AI Just Learned to Read Human DNA

An AI trained on trillions of DNA base pairs has just cracked the code that stumped researchers for decades: reading complex human genomes. What was deemed "impossible" less than a year ago is now open-source reality.

From Bacterial Specialist to Universal Decoder

Late 2025 saw the debut of Evo, an AI that could only decode bacterial genomes. The scientific consensus was clear: "This approach won't work with more complex genomes." Bacteria cluster related genes together—humans don't. Our genetic architecture is messy, scattered, and notoriously difficult to interpret.

The team behind Evo apparently took that as a dare. Today's Evo 2 has been trained on genomes from all three domains of life: bacteria, archaea, and eukaryotes. The result? An AI that can spot regulatory DNA sequences and splice sites that even seasoned geneticists struggle to identify.

What This Actually Means for You

This isn't just about reading genetic code—it's about prediction and creation. Evo 2 can look at a cluster of related genes and suggest what comes next, or propose entirely novel proteins. For drug development, that's potentially revolutionary. Instead of years of trial-and-error protein design, researchers might soon have AI-generated candidates ready for testing.

Biotech investors are paying attention. Companies like Moderna and Ginkgo Bioworks have already pivoted toward AI-driven approaches. With Evo 2 being open-source, smaller players now have access to tools that were previously exclusive to well-funded labs.

The Skeptics Have Valid Points

Not everyone's celebrating. Geneticists worry about AI "hallucinations"—when the system confidently proposes proteins that don't actually work or, worse, could be harmful. There's also the black-box problem: Evo 2 can identify genetic patterns, but it can't always explain why they matter.

Regulators face a dilemma too. How do you approve therapies based on AI-designed proteins when the AI itself can't fully explain its reasoning? The FDA's current framework wasn't built for this scenario.