What if an artificial intelligence could understand DNA better than scientists… without ever being taught what a disease is?
That’s exactly what researchers claim with Evo 2, the largest biological AI model ever created. And its capabilities go far beyond analysis. This system doesn’t just read the code of life—it can write it.
The essential takeaway: Evo 2, a new biological AI model published in Nature, can predict cancer-causing mutations with over 90% accuracy without ever being trained on medical data. Even more striking, it can generate entirely new functional DNA sequences, including synthetic viruses capable of killing antibiotic-resistant bacteria—marking a turning point in how humanity reads and writes the code of life.
An AI That Learned Biology Without Being Taught
Evo 2 is based on the same core idea as modern language models like ChatGPT or Claude. Instead of learning human language, however, it was trained on the language of DNA.
DNA is essentially a sequence of four letters: A, T, C, and G. These letters encode everything about living organisms—from physical traits to disease risks.
To train Evo 2, researchers compiled genetic sequences from more than 128,000 species, totaling around 9.3 trillion DNA letters. The model was trained using over 2,000 NVIDIA GPUs for months and ultimately reached 40 billion parameters.
But here’s the twist: no labels were provided.
The AI was never told what causes disease, what a gene does, or which mutations are dangerous. It simply read raw DNA sequences—and figured everything out on its own.
How Evo 2 “Understands” Life?
The key lies in evolution itself.
All the DNA sequences in the training data come from living organisms. Over billions of years, natural selection has filtered out harmful mutations while preserving functional ones.
This means the dataset is implicitly labeled by nature.
By analyzing patterns across species, Evo 2 learned to identify:
- Sequences that are essential for survival
- Mutations that disrupt biological processes
- Variations that are harmless
In short, it learned the grammar of life without any explicit instruction.
90% Accuracy on Cancer Mutations—Without Medical Training
One of the most striking tests involved the BRCA1 gene, known for its link to breast and ovarian cancer.
The challenge is massive: every human carries thousands of genetic variations, and identifying which ones are dangerous is extremely difficult—even for experts.
Evo 2, without ever being trained on medical datasets, achieved over 90% accuracy in distinguishing between harmless mutations and potentially deadly ones.
In some cases, it even matched or outperformed specialized tools designed specifically for this task.
And unlike traditional systems, it doesn’t need to be retrained for each gene. It generalizes.
Seeing What Other Models Miss
Evo 2 can process up to 1 million DNA letters at once, giving it a long-range understanding of genetic interactions.
This matters because in biology, a gene can be influenced by elements located hundreds of thousands of letters away.
In tests, the model successfully:
- Detected a hidden 100-letter sequence inside 1 million random letters
- Identified subtle mutations that break protein production
- Adapted to organisms with alternative genetic rules—without prior knowledge
It even decoded the genome of the woolly mammoth, a species it had never seen before.
From Reading DNA… to Writing It
This is where things take a dramatic turn.
Evo 2 doesn’t just analyze DNA—it can generate entirely new sequences from scratch.
Researchers asked it to complete segments of human mitochondrial DNA. It produced hundreds of unique sequences, each containing all the necessary components for biological function.
These included:
- Instructions for building proteins
- Functional biological machinery
- Coherent structural organization
Even more impressively, protein structures derived from these sequences were validated using external tools and showed correct folding behavior.
The model also generated:
- A full bacterial genome (~580,000 DNA letters)
- Functional sequences in more complex organisms like yeast
This marks a major leap: AI-generated DNA that works in real biological systems.
Synthetic Viruses That Kill Antibiotic-Resistant Bacteria
Perhaps the most impactful application involves bacteriophages—viruses that infect bacteria but not humans.
Researchers used Evo 2 to design new synthetic phages targeting antibiotic-resistant bacteria, a global health threat responsible for over a million deaths each year.
Out of 285 AI-generated candidates:
- 16 successfully infected and killed target bacteria
- Some outperformed natural phages
- One replicated 65 times more efficiently than its initial population
This is the first time in history that an AI has designed functional biological organisms validated in the lab.
The Risk Nobody Can Ignore
With such power comes an unavoidable question: what prevents misuse?
To reduce risks, researchers excluded all dangerous pathogens from the training data. As a result, Evo 2 performs poorly when asked to generate harmful viruses—it produces incoherent sequences.
But the situation is far from reassuring.
The model is fully open-source. Its code, data, and weights are publicly available.
Already, independent teams have built faster derivatives that incorporate additional biological complexity.
In theory, someone with enough resources could retrain a similar system with dangerous datasets.
No clear answer exists yet for how to manage this risk.
A New Era for Medicine, Agriculture, and Biology
Despite the concerns, the potential benefits are enormous.
In medicine, Evo 2 could:
- Instantly evaluate genetic mutations
- Accelerate disease research
- Enable personalized diagnostics
In agriculture, it could help design crops that are:
- More resilient to climate change
- More nutritious
- More productive
And in science, it represents a foundational shift—from reading DNA to understanding and engineering it.
One of the study’s co-authors compares Evo 2 to an operating system kernel: a base layer upon which countless specialized applications can be built.
We Are No Longer Just Reading Life
The Human Genome Project allowed us to decode DNA.
Evo 2 suggests something far more profound: we are beginning to understand—and rewrite—the language of life itself.
Whether this leads to medical breakthroughs or new risks will depend on how this technology is handled.
But one thing is certain: biology has officially entered the AI era.
