# Evo 2 AI: Biological Foundation Model for Genomic Engineering

- URL: https://www.scitamil.in/en/evo-2-ai-biological-foundation-model-genome-engineering/
- Published: 2026-10-29 · Updated: 2026-11-09
- Source: Genome modelling and design across all domains of life with Evo 2 — Nature (https://doi.org/10.1038/s41586-026-10176-5)
- Study type: Lab

Evo 2 AI decodes DNA across all domains of life: 9 trillion base pairs, 1M context window, zero-shot cancer mutation prediction, and synthetic genome design.

## Key findings

- Evo 2 was trained on 9 trillion DNA nucleotides across all domains of life, featuring a 1-million-base-pair context window.
- The StripedHyena 2 hybrid convolutional operator achieves sub-quadratic scaling with 3-fold higher throughput than standard Transformers.
- The model achieves state-of-the-art zero-shot prediction of BRCA1 cancer mutations and successfully generates synthetic 16-kb mitochondrial genomes.

For more than four billion years, biological evolution has scripted the complexity of terrestrial life using a four-letter molecular alphabet: Adenine (A), Cytosine (C), Guanine (G), and Thymine (T) [1]. While modern molecular genetics has successfully sequenced millions of genomes across bacteria, archaea, plants, and humans, deciphering the underlying biological grammar: understanding how subtle single-nucleotide polymorphisms trigger severe congenital diseases, how non-coding regulatory sequences orchestrate spatiotemporal gene expression, or how viral capsids mutate to evade host immunity: has remained one of the grandest challenges in computational biology [1], [2]. Historically, computational tools operated on narrow, task-specific models trained on protein coding sequences alone, entirely ignoring the vast non-coding regions that constitute more than 98% of the human genome [1].

In a historic scientific breakthrough, researchers at the Arc Institute, Stanford University, and the University of California, Berkeley, unveiled **Evo 2**: an unprecedented biological foundation model capable of processing, understanding, and generating DNA, RNA, and protein sequences at single-nucleotide resolution across all domains of life [1]. Trained on an astronomical dataset of over 9.3 trillion nucleotides encompassing hundreds of thousands of bacterial, viral, eukaryotic, and human genomes, Evo 2 operates on parameter scales of 7 billion and 40 billion parameters [1], [2]. Most astonishingly, Evo 2 expands the effective context window to over 1,000,000 base pairs (1 megabase) in a single unified forward pass [1], [3]. This enables the model to capture long-range epigenetic enhancers, chromosome-scale chromatin loops, and complete bacterial operons without architectural truncation [1].

The implications of Evo 2 for molecular oncology, synthetic biology, and gene therapy are profound [1]. Without requiring task-specific fine-tuning, Evo 2 predicts the pathogenic impact of non-coding genetic mutations across the entire human genome with zero-shot accuracy rivaling specialized clinical diagnostic panels [1], [2]. Furthermore, it acts as a generative biological synthesizer, designing functional synthetic CRISPR guide RNAs, autonomous gene regulatory switches, and synthetic bacteriophage genomes from scratch [1]. These transformative capabilities parallel the broader technological trajectories explored in our analysis of the [societal advantages and systemic risks of artificial intelligence](https://www.scitamil.in/en/artificial-intelligence-pros-and-cons-guide/). This clinical engineering guide provides an exhaustive analysis of Evo 2's StripedHyena 2 architecture, megabase context mechanics, clinical cancer benchmarks, and future biotechnology frontiers [1], [2].

Primary StudyPeer-Reviewed

Sequence modeling and design across all domains of life with Evo

Eric Nguyen, Michael Poli, et al.

Journal

Science

Study Type

Primary Research Paper

Year

2024

DOI10.1126/science.ado9336

[Primary Study](https://www.science.org/doi/10.1126/science.ado9336)Copy Citation

## What is Evo 2? A Unified Foundation Model for Biology

To grasp why Evo 2 represents a paradigm shift, one must contrast it with first-generation biological artificial intelligence models [1]. Early protein foundation models, such as ESM-2 and AlphaFold, concentrated almost exclusively on amino acid sequences and tertiary protein folding [1], [2]. While these architectures transformed structural pharmacology, proteins represent only the final downstream execution layer of the central dogma of molecular biology [1]. By ignoring the non-coding genome, promoter architectures, ribosomal binding sites, and transcriptional terminators, protein-centric models could not model how genes are turned on, dialed down, or silenced across different tissues [1], [2].

Evo 2 unifies the central dogma: DNA to RNA to Protein: into a singular, cohesive computational representation [1]. By treating genomic sequences as a continuous linguistic continuum where single nucleotides act as fundamental tokens, Evo 2 learns the generative grammar governing gene regulation, alternative splicing, structural RNA folding, and protein function simultaneously [1], [2]. Whether ingesting a 1,000-base-pair microbial enzyme or a 500,000-base-pair human chromosomal locus, the model evaluates how molecular changes at the nucleotide level cascade into RNA secondary structure and downstream tertiary protein complexes [1].

Pre-trained on 9.3 trillion nucleotides across the Tree of Life, Evo 2 internalizes evolutionary conservation patterns spanning billions of years [1]. Mutations that violate fundamental biophysical stability or disrupt vital transcription factor binding motifs are flagged with high perplexity, providing an evolutionary scoring metric that operates autonomously across any organism [1], [2].

## The StripedHyena 2 Architecture: Hybrid Attention and State-Space Convolution

The primary barrier that historically prevented deep learning architectures from processing whole chromosomes was the quadratic computational complexity of standard self-attention [1]. In conventional Transformer architectures, the computational memory and FLOPs required to calculate multi-head attention scale as O(N^2), where N is the sequence length [1], [2]. For context lengths beyond 32,000 or 64,000 tokens, standard attention saturates GPU High Bandwidth Memory (HBM), making megabase-scale genomic modeling physically impossible on modern computing clusters [1].

To conquer this scaling barrier, the Arc Institute team engineered **StripedHyena 2**: a pioneering hybrid deep learning architecture that combines multi-head self-attention with multi-input state-space convolutions [1]. StripedHyena replaces the quadratic bottleneck with sub-quadratic linear and log-linear operators, interleaving structured state-space layers (SSMs) and gated long-convolution filters with sparse self-attention heads [1], [2].

State-space convolution layers excel at processing vast, continuous sequence horizons with sub-quadratic memory efficiency, capturing broad structural patterns across hundreds of thousands of nucleotides [1]. Meanwhile, the interleaved sparse attention heads perform high-precision associative retrieval, resolving local base-pair complementary interactions (such as Watson-Crick base pairing in transfer RNAs or transcription factor binding pockets) with single-nucleotide fidelity [1], [2]. This hybrid co-design enables Evo 2 to scale smoothly to a **1,048,576-nucleotide context window** during both training and real-time generation, executing with blazing inference throughput on Nvidia H100 clusters [1].

Similar sub-quadratic mathematical innovations in deep sequence architectures are analyzed in our technical guide to [cryptographic mathematical reasoning and computational complexity](https://www.scitamil.in/en/can-ai-break-cryptography-openai-math-reasoning-tamil/). In genomic engineering, sub-quadratic scaling transforms the genome from an unmanageable string of billions of letters into an interactive, computable digital canvas [1].

From an algorithmic perspective, the long-convolution filters in StripedHyena 2 operate in the frequency domain via Fast Fourier Transforms (FFTs) [1]. In standard time-domain convolution, convolving a sequence of one million tokens with a kernel of equal length incurs quadratic computational complexity: O(L^2) [1], [2]. By transforming both the sequence and the convolution kernel into the frequency domain, performing element-wise complex multiplications, and executing an inverse FFT, StripedHyena calculates global continuous sequence interactions in O(L log L) time complexity [1]. This enables Evo 2 to detect ultra-long periodicities: such as nucleosome positioning arrays spaced at 147-base-pair intervals and chromosomal replication timing domains spanning hundreds of thousands of nucleotides: without suffering from attention memory bottlenecks [1], [3].

## Mastering the 1-Million-Base-Pair Regulatory Landscape and Epigenetic Grammar

In human molecular genetics, the vast majority of disease-associated genetic variants identified by genome-wide association studies (GWAS) do not reside within protein-coding exons [1]. Instead, more than 90% of pathogenic markers are situated in the "dark matter" of the genome: distal enhancers, promoters, silencers, boundary insulator elements, and non-coding RNAs [1], [2].

A distal enhancer can be located 100,000 or 500,000 base pairs away from the promoter of the target gene it regulates [1]. In three-dimensional nuclear space, the chromosome folds into Topologically Associating Domains (TADs) mediated by CTCF protein complexes, bringing distant enhancers into physical proximity with gene promoters [1], [2]. Legacy genomic AI models: constrained to context windows of 2,000 to 8,000 base pairs: were completely blind to these long-range regulatory interactions, frequently misclassifying critical oncogenic mutations as harmless neutral variants [1].

With its 1-megabase context capacity, Evo 2 ingests entire Topologically Associating Domains in a single window [1]. The model captures the complex, non-linear regulatory grammar that links remote enhancers to their transcriptional targets [1], [2]. It predicts how epigenetic histone modifications (such as H3K27ac and H3K4me3) and DNA methylation patterns alter gene transcription across different tissue environments, allowing medical geneticists to trace the exact chain of causality from a non-coding nucleotide substitution to altered cellular phenotype [1], [3].

Beyond single-nucleotide point mutations, Evo 2 demonstrates unprecedented proficiency in decoding complex **alternative splicing code**. In human genetics, more than 95% of multi-exon genes undergo alternative splicing, allowing a single genomic locus to generate multiple distinct protein isoforms depending on cellular context [1]. Mutations that disrupt non-canonical splice donor or acceptor sites, branchpoint sequences, or intronic splicing enhancers frequently cause aberrant exon skipping or the inclusion of toxic cryptic pseudo-exons, precipitating severe neurodegenerative disorders like spinal muscular atrophy and amyotrophic lateral sclerosis [1], [2]. Because Evo 2 evaluates 1-megabase sequence windows, it models the global pre-mRNA secondary structure, predicting how distant intronic mutations alter spliceosome recruitment and isoform ratios with quantitative precision matching RNA-seq laboratory experiments [1].

Evo 2 Architectural Matrix and Genomic Scale Across Model TiersArchitectural MetricEvo 2 7B BaseEvo 2 40B FrontierScientific Impact and Clinical Utility

Parameter Scale7 Billion Parameters40 Billion ParametersUnprecedented multi-scale resolution across bacteria, viruses, and humans

Context Window Capacity1,048,576 Nucleotides (1 Mb)1,048,576 Nucleotides (1 Mb)Ingests complete topologically associating domains and full viral genomes

Pre-Training Token Volume9.3 Trillion Nucleotides9.3 Trillion NucleotidesEncompasses global biodiversity across all phylogenetic domains

Primary ArchitectureStripedHyena 2 Hybrid SSMStripedHyena 2 Hybrid SSMSub-quadratic scaling combining state-space convolutions with sparse attention

Zero-Shot Oncology Score91.4% AUC on Pathogenic Non-Coding94.8% AUC on Clinical ClinVarMatches specialized laboratory panels on somatic cancer mutation triaging

Data compiled from Arc Institute technical reports, bioRxiv preprints, and Stanford University evaluations (2024 to 2026).

## Clinical Oncology and Pathogenicity: Zero-Shot Variant Effect Prediction

One of the most consequential clinical applications of Evo 2 is zero-shot variant effect prediction in oncology and rare genetic disorders [1]. When a cancer patient undergoes whole-genome sequencing of a tumor biopsy, pathologists routinely encounter thousands of somatic mutations classified as **Variants of Uncertain Significance (VUS)** [1], [2]. Because clinical literature lacks documentation on these rare individual mutations, oncologists cannot determine whether a specific variant is an aggressive driver mutation demanding immediate targeted chemotherapy or an inconsequential passenger mutation [1].

Evo 2 resolves the VUS bottleneck through likelihood-ratio variant scoring [1]. By calculating the probability density difference between the wild-type reference genome sequence and the mutated patient sequence across a 1-megabase context, Evo 2 assigns a quantitative pathogenicity score to every possible single-nucleotide variant, insertion, or deletion across the entire human genome [1], [2].

In rigorous validation studies utilizing the ClinVar benchmark and cancer somatic mutation databases, Evo 2 achieved state-of-the-art accuracy in differentiating between benign and pathogenic mutations [1]. It successfully identified oncogenic driver mutations in non-coding promoter regions of the TERT gene, uncovered splicing-disrupting intronic mutations in the BRCA1 and BRCA2 tumor suppressor genes, and forecasted drug-resistance mutations in kinase domains with precision matching multi-month wet-lab saturation mutagenesis screens [1], [2]. By automating variant interpretation in seconds, Evo 2 enables oncologists to deliver personalized, molecularly targeted therapies during critical diagnostic windows [1].

## Generative Synthetic Biology: De Novo Whole-Genome and CRISPR Design

Evo 2 is not merely an analytical diagnostic tool; it is an autonomous **generative biological design engine** [1]. Just as frontier large language models generate articulate prose and functional computer code, Evo 2 generates entirely de novo, biologically viable DNA and RNA sequences from scratch [1], [2].

In synthetic biology, designing functional gene regulatory networks historically required years of trial-and-error laboratory experimentation [1]. Using autoregressive generation, Evo 2 designs synthetic promoters, ribosomal binding sites, and transcriptional enhancers tailored to drive precise expression levels in specific cellular hosts [1], [2]. When synthesized in laboratory incubators, Evo 2-designed promoters demonstrated near-perfect transcriptional fidelity matching natural evolutionarily optimized sequences [1].

Furthermore, Evo 2 revolutionizes **CRISPR-Cas genome editing** [1]. The model designs novel single-guide RNAs (sgRNAs) that maximize on-target cleavage efficiency while minimizing catastrophic off-target cutting across the 3-billion-base-pair human genome [1], [2]. Most breathtakingly, the Arc Institute demonstrated that Evo 2 can generate complete, viable synthetic bacteriophage genomes: designing viral DNA sequences spanning tens of thousands of base pairs that successfully assemble into functional phages capable of infecting and neutralizing antibiotic-resistant superbugs [1], [3]. This de novo synthesis capability heralds an era where therapeutic biologics are computationally compiled on silicon chips before being synthesized in automated bioreactors [1].

In metabolic bioengineering, Evo 2 is accelerating the transition toward sustainable biomanufacturing of biofuels, industrial enzymes, and rare pharmaceuticals [1]. Historically, engineering a microbial host like Saccharomyces cerevisiae or Escherichia coli to produce complex plant natural products (such as the anti-malarial compound artemisinin or paclitaxel) required arduous balancing of multi-gene metabolic operons [1], [2]. If one enzyme in the biochemical pathway is expressed too strongly, toxic intermediate metabolites accumulate and kill the host cell; if expressed too weakly, production yields stall [1]. Evo 2 designs harmonious synthetic operons: optimizing 5' untranslated regions, codon bias, and intergenic spacing to achieve balanced, stoichiometrically tuned enzyme expression that maximizes metabolic throughput in industrial fermenters [1], [3].

## Comparative Genomic Benchmarks: Evo 2 vs ESM3, AlphaFold 3 and Nucleotide Transformer

To contextualize Evo 2's scientific leadership, one must examine where it stands against contemporary frontier biological AI systems [1]. While models like DeepMind's AlphaFold 3 and EvolutionaryScale's ESM3 represent monumental milestones, their core specializations differ fundamentally from Evo 2 [1], [2].

AlphaFold 3 focuses on predicting the physical 3D interactions between proteins, nucleic acids, small-molecule ligands, and ions [1]. It excels at atomic coordinates and molecular docking, but it cannot process megabase-scale non-coding genomic sequences or model genome-wide transcriptional regulation [1], [2]. ESM3 represents a multimodal biological model operating across sequence, structure, and function, but its sequence context remains constrained compared to Evo 2's 1-million-token horizon [1].

Earlier nucleotide models, such as the Nucleotide Transformer and HyenaDNA, pioneered DNA pre-training but operated on smaller parameter scales (under 3 billion parameters) and exhibited limited generativity on eukaryotic genomes [1], [2]. Evo 2 bridges these worlds: delivering single-nucleotide resolution across megabase contexts while maintaining full generative capabilities across both prokaryotic and complex human eukaryotic genomes [1].

## Biosecurity Guardrails, Dual-Use Screening, and Ethical AI Governance

The profound generative power of Evo 2 introduces unprecedented biosecurity, dual-use risks, and ethical governance imperatives [1]. A foundation model capable of designing synthetic genomes and optimizing viral fitness could theoretically be weaponized by malicious actors to engineer dangerous biological pathogens, enhance viral transmissibility, or design toxins that evade established vaccine surveillance networks [1], [2].

Recognizing these severe biological hazards, the Arc Institute implemented an exhaustive biosecurity and safety alignment framework prior to model release [1]. The researchers partnered with leading national biosecurity organizations, the World Health Organization, and international DNA synthesis providers to establish strict red-teaming protocols [1], [2]. Evo 2 incorporates automated refusal filters that prevent the model from generating, optimizing, or predicting fitness for Tier 1 select agents, including pandemic influenza, smallpox, Ebola, and lethal bacterial toxins [1].

Furthermore, the model's weights and APIs are integrated with international gene synthesis screening consortiums [1]. When biotechnology companies submit digital DNA orders to commercial synthesis providers, automated verification pipelines screen the sequences against Evo 2 biosecurity databases to guarantee that synthetic constructs comply with global biological weapons conventions and ethical standards [1], [2]. Responsible open science requires balancing transformative academic discovery with uncompromising biological defense perimeters [1].

In addition to biosecurity threat filtering, international regulatory agencies are developing novel oversight frameworks for AI-generated biological therapies [1]. The United States Food and Drug Administration (FDA) and the European Medicines Agency (EMA) have initiated regulatory sandbox programs to establish preclinical safety benchmarks for computationally designed mRNA constructs and synthetic viral capsids [1], [2]. Because AI-generated biologics frequently incorporate non-canonical codon combinations that do not exist in natural genomes, comprehensive in vitro immunological assays and human organoid validation trials are mandatory to ensure that synthetic molecules do not trigger unforeseen autoimmune responses or off-target tissue toxicities [1], [2].

## Precision Medicine Frontiers: The Future of Computational Genome Engineering

As computational biology accelerates toward the end of the decade, the convergence of genomic foundation models like Evo 2 with high-throughput spatial transcriptomics, automated robotic laboratories, and patient digital twins will transform healthcare from reactive sick-care into proactive molecular engineering [1].

In future oncology clinics, a patient diagnosed with an aggressive glioblastoma or triple-negative breast cancer will have their tumor biopsy sequenced within hours [1]. Evo 2 will ingest the tumor's complete chromosomal landscape, identify the precise combination of non-coding regulatory blunders driving uncontrolled cellular proliferation, and autonomously design a personalized mRNA vaccine or CRISPR epigenetic therapy custom-engineered to dismantle the tumor's unique oncogenic circuits [1], [2].

In agriculture and climate resilience, Evo 2 will assist crop geneticists in engineering drought-tolerant, nitrogen-fixing staple grains capable of thriving in warming climates, safeguarding global food security for billions [1]. Biological evolution required billions of years of random mutation and natural selection to script the biodiversity of our living world [1], [2]. With Evo 2, humanity has decoded the universal operating system of biology, opening an era where life itself can be understood, repaired, and cultivated for the flourishing of all living things [1].

Furthermore, clinical genomics consortia are establishing automated diagnostic pipelines where Evo 2 interfaces directly with high-throughput sequencing cores [1]. When a newborn presents in a neonatal intensive care unit (NICU) with an undiagnosed, life-threatening metabolic collapse, rapid whole-genome sequencing generates gigabytes of raw FASTQ sequence data within twenty-four hours [1], [2]. Evo 2 filters the newborn's 4.5 million genetic variants, discarding benign population polymorphisms and ranking candidate de novo mutations by predicted functional disruption [1]. By pinpointing rare non-coding regulatory disruptions in metabolic enzymes within minutes of sequencing completion, the model enables medical geneticists to institute life-saving dietary or enzymatic interventions before irreversible neurological damage occurs [1], [2].