Evo 2 AI Model Generates Full Microbial Genome

Arc Institute has published Evo 2, an AI model trained on 9.3 trillion DNA nucleotides from over 128,000 complete microbial genomes, in a landmark Nature paper released March 9, 2026. Evo 2 can generate entirely novel, functional microbial genomes from partial sequences and identify disease-causing genetic mutations with over 90% accuracy—capabilities that suggest biology itself has become an information system susceptible to generative AI modeling.

Evo 2: Architecture and Capabilities

Evo 2 is built on the StripedHyena 2 architecture, an efficient transformer variant optimized for long sequences. The model processes up to 1 million DNA nucleotides at once—roughly the length of a small bacterial genome—compared to Evo 1's 131,000 nucleotide context window. This 8x expansion in context length enables the model to capture long-range genomic structure and regulatory relationships.

Training data: 9.3 trillion nucleotides from 128,000+ genomes, dwarfing Evo 1's 300 billion nucleotides. The scale increase is 30x, reflecting Arc Institute's commitment to frontier-scale biological AI.

The model's capabilities are striking:

• Genome generation: Given a partial microbial genome, Evo 2 can complete the sequence, generating plausible, functional bacterial genomes at high accuracy. Arc tested this by withholding genomic regions and comparing model-generated sequences to actual genomes.
• Disease variant prediction: Given a gene sequence and variant information, Evo 2 classifies pathogenic variants with 90%+ accuracy. Arc demonstrated this on BRCA1, the breast cancer susceptibility gene, where Evo 2 outperformed existing variant scoring tools.
• Evolutionary analysis: Evo 2 can infer evolutionary relationships between species by analyzing genomic sequences, suggesting the model has learned universal principles of biological evolution.

Implications for Synthetic Biology

Evo 2 represents a fundamental shift in how biologists approach synthetic biology. Historically, designing new organisms required deep experimental knowledge: understanding which genes to include, how to optimize regulatory regions, what traits would emerge from specific genetic combinations. Evo 2 collapses this knowledge into a learned distribution over plausible genomes.

Biologists can now use Evo 2 as a design tool: specify a desired trait or function (e.g., "produce metabolite X") and ask Evo 2 to generate candidate microbial genomes, then test the most promising candidates in the lab. This shifts synthetic biology from hypothesis-driven design to data-driven exploration.

The economic implications are profound. Synthetic biology—genetically engineering microorganisms to produce pharmaceuticals, biofuels, or specialty chemicals—is a billion-dollar industry. Tools that accelerate design cycles reduce time-to-market and R&D costs. Evo 2 could compress what once took months of experimental work into days of computational design.

Disease Prediction and Personalized Medicine

Evo 2's 90%+ accuracy on disease variant classification has immediate applications in medical genetics. When a patient undergoes genomic sequencing and a rare genetic variant is discovered, clinicians currently face uncertainty: is this variant pathogenic or benign? Evo 2 can answer this question with higher confidence than existing tools.

This capability enables earlier disease detection, more accurate genetic counseling, and more informed decisions about preventive treatment. For complex diseases with polygenic risk (multiple genes contributing small effects), Evo 2 can model how variants interact to determine disease risk.

Open Science Model

Arc Institute released Evo 2 weights and code openly, enabling the global research community to build on the work. This open-science approach contrasts with proprietary AI models from OpenAI and Google, which restrict model access. Evo 2's open release means biologists worldwide can integrate the model into their pipelines, accelerating innovation in synthetic biology, evolutionary analysis, and medical genetics.

The decision to open-source reflects Arc Institute's vision of AI as enabling infrastructure for biology—a public good rather than proprietary advantage. This positioning may provide Arc with long-term credibility in the academic biology community, even if it sacrifices near-term commercial opportunity.

Broader Context: Biology as Information

Evo 2 joins a growing set of tools suggesting that biology has fundamentally become an information science. The genome is a sequence, proteins are structures that can be folded computationally, cellular interactions are networks amenable to graph neural networks. As AI models scale up with more biological data, the boundary between experimental biology and computational prediction blurs.

This shift has profound implications. Biology labs can become leaner—fewer wet-lab experiments, more computational design. Biology careers will increasingly require computational literacy. The pace of biological discovery will accelerate as AI handles routine design and analysis tasks, freeing biologists for high-level conceptual work.

What This Means for AI Engineers

Evo 2 demonstrates that AI engineering is no longer confined to software. The frontier of AI applications lies in transforming entire scientific disciplines through models like Evo that bridge computational and physical domains. For engineers interested in biology, materials science, drug discovery, or synthetic biology, this moment represents an extraordinary opportunity.

If you're preparing for roles at biotech companies, pharmaceutical firms, or synthetic biology startups integrating frontier AI, InterviewAlly helps you practice the systems design and domain-specific thinking that distinguishes engineers capable of bridging AI and biology.