genome, but how a specific patient's genome changes molecular function in that individual.

However, Huang and colleagues did not directly evaluate AlphaGenome Atlas.

Their results therefore do not establish that AlphaGenome Atlas necessarily has the same limitations.

Rather, they identify an important question: to what extent has AlphaGenome Atlas overcome the problems demonstrated in earlier models?

3.2 Cellular States and Environmental Conditions

AlphaGenome uses measurements from many cell types and tissues, but it does not represent every biological condition.

Limitations identified in the original studies include the following.[1][2]

  • Generalization to cell types not represented during training
  • Uneven representation of cell types across molecular modalities
  • Stimulation-dependent variant effects
  • Regulatory processes dependent on transcription factor expression states

For example, in an immune disorder, a variant might have little effect in a resting immune cell but a substantial effect after inflammatory stimulation.

In such a case, predictions based on the available reference conditions may not fully capture what occurs in the relevant disease environment.

This is not necessarily a problem that can be solved simply by increasing model size.

It also concerns how cellular states and environmental information should be incorporated into model inputs.

3.3 Long-Range Regulation Is Not Fully Reproduced

AlphaGenome can process approximately 1 Mb of DNA sequence.

However, being able to process a 1 Mb sequence does not mean that all long-range regulatory effects within that region can be accurately reproduced.

The AlphaGenome paper identifies the need to improve prediction accuracy for distal regulatory elements, particularly those located more than 100 kb away.[2]

Furthermore, interchromosomal interactions, three-dimensional genome organization, and transcription factor availability may not be fully determined by DNA sequence alone.

Consequently, predictions of long-range regulatory effects from AlphaGenome Atlas are valuable for hypothesis generation, but they do not directly establish causal relationships between enhancers and target genes.

3.4 Biological Processes Not Covered by Training Data, Including Noncoding RNAs

The Atlas paper identifies limitations associated with non-polyadenylated RNAs that are not captured by standard poly(A)-selected RNA-seq.[1]

The authors specifically discuss RNU4-2 as an example.

This illustrates a broader issue.

Even a highly capable prediction model may not adequately learn molecular processes that are not captured by the experimental assays used during training.

This is both a limitation of predictive modeling and a limitation inherited from the measurement technologies and datasets underlying functional genomics.

3.5 Covering Approximately 9 Billion SNVs Does Not Mean Covering the Entire Variant Space

AlphaGenome Atlas precomputes predictions for approximately 9 billion single-nucleotide substitutions.

However, the September 2026 Atlas paper also analyzes numerous observed insertions and deletions from population datasets, in addition to the comprehensive SNV analysis.[1]

It would therefore be inaccurate to conclude that the research addresses no variants other than SNVs.

At the same time, comprehensive coverage of SNVs does not imply complete coverage of:

  • Structural variants
  • Repeat expansions
  • Complex insertions and deletions
  • Haplotypes containing multiple variants
  • Individual-specific genomic backgrounds

Furthermore, the publicly distributed dataset does not necessarily include every category of analysis reported in the paper.

The data release information examined as of October 2026 indicates that the downloadable Atlas data has restrictions concerning coverage and conditions of use.[1]

Thus, precomputing approximately 9 billion SNVs is an important achievement, but it does not by itself provide a complete prediction of all effects arising from an individual's genome.

4. What Has Actually Been Validated?

It would be inaccurate to state that AlphaGenome Atlas has not been validated.

The original paper reports evaluations involving variant prioritization, experimental measurements, rare disease cases, and population genetic association analyses.[1]

However, different types of validation support different conclusions.

4.1 Representative Results Reported in the Original Study

The AlphaGenome Atlas paper reports the following results.[1]

Evaluation Reported Result Interpretation
ClinVar intronic SNVs AVI AUPRC of 0.76 Comparator GPN-Star-V: 0.44
ClinVar synonymous SNVs AVI AUPRC of 0.57 Comparator CADD v1.7: 0.35
Saturation genome-editing experiments AVI achieved the highest Spearman correlation in 8 of 10 experiments Evaluates agreement with measured functional effects
Rare noncoding variant association analysis 22% increase in the number of independent association signals discovered Improvement through Atlas-based variant filtering

All of these results are reported by the authors of the original Atlas study.

The evaluation sets, comparison models, and metrics differ, so the numerical results should not be directly compared with one another.

In an experiment involving DNM1, the researchers also evaluated 796 variants within a 265-base-pair region across five cell lines.

The combined AlphaGenome splicing prediction and AVI splicing attribution achieved an AUPRC of 0.943, demonstrating performance close to that of SpliceAI and Pangolin.[1]

These are meaningful results.

Nevertheless, success within specific genomic regions and experimental conditions cannot automatically be generalized to every region of the genome.

4.2 What the Validations Establish—and What They Do Not

The distinction between supported conclusions and unresolved questions can be summarized as follows.

Validation Type What It Supports What It Does Not Yet Establish
Comparison with ClinVar Ability to distinguish known pathogenic and benign variants Diagnostic accuracy for previously unseen patients
Genome editing experiments Agreement between predictions and measured molecular effects for selected variants Quantitative accuracy across the entire genome and all cellular states
DNM1 case study Utility in interpreting an actual disease-related case Improved diagnostic yield across rare diseases generally
UK Biobank analysis Functional filtering can increase the discovery of genetic associations Correct molecular mechanisms for individual variants
SHAP-based explanations Identification of features contributing to AVI scores Causal pathways connecting variants to disease
Integration of evolutionary conservation Improved variant assessment performance Accuracy of AlphaGenome alone as a biological simulator

A particularly important concern is the potential overlap or circularity of information used in AVI evaluation.

AVI uses population frequencies as proxy labels during training and includes features such as evolutionary conservation.

Meanwhile, ClinVar pathogenicity classifications may themselves incorporate population frequency and previously available functional information as evidence.

Therefore, even when training variants and evaluation variants do not directly overlap, related types of information may influence both.

This does not imply that the evaluation is invalid.

However, a high AUPRC on ClinVar should not be interpreted as direct evidence that molecular function predictions have been independently validated through biological experiments.

4.3 What the Comparison with GPN-Star Reveals

GPN-Star provides an important comparison for understanding the role of AlphaGenome Atlas.

GPN-Star models functional constraints using evolutionary history and genomic sequence patterns.

Its Nature paper, published in September 2026, reported benchmark results in which it outperformed AlphaGenome-derived sequence-to-function predictions and other methods for certain noncoding pathogenicity prediction tasks.[6]

However, AVI, as introduced in the Atlas paper, is a different model that integrates AlphaGenome predictions with evolutionary conservation, AlphaMissense, and other features.

The Atlas paper reports that AVI outperformed GPN-Star on some ClinVar benchmarks.[1]

Thus, both of the following statements can be true:

  • GPN-Star outperformed AlphaGenome alone in certain benchmarks.
  • AVI outperformed GPN-Star in other benchmarks.

The key point is that predicting molecular mechanisms and accurately predicting pathogenicity are not the same task.

Similarly, AlphaGenome's standalone performance must be distinguished from the performance of AVI, which incorporates additional information.

Results obtained using different datasets and evaluation metrics do not justify concluding that either GPN-Star or AVI is universally superior.

5. Seven Questions That Require Further Validation

To advance AlphaGenome Atlas and expand its applications in precision medicine and drug discovery, seven questions deserve particular attention.

5.1 Do Predicted Molecular Changes Agree with Experimental Measurements in Direction and Magnitude?

The first question is whether the molecular effects predicted by the model correspond quantitatively to experimentally observed effects.

Previously unseen variants could be introduced using CRISPR-based genome editing and evaluated using RNA-seq, ATAC-seq, splicing assays, and related measurements.

Importantly, evaluation should go beyond the rank correlation of variants.

It should examine:

  • Whether the direction of change is correct
  • Whether the predicted effect magnitude matches the measured effect
  • How prediction errors are distributed
  • Whether performance differs among cell types and molecular modalities

Such validation would help determine whether AlphaGenome Atlas functions not only as a variant-ranking system, but also as a quantitative molecular effect predictor.

5.2 Can Individual Genomes Predict Inter-Individual Gene Expression Differences?

Independent cohorts with matched whole-genome sequencing and RNA-seq data could be used to evaluate how well the model explains differences in the expression of the same gene across individuals.

This is an area where previous sequence-to-expression models have shown substantial limitations.[4][5]

The critical evaluation is not whether the model predicts average expression differences between genes.

It is whether the model captures variation in the same gene among different individuals.

This distinction is essential for applications in precision medicine.

5.3 Can the Model Generalize to Unseen Cells and Stimulation Conditions?

The model should be evaluated in cell types, immune stimulation conditions, and developmental stages that differ from those represented during training.

Its performance could then be compared with models that explicitly incorporate cellular state information.

For example, predictions for the same variant could be evaluated in resting immune cells and in cells exposed to inflammatory stimulation.

Such experiments would help quantify the limits of DNA-sequence-only models and the value of including cellular context as an additional input.

5.4 Can the Model Reproduce Interactions Among Multiple Variants and Haplotypes?

The experimentally measured effects of variant A, variant B, and the combined A+B genotype could be compared.

This would allow researchers to evaluate whether predicted effects are additive or whether nonlinear interactions occur.

When multiple variants are present simultaneously, their combined molecular effects may not equal the sum of their individual effects.

Validating these interactions is necessary for moving from single-SNV predictions on a reference genome toward modeling actual personal genomes.

5.5 Which Information Sources Account for AVI's Performance Improvements?

AVI combines AlphaGenome-derived functional features with evolutionary conservation, AlphaMissense, and other information.

It is therefore important to distinguish the contributions of these information sources.

An ablation study could compare models using:

  • AlphaGenome-derived features only
  • Evolutionary conservation only
  • AlphaMissense only
  • All features combined

Such comparisons would help determine how much independent information is provided by molecular function predictions and evolutionary constraint.

In particular, understanding how much of AVI's performance comes specifically from AlphaGenome's molecular predictions is essential for accurately characterizing the scientific contribution of AlphaGenome Atlas.

5.6 Are High-AVI Variants More Likely to Be Confirmed as Disease-Causing?

Prospective validation should be conducted using previously unclassified variants, including variants of uncertain significance (VUS).

For example, AVI-based rankings could be established in advance and subsequently evaluated using independent functional experiments and clinical evidence.

Performance should also be assessed separately across disease categories and genomic regions.

Strong results on benchmarks composed of previously classified ClinVar variants do not necessarily imply equally strong performance in identifying causal variants in new patients.

5.7 Are the Predicted Molecular Mechanisms Causally Correct?

AVI feature attribution, including SHAP, can indicate which molecular features contribute to a variant's score.

However, this does not establish that the proposed biological causal pathway is correct.

For example, if a variant is predicted to affect a particular transcription factor binding site or enhancer, researchers could experimentally perturb the implicated transcription factor or enhancer to determine whether it mediates the variant's effect.

Such experiments would help move from predictive interpretation toward causal molecular explanation.

5.8 Which Validation Questions Should Be Prioritized?

Among these seven questions, two are especially fundamental.

First, do predicted molecular changes agree with measured changes in direction and magnitude?

If this question remains unresolved, AlphaGenome Atlas may be a powerful variant-ranking system, but its reliability as a quantitative molecular simulator will remain limited.

Second, can an individual's genome predict individual-specific molecular differences?

Without progress on this question, a substantial gap will remain between current sequence-based prediction and the goals of personalized genomic medicine.

6. How Should AlphaGenome Atlas Be Positioned in the Evolution of Biological Simulation?

AlphaGenome Atlas can be viewed as an important advance toward biological simulation.

However, the meaning of "simulation" must be carefully defined.

Predicting molecular functions from DNA sequence is not equivalent to reproducing the temporal dynamics of living systems or predicting disease development.

6.1 Five Stages from Molecular Prediction to Biological Simulation

The broader research trajectory can be organized into five stages.

Stage Prediction Target Current Status
Stage 1: Molecular function prediction Predict RNA, chromatin, and other molecular phenotypes from DNA sequence Central area in which AlphaGenome has demonstrated progress
Stage 2: Variant intervention effect prediction Predict molecular changes caused by single-nucleotide substitutions Core objective of Atlas; partially validated experimentally, but broad generalization remains unresolved
Stage 3: Individual-specific molecular state prediction Incorporate personal genomes, haplotypes, cellular states, and environmental conditions Major unresolved challenge
Stage 4: Dynamic simulation of cells and tissues Model temporal changes, cell-cell interactions, feedback, and drug responses Outside the direct scope of AlphaGenome Atlas
Stage 5: Disease and treatment outcome prediction Predict disease risk, progression, and treatment effects for individual patients Requires additional multilevel causal models and clinical validation

From this perspective, rather than calling AlphaGenome Atlas a "biological simulator," it is more accurate to describe it as a multimodal molecular response prediction resource for DNA sequence variants.

Atlas primarily provides predictions of molecular responses to static sequence changes.

It does not directly simulate the temporal evolution of cellular states or dynamic molecular feedback loops.

6.2 Similarities and Differences Between AlphaGenome Atlas and the AlphaFold Database

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

Both involve large-scale precomputation of AI predictions and their distribution as research resources that can be broadly accessed by scientists.

In this sense, it is reasonable to describe AlphaGenome Atlas as the transformation of sequence-to-function genomics into a research infrastructure.

However, the prediction targets and validation requirements differ.

Predicted variant effects in AlphaGenome Atlas depend strongly on cellular conditions and individual genetic backgrounds.

Therefore, the success of the AlphaFold Database cannot itself be taken as evidence for the biological validity of AlphaGenome Atlas.

The ability to distribute large-scale AI predictions as a scientific resource is distinct from the question of how accurately those predictions reproduce actual biological phenomena.

7. Overall Assessment: Scientific Significance and Future Challenges of AlphaGenome Atlas

The scientific position of AlphaGenome Atlas can be summarized in three major conclusions.

7.1 Its Contribution as a Scientific Research Infrastructure Is Substantial

AlphaGenome Atlas precomputes variant effect predictions across many molecular modalities at genome-wide scale and connects variant prioritization with molecular mechanism interpretation.

This represents a meaningful scientific contribution.

Previously, researchers often needed to identify candidate variants and then run prediction models individually to investigate their functional consequences.

By providing precomputed results for approximately 9 billion SNVs, Atlas makes molecular function prediction more accessible for large-scale variant analysis.

In addition, AVI provides an integrated variant-ranking score together with information about the features contributing to its predictions.

This advances genomic AI from isolated predictive models toward broadly usable research infrastructure.

7.2 Strong AVI Performance Does Not Prove That AlphaGenome Fully Understands Biology

AVI uses evolutionary constraint and population genetic information in addition to molecular function predictions.

Consequently, successful variant ranking and accurate prediction of molecular mechanisms must be evaluated independently.

For example, when AVI assigns a high score to a variant, it is important to determine whether that score is driven by AlphaGenome's molecular predictions, evolutionary conservation, AlphaMissense, or other information.

Furthermore, whether the variant actually causes disease remains a separate question requiring validation.

The usefulness of AVI as a variant-ranking system is not equivalent to demonstrating that AlphaGenome is an accurate simulator of biological processes.

7.3 The Most Fundamental Unresolved Challenge Is Generalization Across Individuals, Cells, and Environments

The deepest unresolved question underlying AlphaGenome Atlas is whether regulatory rules learned from genomic sequences can generalize to actual intervention effects in individual genomes, cellular states, and environmental contexts.

This encompasses several challenges:

  • Generalization from reference genomes to personal genomes
  • Generalization from individual variants to multiple variants and haplotypes
  • Generalization from observed cell types to unseen cellular states
  • Generalization from predicted molecular effects to actual causal intervention effects
  • Connecting molecular changes to disease and treatment outcomes

These problems cannot necessarily be solved simply by expanding the Atlas dataset.

They concern fundamental aspects of the research paradigm itself.

7.4 What Matters Most for Future Research?

Ultimately, four distinct capabilities need to be distinguished.

  1. Ranking variants according to their potential impact
  2. Accurately predicting the molecular changes caused by variants
  3. Explaining the causal pathways through which molecular changes produce disease
  4. Predicting disease progression and treatment responses for individual patients

These objectives form a connected research trajectory, but they are not the same problem.

AlphaGenome Atlas is best understood as a technology that helps bridge the first two stages.

The greatest value of future research may therefore lie not simply in assigning scores to additional variants, but in quantitatively validating predicted molecular effects through independent intervention experiments involving previously unseen variants and through analyses of individual genomic data.

AlphaGenome Atlas is not a completed biological simulator.

However, it provides an important foundation for predicting relationships between DNA sequence and molecular function at scale and for testing the validity of those predictions experimentally.

In the future, integrating evolutionary constraint, molecular function prediction, cellular states, personal genomes, and causal perturbation experiments may enable genomic AI to progress toward more comprehensive biological simulation.

In this sense, the true significance of AlphaGenome Atlas lies not only in generating predictions for approximately 9 billion variants, but also in providing a research foundation for systematically asking—and experimentally testing—what genomic variation actually causes in living biological systems.

This article is based on publicly available research papers, preprints, and technical materials reviewed as of October 29, 2026.

The AlphaGenome Atlas study includes results reported in a preprint that had not yet completed peer review.

The absence of a comprehensive independent benchmark should not be interpreted as a definitive claim that no such evaluation exists. Rather, within the materials reviewed for this article, we did not identify sufficient evidence establishing broad generalization to complete personal genomes or diverse previously unseen cellular conditions.

References

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

[2] Google DeepMind et al., “Advancing regulatory variant effect prediction with AlphaGenome,” Nature, 2026.
https://www.nature.com/articles/s41586-025-10014-0

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

[5] Sasse 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

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

[7] Switzerland Omics, “AlphaGenome Atlas — what would validate it?,” Technical Commentary, September 9, 2026.
https://switzerlandomics.ch/blog/2026-09-09-alphagenome-atlas/