Nucleic Acid and Genome Models
Genome models move life sciences AI from protein sequence toward regulatory sequence, RNA, genome organisation, and cellular context. The unit of modelling is no longer only a protein product. Evo, Evo 2, Semantic Design, AlphaGenome, Nucleotide Transformer, GET, and Orthrus occupy different parts of the genome-model stack: genome-scale generation, regulatory prediction, transcriptional modelling, and RNA representation. These systems differ from protein language models because regulatory function depends on cell type, chromatin context, and measurement modality. The shared discipline is that regulatory function is not the same as disease mechanism.
Use this chapter to:
- Model DNA, RNA, regulatory sequence, genome organization, and sequence-to-function relationships across organisms and contexts.
- Sequence alone rarely proves function; tissue, cell type, chromatin, conservation, perturbation, and measurement design define the claim.
Prerequisites: Foundation Models for Biology for the pretrain-then-transfer pattern; Variant Effect Prediction for the clinical-framework context of non-coding variants.
Summary: Model DNA, RNA, regulatory sequence, genome organization, and sequence-to-function relationships across organisms and contexts. Genome and RNA foundation models are expanding quickly, with stronger evidence for selected regulatory and representation tasks than for broad causal biology.
Key point: Sequence alone rarely proves function; tissue, cell type, chromatin, conservation, perturbation, and measurement design define the claim. Open question: whether sequence models can predict function across tissue, cell state, chromatin context, and organismal biology.
Bottom line: Genome models connect molecular biology to variant interpretation, cell states, development, plant biology, aging, disease mechanisms, and therapeutics.
What is this field trying to solve? Model DNA, RNA, regulatory sequence, genome organization, and sequence-to-function relationships across organisms and contexts.
What is the core idea? Sequence alone rarely proves function; tissue, cell type, chromatin, conservation, perturbation, and measurement design define the claim.
What is the current state of the field? Genome and RNA foundation models are expanding quickly, with stronger evidence for selected regulatory and representation tasks than for broad causal biology.
What do we know, and what remains open? Known reference points include Evo, Evo 2, Nucleotide Transformer, AlphaGenome, Enformer, Basenji, Orthrus, GPN-MSA, GenBank, RefSeq, ENCODE, GTEx, gnomAD, and variant-function benchmarks. What remains open is whether sequence models can predict function across tissue, cell state, chromatin context, and organismal biology.
Why does this matter? Genome models connect molecular biology to variant interpretation, cell states, development, plant biology, aging, disease mechanisms, and therapeutics.
Introduction
Evo was presented as a biological foundation model operating from molecular to genome scale (Nguyen et al., 2024). Evo 2 extended this to 40 billion parameters across all domains of life and moved the result from preprint to Nature in 2026 (Brixi et al., 2026). Semantic design with Evo then showed function-guided de novo gene design in prokaryotic contexts, including generated anti-CRISPR proteins and toxin-antitoxin systems (Merchant et al., 2026). AlphaGenome focuses on regulatory variant-effect prediction from long DNA sequence context (Avsec et al., 2026). Nucleotide Transformer trained on diverse human and multi-species genome data (Dalla-Torre et al., 2025). GET models transcription across human cell types (Fu et al., 2025). These systems differ from protein language models because regulatory function depends on cell type, chromatin context, and measurement modality.
The regulatory-genomics spine
Genome modelling should be read as a progression from short sequence motifs toward long-context regulatory prediction. DeepSEA learned chromatin-feature effects from local sequence (Zhou and Troyanskaya, 2015). Enformer extended the context window so distal regulatory sequence could inform expression and chromatin predictions (Avsec et al., 2021). GET moves from sequence alone toward a transcriptional model across human cell types (Fu et al., 2025). AlphaGenome then frames regulatory variant prediction at megabase context (Avsec et al., 2026).
The lineage matters because “long context” is not a solution by itself. A regulatory model needs enough context to include distal elements, but it also needs the right cell type, assay track, expression state, and perturbational check. A megabase input that lacks the relevant cell-state evidence is still a sequence model, not a full regulatory model.
RNA and DNA models answer different questions
DNA regulatory models often ask how sequence changes affect chromatin, expression, or splicing. RNA models ask about mature RNA sequence properties, structure, stability, translation, localization, or interaction. Orthrus, RhoFold+, and AlphaFold 3 occupy RNA-adjacent territory, but they should not be collapsed into one “genome model” category. RNA therapeutic design adds delivery, immune sensing, chemical modification, and tissue expression constraints that a mature-RNA representation alone does not settle.
This distinction matters for vaccine and RNA therapeutic claims. A sequence model that ranks UTRs or coding sequences is not proving protein expression, immunogenicity, or clinical effect. It is prioritizing constructs for the next assay.
Validation ladder for regulatory claims
The validation ladder for a regulatory prediction starts with held-out sequence or assay tracks and moves toward perturbation. MPRA tests many candidate regulatory sequences outside native chromatin context; CRISPRi tests enhancer-promoter logic in the genome; eQTL and colocalization analyses connect human variation to expression and disease; Perturb-seq and related screens connect genetic perturbation to cell-state response. Each layer answers a different question.
For clinical use, the ladder ends inside a variant-interpretation framework. A high AlphaGenome, Enformer, or GPN-MSA score can prioritize evidence review, but it does not classify a non-coding variant. Population evidence, phenotype fit, functional validation, and laboratory-specific calibration still determine interpretation.
What is demonstrated?
Demonstrated capability includes sequence-to-function prediction for specific genomic tracks and zero-shot or few-shot transfer for selected molecular tasks. Evo demonstrated modelling across DNA, RNA, and proteins in the Science 2024 paper (Nguyen et al., 2024). Evo 2 demonstrated broader all-domains genome modelling, mutational effect prediction, interpretability analyses, and genome-scale generation in Nature (Brixi et al., 2026). Semantic design with Evo demonstrated function-guided generated genes with experimental validation in selected prokaryotic systems (Merchant et al., 2026). AlphaGenome demonstrated regulatory variant-effect prediction using megabase-scale DNA sequence inputs in Nature (Avsec et al., 2026). Enformer demonstrated transformer-based prediction of gene expression and chromatin states from long-range sequence context (Avsec et al., 2021). DeepSEA demonstrated single-nucleotide chromatin impact prediction at scale (Zhou and Troyanskaya, 2015).
RNA modelling is adjacent but not interchangeable with DNA regulatory modelling. Orthrus demonstrated mature RNA representation learning through contrastive pretraining over splice isoforms and mammalian orthologues (Fradkin et al., 2026). RhoFold+ demonstrated RNA 3D structure prediction from sequence using an RNA language-model approach (Shen et al., 2024), while AlphaFold 3 handles RNA inside broader biomolecular-complex prediction. Functional RNA design still needs experimental evidence because secondary structure, tertiary contacts, modification state, and cellular context can all change the relevant conformation.
For non-coding regulatory claims, perturbational assays remain the evidence bridge from predicted track change to biological effect. Massively parallel reporter assays can test many candidate regulatory variants in parallel (Tewhey et al., 2016), and CRISPR interference can map enhancer-promoter links in their chromatin context (Fulco et al., 2016). Genome models help rank hypotheses; they do not replace perturbational validation.
| Evidence Anchor | What It Supports | Practical Constraint |
|---|---|---|
| Evo / Evo 2 | Genome-scale sequence modelling across biological modalities | Training organism coverage defines transfer |
| Semantic design with Evo | Function-guided de novo gene design in tested prokaryotic contexts | Activity claims are specific to the validated systems |
| AlphaGenome | Regulatory variant-effect prediction from long sequence | Human and mouse training context does not equal all biology |
| GET | Human cell-type transcription prediction and regulatory grammar | Transcriptional prediction does not equal phenotype prediction |
| Enformer | Long-range non-coding regulatory prediction | Tissue and cell-type coverage in training shapes outputs |
| Nucleotide Transformer | Multi-task genome foundation model | Evaluation conventions for the modality are still maturing |
| GPN-MSA | Alignment-aware coding and noncoding variant scoring | Variant scores still need clinical-framework interpretation |
| DeepSEA | Single-nucleotide chromatin feature impact | Older; limited context window |
| Orthrus | Mature RNA property representation | RNA representation does not settle delivery or therapeutic behavior |
| RhoFold+ | RNA 3D structure prediction | RNA structure, not general genome interpretation |
What is theoretical?
Theoretical capability includes genome-editing design that forecasts regulatory, transcriptomic, proteomic, and phenotypic outcomes before experiments. The causal path from sequence to phenotype remains too context-rich for routine certainty. Multi-modal foundation models that integrate genome, transcriptome, and protein measurements in one representation are an active research direction with partial demonstrations.
What is beyond current capability?
Beyond current capabilities includes whole-organism phenotype prediction from raw genome sequence alone. Development, environment, epigenetics, microbiome, and measurement context prevent that claim. Reliable clinical action on most non-coding variants outside well-characterised regulatory regions also remains beyond current capabilities.
What would make this more promising?
Genome models become more promising with perturbational validation across organisms, tissues, cell states, genome builds, and assay modalities rather than only held-out sequence tracks. Stronger evidence would show that model-ranked regulatory variants, RNA constructs, or generated genes produce the predicted molecular effects under MPRA, CRISPRi, eQTL, RNA, or organism-specific assays. Clinical claims would need calibrated performance inside an accepted variant-interpretation framework, not a standalone sequence score.
What should researchers, biotech teams, funders, and program leaders do with this?
- Track organism, genome build, cell type, assay, and context window for every genome-model claim.
- Separate coding-variant interpretation from non-coding regulatory interpretation; the evidence standards differ.
- Use perturbational validation (MPRA, deep mutational scanning, CRISPRi) for proposed regulatory edits.
- Keep RNA structure, RNA expression, and RNA therapeutic design as related but distinct tasks.
- Treat peer review and independent reproduction as separate gates. Evo 2 is now peer reviewed; independent reproduction across organism families remains the next credibility test.
- For clinical variant interpretation, integrate genome-model outputs as evidence inside the ACMG/AMP framework rather than as classifications.