Target Audience

  • Researchers and professionals interested in the intersection of comparative genomics, evolutionary biology, machine learning, and deep learning.
  • Readers seeking to understand how genomic foundation models such as GPN-Star, AlphaGenome, and Evo 2 fit into the history of genomic AI and how their underlying approaches differ.
  • Those interested in genetic variant interpretation, disease-associated variant discovery, and AI applications in drug development, including what current models have demonstrated and what remains scientifically unresolved.

Executive Summary

  • GPN-Star integrates classical comparative genomics with Transformer-based genomic language modeling. It learns evolutionary constraints by explicitly incorporating multiple sequence alignments and phylogenetic trees.
  • Despite having approximately 200 million parameters, GPN-Star achieves strong performance across a broad range of coding and noncoding variant prediction benchmarks. It also demonstrates that different evolutionary timescales provide different types of information about functional constraint.
  • However, its approach rests on a fundamental assumption: patterns preserved by natural selection throughout evolutionary history can help predict the functional importance of variants in present-day humans. Evolutionary constraint is not equivalent to molecular function, clinical pathogenicity, or causal effects on disease.
  • Whereas AlphaGenome primarily predicts molecular function and Evo 2 focuses on general-purpose sequence modeling and generation, GPN-Star specializes in evolutionary constraint-based variant interpretation. These approaches may be complementary, although the benefits of integrating them require further validation.
  • GPN-Star's significance extends beyond improved predictive accuracy. It demonstrates the value of incorporating biological knowledge directly into model architectures, while also highlighting unresolved challenges in identifying novel causal variants, predicting experimentally measurable functional effects, and translating evolutionary information into biological and clinical understanding.

1. Where Does GPN-Star Fit in the History of Genomic AI?

On September 9, 2026, the paper "Predicting genome-wide functional constraints with GPN-Star" was published in Nature [1].

GPN-Star (Genomic Pretrained Network with Species Tree and Alignment Representations) is a model that combines DNA sequences, multiple sequence alignments, and phylogenetic trees to predict evolutionary constraints associated with genomic variants.

The study occupies an interesting position in the development of genomic artificial intelligence.

During the 2020s, research increasingly focused on treating DNA as a language and learning its statistical regularities using large neural networks.

One central idea was that sufficiently large models trained on massive amounts of genomic sequence data might automatically acquire the information needed to understand gene function and the effects of genetic variation.

However, genomes have a fundamental characteristic that distinguishes them from natural language.

The genomes of different species are products of evolutionary histories that trace back to common ancestors.

For example, when human and mouse DNA sequences resemble one another, that similarity is not merely a statistical coincidence.

It reflects sequences inherited from common ancestors and subsequently modified through mutation, natural selection, genetic drift, and other evolutionary processes.

Rather than relying entirely on a model to infer these relationships implicitly from sequence data, GPN-Star explicitly provides evolutionary structure through phylogenetic trees and multiple sequence alignments.

Its central design philosophy can therefore be expressed as follows:

By incorporating the evolutionary history that produced genomic sequences, rather than learning only their statistical regularities, we may be able to predict the consequences of genetic variation more accurately.

This is not simply a return to classical comparative genomics.

Traditional evolutionary models such as phastCons and phyloP describe nucleotide substitution processes and evolutionary conservation through relatively explicit mathematical assumptions.

GPN-Star, by contrast, uses evolutionary relationships while allowing a neural network to learn complex sequence-context dependencies.

In this sense, GPN-Star can be understood as an attempt to replace certain assumptions of classical evolutionary models with more flexible, learnable representations.

Understanding this contribution requires examining three major research traditions: comparative genomics, functional genomics, and genomic language modeling.

2. The Evolution of Genomic AI: From Conservation Models to Foundation Models

First Generation: Comparative Genomics — Conserved Regions Are Likely to Be Important

During the 2000s, comparative genomics advanced through the systematic comparison of genomic sequences across species.

The goal was to identify genomic regions that might perform important biological functions.

Representative methods include phastCons and phyloP.

phastCons uses whole-genome multiple sequence alignments and phylogenetic trees to identify regions evolving at relatively low nucleotide substitution rates.

More specifically, it employs a phylogenetic hidden Markov model to represent nucleotide substitution processes and the tendency of conserved genomic regions to occur in contiguous segments [2].

The basic framework can be summarized as:

Species alignment → Evolutionary model → Conservation / constraint score

Nucleotides or genomic regions conserved across distantly related species may have experienced evolutionary constraints that limited their ability to change.

One possible explanation is purifying selection: mutations that negatively affected survival or reproductive success were more likely to be eliminated from populations.

This principle remains fundamental to modern variant interpretation.

However, an important distinction must be made.

Evolutionary conservation is valuable evidence of functional importance, but it does not directly establish the precise molecular function of a sequence or its causal relationship with disease in present-day humans.

This distinction is equally important when evaluating GPN-Star.

Second Generation: Integrating Functional Annotations and Evolutionary Information Through Machine Learning

During the 2010s, researchers began integrating evolutionary conservation with other genomic features to predict variant effects.

Two important examples are CADD and LINSIGHT.

CADD combines evolutionary conservation, gene annotations, functional genomic information, and other features into a unified score for estimating the relative deleteriousness of genetic variants.

The original 2014 study used a Support Vector Machine (SVM) [3].

LINSIGHT represents another important development.

It combines three components:

  • Functional genomics
  • Machine learning
  • Molecular evolution

LINSIGHT connects features derived from functional genomic data with models of molecular evolution to estimate the deleteriousness of noncoding variants [4].

Its underlying research philosophy is closely related to that of GPN-Star.

However, CADD and LINSIGHT rely substantially on explicitly designed features and evolutionary modeling assumptions.

GPN-Star differs by allowing a neural network to learn complex relationships between sequence context and evolutionary information more directly.

Third Generation: Predicting Molecular Function Directly From DNA

The introduction of DeepSEA in 2015 marked a major turning point in the application of deep learning to genomics.

DeepSEA used a Convolutional Neural Network (CNN) to predict functional genomic features directly from DNA sequences [5].

Its prediction targets included:

  • Transcription factor binding
  • DNase hypersensitivity
  • Histone modifications

This established an influential direction in genomic AI known as sequence-to-function modeling.

The basic framework is:

DNA sequence → Neural network → Molecular phenotype

This research trajectory includes:

DeepSEA / Basset
↓
ExPecto
↓
Basenji
↓
Enformer
↓
Borzoi
↓
AlphaGenome

Enformer incorporated Transformer architectures to model regulatory interactions over distances of approximately 100 kb, advancing the ability to learn long-range enhancer–promoter relationships from DNA sequences [6].

Borzoi subsequently extended the sequence context beyond 500 kb.

By directly predicting RNA-seq coverage, it provided a more unified approach to modeling transcription, splicing, polyadenylation, and related processes [7].

AlphaGenome, whose Nature paper was published in January 2026, further expanded this approach by predicting diverse molecular functions from DNA sequences spanning 1 Mb [8].

Its prediction targets include:

  • Gene expression
  • Chromatin accessibility
  • Transcription factor binding
  • Histone modifications
  • Splicing
  • 3D chromatin contacts

AlphaGenome predicts 5,930 functional genomic tracks in humans.

It also achieved performance comparable to or better than the strongest evaluated alternatives on 25 of 26 variant-effect benchmarks [8].

The defining characteristic of this research direction is its attempt to predict how changes in DNA sequence affect measurable molecular functions, such as gene expression or chromatin state.

Fourth Generation: Self-Supervised Learning by Treating DNA as a Language

Inspired by the success of BERT and GPT in natural language processing, researchers began applying language modeling approaches to DNA.

Representative models include:

  • DNABERT
  • Nucleotide Transformer
  • HyenaDNA
  • Caduceus
  • Evo
  • Evo 2

Nucleotide Transformer introduced models containing up to 2.5 billion parameters.

Using self-supervised learning on human genomes and genomic sequences from 850 species, it developed representations that could be fine-tuned for various genomic analysis tasks [9].

Evo extended genomic sequence modeling beyond feature extraction.

It explored learning sequence regularities associated with DNA, RNA, and proteins, as well as generating sequences at genomic scales [10].

Evo 2 further expanded this direction through large-scale sequence modeling across all domains of life.

Its Nature paper was published in March 2026 [14].

Evo 2 addresses zero-shot prediction of human coding and noncoding variant effects, genomic sequence generation, and the analysis of biological information encoded in learned representations.

These developments raised an important scientific question:

If we train sufficiently large models on genomic sequences, will they automatically learn evolutionary constraint?

GPN-Star provides an important perspective on this question.

Large single-sequence genomic language models require substantial computational resources, yet they do not always outperform classical evolutionary models or alignment-based approaches on evolutionary constraint prediction tasks [1].

In other words, increasing model scale does not necessarily produce better predictions of evolutionary constraint.

Fifth Generation: Genomic Language Models That Explicitly Incorporate Evolutionary Information

The GPN family emerged against this background.

Early GPN models attempted to learn evolutionary constraints from genomic sequence distributions.

Subsequent approaches progressively incorporated multiple sequence alignments and phylogenetic information more explicitly.

GPN-Star represents an important development in this progression.

However, the history of genomic AI should not be interpreted simply as a sequence in which each new generation replaces its predecessor.

Comparative genomics, sequence-to-function modeling, and general-purpose genomic language modeling have developed in parallel because they address different scientific questions.

GPN-Star is best understood as a model specializing in evolutionary constraint that connects classical evolutionary information with modern deep learning.

3. The Technical Evolution From GPN to GPN-Star

Understanding GPN-Star requires examining the development of the GPN family itself.

The research trajectory can be summarized as:

GPN → GPN-MSA → PhyloGPN / GPN-Star

PhyloGPN and GPN-Star are best viewed as approaches that address related challenges after GPN-MSA from somewhat different directions.

GPN: Learning From Unaligned Genomes

The original GPN, introduced in 2023, trained a genomic language model using unaligned genomes from Arabidopsis and related plant species.

It demonstrated that self-supervised learning could extract information useful for variant effect prediction from genomic sequences [11].

At this stage, the emphasis was on learning statistical patterns in genomic sequences rather than explicitly modeling evolutionary relationships between aligned positions.

GPN-MSA: Directly Incorporating Multiple Sequence Alignments

GPN-MSA introduced a major conceptual change.

Instead of relying exclusively on implicit learning of evolutionary information, it directly incorporated a 100-way vertebrate whole-genome alignment.

This approach achieved strong variant effect prediction performance across both coding and noncoding regions [12].

It established the value of providing evolutionary comparison data directly to a neural network.

However, GPN-MSA also had limitations, including dependence on particular alignment structures and the need for manual adjustments to species composition.

PhyloGPN: Separating Training-Time Evolutionary Information From Inference-Time Inputs

PhyloGPN explored a different approach to using evolutionary information.

It proposed leveraging phylogeny and alignment information during training while allowing predictions from a single DNA sequence at inference time [13].

The objective was to benefit from evolutionary supervision while simplifying inference requirements.

GPN-Star: Explicitly Modeling the Species Tree

GPN-Star incorporates both whole-genome alignments and species trees.

It uses the evolutionary relationships between species as part of the model architecture rather than treating sequences from different species as unrelated inputs [1].

Its main improvements over GPN-MSA include:

  1. Performing masked nucleotide prediction across multiple species in the alignment, rather than focusing only on humans.
  2. Explicitly incorporating species-tree structure into attention calculations.
  3. Handling alignments with different species compositions more flexibly.

GPN-Star also introduces calibration to address mutation-rate variation.

This is particularly important because observed sequence distributions reflect not only natural selection but also variation in the underlying mutation process.

4. The Core of GPN-Star: Integrating Phylogenetic Trees With Transformers

The technical core of GPN-Star is its incorporation of phylogenetic relationships into attention mechanisms while processing whole-genome alignments.

Its approximate processing pipeline can be represented as follows:

Whole-genome alignment
↓
Species-tree-aware pooling
↓
Phylogeny-aware cross-attention
↓
Nucleotide probabilities
↓
Calibrated variant constraint score

Species Are Not Independent Inputs

When comparing genomes across species, the sequences are not statistically independent.

For example, humans and chimpanzees share a more recent common ancestor than humans and chickens.

Therefore, observing the same nucleotide in two closely related species does not carry the same evolutionary information as observing that nucleotide across distantly related species.

GPN-Star uses attention pooling to aggregate sequences from related species into clade-level representations.

It then uses phylogeny-aware cross-attention, incorporating evolutionary distances to adjust how information from different species contributes to predictions [1].

This enables the model to consider not only which nucleotides are present in an alignment but also the evolutionary relationships among the species carrying them.

What Does Masked Nucleotide Prediction Learn?

GPN-Star uses masked nucleotide prediction to learn conditional probability distributions over nucleotides.

When a nucleotide is masked, the model predicts its identity using surrounding sequence context and evolutionary information from other species.

These predicted probabilities are then used to evaluate alternative alleles at a genomic position.

Intuitively, a substitution that is unlikely under the learned evolutionary sequence context may be less evolutionarily tolerated.

However, this reasoning rests on an important assumption.

A nucleotide assigned a low probability by the model is not necessarily biologically deleterious.

Observed nucleotide distributions are influenced by evolutionary processes other than natural selection.

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

The Role of Mutation-Rate Calibration

GPN-Star learns from DNA sequences observed in nature.

However, the observed distribution of nucleotides reflects at least two distinct processes:

  • The mutation process, which determines how frequently different substitutions arise.
  • The selection process, which influences whether those mutations persist or are removed.

For example, a substitution may be rarely observed because it is deleterious.

Alternatively, it may be rarely observed simply because that type of mutation occurs infrequently.

GPN-Star addresses this issue by estimating background mutation effects from regions considered highly likely to be neutral and using those estimates to calibrate variant scores [1].

This represents an important methodological improvement over approaches that rely more directly on raw sequence probabilities.

Nevertheless, calibration does not establish that mutation processes and natural selection have been perfectly disentangled.

Further investigation is warranted into the effects of local sequence context, CpG mutation patterns, lineage-specific mutation rates, and GC-biased gene conversion.

GPN-Star as a Biology-Informed Foundation Model

GPN-Star is closely aligned with the concept of a biology-informed foundation model.

Rather than requiring the model to discover every relevant regularity from data alone, it incorporates known biological structure as an inductive bias.

The original study reports that a model with approximately 200 million parameters, trained for several days on eight NVIDIA A100 GPUs, achieved strong performance across evolutionary constraint prediction tasks [1].

This suggests that explicitly incorporating evolutionary structure can provide a computationally efficient alternative to scaling up single-sequence genomic language models for certain tasks.

However, this does not establish that biology-informed architectures outperform large general-purpose models across all genomic applications.

What GPN-Star demonstrates is that explicit biological structure can be particularly effective for the specific task of evolutionary constraint prediction.

5. What Has GPN-Star Actually Demonstrated?

Evaluating GPN-Star requires distinguishing its empirically demonstrated contributions from broader expectations about its future applications.

The original study includes extensive evaluations involving coding and noncoding variants, complex-trait heritability, rare variant association testing, and model organisms [1].

Coding Variant Prediction

GPN-Star was evaluated using pathogenic and benign missense variants from ClinVar.

Comparison methods included phyloP, phastCons, CADD, Nucleotide Transformer, Evo 2, and GPN-MSA.

GPN-Star trained on vertebrate alignments, referred to as GPN-Star (V), demonstrated strong predictive performance.

It also performed well across 31 human deep mutational scanning datasets from ProteinGym.

Among genome-wide models, GPN-Star achieved particularly strong results.

However, specialized protein-focused predictors performed slightly better on some ProteinGym benchmarks.

These findings support the value of a genome-wide prediction framework but do not imply that GPN-Star is universally the best model for missense variant prediction.

Noncoding Variant Prediction

One of GPN-Star's most interesting contributions concerns noncoding variants.

The original study compared GPN-Star with sequence-to-function models such as Enformer, Borzoi, and AlphaGenome using pathogenic noncoding variants from OMIM and HGMD.

GPN-Star outperformed these models on the evaluated benchmarks.

The advantage was particularly pronounced for variants located in distal enhancers [1].

This finding is important because distal enhancers may be located far from the genes they regulate, making their functional consequences difficult to predict.

GPN-Star demonstrates that evolutionary sequence constraints can provide useful information for prioritizing variants in such regions, even without directly predicting their molecular effects.

However, identifying variants associated with pathogenicity is different from explaining which genes they affect or precisely how they alter molecular function.

GWAS Fine-Mapping and Complex-Trait Heritability

GPN-Star was also evaluated for variant prioritization in GWAS fine-mapping.

Across more than 100 complex traits, variants prioritized by GPN-Star exhibited strong heritability enrichment.

In particular, the primate-based model, GPN-Star (P), demonstrated strong enrichment for complex traits [1].

This suggests that evolutionary constraints measured over relatively shallow timescales may contain information particularly relevant to present-day human complex traits.

However, heritability enrichment should be interpreted carefully.

Strong enrichment does not establish that individual causal variants have been experimentally identified.

Likewise, fine-mapping probabilities are statistical estimates of causal involvement, not experimental confirmation of causality for every variant.

These results provide valuable evidence of predictive utility, but they do not directly demonstrate that GPN-Star can reliably discover previously unknown disease-causing variants.

Rare Variant Association Testing

The original study also investigated applications to rare variant association testing (RVAT).

Using GPN-Star's evolutionary constraint scores improved the ability to detect associations between rare variants and phenotypes [1].

This suggests that GPN-Star may be valuable not only for scoring individual variants but also for improving broader statistical genetic analyses.

However, further research is needed to determine how consistently these improvements translate into novel biological discoveries, particularly for rare noncoding variants.

Applications to Model Organisms

GPN-Star was also evaluated beyond humans.

The original study trained and evaluated models for five organisms:

  • Mus musculus
  • Gallus gallus
  • Drosophila melanogaster
  • Caenorhabditis elegans
  • Arabidopsis thaliana

Whole-genome alignments containing between 18 and 135 species were used for these organisms [1].

For many model organisms, large collections of clinically classified pathogenic variants are unavailable.

Consequently, evaluations included measures such as the enrichment of rare variants within populations.

For mice, fruit flies, and nematodes, additional evaluations used pathogenic or lethal variant datasets.

GPN-Star outperformed traditional conservation-based methods in these evaluations.

These findings support the broader applicability of its architecture beyond human genomics.

6. Different Evolutionary Timescales Reveal Different Constraints

Another important finding from GPN-Star is that using a broader evolutionary comparison does not necessarily produce better predictions.

The original study compared three major evolutionary ranges:

  • Vertebrates
  • Mammals
  • Primates

Deep Versus Shallow Evolutionary Timescales

Deep evolutionary information can be useful for identifying strong constraints in coding regions and missense variants.

By contrast, relatively shallow evolutionary comparisons involving mammals or primates may provide particularly useful information for noncoding regulatory variation and complex traits.

The GPN-Star study reports strong heritability enrichment across more than 100 complex traits using the primate-based model [1].

These results challenge the idea that evolutionary constraint can always be summarized by a single universal conservation score.

For example, the following interpretations are plausible:

  • Regions conserved for approximately 500 million years may be associated with fundamental biological functions.
  • Regions conserved over approximately 30–100 million years may contain regulatory elements relevant to mammalian biology.
  • Regions showing strong constraint among primates may be associated with regulatory phenotypes shared by species closely related to humans.

These examples illustrate possible interpretations of evolutionary timescales.

They do not establish that the corresponding biological functions originated during those specific periods.

Evolutionary Timescale Is Not the Same as the Time of Functional Innovation

An important distinction is necessary here.

If a primate-based model performs particularly well for a complex trait, this does not prove that the underlying biological function first evolved in primates.

Differences in model performance may also reflect:

  • Evolutionary distances among species
  • Numbers of species included
  • Multiple sequence alignment quality
  • Differences in alignable genomic positions
  • Nucleotide substitution rates and mutation processes

Therefore, the key conclusion is that different evolutionary timescales provide different amounts and types of predictive information depending on the target task.

This is distinct from determining when particular functions first appeared during evolution.

The original study also examined relationships between the effective polygenicity of complex traits and the optimal evolutionary timescale.

For example, mammalian models can be particularly informative for LDL cholesterol, whereas primate models tend to perform well for highly polygenic traits such as schizophrenia [1].

Understanding these relationships may help connect evolutionary constraint with the genetic architecture of complex traits.

7. The Implicit Assumptions Underlying GPN-Star

To understand GPN-Star's scientific significance, it is not enough to examine its predictive performance.

We must also consider the assumptions required for its predictions to be biologically meaningful.

Some assumptions are inherited from classical comparative genomics.

Others arise from using neural networks to learn statistical patterns in evolutionary sequence data.

Assumption 1: Evolutionary Constraint Is a Proxy for Functional Importance Today

The most fundamental assumption underlying GPN-Star is that patterns preserved throughout evolutionary history provide useful evidence about the functional importance of genetic variants in present-day humans.

This assumption has a theoretical foundation in natural selection.

Mutations that negatively affect survival or reproductive success are more likely to be eliminated through purifying selection.

However, the targets of natural selection are not necessarily identical to the phenotypes that are medically important today.

For example, a strongly constrained variant may not cause a particular adult-onset disease.

Conversely, a variant showing weak evolutionary constraint may influence disease risk under specific environmental conditions, at older ages, or in particular genetic backgrounds.

Natural selection generally acts on fitness-related outcomes, including survival and reproductive success.

Clinical medicine, by contrast, focuses on health, disease onset, disease severity, and treatment.

These concepts overlap, but they are not equivalent.

Predicting evolutionary constraint does not directly establish the causal mechanisms of disease in present-day humans.

Assumption 2: Nucleotide Probabilities Contain Information About Natural Selection

GPN-Star learns conditional nucleotide probability distributions.

However, observed sequence distributions reflect not only natural selection but also mutation rates, genetic drift, recombination, gene conversion, and other evolutionary processes.

A low predicted nucleotide probability does not necessarily imply low biological fitness.

GPN-Star addresses this issue through mutation-rate calibration.

Therefore, it would be inaccurate to claim that GPN-Star completely ignores mutation-rate effects.

The remaining question is how thoroughly these effects have been corrected.

Further investigation may be needed into GC content, CpG context, lineage-specific mutation rates, and GC-biased gene conversion.

Assumption 3: Multiple Sequence Alignments Correctly Represent Evolutionary Correspondence

GPN-Star depends heavily on whole-genome alignments.

This requires aligned nucleotides from different species to represent appropriate evolutionary correspondences.

However, genome evolution involves processes such as:

  • Gene duplication and deletion
  • Translocations and inversions
  • Transposable element insertions
  • Expansion and contraction of repetitive sequences
  • Regulatory element turnover

Noncoding regulatory elements are particularly challenging.

Similar functions can sometimes be maintained despite substantial changes in nucleotide sequence.

Conversely, conserved sequences may acquire different regulatory relationships, interacting with different transcription factors or regulating different target genes.

Therefore, sequence homology and functional equivalence are not necessarily the same.

GPN-Star learns from evolutionarily aligned sequences, but it does not explicitly model the complete evolutionary rewiring of gene regulatory networks.

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

GPN-Star uses genomic information from other species to estimate the importance of human variants.

This is a foundational principle of comparative genomics.

However, species differ in their environments, life histories, developmental processes, and genetic backgrounds.

A sequence that experienced strong selection in one lineage may not necessarily have the same functional significance in present-day humans.

GPN-Star partially addresses this issue by comparing vertebrate, mammalian, and primate evolutionary timescales.

Nevertheless, evaluating three evolutionary ranges cannot fully establish the extent to which evolutionary constraints transfer across species.

Assumption 5: The Species Tree Adequately Represents Evolutionary Relationships

GPN-Star explicitly incorporates species trees.

However, a species tree does not perfectly describe the evolutionary history of every genomic region.

Processes such as incomplete lineage sorting and introgression can produce local gene trees that differ from the overall species tree.

Consequently, species-tree-based information aggregation may not be optimal for every genomic region.

An interesting future experiment would compare predictions obtained from a fixed species tree with those incorporating local genealogical relationships.

Assumption 6: Local Sequence Context Contains Sufficient Information About Important Functional Dependencies

GPN-Star can learn relationships between neighboring nucleotides rather than evaluating every position independently.

This is an important improvement over classical site-independent conservation models.

However, the functional effects of genomic variants can depend on long-range regulatory relationships.

Examples include distal enhancer–promoter interactions, three-dimensional genome organization, and cell-state-dependent chromatin structures.

The original study reported relatively limited performance improvements when expanding sequence context length [1].

The authors suggested that fragmentation in reference-based whole-genome alignments could contribute to this limitation.

Therefore, learning local nucleotide dependencies does not establish that GPN-Star fully captures long-range gene regulation.

8. What Do GPN-Star's Learned Nucleotide Dependencies Actually Mean?

One distinctive feature of GPN-Star is its ability to examine learned relationships between nucleotides through nucleotide dependency analysis.

Traditional site-independent models, such as phyloP, primarily evaluate evolutionary constraint at individual genomic positions.

GPN-Star can instead examine how changing a nucleotide at one position affects the model's predicted probability distribution at another position.

The original study identified dependencies corresponding to biological elements such as [1]:

  • Transcription factor binding sites
  • Coding regions
  • Splice donor and acceptor sites
  • Primate-associated regulatory constraints

The researchers also analyzed variants that appeared relatively neutral under classical conservation scores but were predicted to be deleterious by GPN-Star.

These analyses suggest that the model captures information not fully represented by independent position-wise conservation scores.

However, an important question remains unresolved.

Statistical Dependency Is Not Necessarily Functional Interaction

A statistical dependency between two nucleotides does not necessarily indicate a direct biological interaction.

For example, correlated changes between two positions could arise from:

  • Genuine molecular or functional interactions
  • Shared evolutionary history
  • Correlations within the same functional element
  • Shared biases in nucleotide composition or mutation rates
  • Statistical dependencies introduced by alignment structure

One particularly important concept is epistasis.

Epistasis occurs when the combined functional effect of multiple variants differs from what would be expected by simply adding their individual effects.

Whether the nucleotide dependencies learned by GPN-Star accurately reflect experimentally observed epistasis remains a separate question.

This could be investigated through deep mutational scanning or massively parallel reporter assays (MPRAs), comparing experimentally measured effects of single and double mutations with the model's predicted dependencies.

Learning evolutionary covariation is not equivalent to predicting causal functional interactions among multiple mutations.

9. Comparing GPN-Star and AlphaGenome

GPN-Star and AlphaGenome can both contribute to genomic variant interpretation, but their primary prediction targets differ.

Understanding this distinction requires examining what each model learns, not merely comparing benchmark scores.

Comparison of Major Models

Model Primary Learning Target Main Input or Training Information Main Question
GPN-Star Evolutionary constraint DNA, multiple sequence alignments, phylogenetic trees How evolutionarily constrained is this variant?
AlphaGenome Molecular function DNA and functional genomic supervision How does this variant affect expression, splicing, chromatin, or other molecular features?
Evo 2 General sequence distribution Large-scale genomic sequences across diverse organisms How plausible is this sequence under learned biological distributions, and what sequences can be generated?
AlphaMissense Missense pathogenicity Protein structure, evolutionary information, population variation, and related features Is this missense variant likely to be pathogenic?

The Question AlphaGenome Answers

AlphaGenome is a sequence-to-function model that predicts molecular properties from DNA sequences [8].

Its central prediction framework is:

Variant → Molecular phenotype

For example, AlphaGenome might predict:

"This variant may reduce enhancer activity in liver cells and decrease the expression of a particular gene."

This is a hypothetical example intended to illustrate the difference between the models.

AlphaGenome's strength lies in connecting sequence changes with molecular phenotypes, including cell- and tissue-associated effects.

However, a predicted change in molecular function does not necessarily imply that the variant causes disease.

The Question GPN-Star Answers

GPN-Star focuses on a different relationship:

Variant → Evolutionary constraint

For example, it might indicate:

"This nucleotide substitution appears to be poorly tolerated under the learned evolutionary sequence context."

Importantly, GPN-Star's score should not automatically be interpreted as a direct estimate of the strength of purifying selection.

The score is a prediction associated with evolutionary constraint, not a population-genetic selection coefficient.

Superior Pathogenicity Prediction Does Not Mean Universal Superiority Over AlphaGenome

The GPN-Star paper reports that GPN-Star outperformed AlphaGenome on particular pathogenic noncoding variant benchmarks [1].

However, these results describe differences in performance on specific evaluation tasks.

AlphaGenome is primarily designed to predict molecular functions and analyze tissue-specific changes.

GPN-Star focuses on variant prioritization using evolutionary constraint.

The two models therefore provide different types of information.

For example, GPN-Star may help prioritize candidate disease-associated variants.

AlphaGenome may then help investigate which genes or molecular processes those variants affect.

However, the hypothesis that combining both models will consistently improve predictive performance requires direct empirical testing.

10. Comparing GPN-Star and Evo 2

Evo 2 is a large-scale generative genomic model whose Nature paper was published in March 2026 [14].

It learns from large genomic datasets spanning all domains of life and aims to support applications ranging from variant effect prediction to novel sequence generation.

Specialist Versus Generalist

The distinction between GPN-Star and Evo 2 can be summarized as:

GPN-Star = Specialist

Evo 2 = Generalist

GPN-Star focuses on evolutionary constraint-based variant interpretation.

Evo 2 addresses a broader set of tasks, including zero-shot prediction of coding and noncoding variant effects, sequence generation, and analysis of biological information represented within the model.

Evo 2 aims to learn general statistical and biological regularities in DNA sequences that can support diverse downstream applications.

Model Scale Versus Biological Structure

The GPN-Star study demonstrated strong performance using approximately 200 million parameters and training for several days on eight A100 GPUs [1].

Evo 2 uses substantially greater computational resources but targets a broader range of capabilities [14].

These approaches represent two different design philosophies:

  • Explicitly incorporating biological knowledge into model architectures to solve specialized problems efficiently.
  • Using large-scale data and computation to learn representations applicable to many different tasks.

Their relative strengths depend on the prediction task, desired output, computational resources, and available input information.

GPN-Star demonstrates that scaling model size is not the only route to improving genomic AI.

11. Comparing GPN-Star With AlphaMissense and PrimateAI-3D

For coding variants, particularly missense variants, specialized predictors such as AlphaMissense and PrimateAI-3D remain important.

AlphaMissense

AlphaMissense is a model for predicting missense variant effects using information related to protein structure, evolution, and human and primate population variation [15].

Its main objective is to estimate the likelihood that an amino acid substitution is pathogenic.

PrimateAI-3D

PrimateAI-3D incorporates variation data from 233 primate species and three-dimensional protein structure information [16].

It combines structural and evolutionary information to predict the consequences of missense variation.

The Genome-Wide Advantage of GPN-Star

GPN-Star is not limited to protein-coding regions.

The same general framework can score variants in regions such as:

  • Missense variants
  • Synonymous variants
  • UTRs
  • Promoters
  • Enhancers
  • Intergenic regulatory regions

This broad applicability is a major strength.

However, when the task is restricted to missense pathogenicity prediction, specialized models such as AlphaMissense may achieve stronger performance.

The original study also reported that post-processing GPN-Star predictions with gnomAD allele-frequency information substantially improved performance on ClinVar, surpassing PrimateAI-3D and approaching AlphaMissense [1].

This demonstrates the importance of present-day human population variation alongside evolutionary constraint.

However, population frequency can also contribute to the evidence used in clinical variant classification.

Consequently, when evaluating improvements from adding allele-frequency information, it is important to consider the independence between model inputs and benchmark labels.

12. Challenges and Limitations of GPN-Star

GPN-Star achieves strong performance across numerous benchmarks, but these results do not establish that it can explain every type of genetic variation or disease.

The authors themselves acknowledge several important limitations [1].

Dependence on Whole-Genome Alignments

The same evolutionary information that gives GPN-Star much of its predictive power also creates a structural limitation.

Predictions may be more difficult in regions such as:

  • Human-specific regulatory elements
  • Rapidly evolving genomic regions
  • Repetitive sequences that are difficult to align
  • Regions containing structural variation
  • Sequences absent from the reference genome

A particularly important issue is that alignment quality determines which evolutionary information is available to the model.

Therefore, prediction reliability should be assessed alongside the quality of the underlying multiple sequence alignment.

Insertions, Deletions, and Structural Variants

GPN-Star is primarily suited to scoring single-nucleotide variants.

However, clinically important genetic variation also includes insertions and deletions, copy-number variants, inversions, translocations, and repeat expansions.

These variants can alter sequence length or genome structure rather than simply substituting one nucleotide for another.

Because GPN-Star relies on a fixed alignment-based representation, its applicability to these types of variants is limited [1].

By contrast, some single-sequence genomic language models can more naturally process modified input sequences, potentially making them more suitable for certain indel and structural-variation tasks.

Long-Range Genomic Structure

The original study found that increasing model size produced some improvements, whereas extending sequence context length yielded relatively modest gains [1].

The authors suggested that reference-based whole-genome alignments often contain fragmented synteny blocks.

As a result, long input windows may contain sequences that are not actually contiguous in the genomes of other species.

This may interfere with the model's ability to learn long-range regulatory relationships.

Therefore, improving long-context modeling may require advances not only in neural network architecture but also in the representation of alignment data.

Human-Specific Adaptation

GPN-Star primarily exploits evolutionary information shared across species.

However, understanding natural selection within present-day human populations or human-specific evolutionary processes may require different sources of information.

The original paper notes the potential importance of incorporating human population genomes and archaic human genomes to investigate more localized evolutionary processes [1].

Limited Direct Prediction of Molecular Mechanisms

GPN-Star predicts how evolutionarily constrained a variant appears to be.

It does not directly predict molecular mechanisms such as:

  • Which cell type is affected
  • Which transcription factor binding site is altered
  • Which enhancer mediates the effect
  • Which gene changes expression
  • The percentage change in gene expression

These are more directly addressed by sequence-to-function models such as AlphaGenome and Borzoi.

Evolutionary Deleteriousness Is Not Clinical Pathogenicity

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

Even when a variant receives a strong evolutionary constraint score, the model does not independently establish:

  • Which disease the variant is associated with
  • Its penetrance
  • The age of onset
  • Differences between heterozygous and homozygous effects
  • How genetic background or environment modifies the effect

Clinical interpretation therefore requires additional evidence, including patient phenotypes, family segregation, population frequencies, and functional experiments.

GPN-Star scores should not be used alone to determine clinical pathogenicity.

13. What Remains Insufficiently Validated?

GPN-Star has already undergone extensive evaluation involving multiple benchmarks, mutation-rate calibration, nucleotide dependency analyses, and applications to model organisms.

It would therefore be inaccurate to characterize functional prediction or pathogenicity prediction as entirely unvalidated.

The more important question is which claims remain insufficiently established beyond the evaluations already performed.

Major Open Questions

Research Question What Has Been Demonstrated What Remains Unresolved
Experimental functional effects Evaluated using ProteinGym and related datasets How accurately can noncoding expression effects and cell-type dependencies be predicted?
Discovery of novel causal variants Evaluated using existing disease and GWAS datasets Can the model prospectively identify previously unknown causal variants?
Benchmark independence Evaluated across diverse datasets Does its advantage remain after controlling for conservation, allele frequency, and related confounders?
Relationship to natural selection Evaluated through allele-frequency relationships and calibration How closely do scores correspond to selection coefficients or fitness effects?
Alignment robustness Evaluated across species and evolutionary ranges How sensitive are predictions to gaps, alignment errors, and species composition?
Evolutionary timescale Compared across three major phylogenetic ranges Can optimal evolutionary comparisons be selected at finer resolutions?
Nucleotide dependencies Demonstrated consistency with known functional elements Do these dependencies reflect actual epistasis or causal interactions?
Human population generalization Evaluated using human population variation Does performance remain consistent across different ancestry backgrounds?
Drug discovery applications Promising results for genetic variant prioritization Can the model improve target discovery or actual drug development outcomes?

Does Strong Benchmark Performance Generalize to Unseen Variants?

GPN-Star was evaluated using datasets including ClinVar, OMIM, HGMD, GWAS, and ProteinGym.

These datasets are valuable, but their labels represent different forms of evidence.

For example, ClinVar classifications reflect expert interpretations and existing clinical and genetic knowledge.

GWAS fine-mapping provides statistical estimates of causal involvement, but not experimental proof for every variant.

ProteinGym provides experimentally measured variant effects, although the relationship between the measured molecular phenotype and organism-level disease varies across proteins and experimental conditions.

Consequently, strong performance on one benchmark does not guarantee equally strong performance across all variant classes or diseases.

Prospective evaluations involving newly discovered variants are particularly important.

Do Evolutionary Constraint Scores Quantify the Strength of Natural Selection?

GPN-Star predicts evolutionary constraint, but its scores cannot automatically be interpreted as population-genetic selection coefficients.

For example, the model does not directly provide:

  • Selection coefficients
  • Quantitative fitness effects
  • Rates at which deleterious mutations are removed from human populations

The original study demonstrates that the scores contain information associated with natural selection, including relationships with variant frequencies in gnomAD [1].

However, population allele frequencies are also influenced by demographic history, mutation rates, genetic drift, and other factors.

Further work is therefore needed to establish how GPN-Star scores correspond to population-genetic measures of selection.

Do Learned Nucleotide Dependencies Reflect Actual Epistasis?

GPN-Star's nucleotide dependency analysis is an important demonstration of sequence-context modeling.

However, the relationship between learned statistical dependencies and experimentally measured non-additive functional effects remains incompletely understood.

This is not merely a question of predictive accuracy.

It concerns whether the model's internal representations correspond to actual biological mechanisms.

Experiments involving single and double mutations could provide a direct way to investigate this question.

Does Heritability Enrichment Establish Causal Variant Discovery?

Strong heritability enrichment is an important result.

However, enrichment does not necessarily imply accurate identification of individual causal variants.

Results may also be influenced by correlations between prediction scores and existing functional annotations.

Although the original study examined relationships with known annotations, further evaluation across independent populations, new traits, and experimentally validated variants remains valuable.

How Far Has Drug Discovery Utility Been Demonstrated?

GPN-Star may provide useful information for prioritizing variants and genes associated with disease.

However, improved variant prioritization is not equivalent to improved drug development success.

For example, identifying a disease-associated gene does not establish that inhibiting or activating that gene will produce a therapeutic benefit.

Drug target evaluation also requires information such as:

  • Whether variants increase or decrease gene function
  • Whether manipulating the target improves disease phenotypes
  • Which tissues or cell types should be targeted
  • Potential safety risks and adverse effects
  • Whether pharmacological intervention is feasible

GPN-Star has potential applications in drug discovery, but improvements in actual target discovery and therapeutic development require more direct validation.

14. Five Important Directions for Future Research

The limitations discussed above suggest several promising research directions.

Research Direction 1: Validation Against Experimental Noncoding Variant Effects

GPN-Star predictions could be compared with experimentally measured functional effects using methods such as:

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

Particularly important measurements would include:

  • Magnitude of gene expression changes
  • Direction of gene expression changes
  • Cell-type and tissue specificity
  • Identity of affected genes

Such experiments could clarify how closely evolutionary constraint scores correspond to molecular functional effects.

Research Direction 2: Robustness to Phylogenetic Trees and Alignment Structure

The same human variants could be scored using different species compositions, phylogenetic sampling strategies, and alignment quality levels.

It would also be interesting to examine how predictions change when local gene trees differ from the assumed species tree.

This research could clarify how much of GPN-Star's performance reflects robust evolutionary signals rather than particular choices in alignment construction.

Research Direction 3: Connecting Nucleotide Dependencies With Epistasis

The functional effects of single and double mutations could be measured experimentally and compared with the nucleotide dependencies learned by GPN-Star.

This would help determine whether its internal representations reflect genuine functional interactions.

Such work could move genomic language modeling beyond statistical sequence prediction toward a deeper understanding of biological mechanisms.

Research Direction 4: Prospective Validation on Previously Unseen Variants

An especially valuable experiment would involve prioritizing variants absent from existing benchmark datasets, fixing the model's predictions in advance, and then experimentally testing those variants.

GPN-Star could be compared with approaches such as phyloP, phastCons, CADD, and AlphaGenome.

This would assess not only performance on established benchmarks but also the model's ability to contribute to genuinely new biological discoveries.

Research Direction 5: Integrating Evolutionary and Functional Genomic Information

Another important direction is to combine:

  • Evolutionary constraint predictions from GPN-Star
  • Molecular function predictions from AlphaGenome
  • Population allele-frequency information

Performance could then be compared with the use of each model individually.

Such studies should carefully consider overlapping training data, differences in prediction targets, and the independence of evaluation datasets.

The goal would be to determine quantitatively whether evolutionary and functional genomic information are genuinely complementary.

15. Three Major Directions in Contemporary Genomic AI

The current landscape of genomic AI can be understood through three broad research directions.

A. Evolution: Learning Evolutionary Constraint

Representative research progression:

phastCons / phyloP
↓
GPN
↓
GPN-MSA
↓
PhyloGPN / GPN-Star

This direction uses evolutionary conservation and constraints across species to estimate the functional importance of genetic variants.

GPN-Star is an important model within this tradition.

B. Experimental Function: Predicting Molecular Phenotypes

Representative research progression:

DeepSEA / Basset
↓
ExPecto / Basenji
↓
Enformer
↓
Borzoi
↓
AlphaGenome

This direction predicts gene expression, chromatin states, and other molecular functions directly from DNA sequences.

It uses experimental functional genomic data to learn relationships between sequence variation and molecular phenotypes.

C. General Foundation and Generative Models: Learning Broad Sequence Representations

Representative research progression:

DNABERT
↓
Nucleotide Transformer / HyenaDNA / Caduceus
↓
Evo
↓
Evo 2

This direction uses large genomic datasets to build general-purpose models supporting diverse prediction and sequence-generation tasks.

These Three Approaches May Complement One Another

Future genomic AI research may increasingly combine information from:

  • DNA sequence
  • Evolutionary alignments and phylogeny
  • Population variation
  • Functional genomics
  • Cell and tissue context
  • Three-dimensional chromatin and protein structures

Integrating these sources could support models that more comprehensively connect evolutionary constraint, molecular function, and disease-associated phenotypes.

However, combining all these information types into a single foundation model does not guarantee better performance.

Different data modalities represent different biological processes, and their actual complementarity must be demonstrated empirically.

Furthermore, predicting molecular functions from genomic sequences is not equivalent to simulating every biological process within a living organism.

GPN-Star itself is not a model that simulates the entire living system.

Its direct focus is evolutionary constraint-based prediction of genetic variant effects.

16. Conclusion: What GPN-Star Has Changed — and What It Has Not

GPN-Star can be summarized within the history of genomic AI as follows:

GPN-Star integrates classical comparative genomics, represented by approaches such as phastCons and phyloP, with Transformer-era genomic language modeling to advance the prediction of evolutionary constraint.

Its significance extends beyond improving the accuracy of variant effect prediction.

It provides an alternative to the prevailing emphasis during approximately 2023–2025 on scaling increasingly large DNA language models.

By explicitly incorporating evolutionary structure into the architecture, GPN-Star demonstrates that comparatively small models can achieve strong performance on specialized genomic prediction tasks.

It also reveals that different evolutionary timescales provide different forms of information about genomic constraints.

The finding that relatively shallow evolutionary comparisons can be particularly informative for noncoding regulatory variation and complex traits has important implications for future research.

At the same time, the relationship between evolutionary constraint and present-day functional importance remains an open scientific problem.

Further research is required to determine how closely patterns preserved by natural selection correspond to molecular function, disease causation, and the effects of therapeutic interventions in present-day humans.

Similarly, although GPN-Star learns nucleotide dependencies, the extent to which these relationships represent actual functional interactions or epistasis remains to be established.

Future research on GPN-Star can be organized into two broad directions.

The first is understanding evolution more accurately.

This involves investigating the relationships among phylogenetic trees, sequence alignments, mutation rates, natural selection, and nucleotide dependencies to clarify what evolutionary information the model has actually learned.

The second is using evolutionary constraint to better understand contemporary biological function.

This involves connecting evolutionary information with experimental functional genomics, population genetics, and cell- or tissue-specific data to improve predictions of causal variant effects.

The second direction is particularly interesting because it connects the research philosophies underlying GPN-Star and AlphaGenome.

GPN-Star is a powerful predictor of evolutionary constraint.

However, it has not demonstrated that evolutionary information alone can explain every aspect of molecular function or disease causation.

Instead, the study illustrates both the remarkable predictive value of evolutionary information and the scientific questions that evolutionary information alone cannot resolve.

GPN-Star's fundamental contribution is not only that it predicts evolutionary constraints with high accuracy, but also that it demonstrates the value of embedding biologically meaningful structure into AI while clarifying which biological questions remain unresolved.

How evolutionary information can be connected with experimentally measured molecular function may become one of the most important directions for the next generation of genomic AI.

References

[1] Ye, C., Benegas, G., Albors, C., 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] Kircher, M., et al., "A general framework for estimating the relative pathogenicity of human genetic variants," Nature Genetics, 46, 310–315, 2014. https://www.nature.com/articles/ng.2892

[4] 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

[5] 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

[6] Avsec, Ž., et al., "Effective gene expression prediction from sequence by integrating long-range interactions," Nature Methods, 18, 1196–1203, 2021. https://doi.org/10.1038/s41592-021-01252-x

[7] Linder, J., et al., "Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation," Nature Genetics, 2025. https://doi.org/10.1038/S41588-024-02053-6

[8] Avsec, Ž., et al., "Advancing regulatory variant effect prediction with AlphaGenome," Nature, 649, 1206–1218, 2026. https://www.nature.com/articles/s41586-025-10014-0

[9] Dalla-Torre, H., et al., "Nucleotide Transformer: building and evaluating robust foundation models for human genomics," Nature Methods, 2024. https://www.nature.com/articles/s41592-024-02523-z

[10] Nguyen, E., et al., "Sequence modeling and design from molecular to genome scale with Evo," Science, 2024. https://pubmed.ncbi.nlm.nih.gov/39541441/

[11] Song Lab, "GPN: Genomic Pre-trained Network," GitHub repository. https://github.com/songlab-cal/gpn

[12] "GPN-MSA," Nature Biotechnology, 2025. https://www.nature.com/articles/s41587-024-02511-w

[13] "PhyloGPN," arXiv preprint, 2025. https://arxiv.org/abs/2503.03773

[14] Brixi, G., et al., "Genome modelling and design across all domains of life with Evo 2," Nature, 652, 1349–1361, 2026. https://www.nature.com/articles/s41586-026-10176-5

[15] Cheng, J., et al., "Accurate proteome-wide missense variant effect prediction with AlphaMissense," Science, 381, eadg7492, 2023. https://pubmed.ncbi.nlm.nih.gov/37733863/

[16] "PrimateAI-3D," Nature Genetics, 2023. https://www.nature.com/articles/s41588-023-01455-2

Additional Resources

The following resources provide further technical information about GPN-Star.