GPN-Star: Implicit Assumptions, Research Challenges, Limitations, and Open Questions

Target Audience#

• Researchers and readers interested in the intersection of comparative genomics, evolutionary biology, machine learning, and deep learning.
• Those interested in GPN-Star, genomic language models, and a deeper understanding of their scientific significance, technical characteristics, implicit assumptions, and limitations.
• Researchers and practitioners interested in genetic variant effect prediction, interpretation of disease-causing variants, AI applications in drug discovery, and the scientific challenges involved in validating these technologies.

Summary#

• GPN-Star integrates comparative genomics with genomic language modeling. Its most fundamental assumption is that statistical patterns preserved throughout evolutionary history can serve as useful proxies for predicting the functional importance of genetic variants in present-day humans.
• Evolutionary constraint does not necessarily correspond directly to molecular function, clinical pathogenicity, or organismal fitness. Moreover, nucleotide probabilities reflect factors beyond natural selection.
• GPN-Star has several limitations acknowledged in its original publication, including its dependence on whole-genome alignments, difficulties in handling insertions, deletions, and structural variants, and limited improvements from longer sequence contexts.
• Although GPN-Star demonstrates strong performance across existing benchmarks, important questions remain regarding the discovery of previously unknown causal variants, the biological meaning of learned nucleotide dependencies, the quantitative relationship between prediction scores and selection coefficients, and its practical contributions to drug discovery.
• Future research should focus not only on improving evolutionary modeling but also on integrating experimental functional genomics, population genetics, and disease-related information to move from predicting evolutionary constraint toward understanding molecular mechanisms and disease causality.

1. Introduction and Main Findings#

GPN-Star is a genomic language model that integrates classical comparative genomics with Transformer-based modeling to predict variant effects based on evolutionary constraint.

The GPN-Star study, published in Nature on September 9, 2026, demonstrated that explicitly incorporating multiple sequence alignments and phylogenetic relationships can achieve strong predictive performance across several variant effect prediction tasks, outperforming existing evolutionary approaches and large-scale genomic language models in multiple evaluations [1].

The broader scientific positioning of GPN-Star, from classical methods such as phastCons and phyloP to GPN-MSA, AlphaGenome, and Evo 2, has been discussed in a separate article, "The Positioning of GPN-Star."

This article examines GPN-Star from a more critical perspective, focusing on its implicit assumptions, methodological challenges, limitations, and questions that remain insufficiently validated.

The Most Fundamental Assumption Behind GPN-Star#

The central assumption underlying GPN-Star can be expressed as follows:

Statistical patterns in genomic sequences preserved throughout evolutionary history can serve as useful proxies for predicting the functional importance of genetic variants in present-day humans.

This assumption is supported by a long history of research in comparative genomics.

However, it does not hold unconditionally.

Five issues are particularly important.

1. Evolutionary constraint does not necessarily correspond to present-day functional importance. Features preserved by natural selection in the past are not always the same features that are relevant to current diseases or therapeutic targets.
2. Multiple sequence alignments and species trees may not fully capture evolutionary information. Alignment quality, changes in evolutionary rates, and gene duplication can influence predictions.
3. Low nucleotide probability is not equivalent to deleteriousness. The probability distribution learned by the model reflects factors other than natural selection.
4. Strong performance on existing variant datasets does not prove the ability to discover previously unknown disease-causing variants. The effects of benchmark construction and evaluation methodology must be carefully disentangled.
5. There is a substantial gap between predicting evolutionary constraint and predicting molecular function, disease development, or therapeutic intervention outcomes. GPN-Star does not model all of these processes.

Limitations Acknowledged in the Original Publication#

The Discussion section of the original paper explicitly identifies several limitations.

The most important include:

• The requirement for whole-genome alignments at inference time.
• Difficulties in extending the approach to insertions, deletions, and structural variants.
• Limited improvements from incorporating longer sequence contexts.
• The need for further investigation into the model's contributions to actual biological discoveries.

These limitations do not contradict the strong predictive performance demonstrated by GPN-Star [1].

Throughout this article, we distinguish four categories:

• What the paper demonstrates: Claims supported by the presented data and evaluation results.
• What the model implicitly assumes: Biological and statistical assumptions underlying its methodology.
• Limitations explicitly acknowledged by the authors: Constraints described in the original paper, particularly in its Discussion section.
• What remains insufficiently validated: Research questions that cannot yet be resolved from the results presented in the original study.

Importantly, an issue being "not yet validated" does not mean that it has been disproven.

We distinguish between questions for which indirect evidence already exists and those requiring additional independent experiments.

2. Reassessing GPN-Star's Position in the History of Genomic Research#

GPN-Star belongs to a research trajectory extending from classical comparative genomics, represented by phastCons and phyloP, to modern genomic language models.

However, its position within this history deserves more precise consideration.

GPN-Star is not simply a modernization of classical comparative genomics using Transformers.

More precisely, it can be understood as an approach that replaces some of the statistical assumptions of classical molecular evolutionary models with a flexible conditional probability model learned by a neural network.

From Comparative Genomics to GPN-Star#

Research Direction	Central Question	Connection to GPN-Star
Molecular evolution, neutral evolution, and natural selection	What determines the frequencies of genetic variants during evolution?	Fundamental biological assumptions underlying prediction
GERP / phastCons / phyloP	Can evolutionary constraint be detected from substitution rates and sequence conservation?	Use of multispecies comparisons and phylogenetic information
CADD / LINSIGHT	Can different sources of information be integrated to predict variant deleteriousness?	Previous approaches connecting evolutionary constraint and pathogenicity
GPN / GPN-MSA	Can self-supervised learning capture evolutionary constraints in genomic sequences?	Masked nucleotide prediction
GPN-Star	Does integrating sequence context, multiple sequence alignments, and phylogenetic information improve constraint prediction?	Novel architecture and variant scoring

In classical methods such as phyloP and phastCons, assumptions concerning nucleotide substitution rates, phylogenetic relationships, and conservation states are relatively explicit [2].

GPN-Star, by contrast, explicitly incorporates phylogenetic information while allowing a neural network to learn complex dependencies among nucleotides [1].

This flexibility can improve predictive performance and representational capacity.

At the same time, however, it becomes more difficult to map what the model has learned onto interpretable evolutionary parameters used in conventional evolutionary models.

This issue is central to understanding both the strengths and limitations of GPN-Star.

Fundamental Differences from Classical Conservation Models#

The Nature paper demonstrates that GPN-Star captures nucleotide dependencies and sequence-context information that conventional position-wise conservation scores cannot adequately represent [1].

Therefore, it would be inaccurate to characterize GPN-Star as merely a recalculation of sequence conservation.

Its distinctive feature is the ability to learn, through a flexible probabilistic model, how well a nucleotide fits its surrounding sequence context and homologous sequences from other species, while incorporating phylogenetic relationships.

This approach may capture complex sequence patterns that are difficult to represent explicitly using conventional evolutionary models.

However, it also raises an important research question:

To what extent do these learned probabilistic patterns reflect natural selection, functional constraint, or other evolutionary processes?

3. Implicit Assumptions Underlying GPN-Star#

Several important biological and statistical assumptions underlie GPN-Star's predictions.

The following sections examine these assumptions in order of importance.

3.1 Assumption 1: Evolutionary Conservation Is an Indicator of Functional Importance#

This is the most fundamental assumption underlying GPN-Star.

GPN-Star uses the distribution of genomic sequences observed across evolutionary history to estimate how permissible a particular nucleotide substitution is from an evolutionary perspective.

The underlying biological principle is that mutations that disrupt important biological functions tend to be eliminated from populations by natural selection.

However, an important distinction must be made.

The information directly used by GPN-Star and the biological outcomes we ultimately want to understand can be organized as follows.

Information directly used and predicted by GPN-Star:

Genomic sequence distributions observed throughout evolution

↓

Sequence-context-dependent nucleotide permissibility

↓

Predicted evolutionary constraint score

Biological outcomes we ultimately want to infer:

Predicted evolutionary constraint score

↓

Effects on molecular function

↓

Causal effects on disease

↓

Effectiveness of therapeutic interventions

The latter sequence of inferences requires additional assumptions and validation.

For example, a variant may be subject to strong evolutionary constraint without necessarily contributing to a particular adult-onset disease.

Conversely, a variant that is not strongly constrained evolutionarily may still increase disease risk under specific environmental conditions, later in life, or in a particular genetic background.

Furthermore, natural selection generally operates through differences in survival and reproductive success within populations.

These are not necessarily equivalent to an individual's health status or medical importance.

A fundamental unresolved question is how much of the evolutionary constraint detected by GPN-Star actually reflects functional and clinical consequences in present-day humans.

The original study provides indirect evaluations using datasets such as ClinVar, OMIM, HGMD, GWAS, and ProteinGym, but it does not fully resolve this relationship [1].

3.2 Assumption 2: Nucleotide Probabilities Can Be Used to Infer Variant Deleteriousness#

GPN-Star uses masked language modeling to learn the probability distribution of nucleotides conditioned on surrounding sequences and sequences from other species.

Variant scores are then calculated from these learned probabilities.

The implicit assumption is:

A nucleotide substitution assigned a low probability by the model may be less evolutionarily permissible.

However, observed nucleotide frequencies are not determined solely by natural selection.

Other contributing factors include:

• Mutation rates and mutational spectra.
• Local sequence context, including GC content and CpG dinucleotides.
• Genetic drift and effective population size.
• Biases associated with gene conversion and recombination.
• Differences in nucleotide composition and mutational processes across evolutionary lineages.

A particularly important point is that low occurrence probability does not necessarily imply low fitness.

The GPN-Star paper explicitly recognizes this issue and introduces a calibration approach that estimates background mutational effects using regions considered approximately neutral.

The authors report that this correction reduces correlations with mutation rates and improves performance across many benchmarks [1].

Therefore, it would be inaccurate to claim that GPN-Star completely ignores the distinction between mutation rates and natural selection.

The more appropriate question is:

To what extent does the proposed calibration successfully remove the influences of mutation rates and nucleotide composition?

Remaining research questions include whether scores appropriately reflect selection pressures when mutational processes differ among evolutionary lineages or in genomic regions strongly affected by GC-biased gene conversion.

3.3 Assumption 3: Aligned Sequences Correspond Evolutionarily and Functionally#

GPN-Star depends heavily on whole-genome alignments.

Consequently, it is important that aligned nucleotides accurately represent evolutionarily corresponding genomic positions.

However, several processes occur during evolution that complicate this assumption:

• Gene duplication and deletion.
• Structural rearrangements, including translocations and inversions.
• Sequence insertions mediated by transposable elements.
• Expansion and contraction of repetitive sequences.
• Turnover of regulatory elements such as enhancers.

A particularly important distinction is that sequence homology and functional equivalence are not the same concept.

For example, a regulatory element may maintain a similar function across species even when its nucleotide sequence has changed substantially.

Conversely, a conserved sequence may interact with different transcription factors or regulate different target genes in different species.

GPN-Star can learn dependencies among aligned sequences, but it does not directly model the complete evolutionary rewiring of gene regulatory networks.

3.4 Assumption 4: Evolutionary Information Across Species Can Be Transferred to Human Variant Interpretation#

This assumption is fundamental to comparative genomics, but it is especially important for GPN-Star.

For example, even if a genomic sequence is conserved among primates, the reason for that conservation does not necessarily correspond directly to disease causality in humans.

Ecological conditions, life histories, developmental processes, and genetic backgrounds subject to natural selection differ across evolutionary timescales.

The original GPN-Star paper addresses this issue by comparing models trained on three evolutionary groups:

• Vertebrates.
• Mammals.
• Primates.

This represents an important contribution of GPN-Star [1].

However, several distinctions are necessary when interpreting these results.

• Strong predictive performance using deep evolutionary timescales supports the usefulness of information from ancient conserved functions.
• Strong predictive performance using shallower evolutionary timescales supports the usefulness of closely related species for predicting certain human traits.
• However, these results do not establish that the biological functions associated with those traits originated during the corresponding evolutionary period.

For example, if a primate-trained model performs particularly well in GWAS-related analyses, that does not necessarily mean the relevant disease-associated function first emerged in primates.

Differences in predictive performance may also reflect alignment quality, the number of comparable genomic positions, sequence divergence, and mutation rates.

3.5 Assumption 5: A Species Tree Is an Appropriate Representation of Evolutionary Relationships#

GPN-Star emphasizes its explicit use of a species tree.

However, a species tree does not represent every aspect of evolutionary history.

The evolutionary histories of individual genomic regions can differ from the overall species phylogeny.

Examples include:

• Incomplete lineage sorting.
• Introgression.

These phenomena are particularly relevant when many closely related species are included in an analysis.

Using a species tree is a reasonable modeling choice for GPN-Star.

However, the model does not explicitly reconstruct every difference between the species tree and the gene trees of individual genomic regions.

An important question is how much variant-score reliability changes when using a fixed species tree versus incorporating region-specific evolutionary histories.

3.6 Assumption 6: Variant Effects Can Be Inferred from the Available Local Sequence Context#

Because GPN-Star learns dependencies within genomic sequences, it can capture richer information than conservation models that treat individual nucleotide positions independently.

However, some genetic variants influence gene expression and disease through long-range regulatory interactions.

Examples include:

• Interactions between distal enhancers and promoters.
• Three-dimensional genome organization.
• Cell-state-dependent chromatin structure.

The original paper reports that increasing sequence-context length produced relatively modest improvements.

The authors suggest that one contributing factor may be the fragmentation of syntenic blocks in reference-genome-based alignments [1].

Therefore, the ability of GPN-Star to learn local nucleotide dependencies does not demonstrate that it adequately understands long-range gene regulation.

4. Challenges and Limitations Identified in the Original Paper#

This section distinguishes limitations explicitly acknowledged by the authors from broader limitations that follow from the experimental results.

Challenge	Status in the Original Paper	Practical Implications
Dependence on whole-genome alignments	Explicitly acknowledged	Difficult to apply to novel genomes or poorly aligned regions
Insertions, deletions, and structural variants	Explicitly acknowledged	Limited applicability beyond single-nucleotide variants
Long-range sequence context	Limited improvements observed	Constraints on modeling distal regulatory mechanisms
Evolutionary timescale	Optimal models differ by task	A single universal constraint score may be insufficient
Clinical pathogenicity	Explicitly distinguished from evolutionary constraint	Scores alone cannot determine clinical diagnoses
Biological discoveries	Identified as an area for future evaluation	Practical value for discovering unknown causal variants remains to be established

These limitations are discussed in the original Nature publication [1].

4.1 Structural Dependence on Whole-Genome Alignments#

GPN-Star's predictive capabilities depend not only on the model itself but also on the quality of the available comparative genomic information.

Consequently, particular caution is required for:

• Human-specific insertions.
• Repetitive sequences and segmental duplications.
• Genomic regions with ambiguous correspondence across species.
• Sequences absent from the reference genome.

This is not simply a computational cost or input-format issue.

The decision about which genomic regions can be aligned determines what evolutionary information becomes available to the model.

As a result, systematic differences in prediction reliability may arise between genomic regions that are easy to align and those that are difficult to align.

4.2 Strong SNV Prediction Does Not Imply Strong Prediction for All Variant Types#

The original study provides genome-wide prediction scores for possible single-nucleotide substitutions [1].

However, clinically important genetic variation includes more than single-nucleotide variants (SNVs).

Other important variant classes include:

• Insertions and deletions.
• Copy-number variants.
• Inversions.
• Translocations.
• Repeat expansions.

For these variants, not only individual nucleotide identities but also sequence lengths and genomic structures may change.

Therefore, GPN-Star's results should not be generalized into a claim that it can predict the consequences of all types of genetic variation.

4.3 GPN-Star Does Not Directly Predict Clinical Pathogenicity#

A GPN-Star score is not a probability of developing a disease.

Even for a variant receiving a strong constraint score, the score alone cannot determine:

• Which disease is associated with the variant.
• Its penetrance.
• The age at which disease may develop.
• How effects differ between heterozygous and homozygous states.
• How genetic background and environmental conditions modify its effects.

Consequently, clinical applications require integration with patient phenotypes, family segregation data, population frequencies, and functional experiments.

This is more than a practical caution.

It reflects a fundamental difference between what the model estimates and what clinicians ultimately need to know.

5. What Remains Insufficiently Validated?#

The original paper includes evaluations across multiple benchmarks, mutation-rate calibration, nucleotide-dependency analysis, and assessments involving model organisms.

Therefore, it would be inaccurate to describe these areas as entirely untested.

The central question is instead:

Which claims remain unconfirmed when we move beyond the scope of the evaluations already performed?

5.1 Overview of Important Open Questions#

The following assessment is based on the original publication and its implications for future research.

It should not be interpreted as an exhaustive review establishing that no other studies have addressed these questions.

Research Question	Existing Evidence	Remaining Question	Priority
Prediction of experimentally measured functional effects	Evaluated using datasets including ProteinGym	Can GPN-Star quantitatively predict functional effects of noncoding variants?	High
Discovery of previously unknown causal variants	Evaluated using existing disease and GWAS datasets	Can it discover genuinely novel disease-causing variants?	High
Benchmark biases	Evaluated across diverse datasets	Does its advantage remain after controlling for conservation, allele frequencies, and other confounders?	High
Relationship between constraint and selection	Allele-frequency relationships and mutation-rate calibration evaluated	Can selection coefficients be estimated quantitatively?	High
Robustness to alignment quality	Multiple evolutionary groups and alignments evaluated	How sensitive are predictions to gaps, misalignments, and species composition?	High
Optimal evolutionary timescale	Three evolutionary groups compared	Can the optimal evolutionary comparison range be estimated more precisely?	Medium–High
Causal meaning of learned nucleotide dependencies	Consistency with known functional elements demonstrated	Do these dependencies reflect genuine molecular interactions or epistasis?	High
Generalization across populations and ancestry backgrounds	Human population allele-frequency data evaluated	Does predictive accuracy remain consistent across populations?	Medium–High
Practical utility in drug discovery	Indirect evaluation using genetic indicators	Does it improve target discovery or drug development success rates?	High

The following sections examine the most important questions in greater detail.

5.2 What Does Strong Benchmark Performance Actually Establish?#

The original paper evaluates GPN-Star using a broad range of resources and tasks, including:

• ClinVar.
• OMIM.
• HGMD.
• GWAS fine-mapping.
• ProteinGym.
• Heritability analyses.
• Rare variant association testing.

This breadth of evaluation is an important strength of the study [1].

However, these benchmarks do not necessarily provide the same type of ground truth.

ClinVar

ClinVar reflects existing knowledge about clinical pathogenicity.

However, variant classifications may partly incorporate existing computational predictions, population frequencies, and other evidence.

GWAS Fine-Mapping

GWAS fine-mapping estimates the statistical probability that a variant is causal.

However, not every variant identified through fine-mapping has been experimentally validated as causal.

ProteinGym

Deep mutational scanning datasets in ProteinGym provide experimental evidence of variant effects.

However, the molecular functions measured for different proteins are not necessarily equivalent to disease effects at the organismal level.

Therefore, benchmark results must be interpreted in relation to the specific prediction target being evaluated.

In particular, evolutionary conservation may already contribute to existing pathogenicity classifications in some variant datasets.

This raises questions about the independence of the evaluation labels.

Such concerns are not unique to GPN-Star.

Rather, they are a broader methodological challenge for variant prediction models that rely on evolutionary information.

5.3 Does GPN-Star Estimate the Strength of Natural Selection?#

This is an important unresolved theoretical question.

GPN-Star scores have been shown to correlate with evolutionary constraint.

However, these scores do not necessarily provide direct estimates of:

• Selection coefficients, (s).
• Quantitative effects on fitness.
• Rates at which deleterious variants are eliminated from human populations.

The original study uses allele-frequency data from gnomAD and shows that variants predicted to be strongly deleterious tend to occur at lower frequencies.

This provides important evidence that the scores reflect natural selection [1].

However, allele-frequency distributions depend not only on selection but also on demographic history and mutation rates.

The relationship between GPN-Star scores and selection coefficients, as formally defined in population genetics, therefore warrants additional investigation.

In other words, GPN-Star may be an effective predictor of evolutionary constraint without necessarily being a model that directly estimates evolutionary process parameters.

5.4 Do Learned Nucleotide Dependencies Represent Genuine Functional Interactions?#

The original paper demonstrates that GPN-Star learns nucleotide dependencies corresponding to functional sequence features such as:

• Transcription factor binding sites.
• Coding regions.
• Splice sites.

This is an important advance over classical position-wise conservation scores [1].

However, there is a fundamental theoretical distinction:

Statistical dependence is not equivalent to a causal biological interaction.

For example, two nucleotide positions may covary for several possible reasons:

• Genuine molecular or functional interactions.
• Shared evolutionary history.
• Sequence composition biases.
• Correlations arising from belonging to the same functional region.
• Structural properties of multiple sequence alignments.

GPN-Star's nucleotide dependency analysis visualizes conditional dependencies learned by the model.

However, whether these dependencies accurately predict epistasis—non-additive functional effects arising from combinations of mutations—is a separate question.

Proposed Validation

Use massively parallel reporter assays (MPRAs) or deep mutational scanning to measure the functional effects of single and double mutations.

Then compare the experimentally observed interactions with the nucleotide dependencies predicted by GPN-Star.

This would help determine whether learned statistical dependencies correspond to genuine biological interactions.

5.5 Interpreting Strong Heritability Enrichment for Complex Traits#

The original paper reports that the primate-trained model achieves strong heritability enrichment across more than 100 complex traits [1].

This is an interesting and important result supporting the utility of GPN-Star.

However, heritability enrichment measures the degree to which genetic heritability is concentrated among variants prioritized by a particular score or annotation.

Strong heritability enrichment is not equivalent to accurately identifying individual causal variants.

Furthermore, enrichment can be influenced by correlations between prediction scores and existing functional annotations or particular genomic regions.

The GPN-Star paper investigates predictive information that is not readily explained by existing functional annotations.

This is an important strength of the study.

Nevertheless, further validation could include:

• Replication in independent populations and additional traits.
• Conditional comparisons with known functional annotations.
• Experimental validation of individual prioritized variants.

5.6 How Much Has GPN-Star's Utility for Drug Discovery Been Demonstrated?#

Variant interpretation using GPN-Star may eventually support drug discovery.

However, directly demonstrated outcomes must be distinguished from potential future applications.

For example, even if GPN-Star successfully prioritizes disease-associated variants or genes, this does not necessarily mean that inhibiting or activating those genes will produce a therapeutic benefit.

Using such information for therapeutic target development requires additional evidence about:

• Whether a variant increases or decreases gene function.
• Whether manipulating the target improves disease phenotypes.
• Which tissues or cell types should be targeted.
• Potential safety issues and adverse effects.
• Whether the target is pharmacologically tractable.

Therefore, the possibility that GPN-Star can serve as a useful variant-prioritization tool for drug discovery is distinct from demonstrating that it improves actual drug development outcomes.

At present, the former is supported by indirect evidence from genetic analyses, while the latter requires more direct validation.

6. Five Recommended Directions for Further Validation#

Based on the preceding discussion, five research directions appear particularly valuable.

6.1 Study 1: Comparison with Experimentally Measured Noncoding Variant Effects#

Objective

Evaluate the gap between evolutionary constraint and molecular function.

Proposed Approach

Compare GPN-Star scores with experimentally measured functional effects of variants assessed using:

• Massively parallel reporter assays (MPRAs).
• CRISPR interference (CRISPRi).
• Base editing.

Particular attention should be given to:

• The magnitude of gene expression changes.
• The direction of expression changes.
• Cell-type dependence.

This would allow researchers to quantitatively evaluate whether variants predicted to be strongly constrained actually produce substantial functional effects in experimental systems.

6.2 Study 2: Robustness to Species Trees, Species Composition, and Alignment Quality#

Objective

Determine whether model performance reflects genuine evolutionary signals or depends on choices made during input-data construction.

Proposed Approach

For the same set of human variants, systematically modify:

• The number of species included.
• The species selected for comparison.
• The quality of the multiple sequence alignment.

Then measure how much the resulting GPN-Star scores change under each condition.

This would help assess the robustness of predictions and their dependence on the structure of the evolutionary input data.

6.3 Study 3: Relationship Between Nucleotide Dependencies and Epistasis#

Objective

Move from learning statistical nucleotide dependencies toward understanding functional interactions.

Proposed Approach

Use experimental measurements of double and multiple mutations to determine whether sequence-context dependencies learned by GPN-Star correspond to actual non-additive functional effects.

This would test whether nucleotide dependency analysis can serve as a useful approach for understanding molecular interactions.

6.4 Study 4: Prospective Validation Using Previously Uncharacterized Variants#

Objective

Measure the model's ability to support new biological discoveries rather than merely classify variants in existing benchmark datasets.

Proposed Approach

Use GPN-Star to prioritize variants not included in existing benchmarks.

After fixing the predictions in advance, perform experimental validation.

Compare the results with those obtained using:

• Classical conservation-based models.
• Sequence-to-function models.

Such prospective evaluations would provide a more direct assessment of GPN-Star's ability to identify previously unknown functional variants or disease-causing mutations.

6.5 Study 5: Integrating GPN-Star with AlphaGenome and Other Models#

Objective

Quantitatively assess whether evolutionary and molecular functional information are genuinely complementary.

Proposed Approach

Develop an integrated prediction framework combining:

• Evolutionary constraint scores from GPN-Star.
• Functional predictions from AlphaGenome and related models.
• Population allele-frequency information.

Compare the performance of the integrated framework with each individual model.

Particular attention should be given to generalization across:

• Previously unseen genes.
• Different chromosomes.
• New disease contexts.

This would determine whether combining evolutionary and molecular functional information produces predictive capabilities beyond those available from individual models.

6.6 Research Priorities#

If the primary objective is scientific novelty and a deeper understanding of the model, two directions are especially interesting:

• Study 2: Robustness to Species Trees, Species Composition, and Alignment Quality
• Study 3: Relationship Between Nucleotide Dependencies and Epistasis

These address fundamental questions about what GPN-Star actually learns and how its learned representations should be interpreted biologically.

If the primary objective is practical utility in medicine or drug discovery, two other directions may deserve priority:

• Study 1: Comparison with Experimentally Measured Noncoding Variant Effects
• Study 4: Prospective Validation Using Previously Uncharacterized Variants

These studies would more directly evaluate whether GPN-Star predictions can contribute to understanding molecular function and discovering disease-causing variants.

7. Overall Assessment: Scientific Significance and Fundamental Limitations of GPN-Star#

The scientific significance of GPN-Star can be summarized in three major contributions.

First, compared with conventional conservation models, it captures local sequence context and nucleotide dependencies.

Second, by explicitly incorporating phylogenetic relationships into its architecture, it improves predictive performance through an approach that differs from simply scaling up genomic language models.

Third, it demonstrates the importance of evolutionary timescale for variant interpretation across a broad range of prediction tasks [1].

The Most Important Unresolved Question#

The most fundamental unresolved question can be expressed as follows:

To what extent do the patterns learned by GPN-Star represent functional constraints imposed by natural selection, and to what extent do they reflect statistical regularities arising from evolutionary history, mutational processes, or alignment structure?

Furthermore, the degree to which these learned patterns explain actual molecular functions and disease causality remains open to further investigation.

In this sense, the next stage of GPN-Star research can be divided into two broad directions.

7.1 Direction A: Understanding Evolution More Accurately#

This direction involves more rigorous investigation of relationships among:

• Phylogenetic trees.
• Multiple sequence alignments.
• Mutation rates.
• Natural selection.
• Nucleotide dependencies.

For example, researchers could investigate how GPN-Star scores correspond to actual selection coefficients and how predictions change depending on species-tree structures and alignment configurations.

Another important question is whether learned nucleotide dependencies reflect genuine functional interactions.

This direction is not only about improving prediction performance.

It also addresses a fundamental scientific question:

To what extent can neural networks help us understand evolutionary processes themselves?

7.2 Direction B: Understanding Present-Day Biological Functions from Evolutionary Constraint#

This direction aims to combine evolutionary information with experimental functional genomics and disease-related data to predict causal variant effects and biological mechanisms.

Evolutionary constraint estimated by GPN-Star indicates the potential biological importance of a variant.

However, constraint alone does not adequately explain:

• In which cells a variant acts.
• Which genes are affected.
• What molecular changes occur.
• How these changes contribute to disease.

Sequence-to-function models such as AlphaGenome approach these questions from a different perspective.

Therefore, integrating evolutionary constraint with molecular functional predictions may enable richer variant interpretation.

However, whether such integration actually improves predictive performance must be established through independent evaluation.

Direction B is particularly interesting because it represents a point of convergence between the research philosophies underlying GPN-Star and AlphaGenome.

It may therefore become an important direction for future genomic AI research.

7.3 Conclusion: From Evolutionary Constraint Prediction to Understanding Biological Function and Disease Causality#

GPN-Star demonstrates that effectively leveraging evolutionary information can produce variant constraint scores that outperform previous approaches in multiple prediction tasks.

It also provides evidence that explicitly incorporating biologically meaningful structures into machine learning architectures can be effective for specific genomic prediction problems.

However, this result should not be generalized into a claim that biology-informed architectures always outperform large-scale general-purpose models across all genomic AI tasks.

More fundamentally, GPN-Star does not provide a complete answer to the question:

How accurately can we infer present-day biological functions and disease causality from evolutionary information?

Future research should therefore focus not only on improving predictive performance but also on validating the evolutionary and biological meaning of model predictions.

In particular, it will be important to connect evolutionary constraint with experimentally measured molecular functions, population genetics, and disease-associated information.

The ultimate goal is to determine how much these models can contribute to discovering previously unknown functional variants and disease-causing mutations.

The scientific value of GPN-Star extends beyond improvements in variant constraint prediction accuracy.

It also provides a modern machine learning framework for revisiting one of the fundamental questions in genomics: how much can the evolutionary history of DNA sequences tell us about present-day biological function and the causal mechanisms of disease?

References#

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

[2] Siepel, A., et al., “Evolutionarily conserved elements in vertebrate, insect, worm, and yeast genomes,” Genome Research, 15, 1034–1050, 2005.
https://genome.cshlp.org/content/15/8/1034

[3] Huang, Y.-F., Gulko, B., and Siepel, A., “Fast, scalable prediction of deleterious noncoding variants from functional and population genomic data,” Nature Genetics, 49, 618–624, 2017.
https://www.nature.com/articles/ng.3810

[4] “Predicting functional constraints across evolutionary timescales with phylogeny-informed genomic language models,” bioRxiv preprint, 2025. Preprint version of the GPN-Star study.
https://pmc.ncbi.nlm.nih.gov/articles/PMC12458161/

[5] Song Lab, “GPN-Star,” Hugging Face model and data collection.
https://huggingface.co/collections/songlab/gpn-star-68c0c055acc2ee51d5c4f129