Target Audience
- Researchers and practitioners interested in the intersection of comparative genomics, evolutionary biology, machine learning, and deep learning.
- Readers seeking to understand the evolution of genomic language models, genomic foundation models, and technologies for in vivo simulation.
- Professionals interested in AI applications for predicting the functional effects of genetic variants, interpreting disease-associated variants, and developing new therapeutics.
Summary
- GPN-Star combines classical comparative genomics with Transformer-based genomic language modeling to predict evolutionary constraints across the genome.
- Rather than relying on large DNA language models to implicitly learn evolutionary relationships, GPN-Star explicitly incorporates multiple sequence alignments and phylogenetic trees into its architecture, achieving strong predictive performance.
- While AlphaGenome focuses on molecular function prediction and Evo 2 aims for general-purpose sequence understanding and generation, GPN-Star specializes in predicting variant effects through evolutionary constraints.
- One of its important findings is that evolutionary constraint is not a single, universal property: different evolutionary timescales reveal different types of biological function.
- Future genomic foundation models may increasingly integrate evolutionary information, functional genomics, population genetics, cellular context, and structural biology.
1. The Evolution of Machine Learning in Genomics
GPN-Star occupies an important position in the evolution of machine learning-based genomic research.
In simple terms, it represents a shift away from the assumption that sufficiently large DNA language models will automatically learn evolutionary constraints, toward an approach that explicitly incorporates biological structures such as phylogenetic trees and multiple sequence alignments.
However, this is not simply a return to classical comparative genomics.
The fundamental contribution of GPN-Star lies in connecting evolutionary models, such as phastCons and phyloP, with representation learning based on Transformer architectures.
The Nature version of the GPN-Star paper was published on September 9, 2026 [1].
To understand its significance, it is useful to divide the history of machine learning in genomics into five major stages.
First Generation: Comparative Genomics — Conserved Regions Are Likely Important
In the 2000s, methods such as phastCons used whole-genome alignments across multiple species and phylogenetic trees to identify genomic regions with unusually low substitution rates.
These regions were interpreted as evolutionarily conserved sequences.
phastCons uses a phylogenetic hidden Markov model to explicitly represent nucleotide substitution processes and the spatial continuity of conserved regions along the genome [2].
The basic framework can be summarized as follows:
Species alignment → Evolutionary model → Constraint score
The underlying idea is that nucleotides or genomic regions conserved across species over long evolutionary periods are less tolerant of change and may therefore perform important biological functions.
GPN-Star can be viewed as a direct intellectual descendant of this approach, which originated more than two decades ago.
Second Generation: Integrating Multiple Annotations with Machine Learning
During the 2010s, models such as CADD and LINSIGHT emerged.
CADD integrates numerous features, including evolutionary conservation, gene annotations, and epigenomic information, using machine learning to estimate how likely an individual genetic variant is to be deleterious.
The original 2014 paper used a support vector machine (SVM) [3].
LINSIGHT represents another particularly interesting development.
It combines three areas:
- Functional genomics
- Machine learning
- Molecular evolution
In terms of its underlying research philosophy, LINSIGHT is relatively close to GPN-Star [4].
However, CADD and LINSIGHT fundamentally rely on features designed by human researchers.
Third Generation: Predicting Function Directly from DNA with Deep Learning
DeepSEA, introduced in 2015, marked a major turning point.
DeepSEA demonstrated that convolutional neural networks (CNNs) could predict functional genomic properties directly from DNA sequences, including [5]:
- Transcription factor binding
- DNase hypersensitivity
- Histone modifications
This established a research paradigm that can be expressed as:
DNA → Neural network → Functional genomics
This approach forms the foundation of the sequence-to-function modeling paradigm that remains influential today.
The major developments in this lineage can be summarized as:
DeepSEA / Basset
↓
ExPecto
↓
Basenji
↓
Enformer
↓
Borzoi
↓
AlphaGenome
Enformer introduced Transformer-based modeling of regulatory interactions spanning approximately 100 kb, demonstrating that long-range enhancer–promoter relationships could be learned from DNA sequences [6].
Borzoi extended the sequence context to more than 500 kb.
By directly predicting RNA-seq coverage, it moved toward a unified representation of transcription, splicing, and polyadenylation [7].
AlphaGenome, published in 2026, further expanded this paradigm by predicting numerous molecular functions simultaneously from 1 Mb DNA sequences [8].
These functions include:
- Gene expression
- Chromatin accessibility
- Transcription factor binding
- Histone modifications
- Splicing
- 3D chromatin contacts
AlphaGenome predicts thousands of functional genomic tracks simultaneously.
This research lineage represents one of the central directions in modern genomic AI: directly predicting molecular function from DNA sequence.
Fourth Generation: Treating DNA as a Language Through Self-Supervised Learning
Following the success of BERT and GPT in natural language processing, researchers began applying language modeling approaches to DNA sequences.
Nucleotide Transformer introduced models with up to 2.5 billion parameters.
Through self-supervised learning on numerous human genomes and genomes from 850 species, these models learned representations that could be fine-tuned for a variety of genomic tasks [9].
Other models in this broader lineage include:
- DNABERT
- Nucleotide Transformer
- HyenaDNA
- Caduceus
- Evo
- Evo 2
The development from Evo to Evo 2 moved beyond simple feature extraction toward sequence prediction and generation involving DNA, RNA, and proteins.
Evo also demonstrated the ability to generate sequences at genomic scales [10].
However, this development raised an important question:
If a model learns from sufficiently large amounts of DNA sequence data, will it automatically acquire an understanding of evolutionary constraints?
For human variant effect prediction, the answer was not always as favorable as expected.
The GPN-Star paper identifies this as a central motivation: although large single-sequence genomic language models require substantial computational resources, they may still underperform classical evolutionary models or alignment-based models on constraint prediction tasks [1].
This challenge helped motivate the development of the GPN family of models.
Fifth Generation: Genomic Language Models That Explicitly Incorporate Evolution
The GPN research lineage began by exploring whether evolutionary constraints could be learned implicitly from DNA sequences.
Over time, it evolved toward explicitly using multiple sequence alignments and phylogenetic information.
GPN-Star represents an important milestone in this progression.
Rather than simply increasing model size, this approach incorporates biologically meaningful structures into the architecture to improve predictive performance and computational efficiency.
2. The Technical Evolution from GPN to GPN-Star
Understanding the progression of the GPN family is essential to understanding GPN-Star itself.
Its evolution can be summarized as:
GPN → GPN-MSA → PhyloGPN → GPN-Star
GPN: Learning from Unaligned Genomes
The original GPN, introduced in 2023, trained a DNA language model using unaligned genomes from Arabidopsis and closely related plants.
It demonstrated that self-supervised learning could extract information useful for predicting variant effects without requiring explicit supervised labels [11].
At this stage, the objective was to identify functionally important variants from statistical patterns in genomic sequences.
GPN-MSA: Directly Using Multiple Sequence Alignments
GPN-MSA advanced the concept by asking a different question:
Rather than forcing a model to learn evolution implicitly, would it be better to provide multiple sequence alignments directly as input?
GPN-MSA uses a 100-way vertebrate whole-genome alignment.
This approach achieved strong variant effect prediction performance in both coding and non-coding regions [12].
It demonstrated the value of directly supplying comparative evolutionary information to a model.
However, GPN-MSA also had limitations.
It depended heavily on particular alignment structures and required substantial manual adjustment of factors such as species composition.
PhyloGPN: Separating Training Information from Inference Requirements
PhyloGPN addressed a related challenge from a different direction.
During training, the model uses phylogenetic and alignment information.
During inference, however, it can make predictions using only a single DNA sequence [13].
In other words, the model uses evolutionary information as a source of training supervision while simplifying the input requirements at inference time.
GPN-Star: Explicitly Modeling the Phylogenetic Tree
GPN-Star incorporates species trees in addition to whole-genome alignments.
This allows the model to represent not only the nucleotide sequences of multiple species but also the evolutionary relationships among those species [1].
Its major improvements over GPN-MSA include:
- Performing masked nucleotide prediction across multiple species in the alignment, rather than focusing only on humans.
- Explicitly incorporating the species tree into the model architecture.
- Handling alignments with different species compositions more flexibly.
GPN-Star also introduces calibration to account for variation in mutation rates [1].
3. The Core Innovation of GPN-Star: Incorporating Phylogenetic Trees into Transformers
The central technical innovation of GPN-Star is its ability to use not only whole-genome alignments but also the structure of the species tree within its attention mechanism.
Its overall processing pipeline can be summarized as:
Whole-genome alignment
↓
Species-tree-aware pooling
↓
Phylogeny-aware cross-attention
↓
Nucleotide probabilities
↓
Variant constraint score
Species Are Not Treated as Independent Sequences
In some conventional sequence models, DNA sequences from different species are processed simply as multiple inputs.
However, species are not independent of one another.
They share evolutionary histories and have diverged from common ancestors.
GPN-Star groups related species into clades and integrates information through cross-attention mechanisms that account for their phylogenetic relationships [1].
This allows the model to estimate evolutionary constraints while considering both nucleotide-level information and species relationships.
GPN-Star as a Biology-Informed Foundation Model
One of GPN-Star's defining characteristics is that biological knowledge is incorporated directly into the model architecture.
It is therefore better understood as a biology-informed foundation model rather than a biology-free foundation model.
A common direction in large language model development is to improve capability by expanding model size and training data.
GPN-Star takes a different approach.
It incorporates evolution, an established biological structure, as an inductive bias to improve performance on a specific predictive task.
This reflects a broader principle that is becoming increasingly important across AI research:
Instead of simply scaling up a model, incorporating domain-specific structures and knowledge can be an effective way to improve its performance.
4. Different Evolutionary Timescales Reveal Different Biological Functions
Another important finding from GPN-Star is that using a broader evolutionary comparison is not always better.
In classical comparative genomics, including more species and examining longer evolutionary periods can help identify important conserved genomic regions.
However, GPN-Star explicitly compares different evolutionary timescales.
The principal taxonomic groups considered include:
- Vertebrates
- Mammals
- Primates
The Optimal Evolutionary Timescale Depends on the Prediction Task
For strong evolutionary constraints in coding regions and missense variants, deeper evolutionary histories can be useful.
However, for non-coding regulatory variants and complex traits, information from relatively recent evolutionary timescales, such as mammals or primates, can be particularly informative.
The paper reports that models trained on primates showed very strong heritability enrichment across more than 100 complex traits [1].
This finding has important conceptual implications.
Functional constraint is not a single universal property; it depends on evolutionary timescale.
For example:
- Genomic regions conserved for approximately 500 million years may perform fundamental functions essential for survival.
- Regions conserved over approximately 30–100 million years may be associated with mammal-specific regulatory functions.
- Regions strongly constrained among primates may be related to complex regulatory phenotypes particularly relevant to humans.
This suggests that calculating a single conservation score for every variant may not always capture the full biological picture.
The evolutionary timescale over which a sequence is conserved can itself provide valuable information about its function.
5. Comparing GPN-Star with AlphaGenome
GPN-Star and AlphaGenome are both important models in genomic AI, but they differ in what they learn and what they predict.
Although they compete in certain benchmarking tasks, they fundamentally address different questions.
Comparison of Major Models
| Model | Primary Learning Target | Input Information | Main Question |
|---|---|---|---|
| GPN-Star | Evolutionary constraint | DNA, multiple sequence alignments, phylogenetic trees | Is this variant evolutionarily tolerated? |
| AlphaGenome | Molecular function | DNA, supervised functional genomics data | How does this variant affect expression, splicing, or chromatin? |
| Evo 2 | General sequence distribution | Large-scale DNA datasets from diverse organisms | Is this sequence biologically plausible, and can similar sequences be generated? |
| AlphaMissense | Protein pathogenicity | Protein sequence and structure, population variation | Is this missense variant pathogenic? |
AlphaGenome Predicts Molecular Function
AlphaGenome was published in Nature in January 2026.
It predicts molecular functions from 1 Mb DNA sequences, including 5,930 human genomic tracks.
It also demonstrated performance comparable to or better than existing methods on 25 out of 26 variant-effect benchmarks [8].
AlphaGenome primarily addresses the relationship:
Variant → Molecular phenotype
It predicts how DNA sequence changes may affect gene expression, splicing, chromatin states, and other molecular properties.
GPN-Star Predicts Evolutionary Constraint
GPN-Star, in contrast, primarily estimates:
Variant → Evolutionary fitness constraint
It predicts how strongly a nucleotide substitution may be disfavored by evolutionary processes.
For example, consider a particular SNP.
GPN-Star might suggest:
"This nucleotide substitution is inconsistent with strong purifying selection and may therefore be deleterious."
AlphaGenome might suggest:
"This nucleotide substitution may reduce enhancer activity in liver cells and decrease the expression of GENE X."
These are illustrative examples intended to clarify the different roles of the models.
In practical analyses, combining evolutionary constraint predictions from GPN-Star with molecular function predictions from AlphaGenome could provide a more comprehensive interpretation of genetic variation.
Comparing Pathogenic Variant Prediction
In the GPN-Star paper, comparisons involving pathogenic non-coding variants from OMIM and HGMD showed that GPN-Star outperformed sequence-to-function models such as Enformer, Borzoi, and AlphaGenome.
The difference was particularly notable for distal enhancers [1].
However, this does not mean that GPN-Star is universally superior to AlphaGenome.
The two models have different strengths.
- Pathogenicity and evolutionary constraint prediction: GPN-Star demonstrates strong performance.
- Molecular mechanisms and tissue-specific functional effects: AlphaGenome provides important advantages.
Clinical genomic analysis and GWAS interpretation can benefit from both types of information.
6. Comparing GPN-Star with Evo 2
Evo 2 is a large-scale generative genomic model whose Nature version was published in March 2026 [14].
It is a general-purpose genomic sequence model spanning all domains of life and can also perform zero-shot prediction of human coding and non-coding variant pathogenicity.
However, its research objectives differ substantially from those of GPN-Star.
Specialist Versus Generalist
The distinction can be summarized as:
GPN-Star = Specialist
Evo 2 = Generalist
GPN-Star focuses primarily on evolutionary constraint prediction.
Evo 2, by contrast, aims to understand genomic sequences across diverse organisms and generate new DNA sequences, in addition to variant scoring.
Model Scaling Versus Biological Knowledge
The GPN-Star paper reports that a model with approximately 200 million parameters, trained for several days on eight A100 GPUs, can outperform much larger genomic language models on constraint prediction by explicitly incorporating evolutionary information [1].
Evo 2 uses substantially greater computational resources but offers a wider range of capabilities, including DNA generation [14].
Therefore, this comparison is not simply about which model is better.
It represents two different design philosophies:
Incorporating specialized biological knowledge directly into model architecture
versus
Acquiring general-purpose knowledge through large-scale data and computation
This distinction is relevant not only to genomics but also to AI research more broadly.
7. Comparing GPN-Star with AlphaMissense and PrimateAI-3D
When focusing specifically on coding variants, particularly missense variants, specialized models such as AlphaMissense and PrimateAI-3D are highly competitive.
AlphaMissense
AlphaMissense is a specialized model for missense variant prediction that uses information related to protein structure, evolution, and population variation in humans and primates [15].
Its primary objective is to estimate whether an amino acid substitution is likely to be pathogenic.
PrimateAI-3D
PrimateAI-3D uses genetic variation from 233 primate species together with three-dimensional protein structural information [16].
By combining protein structure with evolutionary variation, it represents a different type of specialization from GPN-Star.
GPN-Star's Advantage Is Its Genome-Wide Applicability
A major advantage of GPN-Star is that it is not restricted to protein-coding regions.
Using the same general framework, it can assign scores to:
- Missense variants
- Synonymous variants
- UTRs
- Promoters
- Enhancers
- Intergenic regulatory regions
This broad genomic coverage is one of GPN-Star's most important strengths.
However, when the task is limited to missense pathogenicity prediction, specialized models such as AlphaMissense may still perform better in certain settings.
The GPN-Star paper also highlights the importance of population allele-frequency information through comparisons with AlphaMissense and related models.
Combining GPN-Star scores with gnomAD allele frequencies through post-processing substantially improved predictive performance [1].
This demonstrates that both long-term evolutionary constraint and variation observed in contemporary human populations are important.
8. Three Major Directions in Modern Genomic AI
Current genomic AI research can be organized into three major directions.
A. Evolution: Learning Evolutionary Constraints
phastCons / phyloP
↓
GPN
↓
GPN-MSA
↓
PhyloGPN
↓
GPN-Star
This research lineage uses evolutionary conservation and constraints across species to predict the importance of genetic variation.
GPN-Star represents one of the latest generations of models in this direction.
B. Experimental Function: Predicting Molecular Function
DeepSEA / Basset
↓
ExPecto / Basenji
↓
Enformer
↓
Borzoi
↓
AlphaGenome
This lineage directly predicts molecular properties, such as gene expression and chromatin states, from DNA sequences.
It learns from experimentally measured functional genomic data to estimate how sequence changes affect molecular phenotypes.
C. General Foundation and Generative Models
DNABERT
↓
Nucleotide Transformer / HyenaDNA / Caduceus
↓
Evo
↓
Evo 2
This lineage aims to build general-purpose models capable of performing a wide range of predictions and generating sequences by learning from large and diverse genomic datasets.
These Three Directions May Eventually Converge
Although these research directions have developed somewhat independently, future models may increasingly integrate their respective information sources.
These include:
- DNA sequence
- Evolutionary alignment / phylogeny
- Population variation
- Functional genomics
- Cell / tissue context
- 3D chromatin / protein structure
By incorporating these information sources into unified foundation models, it may become possible to connect genomic sequence variation with molecular functions and ultimately disease-related phenotypes in a more comprehensive way.
This direction is also relevant to the development of technologies for computationally predicting and simulating molecular processes inside living organisms.
However, GPN-Star itself is not a model that simulates all biological processes occurring within a living organism.
Its immediate objective is predicting variant effects through evolutionary constraints.
In this sense, GPN-Star provides important evidence about how evolutionary information should be incorporated into genomic foundation models.
9. Limitations and Practical Applications of GPN-Star
Despite its strengths, GPN-Star also has several clear limitations.
Dependence on Whole-Genome Alignments
One of GPN-Star's most significant limitations is its dependence on whole-genome alignments, which are also a major source of its predictive power.
Evolutionary constraint alone may be insufficient for accurately characterizing:
- Recently emerged human-specific regulatory elements
- Rapidly evolving genomic regions
- Repetitive sequences that are difficult to align
- Regions involving structural variation
When reliable sequence comparisons across species are unavailable, predictions based on evolutionary constraints become more challenging.
It Does Not Directly Predict Molecular Mechanisms
GPN-Star may predict that a particular variant is likely to be deleterious.
However, it does not directly answer detailed mechanistic questions such as:
- In which cell type does the effect occur?
- Which transcription factor binding event is altered?
- Which enhancer mediates the effect?
- Which gene changes its expression?
- By what percentage does gene expression change?
These are the types of questions addressed by sequence-to-function models such as AlphaGenome and Borzoi.
Evolutionary Deleteriousness Is Not the Same as Clinical Pathogenicity
Another important distinction is that evolutionary deleteriousness and clinical pathogenicity are not equivalent.
A variant subject to strong natural selection may not have a simple relationship with a particular disease.
Conversely, a disease-associated variant may not necessarily exhibit strong evolutionary constraint.
For this reason, a GPN-Star score alone should not be treated as sufficient evidence for a clinical pathogenicity determination.
Integrating Multiple Models Is a Natural Practical Strategy
For clinical genomics and disease research, it is useful to consider combining multiple information sources:
GPN-Star score
- AlphaGenome prediction
- Population allele frequency
- Protein-specific predictors
- Phenotype / segregation evidence
GPN-Star provides evolutionary constraint information.
AlphaGenome predicts molecular functional effects.
Models such as AlphaMissense provide predictions specifically related to protein function and pathogenicity.
Population allele frequencies, patient phenotypes, and segregation patterns within families offer additional evidence.
Integrating these sources may help capture information that no single model can fully represent.
In drug development, such integration could also contribute to understanding disease-associated variants and the functions of their associated genes.
However, this should be understood as a potential direction for combining GPN-Star with other variant interpretation technologies, rather than evidence that GPN-Star alone has demonstrated drug discovery performance.
10. Conclusion: GPN-Star Offers a Distinct Philosophy for Designing Genomic Foundation Models
GPN-Star's position in the history of genomic machine learning can be summarized in one sentence:
GPN-Star brings together classical comparative genomics, represented by methods such as phastCons and phyloP, and Transformer-era genomic language modeling.
Its most important contribution is that it offers an alternative to the model-scaling approach that became particularly prominent around 2023–2025.
Rather than simply making DNA language models larger, GPN-Star demonstrates that explicitly incorporating evolutionary structures can enable relatively small models to perform sophisticated variant interpretation.
GPN-Star is a direct successor to GPN-MSA, while also addressing the broader question of how genomic foundation models should be designed from a direction distinct from AlphaGenome and Evo 2.
Its significance extends beyond improving variant effect prediction accuracy.
It demonstrates that incorporating biologically meaningful structures into AI systems can fundamentally influence both model performance and their practical applications.
Future genomic AI research may increasingly integrate evolutionary information, experimentally measured molecular functions, population genetics, cellular and tissue states, and three-dimensional biological structures.
Within this broader research landscape, GPN-Star occupies an important position by demonstrating the value of explicitly modeling evolutionary information.
References
[1] “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] “Advancing regulatory variant effect prediction with AlphaGenome,” Nature, 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] “GPN: Genomic Pre-trained Network,” GitHub repository, Song Lab.
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] “Evo 2,” Nature, 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
[17] “GPN-Star” PubMed preprint, 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12458161/